On Entrepreneurial Learning, Mentoring, and the Logic of ...

192
On Entrepreneurial Learning, Mentoring, and the Logic of Bayes Dissertation Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University By William Robert Stromeyer, B.F.A., M.B.A Graduate Program in Business Administration The Ohio State University 2016 Dissertation Committee: Sharon A. Alvarez, Co-Advisor Raymond A. Noe, Co-Advisor Benjamin A. Campbell Robert B. Lount

Transcript of On Entrepreneurial Learning, Mentoring, and the Logic of ...

1

On Entrepreneurial Learning, Mentoring, and the Logic of Bayes

Dissertation

Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the

Graduate School of The Ohio State University

By

William Robert Stromeyer, B.F.A., M.B.A

Graduate Program in Business Administration

The Ohio State University

2016

Dissertation Committee:

Sharon A. Alvarez, Co-Advisor

Raymond A. Noe, Co-Advisor

Benjamin A. Campbell

Robert B. Lount

2

Copyright by

William Robert Stromeyer

2016

ii

Abstract

This dissertation is comprised of three essays that examine entrepreneurial learning,

entrepreneurial mentoring, and the logic of Bayes and Bayesian analysis. The first essay delves

into the cognitive mechanisms involved in learning under fundamental uncertainty by

entrepreneurs engaged in the process of forming new opportunities. An examination of the

emergence of the pet health insurance marketplace in the United States during the period 2002-

2012 drives a qualitative analysis that integrates propositions concerning the entrepreneurial

process with theoretical assertions from the hierarchical Bayesian theory of learning. The second

essay examines how entrepreneurial career mentoring, mentoring in support of a transition to

entrepreneurial employment, leads to increased entrepreneurial intentions mediated by

entrepreneurial self-efficacy. The final essay provides a commentary and suggestions for best

usage of new techniques developed in Bayesian structural equation modeling, through a Bayesian

based analysis of entrepreneurial self-efficacy.

iii

Acknowledgements

This dissertation would not have been possible without the loving support of my family. My

deepest gratitude to my advisor and dearest friend, Sharon Alvarez. Thank you for guiding me on

this journey, letting me make my own mistake, but always putting me back on the right path. I am

also grateful for the support and encouragement of the rest of my committee. Thank you for your

insights, perseverance, and guidance as I pursued this body of research. Finally, I wish to thank

my fellow PhD students, all the wonderful members of the management department, and the

people of Fisher College for supporting a nurturing, but academically rigorous environment.

iv

Vita

2006 ..................................................................... B.F.A., Rochester Institute of Technology

2007 ..................................................................... M.B.A., Rochester Institute of Technology

2010 to present ..................................................... Graduate Teaching and Research Assistant,

Department of Management & HR, The Ohio

State University

Publications

Stromeyer, W. R., Miller, J. W., Sriramachandramurthy, R., & DeMartino, R. (2015). The

Prowess and Pitfalls of Bayesian Structural Equation Modeling Important Considerations

for Management Research. Journal of Management, 41(2), 491-520.

Miller, J. W., Stromeyer, W. R., & Schwieterman, M. A. (2013). Extensions of the Johnson-

Neyman technique to linear models with curvilinear effects: Derivations and analytical

tools. Multivariate Behavioral Research, 48(2), 267-300.

Stromeyer, W.R., & Barney, J. (2012). Cost-Benefit Analysis. In D. Teece and M. Augier (Eds.)

The Palgrave Encyclopedia of Strategic Management.

Fields of Study

Major Field: Business Administration

Focus: Entrepreneurship

Minor Field: Quantitative Psychology – Judgement & Decision Making

v

Table of Contents

Abstract ............................................................................................................................... ii

Acknowledgments.............................................................................................................. iii

Vita..................................................................................................................................... iv

List of Tables ..................................................................................................................... vi

List of Figures ................................................................................................................... vii

Chapter 1: Entrepreneurial Learning................................................................................... 1

Chapter 2: Entrepreneurial Mentoring .............................................................................. 68

Chapter 3: Bayesian SEM ................................................................................................. 99

References ....................................................................................................................... 148

Appendix A: Prior History in the Pet Health Insurance Market ..................................... 161

Appendix B: In-Depth Timeline of Pet Health Insurance (1977-2012) .......................... 174

Appendix C: SRMR & pseudo-SRMR (pSRMR) .......................................................... 181

Appendix D: 𝚯𝛿 Matrix Estimation ................................................................................ 183

vi

List of Tables

Table 1.1: Risk, ambiguity, & uncertainty ........................................................................ 64

Table 1.2: List of four firms focused on in this study ....................................................... 65

Table 1.3: Data Sources for Study .................................................................................... 66

Table 1.4: State of pet health insurance industry as of early 2000s .................................. 67

Table 2.1: Means, standard deviations, and correlations among study variables ............. 95

Table 2.2: Invariance between calibration and validation models .................................... 96

Table 2.3: Regression parameters for model 4 ................................................................. 97

Table 2.4: Indirect effect for model 4 ............................................................................... 98

Table 3.1: Benefits and cautions for specifying informative priors ................................ 141

Table 3.2: PCS-CFA measurement model fitted using ML estimator ............................ 142

Table 3.3: Bayesian model with informative priors specified for cross-loadings. .......... 143

Table 3.4: Modified Bayesian model .............................................................................. 144

Table 3.5: Cross-Validation of Modified Bayesian model ............................................. 145

Table 3.6: Bayesian model where the 𝚯𝛿 matrix was freely estimated .......................... 146

Table 3.7: Demonstration of priors in context of structural model ................................. 147

vii

List of Figures

Figure 1.1: Iterative nature of explanation building via case based pattern ...................... 16

Figure 2.1: Model of desire and intent for entrepreneurship ............................................ 72

Figure 2.2: Path diagram with coefficients ....................................................................... 86

Figure 3.1: Factor loadings for a perfect cluster solution and a Bayesian model ........... 112

Figure 3.2: Density Plot of 𝚯𝛿 matrix for the estimated PCS-CFA model. .................... 132

Figure 3.3: Density Plot of 𝚯𝛿 matrix for complexity one model .................................. 133

Figure A.1: Adoption rate of pet health insurance .......................................................... 168

Figure A.2: Pet health insurance timeline ....................................................................... 169

1

Chapter1: Entrepreneurial Learning under Fundamental Uncertainty:

An Examination of the Pet Health Insurance Industry

Chapter Abstract

Utilizing an explanation-building case study this paper examines the implications of the

opportunity creation perspective for theorizing in the domain of entrepreneurial cognition. Based

on findings from an in-depth exploration of the pet health insurance industry in the 2000s

propositions are developed regarding the means by which entrepreneurs learn under fundamental

uncertainty. As motivated actors, entrepreneurs develop hypotheses, some of which coincide with

expectations for change in the context, and then test these in the socially-constructed marketplace.

This cycle of experimentation and feedback leads to the refinement of hypotheses and the

updating of beliefs amongst the entrepreneur, as well as within the context permitting a shift in

the social-cultural conversation. As the entrepreneurs understanding of the socially complex

context matures they transition from task-specific learning to a complex integration of multiple

forms of learning.

Introduction

There is a growing appreciation for the role of process in the field of entrepreneurship

(McMullen & Dimov, 2013). Recent work on the formation and exploitation of entrepreneurial

opportunities has focused on the iterative, enactment processes associated with opportunities

(Alvarez & Barney, 2007; Dimov, 2007). Work on resource recombination and the unique

2

application of resources in constrained environments is explicitly process oriented (Baker and

Nelson, 2006). Moreover, work on new venture creation suggests that process is an essential

consideration of nascent venture formation (Gruber, 2007; Dimov, 2011). This recent process

orientation acknowledges entrepreneurial action and learning under conditions of uncertainty as

principal mechanisms by which change is facilitated (Alvarez, Barney, Anderson, 2013).

Fundamental uncertainty, information contexts in which future states and the probability

of these states is unknowable ex-ante (DeQuech, 2006, Knight, 1921) is the backdrop under

which most entrepreneurial process unfolds (Alvarez and Barney, 2005; 2007; Dimov, 2010;

Hmieleski and Baron, 2008). Fundamental uncertainty is a powerful concept in entrepreneurship,

yet with few exceptions, researchers have tended to steer away from the theoretical implications

of this construct (DeQuech, 2000) and there is a paucity of research on learning under conditions

of fundamental uncertainty. In order to maximize the theoretical value of entrepreneurial process

research, scholars need to develop a more in-depth and robust understanding of learning under

these conditions.

However, a dominant assumption of traditional learning literature is that the macro

environment is given and the agent acts within and reacts to that environment (Osman, 2010). In

this view, with few exceptions, the actions of the agent may occur in a complex-dynamic

environment, but macro level environmental changes are attributed to exogenous forces (Funke,

2001). In contrast, a major contribution of entrepreneurial process research is that agents have the

potential to enact meaningful change in the state of the macro environment (Wood & McKinley,

2010). Further it is assumed that the change induced by these agents is undertaken with conscious

intent and some degree of foresight. Yet, it is also acknowledged that the change that emerges

during the entrepreneurial process is not fully definable ex ante (Wiltbank, Dew, Read, &

Sarasvathy, 2006). It is this fundamental paradox that guides the current study, how do

3

entrepreneurs learn under fundamental uncertainty at the same time that they are enacting

structural change given that they may not fully understand the change they are enacting? In

particular this study sets out to examine the implications of this alternative viewpoint for

understanding cognitive learning mechanisms that might underlie how entrepreneurs envision

alternatives, generate potential causal structures underlying these alternatives, and iteratively test

these understandings in the social marketplace during the enactment of opportunities

This paper investigates the cognitive mechanisms of how individuals learn under

conditions of fundamental uncertainty empirically using an in-depth explanation building study

(Eisenhardt, 1989; Tripsas & Gavetti, 2000; Walsh & Bartunek, 2011; Yin, 2009) of the

emergence of the pet health insurance industry in the U.S. during the time period 2002-2012. This

development of an integration of theoretical implications derived from entrepreneurial process

research with cognitive theory regarding the emergence of causal induction and structural form

provides insights into how entrepreneurs learn about the opportunities that they themselves are

forming. The sections that follow provide an overview of the literature and theory that influenced

the extent of this case study.

Implications of Evolutionary Realism and Creation Theory for the Study of Entrepreneurial

Action, Process, and Learning

In the last several years there has been an increasing interest in examining

entrepreneurship from a perspective that makes entrepreneurial opportunities an endogenous

consequence of entrepreneurial action. This opportunity creation perspective (Alvarez & Barney,

2007) is built on an evolutionary realist epistemology (Alvarez & Barney, 2010). This

epistemology brings together components of pragmatic realism (Peirce, 1905), with its focus on

knowledge manifested in the ideal, social construction’s (Berger & Luckmann, 1967) emphasis

on human institutions and shared knowledge, and the evolutionary perspective (Campbell, 1974).

4

These various epistemological underpinnings have implications for the creation perspective’s

assertions regarding entrepreneurial acts, entrepreneurial processes, the nature of uncertainty, and

the emergence of opportunities.

The concept of evolutionary realism emerged out of the work of Charles Pierce (1905),

William James (1907) and John Dewey (Dewey & Bentley, 1949). In particular, Pierce’s work

aimed to strike a balance between the schools of ‘idealism’ (the notion or conviction that the

source and foundation of knowledge is thought itself, i.e. the source of knowledge is the mind

itself) and realism (there is something to be reckoned with that is independent of the mind and is

not constituted by thought alone). While these forms of inquiry are as old as the study of

philosophy, they received increased examination with the work of Berger & Luckmann’s (1967)

investigation of “The Social Construction of Reality”. The central premise of this work was that

persons and groups interacting in social systems generate shared conceptions of reality that

become habituated and generate the institutionalization of knowledge. This work was

subsequently taken by others to an extreme position in which all knowledge is a product of social

agreement and that no states of reality exists outside of that which is conceptualized in a society.

Needless to say, this perspective received much push back.

In contrast Donald Campbell (1960, 1974) and others married the concept of social

construction with earlier work in realism, to develop modern concepts of evolutionary realism.

This perspective asserts that there are aspects of reality that are independent of thought, i.e. forces

such as gravity exist regardless of how or when we perceive them, but there are also many aspects

of human social life that are based on fluid, renegotiated shared meaning. The classic example of

this is the dollar bill, which clearly has a material aspect of paper and ink, but has value only so

much as a shared agreement is maintained through constant re-enactment. This conceptualization

has been made axiomatic, by some authors such as Dopfer & Potts (2004), via the assertions that

5

(1) all existences are bimodal matter-energy actualizations of ideas, (2) all existences associate,

and (3) all existences are processes.

Evolutionary realism’s acknowledgement of social processes in the formation of

knowledge has implications for how we think about markets and the entrepreneurial function.

Crucially rather than being able to postulate markets as a given reality, they are instead the

outcome of a continual process of human action, negotiation, and perception. This perspective

emphasizes the important role of social structures in guiding human interaction and behavior. In

light of the role of shared social structure, one can propose that enacted opportunities may arise

along a spectrum of contexts. With such a priori contexts lying between those that embody well

defined, shared social meaning and those settings where there is a lack of such pre-exiting social

structure. Likewise this also implies that in well-established markets there will be a body of

communally shared knowledge and understanding, but that in those as of yet emerged markets

there may be little known and much yet to create (or negotiate).

The creation perspective proposes that entrepreneurs will face a set of shared

characteristics when attempting to enact opportunities. These include the presence of fundamental

uncertainty, iterations and experimentation as primary mechanisms for generating information

and understanding, and learning by doing within a path-dependent/path-creation framework

(Alvarez, Barney, & Anderson, 2013). If we assume that the presence of a market necessitates the

co-existence of a social structure, then entrepreneurs engaged in opportunity creation processes in

contexts that lack a priori shared social meaning must somehow also be enabling the creation of

this structure. This facilitation of the creation of social structure must occur within the firm, but

more importantly it must also occur for the various related stakeholders with which the firm

interacts. This potential for change also begs the questions: do entrepreneurs know how to change

the status quo, how do entrepreneurs learn if the future they are creating is unknowable ex-ante,

6

and what do we mean by learning under fundamental uncertainty? The first step in examining

these issues is to clearly articulate what is meant by the term ‘fundamental uncertainty’.

Understanding Uncertainty

It is important to clarify what the term fundamental uncertainty means in this particular

study. While it seems that such a term should be rather straight forward there is actually a fair

amount of complexity and confusion below the surface. One of the challenges faced in studying

decision-making and learning under fundamental uncertainty is that the term ‘uncertainty’ has

historically been used to mean many different things across different fields and even within

individual fields. This has led to a situation in which scholars who publish in the area of

‘uncertainty’ may be starting from radically different assumptions and axioms. The following

section gives a brief overview of the relevant terminology and clarifies how various terms are

defined in this study. Table 1.1 (see end of chapter) provides a further breakdown of terminology

related to the concept of uncertainty and illustrates how the same terms may have different uses

and connotations even within the same field. This confounding makes it a challenge to integrate

work and clearly identify gaps in our understanding and knowledge of learning under

‘fundamental uncertainty’.

The three terms of import to this study are risk, ambiguity, and fundamental uncertainty.

In order to maintain consistency the distinctions made by Dequech (2000) amongst these three

interrelated terms are adopted. Risk is present when future events occur with measurable

probability. A clear example of this is a roll of the dice, or playing a lottery with known pay-out

and odds. This concept is the base assumption for utility maximization, most classic economic

theory, the majority of financial theory, and a large portion of the judgment and decision-making

literature. Work in these areas often examines normative or prescriptive behavior, while a

7

counter-stream examines why individuals show regular deviations from ‘optimal’ behavior (i.e.

biases and heuristics). Most of the decisions faced in life are not truly risk in a pure sense, but as a

theoretical tool and an approximation of what we face the concept certainly has validity.

In contrast ambiguity “is uncertainty about the probability, created by missing

information that is relevant and could be known” (DeQuech, 2000: pg 45). This is generally taken

to mean that an agent knows what the potential outcomes can be, but is unable to access adequate

information to make even subjective probabilistic estimates for their occurrences. Ambiguity can

be further broken down into substantive uncertainty wherein the lack of all information inhibits

the ability to make decisions with certain outcomes and procedural uncertainty wherein

limitations of the computational and cognitive abilities of the agent prevents the agent from

electing optimal choices given the available information. This complexity is akin to Simon’s

notion of bounded rationality (substantive and procedural rationality). It should be noted that both

neo-classical and main-stream economics tends to place “Knightian uncertainty” in the ambiguity

silo (DeQuech, 2006).

Fundamental uncertainty is “characterized by the possibility of creativity and structural

change and therefore by significant indeterminacy of the future” (DeQuech, 2000: pg. 48). “The

list of possible events is not predetermined or knowable ex ante, as the future is yet to be created”

(DeQuech 2006: pg.112). This form of uncertainty implies that models of normative and

prescriptive behavior may not be applicable. In the absence of fundamental uncertainty it makes

sense to behave by rule-guided conventions as they have been enshrined in the utility

maximization paradigm. However, under fundamental uncertainty unconventional acts may lead

to innovation, competitive advantage, and structural change. The field of economics has for the

most part not attempted to address this conceptualization of uncertainty. The set of axioms and

assumptions needed to generate meaningful analytical models have not been identified at this

8

time and it is likely that even if they are the resulting models will not be tractable (DeQuech,

2006). Similarly there has only been minimal inquiry into this area in the field of psychology. The

vast majority of studies completed in the field of judgment and decision-making are based on

experimental approaches. Designing experiments that both permit an open state space and

simultaneously provide control comparative conditions has proven to be exceptionally difficult.

Recently there has been some ground gained in this area (Payzan-LeNestour & Bossaerts, 2011),

but there is much work still to be done.

A major critique of the concept of fundamental uncertainty, at least from a theoretical

perspective, is the belief that fundamental uncertainty necessarily implies an ‘anything goes’

theory of behavior (Coddington, 1982; DeQuech, 2000). This reductionist, theoretical, nihilism is

only appropriate if we assume that fundamental uncertainty is synonymous with ‘total ignorance’.

While this study acknowledges that fundamental uncertainty does imply unknowable future

states, it also recognizes that the world we live in is full of constraints and enablers and thus it is

not a world of ‘anything goes’. Further actors bring a body of prior knowledge and awareness

with them when they face fundamental uncertainty, providing both tools to address this

uncertainty, but likewise biases and assumptions that may restrict the set of considered choices. In

essence this perspective asserts that there can be degrees of fundamental uncertainty, it is not a

pure binary condition of either risk or total ignorance.

Another crucial aspect that facilitates movement away from the nihilistic perspective is

the acknowledgement that opportunity creation, as embedded in creation theory, is a path-

dependent, emergent phenomenon (Arthur, 1989; Garud & Karnoe, 2001; Mintzberg & Waters,

1985). This immediately adds an important temporal component to the entrepreneurial process,

such that opportunities do not simply appear with a snap of the fingers. Rather entrepreneurs in

their experimentations iterate new understandings and new market potentials, oscillating between

9

successes and failures without always knowing why. This experimentation and exploration in

undefined, or ill-defined contexts is a crucial area for entrepreneurship literature to more deeply

address.

The fundamental nature of such inquiry gets at the heart of entrepreneurship and further

developments in these areas will clearly articulate the field’s unique contribution to the other

streams of business, economics, and psychology literature. At the same time, the challenges that

economics and psychology has faced in addressing issues of decision-making, judgment, and

learning under fundamental uncertainty highlights the difficulty faced in undertaking such a task.

A priori it is not clear what route theoretically should be pursued in addressing these issues. This

study has chosen to focus on the role of learning under fundamental uncertainty, and thus the next

section lays out a set of inter-related theoretical frameworks that shed light on the phenomenon of

learning under uncertainty. These frameworks were chosen in iteration with the data originating

from the qualitative study. In essence the stories emerging from the study of the emergence of the

pet health insurance market were used to drive questions of interest and insights into the

phenomenon, which were subsequently ruminated on and distilled into aspects needing

explanation. In tandem potential explanations and concepts from the current learning literature

were explored in an effort to seek meaningful connections with aspects emerging from the

qualitative data.

Avenues of Inquiry for Learning under Uncertainty

A significant portion of the learning literature makes the assumption that the macro

environment is a state separate from the agent; not in that the agent lies outside the macro

environment rather that the agent’s effects on the macro environment are minor enough that they

are of little consequence (Shanks, 2010). In contrast, the major contribution of the creation

10

perspective of entrepreneurial opportunities, is that agents not only have the potential to enact

meaningful change in the macro environment (i.e. the market), but that they actually do facilitate

the emergence of this change. It is this fundamental paradox that guides the current study, how

can entrepreneurs learn under fundamental uncertainty at the same time that they are enacting

structural change (particularly given that they may not understand the change they are inducing)?

Entrepreneurship opens a new avenue of exploration in that it is specifically focused on

an agent who is not only attempting to profit from structural change, but may also be a potential

source of the creativity and innovation that predicates this change. This observation in regards to

the learning literature led to the formation of a set of interrelated fundamental questions that

guided both the qualitative analysis of the pet health insurance data and the theory domains that

were explored.

The first and most significant question is “Under conditions of fundamental uncertainty,

how do entrepreneurs learn from their experiments and how are these experiments conceived?”

Closely related, the answer to such a question would address issues of what agent and

environmental aspects both constrain and enable this process of learning (lest we slip into the

domain of “anything goes”). Secondly, “Overtime, how is noisy feedback both from the

‘experiments’ and from the environment interpreted?” In conditions of uncertainty what are the

mechanism that allow entrepreneurs to extract ‘the signal from the noise.’ Finally, “How does

social context influence the learning process and how does the process influence social context?”

Substantial research addresses the first part of this question, but we have little understanding of

the later.

11

Frameworks for Understanding Learning under Uncertainty

The information-processing perspective in cognitive research focuses on those

mechanisms of cognition that influence what information is perceived, how it is processed, what

response is elicited, and how this response is enacted through behavior (Newell & Simon, 1956).

In regards to learning, two divergent streams tend to dominate this field of inquiry (Courville,

Daw, & Touretzky, 2006). One is based on knowledge-independent, statistical mechanisms of

inference that rely on underlying processes of similarity and association. These cognitive

interpretations envision learning as the testing and refinement of probabilities. Models of learning

arising from this first stream tend to be based on the rational or bounded-rational actor (Langley

& Simon, 1981; Simon, 1982) or on the mechanisms that circumvent rationality, i.e. biases and

heuristics (Tversky & Kahneman, 1974).

The other stream is task-specific, dependent on robust domain-specific knowledge, and

relies on processes driven by representations and intuitions (Tenenbaum, Griffiths, & Kemp,

2006). This cognitive perspective depicts learning arising from intuitive theories, schemas, and

knowledge structures (Markus. 1977). Models from this second stream are found primarily in

research on task-specific learning, developmental cognition, and experimental choice models

(Courville, Daw, & Touretsky, 2006; Tenenbaum, Kemp, Griffiths, & Goodman, 2011).

Information-processing based theories tend to be the preferred avenue for experimental

work and theoretical development in judgment and decision-making research, cognitive

modeling, learning studies, and computational simulation (Shanks, 2010). Such theories provide

formalized mechanisms for how information is perceived, encoded, retrieved, processed, and

outputted. This formalization permits the falsification of proposed theory through

experimentation, ease of testing proposed contingency mechanisms (i.e. such as affect, priming,

etc.), and robust specification in computer driven simulation (which is also amenable to cross-

12

validation with data from human experimentation). Although this formalization provides many

beneficial avenues, one of the stronger critiques of this line of inquiry has been its inability to

integrate the complexity of context, particularly when the task domain might be ambiguous, task

ordering is non-sequential, and information is contradictory (Clark, 2016; Griffiths &

Tenenbaum, 2009). However, recent advances in the area of developmental, cognitive learning

theory based on hierarchical Bayesian representations, supported by emerging findings from

biological neuroscience, machine learning, and artificial intelligence, are permitting an integrated

investigation of the emergence of complex, abstract human knowledge domains (i.e. areas such

as language formation, grammars, object taxonomies, the acquisition of new cause-effect

relationships) (Jacobs & Kruschke, 2010). These approaches marry aspects of the two previously

articulated, divergent streams of information processing to hypothesis the mind as a probabilistic,

computational machine capable of utilizing abstract knowledge to infer domain and task specific

attributes, while simultaneously leading to the emergence of generalization learning and the

formation of causal hypotheses.

At its heart, the hierarchical Bayesian cognitive approach to learning is based on the

fairly simple logic of Bayes rule, but this belays a rich complexity that permits approaches to

learning that are able to side-step the classic either-or dichotomy of general abstraction vs.

domain specific and nativism vs. empiricism, (Tenenbaum et al., 2011). The basic logic of Bayes

rule is that the observation of new data permits the formation of an updated belief for any

underlying hypothesis (posterior distribution), which in turn is a function of initial beliefs about

that hypothesis (the prior) and of the character and implications of the observed data (the

likelihood). Bayes rule provides the logic for inferential updating, which in of itself is useful but

hardly new. The contribution of recent advances in the cognitive understanding of development

and learning is the focus on the hypothesis space that provides the relationships that are to be

13

examined (i.e. the priors in the Bayes rule equation) (Perfors, Tenenbaum, Griffiths, & Xu, 2011).

The hypothesis space represents the range of potential hypotheses that an individual considers

when confronted with a new learning task. It dictates both what hypotheses will be examined, as

well as explains where these hypotheses originate from in relation to the individual’s existing

knowledge.

In particular the theory of hierarchical Bayesian learning has moved beyond the classic

use of Bayesian learning in which a single task or fixed set of tasks is analyzed at one level of

inference. These earlier models of learning, denoted by the term discriminative models, focus

only on the data relevant to the specific task or conditioning response at hand. Hierarchical

models focus instead on a generative approach by marrying structured knowledge, domain

insight, and statistical inference in order to explain how individuals generalize from sparse and

sometimes contradictory data to form both domain-specific inferences and higher-level abstract

understanding (Tenenbaum et al., 2006). This is accomplished by articulating the hypothesis

spaces as a multi-tiered knowledge structure. At the lowest level the hypothesis space is related to

the particular question, task, or domain-specific character of interest. Each upper level of the

hypothesis space imposes a logic based on abstraction and structured knowledge which dictates

the possible domain of the lower levels. Such a structure permits the learning of complex

knowledge entities, such as new abstract causal frameworks, new representational schema, and

categorical taxonomies. These complex entities in turn accelerate the learning of specific relations

at the domain-specific level, a feature which has come to be known colloquially as the ‘blessing

of abstraction” (Griffiths & Tenenbaum, 2009).

The mathematical nature of such models quickly becomes extremely complex,

particularly when researchers attempt to examine how such inferential mechanisms may actually

be accomplished in the brain (Gershman, Blei, & Niv, 2010). As the focus of this study is the

14

generation of a conceptualization of learning under fundamental uncertainty, this mathematical

complexity will not be directly addressed, and rather the focus will be kept on the concepts of

multiple levels of hypothesis generation, the role of prior knowledge, and the social structure

surrounding the entrepreneur. Most of the hierarchical Bayesian cognitive literature is a mix of

formal mathematical modeling and experimental data, often with the aim of examining how well

a proposed models explains the experimental data (Jones & Love, 2011).

There are limitations to the hierarchical Bayesian approach as it has been implemented to

date. Utilization of this theoretical approach in the current cognitive literature implies a learner

who is constrained by an exogenously given task. For our purposes we need to acknowledge

constraints for learning under fundamental uncertainty, lest we fall prey to ‘anything goes’, but

these constraints may not be fully explicable a priori and they may be subject to change over

time. The ad hoc, status quo challenging processes of entrepreneurial creation imply that

significant aspects of the opportunity are learned through experimentation and iteration. However

the information available in such a setting will be notoriously noisy and multiple sources will

compete for attention. This makes it challenging for the entrepreneur to generate causal

attributions or to develop a coherent logic as to the underling latent principles at play, and yet this

happens anyhow.

Implementing this shift in the boundary conditions of the theory supporting the

hierarchical Bayesian conceptualization of learning requires careful consideration to account for

the notion of fundamental uncertainty and entrepreneurial process, whilst not jettisoning the

primary logic underlying this approach. An examination of the emergence of the pet health

insurance industry was used to guide this exploration and theory integration. This qualitative

study was used as a means to explore how propositions can be put together in an iterative manner,

15

going back and forth between received theory and the data to generate a unifying concept of

entrepreneurial learning under uncertainty.

Research Method

This study used a qualitative research approach (Gephart, 2004; Stake, 2005). As

entrepreneurship is a younger field of study and many of the questions of interest are focused on

temporal processes (i.e. learning), the rich empirical data provided by qualitative methods permits

the illumination of aspects not previously recognized in current theory. Suddaby (2006)

highlights this methodology as a viable means for extending extant theory and for addressing

theory that may be incomplete in regards to complex phenomenon. Similarly when theoretical

constructs are deployed into new domains (i.e. transitioning learning theory that was developed

under the assumptions of ambiguity into the domain of fundamental uncertainty) qualitative

methods can be a viable approach for articulating relevant theoretical boundaries and imperatives

(Siggelkow, 2007).

In particular this study utilized a pattern-matching analytical technique in order to derive

a series of explanations for the phenomenon under observation (Wynn & Williams, 2012; Yin,

2009). Figure 1.1 presents a pictorial representation of the iterative process used in developing

initial propositions, comparing these against case findings, refining propositions and reiterating

the process of questioning the data for alternative explanations. “To ‘explain’ a phenomenon is to

stipulate a presumed set of casual links about it, or ‘how’ or ‘why’ something happened” (Yin,

2009: p. 141). As this technique relies heavily on narrative development, it is well suited to

embedding tests of theoretical statements or propositions within rich descriptive content

(Eisenhardt & Graebner, 2007). This permits the gradual development and refinement of a series

of ideas, while also entertaining the possibility of rival explanations and significant contingencies.

16

Figure 1.1: Iterative Nature of Explanation Building via Case Based Pattern-Matching (Adapted

from Yin, 2009: p. 143)

Data Setting & Sources

This study used data from the emergence of four firms (one primary and three

comparative) in the pet health insurance industry in the United States during the period of 2000-

2012, supplemented by validating information from several other institutions, sources, and related

parties, and the ongoing social change that surrounded this industry to examine the processes by

which entrepreneurs learn under fundamental uncertainty during the enactment of new

opportunities. Table 1.2 (see end of chapter) provides information about these four firms, which

will be referred to by the fictitious pseudo-names: Toto (primary case), Asta, Gromit, and Snowy

(comparative cases). This table identifies both when the founders began working on the initial

concept of pet health insurance as a market offering and when the founders legally created a firm

to begin the process of formalizing financing and regulatory approvals. The founders of these

four firms had varied levels of knowledge in regards to veterinary science and the practice of

veterinary medicine. Likewise, across the firms the founders had varying prior exposure to the

1) Making initial

theoretical statements or

initial propositions

2) Comparing the findings of an initial

investigation against such statements or propositions

3) Revising the statements

or propositions

4) Comparing other

contextual details of the case against the revisions

5) Comparing the revisions

to the findings of a second, third, or more cases

6) Revising the statements

or propositions, accounting

for contingencies

17

operation of an insurance entity and the creation of insurance products. Finally the table provides

the self-reported reason for why the founders first became involved with pet health insurance.

Although pet health insurance has been available in the United States since the 1980’s,

the emergence of a clearly identifiable industry did not occur until the 2000’s (some would argue

that is still not a ‘Category’). The thought that pet health insurance is not yet a recognized

‘category’ in the common vernacular or in the US insurance industry in general was a comment

that was echoed by two different CEOs and by insurance regulators. Pet health insurance is an

appropriate context in which to investigate the research questions as it is socially complex, is

embedded in changing social perceptions of how we view pets and our relationships with them,

demonstrates a spectrum of iterative approaches to implementing business models, written

documents from the process are still readily available, and is recent enough as to provide direct

access to involved actors.

In order to triangulate the data and provide greater validity, data was collected from both

within the studied firms and from external sources. Internal interviews were conducted with

founders, staff, and the immediate parties that they dealt with in forming their firm (i.e. funding

sources, underwriters, etc.). External interviews were conducted with parties that had an active

engagement with the industry, such as regulators, veterinarians, and clients. Interviews were

further supplemented with regulatory records, internal firm documents, newspaper and trade

journal archives, veterinary trade group commentaries, and other contemporaneous evidence.

Table 1.3 (see end of chapter) presents a summary of data sources used for this study.

Analytic Strategy

In order to stay true to the explanation building modality this study proceeded in stages.

While it was known that the study would be focused on the question of how entrepreneurs learn

under fundamental uncertainty, it was unclear at the start what theoretical learning frameworks

18

might guide the eventual data analysis and what level of unit of analysis would best illuminate

entrepreneurial learning mechanism. As opposed to a pure grounded-theory approach that might

look to generate de-novo theory this study was informed from the start by the assumptions of

opportunities that guide entrepreneurial action (Alvarez, Young, Wooley, 2015). The propositions

of opportunity process theory were used as guide posts in the initial exploration of the pet health

insurance industry context and to illuminate what might be relevant to understanding learning in

uncertain contexts (step 1).

This guided exploration of the pet health insurance industry revealed that there had been

significant market and social uncertainty regarding the future viability of pet health insurance as a

service in the early 2000s. As the data shows a variant of pet health insurance was already

available in the USA, but many viewed it as a failed industry, a hurdle that a new round of

entrepreneurs would have to overcome. This initial examination of the data revealed important

aspects that would have to be addressed in the data analysis including the role of prior beliefs

(both the entrepreneur and stakeholders), the process of belief updating, the formation of

alternative hypotheses, the generation of causal understanding, iterative testing to extract

information from alternatives, the differential functions of both task and abstract learning, and the

interplay between these two types of learning (step 2).

Utilizing these primary guideposts the body of cognitive learning literature was examined

and the logic of hierarchal Bayesian learning was identified as a viable theory with explanatory

power to articulate relationships amongst these components (step 3). Equipped with the dual

theories of opportunity process and hierarchal Bayesian learning, data from one firm that emerged

recently in the pet health insurance industry was examined in-depth for specific incidents in

which the founders devised new alternatives to the status quo and through iterative enactment

learned about their viability. Early in the analysis, data was organized in a chronological manner

19

for each firm in order to detect meaningful phenomenon that would articulate how entrepreneurial

learning develops and how this relates to changes in uncertainty over time (Miles & Huberman,

1994). Later, this within-firm data was subdivided along the boundaries of three identified firm

resources, while maintaining the original chronological ordering to preserve the interactions

between these resources and the meaningful context.

While the learning process is ubiquitous throughout the opportunity creation process, at

this point in the study it was discovered that the initial firm level case contained meaningful sub-

cases that could help guide the analysis by facilitating pattern-matching. In order to generate a

consistent and analyzable case-logic it was elected to treat particular business resources

developed by the focal firm as units of learning. The initial genesis, development, adaptation, and

eventual deployment of each of these resources provided a structure by which to identify the

underlying sequences of entrepreneurial learning as they progressed from initial simple

hypotheses to complex integration. These learning units were identified both via how the

entrepreneurs intuitively presented aspects of the business as distinctive domains during the

various interviews and from triangulation with firm documents and the other available evidence.

While there were certainly more learning units at play than are examined here, it was elected to

examine a subset that played a significant role in facilitating the entrepreneur’s alternative vision

for the social status quo, those that were crucial to the structure of the emerging opportunity, and

those that played the most important role in bringing other members of the social interaction into

alignment with the entrepreneur’s vision. This initial case analysis highlighted the prevalence of

fundamental uncertainty, the important role of generative causal structures, the role of noisy

feedback, and the role of constraints and enablers for examining the question of learning under

fundamental uncertainty (step 4).

20

This within-firm theory development was then challenged by examining learning

outcomes for three additional firms, in order to test the previously developed explanations against

similarities and differences across-firms. The various founders of these firms brought different

background experiences and capabilities to their ventures, and interacted with a different cast of

internal and external actors (although there was some overlap in particular domains, such as

regulators) leading to different contexts within which learning occurred. In order to support

consistency in the analysis, these three comparative firms were examined in regards to how their

founders created analogous firm resources during the opportunity formation process (step 5).

These cross-case comparisons permitted the illumination and refinement of the

theoretical propositions in regards to similarities and differences in the relevant contexts and

highlighted relevant contingencies (step 6). This re-examination of developing theoretical

propositions under varying contexts serves to strength the pattern-matching approach by

delineating when patterns hold and when and why they fall apart (Locke, 2001). This in turn leads

to the caveat that the propositions presented in this manuscript are the culmination of the entire

process and for parsimony the intermediate propositions are not articulated, excepting a section

on contingencies emerging out of step five.

As a clarification for the reader the analyses in this study are focused on the learning

processes of the firm founders. The number of primary founders ranged from one to three across

the four studied firms. For simplification purposes this study treats founders from a single firm as

a homogenous learning unit, rather than attempting to parse the specific learning of each

individual founder. This parsimonious step was taken in order to facilitate a focus on the process

of hypothesis formation and testing, the emergence of causal structures, the paradox of learning

from noisy feedback, and the influence of fundamental uncertainty without simultaneously having

to account for group/team level effects generated by individuals. As the founders of these firms

21

worked very closely together with constant engagement and the learning occurred over many

years it is believed that such individual-team level effects can be safely parceled out from the

phenomenon of interest. For ease of prose this paper will often use just the pseudo-name of the

firm as a proxy for the founder(s). It is expected that the use of the firm pseudo-names will ease

the burden of the reader who would otherwise have to keep track of the various founders’ names.

Study Landscape and Prior History

While this study is concerned with the processes of learning under fundamental

uncertainty, as examined in the emergence of a focal firm and three comparative firms within the

pet health insurance industry in the 2000s, it is important to understand some broader trends in

social, demographic, technical, and regulatory change that influenced the processes investigated.

These changes had direct implications for the social milieu that guided, constrained, enabled, and

inspired the involved entrepreneurs and those who they interacted with in resourcing their firms.

Two primary social-level changes included the changing role of the pet as a member of the family

and huge advancements in both the way that veterinary medicine was practiced and its perceived

value. While Americans have always had an abiding fascination with pets (Grier, 2006, p. 12), the

rate of change in both of these areas picked up speed throughout the eighties and nineties.

Between 1979 and 2009 the US dog population increased from an estimated 49 million to

77.5 million and likewise a similar increase occurred in the cat population. This period also

witnessed a marked increase in the amount that pet owners were willing to spend on their pets

and the degree to which they identified pets as an integral member of the family. The 1980s and

1990s were periods of dramatic upheaval in veterinary medicine; with the emergence of

veterinary specialists and the increased adoption of advanced medical techniques that in the past

were found only in the domain of human medicine. Together these changes led to a major

22

increase in the demand for advanced veterinary care, which came with major increases in the cost

of care. For those readers who are interested appendix A provides a more thorough discussion of

these issues and also provides in-depth information about the early days of pet health insurance.

This data reveals the depth and magnitude of the uncertainty (primarily social) faced by the pet

health insurance entrepreneurs examined in this study.

Summary of Prior Industry History and its Implications for this Study

In 1982, pet health insurance was introduced by Veterinary Pet Insurance (VPI) (should this

get a pseudo name as well?) into the United States, but it grew anemically up through the start of

the 2000s. While pet insurance stayed under the radar for most Americans, this time period had

important implications for the nature of the context within which a new wave of firms

subsequently co-created opportunities during the 2000s. In order to articulate these issues the

following section provides some needed background information that explains the environment

that the firms under study confronted. The initial goal for the formation of VPI was to reduce the

incident of ‘economic euthanasia’, which is when a pet owner elects to have an animal put down

rather than pay for veterinary care. For a broad range of reasons, which are articulated in more

depth in Appendix A, VPI created an insurance product that according to the founder was

“fundamentally flawed.” While VPI survived as a firm in order to enjoy the subsequent rebirth of

the industry in the 2000s, it was kept alive through repeated cash-infusions having never had a

year of positive cash flow. The major downside of the ‘flawed’ insurance products of VPI was

that it created negative sentiment for the concept of pet health insurance in the US amongst

veterinarians, regulators, insurance underwriters, and pet owners (i.e. the complete body of

significant external stakeholders).

23

The prior history of pet health insurance, for the period 1982-2001, provides an explication of

the complex context within which a new wave of entrepreneurs reimagined and reinvigorated the

industry. The industry that subsequently emerged during the 2000s looked nothing like the one

that VPI earlier brought to the market. However in order to understand the cognitive learning

processes that these new entrepreneurs engaged, we must account for the socio-cognitive

framework that was already on the ground. The unfolding of a shared consensus (Canon-Bowers

& Salas, 2001; Lee, 2001) around the failure of pet health insurance had led to mutually held

beliefs along several dimensions that had the potential to derail any new attempts at the concept

of pet health insurance. Such embedded negative associations were acting as a repository of

shared knowledge, one that was functioning as a short circuit for efforts to reinvigorate the

industry. This strongly shared negative sentiment generated substantial uncertainty for those

entrepreneurs in the 2000s who envisioned alternatives for the exiting status quo. All of the

various firm founders interviewed related a similar sentiment as captured by one in particular:

“there was really a lot of uncertainty about what might work, or how we might make it happen.”

Table 1.4 (see end of chapter) provides an overview of the conflicted socio-cultural

environment that founders engaged in the entrepreneurial process faced at the beginning of the

2000s. In particular it highlights the overall impressions that various parties had for the future of

pet health insurance, and the negative associations that had been formed. These knowledge

structures, shared understandings, had become fairly solidified over twenty years and as can be

seen appear to be rather unfavorable to the re-emergence of the pet health insurance industry.

In order for entrepreneurs to break out of the mold of how pet health insurance was

perceived at the start of the 2000s they needed to integrate several aspects of the context. Firstly

they needed to not only comprehend these negative sentiments and the reason for their existence,

but then they needed to envision alternatives. With such concepts in mind they then needed to

24

propose and tests solutions (based on perceived causal structures) in the marketplace (Alvarez &

Barney, 2007). Relying on noisy feedback, they needed to nonetheless infer results and the

reasons underlying these results; which would then guide them in further iterations. If successful

they would eventually enact a change in these prior sentiments, leading to new ones that would

consider pet health insurance as a desirable state of affairs.

The section that follows recounts the history of how one firm, the primary case (Toto),

went about this task. This single case overview is used to illuminate three firm resources that

emerged during the opportunity formation process. These three identified resources were

fundamental to the nature of the specific firm that eventually emerged and possessed the

significant benefit that variants of them were developed by each of the other three firms used in

the comparative stage of the analysis (Snowy, Asta, and Gromit). This overview of Toto is

followed by a short section introducing the three comparative firms and articulating the rationale

of each founder for initially pursuing the pet health insurance concept. For brevity’s sake a

complete history of each of the three comparative firms is not provided, but rather highlights and

insights from this additional data analysis is included in the later theoretical and proposition

development.

Toto Pet Insurance: Brief Firm History

Toto Pet Insurance had its 2002, pre-firm days start amongst a group of MBA students at

Wharton College, University of Pennsylvania. Bob and Sara (names changed), a couple from the

UK, during their studies had a cat that became very ill and required emergency veterinary care.

The cost of treatment for this animal was close to $5,000 and they had to pay for it out of pocket.

Being from the UK, where there was already a well-established pet health insurance market, they

wondered why pet health insurance in the US was relatively unknown amongst the public and

25

generally disliked by the veterinary community. With the help of Simon, a fellow MBA student,

they developed a basic business plan, which they subsequently entered in the MBA jungle

business plan challenge at Wharton. While the early plan was ambitious it was out of sync with

the regulatory structures for how insurance products can be developed and sold in the US. Further

many of the assumptions used in the plan were not in line with the reality on the ground. It had

not sunk in yet with the team how badly the pet health insurance marketplace had been damaged

by prior attempts by other firms.

While this first business plan made it through the initial competition, it was realized that

there was inadequate knowledge amongst the team concerning the insurance industry in general.

This is when the team was connected with Karen, a fellow student, who had previously worked as

an actuarial specialists for Canada Life and had twelve years of insurance industry experience.

With her help the business plan was redrafted with a focus on building a quality pet health

insurance product that was priced “using actuarial principles from the outset.” The four MBA

students and their new concept joined Wharton’s venture initiation program. As part of this

undertaking they were able to enlist the advisory board assistance of a well-known veterinarian

with ties to the professional journal Veterinary Economics, and a successful property & casualty

insurance entrepreneur. At this time they also began the lengthy and challenging process of

beginning to develop actuarial models to support the types of policies they envisioned for the

marketplace. They went on to win the Wharton business plan competition in 2003, beating out

various technology and biotechnology offerings.

Shortly after this success, disagreement amongst the team members on how to proceed

and increasing team acrimony led to the dissolution of the group. Bob and Sara left to start their

own pet health insurance firm based on licensed material from the UK. Karen and Simon struck

out on their own, spending the next year writing the business plan for their enterprise concept,

26

now called Toto Pet Health Insurance. This work involved substantial efforts to finalize dog and

cat actuarial models that supported the policy concept, pricing the new product, searching for

insurance partners that might be willing to be involved as underwrites, and traveling the funding

circles. In order to support himself and gain more insurance industry experience Simon got a job

at Progressive insurance, while Karen moved to Ohio with her family. In September, 2004 an

Ohioan non-profit development fund based in Cleveland selected Toto as one of three firms

amongst 150 for start-up support funding.

With the support of this funding Karen and Simon were able to refine the concept and in

October, 2005 they signed a letter of intent with Lloyd’s of London syndicate to provide

underwriting for Toto’s pet health insurance policies. However, it was still another year until

finally in October 10, 2006 Toto sold its first policy which covered Karen’s newly adopted cat. In

March, 2007 Toto launched its ecommerce website which provided an informative portal for

potential and current customers, as well as for veterinarians and other related parties. Based on

successful launch and growth, a venture capital firm stepped in as a partner in 2008 providing

needed capital and support. In June 2011, Toto reached break even with the continuing prospect

for strong growth.

The next section examines the development of three specific firm resources (units of

learning) and the events that surrounded them that were integral to the emergence of Toto and the

formation of the opportunity that they enacted. Narrowing the focus to these three resources

permits the development of propositions and theory regarding how entrepreneurs learn under

uncertainty. Variants of the three identified resources are also found in the stories of the three

comparative firms.

27

Resource 1: The Actuarial Model

When discussing insurance policies, it is important to know that an actuarial model lies at

the heart of insurance products. Insurance is premised on the notion that the policy holder in

return for providing a premium is entitled to compensation after the occurrence of a covered

event. An actuarial model is used to create and price insurance policies by matching the risk of

coverable events, the outgoing payment in response to event occurrence, and the inflow from

received premiums. There are many ways an actuarial model can be structured. Amongst property

and causal insurance a primary difference is found in how the underlying actuarial model

determines the payout amount and the list of covered events. With pet health insurance the

primary inputs are the morbidity rates of accidents and illnesses amongst cats and dogs, and the

cost of associated veterinary treatments and interventions. At its core the theory and structuring of

an actuarial model is not very complicated, but this simplification belays significant difficulties

and complexities in real world implementation.

Early on the founders of Toto realized that they had the opportunity to provide a new

alternative to the then currently available pet health insurance products (i.e. policies from VPI).

When VPI was initially brought to market in the 1980s the firm elected to use a schedule of

benefits actuarial model. This type of model pays out a set amount in response to a covered event,

regardless of the actual treatment provided by the veterinarian or the billing practices of the

veterinarian. As further articulated in appendix A, VPI’s choose this model because it was

implementable with the technology that they had at the time, meet their goals of keeping the cost

of the policies low, and allowed the product to be brought to the market quickly. The major

downside was that over time increasing mismatches between what VPI paid out for veterinary

care and what veterinarians actually billed led to angry customers, substantial policy / customer

turn-over, and the formation of negative sentiment towards pet health insurance.

28

Toto’s initial business plans underestimated how challenging shifting the actuarial model

underlying pet health insurance ended up being. This observation was put forth unprimed by the

founders during one of the early interviews. As Table 1.4 illustrates there was significant negative

sentiment for the concept. Several key players who needed to be involved in the process of

recreating the pet health insurance industry had seen prior missteps and were wary of or actively

avoided future involvement. The initial business concept attempted to address these issues by

envisioning a firm that would engage these various parties in a dialogue about the potential value

of insurance for the veterinary industry, the insurance industry, and the pet owner.

While the very earliest business plans of Toto do not discuss the particulars of how

policies would be developed and priced, with the addition of Karen this aspect became a

significant part of the opportunity creation process. It was realized that if the conversation about

pet health insurance was going to change the product needed to be redesigned from the ground

up. With her significant prior experience in actuarial practice, Karen proposed the concept of a

percentage of bill policy. VPI’s policies were still based on the schedule of benefits model with

its inherent limitations. Over the years VPI’s polices had become more and more complicated as

problems occurred and were resolved through add-ons that eventually led to policies where it was

essentially impossible for the customer to know how much of a claim was going to be refunded.

Toto believed that a percentage of bill policy approach would solve this problem. Such a

concept is also seen in human health insurance, often under the name of co-insurance. In this type

of setup a policyholder buys a policy that will pay a fixed percentage of a claim, such amounts

often range from 70-90%. When a veterinary bill is paid by the policyholder they submit a claim

to Toto, which then reimburses the agreed on percentage (thus the policyholder pays a co-

insurance in the amount of 10-30%). The advantages to such an approach are that it keeps claim

pay-outs current with veterinary service market-pricing and payment is automatically adapted to

29

local market conditions. This is of great benefit to the policyholder who no longer needs to worry

about how their vet prices relative to the schedule of benefits, and who no longer needs to

understand all of the veterinary terminology that underlies such a schedule.

The disadvantages are that such an approach is more data intensive and involves

significantly more complex actuarial methods as compared to the schedule of benefits approach

as implemented by VPI. Without a deep understanding of this complexity and proper risk

management policies can quickly get out of hand. Toto faced several challenges in enacting this

resource and admitted that they had almost given up several times. The first major hurdle they

faced was two-fold, no one on the team had an intimate understanding of the veterinary industry

or veterinary terminology, and they had no data on which to build the cat and dog morbidity

tables that would be needed for policy pricing. What they did have was persistence, the ability to

envision an alternative solution, and fortuitous network connections. An economics professor at

Wharton, who had been conducting a multi-year study of the veterinary industry, agreed to share

his data with the team. “If we hadn’t been at Wharton, we would never have met him and he

probably would have never shared with us.” Armed with this data the team members immersed

themselves in reading veterinary materials and recruited vets to their advisory board who helped

them start the process of understanding veterinary procedures.

With these initial steps the Toto team was able to begin the formalization of their first

version of the actuarial models. However the realities of the marketplace began to hit home: “a

hard market hit home as we were turned down or ignored for the most part.” It seems that the

other parties needed for the creation of an opportunity simply didn’t want to be part of the

process. Regulators had through their experiences with VPI become fairly suspicious of the

validity of the concept of pet health insurance. VPI had faced regulatory investigations in the

1990s with eventual resolution of outstanding concerns by the 2000s. The regulatory apparatus

30

(which in the USA is state by state, but fairly homogenous in standards across the country) was

satisfied with the concept of polices based on a schedule of benefits actuarial model. There was

outstanding data to support that policies from VPI, while not necessarily well-loved by

customers, were at least meeting regulatory requirements and achieving acceptable claims pay-

out levels. A percentage of bill model meant something new and untried, with regulators being

well-aware that such policies entail more risk and put more demands on the firms that administer

them. As is usual in such situations they put up many hurdles for Toto to clear.

Along with regulatory approval, Toto was also in the position of having to convince an

established insurance entity to provide underwriting services. Underwriting requires a several

million dollar capital reserve, something that simply was not feasible for Toto to undertake itself.

Of course underwriters had many of the same reservations as the regulators. Further these firms

knew that if they underwrote a product and it failed they would be on the hook, or if it was

disliked their association might generate negative sentiment in the marketplace towards their

other insurance offerings. At the same time if a new insurance product was well structured and

had good market penetration it could provide a robust revenue stream for the underwriter, albeit a

small stream in comparison to existing insurance offerings. Through much iteration and fine-

tuning Toto was eventually able to convince Lloyd’s of London to sign on as their initial

underwriter and received initial regulatory approval from the state of Ohio (subsequently

registering in all fifty states). It took roughly two years for Toto to court an underwriter and three

years to finally clear regulatory hurdles in order to sell its first policy.

The eventual acceptance of the percentage of bill policy structure by underwriters and

regulators was an important step in the opportunity creation process. Equally important were the

socio-cultural shifts that permitted this event and the spillover that occurred into other

relationships that the firm was cultivating. A percentage of bill policy meant that a policy could

31

be structured from the ground up to provide the features that Toto thought would be most

appealing to the market (it is important to keep in mind that the knowledge of what was appealing

to the market wasn’t a given, but rather took significant time and effort to cultivate). Observation

of the failures of VPI revealed many aspects of the prior product that had alienated customers and

vets alike. Toto’s primary competency in actuarial sciences and the data they now had, allowed

them to design policies that permitted full customizability, guaranteed renewal for the life of the

pet, level premiums for life, and variable maximum annual payout limits. This flexibility

permitted the creation of a product that was in harmony with the rest of the mission of Toto and

integrated eloquently with the other crucial firm resources.

Resource 2: Ecommerce Website

Initially it was believed that the best way to market policies was by a “two-pronged” strategy.

The most important community to breakthrough to was the veterinary community, as they served

the role of primary gate keeper to the pet owner, acting as a portal for information and

recommendations to potential customers. Toto initially planned to entice veterinarians to become

involved by offering a monetary incentive tied to policy sales. While monetary incentives seem

like a ready solution to the problem of enlisting the help of vets there were two major problems

with this approach. First such monetary compensation is not permitted within the US regulatory

system, as only licensed agents & brokers are allowed to provide advice or recommendations on

the purchase of insurance. Second, vets neither had the training for this, nor did they want to

spend the time to become licensed, and likewise they had no appetite for the related liability. Toto

surmised that active institutional efforts to change the regulatory system to permit the utilization

of economic incentives would have inevitably been futile. The second prong was to target

pedigree dog breeders with short-term complementary plans to be given when a pet was sold to a

32

new owner. While such an approach was legal, Toto eventually concluded that alone this step

would be insufficient to drive adequate policy sales as there were simply too few pedigree pets

sold relative to the overall dog & cat population which is mostly mixed-breed.

With the development of the actuarial model and more time spent probing stakeholder

responsiveness (testing hypotheses), a multi-tiered marketing strategy was developed. This

included efforts to further public relations, promotion by word of mouth, corporate giving to cat-

dog charities, marketing alliances and affiliations, and most importantly internet marketing. Time

in the trenches revealed to the founders that prior efforts at direct mass-marketing by classic

channels (newspaper, TV, etc.) had been fatal to several of the prior short-lived firms that were

involved in the early years of the pet health insurance industry. The cost to conduct such forms of

advertising simply did not generate adequate sales, particularly with the uphill battle against the

already existing negative sentiment towards the entire concept of pet health insurance. Owing to

the scalability of websites and their ability to be easily updated as needed, the internet was elected

as the primary distribution channel for Toto both in regards to marketing, but also as a sales

engine.

Partly for employment in the period when Toto was not up and running yet, and partly to gain

valuable experience Simon had sought employment at Progressive insurance. This firm was

known, and still is known, for innovativeness in the insurance industry, the introduction of

customizable auto insurance policies (Name Your Price Tool), and a strong internet presence.

With Simon’s experience on the front-end and Karen’s expertise in claims handling and pricing,

they developed and launched a website that allowed customers to explore the product. The

website also provided educational material that explained how veterinary care worked and how

pet health insurance policies could be integrated in a manner such that they actually paid-out on

claims in a manner that the customer would understand. Importantly the website also provided

33

free policy pricing where a customer could enter their pets information, the location they lived in,

and pick a customizable level of coverage. This function provided live feedback on the fly so that

customers could explore different options that might best suit their budget. Additionally the

website provided a direct route to actually purchase the policy and complete the enrollment

process for placing a specific pet onto a policy. Later development advanced the website to

integrate with the claims processing system, providing customers with direct online account

access, expedited claims processing, and trackable, commented policy histories.

Resource 3: Claims Processing System

The earliest plans of Toto did not address the issue of claims processing, a fundamental

step in the insurance process, however it was eventually recognized that this would need to be

addressed. At the outset it was believed that the claims handling process could be readily

outsourced under the assumption that this was a fairly routine task (many existing insurance firms

already utilized outsourced claims processing) and could be implemented through a claims

manual stipulating Toto’s policies. The founders felt that they didn’t have the internal training

and implementation capabilities and that they were not likely to get adequate funding to support

the hiring and training of a claims staff.

As they explored the landscape further and started to understand how radically their

concepts for pet health insurance were departing from the old ways, it became more and more

apparent that they needed to develop internal claims capabilities. “The Toto Pet Insurance

program is unique: efficient, customer-friendly claims processing is central to our operating

tenets. With pet insurance there is little in the way of past practices to rely upon in the US and

there are no third-party administration that specializes in pet insurance coding or claims

management.” Exiting third party services simply didn’t have the requisite knowledge to

34

understand veterinary procedures and the claims process that would be needed to support the

percentage of bill actuarial model.

Importantly, Toto eventually came to realize that it is at the point of claims payment that

the customer truly judges a pet health insurance policy. If they received a fair payment that

matched their expectations they would be content, if they received an unexpected payment that

was based on complex and convoluted claims information they would be upset. As heard from

one prior VPI executive: “the only happy VPI policyholders were the ones that never filed a

claim.”

In order to implement the claims handling process Toto developed a claims

administration system with a backbone based on automation. With the frequent, small

transactions expected from pet health insurance policies it was realized that automation would be

crucial to controlling costs. Further such a system could directly integrate with the actuarial

models, feeding in claims information, and pulling out policy quotations. The ability to data mine

claims information was important for driving the businesses analytical engine that could be used

to develop projections of earnings by source, sales analysis, a range of other crucial metrics, and

provide information necessary for the development and refinement of future insurance policy

offerings.

Comparative Firms: A Brief Overview of Snowy, Asta, & Gromit

The prior section provides a brief history of the formation of the firm Toto (the primary

case) and initial insights into the experience of the founders during the opportunity formation

process. Three resources generated by the entrepreneurial process were highlighted, as they

played a central role in the development and structuring of Toto’s insurance offerings, facilitated

pivotal integration with stakeholders, and helped to illuminate the process of entrepreneurial

35

learning. As these identified resources are fundamental components of any insurance product

variants of them were also developed by the three comparative firms examined in this study. In

total this permitted pattern-matching the processes of learning under fundamental uncertainty

across 12 units of learning. The emergence of these three comparative firms substantially

overlapped the time period during which Toto was maturing from an idea for a business plan

contest to a fully operational business.

The first of these, Snowy, was launched in response to the founder’s recent success with an

unrelated consumer-level startup. This individual was looking for a challenge, enjoyed and

actively sought out business experiences where he knew nothing or very little a priori, and had

witnessed the thriving pet health insurance market in Europe during travels for his prior business.

Although not at first, the development of the three resources examined in this study was strongly

influenced by the necessity that they would support the eventual formation of a ‘monoline’

insurance company. Usually a specialty insurance policy, like pet health insurance, is written by

the issuing firm, but underwritten by an existing large capital provider. All of the existing pet

health insurance companies, excepting Snowy, are structured in this manner. In contrast, a

monoline insurer places both policy issuance and underwriting under one roof. For Snowy, the

founder gradually came to believe that this alternative structure would permit a hyper-focus on

cost-minimization efficiency, maximization of policy issuance flexibility, positioning of the firm

for future lock-in product offerings, and rapid progress towards IPO or M&A. This strong focus

of the founder did not emerge for several years into his efforts to launch his firm, as he too faced

the same fundamental uncertainty that Toto’s and the other firms’ founders were confronting.

Asta, the second comparative case, was established by a founder who had previously played a

central role in the formation of VPI in the 1980s. This founder emphasized the missteps of VPI

and what might have been done differently. He saw this new firm as a chance to do it right

36

without all the cemented cultural and institutional norms of an existing entity. For Asta the

founder had a desire for growth tempered with the necessity to provide the best product to the

specific customer base the might want a mid-range policy (mid-range price and mid-range

coverage). For this firm the implementation of the three resources focused on the development of

this mid-range offering for the well-informed customer.

The final comparative case, Gromit, was initiated as a side-project for a large existing firm in

the pet supplies industry. This firm initially envisioned pet health insurance as an extension

offering to their current broad portfolio of pet products, animal feeds, and branded veterinary

care. The formation of a stand-alone pet health insurance firm was placed under the purview of a

president tasked with its development and integration with the parent firm. This manager

approached three individuals to acts as firm founders, two of whom had prior experience with

VPI. From the start Gromit’s attempted implementation of a new offering in the pet health

insurance marketplace was the least well-developed. This struggle was reflected in the formation

of the three resources of interest and the slowed pace of learning amongst the founders. While

Gromit eventually launched an insurance policy offering, conflicting demands on the firm led to

its demise during the time period in which this study was undertaken.

Armed with this basic history of the primary case and the three comparative cases,

specific attention is now turned to processes by which entrepreneurs learn under fundamental

uncertainty.

Entrepreneurial Learning under Fundamental Uncertainty

The prior section provided a brief history of the formation of the firm Toto and initial insights

into the experience of the founders during the opportunity formation process in the pet health

insurance industry. Three resources with economic relevance were highlighted, as they played a

37

central role in the development and structuring of Toto’s eventual insurance offerings, and

facilitated pivotal integration with stakeholders. Two of these identified resources, the actuarial

model and the claims processing system, are fundamental components of any insurance product.

For this reason variants of both of these resources were developed by the comparative firms.

Although the functionality and business domain of the third resource, the e-commerce website,

could have been met with alternative resources (door-to-door sales, advertising and sales through

exciting insurance brokers) isomorphic market forces led the three comparative firms to develop

their own dedicated e-commerce websites.

In summary, the actuarial model dictated the fundamental structure of the policy governing

what claims would be covered, how much would be paid out for covered claims, and policy

pricing that supported the generation of a profit while assuring a continuous and renewable

capital pool. The claims processing system provided the backbone by which policies were

administered and policyholder obligations were fulfilled. The ecommerce website served as a

unification platform, providing the primary method of communication with active policy holders

and a means for acquisition of new customers, while simultaneously integrating bi-directional

information from both the actuarial model and the claims processing system. This level of

complexity and integration however did not simply spring into fully integrated existence, but

emerged gradually during the co-creation of the opportunity. On average it took the various

founders five to six years to develop, structure, and deploy these resources.

By examining the chronological history of the development of each resource a picture of how

entrepreneurs learn under fundamental uncertainty during the formation of opportunities

coalesced. Iterating between the data regarding Toto’s development of the actuarial model and the

assumptions, predictions, and propositions of entrepreneurial process theory a viable temporal

delineation emerged. Subsequently, Toto’s development of its claims processing system and its e-

38

commerce website were also mapped into this organizing approach. Once the data was organized

in this manner particular features of each temporal category became more prominent and an

overarching structure became apparent. Mapping the other firms in a similar manner revealed the

same structure emerging, albeit with some contingencies around prior experience and the firm’s

position in the social conversation that surrounds the emergence of a new opportunity.

Earliest days

In the earliest days of the opportunity formation process there is little in the way of social

structures to guide the entrepreneur (Berger & Luckmann, 1967). The entrepreneur has many

dreams, and visions of what might come to be, but there is a dearth of information to validate

these ideations (Alvarez & Barney, 2010). If as argued, the emergence of an opportunity is the

creation of a new social arrangement, then the entrepreneur faces the uphill battle of enlisting

others into a shared reality (Kaplan, 2008). However, exiting social structures are not monolithic,

and the entrepreneur will come into contact with many different parties who hold different

viewpoints and respond to the entrepreneur in differing ways (Searle, 1995). This means that it

will be challenging for the entrepreneur to determine when the new social contract they are

putting forth is receiving mixed response because it is not the right solution or that they are

talking to the wrong people. Further, external parties themselves are not temporally homogenous

in their response to proposed alternatives, they can be swayed this way and that (Woolley, 2013),

and they may not even understand their own preferences (Lichtenstein & Slovic, 2006).

The pet health insurance industry posed a further challenge in that a social structure that

had already been formed was significantly antithetical to the re-launch of the industry. The four

firms examined in this study came to the opportunity formation process with different stores of

prior knowledge about what had previously transpired in the industry. The founders of Toto had

39

extensive experience in casualty and property insurance, but knew little about veterinary

medicine or the pet health insurance industry’s history. The founder of Snowy, who originally

started a pet health insurance firm in Canada before eventually making the move to the USA,

began his exploration of the opportunity with no knowledge about either insurance or veterinary

medicine. Although his success in Canada provided him with the experience of bringing a pet

health insurance firm to that market, the knowledge he gained there was not as readily

transferable to the USA market as he initially expected. Snowy: “After what I had accomplished

in Canada, I thought the move down south would be fairly straight forward, it wasn’t and it took

me several years to figure out why”

The founders of the two firms, Asta and Gromit, both had prior experience at VPI and

thus they were aware that pet health insurance had never successfully caught on in the USA,

although many efforts had been made. Both of these groups of founders had extensive veterinary

experience, but their insurance experience had been with a firm that was never able to innovate

itself out of its troubles and one that carried dysfunctional cultural norms. Unlike the founders of

Toto who had seen highly successful insurance firms up close in person or the founder of Snowy

who started with a blank slate, the founders of Asta and Gromit needed to unlearn what they had

been exposed to at VPI and develop a deeper understanding of what caused the problems at VPI

beyond their currently held beliefs. Both Asta and Gromit envisioned their new firms as a chance

to “get it right the second time around” at the same time they were aware that they faced

significant challenges in the inertia that had already been formed in the industry.

Turning back to the resources that each of these firms developed, we can see the

consequences of both the unusual social structure that had emerged around the value of pet health

insurance (or perhaps the destructive value) and the prior knowledge that each founding team

brought to the table. In the earliest days of the formation of these resources the various firms

40

envisioned the resources purely as utilitarian and necessary for the conduct of business. All four

firms initially focused on the creation of the actuarial model as this is the defining lynch pin of an

insurance offering and is necessary for going through the process of underwriting and regulatory

approval. Eventually all four firms turned away from VPI’s approach based on a schedule of

benefits, and instead adopted the percentage of bill model.

The founder of Snowy, who knew nothing about insurance, studied what other insurance

firms were doing, realized he needed to know more about actuarial models and began educating

himself. His reason for choosing a percentage of bill model was “I just looked at what all the

other property and causal insurance firms were doing and decided I better do the same thing.”

The founders of Toto initially elected the percentage of bill model because “there has been a trend

over the last decades for all insurance products of this type to move towards percentage of bill as

actuarial models have matured and the technology to manage them has improved … as more

competitors started doing this customers began to expect it and if you didn’t do the same you

were no longer in the game.” The founder of Asta elected the percentage of bill model early in his

development of the actuarial model because “I had seen what it was like in the early days of VPI

with a warehouse full of files needing to be accessed, it was a nightmare, we choose the schedule

of benefits because it was the only thing we knew how to do at the time … this time around with

Asta I had a feeling that this is the way to do it right in regards to a robust policy.” After the

founders of Gromit were approached by the large parent company, they realized “we have a

chance to try to do it again and get it right this time … the schedule of benefits model at VPI has

a historic cryptic nature with lots of exclusions and etcetera which lead to limited reimbursements

… a negative experience for the client meant a negative experience for the vet”.

Although all of the entrepreneurs envisioned the actuarial model as a means to offer

something new to the marketplace, early on the focus was on the nuts and bolts of how to

41

construct such a model, where the data might come from, what insurance risks to take on and

what to exclude from the policies, etc. There was little if any outwards facing engagement with

others beyond those needed for the actual implementation of the models. Snowy: “I just put my

noise to the grindstone and learned everything I could about actuarial models, without ever

looking up to think about what it meant beyond the policy”. When there was outward engagement

it rarely provided useful information. Gromit: “We realized in the early days that talking to others

could be detrimental to what we were trying to do, you get excited about achieving a

breakthrough and the response is either mildly positive from someone who already thinks pet

health insurance is needed or blanket fear, perhaps caused by severe ignorance, from those vets

who still hate you.”

Returning to entrepreneurial process theory, it is logical that entrepreneurs in the early days

would be focused inward and would tend to hyper-focused on the details of the task at hand. The

context they face is extremely noisy and contradictions abound. At this stage the entrepreneur is a

hypothesis maker, in whom prior experience and imagination dominates the social reality at hand.

The entrepreneur is testing their own hypotheses about what might come to be, but the potential

hypothesis space is so large that the individual can only focus on small pieces at a time. The

existing social structure still dramatically constrains the feedback they can receive from their

hypotheses, but the complexities of this social structure is poorly understood by the entrepreneur.

The entrepreneur is able to make small local changes, receive feedback, and update to the best of

their abilities. This is the beginning of learning. This insight leads to the first proposition:

Proposition 1

Under conditions of fundamental uncertainty task and domain specific learning will be used by

the entrepreneurs in the earliest of stages of hypotheses formation. The uncertainty is deeply

42

pervasive and these early hypotheses are superficial and the resulting feedback is noisy and

difficult to interpret.

Emergence of abstract learning

As the development of the actuarial models progressed, the founders of the various firms

moved into a new stage of the entrepreneurial process. In this stage they became more intimately

involved with outside parties and those who would eventually become critical stakeholders in the

industry. Some of these individuals were discovered accidently, some were actively recruited, and

others were brought into the discussion owing to institutional and regulatory forces. These

various parties brought their own beliefs about the value and validity of pet health insurance.

They shared their viewpoints of the current social order. At this point the various founders started

engaging in a manner that extended beyond simply accomplishing the task at hand, and instead

started crafting a narrative of what they hoped to accomplish in the industry (Searle, 1995). This

narrative however was still in its infancy and lacked strong consensus amongst the involved

parties.

One of the first major hurdles that the firms faced was receiving regulatory approval to

offer their insurance products to the market. In the United States regulatory approval for

insurance policies is granted state by state. Amongst the various states the firms faced different

responses. All of the states required the various firms to go through multiple iterations of the

screening process before finally granting approval. For some of these states pet health insurance

“wasn’t high on the radar” and the involved regulators would address the relevant paperwork in

less than a timely manner. Other states were outright hostile to the idea, although they hid this

opposition behind a wall of bureaucracy.

43

State regulator: “I still remember the first time we dealt with that firm, they were the first

to apply for approval based on that form of actuarial model (percentage of bill). I didn’t know the

exact reason why, but my supervisor told me to make it hard for them to get approved. Now, I

couldn’t simply say no, because we have to justify what we flag in approvals, but you know you

can do that in a way that makes it a challenge … I found out later that my supervisor’s boss was

the one who wanted to deny approval. Not sure, but he had experience many years ago with pet

insurance and I think medical friends of his opposed the idea.”

On average it took each of the firms roughly two to three years to go through the regulatory

process with the first state. Subsequent approvals were faster, as in general, states weigh the

evidence of what another state has done in the regulatory process in a positive manner.

During the process of pursing regulatory approval from the various states in which each

firm planned to first launch, the founders were also beginning the process of developing the

claims processing systems and the e-commerce websites for their firms. While these undertakings

still exhibited a significant amount of task-dependent learning, the process of abstract learning

was starting to be recognizable. These two resources were not simply about meeting their planned

utilitarian need, but they also embedded the founders growing understanding of the social

complexity that surrounds people’s connections with their pets, the value that society places on

veterinary care, and the missteps of prior attempts at pet health insurance.

Snowy: “It was around the time that we were going through regulatory approval in

Canada that I realized just how fouled up things had been in both Canada and the US. The

regulators are pushing back on you, and you know that part of that is just their job, but also I got

the sense that there was real unease with the whole thing. Too many bad apples in the past, lots

of older pet owners who I would never be able to get back to the table, and vets who wouldn’t

even talk about it. The same thing happened when we moved into the USA. You would think after

44

having success in the Canada it would have been an easy move, it wasn’t. I talked to over a

hundred venture capital and private equity shops before one was willing to actually deal with us.

We had to show them not only that our various business functions worked, but that the story we

were telling was actual market potential”

Gromit: “I wanted to know what the roadblock to pet health insurance was. Why

wouldn’t vets accept this paradigm? I think there were two big factors inhibiting adoption. The

first is the assumption that pet insurance is equivalent to human insurance. That means a fear of

an HMO regulatory structure. I was at the Iowa conference and a vet in the audience said that a

physician friend had told him to run away, there would be massive constraints. The second factor

is the fault of pet health insurance itself. The prior negative experience had poisoned the well.

The inertia at VPI was just too strong and we kept doing the wrong thing. My partner says that

the problem with VPI is that they have pissed in the chili (a southern expression meaning that

everyone contributes to the night’s meal, but one party ruins it for everyone else).”

As part of resource development, all of the firms began to actively engage the veterinary

community as they came to understand that the veterinarian plays a significant role in the

customer’s decision to purchase pet health insurance. If a vet strongly opposes the concept then

the various firms would be unable to place advertisement materials in their offices and if asked

the vet would steer clients away from the product. Many older vets who had experienced VPI

were still strongly opposed to pet health insurance, with one vet stating that “it is one of the worst

things to ever happen to veterinarian medicine.” Younger vets who had not been exposed to this

earlier history, might be amenable to the concept but they did not understand the product and

were hesitant to make any recommendations. The various firms all started efforts to engage the

veterinary community through providing educational lectures at veterinary schools, symposium at

veterinary conferences, and office visits. This higher level understanding of the social structure

45

around veterinary care was facilitating the formation of causal understandings on the part of the

various founders.

Through this engagement with the veterinary community the various founders developed

a deeper understanding of the relationship between vet and client, and the relationship that clients

had with their pets. For example, Toto came to understand pet owners as “pet parents”. “The

strengthening of the bond between human and animal, had moved beyond the relationship

between master and subject, to loved family member. As we came to understand this we came to

realize that pet health insurance isn’t just about paying for medical care, it is about your

commitment to the ones you love.” This deeper understanding was reflected in the development

of the claims processing system and the e-commerce website. Although Toto initially planned to

outsource claims processing, they discovered that existing entities that provided this service

simply did not have the capabilities to handle veterinary claims data. This pushed them to develop

internal capabilities to accomplish this task, but as reflected in their writings and business plan at

the time the desire to turn inwards was more than simply needing to fulfill a business function.

They envisioned the claims processing function as a vital touch point with the customer, one in

which they could demonstrate that they were an important member in the “pet parent”

relationship. Likewise, the e-commerce website became more than simply a tool to drive

marketing and sales, but also a platform for relationship management and a means to keep abreast

of the community of “pet parents” they were working to foster. The three other firms

demonstrated similar advances in the complexity of the thought and rationale put forth for the

decisions they were making in regards to the firm resources they were developing.

During this later stage of the early development of the opportunity, entrepreneurs’ prior

hypothesis testing has generated feedback that aided them in refining their initial

conceptualizations of the tasks at hand. At the same time this hypothesis testing and task specific

46

learning has begun to illuminate the social constraints that currently surround the potential

opportunity. The entrepreneurs’ increasing awareness of the role of social context is beginning to

illuminate what aspects of the social structure have plasticity and might be modified. The cross-

talk amongst the firm, outside parties, and stakeholders is beginning to share some degree of

consensus, but strong and differing opinions still hold sway. Entrepreneurs still face the

challenging task of determining what feedback means above and beyond the task at hand and they

still struggle to engage others in dialogue supported by shared understanding. This insight leads

to the second proposition:

Proposition 2

Under conditions of fundamental uncertainty general abstract learning will start to be used by the

entrepreneurs in the later part of the early stages of hypotheses formation. The uncertainty is still

deeply pervasive and the resulting feedback continues to be noisy and difficult to interpret.

Integration & Formalization

As the four firms studied here began to receive regulatory approval and the resources

they needed to support business functions came to fruition, the new market for pet health

insurance went online. For all of the firms the initial days were slow. Toto sold one policy

initially to one of its founders for her cat. Gromit struggled in its first month with anemic sales

and Asta mostly spent its time trying to entice new customers rather than selling policies.

Snowy’s reputation from Canada gave it a boost in the US market, although initial sales were

below projected targets. However for the three firms, Toto, Snowy, and Asta continued efforts in

the marketplace began to quickly show surprising results with commensurately staggering growth

rates. Gromit became mired in the demands of its large corporate owner and although the

47

founders were equally engaged in the learning process as their brethren in the other firms, the

actions they could take were too constrained. Gromit was unable to generate the growth that the

other firms achieved and eventually withdrew from the marketplace.

Interestingly it was at the point that these various firms brought their policies to the

marketplace that the pace of learning substantially accelerated. By engaging various stakeholders

in the process of formalizing the format and structure of the insurance offering the firms had built

a community of interested parties that shared a common language. Major veterinary organizations

had come back to the table with a willingness to view what these firms were doing, separated

from the problems of the past. These organizations were pushing information and opinion out to

their members that showed that the ‘new’ pet health insurance was fundamentally different and

might offer a means for veterinary offices to aid their clients in light of increasing costs for

veterinarian medicine. Likewise the various state regulators were now on board with the new

industry. Asta: “it was clear to the regulators that we were becoming more important, sometimes

they would require negotiation, but for the most part they were letting us do it our way.”

Learning was still ongoing, the process hadn’t stopped. It was around this time that

several of the firms got together informally and began discussing how to support and grow the

industry. Out of this an industry trade group was formed. This group provided a platform to

engage in industry promotion and lobbying, but more importantly it served as a mechanism to

share learning amongst the firms. “It seemed like we had everything all figured out, ready to go,

like you could just throw the switch and start printing money, but we quickly found out that there

was much we didn’t know.” “We knew how to do everything but the most important, we didn’t

know how to sell this thing.” The trade group was a platform to share what worked and what

didn’t, to debate how to move forward, and a place to both compete and support one another.

48

Through the trade group firms were sharing task specific know-how, bolstered by higher level

concepts that supported the validity of what had been learned.

At this stage the various firms began demonstrating the ability to seamlessly engage both

task-specific learning and abstract learning simultaneously. Integration of their varied

entrepreneurial visions with reformation of the social constraints that surrounded the industry had

created the opportunity to engage in new market activity. Increasing awareness of their ability to

influence and shift social constraints in tandem with others, empowered the founders.

Gromit: “We were changing the attitude of the profession (veterinarians) by giving them

a better experience. We realized that one way to do this was to support the push for more

education about pet health insurance … The fascinating part was that at the same time I was

focused on getting vet school talks done, we were taking everything we heard there and feeding it

back into the firm.”

Snowy: “I remember it as one of those moments were a light goes off. All of a sudden

things start to make sense, you understand that the reason something isn’t working is because

something is broken elsewhere. Initially there is too much going on, you get distracted and go this

way and that. Once it all comes together you are no longer just playing with pieces.”

For Toto, as for the other firms, this was also the time when differentiation became more

salient and each firm began to hone its own specific strength. Toto’s advantage lay in

information. The resources they had been creating in the early days were individually robust, but

as they came to understand the industry more deeply they created deep linkages between these

resources. The actuarial model supported the policy formation, but it was also linked to the e-

commerce website permitting the production of on the spot policy quotations with a huge range

of options for potential customer. The claims processing system was linked into both of these as

well, providing a wealth of information as to what was happening in veterinarian care and the

49

performance of Toto’s insurance offerings. This allowed the firm to tailor its offerings to the

customers that it wanted, while shifting those individuals who weren’t willing to pay for their risk

profile to other firms. Without the integration of task-specific and abstract learning this

competitive differentiation would not have been attainable by Toto. Toto: “It wasn’t just about

having a good product, it was about having the best. Behind the scenes you have to get everything

right, but to the pet parent you have to come across as transparent, honest, genuine, and

committed to the relationship. Our strength in information is what let us accomplish this. It was

integration across all facets of the firm.”

In the later stages of the opportunity formation process broadly shared social consensus

emerges as a new way of doing business begins to take hold in the marketplace (Lounsbury &

Crumley, 2007). The various parties involved in the market process are approaching common

ground and shared beliefs are formed (Porac & Baden-Fuller, 1989). Dissent may still be present,

but the involved parties can now communicate with a common language and the aspects on which

they disagree are more readily understood (Searle, 1995). The dual features of both isomorphic

tendencies and competitive differentiation begin to be recognizable driven by the emergence of

shared institutions (regulations), the formation of common stakeholder vernacular, and path-

dependent resource differentiation (Aldrich & Fiol, 1994; Dacin, 1997; Garud & Karnoe, 2001).

Learning is becoming robust and there is a smooth flow between task-dependent efforts and

abstract exploration. These insights lead to proposition three:

Proposition 3

After initial hypotheses testing has generated new knowledge and data allowing the entrepreneurs

to distinguish results from noise the entrepreneurs begin to understand the social complexity of

50

the context and how their previous actions have begun to shape the context. Entrepreneurs now

use task and domain specific learning and general abstraction learning in an integrated manner.

Integrating Study Findings with Extant Learning Theory

This examination of the data has revealed important aspects that need to be addressed in

furthering our understanding of learning under fundamental uncertainty including the role of prior

beliefs the process of belief updating, the formation of alternative hypotheses from the status quo,

the generation of causal understanding, iterative testing to extract information from alternatives,

the differential functions of both task and abstract learning, and the interplay between these two

types of learning.

This proposed temporal structure envisions three broad epochs in the evolution of how

entrepreneurs learn during the opportunity formation process. Given the initial assumption of

fundamental uncertainty, as an opportunity is formed aspects of the uncertainty begin to resolve

and the information available to the entrepreneur undergoes fundamental shifts. This shift in the

character of information leads to changes in how the entrepreneur relates with the environment

and to changes in the efficiency and type of learning employed. The primary forms of learning

identified included task/domain specific (i.e. the answer to a specific question or the solution to a

particular problem) and higher-level abstract learning (i.e. organizing principles, social

structures). The need to integrate these two forms of learning and the need to articulate a

mechanism by which an individual could transition between them highlights a linkage between

entrepreneurial process theory and the logic underlying the hierarchical Bayesian theory (HBT) of

learning.

HBT assumes that the process of learning integrates both task-specific learning and

generalized abstract learning (Kemp & Tenenbaum, 2008). An individual engaged in a specific

51

activity proposes hypotheses for the results they see based on higher order organizing principles.

These higher organizing principles both enable the formation of task-specific hypotheses, but

they also constrain the type and variety of alternative hypotheses (Griffiths & Tenenbaum, 2009).

As information from a task is collected beliefs about the potential hypotheses that could underlie

that task are updated. At the same time the refinement of these local, task-specific hypotheses

leads to refinement of the higher order principles (Tenenbaum, Griffiths, & Kemp, 2006). This

hierarchy of knowledge accelerates the process of local learning and facilities the ability to carry

the lesson learned in one domain into other domains of similar character (Kemp & Tenenbaum,

2008). At its core the HBT approach to learning is about the formation of generative models. An

individual is capable of more than simply describing what is happening, but they are also able to

inductively reason potential causes for what is happening (Steyvers, Tenenbaum, Wagenmakers,

& Blum, 2003). The ability to form generative models allows the individual to extract

relationship properties well in advance of the data that would be necessary to achieve this by pure

probabilistic means (Kemp, Perfors, & Tenenbaum, 2007). The early imposition of an inductive

causal understanding on the part of the individual, also provides the individual with robust

mechanisms to differentiate signal from noise (Payzan-LeNestour & Bossaerts, 2011). As with all

Bayesian approaches to learning the process of belief updating is rationale, but it is important to

note that the individual is rationale to the information they receive, not in any global manner. This

means that HBT learning is compatible with other entrepreneurial logics such as effectuation

(Sarasvathy, 2011) and other conceptualizations of learning (Holcomb, Ireland, Holmes, & Hitt,

2009). Further there is no reason that various forms of information screening mechanisms

(Hutchins, 1995), such as biases and heuristics, cannot play a significant role in the information

gathering and processing components of the HBT approach (Gigerenzer & Todd, 1999;

Gigerenzer, 2000; Payne, Bettman, & Johnson, 1993).

52

In examining the linkages between the case-study derived propositions and the HBT

approach we can identify some modification and contingencies to include when applying this

theory. The entrepreneurs from the case study also faced a hierarchy of knowledge. For them the

most significant higher order principles were the social structure that surrounded the old

perceptions of pet health insurance and the newly emerging social consensus about what the new

pet health insurance could mean. Much like Bayesians they brought their prior knowledge to the

problem at hand. However we must be cautious in our assumptions of the importance of prior

knowledge for the entrepreneurial process. Unlike a simple learning task, the process of

opportunity creation unfolds over a significantly lengthy time period. Yesterday’s new

discoveries become today’s prior knowledge. Although the various founders began with different

prior knowledge backgrounds within a few years of starting they all shared a similar set of

perceptions of what had gone wrong with the industry and how their firms might play a role in the

new marketplace. In situations where multiple groups of entrepreneurs are working towards the

same opportunity set, it is possible that initial prior knowledge plays only a very small role in

what transpires over the coming years of the opportunity formation process.

The HBT literature to date has been applied to clearly defined tasks, wherein the learning

individual may not know the potential dynamic nature of the system, but where the dynamic

component is externally pre-determined (Navarro, Newell, & Schulze, 2015). Such tasks include

activities such as sorting objects into taxonomies, the formation of causal relationships about the

hidden relationships between variables, learning names for things, and other forms of inductive

reasoning (Griffiths & Tenenbaum, 2009). Providing robust explanations for how the human

mind is able to generate such understandings is no small undertaking, but in relation to the

questions we examine in entrepreneurship such tasks are certainly more circumscribed. In

extending the concepts behind HBT to further our understanding of learning under fundamental

53

uncertainty aspects of the formalized nature of HBT models will need to be pushed to their

extreme boundaries.

Current HBT models envision a hypothesis space that is informed by higher and higher

levels of organizing principles. These various levels are directly linked to one another and as such

any task based learning necessarily effects all levels of the knowledge hierarchy. This study

argues that initially entrepreneurs learning under conditions of uncertainty rely almost exclusively

on task-based learning as the complexities of the social structure they face are too great and they

do not yet understand their potential role in effecting higher level changes. If one extends the

HBT concept of higher-order organizing principles to include not only direct linkages with lower

levels (where task based-learning resides) but also to include complex linkages amongst the

higher levels, then it is possible that a greater volume of task-based learning is required before the

refinement of higher levels occurs as compared to classic HBT models. Likewise as abstract

learning accelerates in the later stage of entrepreneurial hypothesis testing, then the resolution of

understanding of the complex linkages at the higher levels will accelerate the performance of

learning at the task level. The insight in this study point towards the potential to revisit HBT

concepts in a formalized manner to both strengthen the logic herein, but also as an opportunity for

entrepreneurship literature to contribute back to the developmental cognitive literature that

brought forth HBT.

Conclusion

This study has examined the emergence of new opportunities created in the pet health

insurance industry by motivated entrepreneurs who had visions to radically transform the

perceptions of what the product provided and how it could be implemented. In particular the

phenomenon of how entrepreneurs learn under fundamental uncertainty was examined. In the

54

process of creating firm resources the entrepreneurs also enacted significant change in the

context, bringing several parties that were previously opposed to the concept back to the table.

Repeatedly throughout interviews with firm founders in this sector it was mentioned by these

individuals that they didn’t think of themselves as anything particularly special, and yet the

change in the industry was remarkable. Likewise an often heard comment was “we didn’t really

know what we were doing at the start” and yet these individuals under conditions that would have

deterred many, learned new ways of doing things and in the process re-introduced a concept to

the US market that many believed had already failed. In this sense the entrepreneurs were also

changed by the process itself, with most stating that they couldn’t possibly have imagined were

they ended up.

In the years since the ‘re-launch’ of pet health insurance by Toto and a group of other firms

who were implementing similar, although not identical, policy structures and business strategies

the industry has undergone a significant rebirth. As of 2005 the penetration rate of pet health

insurance was estimated at less than 1% of the cat and dog population. By 2007, with the entrance

of a small group of firms doing things in a new way, sales of pet health insurance grew to an

estimated $230 million from an estimated $120 million in 2004, a 92% increase. VPI, in response

to the actions of these new entrants, began the process of redesigning its policies, claims handling

processes, and communication with customers. As of 2000 VPI sales were roughly 200,000

policies annually, this number grew to an estimated 415,000 by 2007, finally breaking the

company out of its long stall. Fetch, Inc. (a contemporary of Toto) saw its revenue of $812,000 in

2007, grow to an estimated $18.7 million by 2011, an outstanding 2203% increase. As of 2012

USA Today reported that there were eleven companies offering pet health insurance in the US

market, with sector revenue of $303 million in 2009.

55

An information processing cognitive approach was utilized in this paper as a means to

understand how entrepreneurs integrate information from the context with their own inductions

and intuitions (Clark, 1997). Positing the entrepreneur as a motivated actor seeking to enact

opportunities in a socially constructed marketplace provides both challenges and options for

examining the mechanisms that underlie the entrepreneurial process. While there has certainly

been progress in understanding how individuals make decisions under uncertainty, this work

tends to focus on uncertainty defined either as risky choice or parametric uncertainty (Knight,

1921). The theoretical assertion that entrepreneurs may endogenously influence the context as a

means of generating opportunities, means that we must address the issue of decision-making and

learning under fundamental uncertainty. This fundamental uncertainty means that at certain time

points of the entrepreneurial process it will not be possible to know what will be most relevant to

the future, i.e. what information is most pertinent may not be identifiable (as some of this

information may not yet even exist). At the same time fundamental uncertainty doesn’t mean

anything goes. Markets are not socially created by entrepreneurs alone, rather they are created by

the interaction between entrepreneurs and the many others involved in the process (Garud &

Karnoe, 2001). This means that there are parts of the environment that are both exogenous and

endogenous to the entrepreneurial process, and that even those parts that are endogenous are not

under the strict control of the entrepreneur. Likewise learning in this setting isn’t just about

searching the environment for the optimal solution, but rather entails repeated iteration in an

effort to understand how to integrate with the socio-cultural milieu. This moves the investigation

of entrepreneurship out of the realm in which exploration and exploitation only occurs in a given

external context, a land of ‘search’ (March, 1991)

In order to illuminate concrete cognitive processes this study focused on the

manifestation of resources that facilitated new opportunities. The resources that were examined

56

embodied both tangible and intangible components. For example the actuarial model is

manifested in a set of tangible data artifacts and physical output. But the true value of the

actuarial model lies in its ability to support the firm’s alternative vision of a transparent, easy to

understand, and reliable insurance product. This is what the consumer ended up developing a

belief about, not how the policy actually was operationalized in the background. Tangible aspects

of resources can embody technical and creative breakthroughs which are vital to the future

success of the firm within the opportunity it is creating. However alone these aspects are

insufficient to generate value as they do not provide adequate social understanding to consumers

in order to warrant their engagement in transactions for the firms potential outputs. The intangible

aspects of the resources provide a means by which entrepreneurial creativity is manifested in

social meaning. By generating social meaning they permit the formation of consumer demand,

thus inherently they are also serving as part of the mechanism that generates consumer need

(clearly within a context that was already favorable to increased spending on pet care).

Clearly this paper’s conceptual development is a simplification of an extremely complex

process. In assuming that both individual entrepreneurs and teams of entrepreneurs act as single

entities it removed interesting aspects of how groups process information in interdependent ways

and the effect of different motivational tendencies (DeDreu & Nijstad, 2008). Likewise the

knowledge domain of a group is not a simple additive function of the group members. Rather

diversity of knowledge and depth of knowledge should play different roles in how groups

perceive both exiting alternatives in the given environment along with how they imagine not as of

yet exiting alternatives (Taylor & Greve, 2006).

Further work is needed to understand what the cognitive mechanisms are for how

entrepreneurs envision novel alternatives to the status quo, an area that creativity research has

long struggled to articulate. The philosophy of art speaks of the moment of ‘incept’ (Beardsley,

57

1965), Freud talked of “phantasying” (1908), and more recently scholars such as T. Amabile have

examined how creativity is influenced (1994). In more recent work Amabile & Muller (2008: 33-

34) argue:

Creativity research has enjoyed only a slightly better reputation among the broader

group of psychology scholars, management scholars, and business leaders. Many who

are unfamiliar with recent advances in the field assume that is has little broad relevance

because its focuses only on the arts (and perhaps the sciences), has little validity because

creativity is too ill defined, ephemeral, and “soft” to study rigorously, and provides little

practical applicability because creativity cannot be influenced. But they are wrong.

Likewise, there is clearly a role for prior experience, put perhaps it is does not have as

significant a role to play as it does in the discovery perspective (Shane, 2000; Shane &

Venkataraman, 2000). Prior experience may certainly influence the environmental cues that are

attended to by entrepreneurs (Tversky & Kahnmen, 1974), but little is known about how prior

experience will influence the imagined alternatives and the link between prior experience and acts

that modify the social setting. For example, one of the new pet health insurance firms, was started

by a previously successful entrepreneur who had no experience in either the veterinary field or

the insurance industry. At the same time this study illustrates that the process of market formation

may lead entrepreneurs from different backgrounds towards mutually shared prior knowledge.

Another avenue that might provide further fruitful inquiry is an examination of the

interplay between entrepreneurial motivation, information-processing, and persistence. Clearly

the entrepreneurs in this study faced much more significant odds than they initially perceived, the

deck was heavily stacked against them. Yet even in light of many negative signals there are those

who persisted and in the process were amongst the group that resurrected the industry. The

motivation to succeed was huge and the belief in the new way of structuring the product was

58

robust. One limitation of this study is that while some failed firms were observed, their failure

occurred in the mid-1990s, in-depth data wasn’t available to understand why their motivation and

persistence wasn’t adequate. This limitation may not be severe as from what can be gleaned from

the data these firms were simply imitating VPI, and thus were unlikely attempting the effort to

shift the social conversation. On the other hand, the specter of unobserved entrepreneurial entities

that never made it to the firm formation stage in the 2000s, presents a data challenge that cannot

be directly handled. Did these pre-firms not survive because the entrepreneurs had the wrong mix

of motivation and persistence, were they trying to engage in a different social conversation, did

they fail to learn from their hypothesis testing, could they not differentiate the signal from the

noise, was it simply errors of execution, or was it some combination of all of these?

59

References: Chapter 1

Aldrich, H. E., & Fiol, C. M. (1994). Fools rush in? The institutional context of industry creation.

Academy of management review, 19(4), 645-670.

Alvarez, S. A., & Barney, J. B. (2005). How do entrepreneurs organize firms under conditions of

uncertainty?. Journal of management, 31(5), 776-793.

Alvarez, S. A., & Barney, J. B. (2007). Discovery and creation: Alternative theories of

entrepreneurial action. Strategic entrepreneurship journal, 1(1‐2), 11-26.

Alvarez, S. A., & Barney, J. B. (2010). Entrepreneurship and epistemology: The philosophical

underpinnings of the study of entrepreneurial opportunities. The Academy of

Management Annals, 4(1), 557-583.

Alvarez, S. A., Barney, J. B., & Anderson, P. (2013). Forming and exploiting opportunities: The

implications of discovery and creation processes for entrepreneurial and organizational

research. Organization Science, 24(1), 301-317.

Alvarez, S. A., Young, S. L., & Woolley, J. L. (2015). Opportunities and institutions: a co-

creation story of the king crab industry. Journal of Business Venturing, 30(1), 95-112.

Amabile, T. M., Conti, R., Coon, H., Lazenby, J., & Herron, M. (1996). Assessing the work

environment for creativity. Academy of management journal, 39(5), 1154-1184.

Amabile, T., & Mueller, J. (2008). Studying Creativity, its Processes, and its Antecedents. In J.

Zhou, & C. Shalley, Handbook of Organizational Creativity (pp. 33-64). New York:

Lawrence Erlbaum Associates.

Arthur, W. B. (1989). Competing technologies, increasing returns, and lock-in by historical

events. The economic journal, 99(394), 116-131.

Baker, T., & Nelson, R. E. (2005). Creating something from nothing: Resource construction

through entrepreneurial bricolage. Administrative science quarterly, 50(3), 329-366.

Beardsley, M. C. (1965). On the creation of art. The Journal of Aesthetics and Art Criticism,

23(3), 291-304.

Berger, P. L., & Luckmann, T. (1967). The Social Construction of Reality: A Treatise in the

Sociology of Knowledmann. Anchor books.

Campbell, D. T. (1960). Blind variation and selective retentions in creative thought as in other

knowledge processes. Psychological review, 67(6), 380.

Campbell, D. T. (1974). Evolutionary Epistemology. In P. A. Schilpp, The Philosophy of Karl

Popper, Vol 14 (pp. 413-463). La Salle: Open Court.

60

Cannon‐Bowers, J. A., & Salas, E. (2001). Reflections on shared cognition. Journal of

Organizational Behavior, 22(2), 195-202.

Clark, A. (1997). Being There. Cambridge: MIT Press.

Clark, A. (2016). Surfing uncertainty. Oxford: Oxford University Press.

Coddington, A. (1982). Deficient foresight: a troublesome theme in Keynesian economics. The

American Economic Review, 72(3), 480-487.

Courville, A. C., Daw, N. D., & Touretzky, D. S. (2006). Bayesian theories of conditioning in a

changing world. Trends in cognitive sciences, 10(7), 294-300.

Dacin, M. T. (1997). Isomorphism in context: The power and prescription of institutional norms.

Academy of Management journal, 40(1), 46-81.

De Dreu, C. K., Nijstad, B. A., & van Knippenberg, D. (2008). Motivated information processing

in group judgment and decision making. Personality and Social Psychology Review,

12(1), 22-49.

Dequech, D. (2000). Fundamental uncertainty and ambiguity. Eastern Economic Journal, 26(1),

41-60.

Dequech, D. (2006). The new institutional economics and the theory of behaviour under

uncertainty. Journal of Economic Behavior & Organization, 59(1), 109-131.

Dewey, J., & Bentley, A. (1949). Knowing and the Known. Boston: Beacon Press.

Dimov, D. (2007). Beyond the single‐person, single‐insight attribution in understanding

entrepreneurial opportunities. Entrepreneurship Theory and Practice, 31(5), 713-731.

Dimov, D. (2010). Nascent entrepreneurs and venture emergence: Opportunity confidence,

human capital, and early planning. Journal of Management Studies, 47(6), 1123-1153.

Dimov, D. (2011). Grappling with the unbearable elusiveness of entrepreneurial opportunities.

Entrepreneurship Theory and Practice, 35(1), 57-81.

Dopfer, K., & Potts, J. (2004). Evolutionary foundations of economics. Evolution and economic

complexity, 3-23.

Eisenhardt, K. M. (1989). Building theories from case study research. Academy of management

review, 14(4), 532-550.

Eisenhardt, K. M., & Graebner, M. E. (2007). Theory building from cases: Opportunities and

challenges. Academy of management journal, 50(1), 25.

Freud, S. (1908). Creative writers and day-dreaming. Standard edition, 9(1).

Funke, J. (2001). Dynamic systems as tools for analysing human judgement. Thinking &

Reasoning, 7(1), 69-89.

Garud, R., & Karnøe, P. (2001). Path creation as a process of mindful deviation. Path dependence

and creation, 138.

61

Gephart, R. P. (2004). Qualitative research and the Academy of Management Journal. Academy

of Management Journal, 47(4), 454-462.

Gershman, S. J., Blei, D. M., & Niv, Y. (2010). Context, learning, and extinction. Psychological

review, 117(1), 197.

Gigerenzer, G. (2000). Adaptive thinking: Rationality in the real world. Oxford University Press,

USA.

Gigerenzer, G., & Todd, P. M. (1999). Simple heuristics that make us smart. Oxford University

Press, USA.

Grier, K. C. (2006). Pets in America: A History. Chapel Hill: The University of North Carolina

Press.

Griffiths, T. L., & Tenenbaum, J. B. (2009). Theory-based causal induction. Psychological

review, 116(4), 661.

Gruber, M. (2007). Uncovering the value of planning in new venture creation: A process and

contingency perspective. Journal of Business Venturing, 22(6), 782-807.

Hmieleski, K. M., & Baron, R. A. (2008). When does entrepreneurial self‐efficacy enhance

versus reduce firm performance?. Strategic Entrepreneurship Journal, 2(1), 57-72.

Holcomb, T. R., Ireland, R. D., Holmes Jr, R. M., & Hitt, M. A. (2009). Architecture of

entrepreneurial learning: Exploring the link among heuristics, knowledge, and action.

Entrepreneurship Theory and Practice, 33(1), 167-192.

Hutchins, E. (1995). Cognition in the Wild. MIT press.

Jacobs, R. A., & Kruschke, J. K. (2011). Bayesian learning theory applied to human cognition.

Wiley Interdisciplinary Reviews: Cognitive Science, 2(1), 8-21.

James, W. (1975). Pragmatism (Vol. 1). Harvard University Press.

Jones, M., & Love, B. C. (2011). Bayesian fundamentalism or enlightenment? On the explanatory

status and theoretical contributions of Bayesian models of cognition. Behavioral and

Brain Sciences, 34(04), 169-188.

Kaplan, S. (2008). Framing contests: Strategy making under uncertainty. Organization Science,

19(5), 729-752.

Kemp, C., Perfors, A., & Tenenbaum, J. B. (2007). Learning overhypotheses with hierarchical

Bayesian models. Developmental science, 10(3), 307-321.

Kemp, C., & Tenenbaum, J. B. (2008). The discovery of structural form. Proceedings of the

National Academy of Sciences, 105(31), 10687-10692.

Knight, F. H. (1921). Risk, uncertainty and profit. New York: Hart, Schaffner and Marx.

Lee, B. P. (2001). Mutual knowledge, background knowledge and shared beliefs: Their roles in

establishing common ground. Journal of pragmatics, 33(1), 21-44.

Lichtenstein, S., & Slovic, P. (Eds.). (2006). The construction of preference. Cambridge

University Press.

62

Locke, K. (2001). Grounded theory in management research. Sage.

Lounsbury, M., & Crumley, E. T. (2007). New practice creation: An institutional perspective on

innovation. Organization studies, 28(7), 993-1012.

March, J. G. (1991). Exploration and exploitation in organizational learning. Organization

science, 2(1), 71-87.

Markus, H. (1977). Self-schemata and processing information about the self. Journal of

personality and social psychology, 35(2), 63.

McMullen, J. S., & Dimov, D. (2013). Time and the entrepreneurial journey: the problems and

promise of studying entrepreneurship as a process. Journal of Management Studies,

50(8), 1481-1512.

Mintzberg, H., & Waters, J. A. (1985). Of strategies, deliberate and emergent. Strategic

management journal, 6(3), 257-272.

Mehlhorn, K., Newell, B. R., Todd, P. M., Lee, M. D., Morgan, K., Braithwaite, V. A., &

Gonzalez, C. (2015). Unpacking the exploration–exploitation tradeoff: A synthesis of

human and animal literatures.

Newell, A., & Simon, H. A. (1956). The logic theory machine--A complex information

processing system. Information Theory, IRE Transactions on, 2(3), 61-79.

Osman, M. (2010). Controlling uncertainty: a review of human behavior in complex dynamic

environments. Psychological bulletin, 136(1), 65.

Payne, J. W., Bettman, J. R., & Johnson, E. J. (1993). The adaptive decision maker. Cambridge

University Press.

Payzan-LeNestour, E., & Bossaerts, P. (2011). Risk, unexpected uncertainty, and estimation

uncertainty: Bayesian learning in unstable settings. PLoS Comput Biol, 7(1), e1001048.

Peirce, C. S. (1905). What Pragmatism Is. The Monist, 161-181.

Perfors, A., Tenenbaum, J. B., Griffiths, T. L., & Xu, F. (2011). A tutorial introduction to

Bayesian models of cognitive development. Cognition, 120(3), 302-321.

Porac, J. F., Thomas, H., & Baden‐Fuller, C. (1989). Competitive groups as cognitive

communities: The case of Scottish knitwear manufacturers*. Journal of Management

studies, 26(4), 397-416.

Sarasvathy, S. D. (2001). Causation and effectuation: Toward a theoretical shift from economic

inevitability to entrepreneurial contingency. Academy of management Review, 26(2), 243-

263.

Searle, J. R. (1995). The construction of social reality. Simon and Schuster.

Shane, S. (2000). Prior knowledge and the discovery of entrepreneurial opportunities.

Organization science, 11(4), 448-469.

Shane, S., & Venkataraman, S. (2000). The promise of entrepreneurship as a field of research.

Academy of management review, 25(1), 217-226.

63

Shanks, D. R. (2010). Learning: From association to cognition. Annual review of psychology, 61,

273-301.

Siggelkow, N. (2007). Persuasion with case studies. Academy of management journal, 50(1), 20-

24.

Simon, H. A. (1982). Models of bounded rationality: Empirically grounded economic reason

(Vol. 3). MIT press.

Langley, P., & Simon, H. A. (1981). The central role of learning in cognition. Cognitive skills and

their acquisition, 361-380.

Stake, R.E. 2005. Qualitative Case Studies pp. 443-466. In Sage Handbook of Qualitative

Research, 3rd Edition. Denzin, N.K. & Lincoln, Y.S. (eds.) Sage Publications: Thousand

Oaks, CA.

Steyvers, M., Tenenbaum, J. B., Wagenmakers, E. J., & Blum, B. (2003). Inferring causal

networks from observations and interventions. Cognitive science, 27(3), 453-489.

Suddaby, R. (2006). From the editors: What grounded theory is not. Academy of management

journal, 49(4), 633-642.

Taylor, A., & Greve, H. R. (2006). Superman or the fantastic four? Knowledge combination and

experience in innovative teams. Academy of Management Journal, 49(4), 723-740.

Tenenbaum, J. B., Griffiths, T. L., & Kemp, C. (2006). Theory-based Bayesian models of

inductive learning and reasoning. Trends in cognitive sciences, 10(7), 309-318.

Tenenbaum, Joshua B., Charles Kemp, Thomas L. Griffiths, and Noah D. Goodman. "How to

grow a mind: Statistics, structure, and abstraction." science 331, no. 6022 (2011): 1279-

1285.

Tripsas, M., & Gavetti, G. (2000). Capabilities, cognition, and inertia: Evidence from digital

imaging. Strategic management journal, 21(10-11), 1147-1161.

Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases.

science, 185(4157), 1124-1131.

Walsh, I. J., & Bartunek, J. M. (2011). Cheating the fates: Organizational foundings in the wake

of demise. Academy of Management Journal, 54(5), 1017-1044.

Wiltbank, R., Dew, N., Read, S., & Sarasvathy, S. D. (2006). What to do next? The case for non‐

predictive strategy. Strategic management journal, 27(10), 981-998.

Wood, M. S., & McKinley, W. (2010). The production of entrepreneurial opportunity: a

constructivist perspective. Strategic Entrepreneurship Journal, 4(1), 66-84.

Woolley, J. L. (2014). The creation and configuration of infrastructure for entrepreneurship in

emerging domains of activity. Entrepreneurship theory and practice, 38(4), 721-747.

Wynn Jr, D., & Williams, C. K. (2012). Principles for conducting critical realist case study

research in information systems. Mis Quarterly, 36(3), 787-810.

Yin, R. K. (2009). Case study research: Design and methods, 4th. Thousand Oaks.

64

Term Conceptualization Used

in This Paper Implications Alternative Names

Risk

Probability of all

outcomes known,

effect/type of each

outcome known

Normative and predictive

utility maximization

holds

Uncertainty

Irreducible uncertainty

Lotteries

Ambiguity

Probability of some or all

outcomes are unknown,

effect/type of each

outcome known

Events are pre-

determined and

knowable, but some

aspect of agent

(computational &

cognitive limits) or

environment (complexity)

prevents full acquisition

of information (i.e.

bounded rationality)

Normative and predictive

learning and decision

models based on the

notion of risk don’t hold

owing to ambiguity

aversion (Ellsberg

paradox, 1961)

Uncertainty

Parametric uncertainty

Estimation uncertainty

Weak uncertainty

Knightian Risk

Knightian uncertainty

Savage’s uncertainty

Substantive uncertainty

Procedural uncertainty

Fundamental

Uncertainty

Probability of outcomes is

unknown, effect/type of

outcomes unknown

Both risk and ambiguity

models don’t hold owing

to changeability of the

state space (i.e. structural

change).

Uncertainty

Unexpected uncertainty

Knightian uncertainty

Strong uncertainty

Structural uncertainty

Radical uncertainty

Genuine uncertainty

Table 1.1. Risk, Ambiguity, & Uncertainty

65

Primary Case Comparative

Case One

Comparative

Case Two

Comparative

Case Three

Pseudo-name Toto Snowy Asta Gromit

Year Founder(s)

Started Working

on Idea

2002 2000 (Canada)

2004 (USA)

2003 2005

Year of US Firm

Formation

2006 2007 2006 2008

# of Founders 2 1 1 3

Prior Veterinary

Experience

None None Extensive Extensive

Prior Insurance

Industry

Experience

Extensive None Medium Medium

Impetus for

Initial Concept

Extremely high

bill for

veterinary

services

Prior

entrepreneurial

exit, looking to

do something

new

Getting it right

the second time

around

Extension of

parent firm’s

product portfolio

Table 1.2. List of four firms focused on in this study

Note: Asta was the Charles’ wire fox terrier in the book and several movies starting with the

“Thin Man” in the 1930s. Gromit is Wallace’s companion in several animated claymation movies

created by Nick Park. From the Belgian comic books by Hergé Remi, Snowy is a fox terrier that

faithfully follows Tintin in his adventures around the world. Toto, a terrier, is Dorothy’s intrepid

companion in L. Frank Baum’s series of Oz children books.

Note: Although Snowy had its start in Canada, this study focuses more narrowly on its preparation

and subsequent expansion into the United States. This period also corresponds with the firm’s

main growth phase and acquisition of significant funding. Further, while certain aspects of the

Candian operations were adaptable to the US market, many aspects required novel learning.

66

Table 1.3: Data Sources for Study

67

Table 1.4: State of Pet Health Insurance Industry as of Early 2000s

Group

Knowledge & Sentiment Representative Data

Pet-Owners Few owners have heard of pet health insurance, those who have often don’t understand the

product.

Policyholder sentiment is mixed with some

positive feedback, but also many negative

experiences and hostility.

“Consumers didn’t know the asymmetry of the risk and information, there were shifting underwriter and claims

policies.”

“I get mixed reviews on it. Reimbursements seems to be

at the whim of someone at the other end of the telephone

… many people hesitate to get it.”

Veterinarians Recognition of rising cost of care and rising cost of student debt from veterinary training.

Desire to balance providing the best care with care that the owner can afford, ethical dilemma.

Anger at guilt by association effects from available pet health insurance products.

Fear of going down the same road as human health insurance, desire to avoid HMO system

that tells Vets what they can and can’t do

“The American Veterinary Medical Association supports pet insurance, calling such coverage ‘important to the

future of the veterinary profession’s ability to provide

high quality and up-to-date veterinary services”

“…it is the worst thing that ever happened to veterinary

medicine” – prominent older vet

“a vet in the audience said that physician friends had told

him to run away, there would be massive constraints”

“…lack of understanding difference between HMO and

indemnity.”

Regulators &

Underwriters

Wary of alternative solutions as current solutions

have been problematic

Small, niche product that has to be regulated

owing to legislation but gets passed to the low

man on the totem

Awareness of prior history of claims problems

and firm investigations in the industry

“…company (VPI) has had some problems with the state

insurance department, which has required it to increase

its capital reserves.”

“We didn’t really want to deal with it again … the new

guy got stuck with it.” – state regulator

“California Insurance Commissioner John Garamendi …

recently filed charge against Veterinary Pet Insurance Co (VPI)”

Insurance

Industry

Perception of category as insignificant and

problematic

Perception of category as a joke

Perception of category that succeeded elsewhere, but for structural reasons would not in the US

“There was a person who walking in looking for a

Lloyd’s representative saying he had invented a policy to insure the world against nuclear war … along with some

esoteric forms of coverage like pet health insurance.”

“Should aliens kidnap an earthling … more than 100,000

US citizens have taken out insurance against just this

possibility” in the same article as a discussion of pet

health insurance

“Pet Insurance, for example, grabbed the headline in May

when Patsy Bloom sold the company she founded 20 years ago – PetPlan – for pounds 16m to Cornhill (UK)”

Economists Perception of coverage as a junk product that is

not needed.

Assumed causal link between the creation of an

insurance product an subsequent inflation in the pricing of veterinary care

Misallocation of societal resources

“If you are really worried that someday you will have a

big veterinary bill, put $50 a year away in a bank account

and collect interest on it.” “This is in the junk coverage category”

“If everybody buys insurance we will get CAT scans for cats and dog scans for dogs and all kinds of crazy

machines for pets that nobody would ever have thought of using … and pet owners will pay for it.”

“But Orin Kramer, an economist and consultant in Princeton, NJ, who specialized in insurance issues, says

that widespread insurance for pets may have results that

mirror human health care strikingly.”

68

Chapter 2: The Influence of Mentoring on Entrepreneurial Self-Efficacy and the Desire to

Become an Entrepreneur

Chapter Abstract

This study proposes and tests a model of the relationships between entrepreneurial career

mentoring, traditional career mentoring, and the desire and intent to become an entrepreneur.

Career commitment, career satisfaction, and entrepreneurial self-efficacy were examined as

mediators. The sample included 4,027 university alumni who provided survey data. A multi-

group analysis strategy including calibration and validation samples was used to test the model.

The results support the model fit and study hypotheses. Both types of mentoring were positively

related to entrepreneurial self-efficacy. However, entrepreneurial career mentoring had a positive

relationship with desire and intent to become an entrepreneur while traditional career mentoring

had a negative relationship. The implications of the results for mentoring and entrepreneurship

research and practice are discussed.

Introduction

The choice to become an entrepreneur is a daunting notion for most individuals because

they may feel uncertain if they have the personal, financial, and social resources needed to be

successful. However, despite their uncertainty, individuals do intentionally choose to shift from

an organizational to entrepreneurial career (Bird, 1988; Katz & Gartner, 1988; Krueger &

69

Brazeal, 1994). Researchers have shown that attitudes towards entrepreneurial behavior are an

important predictor of intentions to become an entrepreneur (Douglas & Shepherd, 2002;

Krueger, Reilly, & Carsrud, 2000). The theory of social cognition stresses the role of self-efficacy

as a primary determinant of motivation, willingness to engage, and perseverance in undertaking

tasks (Bandura, 1977). Many studies of entrepreneurship have utilized self-efficacy as a predictor

of intentions (Hmieleski & Baron, 2008). More specifically, studies have shown a relationship

between entrepreneurial self-efficacy (ESE) and entrepreneurial career preferences (Chen et al.,

1998; DeNoble et al., 1999; Krueger, Reilly, & Carsrud, 2000; Segal, Borgia, & Schoenfeld,

2002).

Individuals contemplating career decisions often rely on their mentors for advice,

encouragement, and to serve as a sounding board for their ideas (Kram, 1983). However, we

know little about the impact of mentoring on individual’s intentions to engage in an

entrepreneurial career, i.e. entrepreneurial career intentionality. There are three reasons for our

lack of understanding about the role of mentoring in entrepreneurial career decisions. First,

mentoring research has traditionally explored career advancement within existing companies. We

refer to this form of mentoring as traditional career mentoring. Traditional career mentoring

occurs both formally and informally within existing companies and should, at least logically, be

inversely related to a decision to become an entrepreneur. Entrepreneurial career mentoring, on

the other hand, is substantially different from traditional career mentoring. It consists of

mentoring that encourages departure from the corporate setting and ‘transitioning’ into

entrepreneurship.

A second reason for our lack of understanding about the role of mentoring in

entrepreneurial career decisions is that the very notion of entrepreneurial career mentoring is

conceptually confounded with more general notions of entrepreneurial mentoring. Entrepreneurial

70

mentoring includes the mentor-provided networks, relationships, expertise and assistance

provided to entrepreneurs already operating within the entrepreneurial process (e.g. Deakins,

Graham, Sullivan, & Whittam, 1998; St-Jean & Audet, 2012; Sullivan, 2000). This form of

mentoring is not career-decision focused, but rather focused on aiding the success of the

entrepreneur and their venture. Entrepreneurial mentoring thus occurs after the decision to

become an entrepreneur, while entrepreneurial career mentoring occurs before this decision.

Although these two forms of entrepreneurial mentoring may not be mutually exclusive, they

differ in the types of activities and outcomes received by protéges.

Third, research shows that protégés self-esteem and general and contextual self-efficacy

increase as a result of participating in mentoring relationships (Waters, McCabe, Kiellerup, &

Kiellerup, 2002). These improvements in self-efficacy are positively related to intentions to

engage in related actions or behaviors (Byrne & Keefe, 2002). However, the generalizability of

results of current mentoring research to entrepreneurial mentoring is limited because most studies

have focused on mentoring relationships in which both the mentor and the protégé are from the

same organization, i.e., traditional career mentoring. Also, studies have focused on identifying the

career and psychosocial outcomes that can benefit the protégés while they remain with the current

organization such as increased affective commitment and satisfaction with their current role (see

Noe, Greenberger, and Wang, 2002). In contrast, mentoring in which an individual or group

provides guidance and support to a mentee’s decision to become an entrepreneur, i.e.

entrepreneurial career mentoring, is likely to occur outside traditional organizational boundaries

and influences the protégés intentions to leave their current organization to start a new business

venture. Finally, although research suggests that mentoring enhances protégés self-efficacy, it is

unknown if it further influences the desire and intent to choose an entrepreneurial career.

71

The purpose of this study is to examine the relationships among traditional and

entrepreneurial career mentoring, entrepreneurial self-efficacy, career commitment, career

satisfaction, and an individual’s desire and intention to become an entrepreneur. This contributes

to our understanding of mentoring and entrepreneurship in several ways. The study offers insights

into the role of mentoring by investigating entrepreneurial career mentoring; a type of mentoring

that has received little research attention. Further, the outcome variable used in the study,

intention to become an entrepreneur, answers calls for considering a broader range of personal

outcomes in mentoring research (Kram and Ragins, 2007) and contributes to our understanding of

the role of mentoring in entrepreneurial career decision-making. Finally, the study provides

insight into the mechanism through which mentoring may influence intent to choose an

entrepreneurial career by investigating entrepreneurial self-efficacy as a potential mediator.

Figure 2.1 presents the conceptual model for the study. Below we discuss the theoretical

background for the model and the study hypotheses.

72

Figure 2.1: Model of Desire and Intent for Entrepreneurship

Entrepreneurial Career Mentoring and Traditional Career Mentoring

Entrepreneurial career mentoring refers to a relationship in which a senior more

experienced mentor provides encouragement, guidance, and feedback to a less experienced

individual in regards to the transition to an entrepreneurial position. This transition requires

cognitive, skill-based, and affective learning which can be facilitated by a mentor (St-Jean &

Audet, 2012; Sullivan, 2000). The mentor provides feedback that enables the prospective

entrepreneur to reflect on their actions, choices, attitude, and intended behavior (Sullivan, 2000).

The mentor may also serve as a role model to the protégé, increasing the saliency and desirability

of the ‘entrepreneurial life style’ (Scherer et al., 1989; Scherer et al., 1991). The mentoring

Career Commitment

Entrepreneurial

Career Mentoring

Traditional Career

Mentoring

Career Satisfaction

Desire & Intent

for

Entrepreneurship

+

+

+

+

+

+

+

-

-

ESE-Searching

ESE-Planning

ESE-Marshaling

ESE-People

ESE-Financing

73

relationship likely fosters entrepreneurship as an option to current employment for the protégé by

increasing the salience of its potential benefits and challenges.

Hypothesis 1: Entrepreneurial career mentoring will be positively related to the desire and

intent to become an entrepreneur.

Traditional career mentoring refers to “an intense interpersonal exchange between a

senior experienced colleague (mentor) and a less experienced junior colleague (protégé) in which

the mentor provides support, direction, and feedback regarding career plans and personal

direction” (Russell & Adams, 1997, p. 2). Congruent with the majority of mentoring research this

study constrains career mentoring to mentoring that occurs in an organizational setting and is

primarily focused on benefitting the organization and the individual’s career within the

organization. Kram (1983) asserted that career mentors provide their protégés with career and

psychological support. Career support is provided through coaching, sponsorship, protection,

exposure, the assignment of challenging work, and advocacy. Psychosocial support is provided

through role modeling, counseling, confirmation, and friendship. Traditional career mentoring has

been shown to be inversely related to a protégé’s intentions to turnover and subsequent turnover

(Lankau & Scandura, 2002; Viator & Scandura, 1991)).

Traditional career mentoring has been advocated as a means to accelerate the process of

organizational socialization, increase the retention of high performing individuals, and generate

organizational commitment (Payne & Huffman, 2005). Through organizational socialization

individuals develop an understanding of the organization’s goals and values leading to greater

affective commitment (Griffeth et al., 2000). Repeated exposure to an organizational culture may

lead to a shift in an individual’s goals and values such that they become more congruent with the

espoused organizational values (O’Reilly & Chatman, 1996). Insuring congruence with

organizational values and developing commitment are the mechanisms through which mentoring

74

inhibits turnover intentions. As a result, because an individual’s decision to become an

entrepreneur or move to self-employment necessitates leaving the current organization it is

proposed that exposure to traditional career mentoring will suppress the desire and intention to

engage in this behavior.

Hypothesis 2: Traditional career mentoring will be negatively related to the desire and intent

to become an entrepreneur.

Self-Efficacy & Entrepreneurial Self-Efficacy

The concept of self-efficacy, an individual’s belief in their ability to accomplish tasks

within a particular domain, has played a central role in theories of social learning and social

cognition (Wood & Bandura, 1989). The expectations and motivation that arises from self-

efficacy influence coping behaviors, the degree that effort will be expended, tolerance to

adversity, goal setting, and the choice of actions to undertake (Bandura, 1977; Gist, 1987).

According to social cognition theory, self-efficacy is posited as a central mechanism for the

enactment of human agency (Bandura, 1982, 1989, 2001).

Social cognition theory recognizes that self-efficacy is not a static trait, but rather

malleable based on external and internal influences. Self-efficacy is generally conceptualized to

be task or event specific, i.e., an individual can have a high level of self-efficacy in one domain,

but low self-efficacy in another. This does not mean that the formation of self-efficacy in

response to a particular task is confined to only that specific task. The generative capability of

self-efficacy (Bandura, 1982) asserts that the formation of self-efficacy for one task can influence

the formation of self-efficacy for related tasks. This influence is attenuated as the similarity in

tasks declines and task independence increases.

75

Entrepreneurial self-efficacy (ESE) refers to an individual’s belief in their personal

capabilities related to the formation of a new venture (Boyd & Vozikis, 1994). This specification

of self-efficacy is based on the assumption that the entrepreneurial process involves a range of

inter-related tasks that are unique to such a degree that they cannot be readily captured in a

general measure of self-efficacy (Chen, Greene, & Crick, 1998).

Entrepreneurship is a multi-phase process (McGee et al., 2009; Mueller and Goic, 2003).

The phases of the process model include searching and evaluating the opportunity, developing the

business concept, acquiring needed resources, and managing the venture (Stevenson, Roberts, and

Grousbeck, 1985). During the searching phase the entrepreneur develops a novel idea or

identifies a market opportunity. As part of this process the entrepreneur relies on their creativity

and innovativeness to explore many alternatives. The planning phase (developing the business

concept & assessing required resources) involves formalizing the entrepreneurial concept into an

implementable plan that fits within the entrepreneur’s abilities and goals. During the marshaling

phase (acquiring needed resources) the entrepreneur acts to gain control over the resources

needed to implement the business. The implementing stage (managing and harvesting the venture)

is focused on managing the venture and assuring its successful growth past incubation. The

implementing stage has been conceptualized as including managing people (implementing-

people) and managing the finances of the business (implementing-finance). Entrepreneurs vary in

the extent to which they believe they will be successful in each phase of the entrepreneurial

process. As a result, it is necessary to separately consider individual’s self-efficacy for each phase

of the entrepreneurship process.

Mentors likely exert their influence on entrepreneurship through influencing protégés

self-efficacy (Bandura, 1982; Waters et al., 2002). Mentors may have protégés engage in

activities that expose them to entrepreneurial activities and provide them with a sense of

76

accomplishment or mastery experiences. In these situations the mentor provides an outlet for the

protégé to experiment in a career transition without incurring its full risk, thus increasing the

likelihood of success and mitigating the negative consequence of failure. Interactions with the

mentor also provide the protégé with vicarious experiences, e.g., stories, related to successful

entrepreneurship. This heightens the protégé’s sense that they too will be successful if they

choose to engage in entrepreneurial behaviors. Further, the encouragement and engagement of a

mentor likely serves as a source of verbal and social persuasion to assure the protégé that they

possess the necessary skills and attributes for success. Finally, a mentor who is able to

communicate their experiences, present opportunities to the protégé, and provide feedback and

assurance can generate a positive change in the protégé’s attitude toward and willingness to

engage in entrepreneurship.

Entrepreneurial career mentoring will likely influence all aspects of ESE. The exposure

to tasks related to the entrepreneurial process will heighten the protégé’s perception that they

have the ability to engage and persevere in these domain related tasks. Also, the presence of a

mentor, a supportive other who is respected and admired, enhances the protégé’s self-perceptions

that they are capable of success in entrepreneurial tasks.

Hypothesis 3: Entrepreneurial career mentoring will be positively related to searching,

planning, marshalling, implementing-people, and implementing-finance dimensions of ESE.

Individuals with a high level of self-efficacy within a domain are more likely to engage

and persist in tasks related to that domain (Gist & Mitchell, 1992; Chen, Gully, Eden, 2004).

Studies have demonstrated that higher levels of entrepreneurial self-efficacy are associated with

increases in individual’s intentions to engage in entrepreneurial activities and behaviors (Baum &

Locke, 2004; Chen et al., 1998; Zhao, Seibert, & Hills, 2005). Although we cannot assume that

the formation of intentions will necessarily lead to the career decision to become an entrepreneur,

77

intent has been shown to have a strong influence on subsequent actions (Armitage & Connor,

2001). As a result, increases in individual’s ESE will likely lead to a greater desire and intent to

become an entrepreneur.

Hypothesis 4: Searching, planning, marshalling, implementing-people, and implementing-

finance dimensions of ESE will be positively related to the desire and intent to become an

entrepreneur.

In combination, Hypotheses 1, 3, and 4 suggest that the positive relationship between

entrepreneurial career mentoring and desire and intent is mediated through ESE.

Hypothesis 5: The relationship between entrepreneurial career mentoring and the desire and

intent to become an entrepreneur is mediated through the searching, planning, marshalling,

implementing-people, and implementing-finance dimensions of ESE.

Career Commitment and Career Satisfaction

Career commitment refers to one’s attitude towards a profession or vocation. As such, career

commitment is related to a broader range of referents than is organizational commitment (Blau,

1985). Allen et al. (2004) in a meta-analysis of mentoring notes that the most consistent benefit of

mentoring is probably “the impact on affective reactions to the workplace and positive

psychological feelings regarding one’s career” (p.132). These positive changes in affective

reactions toward the workplace and the generation of feelings of commitment have been shown to

lead to higher levels of satisfaction (Aryee & Chay, 1994). Exposure to traditional career

mentoring is also likely to heighten one’s sense of capability for managing and implementing

challenging undertakings. The capabilities required for engaging in more complex or managerial

tasks within an organizational setting including acquiring resources and managing people and

finances are similar to those required in managing a new venture. As a result, it traditional career

78

mentoring influences protégés self-efficacy related to acquiring resources and managing people

and finances as well as their commitment to and satisfaction with their current career.

Hypothesis 6: Traditional career mentoring will be positively related to marshaling,

implementing-people, and implementing-finance dimensions of ESE, career commitment, and

career satisfaction.

An increase in ESE resulting from career mentoring will likely be associated with a

positive increase in desire and intent to become an entrepreneur. Improvements in an individual’s

believe in their capabilities to persevere in the face of challenging and uncertain tasks likely leads

to beliefs that one will succeed as an entrepreneur. However, an increase in career commitment

and career satisfaction should reinforce the benefits of the current position and reduce the desire

to leave the current organization for an entrepreneurial career.

Hypothesis 7: Career commitment and career satisfaction will be negatively related to the

desire and intent to become an entrepreneur. Marshaling, implementing-finance, and

implementing-people dimensions of ESE will be positively related to the desire and intent to

become an entrepreneur.

In combination Hypotheses 2, 6, and 7 suggest that the negative relationship between traditional

career mentoring and desire and intent to become an entrepreneur is mediated through three

dimensions of ESE, career commitment, and career satisfaction.

Hypothesis 8: The relationship between traditional career mentoring and the desire and

intent to become an entrepreneur is mediated through marshaling, implementing-finance, and

implementing-people dimensions of ESE, career commitment, and career satisfaction.

79

Method: Sample and Procedure

A survey was administered to a large and diverse population of college-educated

individuals who were in different career stages. Survey data was collected from the alumni of a

large Northeastern university in the United States. Approximately 70,000 potential respondents

were contacted via email to participate in the study. 5,300 participants completed the survey

representing a response rate of 7.6%. An examination of educational and demographic variables

revealed no potential response bias.

Because the study purpose was to examine factors that influence desire and intent to

become an entrepreneur, respondents who classified themselves as an entrepreneur or self-

employed were excluded. The remaining sample included 4,027 participants ranging in age from

16 to 72 (mean age =35.52). Males accounted for 64.4% of the sample, which is typical of the

university’s alumni.

Measures

Entrepreneurial Self-Efficacy (ESE). ESE was assessed using a measure that focuses on

the five specific tasks that entrepreneurs engage in when launching a business venture (McGee et

al., 2009). The measure used a seven point Likert-type response scale (1=Disagree to 7=Agree).

Sample items included “Think of new ideas for a product or service” and “Design an effective

marketing campaign for a new product or service.” The scales representing each task included

searching ( = .84), planning ( = .83), marshaling ( = .83), implementing-people ( = .92), and

implementing-finance ( = .93).

Career Commitment. Blau’s (1985) career commitment scale was adapted for use. Items

were reworded to be more industry agnostic. The measure used a seven point Likert response

80

scale (1=Disagree to 7=Agree). Sample items included “I want a career in my current industry”

and “If I could start again I would not choose this field”( = .80).

Desire and Intent for Self-Employment or Entrepreneurship (D&I). This four-item

measure assessed an individual’s desire and intent for job and career, conditions that are

associated with becoming an entrepreneur ( = .78). The four items included a desire to be an

entrepreneur or to be self-employed, a desire for company ownership, a desire to be free from

close supervision (1=Not at all important to 7=Very Important), and intention to become an

entrepreneur or to transition to self-employment in the next five years (1=Not Likely to 7=Very

Likely).

Career Satisfaction. Two items were used to assess current employment satisfaction ( =

.80). The items included “How satisfied are you with your current job” and “Overall, how

satisfied are you with your current career” (1=Not Satisfied to 7=Very Satisfied).

Entrepreneurial Career Mentoring. One item asked participants to estimate the amount

of mentoring that they have received for starting a new business or for being an entrepreneur. A

seven point Likert-type response scale was provided (1=Very little to 7=A lot).

Traditional Career Mentoring. One item asked participants to estimate the amount of

mentoring that they have received in their career or field of professional employment. A seven

point Likert-type response scale was provided (1=Very little to 7=Alot).

Demographics. Age and gender were collected through either the survey or via matching

to university records. Age and gender were included as control variables in this study because

they have both been shown to have an influence on self-efficacy, desire, and intent (Betz &

Hackett, 1981; Maurer, 2001; Wilson et al., 2007).

81

Analytical Strategy

A calibration and validation based data analysis strategy was used because of the large

sample size and the desire to reduce problems associated with excessive model fitting (Cudeck &

Browne, 1983). Calibration and validation with a holdout sample relies on testing and fitting the

model to the calibration sample. The model specified using the calibration sample is then

assessed against the validation holdout sample to confirm that it fits this new data equally well.

This approach adds rigor to the data analysis while improving the generalizability of the tested

model.

The data was split into a calibration sample (n=2,500) and a validation sample (n=1,527)

by random assignment. All descriptive statistics and modeling leading up to the validation stages

are from the calibration sample only.

Results

Table 2.1 (see end of chapter) presents the means, standard deviations, and correlations

for the study variables. There was variability in the amount of traditional career mentoring and

entrepreneurial career mentoring received by the study respondents. Twenty percent of

respondents reported that they had received some entrepreneurial mentoring, 8.2% had received a

great amount of entrepreneurial mentoring, and 71.8% had received very little or none. Also,

72.6% of respondents reported that they had received some traditional mentoring, 12.8% had

received a great amount of traditional mentoring, and 14.6% had received very little or none. The

values shown in Table 2.1 represent the composite average scale scores of the respective items. In

the structural equation modeling analysis, the scale scores are treated as latent variables

82

Testing the Model

To test the study hypotheses, D&I was the dependent variable, entrepreneurial career

mentoring and traditional career mentoring were independent variables, the five dimensions of

ESE, career commitment, and career satisfaction were mediators. Age and gender were included

as covariates.

AMOS 19 was used to test the measurement model (Arbuckle, 2010). This model

included all of the multi-item measures previously tested as well as dummy-latent variables for

the single-item variables. To test model fit all latent variables were allowed to covary. The model

fit was good as indicated by the relevant fit indices (χ2=3828.884, df=432, RMSEA=.056,

CFI=.926, NFI=.918).

The measurement model was respecified with regression pathways to test the study

hypotheses. A nested model comparison approach was used because the model proposes that

specific pathways are important for the link between mentoring and D&I. The initial model

tested (Model A) included all possible regression pathways between the antecedents (age, gender,

entrepreneurial career mentoring, traditional career mentoring), mediators (ESE dimensions), and

outcome (DSI). This test of the full structural model yielded a model with good fit (χ2=3828.884,

df=432, RMSEA=.056, CFI=.926, NFI=.918).

To examine the model specified in Figure 2.1 the paths between entrepreneurial career

mentoring and career commitment and career satisfaction, and the paths between traditional

career mentoring and the searching and planning dimensions of ESE were constrained to zero

(Model B). All of these paths were non-significant in the full structural model (Model A). As

anticipated, a nested model comparison between Model A and Model B revealed that it provided

83

equivalent fit based on the delta CFI criteria (Δχ2=7.0549, df=4, p=.133, ΔCFI=.000).1 Model B

also provided a good fit to the data (χ2=3835.943, df=436, RMSEA=.056, CFI=.926, NFI=.917).

We chose to use Model B for testing with the validation sample and for testing the hypotheses

because of the invariance equivalence and it is more parsimonious than Model A.

Model Validation Using Multi-Group Analysis

We assessed the fit of the model specified in the first stage against the holdout sample. A

multi-group analysis strategy was used in which the same model was tested across both samples

and increasing constraints were imposed. For this analysis, Model B was specified for both the

calibration and validation sample. Estimation of model fit was determined concurrently from both

sets of data. A series of increasingly constraining invariance tests were imposed between the

calibration and validation samples.

The two most important comparisons were tests for configural invariance and metric

invariance. The unconstrained, configural fit of Model B estimated from both calibration and

validation was good (Model 1) (χ2=6501.2, df=877, RMSEA=.040, CFI=.922, NFI=.911). Metric

invariance (Model 2) demonstrated that the measurement loadings (i.e. the loadings of factors on

items) are equivalent between the two groups. A nested comparison between these two models

reveals that the models were equivalent by the delta CFI criteria (Δχ2=30.924, df=21, p=.086,

ΔCFI=.000). Structural invariance was confirmed by further constraining both the structural

weights (Model 3) and the structural covariances (Model 4) equivalent between the groups.

Constraining the structural weights demonstrated equivalence (Δχ2=29.629, df=31, p=.537,

ΔCFI=.000), as did constraining the structural covariances (Δχ2=21.902, df=10, p=.016,

1 The delta CFI method is useful in situations where the sample size is very large and delta χ2 becomes overly sensitive.

Cheung and Rensvold (2002) recommend that if the change in CFI between two nested models is less than or equal to

0.01 then the models can be treated as if they are invariant.

84

ΔCFI=.000). These results showed that Model B was configural, metric, and structural invariant

between the calibration and validation sample. Table 2.2 (see end of chapter) provides the details

of the calibration and validation.

Demonstrating invariance of the model between the calibration and validation samples

showed that the model fits equally well in the sample in which model exploration occurred as

well as a sample that was not involved in model specification. To examine the proposed

hypotheses it was necessary to examine the regression parameters, the total effects, and the

indirect effects. Because the potential of non-normality exists, particularly in regards to the

indirect effects and the gender dichotomy, we used bootstrapping to derive empirical estimates of

parameters. Tables 2.3 and 2.4 (see end of chapter) report the bootstrapping estimates from

Model 4 (Model 4 was used because it provided the most invariance constraints). The estimates

and confidence intervals were derived from 1,000 bootstrap samples and represent unstandardized

coefficients (Preacher & Hayes, 2008).

Results of Hypothesis Testing

Age and gender, the control variables, had significant relationships with D&I (b=-.011,

95% CI -.018, -.007 and b=.434, 95% CI .313, .542, respectively). Hypotheses 1 and 2 described

the relationship between entrepreneurial career mentoring and traditional career mentoring and

D&I. Both Hypotheses 1 and 2 were supported. We found a positive and significant relationship

between entrepreneurial career mentoring and D&I (b=.515, 95% CI .457, .576). Also, as

hypothesized the effect of traditional career mentoring on D&I was negative and significant (b=-

.131, 95% CI -.161, -.098).

Hypotheses 3 predicted a positive relationship between entrepreneurial career mentoring

and the searching, planning, marshaling, implementing-people, and implementing-finance

85

dimensions of ESE. Hypothesis 3 was supported. As shown in Table 2.3, we found significant

and positive relationships between entrepreneurial career mentoring and searching (b=0.306, 95%

CI. .271, .341), planning (b=0.377, 95% CI. .334, .415), marshaling (b=0.282, 95% CI. .251,

.313), implementing-people (b=0.115, 95% CI .088, .139), and implementing-finance (b=0.231,

95% CI. 0.184, .270) dimensions of ESE.

Hypothesis 4 predicted a positive relationship between D&I and each of the ESE

dimensions. Hypothesis 4 was partially supported. Searching and planning dimensions of ESE

were significantly related to D&I search, b=.339, 95% CI .227, .411; planning, b=.330, 95% CI

.224, .429), but the other three ESE dimensions were not (see Table 2.3). These results extend to

the mediation hypothesis (Hypothesis 5). Hypothesis 5 was only partially supported. Only

searching (b=.104, 95% CI .081, .130) and planning (b=.124, 95% CI .083, .164) dimensions of

ESE were significant. The complete mediation results are presented in Table 3.4.

Hypotheses 6 suggested a positive relationship between traditional career mentoring and

marshaling, finance, and people dimensions of ESE and career commitment and career

satisfaction. Hypothesis 6 was partially supported. Although the relationships between career

mentoring and marshalling (b=.026, 95% CI .013, .042) and people (b=.031, 95% CI .018, .047)

dimensions of ESE were significant, the path between career mentoring and the finance

dimension of ESE was not. The estimates were positive and significant for career commitment

(b=.141, 95% CI .119, .167) and career satisfaction (b=.224, 95% CI .195, .253). Similarly,

Hypothesis 7 was only partially supported. Marshaling, people, and finance dimensions of ESE

were not significantly related to D&I, but both career commitment (b=-.142, 95% CI -.212, -.060)

and career satisfaction (b=-.118, 95% CI -.182, -.063) were significantly and negatively related to

D&I. Estimates of the indirect effects between traditional career mentoring and D&I provide

partial support for Hypothesis 8. The indirect effects through the three ESE dimensions were not

86

significant, but we found a significant and negative indirect effect for both career commitment

(b=-.020, 95% CI -.032, -.009) and career mentoring (b=-.027, 95% CI -.042, -.014). Figure 2.2

shows the individual path coefficients and the full path diagram with the respective path

coefficients.

Figure 2.2: Path Diagram with Coefficients

Note: All coefficients are significant at p<.05, except those marked with ns.

Entrepreneurial

Career

Mentoring

Professional

Career

Mentoring

Career Satisfaction

Career Commitment

Desire & Intent

for

Entrepreneurship .015ns

-.142

-.118

.141

.224

.339

.028ns

.330

.031ns

-.038ns

.306

.282

.115

.377

.231

.026 .031

.277

-.084

ESE-Marshaling

ESE-People

ESE-Financing

ESE-Searching

ESE-Planning

87

Discussion

Overall, this study contributes to our understanding of how mentoring can influence

career change involving entrepreneurship. Specifically, we found that traditional career mentoring

increases protégés satisfaction and commitment to their current career while entrepreneurial

career mentoring increases the desire and intent to become an entrepreneur. This suggests that it

is necessary to abandon a “one size fits all’ approach when considering how mentoring influences

career change, at least in the case for individual’s considering a career change to

entrepreneurship.

Our results for ESE as a mediating variable adds to our understanding of the mechanisms

through which mentoring can influence desire and intent to become an entrepreneur. Traditional

career mentoring influenced the two dimensions of ESE that deal with marshaling resources and

managing people. However, neither of these dimensions a significant influence on desire and

intent to engage in entrepreneurship. Entrepreneurial career mentoring had a significant influence

on entrepreneurial desire and intent through searching and planning efficacy, dimensions related

to entrepreneurial ideation. This suggest that the career and psychosocial functions that mentors

need to provide protégés to entice them to change to entrepreneurial careers do not completely

overlap with those typically investigated in mentoring research, i.e., serve as a sounding board,

provide guidance about ideas, help develop their business concept, and identify necessary

resources.

One practical implication of the results is that a specific type of mentoring,

entrepreneurial career mentoring, can enhance individual’s desire and intent to become an

entrepreneur. Initiatives designed to work with aspiring entrepreneurs to get their new ideas and

products to the market through starting new businesses may benefit from providing them with

access to an experienced entrepreneur who can serve as a mentor. The mentor should provide

88

guidance on how to successfully make the transition to an entrepreneurial career, increasing

protégés motivation to search and plan for an entrepreneurship career.

Study Limitations & Future Research

The results and conclusions of this research should be interpreted with caution for several

reasons. First, cross-sectional data was used to test the study hypotheses. The causal directions

proposed in the model could be reversed. For example, it could be that individuals who possess

high levels of desire and intent for self-employment may actively seek out mentoring. The

directionality issue is partially addressed by examining the mediating effect of self-efficacy. It is

logical to assume that mentoring improves self-efficacy, which in turn, drives desire and intent. It

is far less likely the case that desire and intent directly lead to changes in self-efficacy. Also,

increases in entrepreneurial self-efficacy are undoubtedly influenced by other factors and

experiences, which in turn, increase desire and intent to become an entrepreneur.

Another potential limitation of the cross-sectional data is the presence of common

method variance. The study attempted to address this in the design of the survey by separating

related constructs across the survey and by changing the response format across question clusters.

As a robustness check the models presented in this study were examined with the inclusion of a

CFA marker test (Richardson et al., 2009). This test was conducted with a four-item

personal/lifestyle orientation scale that prior research has shown is not related to entrepreneurial

status. Results revealed no statistical detection of either congeneric common method variance

(ΔCFI=0.000) or non-congeneric common method variance (ΔCFI=-0.002).

Second, we used single-item measures to assess entrepreneurial and traditional career

mentoring. Single-item measures are often presumed to have unacceptably low reliability that

cannot be estimated. However work by Wanous & Reichers (1997) and Gardner, Cummings,

Dunham, and Pierce (1998) has shown that single-item measures may actually exhibit acceptable

89

reliabilities. In this study we sought to focus on participants’ overall exposure to two types of

mentoring, a unidimensional report, rather than on varied characteristics of the mentoring process

or individual perceptions of the mentoring experience (Wanous & Hudy, 2001). An analysis

using reliability correction for these two single items leads to the same results of significant and

non-significant outcomes. This analysis was conducted with both mentoring indicators attenuated

as if their reliabilities were .70 (Coffman & MacCallum, 2005).

Eby et al. (2008) provided a meta-analysis of the effects of mentoring across a broad

range of outcomes. The form of mentoring examined by Eby et al. (2008) is similar to what we

consider as traditional career mentoring in this study. Eby et al. (2008) provides a basis for

comparison of how the effects found in our study compare to prior studies. If we assume that

intent to become an entrepreneur is similar to intent to withdraw from the current organization, a

comparison of the effect size of traditional career mentoring on desire and intent (b=-.131) falls

within one standard deviation of the corresponding effect size for withdraw intentions

documented in Eby et al. (2008) (rc=-.10, s.d.=.03) and within the 95% C.I. for this meta effect (-

.15 to -.05). A similar relationship is shown for the relationship between traditional career

mentoring, career commitment and career satisfaction. This similarity of the effect sizes supports

the integrity of the single-item measures used in this study.

A third limitation is that this study only examined intentions rather than the actual

decision to become an entrepreneur. Also, we did not examine the relationship between

mentoring and success as an entrepreneur. Prior studies have shown links between entrepreneurial

self-efficacy and willingness of individuals to persevere in entrepreneur settings (Baum & Locke,

2004). Future research needs to explore how changes in self-efficacy influence changes in

subjects’ motivations, and how events that occur during the entrepreneurial process influence

self-efficacy (Forbes, 2005). While venture performance depends on many factors (Hmieleski &

90

Baron, 2008), prior research has shown a tentative link between supportive mentoring and

venture success (Deakins, Graham, Sullivan, & Whittman, 1998). Future research should explore

what types of mentoring are needed during particular stages of venture formation, what forms of

mentoring are best suited to individuals’ styles, and how mentor-protégé relationships influence

the mentoring process (Ragins, Cotton, & Miller, 2000).

Considering the different influence of traditional career mentoring and entrepreneurial

career mentoring on desire and intent to become an entrepreneur suggests that future research

needs to examine the impact of antecedents that increase ESE but reduce the desire and intent to

pursue an entrepreneurial career. This is especially important for understanding how

organizations can motivate individuals to engage in entrepreneurial activities that benefit the firm,

but do not encourage the employees to leave to start their own businesses. Such antecedents

might involve participation in specific type of mentoring programs as well as communities of

practice, work-related projects, development and training activities, and rewards and recognition

for creative and innovative ideas. Also, further conceptual development is necessary to better

define and understand the nature of entrepreneurial career mentoring and its relationship to the

larger career mentoring literature. Significant questions remain regarding the nature of

entrepreneurial career mentoring and the role of family, friends, associations, trade groups, and

government programs.

91

References: Chapter 2

Allen, T.D., Eby, L.T., Poteet, M.L., Lentz, E., & Lima, L. (2004). Career benefits associated

with mentoring for protégés: a meta-analysis. Journal of Applied Psychology, 89, 127-

136.

Arbuckle, J.L. (2010) Amos (Version 19) [Computer Program]. SPSS, IBM.

Armitage, C.J., & Conner, M. (2001). Efficacy of the theory of planned behavior: a meta-analytic

review. The British Journal of Social Psychology, 40(4), 471-499.

Aryee, S., & Chay, Y.W. (1994). An examination of the impact of career-oriented mentoring on

work commitment attitudes and career satisfaction among professional and managerial

employees. British Journal of Management, 5, 241-249.

Bandura, A. (1977). Self-efficacy: toward a unifying theory of behavioral change. Psychological

Review, 84, 191–215.

Bandura, A. (1982). Self-efficacy mechanism in human agency. American Psychologist, 37(2),

122-147.

Bandura, A. (1989). Human agency in social-cognitive theory. American Psychologist, 44, 1175–

1184.

Bandura, A. (2001). Social cognitive theory: an agentic perspective. Annual Review of

Psychology, 52, 1-26.

Baum, J. & Locke, E. (2004). The relationship of entrepreneurial traits, skill, and motivation to

subsequent venture growth. Journal of Applied Psychology, 89(4), 587–598.

Betz, N. & Hackett, G. (1981). The relationship of career-related self-efficacy expectations to

perceived career options in college men and women. Journal of Counseling Psychology,

28, 399–410.

Bird, B. (1988). Implementing entrepreneurial ideas: the case for intention. Academy of

Management Review, 13(3), 442–453.

Blau, G. 1985. The measurement and predication of career commitment. Journal of Occupational

Psychology, 58, 277-288.

Boyd, N. & Vozikis, G. (1994). The influence of self-efficacy on the development of

entrepreneurial intentions and actions. Entrepreneurship Theory and Practice, 18(4), 63–

77.

Byrne, M. & Keefe, M. (2002). Building research competence in nursing through mentoring.

Journal of Nursing Scholarship, 4th Quarter, 391-396.

Chen, G.C., Greene, P.G., & Crick, A. (1998). Does entrepreneurial self-efficacy distinguish

entrepreneurs from managers? Journal of Business Venturing, 13, 295–317.

92

Chen, G., Gully, S.M., & Eden, D. (2004). General self-efficacy and self-esteem: toward

theoretical and empirical distinction between correlated self-evaluations. Journal of

Organizational Behavior, 25, 375–395.

Cheung, G.W., & Rensvold, R.B. (2002). Evaluating goodness-of-fit indexes for testing

measurement invariance. Structural Equation Modeling, 9(2), 233-255.

Coffman, D.L., & MacCallum, R.C. (2005). Using parcels to convert path analysis models into

latent variable models. Multivariate Behavioral Research, 40(2), 235-259.

Cudeck, R., & Browne, M.W. (1983). Cross-validation of covariance structures. Multivariate

Behavioral Research, 18, 147-167.

Deakins, D., Graham, L., Sullivan, R., & Whittam, G. (1998). New venture support: an analysis

of mentoring support for new and early stage ventures. Journal of Small Business and

Enterprise Development, 5(2), 151-161.

De Noble, A. F., Jung, D., & Ehrlich, S. B. (1999). Entrepreneurial self-efficacy: The

development of a measure and its relationship to entrepreneurial action. Frontiers of

entrepreneurship research, 1999, 73-87.

Douglas, E., & Shepherd, D. (2002). Self-employment as a career choice: attitudes,

entrepreneurial intentions, and utility maximization. Entrepreneurial Theory and

Practice, 26(3), 81-90.

Eby, L.T., Allen, T.D., Evans, S.C., Ng, T., & DuBois, David. (2008). Does Mentoring Matter? A

Multidisciplinary Meta-Analysis Comparing Mentored and Non-Mentored Individuals.

Journal of Vocational Behavior. 72(2), 254-267.

Forbes, D.P. (2005). The effects of strategic decision making on entrepreneurial self-efficacy.

Entrepreneurship Theory and Practice, 29(5), 599-626.

Gardner, D.G., Cummings, L.L., Dunham, R.B., & Pierce, J.L. (1998). Single-item versus

multiple-item scales: an empirical comparison. Educational and Psychological

Measurement, 58(6), 898-915.

Gist, M. (1987). Self-efficacy: Implications for organizational behavior and human resource

management. Academy of Management Journal, 12, 472–485.

Gist, M. E., & Mitchell, T. R. (1992). Self-efficacy: a theoretical analysis of its determinants and

malleability. Academy of Management Review, 17, 183–211.

Griffeth, R.W., Hom, P.W., & Gaertner, S. (2000). A meta-analysis of antecedents and correlates

of employee turnover: update, moderator tests, and research implications for the next

millennium. Journal of Management, 26, 463-488.

Hmieleski, K.M. & Baron, R.A. (2008). When does entrepreneurial self-efficacy enhance versus

reduce firm performance? Strategic Entrepreneurship Journal, 2(1), 57-72.

Katz, J., & Gartner, W.B. (1988). Properties of emerging organizations. The Academy of

Management Journal, 13(3), 429-441.

Kram, K.E. (1983). Phases of the mentor relationship. Academy of Management Journal, 26, 608-

625.

Kram, K.E. & Ragins, B.R. (2007). The landscape of mentoring in the 21st century. In Ragine,

B.R. & Kram, K.E. (eds.), The handbook of mentoring at work, pps. 659-692, Thousand

93

Oaks, CA: Sage Publications.

Krueger, N.F., Jr., & Brazeal, D.V. (1994). Entrepreneurial potential and potential entrepreneurs.

Entrepreneurship Theory & Practice, 18(3), 91–104.

Krueger, N.F., Jr., Reilly, M.D., & Carsrud, A.L. (2000). Competing models of entrepreneurial

intentions. Journal of Business Venturing, 15, 411–432.

Lankau, M. & Scandura, T.A. (2002). An investigation of personal learning in mentoring

relationships: content, antecedents, and consequences. Academy of Management Journal,

45, 779-790.

Maurer, T. J., (2001). Career-relevant learning and development, worker age, and beliefs about

self-efficacy for development. Journal of Management, 27, 123-140.

McGee, J. E., Peterson, M., Mueller, S. L., & Sequeira, J.M., (2009). Entrepreneurial self-

efficacy: refining the measure. Entrepreneurship Theory & Practice, 33, 965-988.

Mueller, S.L. & Goic, S. (2003). East-west differences in entrepreneurial self-efficacy:

implications for entrepreneurship education in transition economies. International

Journal of Entrepreneurship Education, 1, 613–632.

Noe, R.A., Greenberger, D.B., & Wang, S. (2002). Mentoring: what we know and where we

might go. In (Ed.), Research in Personnel and Human Resources Management, Volume

21 (pp. 129-173). Emerald Group Publishing Limited.

O'Reilly, C.A., & Chatman, J.A., (1996). Culture as social control: corporations, cults, and

commitment. In B. Staw and L. Cummings (Eds.), Research in organizational behavior,

Volume 18 (pp. 157-200). Greenwich, CT.: JAI Press.

Payne, S.C., & Huffman, A.H. (2005). A longitudinal examination of the influence of mentoring

on organizational commitment. The Academy of Management Journal, 48, 158-168.

Preacher, K. J., & Hayes, A. F. (2008). Asymptotic and resampling strategies for assessing and

comparing indirect effects in multiple mediator models. Behavior Research Methods, 40,

879-891.

Ragins B.R., Cotton, J.L., & Miller, J.S. (2000). Marginal mentoring: the effects of type of

mentor, quality of relationship, and program design on work and career attitudes.

Academy of Management Journal, 43, 1177-1194.

Richardson, H.A., Simmering, M.J., & Sturman, M.C. (2009). A Tale of Three Perspectives:

Examining Post Hoc Statistical Techniques for Detection and Correction of Common

Method Variance. Organizational Research Methods. 12, 762-800.

Russell, J.E.A., & Adams, D.M. (1997). The changing nature of mentoring in organizations: an

introduction to the special issue on mentoring in organizations. Journal of Vocational

Behavior, 51, 1-14.

Segal, G., Borgia, D., & Schoenfeld, J. (2005). The motivation to become an entrepreneur.

International Journal of Entrepreneurial Behaviour & Research, 11, 42-57.

Scherer, R., Adams, J., Carley, S., & Wiebe, F. (1989). Role model performance effects on

development of entrepreneurial career preference. Entrepreneurship Theory & Practice,

13, 53–71.

Scherer, R.F., Brodinski, J.D., & Wiebe, F. (1991). Examining the relationship between

94

personality and entrepreneurial career preference. Entrepreneurship & Regional

Development, 3, 195-206.

Stevenson, H.H., Roberts, M.J., & Grousbeck,H.I. (1985). New Business Ventures and the

Entrepreneur. Burr Ridge, IL: Richard D. Irwin.

St-Jean, E., & Audet, J. (2012). The role of mentoring in the learning development of the novice

entrepreneur. International Entrepreneurial Management Journal, 8, 119-140.

Sullivan, R. (2000). Entrepreneurial learning and mentoring. International Journal of

Entrepreneurial Behavior & Research, 6, 160-175.

Viator, R.E. & Scandura, T.A. (1991). A study of mentor protégé relationships in large public

accounting firms. Accounting Horizons, 5, 20-30.

Wanous, J.P., Reichers, A.E., & Hudy, M.J. (1997). Overall job satisfaction: how good are single

item measures? Journal of Applied Psychology, 82, 247-252.

Wanous, J.P., & Hudy, M.J. (2001). Single item reliability: a replication and extension.

Organizational Research Methods, 4, 361-375.

Waters, L., McCabe, M., Kiellerup, D., & Kiellerup, S. (2002). The role of formal mentoring on

business success and self-esteem in participants of a new business start-up program.

Journal of Business and Psychology, 17, 107-121.

Wilson, F., Kickul, J., & Marlino, D. (2007). Gender, entrepreneurial self-efficacy, and

entrepreneurial career intentions: Implications for entrepreneurship education.

Entrepreneurship Theory & Practice, 31, 387– 406.

Wood, R., & Bandura, A. (1989). Social cognitive theory of organizational management.

Academy of Management Review, 14, 361–381.

Zhao, H., Seibert, C., & Hills, C. (2005). The mediating role of self-efficacy in the development

of entrepreneurial intentions. Journal of Applied Psychology, 90, 1265–127

95

Mean SD ECM TCM ES EP EM EIP EIF CC CS D&I

Entrepreneurial Career

Mentoring (ECM) 1.49 1.04 n/a

Traditional Career

Mentoring (TCM) 3.19 1.74 .234** n/a

ESE Search (ES) 5.13 1.15 .276** .060** 0.84

ESE Plan (EP) 4.19 1.42 .294** -.002 .567** 0.83

ESE Marshall (EM) 4.94 1.24 .298** .086** .615** .679** 0.83

ESE Imp People (EIP) 5.55 1.03 .182** .059** .431** .459** .609** 0.92

ESE Imp Finance (EIF) 4.75 1.50 .172** .000 .242** .511** .388** .439** 0.93

Career Commitment (CC 4.83 1.26 .027 .170** .069** -.033 .112** .062** -.019 0.80

Career Satisfaction (CS) 5.35 1.44 .049** .216** .057** .001 .115** .103** .035 .527** 0.78

Desire & Intent (D&I) 3.92 1.39 .292** -.074** .403** .404** .342** .218** .213** -.126** -.128** 0.78

Table 2.1 Means, standard deviations, and correlations among study variables

Note: N=2500; *p<.05, **p<.01. Scale reliabilities are presented on the diagonal. ESE= Entrepreneurial self-efficacy.

95

96

Model df Δdf Χ2 Δ Χ2 p CFI ΔCFI NFI RMSEA SRMR

1-Configural

Invariance 877 6501.2 .922 .911 .040 .0474

2-Measurement

Weights 898 21 6531.5 30.294 .086 .922 .00 .911 .039 .0478

3-Structural

Weights 929 31 6561.1 29.629 .537 .922 .00 .910 .039 .0483

4-Structural

Covariance 939 10 6583.0 21.902 .016 .922 .00 .910 .039 .0484

Table 2.2: Multi-Group Analysis of Invariance between Calibration and Validation Models

Note: Each model is nested in the prior model directly above it

97

Antecedent Outcome beta 95% LLCI 95% ULCI

Entrepreneurial

Career Mentoring

ESE Search .306* .271 .341

ESE Plan .377* .334 .415

ESE Marshall .282* .251 .313

ESE Imp People .115* .088 .139

ESE Imp Finance .231* .184 .270

Desire & Intent .277* .220 .335

Traditional

Career Mentoring

ESE Marshall .026* .013 .042

ESE Imp People .031* .018 .047

ESE Imp Finance .015 -.009 .036

Career Commitment .141* .119 .167

Career Satisfaction .224* .195 .253

Desire & Intent -.084* -.115 -.052

Age ESE Search .007* .003 .011

ESE Plan .019* .015 .022

ESE Marshall .006* .003 .010

ESE Imp People .015* .013 .018

ESE Imp Finance .021* .017 .026

Career Commitment .005* .002 .010

Career Satisfaction .014* .009 .020

Desire & Intent -.018* -.023 -.013

Gender ESE Search .135* .053 .217

ESE Plan .041 -.047 .133

ESE Marshall .007 -.064 .077

ESE Imp People -.086* -.144 -.035

ESE Imp Finance .099 .011 .186

Career Commitment -.001 -.086 .089

Career Satisfaction .213* .101 .316

Desire & Intent .393* .294 .499

ESE Search Desire & Intent .339* .272 .411

ESE Plan Desire & Intent .330* .224 .429

ESE Marshall Desire & Intent .028 -.199 .166

ESE Imp People Desire & Intent -.038 -.135 .066

ESE Imp Finance Desire & Intent .031 -.024 .086

Career Commitment Desire & Intent -.142* -.212 -.060

Career Satisfaction Desire & Intent -.118* -.182 -.063

Table 2.3 Regression Parameters for Model 4

Note: *p<.05. LLCI = lower limit of confidence interval, ULCI = upper limit of confidence

interval, ESE= Entrepreneurial self-efficacy.

98

Antecedent Mediating Pathway beta 95% LLCI 95% ULCI

Entrepreneurial

Career Mentoring

ESE Search .104* .081 .130

ESE Plan .124* .083 .164

ESE Marshall .008 -.034 .048

ESE Imp People -.004 -.016 .007

ESE Imp Finance .007 -.005 .021

Traditional

Career Mentoring

ESE Marshall .001 -.003 .005

ESE Imp People -.001 -.005 .002

ESE Imp Finance .000 .000 .002

Career Commitment -.020* -.032 -.009

Career Satisfaction -.027* -.042 -.014

Table 2.4: Indirect Effect for Model 4

Note: *p<.05. Desire and intent to become an entrepreneur is the outcome for the indirect effects.

ESE= Entrepreneurial self-efficacy.

99

Chapter 3: Tensions between Theory and Data: Integrating Subjective Interpretations in a

Bayesian Structural Equation Modeling Examination of Entrepreneurial Self-Efficacy

Chapter Abstract

Recent advances using small-variance priors in conjunction with MCMC estimation,

termed Bayesian structural equation modeling (BSEM), have the potential to create a new

paradigm in scale development, measurement modeling, and structural testing in covariance

based structure modeling (CSM). However, theoretical and statistical considerations have been

raised about BSEM which need to be addressed before BSEM is adopted by scholars. This article

aims to discuss these concerns, provide an overview of the method, and develop guidelines on

how to best utilize BSEM in order to realize its full potential. Using a large dataset, this article

employs a BSEM approach to validate a multidimensional measure of entrepreneurial self-

efficacy. This example illustrates how this technique can be used to address complex

measurement structures. Drawing on factor analytic theory, important issues and the appropriate

application of the technique along with guidelines are presented and discussed. At a more

fundamental level this article illuminates the tension between pragmatic subjective interpretations

of data that give precedence to theory and view models as imperfect simplifications of complex

phenomenon, as opposed to positivist views that more heavily weight the voice of the data and

view the potential models that generate this data as knowable, finite entities.

100

Introduction

The development and dissemination of new statistical techniques often creates

challenges. New statistical methods sow promising theoretical seeds through their potential to (1)

allow for the testing of hypotheses in ways more consistent with the underlying theory, (2) relax

unrealistic model assumptions, and (3) allow for previously un-testable relationships to be

empirically examined. On the other hand, the introduction of new approaches may reap a crop of

dubious findings if they are inappropriately utilized (MacCallum, Edwards, & Cai 2012; Muthén

& Asparouhov 2012b). Furthermore, the capabilities of new techniques may clash with existing

methodological paradigms, creating frustration for writers, reviewers, and editors alike.

The use of Bayesian approaches to conduct covariance structure modeling2 (CSM) for

multidimensional reflective constructs3 (MRCs) provides an exemplar of these tensions. A key

benefit of Bayesian CSM is that the measurement model of MRCs can be more realistically

specified by allowing for the estimation of cross-loadings (Muthén & Asparouhov 2012a) as

opposed to specifying that each observed variable has only one loading as in most CFA models,

which Browne (2001) terms a perfect cluster solution (PCS). Failure to incorporate these cross-

loadings can result in poor model fit and inflated correlations between the underlying latent

variables (Marsh et al. 2009, 2010). On the other hand, Bayesian CSM also allows researchers to

model all correlated unique variances (CUVs) between observed measures, which can raise

2 We utilize the term covariance structure modeling (CSM) throughout this manuscript rather than the

acronym SEM to avoid confusion given that many scholars use terms such as CFA to refer to the

measurement model and SEM to refer to (potentially) the exact same model with regression pathways in

place of correlations between the latent variables (Anderson & Gerbing 1988). The term CSM

encompasses both measurement and structural models.

3 Law, Wong, & Mobley (1998:741) define a multidimensional construct as a construct which contains “a

number of interrelated attributes or dimensions and exists in multidimensional domains.” MRCs occupy

an important place in several organizational and managerial theories, examples of which include the “Big

5” (Marsh et al. 2010), relational norms (Heide & John 1992), and organizational citizenship behavior

(Organ 1988).

101

important concerns about the theoretical meaning of such solutions (Rindskoff 2012) and the

value of such models given they may fit any data structure (MacCallum 2003). Furthermore, the

practice of specifying cross-loadings is contradictory to the current reflective measurement

paradigm developed in the 1980s by marketing scholars (Anderson, Gerbing, & Hunter 1987;

Gerbing & Anderson 1988) and propagated today by well-accepted textbooks such as Hair et al.

(2010). According to this perspective “A necessary condition for assigning meaning to estimated,

latent variables is that the measures posited as alternative indicators of each construct be

acceptably unidimensional,” (Anderson et al. 1987: 435). Thus a paradox exists between the

emerging Bayesian CSM paradigm4 and the current paradigm, which we term the “CFA

paradigm.” This paradox (Poole & Van de Ven 1989) requires illumination and resolution.

This manuscript addresses the dialectic created by the emerging paradigm of Bayesian

CSM. Manuscripts (Kaplan & Depaoli 2012; Muthén & Asparouhov 2012a; Scheines, Hoijtink,

& Boomsma 1999), books (Lee 2007), and technical reports (Asparouhov & Muthén 2010;

Dunson, Palomo, & Bollen 2005) provide an extensive overview of the mathematical foundations

and implementation of Bayesian approaches to conduct Bayesian CSM. However important

theoretical questions remain unaddressed, such as (1) when and why are some cross-loadings

permissible, (2) for cross-sectional research, is their utility in correlating the unique variances of

the manifest variables, and (3) what latitude should be granted to the analyst in specifying priors

on structural relationships? Addressing these questions gets to the core principles of the

philosophy of modeling (MacCallum 2003), measurement (Bagozzi 2011), and theory testing

(Roberts and Pashler 2000).

4 It should be noted that the “Bayesian CMS paradigm” is based off of the work of Thurstone (1947) and

classic psychology. While the utilization of Bayesian techniques for model estimation in a CSM context are

novel, the underlying philosophy of measurement represented in this approach shares a kinship with the

earliest solutions to latent variable measurement models. As in many settings, we can clearly see the ‘swing

of the pendulum’ between paradigms representing contrasting viewpoints (Kuhn, 1977)

102

The remainder of this manuscript is structured in several sections. First, a brief overview

of the Bayesian and frequentist approaches to statistical inference is provided for the unfamiliar

reader. This discussion is utilized as a springboard to reconcile the clear inconsistencies between

authors such as Browne (2001), Marsh et al. (2009) and Muthén & Asparouhov (2012a) who

argue that manifest variables can exhibit multiple large factor loadings and work by Anderson et

al. (1987), Gerbing & Anderson (1988), and Hair et al. (2010) that strongly advocates for the

requirement of unidimensionality. Drawing on works concerning measurement theory (Bagozzi

2007, 2011; Edwards & Bagozzi 2000), the theoretical meanings of cross-loadings are examined

in order to provide a more nuanced, theoretically driven rationale for when and why cross-

loadings are permissible. A review of the benefits and cautions for incorporating informative

priors when conducting Bayesian CSM is next presented, and these issues are empirically

demonstrated using survey data from a study of entrepreneurship. The discussion section

provides additional reflection on these issues, in particular noting that (1) when studying

multidimensional constructs, obtaining measures with high reliability should take precedence

over unidimensionality and (2) for cross-sectional data, researchers are urged to avoid specifying

a prior distribution to correlate manifest variable unique variances as it appears that freeing these

parameters is almost assured to guarantee perfect model fit and obscure potentially important

relationships. Lastly, the manuscript concludes by noting the subjective nature of models and the

importance of recognizing all models are incomplete representations of a complex reality.

Bayesian Modeling of Covariance Structures

The two greatest differences between Bayesian and frequentitst approaches are their

treatment of the nature of population parameters5 and the incorporation of prior information. In

5 Frequentists view parameter values as fixed in the population, and through random sampling, estimates of

these parameters are obtained. Armed with parameter estimates and associated standard errors, inferences

103

regards to the first difference, Frequentists view parameter values as fixed in the population and

through random sampling estimates of these parameters are obtained. In Bayesian inference

population parameters are not fixed but rather are treated as random variables with their own

distributions (Bolstad 2007). This distinction results in interpretational differences between

frequentist confidence intervals and Bayesian credibility intervals (Yuan & MacKinnon 2009).

The second difference between frequentist and Bayesian statistics is that the former is

agnostic to prior information concerning population parameters whereas the latter allows for the

incorporation of prior information. In Bayesian statistics prior information about the distribution

of parameters is updated with new data to produce a posterior distribution of the parameters. This

can be written following Scheines et al. (1999, 38) as:

𝑝(𝜽|𝑦) =𝑝(𝑦|𝜽)𝑝(𝜽)

∫𝑝(𝑦|𝜽)𝑝(𝜽)𝑑𝜽 ∝ 𝑝(𝑦|𝜽)𝑝(𝜽) (1)

In Equation 1 𝑝(𝜽|𝑦) is termed the posterior distribution of 𝜽, which represents the

distribution of parameters incorporating the data and 𝑝(𝑦|𝜽) is the distribution of the data given

parameters, which is equivalent to the likelihood of the parameter estimates given the data,

denoted 𝐿(𝜽|𝑦). Equation 1 shows that this likelihood is weighted by the prior distribution of the

are made whether parameters differ from a specified value, most often zero, in the population. As an

example, imagine that a sample regression coefficient’s estimate is 0.35 with a standard error of 0.10,

resulting in a 95% confidence interval of [0.15, 0.55]. Counter-intuitively, this confidence interval says

nothing about the probability that the true population parameter falls in this range. Rather, the

interpretation of a confidence interval is based on the infinite repeated sampling framework: “If we were to

draw repeated samples from the population and calculate the confidence interval many times, we would

assume that 95% of these intervals would contain the parameter. Given that this interval does not contain

zero, we will infer that the parameter differs from zero in the population.”

In Bayesian inference population parameters are not fixed but rather are treated as random variables with

their own distributions (Bolstad 2007). As such, statistical inference is more straightforward in that a

Bayesian credibility interval is probabilistic in that a 95% credibility interval of [0.15, 0.55] can be

interpreted by saying “there is a 95 percent probability that the population parameter falls within this

interval.” Such inference is more in line with how most individuals think about statistical results.

104

parameters, 𝑝(𝜽), and lastly ∫𝑝(𝑦|𝜽)𝑝(𝜽)𝑑𝜽 is termed the marginal distribution 6. Equation 1 is

often rewritten as:

𝑝(𝜽|𝑦) ∝ 𝐿(𝜽|𝑦)𝑝(𝜽) (2)

The symbol ∝ can be interpreted as “proportional to” indicating that the posterior

distribution of the parameters given the data is proportional to the likelihood of the parameters

weighted by the prior information about the distribution of the parameters. This ability to

incorporate prior information about parameters is a defining characteristic and advantage of

Bayesian statistics. As noted by Bolstad (2007, xxi) “The ‘objectivity’ of frequentists statistics

has been obtained by disregarding any prior knowledge about the process being measured…

Throwing away this prior information is wasteful of information… Bayesian statistics uses both

sources of information.” Apart from incorporating researchers’ intuition and earlier findings,

prior information improves the precision of posterior estimates (Yuan & MacKinnon 2009). In

situations where little information is available, diffuse (also called non-informative) prior

distributions can be specified, in which case point estimates (i.e. the mean) of posterior

distributions approach maximum likelihood estimates in large samples (Dunson et al. 2005).

Despite these advantages, as noted by Kaplan & Depaoli (2012), a primary limitation in

the application of Bayesian methods to complex models has been the challenge of developing the

posterior distribution of parameters given the mathematical intractability of high dimension

integrals. With increased computing capabilities and the development/refinement of Markov

Chain Monte Carlo (MCMC) techniques utilizing the Gibbs sampler (Geman & Geman 1984)

6 The challenge of analytically solving for 𝑝(𝜃|𝑦) is that this can require high dimensional integration for

the marginal distribution. However, the use of Markov Chain Monte Carlo simulation allows for us to

make “draws” from the posterior distributions of interest. Such draws are often conducted using the Gibbs

sampler because while the distribution of the posterior may not be known, the conditional distribution of

the posterior given the data and other model parameters is known (Asparouhov & Muthén 2010; Dunson et

al. 2005; Edwards 2010).

105

combined with their incorporation into user-friendly software (WinBugs, Mplus, R), estimating

complex models using Bayesian approaches has become accessible to applied researchers who

are not trained statisticians/mathematicians. Readers interested in a technical discussion of

MCMC and the Gibbs sampler implemented in CSM are referred to Asparouhov & Muthén

(2010), Dunson et al. (2005), Edwards (2010), and Lee (2007).

For the applied researcher studying MRCs, the increased availability of user-friendly

software to conduct Bayesian inference affords increased modeling capabilities including

estimating models that would not be identified using frequentist estimators (Scheines et al. 1999).

Furthermore the incorporation of prior information, particularly knowledge of the magnitude of

cross-loadings, allows for more realistic modeling of covariance structures (Muthén &

Asparouhov 2012a). However to apply this approach one must reject the current paradigm of

unidimensional measurement as espoused by Gerbing & Anderson (1988) and Hair et al. (2010).

Hair et al. (2010: 674) state “the existence of significant cross-loadings is evidence of a lack of

construct validity” and recommend that “You…should not run CFA models that include…cross-

loadings…evidence that a significant cross-loading exists also shows a lack of discriminant

validity” (675). However, Marsh et al. (2009, 447) have a diametrically opposite view:

“Although there are advantages to having “pure” items that load on a single factor, this is clearly

not a requirement of a well-defined, useful factor structure, nor even a requirement of traditional

definitions of “simple structure” in which nontarget loadings are ideally small relative to target

loadings but not required to be zero.” This presents a paradox (Poole & Van de Ven 1989) that

must be rectified if the flexibility of Bayesian CSM in management research is to be accepted and

subsequently leveraged.

106

Rectifying the ‘old’ and ‘emerging’ measurement paradigms

Is unidimensionality a required property of manifest variables as argued by Anderson et

al. (1987) and Hair et al. (2010), or has this paradigm emerged owing to a fundamental

misunderstanding of the relationship between RLVs and their manifest indicators? To answer

this question, two issues must be examined: (1) the nature of the relationship between a RLV and

observed measures and (2) the source of the theoretical meaning of a RLV. In regards to the first

point, auxiliary measurement theory provides the logical rationale for why a given RLV should

be related to an observed measure (Edwards & Bagozzi 2000). 7 Auxiliary measurement theory

arises from the RLV’s theoretical definition, which provides a parsimonious, precise meaning for

the concept represented by this RLV (Bollen 1989). In other words, one can think, “given this

RLV’s definition, does it make sense that this RLV would be related to this observed measure?”

In regards to the second point, the theoretical meaning of the RLV arises from the

concept the RLV represents (Bollen 2011). Importantly as Bagozzi (2007, 2011) notes, RLVs

contain “surplus” meaning beyond the empirical meaning created through their

operationalization. This recognition that RLVs have surplus meaning is contradictory to the

7 There is some disagreement about the relationship between a RLV and an observed indicator.

Traditionally scholars have stated that the RLV “causes” variation in the observed indicator (Bollen an

Lennox 1991; Howell, Breivik, & Wilcox 2007). Bagozzi (2007, 2011), on the other hand, avoids the use

of causal language and instead states that correspondence rules indirectly connect a latent variable with

observed measures:

“an auxiliary hypothesis concerning theoretical mechanisms, empirical criteria, and a rule

connecting the mechanisms and criteria. A correspondence rule is a complex conceptualization

consisting of a logical expression, some theoretical meaning, and some empirical meaning. It

bridges the abstract meaning that should be specified formally in the latent variable and the

observational meaning residing in the empirical operations defining the manifest variable…A

connection thus exists between a latent variable and manifest variable, but the connection is an

indirect one. Latent variables are not identified or defined by manifest variables; manifest variables

provide only part of the meaning of the latent variable.”

While this distinction does not impact the core arguments we make in this manuscript, due to this

contention in measurement theory, we avoid the use of causal language when describing the relationship

between a RLV and an observed measure.

107

aforementioned Anderson et al. (1987: 435) excerpt because their statement implies that a RLV’s

meaning is a direct function of its indicators8. One way to understand that a RLV has surplus

meaning is that the exclusion of one or more measures does not alter the theoretical definition of

said latent variable (Bollen and Lennox 1991), which is contradictory to the argument that an

RLV’s theoretical meaning is a function of its indicators9.

Apart from unidimensionality not being a requirement to establish the meaning of a RLV,

the requirement that measures be unidimensional is not consistent with the original literature on

factor analysis, particularly Thurstone’s (1947) work on simple structure. As noted by Browne

(2001), Thurstone’s argument for simple structure was that the matrix of factor loadings needed

to be easily interpretable, with the core requirement being that a manifest variable not be allowed

to load on all latent factors. In other words, if there are m latent factors, then each manifest

variable can have at most m – 1 “large” loadings. If all items have only one large loading then

the matrix of factor loadings is said to have a “perfect cluster solution” (Browne 2001), which is

the most restrictive form of simple structure, but not the only form of permissible simple

structure. An example of a more complex pattern of loadings is Holzinger & Swineford’s (1937)

bi-factor model whereby each manifest variable is posited to load onto a general factor and one

8 Further evidence that these scholars adopt the conceptualization that a RLV’s definition is a direct

function of its indicators can be seen in Gerbing & Anderson’s (1988: 189) statement that “Factors in an

exploratory analysis do not correspond directly to the constructs represented by each set of indicators

because each factor from an exploratory factor analysis is defined as a weighted sum of all observed

variables in the analysis.”

9 To provide more detail, imagine that there is a series of 10 indicators that operate as indicators for a given

RLV, but for parsimony, the researcher only includes a set of four on the measurement instrument.

According to the Anderson et al. (1987), the theoretical meaning of this RLV would differ from study to

study depending on the series of indicators selected, which is actually consistent with the determination of

meaning for a formative construct (Howell et al. 2007). However, as noted by Bagozzi (2011) and

Edwards (2011) one advantage of using RLVs over formative constructs is that they are generalizable

beyond the empirical operationalization in a given study precisely because indicators are interchangeable.

Furthermore, Anderson et al.’s (1987) conceptualization of an RLV’s definition being a function of its

indicators is inconsistent with the ontological necessity that an RLV exist independent of its measures

(Borsboom, Mellenbergh, & van Heerden 2003)

108

specific factor. Bi-factor models, which are not permitted by Gerbing & Anderson’s (1988) and

Hair et al.’s (2010) logic, have recently seen increased application in the education and medical

literature as reviewed by Reise (2012). A second example of a more complex structure is

Thurstone’s “box” data, where the majority of measures exhibit a complexity greater than one10

(Browne 2001).

A natural question following from the above arguments is when are more complex factor

structures permissible? An answer can be found by drawing on the aforementioned concept of

auxiliary measurement theory: when two RLVs have related theoretical domains, such as with

MRCs, cross-loadings would be permissible given there is theoretical justification for such

loadings (Marsh et al. 2009). On the other hand, if two RLVs have distinct theoretical domains,

then large-magnitude cross-loadings would suggest that the measures are not operating as the

researcher expects, which requires refinement of the current measurement theory. However,

there is no definitive statistical test whether cross-loadings are permissible; rather, such cross-

loadings must be justified by the researcher and their inclusion or exclusion is ultimately a

subjective judgment. Thus, like most statistical urban legends (Spector 2006), there are “kernels”

of truth underlying the use of unidimensional measure. Complexity one measures, and

consequently perfect cluster solutions, encourage researchers to formalize measurement

instruments and develop scales with an easily-interpretable simple structure and avoid

theoretically-inconsistent cross-loadings and increase the parsimony of measurement models

(Asparouhov & Muthén 2009).

An important lesson can be taken from examining the emergence of the CFA

measurement paradigm: readily-available and easily-implemented software, combined with

10 An observed variable’s complexity refers to the number of large loadings that it is posited to have on the

set of m latent variables. As noted by Browne (2001) and Myers, Ahn, & Jin (2013) complexity one

measures have dominated applied applications in the social sciences.

109

unchallenged recommendations, can lead to a paradigm inconsistent with the theory underlying

the statistical procedure. In his review of the history of factor rotation Browne (2001) notes that

when factor rotation was completed by hand researchers were able to include his/her subjective

knowledge at each step of the process. He then notes that with the advent of computers:

“First of all the time consuming aspect of factor rotation was eliminated. Rotating

factor matrices became quick and easy. Secondly the opportunity for use of

background knowledge concerning the variables during the rotation process was

eliminated. Some regarded this as a desirable change of direction to greater

objectivity, since the rotation process was no longer influenced by the investigator

and depended only on the choice of rotation algorithm.” (113)

Browne (2001), citing Yates (1987), notes that the ability of rotation algorithms to

recover perfect cluster solutions rather than more complex structures such as Thurstone’s (1947)

“box” data was an important cause of applied researchers seeking measures that exhibited a

complexity of one. Asparouhov & Muthén (2009) further note that with the development of CFA

in the late 1960s and popularization of LISREL in the 1970s, combined with the limitations of

EFA for conducting structural analysis, researchers shifted to utilizing CFA models. Perfect

cluster solutions were preferable given the necessity in CFA of a priori specifying the factor

structure, which when augmented with the recommendations of Anderson et al. (1987), Anderson

& Gerbing (1988), Gerbing & Anderson (1988), and the popular text by Hair et al. (2010),

resulted in a “perfect storm” that has led us to the CFA paradigm. As House (1996: 333, 346)

articulated: “Clearly, social scientists need to escape the boundaries of prevailing paradigms and

to question prevailing wisdom,” lest we “…get trapped in our measurement system and apply it

blindly to new questions for which it is inappropriate.” However, by the same token, clearly

articulating key benefits and pitfalls of the Bayesian CSM paradigm is essential to help

researchers best leverage these techniques.

110

Incorporation of prior information into Bayesian CSM

Given the unfamiliarity of many researchers with the specification of prior distributions

in Bayesian analysis, this section provides an overview of the benefits and cautions of specifying

prior distributions on three sets of parameters: (1) factor loadings, (2) correlated unique variances,

hereafter noted as CUVs, and (3) structural coefficients. Attention is limited to these parameters

because decisions about specifying priors for these parameters will be necessary when conducting

the structural analysis of any model, whereas specifying priors on parameters such as direct

effects between independent variables and the intercepts of endogenous manifest variables, as in

Muthén and Asparouhov (2012a), are less-frequently seen in management research.

As a broad overview, a few words about priors in the context of CSM need mentioning.

First, as noted by Yuan and MacKinnon (2009: 306) “A strong prior that dominates the likelihood

usually is not recommended. The inference should be mostly driven by currently observed data.”

Thus, as expanded on in the discussion, it is the responsibility of the researcher to disclose what

priors, both diffuse and informative, were specified and the robustness of the results to the

specification of different priors. Second, different types of parameters are specified to have

different prior distributions. Typically intercepts, factor loadings, and structural regression

coefficients are specified as using normal distributions, an individual variance parameter (such as

a unique variance or error variance) is specified using an inverse gamma distribution, and

covariance/correlation matrixes are specified using an inverse Wishart distribution (Asparouhov

& Muthén 2010). Third, as summarized in Table 3.1 (see end of chapter), there are distinct

theoretical implications for specifying informative priors for these different parameter types.

Factor Loadings

The use of informative priors on factor loadings represents one of the most valuable

features of Bayesian CSM relative to frequentist CFA approaches. When conducting traditional

111

CFA with complexity one indicators, researchers specify one freely-estimated factor loading and

fix the remaining loadings at zero for a given manifest variable. However, the requirement that

loadings be fixed to zero in the population is often unrealistic as Asparouhov & Muthén (2009:

398) note “although technically appealing, CFA requires strong measurement science that is often

not available in practice. A measurement instrument often has many small cross-loadings that are

well motivated by either substantive theory or by the formulation of the measurements.” In

Bayesian CSM, researchers can more realistically model the hypothesized factor structure by

specifying informative priors for expected small loadings and use diffuse priors for expected

large loadings (Muthén & Asparouhov 2012a). For example, imagine a researcher has data for a

multidimensional construct measured by 12 items each assumed to have a complexity of one that

serve as indicators for three RLVs. Assuming that the variances of the latent variables are fixed

to one for identification, Figure 3.1 Panel A presents the traditional CFA specification with large

loadings denoted with a ‘?’ and remaining loadings fixed to zero whereas Figure 1 Panel B shows

a Bayesian CSM specification with large loadings also denoted as a ‘?’ (i.e., diffuse priors are

utilized) and remaining loadings specified using an informative normal prior11 with mean equal to

zero and a standard deviation of 0.1012.

11 The reader will notice that by frequentist standards Panel B is not identified because with the factor

correlations freely estimated, at least two zero loadings would need fixed in each column (Bollen 1989).

However as noted by Dunson et al. (2005) identification in a Bayesian framework is distinct in that

identification means that the posterior distribution can be updated from the data. Using informative priors

can allow for this updating to occur, which provides an explanation for why Bayesian methods can allow

for the estimation of CSMs that would have previously been unidentified (Muthén & Asparouhov 2012a;

Scheines et al. 1999).

12 Given that the latent variables are assumed to have a variance of 1.0, the factor loadings are standardized

in this setting. Specifying that the standard deviation of the factor loading is 0.10 would indicate that the

loading is expected to have a value between -0.20 and 0.20, which we believe can be considered

inconsequential from a substantive perspective.

112

Panel A: Traditional CFA Model Panel B: Bayesian Model

(

? 0 0? 0 0? 0 0? 0 00 ? 00 ? 00 ? 00 ? 00 0 ?0 0 ?0 0 ?0 0 ?)

(

? 𝑁(0,0.1) 𝑁(0,0.1)? 𝑁(0,0.1) 𝑁(0,0.1)? 𝑁(0,0.1) 𝑁(0,0.1)? 𝑁(0,0.1) 𝑁(0,0.1)

𝑁(0,0.1) ? 𝑁(0,0.1)𝑁(0,0.1) ? 𝑁(0,0.1)𝑁(0,0.1) ? 𝑁(0,0.1)𝑁(0,0.1) ? 𝑁(0,0.1)𝑁(0,0.1) 𝑁(0,0.1) ?𝑁(0,0.1) 𝑁(0,0.1) ?𝑁(0,0.1) 𝑁(0,0.1) ?𝑁(0,0.1) 𝑁(0,0.1) ? )

Figure 3.1: Specification of the factor loadings for a perfect cluster solution (PCS) CFA model

where small loadings are fixed to zero (Panel A) and a Bayesian model where small loadings are

given a small-variance normal prior with mean equal to zero and a standard deviation equal to

0.1.

113

As seen in Figure 3.1, the Bayesian specification using informative priors provides a

more realistic representation of the measurement model in that all small loadings need not be

fixed to a value of zero. Apart from providing a more realistic representation of the measurement

model, two statistical benefits arising from the use of small-variance cross-loadings are (1)

improved model fit and (2) reduced correlation between the latent variables13. Furthermore,

imagine that Panel B is estimated on pilot data and the researcher finds that three loadings which

had been a priori hypothesized to be small are large enough to be of substantive concern. Upon

collecting a subsequent dataset these three loadings could be freely-estimated using diffuse

priors14. In addition, information concerning the magnitudes of the cross-loadings from the pilot

study could be incorporated in subsequent research to improve the precision of parameter

estimates.

However the use of small-variance priors for factor loadings presents a drawback by

imposing increased methodological and theoretical challenges as summarized in Table 3.1 Panel

A. First, if the researcher specifies informative priors with too large a variance (i.e. standard

deviation of 0.30 rather than 0.10 in our example) the model may not converge (Muthén &

13 Marsh et al. (2009, 2010) note that many measurement instruments such as the Big Five personality scale

display unacceptable fit when modeled using CFA models where all indicators have a complexity of one.

Furthermore, in order to recreate the underlying covariance matrix when there are multiple small cross-

loadings, the correlations between the latent variables will be inflated, which can threaten the discriminant

validity of the constructs.

14 Concurring with MacCallum et al. (2012), we caution researchers that simply freeing these three

parameters in the original sample is akin to utilizing modification indices to conduct a specification search

(MacCallum et al. 1992). It should be noted there are differences between this Bayesian approach and

using modification indices to conduct specification searchers using ML estimation. As noted by Steiger

(1990), modification indices, also termed Lagrangian multipliers (Bollen 1989), are calculated assuming

that the remainder of the model is properly specified. As a result, the resulting change in the χ2 from

freeing a given parameter based in its modification index may not equal the reported value of the

modification index. Thus, as noted by Muthén & Asparouhov (2012a) the Bayesian approach using small-

variance priors provides a more complete picture than using ML modification indices. However, it is

important to remember that the researcher is still modifying the originally hypothesized model as noted by

MacCallum et al. (2012).

114

Asparouhov 2012a). In other words, this approach is unlike exploratory factor analysis (EFA)

where the analyst lets the data speak for itself, given that a degree of prior knowledge about

loadings and the number of factors is required. Second, as of this point the use of normal priors

for factor loadings has been assumed, which have a range of minus infinity to positive infinity.

However, a researcher could knowingly or unknowingly specify different distributions (i.e.

uniform) on small loadings that “force” hypothesized small loadings to be small, which

underscores why reporting priors is critical. Lastly, as argued in the previous section, practically-

significant cross-loadings need to have a theoretical justification, which places an increased

burden on the researcher to defend such cross-loadings.

Correlated Unique Variances (CUVs)

As shown in Table 3.1 Panel B, specifying informative priors for the 𝚯𝛿 matrix15 to

correlate the unique variances (UVs) of the observed variables has different theoretical

implications from using informative priors for factor loadings. It should be noted this is not

referring to instances where researchers allow for individual CUVs due to a priori expectations

such as longitudinal designs (Bollen 1989) or item-wording effects such as reverse scoring

(Marsh et al. 2010), which are permissible. Rather, this refers to the specification of an

informative inverse-Wishart distribution for 𝚯𝛿 that allows for all parameters in this matrix to be

estimated as in Muthén & Asparouhov (2012a).

To understand the distinct theoretical implication from modeling 𝚯𝛿, it is important to

define the statistical meaning of a CUV. Imagine item X1 has a positive loading onto Factor1

15 𝚯𝛿is the LISREL notation for the symmetric matrix for the unique variances of the observed variables on

the diagonal and the correlations between the unique variances on the off-diagonal. In the factor analysis

model 𝚯𝛿is assumed to be a diagonal matrix, which means that the partial correlation between the observed

variables, holding the latent factors constant, is assumed to be zero (MacCallum & Tucker 1991).

115

and X4 has a positive loading on Factor2. If Factor1 and Factor2 are positively correlated, a

positive CUV indicates that the model under-estimates the correlation between X1 and X4

whereas a negative CUV indicates that the model over-estimates this correlation (Gerbing &

Anderson 1984). If Factor1 and Factor2 are negatively correlated, a positive CUV indicates that

the model over-estimates the correlation whereas a negative CUV indicates that the model under-

estimates this correlation. Given this statistical meaning the first cause for apprehension is that,

unlike cross-loadings, there is theoretical ambiguity as to the cause of CUVs. For example,

without a priori expectations, is the CUV between X1 and X4 the result of a method factor such

as social desirability bias, is there another underlying latent variable that is not modeled such as a

second order-factor (Gerbing & Anderson 1984), or is the CUV the result of sampling error

(Muthén & Asparouhov 2012a)? Compounding matters, when 𝚯𝛿 is estimated using an

informative prior the researcher has to contend with p(p-1)/2 CUVs which can create severe

interpretational challenges when many of these CUVs are large16. Rindskoff’s (2012: 338)

statement is apropos to summarize this problem in the context of Muthén & Asparouhov (2012a)

analysis of a male and female sample of the Big Five where there were 17 and 37 significant

CUVs:

“If I were a personality researcher, I do not think I would be happy with a “Big 5

plus moderate to small 27, give or take 10” theory. If there are supposed to be

five factors, then the number of failures to fit the model (by adding extra

correlations) should be small, and either their numerical value should be small or

there should be a theoretical explanation for why these residuals are correlated.

Of course, this theoretical explanation would be post hoc (if it were known ahead

of time, the model would have included the expected parameters.) In this case,

the theoretical explanation would be tentative and would need to be corroborated

on a different data set.”

16 Given that statistical significance is a function of sample size, we believe that a focus on the magnitude

of the CUVs is more important than whether they are statistically significant. For example, in a large

sample a CUV of 0.10 may be significant whereas in a small sample a CUV of 0.20 may not be significant,

even though the CUV of 0.20 is more important from a practical standpoint.

116

A second reason for apprehension with the estimation of 𝚯𝛿 is that, as will subsequently

be demonstrated, it appears that estimation of this matrix may result in outstanding model fit for

any specified model. For example in Muthén & Asparouhov (2012a) when the authors estimate

𝚯𝛿 the hypothesized structure of the Big Five fits perfectly based on the PPC17 fit criteria.

Furthermore, all factor loadings specified with diffuse priors exhibited large loadings whereas

small loadings specified with zero mean and small variance informative priors were near zero.

The concern with this procedure is that if the estimation of 𝚯𝛿 will result in perfect model fit even

with model misspecification, then the model has little value. MacCallum (2003: 131) captures

this well in regards to the flexibility of models by stating “In practice, if a highly flexible model

fits observed data well, support is still weak, because the model would fit a wide range of data

well. On the other hand, if an inflexible model were found to fit well in an empirical study,

support for that model would be stronger. If two models were found to fit equally well, we

should prefer the one that is less complex or flexible” and further notes “the evaluation of a given

model should take into account the capacity of the model to fit a wide array of data. And models

should be devalued to the extent that they are able to achieve good fit to nearly any data” (133).

17 PPC stands for posterior predictive checking, which is a common way to establish fit for Bayesian CSM

models. Essentially, the PPC is calculated at each k iteration of the MCMC whereby a discrepancy

function is calculated for the parameter estimates from the kth iteration and the observed data and at the

same iteration a simulated dataset of the same size as the sample data is generated from the specified model

and the parameters estimated in the kth iteration. A discrepancy function is also calculated for this

simulated dataset given the model parameters. A PPC confidence interval not containing zero indicates

that the discrepancy function for the observed data given the parameters fits better than the simulated data

given the parameters (Asparouhov & Muthén 2010).

The posterior distribution of the predicted data can be mathematically from Kaplan & Depaoli (2012) as:

𝑝(𝑦𝑟𝑒𝑝|𝑦) = ∫𝑝(𝑦𝑟𝑒𝑝|𝜃) ∗ 𝑝(𝜃|𝑦)𝑑𝜃

As they note, given that 𝑝(𝜃|𝑦) is proportional to 𝑝(𝑦|𝜃) ∗ 𝑝(𝜃), PPC incorporates both uncertainty about

the model parameters and the data. As before, the core idea underlying PPC is that the replicated data

should closely match the observed data, and if not, this indicates a problem with the model. However, it is

important to note that PPC does not take model parsimony into account, unlike the Bayesian and Deviance

Information Criteria (BIC and DIC).

117

Structural Regression Parameters

Regression parameters between observed or latent variables represent a “middle ground”

of sorts with equally-weighted benefits and cautions for using informative priors as summarized

in Table 1 Panel C. The benefit of specifying informative priors is that the incorporation of

information from previous studies (Bolstad’s 2007) allows for more precise estimates of

regression parameters (Yuan & MacKinnon 2009) and allows the researcher to incorporate

his/her knowledge of the substantive area (Rindskopf 2012). This use of prior information is

especially valuable in studies with small sample sizes given this prior information can drastically

improve the precision of the estimated posterior (Yuan & MacKinnon 2009)

At the same time, it is important for researchers to recognize that the results from study to

study are rarely interchangeable due to different samples, operationalizations of measures, and

included covariates. As a result prior information should be discounted by using a larger variance

when specifying informative priors (Yuan & MacKinnon 2009). A second challenge exists when

theory predicts that variable X should have a stronger impact on Y than variable M does on Y.

While such tests are critical to allow for theory pruning (Leavittt et al., 2010) and improve the

precision of our theorizing (Edwards & Berry 2010), researchers face the challenge that they can

(1) let the data speak for itself by using diffuse priors for these regression parameters or (2)

specify informative priors to incorporate the theory’s predictions whereby the X to Y parameter

has a larger mean than the M to Y parameter. The clear tradeoff is that by using the informative

priors the researcher may “force” the model to say what the theory predicts, but at the same time,

incorporation of prior knowledge is a core benefit of Bayesian CSM (Kaplan & Depaoli 2012).

Methodology - Research Setting

As part of a larger scale study on the antecedents and consequences of entrepreneurship a

questionnaire-based survey was administered to the alumni of a large US university in the

118

summer of 2011. Among the measures collected was a multi-dimensional entrepreneurial self-

efficacy scale developed by McGee et al. (2009). The concept of self-efficacy has played a

central role in theories of social learning and social cognition (Wood & Bandura 1989). Self-

efficacy can be adequately summarized as one’s belief in one’s ability to accomplish tasks within

a domain. The expectations and motivation that arises from an individual’s self-efficacy have an

influence on that individuals’ coping behaviors, expended effort, adversity tolerance, goal setting,

and choice of actions (Bandura 1977; Gist 1987). When self-efficacy is used to appraise

individuals’ belief in their personal capabilities related to the formation of a new venture, it is

further delineated as entrepreneurial self-efficacy (abbreviated as ESE) (Boyd & Vozikis 1994).

This specification of self-efficacy is based on the assumption that the entrepreneurial process

involves a range of inter-related tasks that are unique to such a degree that they cannot be readily

captured in a general measure of self-efficacy (Chen, Greene, & Crick 1998).

McGee et al.’s development of an ESE scale specified a five-factor PCS-CFA solution

comprising the dimensions of search, plan, marshal, implement-finance, and implement-people.

This scale conceptualizes the process of entrepreneurship as a multi-staged life cycle. Stevenson,

Roberts, and Grousbeck (1985) proposed a process model that separates new venture creation into

multiple phases: evaluating the opportunity, developing the business concept, acquiring needed

resources, and managing the venture. During the searching phase (evaluating the opportunity,

Stevenson’s et al.’s term) the entrepreneur develops a novel idea or identifies a market

opportunity. As part of this process the entrepreneur relies on their creativity and innovativeness

to explore many alternatives. The planning phase (developing the business concept & assessing

required resources) is focused on formalizing the entrepreneurial concept into an implementable

plan that fits within the entrepreneur’s abilities and goals. During the marshaling phase (acquiring

needed resources) the entrepreneur acts to gain control over the resources needed to implement

the business. The implementing stage (managing and harvesting the venture) is focused on

119

managing the venture and assuring its successful growth past incubation. The implementing stage

has been conceptualized involving both an aspect of managing people (implementing-people) and

managing the finances of the business (implementing-finance).

A CFA analysis presented in their original article showed acceptable fit (CFI= .96, TLI=

.95, RMSEA= .06) with good factor loadings (range 0.70-0.92). However, a pressing concern was

that the inter-factor correlations showed several extreme values (range 0.55-0.94, median= 0.70).

These high factor correlations raise the concern that items might be cross loading and that the

discriminant validity of the proposed factors was not robust. The new data collected as part of this

large-scale study provided the perfect opportunity to contrast the utility of the ICS-CFA approach

with a Bayesian approach to measurement modeling, in the context of a complex multi-factored

measure.

Data Collection

Approximately 70,000 potential respondents were queried for participation via email,

with 7,891 participants completing the survey instruments (a response rate of ~11.3%). As this

study is primarily exploratory it was elected to drop any respondent who did not totally complete

the ESE scale rather than to rely on an imputation strategy. This election was made as imputation

in a Bayesian context is a unique area of inquiry that this article will not attempt to address (see

Rubin, 1996). This list-wise deletion strategy reduced the sample to 6,306 respondents. In order

to focus the measurement model on a consistent population, it was elected to remove those

individuals who reported that they were either retired, unable to work, or were already an

entrepreneur or self-employed. This resulted in a final sample size of 4,041 respondents. For

analysis purposes two random samples of 500 participations was drawn from this final sample, in

order to create sample sizes more in line with average study populations. Note the results

presented here are relatively consistent across sample sizes ranging from 200-500.

120

CFA Analysis

The first step in the analysis was to fit a CFA model to the data to determine if the

originally hypothesized perfect cluster solution displayed acceptable fit and discriminant validity.

The raw data were used as input into Mplus Version 7. (Muthén & Muthén 2012) To provide

comparability to the majority of published studies, the default ML algorithm was used for

estimation18. The standardized factor loadings and inter-factor correlation for this model are

displayed in Table 3.2 (see end of chapter).

Table 3.2 indicates that while all items have high-standardized loadings on their

respective constructs, overall model fit is highly suspect. Specifically, the CFI (0.900) is well

below the 0.95 recommendation (Hu & Bentler 1999), the RMSEA (0.100) point estimate

indicates unacceptable fit based on Browne & Cudeck’s (1992) guidelines, and the SRMR is

large at 0.075. An examination of the inter-factor correlations reveals concerns about

discriminant validity of the RLVs given the large correlations (r > 0.60) between search & plan,

plan & marshal, and marshal & implement people19. Finding that this specification of ESE

displays unacceptable fit illustrates Marsh et al.’s (2010) contentions that many MRCs are

difficult to model using the standard PCS-CFA model. This provided the opportunity to explore

the use of different approaches to Bayesian CSM to successfully model ESE.

Bayesian Analysis – Measurement Model

18 Mplus has several “robust” estimators including the MLR estimator that applies the Sattora-Bentler

correction to the χ2 test static to address nonnormality and utilizes a sandwich estimator to calculate the

standard errors of the estimated parameters (Muthén & Muthén 2012).

19 It should be noted that the correlations between the LVs reported in Table 3.2 are substantially less than

those reported in McGee et al.’s (2009) original article.

121

Given the overlap of the theoretical definitions of the facets of ESE, a Bayesian model

was developed similar to Figure 3.1 Panel B where all posited zero loadings were modeled using

informative priors with a mean of zero and a standard deviation of 0.14120. Since the setting of

the prior is sensitive to the scale of the model, all observed variables were standardized prior to

the analysis, which is permissible given that the model is scale free21. The hypothesized large

loadings were specified using Mplus’ default diffuse normal prior. Estimation was completed

using the default Gibbs sampler. Convergence was evaluated through examination of the PSR,

evaluating trace plots of parameters, and evaluating the autocorrelation plots of parameters. The

median values for the estimated parameters are reported below in Table 3.3 (see end of chapter).

One of the disadvantages of Bayesian CSM is that model fit indices are not as developed,

i.e. fit statistics such as CFI and RMSEA are not available (Levy 2011). Based on the PPC

criteria, model misfit is suggested as the PPC confidence interval does not contain zero [231.2,

339.1], with the model’s DIC = 20,344 and BIC = 21,003. However, as argued by Gelman

(2003) and Levy (2011), PPC can be viewed more as a diagnostic tool than as a measure of model

fit per se. Under this philosophy, the PPC for the model where cross-loadings were estimated fits

better than the model (results not reported) where degenerate priors22 with a value of zero were

utilized, resulting in a PPC confidence interval of [654.5, 750.8].

20 In Mplus normal priors are specified using the mean and variance; for simplicity we set the variance at

0.02, which corresponds to a standard deviation of 0.141.

21 Scale free means that the results from estimating a given model will not change if the observed data is

transformed in a linear fashion (Cudeck 1989). The ML estimator is scale free whereas other estimators

such as ULS (unweighted least squares) are not scale free (Bollen 1989). In the context of all of our

present models we are not imposing parameter restrictions such as equality constraints which allow us to

linearly transform the raw data from its original metric to conduct the analysis. Cudeck (1989) provides a

more in-depth examination of the topic of this issue.

22 Degenerate priors do not have a distribution and are a fixed point (MacCallum et al. 2012). As such,

estimating the original measurement model using degenerate priors with a value of zero is equivalent to

estimating the model reported in Table 3.1 using the Gibbs sampler rather than maximum likelihood. In

other words, this is a PCS-CFA model estimated in a Bayesian framework.

122

Owing to the complexity of this measurement model, i.e. five factors with nineteen

observed measures, it is not surprising that the proposed solution fails to totally replicate the data

even when allowing for cross-loadings given that all models are to some degree incorrect

(Cudeck & Henly 1991; MacCallum 2003). While fit statistics such as CFI and RMSEA are

unavailable in a Bayesian framework, a proxy for SRMR can be calculated for this measurement

model. However, in Bayesian CSM SRMR is actually a distribution (Levy 2011), but Mplus

does not yet have the capability of calculating SRMR during MCMC iterations to develop this

distribution. Given this limitation in this study SRMR is calculated using the median values for

the parameters, which we term pseudo-SRMR (pSRMR). Details about the calculation of

pSRMR are provided in Appendix C. Using the pSRMR to evaluate model fit, the model with

degenerate priors demonstrates poor fit with a pSRMR = 0.075 (equivalent to the ML-CFA

model) whereas for the model with cross-loadings the pSRMR = 0.028, indicating that the model

with cross-loadings better replicates the observed data correlation matrix.

Returning to Table 3.3, calls attention to two findings. First, the correlations between the

LVs are smaller, which is to be expected (Marsh et al. 2009; Muthén & Asparouhov 2012a) and

reduces concerns about discriminant validity. The average correlation is 0.518 and the largest

correlation is 0.714. Second, there is evidence that there are three cross-loadings (S3 on Plan, P1

on Search, and P4 on Marshall) that have practical significance (i.e. > 0.30) and warrant further

examination. Examining the wording of these items in the McGee at al. scale it becomes

apparent that this might be a model in which items should be expected to show theoretically

meaningful cross-loadings. The wording of item S3: “Design a product or service that will satisfy

customer needs and wants” (McGee et al., pg. 978) would appear to refer to both the constructs of

Search (i.e. identifying a new opportunity or market) and Plan (formalizing the entrepreneurial

concept). Likewise, Item P1: “Estimate customer demand for a new product or service” is clearly

related to Plan for a new business, but may also be considered by some to be a central aspect of

123

identifying an opportunity (Search). Lastly, P4: “Design an effective marketing/advertising

campaign for a new product or service” requires the ability to Plan, but the act of engaging in

efficient action can also be tied to Marshal (acquiring needed resources).

Based on reasoning these three important cross-loadings it was elected to conduct the

analysis a second time with these three cross-loadings freely estimated. This was accomplished

by replacing the respective zero mean, small variance informative priors with diffuse priors.

Results in Table 3.4 reveal that freeing up these loadings allows the model to place larger weights

on them; in each case the freely estimated loading is greater than in the previous model. In terms

of model fit, the models are quite similar with the PPC, BIC, and pSRMR being nearly identical.

However, importantly, there is an improvement in the discriminant validity of the LV

correlations. The average correlation is 0.483 and the largest correlation is 0.632 as compared to

the prior Bayesian model with values of 0.518 and 0.714. This substantial improvement leads to

the specification of this model as the final measurement structure.

In order to avoid the possibility that the previous results are an outcome of sample

specific characteristics, a calibration/validation strategy was implemented per the

recommendations of MacCallum et al. (2012). The validation was completed by drawing a

second independent sample of 500 observations. At this time multi-group analysis is not yet

available in Bayesian CSM (it can be partially implemented through mixture models, but that is

beyond the scope of this article). Nonetheless the results in Tables 3.5a & 3.5b reveal that the

measurement model with three complexity two indicators fits both samples in a roughly

equivalent manner.An alternative Bayesian approach to address the misfit of the PCS-CFA model

is to apply a technique not possible using ML approaches. This approach allows for the

estimation of the off-diagonal entries of the 𝚯𝛿 matrix, through specification of an informative

prior. Estimation of 𝚯𝛿 is not possible using ML due to the fact that the model would have

negative degrees of freedom, but is possible in Bayesian approaches due to the difference

124

between frequentist and Bayesian identification (see Endnote x). In order to estimate 𝚯𝛿 one can

follow Muthén & Asparouhov’s (2012a) example of specifying an inverse-Wishart prior

distribution for 𝚯𝛿 where the parameters of this distribution are ~IW(I, p+6), given p is the

number of observed measures.

In order to understand the implications of estimating 𝚯𝛿, a measurement model was

estimated where all observed measures were assumed to have a complexity of one, small cross-

loadings were permitted, and 𝚯𝛿 was freely estimated. This model is equivalent to the one

previously presented in Table 3.3, with the addition of the estimation of all CUVs. The results

from this model are reported in Table 3.6.

This model displays outstanding fit judged by the PPC criteria; the 95% PPC confidence

interval contains zero, indicating the replicated data closely matches the sample data, which is not

surprising given the enormous increase in the number of parameters. The pSRMR value further

indicates excellent fit (pSRMR=0.018). However, examination of the factor loadings reveals an

important concern: with the CUVs estimated evidence that S3 and P4 have a complexity greater

than one is not apparent, which would lead one to conclude on a different measurement structure.

Furthermore, the outstanding model fit raises the concern highlighted earlier: can estimation of

𝚯𝛿 essentially allow the researcher to obtain acceptable fit for the model he/she desires? To

check this concern, the PCS-CFA model where all cross-loadings were specified to have

degenerate priors was run, with the addition of allowing for the estimation of 𝚯𝛿. These results

(not reported) lend credence to this concern: the PPC confidence interval contains zero [-60.0,

54.1] and the pSRMR for this model is outstanding (pSRMR = 0.013).

Bayesian Analysis – Structural Model

Having explored issues in regards to the use of priors for measurement models specified

in a Bayesian CSM context, attention is now turned to their use in relation to structural pathways.

125

In order to demonstrate the previous observation that informative priors can improve parameter

estimation, or when misused can produce corrupted results, three example are presented. In each

case the complexity two measurement model specified in Table 3.4 is utilized for the structure of

ESE. Since prior research has shown that self-efficacy (Bandura, 1991) is predictive of an

individual's intent to engage in an activity, a simple structural model is tested in which the five

dimensions of ESE are used to predict an individual's desire and intent to become an entrepreneur

(measured by four items). It should be noted that the result presented here should not be

interpreted as research results (several important covariates have been excluded in order to

simplify the example), but rather serve to illustrate the principles of specifying structural priors.

Table 3.7 presents results from specifying this model with diffuse priors, informative

priors, and inappropriate priors. The first two of these are used to present a hypothetical case in

which data collected from a pilot or pre-study is used to develop informative priors for

subsequent estimation of a second sample. Using the first group of 500 observations, the

structural model was estimated with diffuse priors as such priors are most appropriate when there

is little prior guidance on potential effects sizes or the researcher wishes to let the data speak the

loudest. The parameters estimates derived from this estimation and the standard deviation of these

estimates were then used to generate informative priors for estimating the same model in the

second group of 500 observations. In-line with Yuan & McKinnon's (2009) recommendation to

inflate the standard deviation specified for the prior a value of 150% of the observed standard

deviation was used.

Examining the results of estimating the structural model in group 2, it can be seen that the

inclusion of the informative priors leads to improvement in the precision of the parameter

estimates for the structural regressions. While the median values are different, which would be

expected since they are from a new sample, the 95% C.I.s on are roughly 73% the width of the

same intervals estimated in group 1with diffuse priors. Note that this is not just an effect of being

126

a different sample, if these same priors were used for estimating the original set of observations a

similar, if not superior, improvement in the width of the 95% C.I.s would be noted.

The last column in Table 3.7 presents an example in which inappropriate priors have

been used in the context of a structural model. This example shows that if a researcher believed

that only the last two dimensions of ESE (implement-people and implement-finance) should be

predictive of desire and intent, that specification of strong priors can be used to force the desired

results. In this case near-degenerate priors were specified for the first three regression pathways

and informative priors with large means were specified for the desired significant pathways.

These strong priors dominate the data, and lead to the radically skewed results. An examination

of the model comparison statistics reveals that this model provides a dramatically inferior fit in

comparison to the prior two examples. However if this comparison was not provided, or the fact

that strong informative priors were used was not revealed, a reader may not be aware of the

sleight-of-hand going on here. While this example is clearly contrived it should highlight the

caution necessary in using informative priors.

Discussion

The increased availability and usability of Bayesian techniques for modeling covariance

structures, especially MRCs, could represent a watershed methodological moment for

management researchers similar to the development and diffusion of ML-based CSM during the

1970s and 1980s. Bayesian approaches afford researchers a degree of flexibility previously

unseen, but as with all new statistical approaches, such flexibility can come at a cost. The

Bayesian CSM paradigm allows for, and indeed advocates for, the specification of measurement

models for MRCs that are not permitted in the current CFA measurement paradigm. This

manuscript has sought to address these tensions by rectifying the differences between the

Bayesian and CFA measurement paradigms along with highlighting the benefits and cautions

127

associated with Bayesian CSM pertaining to the specification of informative priors on factor

loadings, estimation of CUVs, and structural pathways. From the results of the empirical

analysis, in this section two key findings are described in more detail: (1) having reliable

measures trumps unidimensional measures when modeling MRCs and (2) estimating all CUVs

using informative priors should be limited to a diagnostic role and not included as standard

practice until this approach is subjected to further simulation analysis and inquiry.

Reliable Measures over Unidimensional Measures (Within Reason)

The reliability of observed measures can be thought of as the proportion of variance of

the observed measure explained by the set of m latent variables23 (Bollen 1989). Numerous

benefits exist for using highly-reliable measures. First, as analytically demonstrated by

MacCallum & Tucker (1991) and shown empirically via simulation in MacCallum et al. (2001),

the primary driver of model misfit arises from items having low reliability24. Second, holding

sample size constant, there is a greater probability of factor extraction when items have a high

23 In CSM there are three sources of variance for an observed measure: (1) common variance (termed

communality) explained by the latent variable, (2) specific variance, which is variance associated with an

individual item, and (3) error variance, which arises from imperfect measurement. We will define common

variance as 𝜎𝐶2, specific variance as 𝜎𝑆

2, and error variance as 𝜎𝐸2. The total variance of an item (𝜎𝑇

2) is the

sum of these three sources of variance.

The reliability of an observed measure is the ratio of the common and specific variance over the measure’s

total variance. However, in most CSM applications specific variance is unknown, and thus specific

variance and error variance are summed together in what is termed unique variance (𝜎𝑈2). This unique

variance is the diagonal entry on the 𝚯𝛿 matrix. As such, the statement that an item’s reliability can be

found as 𝜎𝐶2/𝜎𝑇

2 is an underestimate of the true reliability as the specific variance is not included in the

numerator (Bollen 1989).

24 Two important assumptions in the common factor model are (1) unique variances are uncorrelated and

(2) there is no correlation between the latent factors and unique variances (MacCallum & Tucker 1991).

However, these assumptions apply to the population model; when fitting the common factor model to a

sample covariance or correlation matrix, these assumptions are unlikely to hold due to sampling variability.

Given that a low reliability of an item implies that the elements of the vector of unique variances will be

large, violating these assumptions when there is low reliability is more serious when there is high

reliability.

128

reliability (MacCallum et al. 1999). Third, high reliability items reduce the negative

consequences of multicollinearity in CSM, with Grewal, Cote, and Baumgartner (2004: 527)

stating “Probably the most important safeguard against the damaging effects of multicollinearity

is to make sure that all constructs are measured as reliably as possible.” Consistent with our

views, Asparouhov & Muthén (2009: 430) state “One can argue that it is more important to find

accurate measurements than to find a pure set of measurements.”

Despite the importance of utilizing highly reliable measures, several caveats need to be

made to the argument that highly reliable measures are more important than having

unidimensional measures. First, this argument is most applicable when studying MRCs given

that cross-loadings are more likely to be theoretically permissible due to measures having overlap

with interrelated RLVs’ theoretical definitions. Such an example occurred with the S3 and P1

measures loading onto the search and plan RLVs. As the McGee et al. scale is derived from a life-

stage model of the entrepreneurial process, one would expect there to be some theoretical bleed

over between constructs representing various stages. Further, according to R2 (0.58 for S3 and

0.64 for P1), the complexity two measurement structure explains a large portion of the observed

variance. Given the theoretical justification for these cross-loadings, we would not want to

discard these measures25 given the large R2. Second, following Thurstone (1947), no measure

should load onto all m RLVs given the minimum requirement for simple structure, according to

Browne (2001), is that a measure can have at most m-1 large loadings. Third, the factor loading

matrix should be easily interpretable, which is purpose of identifying a simple structure

(Thurstone 1947). Fourth, while cross-loadings are permissible, a measure should display a large

25 Hair et al. (2010: 675) contend that “evidence that a significant cross-loading exists…shows a lack of

discriminant validity.” This statement is not valid as discriminant validity is the property of the RLV, not

manifest measures. Their statement is consistent with the aforementioned Anderson et al. (1987) quote that

the empirical operationalization of an RLV determines its definition, which has already been shown to be

inconsistent with factor analytic theory (Browne 2001; Thurstone 1947).

129

loading on the RLV it is intended to empirically operationalize. Absence of a hypothesized large

loading suggests interpretation confounding (Burt 1976), which implies that a disconnection

between the theoretical concept the RLV represents and its empirical operationalization26. Thus,

one should not interpret the argument that reliable measures are more valuable than

unidimensional measures as an argument to blindly utilize observed measures with the greatest

explained variance, per these above reasons.

Correlated Unique Variances: Concerns & Recommendations

As previously noted, the ability to freely estimate all CUVs using an informative inverse-

Wishart prior is unique to Bayesian CSM as such an analysis would be impossible using

frequentist approaches. However, this is an example of new modeling flexibility that raises

important theoretical concerns. First, an inherent assumption necessary to estimate 𝚯𝛿 and

maintain that the hypothesized factor loadings and specified RLV relationships have theoretical

meaning is that model misfit arises solely from sampling error and unimportant minor factors.

This assumption allows the researcher to treat 𝚯𝛿 as a “vacuum” that captures all noise arising

26 One can think of interpretational confounding as occurring when an observed measure that is posited to

be influenced by a RLV does not exhibit a large loading. As noted by Bollen (1989), the first process for

developing items representative of a RLV is to articulate a theoretical definition of the RLV to establishing

the meaning of the concept the RLV represents. Given this theoretical definition, the researcher then

develops a set of observed measures that are expected to be highly correlated given the theoretical

definition of the RLV (Bollen & Lennox 1991). Interpretational confounding would occur is a measure

that is expected to be highly correlated with the other items is not, and thus would display a small factor

loading. The theoretical issue that arises is that the observed measure and theoretical definition of the RLV

are thus disconnected because, given the definition of the RLV, the researcher would have expected the

observed measure to be highly correlated with the other measures. This then raises the question of why the

measure was not correlated with the other measures, which implies that the original auxiliary measurement

theory is in need of modification.

A second example of interpretational confounding would occur if a researcher articulates a theoretical

definition for an RLV but then operationalizes the RLV using a set of measures that are not consistent with

the theoretical definition of the RLV. In this instance there is interpretational confounding even if the

measures are highly correlated because there is little correspondence between the observed measures and

the theoretical definition of the RLV.

130

from model error27 and sampling error (Cudeck & Henley 1991; MacCallum and Tucker 1991).

Thus estimation of 𝚯𝛿 is problematic if model misfit is the result of the researcher failing to

model theoretically-meaningful effects, such as including additional structural paths between

RLVs or modeling an additional RLV. The previous examples demonstrated this in that

estimating 𝚯𝛿 would lead the researcher to affirm that a PCS-CFA model fits the data perfectly.

A second problem as shown in Muthén & Asparouhov’s (2012a) Big Five example and

with the models reported earlier where 𝚯𝛿 was estimated is that the PPC criterion and pSRMR

are rendered “worthless” measures of fit as estimation of the 𝚯𝛿 allows the researcher’s specified

model to replicate the data (covariance matrix). Appendix D further examines this issue by

asking the question can estimation of the 𝚯𝛿 matrix allow a grossly ill-specified model to

nonetheless recreate the underlying data. The ability of a CSM model with 𝚯𝛿 estimated to fit

any data raises a concern parallel to Roberts and Pashler’s (2000: 359) caution about model fit

affirming theory “Theorists who use good fit as evidence seem to reason as follows: If our theory

is correct, it will be able to fit the data; our theory fits the data therefore it is more likely that our

theory is correct. However, if a theory does not constrain possible outcomes, the fit is

meaningless.” Their logic drove the aforementioned quote from MacCallum (2003) that a model

that can fit any data has little value.

Since the fit of a model with 𝚯𝛿estimated has little practical meaning, does this technique

in fact have any value? Our position is that it is too early to answer this question definitively, as

further research is needed to identify what additional knowledge may be gleamed from estimating

all CUVs. For the time being we strongly recommend against the use of this technique as a

27 Traditionally it is assumed that the common factor model fits perfectly in the population and thus all

misfit is a function of random sampling. However, as MacCallum & Tucker (1991) and Cudeck & Henley

(1991) argue, the assumption that the posited model fits perfectly in the population is untenable due to

many factors such as failure to include minor factors, nonlinear relationships, and violating the assumption

that all observations are homogeneous (MacCallum et al. 2001).

131

method for validating measurement structures. However with adequate caution to avoid sample-

dependent alterations, one potential use for this technique might be to help identify why a

measurement scale is not operating according to a priori expectations. Figure 3.2 and 3.3 present

matrix-density plots of the estimated 𝚯𝛿 matrix for the model specified as a PCS-CFA structure

(Figure 3.2) and the model where all observed measures had a complexity of one and cross-

loadings were specified with informative priors (Figure 3.3). These plots represent the estimated

CUVs via gray tones, with darker colors indicating larger values. As can be seen in the plot from

the PCS-CFA Model, there are many large CUVs, with several concentrated clusters. All of these

non-white colors indicate areas where the proposed model was unable to accurately explain the

relationships between the observed measures. The plot of the model that includes cross-loadings

shows that more of the underlying relationships are being explained by the model and that most

of the problematic clusters have been resolved.

132

Figure 3.2: Density Plot of 𝚯𝛿Matrix for the estimated PCS-CFA model. Darker off-diagonal

elements indicate entries with a larger absolute value. The numbers 1-19 are in numerical order

for the observed measures, thus 1 = S1, 2 = S2,...,19 = IF3.

1 5 10 15 19

1

5

10

15

19

1 5 10 15 19

1

5

10

15

19

133

Figure 3.3: Density Plot of 𝚯𝛿Matrix for the estimated model where all observed measures had a

complexity of one and cross-loadings were specified using informative priors. Darker off-

diagonal elements indicate entries with a larger absolute value. The numbers 1-19 are in

numerical order for the observed measures, thus 1 = S1, 2 = S2,...,19 = IF3.

1 5 10 15 19

1

5

10

15

19

1 5 10 15 19

1

5

10

15

19

134

Limitations

In this manuscript attention has been purposely focused on the theoretical meanings and

implications inherent in new options provided by a Bayesian approach to CSM. In order to avoid

a long side-track several additional issues have been left uncovered or passed over quickly. The

actual implementation and execution of MCMC and the Gibbs sampler was left unexplored,

several excellent sources are available for readers interested in the technicalities underlying these,

see Asparouhov & Muthén (2010), Dunson et al. (2005), Edwards (2010), and Lee (2007).

Further examination of measurement models was restricted to cross-sectional, reflective, multi-

dimensional scales. Alternatives to such specifications exist including the bi-factor model,

MIMCs, and various mixture models. In regards to longitudinal models the issues addressed here

are still pertinent, although additional complexities are introduced. One of these issues, which is

also relevant in multi-group research designs, is the role of invariance (both measurement and

temporal). Muthén & Asparouhov (2013), along with other scholars, are examining issues related

to the role of near-invariance. In much the same way that informative priors on cross-loadings

removes the restrictive assumption of unidimnesionality, such approaches relax strict invariance

assumptions.

Conclusions

Modeling covariance structures using Bayesian approaches, particularly the

BSEM technique outlined by Muthén & Asaporouhov (2012a), combined with the

increased synthesis of categorical and continuous latent variable models (Muthén 2002),

appears to herald the dawn of a second generation of CSM (Kaplan & Depaoli 2012).

However, to best leverage the flexibility provided by Bayesian CSM, as a community we

must rectify the contradictory arguments of the Bayesian CSM paradigm and the current

CFA paradigm and establish the boundaries of modeling flexibility to ensure that models

135

remain theoretically meaningful. This manuscript has sought to address these issues by

arguing that Bayesian CSM is more consistent with factor analytic theory than the current

CFA-dominated approach and articulating the theoretical implications for utilizing

informative priors to estimate cross-loadings and correlated unique variances. To

conclude, we would like to leave the reader with the following quotation from Cudeck &

Henly (1991: 512) which provides an elegant summary of our core points:

“In the study of mathematical models, the process of developing and justifying a

model is the most fundamental of issues, because every other feature associated

with the use of quantitative models is influenced by the final form of the

structure. Yet no model is completely faithful to the behavior under study.

Models usually are formalizations of processes that are extremely complex. It is

a mistake to ignore either their limitations or their artificiality. The best one can

hope for is that some aspect of a model may be useful for description, prediction,

or synthesis. The extent to which this is ultimately successful, more often than

one might wish, is a matter of judgment.”

136

References: Chapter 3

Anderson, J. C., Gerbing, D. W., & Hunter, J. E. (1987). On the Assessment of Unidimensional

Measurement: Internal and External Consistency, and Overall Consistency Criteria.

Journal of Marketing Research, 24(4), 432-437.

Asparouhov, T., & Muthén, B. (2010). Bayesian Analysis Using Mplus: Technical

Implementation. Los Angeles, CA: Muthén & Muthén.

Bagozzi, R. P. (2007). On the Meaning of Formative Measurement and How It Differs From

Reflective Measurement: Comment of Howell, Breivik, and Wilcox (2007).

Psychological Methods, 12(2), 229-237.

Bagozzi, R. P. (2011). Measurement and Meaning in Information Systems and Organizational

Research: Methodological and Philosophical Foundations. MIS Quarterly, 35(2), 261-

292.

Bandura, A. (1977). Self-Efficacy: Toward a Unifying Theory of Behavior Change.

Psychological Review, 84(2), 191-215.

Bollen, K. A. (1989). Structural Equations with Latent Variables. New York: Wiley.

Bollen, K. A. (2011). Evaluating Effect, Composite, and Causal Indicators in Structural Equation

Models. MIS Quarterly, 35(2), 359-372.

Bollen, K., & Lennox, R. (1991). Conventional Wisdom on Measurement. Psychological

Bulletin, 110(2), 305-314.

Bolstad, W. M. (2007). Introduction to Bayesian Statistics (Second ed.). Hoboken, NJ: John

Wiley & Sons.

Borsboom, D., Mellenbergh, G. J., & van Heerden, J. (2003). The Theoretical Statis of Latent

Variables. Psychological Review, 110(2), 203-219.

Boyd, N. G., & Vozikis, G. S. (1994). The Influence of Self-Efficacy on the Development of

Entrepreneurial Intentions and Actions. Entrepreneurship Theory & Practice, 18(4), 63-

77.

Browne, M. W., & Cudeck, R. (1992). Alternative Ways of Assessing Model Fit. Sociological

Methods & Research, 21(2), 230-258.

Browne, M. W., & Mels, G. (1998). Path Analysis: Ramona. In: SYSTAT for Windows:

Advanced Applications (Version 8). Evanston, IL: SYSTAT.

137

Burt, R. S. (1976). Interpretational Confounding in Unobserved Variables in Structural Equation

Models. Sociological Methods and Research, 5(1), 3-52.

Chen, C. C., Greene, P. G., & Crick, A. (1998). Does Entrepreneurial Self-Efficacy Distinguish

Entrepreneurs from Managers? Journal of Business Venturing, 13(4), 295-316.

Cudeck, R. (1989). Analysis of Correlation Matrices Using Covariance Structure Models.

Psychological Bulletin, 105(2), 317-327.

Cudeck, R., & Henly. (1991). Model Selection in Covariance Strucutres Analysis and the

"Problem" of Sample Size: A Clarification. Psychological Bulletin, 109(3), 512-519.

Dunson, D. B., Palomo, J., & Bollen, K. (2005). Bayesian Structural Equation Modeling

(Technical Report). Research Triangle Park, NC: Statistical and Applied Mathematical

Sciences Institute.

Edwards, J. R. (2011). The Fallacy of Formative Measurement. Organizational Research

Methods, 14(2), 370-388.

Edwards, J. R., & Bagozzi, R. P. (2000). On the Nature and Direction of Relationships Between

Constructs and Measures. Psychological Methods, 5(2), 155-174.

Edwards, M. C. (2010). A Markov Chain Monte Carlo Approach to Confirmatory Item Factor

Analysis. Psychometrika, 75(3), 474-497.

Gelman, A. (2003). A Bayesian Formulation of Exploratory Data Analysis and Goodness-of-Fit

Testing. International Statistical Review, 71(2), 369-382.

Geman, S., & Geman, D. (1984). Stochastic Relaxation, Gibbs Distributions, and the Bayesian

Restoration of Images. IEEE Transactions on Pattern Analysis and Machine Intelligence,

6, 721-741.

Gerbing, D. W., & Anderson, J. C. (1984). On the Meaning of Within-Factor Correlated

Measurement Errors. Journal of Consumer Research, 11(1), 572-580.

Gerbing, D. W., & Anderson, J. C. (1988). An Updated Paradigm for Scale Development

Incorporating Unidimensionality and Its Assessment. Journal of Marketing Research,

25(2), 186-192.

Gist, M. E. (1987). Self-Efficacy: Implications for Organizational Behavior and Human Resource

Management. Academy of Management Review, 12(3), 472-485.

Grewal, R., Cote, J. A., & Baumgartner, H. (2004). Multicollinearity and Measurement Error in

Structural Equation Models: Implications for Theory Testing. Marketing Science, 23(4),

519-529.

Hair, J. F., Black, W. C., Babin, B. J., & Anderson, R. E. (2010). Multivariate Data Analysis (7th

ed.). New York: Prentice Hall.

Heide, J. B., & John, G. (1992). Do Norms Matter in Marketing Relationships. Journal of

Marketing, 56(2), 32-44.

138

Holzinger, K. J., & Swineford, F. (1937). The Bi-Factor Method. Psychometrika, 2(1), 41-54.

Howell, R. D., Breivik, E., & Wilcox, J. B. (2007). Reconsidering Formative MEasurement.

Psychological Methods, 12(2), 205-218.

Hu, L.-t., & Bentler, P. M. (1999). Cutoff Criteria for Fit Indexes in Covariance Structure

Analysis: Conventional Criteria Versus New Alternatives. Structural Equation Modeling,

6(1), 1-55.

Kaplan, D., & Depaoli, S. (2012). Bayesian Structural Equation Modeling. In R. H. Hoyle (Ed.),

Handbook of Structural Equation Modeling. New York: Guilford Press.

Kuhn, T. S. (1977). The Essential Tension: Selected Studies in Scientific Tradition and Change.

Chicago, IL: University of Chicago Press.

Leavitt, K., Mitchell, T. R., & Peterson, J. (2010). Theory Pruning: Strategies to Reduce Our

Dense Theoretical Landscape. Organizational Research Methods, 13(4), 644-667.

Lee, S.-Y. (2007). Structural Equation Modeling: A Bayesian Approach. West Sussex, UK:

Wiley.

Levy, R. (2011). Bayesian Data-Model Fit Assessment for Structural Equation Modeling.

Structural Equation Modeling, 18(4), 663-685.

MacCallum, R. C. (2003). Working with Imperfect Models. Multivariate Behavioral Research,

38(1), 113-139.

MacCallum, R. C., & Tucker, L. R. (1991). Representing Sources of Error in the Common-Factor

Model: Implications for Theory and Practice. Psychological Bulletin, 109(3), 502-511.

MacCallum, R. C., Edwards, M. C., & Cai, L. (2012). Hopes and Cautions in Implementing

Bayesian Structural Equation Modeling. Psychological Methods, 17(3), 340-345.

MacCallum, R. C., Widaman, K. F., Preacher, K. J., & Hong, S. (2001). Sample Size in Factor

Anlaysis: The Role of Model Error. Multivariate Behavioral Research, 36(4), 611-637.

MacCallum, R. C., Widaman, K. F., Zhang, S., & Hong, S. (1999). Sample Size in Factor

Analysis. Psychological Methods, 4(1), 84-99.

Marsh, H. W., Muthén, B., Asparouhov, T., Lüdtke, O., Robitzsch, A., Morin, A. J., &

Trautwein, U. (2009). Exploratory Structural Equation Modeling, Integrating CFA and

EFA: Application to Students' Evaluations of University Teaching. Structural Equation

Modeling, 16(3), 439-476.

Marsh, H. W., Muthén, B., Morin, A. J., Lüdtke, O., Asparouhov, T., Trautwein, U., &

Nagengast, B. (2010). A New Look at the Big Five Factor Structure Through Exploratory

Structural Equation Modeling. Psychological Assessment, 22(3), 471-491.

McGee, J. E., Peterson, M., Mueller, S. L., & Sequeira, J. M. (2009). Entrepreneurial Self-

Efficacy: Refining the Measure. Entrepreneurship Theory & Practice, 33(4), 965-988.

139

Muthén, B. O. (2002). Beyond SEM: General Latent Variable Modeling. Behaviormetrika, 29(1),

81-117.

Muthén, B., & Asparouhov, T. (2012a). Bayesian Structural Equation Modeling: A More Flexible

Representation of Substantive Theory. Psychological Methods, 17(3), 313-335.

Muthén, B., & Asparouhov, T. (2012b). Rejoinder to MacCallum, Edwards, and Cai (2012) and

Rindskopf (2012): Mastering a New Method. Psychological Methods, 17(3), 346-353.

Muthén, B., & Asparouhov, T. (2013). BSEM Measurement Invariance Analysis (Mplus Web

Notes: No. 17). Los Angeles, CA: Muthén & Muthén.

Muthén, L. K., & Muthén, B. O. (2012). Mplus User's Guide (Seventh ed.). Los Angeles, CA:

Muthén & Muthén.

Myers, N. D., Ahn, S., & Jin, Y. (2013). Rotation to a Partially Specified Target Matrix in

Exploratory Factor Analysis: How Many Targets? Structural Equation Modeling, 20(1),

131-147.

Organ, D. W. (1988). Organizational Citizenship Behavior: The Good Soldier Syndrome.

Lexington, MA: Lexington Books.

Poole, M. S., & Van de Ven, A. H. (1989). Using Paradox to Build Management and

Organization Theories. Academy of Management Review, 14(4), 562-578.

Reise, S. P. (2012). The Rediscovery of Bifactor Measurement Models. Multivariate Behavioral

Research, 47(5), 667-696.

Rindskopf, D. (2012). Next Steps in Bayesian Structural Equation Models: Comments on,

Variations of, and Extensions to Muthén and Asparouhov (2012). Psychological

Methods, 17(3), 336-339.

Roberts, S., & Pashler, H. (2000). How Persuasive Is a Good Fit? A Comment on Theory Testing.

Psychological Review, 107(2), 358-367.

Rubin, D. B. (1996). Multiple Imputation After 18+ Years. Journal of the American Statistical

Association, 91(434), 473-489.

Rupp, A. A., Dey, D. K., & Zumbo, B. D. (2004). To Bayes or Not to Bayes, From Whether to

When: Applications of Bayesian Methodology to Modeling. Structural Equation

Modeling, 11(3), 424-451.

Scheines, R., Hoijtink, H., & Boomsma, A. (1999). Bayesian Estimation and Testing of Structural

Equation Models. Psychometrika, 64(1), 37-52.

Stevenson, H. H., Roberts, M. J., & Grousbeck, H. I. (1985). New Business Ventures and the

Entrepreneur. Burr Ridge, IL: Richard D Irwin.

Thurstone, L. L. (1947). Multiple Factor Analysis. Chicago: University of Chicago Press.

Wood, R., & Bandura, A. (1989). Social Cognitive Theory of Organizational Management.

Academy of Management Review, 14(3), 361-384.

140

Yates, A. (1987). Multivariate Exploratory Data Analysis: A Perspective on Exploratory Factor

Analysis. Albany, NY: State University of New York Press.

Yuan, Y., & MacKinnon, D. P. (2009). Bayesian Mediation Analysis. Psychological Methods,

14(4), 301-322.

141

14

1

Parameter Type Panel A: Factor Loadings Panel B: Correlated Unique Variances Panel C: Regression Parameters

Benefits

•Ability to more realistically specify the

measurement model given the imprecision of

most scales by specifying small-variance

priors for loadings that are expected to be

small.

•Ability to incorporate prior information from

previous studies and/or pilot studies

concerning the magnitude of loadings.

•No clearly known theoretical benefits

•Potential benefits for diagnostics and scale

development, but utility unknown and not

validated.

•Ability to incorporate prior information from

previous and/or pilot studies to improve the

precision of estimates and incorporate

substantive knowledge

Cautions

•Use of priors with too large a variance for

expected small loadings can result in the

model not being identified.

•If utilizing a set of different small-variance

priors, researchers should report which priors

resulted in the best model fit and whether

substantive conclusions about the

measurement model change depending on the

choice of prior.

•Specification of priors with a limited range

(i.e. uniform) could be utilized to "force"

problematic loadings to fit the researcher's

hypothesized model.

•Cross-loadings must have theoretical

justification

•Conceptual ambiguity about the theoretical

meaning of a correlated unique variance—the

correlated unique variance could be the result

of sampling error, a common method factor, or

another underlying latent variable

•Can result in models with limited theoretical

meaning

•Potential evidence that freeing the Θδ matrix

may result in perfect model fit for the

researcher's specified model

•To our knowledge no simulation study has

examined the theoretical implications for

freeing the Θδ matrix, with current simulation

only examining the robustness of different

sampling algorithms

•Important to consider the equivalence of

previous studies to specify priors; in most

cases given heterogeneous samples and

variables, prior information should be

discounted

•Researchers can, knowingly or unknowingly,

specify small-variance priors that "force"

structural pathways to be statistically (non)-

significant rather than letting the new data be

the primary driver of inference

Table 3.1: Benefits and cautions from specifying informative priors on factor loadings (Panel A), correlated unique variances (Panel B), and structural

pathways (Panel C)

142

PCS-CFA Model (ML)

Search Plan Marshal

Implement

People

Implement

Finance

S1 0.887

S2 0.942

S3 0.675

P1 0.716

P2 0.783

P3 0.763

P4 0.707

M1 0.851

M2 0.738

M3 0.781

IP1 0.880

IP2 0.848

IP3 0.851

IP4 0.777

IP5 0.789

IP6 0.736

IF1 0.913

IF2 0.963

IF3 0.817

Fit Statistics: χ2=845.0; DF=142; RMSEA=0.100; CFI=0.900; SRMR=0.075

Latent Variable Correlation Matrix

Search Plan Marshal

Implement

People

Implement

Finance

Search 1

Plan 0.633 1

Marshal 0.569 0.789 1

Implement

People 0.420 0.524 0.639 1

Implement

Finance 0.269 0.617 0.443 0.443 1

Table 3.2: PCS-CFA measurement model fitted using the ML estimator.

143

Bayesian Model w/ Small Variance Cross-Loadings (Original Structure)

Search Plan Marshal

Implement

People

Implement

Finance

S1 0.922 0.017 -0.056 -0.028 -0.009

S2 1.007 -0.104 -0.008 0.010 0.032

S3 0.415 0.326 0.096 0.057 -0.031

P1 0.402 0.430 0.077 0.026 -0.046

P2 -0.037 0.988 -0.187 0.033 0.000

P3 -0.087 0.788 -0.025 -0.035 0.157

P4 -0.059 0.520 0.366 -0.064 -0.030

M1 0.015 0.070 0.825 -0.023 -0.009

M2 -0.057 0.092 0.655 0.077 -0.011

M3 0.023 -0.103 0.771 0.075 0.064

IP1 -0.051 -0.022 -0.067 0.969 -0.007

IP2 -0.113 0.021 0.041 0.858 0.015

IP3 -0.077 0.045 -0.113 0.955 -0.029

IP4 0.124 -0.015 0.014 0.699 0.054

IP5 0.049 -0.013 0.073 0.759 -0.061

IP6 0.045 -0.122 0.045 0.745 0.013

IF1 0.028 -0.086 0.043 0.025 0.927

IF2 0.028 -0.025 -0.003 -0.056 1.012

IF3 -0.054 0.126 -0.055 0.028 0.771

Free Para.=143; PPC= 231.2-339.1; DIC=20,344; BIC=21,003; pSRMR=0.028

Latent Variable Correlation Matrix

Search Plan Marshal

Implement

People

Implement

Finance

Search 1

Plan 0.586 1

Marshal 0.565 0.714 1

Implement

People 0.448 0.520 0.624 1

Implement

Finance 0.259 0.589 0.404 0.466 1

Table 3.3: Bayesian model with informative priors specified for cross-loadings. Factor loadings

in bold were freely estimated using diffuse priors.

144

Modified Bayesian Model w/ Small Variance Cross-Loadings

Search Plan Marshal

Implement

People

Implement

Finance

S1 0.911 0.060 -0.059 -0.039 -0.001

S2 0.984 -0.039 -0.029 -0.002 0.042

S3 0.437 0.389 0.081 0.056 -0.063

P1 0.470 0.415 0.087 0.020 -0.053

P2 0.051 0.888 -0.088 0.034 -0.014

P3 -0.017 0.689 0.068 -0.035 0.149

P4 -0.022 0.406 0.511 -0.096 -0.028

M1 0.032 0.062 0.842 -0.032 -0.018

M2 -0.040 0.082 0.671 0.069 -0.021

M3 0.025 -0.091 0.763 0.066 0.059

IP1 -0.041 -0.013 -0.057 0.951 -0.002

IP2 -0.099 0.017 0.059 0.841 0.020

IP3 -0.062 0.056 -0.106 0.942 -0.028

IP4 0.132 0.011 0.006 0.689 0.055

IP5 0.059 -0.004 0.080 0.742 -0.057

IP6 0.045 -0.102 0.039 0.731 0.020

IF1 0.016 -0.069 0.033 0.029 0.925

IF2 0.017 -0.013 -0.010 -0.049 1.008

IF3 -0.048 0.115 -0.040 0.033 0.768

Free Para.=143; PPC= 227.4-334.9; DIC=20,339; BIC=21,001; pSRMR=0.031

Latent Variable Correlation Matrix

Search Plan Marshal

Implement

People

Implement

Finance

Search 1

Plan 0.446 1

Marshal 0.549 0.632 1

Implement

People 0.429 0.452 0.618 1

Implement

Finance 0.239 0.597 0.415 0.452 1

Table 3.4: Modified Bayesian model with informative priors specified for cross-loadings. Factor

loadings in bold were freely estimated using diffuse priors.

145

Group 1 Group 2

Latent

Factor

Range of

Primary

Loadings

Range of Cross-

Loadings

Range of

Primary

Loadings

Range of Cross-

Loadings

min max min max min max min max

Search 0.437 0.984 -0.099 0.132 0.440 0.931 -0.110 0.155

Plan 0.389 0.888 -0.102 0.115 0.318 0.978 -0.065 0.092

Marshal 0.511 0.842 -0.106 0.087 0.503 1.050 -0.090 0.152

Imp-People 0.689 0.942 -0.096 0.069 0.650 0.976 -0.094 0.177

Imp-Finance 0.768 1.008 -0.063 0.149 0.784 1.003 -0.086 0.169

Group 1 Group 2

Free Parameters 143 143

PPC 231.2 - 339.1 227.4 - 334.9

DIC 20344 20339

BIC 21002 21001

pSRMR 0.034 0.035

Table 3.5: Comparison of Modified Bayesian model between Group 1 (calibration) and Group 2

(validation). Top portion shows the range of observed primary and cross loadings. Bottom portion

shows the model fit results.

146

Bayesian Model w/ Small Variance Crossloadings & CU

Search Plan Marshal

Implement

People

Implement

Finance

S1 0.937 -0.005 -0.049 -0.035 -0.011

S2 0.912 -0.044 0.004 0.016 0.002

S3 0.463 0.219 0.115 0.060 0.035

P1 0.323 0.464 0.082 0.046 -0.023

P2 -0.033 0.887 -0.065 0.005 0.028

P3 -0.057 0.812 -0.033 -0.028 0.131

P4 -0.035 0.767 0.112 -0.014 -0.064

M1 0.054 0.098 0.755 0.003 -0.014

M2 -0.049 0.002 0.846 0.005 -0.008

M3 -0.013 -0.072 0.864 0.036 0.024

IP1 -0.013 -0.029 -0.037 0.907 0.005

IP2 -0.066 0.023 0.038 0.831 0.011

IP3 -0.052 -0.019 -0.037 0.903 -0.006

IP4 0.067 -0.011 0.034 0.755 0.037

IP5 0.033 0.009 0.026 0.809 -0.047

IP6 -0.007 -0.036 -0.044 0.845 -0.026

IF1 0.006 -0.040 0.008 0.037 0.906

IF2 0.006 0.020 -0.009 -0.025 0.920

IF3 -0.022 0.033 -0.012 -0.003 0.852

Free Para.=314; PPC= -58.6-58.1; DIC=20,134; BIC=21,699; pSRMR=0.018

Correlation Matrix: BSEM Model w/ Small Variance Crossloadings

Search Plan Marshal

Implement

People

Implement

Finance

Search 1

Plan 0.579 1

Marshal 0.545 0.669 1

Implement

People 0.451 0.497 0.605 1

Implement

Finance 0.288 0.564 0.425 0.459 1

Table 3.6: Bayesian model with informative priors specified for factor loadings and correlated

unique variances. Factor loadings in bold were freely estimated using diffuse priors.

147

Diffuse Prior (Grp 1) Informative Prior (Grp 2) Inappropriate Prior (Grp 1)

Free Para. 160 160 160

Est. Para. 126 124 254

PPC 279-397 274-394 350-1559

BIC 26095 25990 26427

DIC 25360 25252 25941

Median 95% C.I. Median 95% C.I. Median 95% C.I.

Search 0.342 0.196 0.468 0.368 0.265 0.459 0.038 -0.025 0.100

Plan 0.451 0.337 0.549 0.394 0.314 0.467 0.013 -0.074 0.098

Marshall 0.351 0.217 0.473 0.368 0.283 0.447 0.042 -0.021 0.109

Imp-Ppl 0.248 0.118 0.369 0.221 0.125 0.316 0.502 0.402 0.588

Imp-Fn 0.350 0.216 0.464 0.250 0.147 0.344 0.680 0.504 0.762

Table 3.7: Demonstration of priors in the context of a structural model. Median parameter estimates are the regression pathways between each

latent factor of ESE and the individual's desire and intent to become an entrepreneur.

14

7

148

References

Aldrich, H. E., & Fiol, C. M. (1994). Fools rush in? The institutional context of industry creation.

Academy of management review, 19(4), 645-670.

Allen, T.D., Eby, L.T., Poteet, M.L., Lentz, E., & Lima, L. (2004). Career benefits associated

with mentoring for protégés: a meta-analysis. Journal of Applied Psychology, 89, 127-

136.

Alvarez, S. A., & Barney, J. B. (2005). How do entrepreneurs organize firms under conditions of

uncertainty?. Journal of management, 31(5), 776-793.

Alvarez, S. A., & Barney, J. B. (2007). Discovery and creation: Alternative theories of

entrepreneurial action. Strategic entrepreneurship journal, 1(1‐2), 11-26.

Alvarez, S. A., & Barney, J. B. (2010). Entrepreneurship and epistemology: The philosophical

underpinnings of the study of entrepreneurial opportunities. The Academy of

Management Annals, 4(1), 557-583.

Alvarez, S. A., Barney, J. B., & Anderson, P. (2013). Forming and exploiting opportunities: The

implications of discovery and creation processes for entrepreneurial and organizational

research. Organization Science, 24(1), 301-317.

Alvarez, S. A., Young, S. L., & Woolley, J. L. (2015). Opportunities and institutions: a co-

creation story of the king crab industry. Journal of Business Venturing, 30(1), 95-112.

Amabile, T. M., Conti, R., Coon, H., Lazenby, J., & Herron, M. (1996). Assessing the work

environment for creativity. Academy of management journal, 39(5), 1154-1184.

Amabile, T., & Mueller, J. (2008). Studying Creativity, its Processes, and its Antecedents. In J.

Zhou, & C. Shalley, Handbook of Organizational Creativity (pp. 33-64). New York:

Lawrence Erlbaum Associates.

Anderson, J. C., Gerbing, D. W., & Hunter, J. E. (1987). On the Assessment of Unidimensional

Measurement: Internal and External Consistency, and Overall Consistency Criteria.

Journal of Marketing Research, 24(4), 432-437.

Arbuckle, J.L. (2010) Amos (Version 19) [Computer Program]. SPSS, IBM.

Armitage, C.J., & Conner, M. (2001). Efficacy of the theory of planned behavior: a meta-analytic

review. The British Journal of Social Psychology, 40(4), 471-499.

Arthur, W. B. (1989). Competing technologies, increasing returns, and lock-in by historical

events. The economic journal, 99(394), 116-131.

149

Aryee, S., & Chay, Y.W. (1994). An examination of the impact of career-oriented mentoring on

work commitment attitudes and career satisfaction among professional and managerial

employees. British Journal of Management, 5, 241-249.

Asparouhov, T., & Muthén, B. (2010). Bayesian Analysis Using Mplus: Technical

Implementation. Los Angeles, CA: Muthén & Muthén.

Bagozzi, R. P. (2007). On the Meaning of Formative Measurement and How It Differs From

Reflective Measurement: Comment of Howell, Breivik, and Wilcox (2007).

Psychological Methods, 12(2), 229-237.

Bagozzi, R. P. (2011). Measurement and Meaning in Information Systems and Organizational

Research: Methodological and Philosophical Foundations. MIS Quarterly, 35(2), 261-

292.

Baker, T., & Nelson, R. E. (2005). Creating something from nothing: Resource construction

through entrepreneurial bricolage. Administrative science quarterly, 50(3), 329-366.

Bandura, A. (1977). Self-Efficacy: Toward a Unifying Theory of Behavior Change.

Psychological Review, 84(2), 191-215.

Bandura, A. (1977). Self-efficacy: toward a unifying theory of behavioral change. Psychological

Review, 84, 191–215.

Bandura, A. (1982). Self-efficacy mechanism in human agency. American Psychologist, 37(2),

122-147.

Bandura, A. (1989). Human agency in social-cognitive theory. American Psychologist, 44, 1175–

1184.

Bandura, A. (2001). Social cognitive theory: an agentic perspective. Annual Review of

Psychology, 52, 1-26.

Baum, J. & Locke, E. (2004). The relationship of entrepreneurial traits, skill, and motivation to

subsequent venture growth. Journal of Applied Psychology, 89(4), 587–598.

Beardsley, M. C. (1965). On the creation of art. The Journal of Aesthetics and Art Criticism,

23(3), 291-304.

Berger, P. L., & Luckmann, T. (1967). The Social Construction of Reality: A Treatise in the

Sociology of Knowledmann. Anchor books.

Betz, N. & Hackett, G. (1981). The relationship of career-related self-efficacy expectations to

perceived career options in college men and women. Journal of Counseling Psychology,

28, 399–410.

Bird, B. (1988). Implementing entrepreneurial ideas: the case for intention. Academy of

Management Review, 13(3), 442–453.

Blau, G. 1985. The measurement and predication of career commitment. Journal of Occupational

Psychology, 58, 277-288.

Bollen, K. A. (1989). Structural Equations with Latent Variables. New York: Wiley.

150

Bollen, K. A. (2011). Evaluating Effect, Composite, and Causal Indicators in Structural Equation

Models. MIS Quarterly, 35(2), 359-372.

Bollen, K., & Lennox, R. (1991). Conventional Wisdom on Measurement. Psychological

Bulletin, 110(2), 305-314.

Bolstad, W. M. (2007). Introduction to Bayesian Statistics (Second ed.). Hoboken, NJ: John

Wiley & Sons.

Borsboom, D., Mellenbergh, G. J., & van Heerden, J. (2003). The Theoretical Statis of Latent

Variables. Psychological Review, 110(2), 203-219.

Boyd, N. & Vozikis, G. (1994). The influence of self-efficacy on the development of

entrepreneurial intentions and actions. Entrepreneurship Theory and Practice, 18(4), 63–

77.

Boyd, N. G., & Vozikis, G. S. (1994). The Influence of Self-Efficacy on the Development of

Entrepreneurial Intentions and Actions. Entrepreneurship Theory & Practice, 18(4), 63-

77.

Browne, M. W., & Cudeck, R. (1992). Alternative Ways of Assessing Model Fit. Sociological

Methods & Research, 21(2), 230-258.

Browne, M. W., & Mels, G. (1998). Path Analysis: Ramona. In: SYSTAT for Windows:

Advanced Applications (Version 8). Evanston, IL: SYSTAT.

Burt, R. S. (1976). Interpretational Confounding in Unobserved Variables in Structural Equation

Models. Sociological Methods and Research, 5(1), 3-52.

Byrne, M. & Keefe, M. (2002). Building research competence in nursing through mentoring.

Journal of Nursing Scholarship, 4th Quarter, 391-396.

Campbell, D. T. (1960). Blind variation and selective retentions in creative thought as in other

knowledge processes. Psychological review, 67(6), 380.

Campbell, D. T. (1974). Evolutionary Epistemology. In P. A. Schilpp, The Philosophy of Karl

Popper, Vol 14 (pp. 413-463). La Salle: Open Court.

Cannon‐Bowers, J. A., & Salas, E. (2001). Reflections on shared cognition. Journal of

Organizational Behavior, 22(2), 195-202.

Chen, C. C., Greene, P. G., & Crick, A. (1998). Does Entrepreneurial Self-Efficacy Distinguish

Entrepreneurs from Managers? Journal of Business Venturing, 13(4), 295-316.

Chen, G., Gully, S.M., & Eden, D. (2004). General self-efficacy and self-esteem: toward

theoretical and empirical distinction between correlated self-evaluations. Journal of

Organizational Behavior, 25, 375–395.

Chen, G.C., Greene, P.G., & Crick, A. (1998). Does entrepreneurial self-efficacy distinguish

entrepreneurs from managers? Journal of Business Venturing, 13, 295–317.

Cheung, G.W., & Rensvold, R.B. (2002). Evaluating goodness-of-fit indexes for testing

measurement invariance. Structural Equation Modeling, 9(2), 233-255.

151

Clark, A. (1997). Being There. Cambridge: MIT Press.

Clark, A. (2016). Surfing uncertainty. Oxford: Oxford University Press.

Coddington, A. (1982). Deficient foresight: a troublesome theme in Keynesian economics. The

American Economic Review, 72(3), 480-487.

Coffman, D.L., & MacCallum, R.C. (2005). Using parcels to convert path analysis models into

latent variable models. Multivariate Behavioral Research, 40(2), 235-259.

Courville, A. C., Daw, N. D., & Touretzky, D. S. (2006). Bayesian theories of conditioning in a

changing world. Trends in cognitive sciences, 10(7), 294-300.

Cudeck, R. (1989). Analysis of Correlation Matrices Using Covariance Structure Models.

Psychological Bulletin, 105(2), 317-327.

Cudeck, R., & Browne, M.W. (1983). Cross-validation of covariance structures. Multivariate

Behavioral Research, 18, 147-167.

Cudeck, R., & Henly. (1991). Model Selection in Covariance Strucutres Analysis and the

"Problem" of Sample Size: A Clarification. Psychological Bulletin, 109(3), 512-519.

Dacin, M. T. (1997). Isomorphism in context: The power and prescription of institutional norms.

Academy of Management journal, 40(1), 46-81.

De Dreu, C. K., Nijstad, B. A., & van Knippenberg, D. (2008). Motivated information processing

in group judgment and decision making. Personality and Social Psychology Review,

12(1), 22-49.

De Noble, A. F., Jung, D., & Ehrlich, S. B. (1999). Entrepreneurial self-efficacy: The

development of a measure and its relationship to entrepreneurial action. Frontiers of

entrepreneurship research, 1999, 73-87.

Deakins, D., Graham, L., Sullivan, R., & Whittam, G. (1998). New venture support: an analysis

of mentoring support for new and early stage ventures. Journal of Small Business and

Enterprise Development, 5(2), 151-161.

Dequech, D. (2000). Fundamental uncertainty and ambiguity. Eastern Economic Journal, 26(1),

41-60.

Dequech, D. (2006). The new institutional economics and the theory of behaviour under

uncertainty. Journal of Economic Behavior & Organization, 59(1), 109-131.

Dewey, J., & Bentley, A. (1949). Knowing and the Known. Boston: Beacon Press.

Dimov, D. (2007). Beyond the single‐person, single‐insight attribution in understanding

entrepreneurial opportunities. Entrepreneurship Theory and Practice, 31(5), 713-731.

Dimov, D. (2010). Nascent entrepreneurs and venture emergence: Opportunity confidence,

human capital, and early planning. Journal of Management Studies, 47(6), 1123-1153.

Dimov, D. (2011). Grappling with the unbearable elusiveness of entrepreneurial opportunities.

Entrepreneurship Theory and Practice, 35(1), 57-81.

Dopfer, K., & Potts, J. (2004). Evolutionary foundations of economics. Evolution and economic

complexity, 3-23.

152

Douglas, E., & Shepherd, D. (2002). Self-employment as a career choice: attitudes,

entrepreneurial intentions, and utility maximization. Entrepreneurial Theory and

Practice, 26(3), 81-90.

Dunson, D. B., Palomo, J., & Bollen, K. (2005). Bayesian Structural Equation Modeling

(Technical Report). Research Triangle Park, NC: Statistical and Applied Mathematical

Sciences Institute.

Eby, L.T., Allen, T.D., Evans, S.C., Ng, T., & DuBois, David. (2008). Does Mentoring Matter? A

Multidisciplinary Meta-Analysis Comparing Mentored and Non-Mentored Individuals.

Journal of Vocational Behavior. 72(2), 254-267.

Edwards, J. R. (2011). The Fallacy of Formative Measurement. Organizational Research

Methods, 14(2), 370-388.

Edwards, J. R., & Bagozzi, R. P. (2000). On the Nature and Direction of Relationships Between

Constructs and Measures. Psychological Methods, 5(2), 155-174.

Edwards, M. C. (2010). A Markov Chain Monte Carlo Approach to Confirmatory Item Factor

Analysis. Psychometrika, 75(3), 474-497.

Eisenhardt, K. M. (1989). Building theories from case study research. Academy of management

review, 14(4), 532-550.

Eisenhardt, K. M., & Graebner, M. E. (2007). Theory building from cases: Opportunities and

challenges. Academy of management journal, 50(1), 25.

Forbes, D.P. (2005). The effects of strategic decision making on entrepreneurial self-efficacy.

Entrepreneurship Theory and Practice, 29(5), 599-626.

Freud, S. (1908). Creative writers and day-dreaming. Standard edition, 9(1).

Funke, J. (2001). Dynamic systems as tools for analysing human judgement. Thinking &

Reasoning, 7(1), 69-89.

Gardner, D.G., Cummings, L.L., Dunham, R.B., & Pierce, J.L. (1998). Single-item versus

multiple-item scales: an empirical comparison. Educational and Psychological

Measurement, 58(6), 898-915.

Garud, R., & Karnøe, P. (2001). Path creation as a process of mindful deviation. Path dependence

and creation, 138.

Gelman, A. (2003). A Bayesian Formulation of Exploratory Data Analysis and Goodness-of-Fit

Testing. International Statistical Review, 71(2), 369-382.

Geman, S., & Geman, D. (1984). Stochastic Relaxation, Gibbs Distributions, and the Bayesian

Restoration of Images. IEEE Transactions on Pattern Analysis and Machine Intelligence,

6, 721-741.

Gephart, R. P. (2004). Qualitative research and the Academy of Management Journal. Academy

of Management Journal, 47(4), 454-462.

Gerbing, D. W., & Anderson, J. C. (1984). On the Meaning of Within-Factor Correlated

Measurement Errors. Journal of Consumer Research, 11(1), 572-580.

153

Gerbing, D. W., & Anderson, J. C. (1988). An Updated Paradigm for Scale Development

Incorporating Unidimensionality and Its Assessment. Journal of Marketing Research,

25(2), 186-192.

Gershman, S. J., Blei, D. M., & Niv, Y. (2010). Context, learning, and extinction. Psychological

review, 117(1), 197.

Gigerenzer, G. (2000). Adaptive thinking: Rationality in the real world. Oxford University Press,

USA.

Gigerenzer, G., & Todd, P. M. (1999). Simple heuristics that make us smart. Oxford University

Press, USA.

Gist, M. (1987). Self-efficacy: Implications for organizational behavior and human resource

management. Academy of Management Journal, 12, 472–485.

Gist, M. E. (1987). Self-Efficacy: Implications for Organizational Behavior and Human Resource

Management. Academy of Management Review, 12(3), 472-485.

Gist, M. E., & Mitchell, T. R. (1992). Self-efficacy: a theoretical analysis of its determinants and

malleability. Academy of Management Review, 17, 183–211.

Grewal, R., Cote, J. A., & Baumgartner, H. (2004). Multicollinearity and Measurement Error in

Structural Equation Models: Implications for Theory Testing. Marketing Science, 23(4),

519-529.

Grier, K. C. (2006). Pets in America: A History. Chapel Hill: The University of North Carolina

Press.

Griffeth, R.W., Hom, P.W., & Gaertner, S. (2000). A meta-analysis of antecedents and correlates

of employee turnover: update, moderator tests, and research implications for the next

millennium. Journal of Management, 26, 463-488.

Griffiths, T. L., & Tenenbaum, J. B. (2009). Theory-based causal induction. Psychological

review, 116(4), 661.

Gruber, M. (2007). Uncovering the value of planning in new venture creation: A process and

contingency perspective. Journal of Business Venturing, 22(6), 782-807.

Hair, J. F., Black, W. C., Babin, B. J., & Anderson, R. E. (2010). Multivariate Data Analysis (7th

ed.). New York: Prentice Hall.

Heide, J. B., & John, G. (1992). Do Norms Matter in Marketing Relationships. Journal of

Marketing, 56(2), 32-44.

Hmieleski, K. M., & Baron, R. A. (2008). When does entrepreneurial self‐efficacy enhance

versus reduce firm performance?. Strategic Entrepreneurship Journal, 2(1), 57-72.

Hmieleski, K.M. & Baron, R.A. (2008). When does entrepreneurial self-efficacy enhance versus

reduce firm performance? Strategic Entrepreneurship Journal, 2(1), 57-72.

Holcomb, T. R., Ireland, R. D., Holmes Jr, R. M., & Hitt, M. A. (2009). Architecture of

entrepreneurial learning: Exploring the link among heuristics, knowledge, and action.

Entrepreneurship Theory and Practice, 33(1), 167-192.

154

Holzinger, K. J., & Swineford, F. (1937). The Bi-Factor Method. Psychometrika, 2(1), 41-54.

Howell, R. D., Breivik, E., & Wilcox, J. B. (2007). Reconsidering Formative MEasurement.

Psychological Methods, 12(2), 205-218.

Hu, L.-t., & Bentler, P. M. (1999). Cutoff Criteria for Fit Indexes in Covariance Structure

Analysis: Conventional Criteria Versus New Alternatives. Structural Equation Modeling,

6(1), 1-55.

Hutchins, E. (1995). Cognition in the Wild. MIT press.

Jacobs, R. A., & Kruschke, J. K. (2011). Bayesian learning theory applied to human cognition.

Wiley Interdisciplinary Reviews: Cognitive Science, 2(1), 8-21.

James, W. (1975). Pragmatism (Vol. 1). Harvard University Press.

Jones, M., & Love, B. C. (2011). Bayesian fundamentalism or enlightenment? On the explanatory

status and theoretical contributions of Bayesian models of cognition. Behavioral and

Brain Sciences, 34(04), 169-188.

Kaplan, D., & Depaoli, S. (2012). Bayesian Structural Equation Modeling. In R. H. Hoyle (Ed.),

Handbook of Structural Equation Modeling. New York: Guilford Press.

Kaplan, S. (2008). Framing contests: Strategy making under uncertainty. Organization Science,

19(5), 729-752.

Katz, J., & Gartner, W.B. (1988). Properties of emerging organizations. The Academy of

Management Journal, 13(3), 429-441.

Kemp, C., & Tenenbaum, J. B. (2008). The discovery of structural form. Proceedings of the

National Academy of Sciences, 105(31), 10687-10692.

Kemp, C., Perfors, A., & Tenenbaum, J. B. (2007). Learning overhypotheses with hierarchical

Bayesian models. Developmental science, 10(3), 307-321.

Knight, F. H. (1921). Risk, uncertainty and profit. New York: Hart, Schaffner and Marx.

Kram, K.E. & Ragins, B.R. (2007). The landscape of mentoring in the 21st century. In Ragine,

B.R. & Kram, K.E. (eds.), The handbook of mentoring at work, pps. 659-692, Thousand

Oaks, CA: Sage Publications.

Kram, K.E. (1983). Phases of the mentor relationship. Academy of Management Journal, 26, 608-

625.

Krueger, N.F., Jr., & Brazeal, D.V. (1994). Entrepreneurial potential and potential entrepreneurs.

Entrepreneurship Theory & Practice, 18(3), 91–104.

Krueger, N.F., Jr., Reilly, M.D., & Carsrud, A.L. (2000). Competing models of entrepreneurial

intentions. Journal of Business Venturing, 15, 411–432.

Kuhn, T. S. (1977). The Essential Tension: Selected Studies in Scientific Tradition and Change.

Chicago, IL: University of Chicago Press.

Langley, P., & Simon, H. A. (1981). The central role of learning in cognition. Cognitive skills and

their acquisition, 361-380.

155

Lankau, M. & Scandura, T.A. (2002). An investigation of personal learning in mentoring

relationships: content, antecedents, and consequences. Academy of Management Journal,

45, 779-790.

Leavitt, K., Mitchell, T. R., & Peterson, J. (2010). Theory Pruning: Strategies to Reduce Our

Dense Theoretical Landscape. Organizational Research Methods, 13(4), 644-667.

Lee, B. P. (2001). Mutual knowledge, background knowledge and shared beliefs: Their roles in

establishing common ground. Journal of pragmatics, 33(1), 21-44.

Lee, S.-Y. (2007). Structural Equation Modeling: A Bayesian Approach. West Sussex, UK:

Wiley.

Levy, R. (2011). Bayesian Data-Model Fit Assessment for Structural Equation Modeling.

Structural Equation Modeling, 18(4), 663-685.

Lichtenstein, S., & Slovic, P. (Eds.). (2006). The construction of preference. Cambridge

University Press.

Locke, K. (2001). Grounded theory in management research. Sage.

Lounsbury, M., & Crumley, E. T. (2007). New practice creation: An institutional perspective on

innovation. Organization studies, 28(7), 993-1012.

MacCallum, R. C. (2003). Working with Imperfect Models. Multivariate Behavioral Research,

38(1), 113-139.

MacCallum, R. C., & Tucker, L. R. (1991). Representing Sources of Error in the Common-Factor

Model: Implications for Theory and Practice. Psychological Bulletin, 109(3), 502-511.

MacCallum, R. C., Edwards, M. C., & Cai, L. (2012). Hopes and Cautions in Implementing

Bayesian Structural Equation Modeling. Psychological Methods, 17(3), 340-345.

MacCallum, R. C., Widaman, K. F., Preacher, K. J., & Hong, S. (2001). Sample Size in Factor

Anlaysis: The Role of Model Error. Multivariate Behavioral Research, 36(4), 611-637.

MacCallum, R. C., Widaman, K. F., Zhang, S., & Hong, S. (1999). Sample Size in Factor

Analysis. Psychological Methods, 4(1), 84-99.

March, J. G. (1991). Exploration and exploitation in organizational learning. Organization

science, 2(1), 71-87.

Markus, H. (1977). Self-schemata and processing information about the self. Journal of

personality and social psychology, 35(2), 63.

Marsh, H. W., Muthén, B., Asparouhov, T., Lüdtke, O., Robitzsch, A., Morin, A. J., &

Trautwein, U. (2009). Exploratory Structural Equation Modeling, Integrating CFA and

EFA: Application to Students' Evaluations of University Teaching. Structural Equation

Modeling, 16(3), 439-476.

Marsh, H. W., Muthén, B., Morin, A. J., Lüdtke, O., Asparouhov, T., Trautwein, U., &

Nagengast, B. (2010). A New Look at the Big Five Factor Structure Through Exploratory

Structural Equation Modeling. Psychological Assessment, 22(3), 471-491.

156

Maurer, T. J., (2001). Career-relevant learning and development, worker age, and beliefs about

self-efficacy for development. Journal of Management, 27, 123-140.

McGee, J. E., Peterson, M., Mueller, S. L., & Sequeira, J. M. (2009). Entrepreneurial Self-

Efficacy: Refining the Measure. Entrepreneurship Theory & Practice, 33(4), 965-988.

McGee, J. E., Peterson, M., Mueller, S. L., & Sequeira, J.M., (2009). Entrepreneurial self-

efficacy: refining the measure. Entrepreneurship Theory & Practice, 33, 965-988.

McMullen, J. S., & Dimov, D. (2013). Time and the entrepreneurial journey: the problems and

promise of studying entrepreneurship as a process. Journal of Management Studies,

50(8), 1481-1512.

Mehlhorn, K., Newell, B. R., Todd, P. M., Lee, M. D., Morgan, K., Braithwaite, V. A., &

Gonzalez, C. (2015). Unpacking the exploration–exploitation tradeoff: A synthesis of

human and animal literatures.

Mintzberg, H., & Waters, J. A. (1985). Of strategies, deliberate and emergent. Strategic

management journal, 6(3), 257-272.

Mueller, S.L. & Goic, S. (2003). East-west differences in entrepreneurial self-efficacy:

implications for entrepreneurship education in transition economies. International

Journal of Entrepreneurship Education, 1, 613–632.

Muthén, B. O. (2002). Beyond SEM: General Latent Variable Modeling. Behaviormetrika, 29(1),

81-117.

Muthén, B., & Asparouhov, T. (2012a). Bayesian Structural Equation Modeling: A More Flexible

Representation of Substantive Theory. Psychological Methods, 17(3), 313-335.

Muthén, B., & Asparouhov, T. (2012b). Rejoinder to MacCallum, Edwards, and Cai (2012) and

Rindskopf (2012): Mastering a New Method. Psychological Methods, 17(3), 346-353.

Muthén, B., & Asparouhov, T. (2013). BSEM Measurement Invariance Analysis (Mplus Web

Notes: No. 17). Los Angeles, CA: Muthén & Muthén.

Muthén, L. K., & Muthén, B. O. (2012). Mplus User's Guide (Seventh ed.). Los Angeles, CA:

Muthén & Muthén.

Myers, N. D., Ahn, S., & Jin, Y. (2013). Rotation to a Partially Specified Target Matrix in

Exploratory Factor Analysis: How Many Targets? Structural Equation Modeling, 20(1),

131-147.

Newell, A., & Simon, H. A. (1956). The logic theory machine--A complex information

processing system. Information Theory, IRE Transactions on, 2(3), 61-79.

Noe, R.A., Greenberger, D.B., & Wang, S. (2002). Mentoring: what we know and where we

might go. In (Ed.), Research in Personnel and Human Resources Management, Volume

21 (pp. 129-173). Emerald Group Publishing Limited.

O'Reilly, C.A., & Chatman, J.A., (1996). Culture as social control: corporations, cults, and

commitment. In B. Staw and L. Cummings (Eds.), Research in organizational behavior,

Volume 18 (pp. 157-200). Greenwich, CT.: JAI Press.

157

Organ, D. W. (1988). Organizational Citizenship Behavior: The Good Soldier Syndrome.

Lexington, MA: Lexington Books.

Osman, M. (2010). Controlling uncertainty: a review of human behavior in complex dynamic

environments. Psychological bulletin, 136(1), 65.

Payne, J. W., Bettman, J. R., & Johnson, E. J. (1993). The adaptive decision maker. Cambridge

University Press.

Payne, S.C., & Huffman, A.H. (2005). A longitudinal examination of the influence of mentoring

on organizational commitment. The Academy of Management Journal, 48, 158-168.

Payzan-LeNestour, E., & Bossaerts, P. (2011). Risk, unexpected uncertainty, and estimation

uncertainty: Bayesian learning in unstable settings. PLoS Comput Biol, 7(1), e1001048.

Peirce, C. S. (1905). What Pragmatism Is. The Monist, 161-181.

Perfors, A., Tenenbaum, J. B., Griffiths, T. L., & Xu, F. (2011). A tutorial introduction to

Bayesian models of cognitive development. Cognition, 120(3), 302-321.

Poole, M. S., & Van de Ven, A. H. (1989). Using Paradox to Build Management and

Organization Theories. Academy of Management Review, 14(4), 562-578.

Porac, J. F., Thomas, H., & Baden‐Fuller, C. (1989). Competitive groups as cognitive

communities: The case of Scottish knitwear manufacturers*. Journal of Management

studies, 26(4), 397-416.

Preacher, K. J., & Hayes, A. F. (2008). Asymptotic and resampling strategies for assessing and

comparing indirect effects in multiple mediator models. Behavior Research Methods, 40,

879-891.

Ragins B.R., Cotton, J.L., & Miller, J.S. (2000). Marginal mentoring: the effects of type of

mentor, quality of relationship, and program design on work and career attitudes.

Academy of Management Journal, 43, 1177-1194.

Reise, S. P. (2012). The Rediscovery of Bifactor Measurement Models. Multivariate Behavioral

Research, 47(5), 667-696.

Richardson, H.A., Simmering, M.J., & Sturman, M.C. (2009). A Tale of Three Perspectives:

Examining Post Hoc Statistical Techniques for Detection and Correction of Common

Method Variance. Organizational Research Methods. 12, 762-800.

Rindskopf, D. (2012). Next Steps in Bayesian Structural Equation Models: Comments on,

Variations of, and Extensions to Muthén and Asparouhov (2012). Psychological

Methods, 17(3), 336-339.

Roberts, S., & Pashler, H. (2000). How Persuasive Is a Good Fit? A Comment on Theory Testing.

Psychological Review, 107(2), 358-367.

Rubin, D. B. (1996). Multiple Imputation After 18+ Years. Journal of the American Statistical

Association, 91(434), 473-489.

158

Rupp, A. A., Dey, D. K., & Zumbo, B. D. (2004). To Bayes or Not to Bayes, From Whether to

When: Applications of Bayesian Methodology to Modeling. Structural Equation

Modeling, 11(3), 424-451.

Russell, J.E.A., & Adams, D.M. (1997). The changing nature of mentoring in organizations: an

introduction to the special issue on mentoring in organizations. Journal of Vocational

Behavior, 51, 1-14.

Sarasvathy, S. D. (2001). Causation and effectuation: Toward a theoretical shift from economic

inevitability to entrepreneurial contingency. Academy of management Review, 26(2), 243-

263.

Scheines, R., Hoijtink, H., & Boomsma, A. (1999). Bayesian Estimation and Testing of Structural

Equation Models. Psychometrika, 64(1), 37-52.

Scherer, R., Adams, J., Carley, S., & Wiebe, F. (1989). Role model performance effects on

development of entrepreneurial career preference. Entrepreneurship Theory & Practice,

13, 53–71.

Scherer, R.F., Brodinski, J.D., & Wiebe, F. (1991). Examining the relationship between

personality and entrepreneurial career preference. Entrepreneurship & Regional

Development, 3, 195-206.

Searle, J. R. (1995). The construction of social reality. Simon and Schuster.

Segal, G., Borgia, D., & Schoenfeld, J. (2005). The motivation to become an entrepreneur.

International Journal of Entrepreneurial Behaviour & Research, 11, 42-57.

Shane, S. (2000). Prior knowledge and the discovery of entrepreneurial opportunities.

Organization science, 11(4), 448-469.

Shane, S., & Venkataraman, S. (2000). The promise of entrepreneurship as a field of research.

Academy of management review, 25(1), 217-226.

Shanks, D. R. (2010). Learning: From association to cognition. Annual review of psychology, 61,

273-301.

Siggelkow, N. (2007). Persuasion with case studies. Academy of management journal, 50(1), 20-

24.

Simon, H. A. (1982). Models of bounded rationality: Empirically grounded economic reason

(Vol. 3). MIT press.

Stake, R.E. 2005. Qualitative Case Studies pp. 443-466. In Sage Handbook of Qualitative

Research, 3rd Edition. Denzin, N.K. & Lincoln, Y.S. (eds.) Sage Publications: Thousand

Oaks, CA.

Stevenson, H. H., Roberts, M. J., & Grousbeck, H. I. (1985). New Business Ventures and the

Entrepreneur. Burr Ridge, IL: Richard D Irwin.

Stevenson, H.H., Roberts, M.J., & Grousbeck,H.I. (1985). New Business Ventures and the

Entrepreneur. Burr Ridge, IL: Richard D. Irwin.

Steyvers, M., Tenenbaum, J. B., Wagenmakers, E. J., & Blum, B. (2003). Inferring causal

networks from observations and interventions. Cognitive science, 27(3), 453-489.

159

St-Jean, E., & Audet, J. (2012). The role of mentoring in the learning development of the novice

entrepreneur. International Entrepreneurial Management Journal, 8, 119-140.

Suddaby, R. (2006). From the editors: What grounded theory is not. Academy of management

journal, 49(4), 633-642.

Sullivan, R. (2000). Entrepreneurial learning and mentoring. International Journal of

Entrepreneurial Behavior & Research, 6, 160-175.

Taylor, A., & Greve, H. R. (2006). Superman or the fantastic four? Knowledge combination and

experience in innovative teams. Academy of Management Journal, 49(4), 723-740.

Tenenbaum, J. B., Griffiths, T. L., & Kemp, C. (2006). Theory-based Bayesian models of

inductive learning and reasoning. Trends in cognitive sciences, 10(7), 309-318.

Tenenbaum, Joshua B., Charles Kemp, Thomas L. Griffiths, and Noah D. Goodman. "How to

grow a mind: Statistics, structure, and abstraction." science 331, no. 6022 (2011): 1279-

1285.

Thurstone, L. L. (1947). Multiple Factor Analysis. Chicago: University of Chicago Press.

Tripsas, M., & Gavetti, G. (2000). Capabilities, cognition, and inertia: Evidence from digital

imaging. Strategic management journal, 21(10-11), 1147-1161.

Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases.

science, 185(4157), 1124-1131.

Viator, R.E. & Scandura, T.A. (1991). A study of mentor protégé relationships in large public

accounting firms. Accounting Horizons, 5, 20-30.

Walsh, I. J., & Bartunek, J. M. (2011). Cheating the fates: Organizational foundings in the wake

of demise. Academy of Management Journal, 54(5), 1017-1044.

Wanous, J.P., & Hudy, M.J. (2001). Single item reliability: a replication and extension.

Organizational Research Methods, 4, 361-375.

Wanous, J.P., Reichers, A.E., & Hudy, M.J. (1997). Overall job satisfaction: how good are single

item measures? Journal of Applied Psychology, 82, 247-252.

Waters, L., McCabe, M., Kiellerup, D., & Kiellerup, S. (2002). The role of formal mentoring on

business success and self-esteem in participants of a new business start-up program.

Journal of Business and Psychology, 17, 107-121.

Wilson, F., Kickul, J., & Marlino, D. (2007). Gender, entrepreneurial self-efficacy, and

entrepreneurial career intentions: Implications for entrepreneurship education.

Entrepreneurship Theory & Practice, 31, 387– 406.

Wiltbank, R., Dew, N., Read, S., & Sarasvathy, S. D. (2006). What to do next? The case for non‐

predictive strategy. Strategic management journal, 27(10), 981-998.

Wood, M. S., & McKinley, W. (2010). The production of entrepreneurial opportunity: a

constructivist perspective. Strategic Entrepreneurship Journal, 4(1), 66-84.

Wood, R., & Bandura, A. (1989). Social Cognitive Theory of Organizational Management.

Academy of Management Review, 14(3), 361-384.

160

Wood, R., & Bandura, A. (1989). Social cognitive theory of organizational management.

Academy of Management Review, 14, 361–381.

Woolley, J. L. (2014). The creation and configuration of infrastructure for entrepreneurship in

emerging domains of activity. Entrepreneurship theory and practice, 38(4), 721-747.

Wynn Jr, D., & Williams, C. K. (2012). Principles for conducting critical realist case study

research in information systems. Mis Quarterly, 36(3), 787-810.

Yates, A. (1987). Multivariate Exploratory Data Analysis: A Perspective on Exploratory Factor

Analysis. Albany, NY: State University of New York Press.

Yin, R. K. (2009). Case study research: Design and methods, 4th. Thousand Oaks.

Yuan, Y., & MacKinnon, D. P. (2009). Bayesian Mediation Analysis. Psychological Methods,

14(4), 301-322.

Zhao, H., Seibert, C., & Hills, C. (2005). The mediating role of self-efficacy in the development

of entrepreneurial intentions. Journal of Applied Psychology, 90, 1265–127

161

Appendix A: Prior History in the Pet Health Insurance Market

Study Landscape and Prior History

While this study is concerned with the processes of resource co-creation, as examined in

the emergence of a particular firm within the pet health insurance industry in the 2000s, it is

important to understand some broader trends in social, demographic, technical, and regulatory

change that influenced the processes investigated. These changes have direct implications for the

social milieu that guides, constrains, and inspires the involved entrepreneurs and those who they

interacted with in resourcing their firms. These shifts can be grouped into families of related

changes, with two primary social-level changes including the changing role of the pet as a

member of the family and huge advancements in both the way that veterinary medicine is

practiced and its perceived value. While Americans have always had an abiding fascination with

pets (Grier, 2006), the rate of change in both of these areas picked up speed throughout the

eighties and nineties. In 1982, pet health insurance was introduced into the United States, but

grew anemically through this same period. While pet insurance stayed under the radar for most

Americans, this time period had important implications for the nature of the context within which

a new wave of firms subsequently co-created resources during the 2000s. In order to articulate

these issues the following sections provide some needed background information that explains the

environment that the firm under study confronted.

162

The Changing Role of Pet as Property and Pet as Family Member

America has long been a nation of pet-owners, even from the earliest days of the

European settlers (Grier, 2006: p.2). Spanish conquistadors and other European groups brought

dogs with them as war beasts, guards, workers, and companions; along with a host of other small

animals, including domestic cats and birds. While there is a long, interwoven history of

Americans and their love for dogs and more recently cats, changes in the nature of these

relationships have accelerated in the last several decades. Huge breakthroughs in crop sciences

and infrastructure in the last century allowed America to conquer the persistent threat of

widespread hunger and food availability volatility. Along with the invention of two of the more

taken for granted conveniences of modern pet care, clay-based cat litter and viable multi-life

stage flea control (Grier, 2006: p. 87). These changes dramatically increased the rate at which

households welcomed domesticated animals into their homes.

While exact statistics for the number of dogs, cats, and families with pets are not

available, it should be observed that the population of both has been on a steady increase over the

last three decades. The US dog population between 1979 and 2009 increased from an estimated

49 million to 77.5 million28, with 58.3% of families reporting at least one pet as of 2001 and 62%

by 2003.29 On average the cat population has mirrored the growth in the dog population, although

it probably contains roughly ten million more animals.

This same time period also saw a dramatic increase in the actual amount that pet owners

reported spending on their pets, their willingness to spend on their pets, and the offerings

available from the pet industry. The pet supply industry has grown steadily over the last decades,

with total U.S. pet industry expenditures growing from $17 billion in 1994 to $53.3 billion in

28 Give cite for this information 29 AVMA Survey 2011

163

201230. Petsmart, the nation’s largest pet supply chain, was started in 1987 with two stores and

subsequently IPO’ed in 1994 with 107 stores. By 2002 the chain had 600 stores and net sales

totaling $2.7 billion (Grier, 2006, p. 270), with sales increasing to $6.8 billion in 2012.

The social manner by which pets are identified within the family has also changed, and

continues changing. Pets have always been considered property in the United States, as reflected

in the law and the very nature of the term ‘pet-owner’. However it is becoming more common for

pet owners to report that they consider their pet to be a member of the family or a best friend31

Recently some pet owners have begun rejecting the term ‘pet owner’ in favor of the more familial

‘pet parent’. For a period of time the Humane Society of America attempted to only use the term

‘companion animal’ in the believe that ‘pet’ denoted a dominant form of ownership and implied a

hierarchical relationship (Grier, 2006, p. 7).

While there is no one clear reason for this change in how Americans allocate their income

and time, several factors have been identified as potential explanations. Amongst these are

changes in the demographic composition of the population with more families consisting of a

smaller number of children, a larger elderly population, and the tendency for families to fragment

into separate units. It is speculated that animal companionship serves an important social role for

the estimated 25% of US individuals who live on their own. Likewise changes in discretionary

income, the availability of ready-made pet foods and pet care products, the quality of veterinary

care and human medical care (extending live spans in both humans and animals) have contributed

to the ease with which individuals may now possess pets. There is also a general change in how

society views its role as a steward of the planet and by extension the animals that we bring under

our domain. While change is never uniform across a society, the overall trend of pet owner’s

willingness to spend (and dote) on their pets has not been restricted only to the affluent.

30 http://www.americanpetproducts.org/press_industrytrends.asp 31 APPMA Survey, 2004

164

Changes in the Veterinary Profession

Veterinary medicine has changed dramatically in the last several decades as well. The

field has become more professionalized, more advanced, and more in-line with the current state

of human medicine. These rapid changes in veterinary medicine have been partly driven by an

increased adoption of techniques from human medicine, with their associated complexities and

high costs, and partly by the previously cited willingness of pet owners to incur the costs for these

more advanced practices. A 1980 report from the American Kennel Club cited that veterinary

bills had on average doubled in the prior five years32, this pace has remained relatively consistent

over the last three decades. The cost of veterinary medicine has increased annually at roughly

5.6% since 200033.

The 1980s and 1990s saw two large shifts in veterinary practice, one of these being the

emergence of specialists, and the other the appearance of veterinary care facilities that possessed

high-end technology akin to human hospitals. As heard from one interviewee this time period was

the end of the “James Herriot era.”34 As demand for quality veterinary medicine increased and

more techniques were transferred into veterinary medicine from human medicine (although in

many cases such techniques might have originally been developed with animals) it became

possible for veterinarians to specialize in narrower fields of medicine. These shifts in demand for

advanced veterinary care made it possible for specialists to cover the additional cost of the

associated training. Likewise these specialists needed access to equipment that was too expensive

for any one small practice to afford (i.e. Ct scan, etc.). While initially located around veterinary

32 Newsweek, September 15, 1980. “Man Insures Dog: How Pets Get Vets.” 33 Bureau of Labor Statistics, US Department of Labor, Consumer Price Index 2010 edition 34 James Herriot (pen name) was an English veterinarian who wrote a well-loved series of books about his

experience as a small-town country vet. His stories centered on his adventures as a vet in which any day he

might be called on to tend the upset stomach of a small, spoiled lap dog or spend his evening assisting a

farmer with a cow’s breach birth.

165

teaching and research schools, high-end animal medical centers sprung up around the country.35

These centers provided a location for specialists to receive referral traffic from local veterinary

practices, provide emergency case management, and facilitate procedures that were unavailable at

the average veterinarian’s office.36

While these specialists had, on average, received the same level of training (and likewise

incurred similar debt) as their human-medicine counterparts the procedures they completed were

billed at a fraction of what equivalent care at a doctor’s office would have cost.37 This trend of

improving quality of care, cost of service, and minimal remuneration continues to be a primary

concern amongst the veterinary profession. As one veterinarian interviewed for this study

commented: “We often ethically feel torn between providing the best care that our training has

prepared us for, and providing the care that the pet owner can actually afford.” This echoes a

sentiment that has been iterated throughout the last several decades: “… the possibilities afforded

by high-tech pet care and its costs create difficult ethical questions.”38

This ethics of this line of inquiry extended to the broader societal concern of where resources

should be allocated. As these seismic shifts were ongoing in the practice of veterinary medicine,

more than once the concern that it is inappropriate to spend thousands treating a sick dog when

there are people in desperate need of medicine echoed through the field39. Yet growth in demand

for such services remained unabated and the veterinary field responded by graduating more

specialists and increasing the overall level of training for general veterinarians.40 By 1998,

amongst the 60,000 vets in the USA there were 5,600 specialists in such diverse areas as

35 St Petersburg Times, September 11, 1988. “Extraordinary Pet Care on Rise” 36 The Toronto Star, October 23, 1990. “Say Woof” 37 The Globe and Mail, January 17, 1985. “Clawed by Pet Owners” 38 Newsweek, May 20, 1991. “In No Time, Back on All Four Feet” 39 The Washington Post, July 2, 1991. “High-tech Medicine for Pets: How Much Are Owners Willing to

Spend?” 40 Kiplinger’s Personal Finance Magazine, July 1997. “Money (Ouch!) Can Cure Fido”

166

endocrinology, cardiology, toxicology, and psychology.41 As the veterinary field and their clients

(by which is meant the pet owners) discovered that interventions taken from human medicine

could dramatically affect outcomes, the rate of technology transfer accelerated. This adoption

included pharmaceuticals, which Novartis estimated in 1999 as a $3 billion market in the US

alone, growing at 20 percent annually.42 Interestingly amongst all these sweeping changes in the

field of veterinary medicine, one aspect that has remained essentially the same is that it is the only

major medical field that is paid for, essentially solely, by client’s discretionary cash flow.

Some Details about Pet Health Insurance

As this study is not particularly about the intricacies of insurance, but rather the

relationships between entrepreneurs and resources, some simplifications will be utilized in

regards to discussing insurance. The pricing, issuance, and regulation of insurance is a complex

field, a complexity that would most likely get in the way. In order to address this complexity

some aspects of the story have been simplified when it was felt that such simplification would not

compromise the data or the theorizing.

Pet health insurance is an insurance product designed to defray the potential medical

costs that can occur when a pet sustains an injury or an illness that requires veterinary care. Like

most insurance products it is premised on the notion that the pooling of risks allows policy

holders to pay in a steady stream of premiums and receive compensation when a covered event

occurs. Unlike more commonly available livestock insurance, which is designed to provide the

insured party with economic value coverage, pet health insurance is designed as a tool for

absorbing unanticipated veterinary expenses.

41 The Times, April 6, 1998. “Bright Eyes and Bushy Tails Cost US Pounds 6 Billion” 42 Newsweek, October 11, 1999. “When Pets Pop Pills”.

167

Like many smaller, niche insurance products, policies are sold by the issuer (the pet

health insurance firm in this case), but are underwritten by larger, established multiline insurance

firms. These smaller firms are called MGAs (managing general agents) and are responsible for

marketing and sales, issuing policies, handling claims, risk assessment and pricing. An MGA

manages the day to day of providing the insurance product and pays a share to the underwriter for

providing capital coverage as mandated by regulation and other services. In order for an MGA to

remain in business they must offer a product that appeals to customers and also one that works

economically. Built into the premium for a policy is the actuarial estimate of how much will be

paid out to settle claims, a fee to be paid to the underwriter, any regulatory or taxation fees, the

estimated expenses that the MGA will incur in servicing the policy, and any remaining profit for

the MGA. The largest portion of the premium can be linked to the expected claims payouts,

which is guided by the underlying actuarial model. For an MGA to remain viable their actuarial

model must be accurate relative to the product type they are offering, otherwise the MGA may

find itself in a situation where it has to pay out more than it budgeted. Such a situation would lead

to the underwriter having to cover the difference, an occurrence the underwriter clearly wants to

avoid. Further the amount that a pool of insurance policies is expected to pay out is governed by

the regulatory authority of the state. Successful niche, insurance products thread the needle

between providing a well-priced product that attracts consumers and also providing adequate

margins to account for expenses and profit, all the while being squeezed by regulators and

underwriters.

168

The Early Emergence of Pet Health Insurance (1980s-1990s)

While pet health insurance had been previously tried in certain Scandinavian countries,

this story starts in 1977 with the formation of Pet Plan in the UK.43 Patsy Bloom, a former charity

worker, and David Simpson launched the business with 500 pounds in capital and the believe that

the “I had to sell the concept before I could sell the product.”44 Mrs. Bloom was renowned for her

ability to enlist the involvement of others (in particular the veterinary community) as she traveled

the country with her dog Annie on a perpetual sales pitch. As Figure A.1 illustrates, the growth of

pet insurance in the UK was phenomenal, growing steadily through the late nineties and then

exploding thereafter. By 1993 Pet Plan was collecting premiums amounts in the tens of millions

and in 1996 was acquired by a large multi-line insurance firm, Cornhill Insurance.45 Several other

firms entered the market, including Tesco the multinational grocery and general merchandise box

store. By 2013 market penetration is approaching 25% in both the cat and dog markets.

Figure A.1: Adoption rate of pet health insurance in USA, Canada, and Britain between 1979 and 2011.

Rate is determined by the number of outstanding dog policies divided by the owned dog population in each

region. Some values have been interpolated from known data based on estimated growth rates.

43 The Guardian, May 27, 1989. “Cover for creatures great and small – Vets’ bills are proving a growing

burden for pet owners” 44 The Times, April 17, 1993. “Dogged determination wins through” 45 The Independent, May 2, 1996. “Cornhill buys Pet Plan for pounds 32.5m”

0.0%

5.0%

10.0%

15.0%

20.0%

25.0%

30.0%

Pet Health Insurance Penetration Rate:

1979-2011

USA

Canada

Britian

169

The start of pet health insurance in the USA was not the same rosy picture. Pet health

insurance got its start in 1982 with the founding of Veterinary Pet Insurance (VPI) by Jack

Stephens, DVM with funds raised from several hundred veterinary practices in California.46 The

main driving force behind this joint effort was the desire to reduce the occurrence of economic

euthanasia (when a pet’s owner elects to have a pet put down rather than incur the cost of

veterinary treatment).47 Growth in policy sales for VPI was anemic, with Dr. Stephens admitting

that his company lost money every year between 1982 and 2000.48 During this same time period

several firms attempted to enter the marketplace with similar offerings, and either failed after a

short time or never made it past the regulatory approval phase. A similar story played out in

Canada with several firms attempting entry, a few limping along and most simply failing outright.

Figure A.2: Pet Health Insurance Timeline for USA, Canada, and Britain

Contemporaneous data from these time periods shows that many characteristics of the

cultures, regulatory structures, and institutional regimes were nearly identical in the US, UK, and

Canada (including aspects such as pet owners willingness to pay for veterinary care, perceptions

of pet as a family member, other forms of insurance available and regularly purchased, regulation

46 The Washington Post, February 18, 1982. “Pets: Mr. Rover, Your Policy?” 47 The New York Times, July 15, 1982. “Insurance for the family pet” 48 The New York Times, June 30, 2002. “Break a leg, Fluffy, if you have insurance”

Pet Plan luanched in

Britian

Lloyd's of London

underwirintg Pet Plan

VPI launched in USA

AHIA launched in USA

PetPlan & Pet Sure

launced in Canada

PetPlan & Petsure

consolidated

AHIA fails

Pet Plan, UK sold to

Cornhill

VPI finally makes a profit

UK market saturated

Trupanion & PetPlan

launched in USA

Embrace & ASPCA

launched

Nationwide acquires VPI

VPI market share falls

from >90% to ~50%

1975 1980 1985 1990 1995 2000 2005 2010 2015

Pet Health Insurance Timeline

170

and pricing of veterinary care, etc.).49 Given very similar contexts it is a paradox that the rate of

adoption of pet health insurance was so very divergent amongst these countries.

Whilst VPI eventually became a successful business, it became cash flow positive in the

2000s and was later acquired by Nationwide Insurance in 2008, the history it laid down in the

process had important implications for the research question. For the vets who initially funded

and formed VPI an overriding goal was to reduce the occurrence of economic euthanasia and to

provide pet owners an alternative means to pay for care. They felt the best way to accomplish this

goal was to introduce a product that was as cheap as possible and thus would get into the most

hands. While they were not oblivious to the economics of running a business, many of them being

veterinary practice owners, neither were they experts in the field of insurance.

The primary pricing mechanism of an insurance product is the underlying actuarial model.

An actuarial model provides a set of probabilistic estimates for how likely a covered event will

occur during the life of the policy. Based on a pool of policyholders one can calculate the

expected amount that will be paid out across the pool. Individual policies are then priced such

that there are adequate reserves to cover the expected payout plus additional monies to provide

for fees, expenses, and profit to the insurance entities. There are many forms of models to choose

from; in this case a schedule of benefits model was elected. A schedule of benefits policy is

designed to provide a given payment of y dollars for a given procedure x (for example a policy

might pay $200 for the removal of a foreign object from a dog’s stomach). The advantage to such

a model is that it is less data intensive than many of the other options. In this case VPI needed to

estimate the average occurrence rate of each illness or accident (called morbidity) that the policy

covered and make an overall average estimate of the amount that should be paid out to cover the

related treatment. The disadvantage of such a model is that it assumes that care is priced the same

49 These findings are from various studies and reports put out by both veterinary (AVMA, NCVEI, BVA,

CVMA) and animal welfare organizations (ASPCA, RSPCA, SPCA) in all three countries.

171

for all policyholders (i.e. vets charge the same for the same procedure) and if payouts for

procedures are not adjusted often enough they can quickly fall out of sync with the market. It can

be seen that in some cases policyholders will file for a claim and receive reimbursement that is

most or all of what they paid out for treatment. Such cases will occur when a veterinarian’s fees

are in line with the schedule of benefits model’s assumptions. Other policyholders will receive

only a fraction of their claim when for various reasons a veterinarian’s fees are greater than those

assumed by the schedule of benefits. Further it can be seen that if on average veterinarians’ fees

increase (things never seem to get cheaper) and the model is not adjusted, on average

policyholders will receive a smaller reimbursement as a percentage of their bill.

So why did VPI elect a schedule of benefits model? First, as previously mentioned, such

a model is less data intensive and requires less back-office actuarial modeling to upkeep. More

complex models, such as the percentage of bill, require both morbidity data as well as pricing

data by geographic regions (with finer grain data leading to more differentiable pricing). Not only

is it more difficult to get this greater depth of information, it is harder to process this data into

interpretable models. In the early 1980s, when VPI was started the founders and their staff simply

didn’t have the horsepower to support these more complex models, both in terms of

computational and actuarial capabilities. Secondly, VPI was started in California, a state well

known for its heavy regulatory framework. At the time of VPI’s formation, regulators pushed VPI

towards the schedule of benefits model, under many of the same concerns previously expressed.

Unfortunately the downsides of the schedule of benefits model outweighed the

advantages in implementing the product, particularly when it came to customers and their

relationship with veterinarians. Few, if any, customers (i.e. pet owners who bought a VPI policy)

understood the schedule of benefits model or if they understood the concept, they could not

navigate the minutiae of veterinary terminology. In essence, this policy format imposed a

significant information asymmetry on the client. When an animal was treated and care was paid

172

for the customer had no idea how much they were going to receive back when they eventually

filed a claim with VPI. This led to situations in which a pet owner would take their animal to the

vet, receive care and pay for treatment, then receive a reimbursement some time later that was a

fraction of the paid bill (estimated around 50-55% by firm founder during this time period).

Inevitably this upset the pet owner who would then accuse the vet of over-charging, engaging in

over-care, or otherwise being dishonest. Further it was not uncommon for clients to assume,

truthfully or not, that a vet had given their approval and recommendation for the insurance policy.

Such assumptions led to increased acrimony between client and vet. Vets for their part didn’t

have the resources to aid clients in navigating the insurance process and resented being the

accused party.

Further issues exacerbated the formation of negative sentiment towards the concept of pet

health insurance. VPI felt that its policies might still be priced too highly, even with the schedule

of benefits model. The easiest way to deal with this issue in insurance is to add exemptions to the

policy, things that the policy will not cover. A primary exemption that was made was the

exclusion of pre-exiting conditions, such an exemption is needed in order avoid the situation in

which only pet owners with known issues would buy policies. Insurance, of this sort, is designed

to pay for unexpected expenses, in the case of pre-exiting conditions clients already know that

they will incur costs and thus have an incentive to a buy a policy that will shift some of the cost

onto other policyholders. This exemption was reasonable and well understood by clients; most

pet-health insurance policies to this day exclude pre-exiting conditions. Other exclusions found in

VPI policies were much more problematic. In particular policies excluded breed specific,

predisposed diseases. For example the spitz family of dogs has a genetic predisposition to hip

dysplasia (abnormal development of the hip joint that leads to poor fit between the ball and

socket, often leading to painful dislocation injuries). Such exclusions remove a known pool of

risk for the policy originator, thus allowing for an overall reduction of policy pricing.

173

Unfortunately they impose yet more information asymmetries on the policyholders, most of who

will have no knowledge of the prevalence or perhaps even existence of predispositional diseases.

These policy exemptions, other steps taken to reduce the cost of policies, and other

various factors led to policy products that were confusing for both customers and vets.

Additionally the few other firms that attempted entry during this period essentially copied VPI’s

business model. None of these firms survived, but they did act to reinforce the assumptions of

vets, clients, regulators, economists, and other interested parties that pet health insurance simply

wasn’t a viable product. The overall sentiment concerning pet insurance as of 2000 was very

negative, with many vets viewing the concept as anathema to the practice of veterinary care.

There was much information that provided a clear negative signal for the viability of pet health

insurance as a service product (at least as it was understood at the time). However, in a broader

context there was clear information that pet owners had a greater willingness and desire to

provide quality veterinary care for their companions and that vets were willing to and wanted to

provide this care. However, the general notion of insurance, in particular the role of human health

insurance, was also playing a confounding role into how the market for pet health insurance

might be viewed. In summary, entrepreneurs faced many contradictory signals, as did regulators,

underwriters, vets, legislators, pet owners, and other related parties.

174

Appendix B: In-Depth Timeline of Pet Health Insurance (1977-2012)

USA, Canada, Great Britain

Since 1945

o According to the American Kennel Club at least 30 companies have testes the market for pet

insurance and no has succeeded (as of Sept. 1980)

o According to Guy Hodge, director of information services for the Human Society of the United

States, more than 35 companies have come and gone in this sector since 1945 (as of Feb. 1982).

Plans failed from such factors as undercapitalization, inadequate actuarial information, and lack of

support from veterinarians.

1977

o Patsy Bloom and David Simpson launch Pet Plan in the UK with 500 pounds in capital. Patsy

promotes this ‘new’ product along with her dog, Annie, by traveling around dog shows and

veterinary practices. “I had to sell the concept before I could sell the product”. She sponsors local

meetings on the British Small Animal Veterinary Association on condition that she can give a talk

about her product. Originally underwritten by Dog Breeders Insurance Company, with 1,300

policies sold in the first year. By 1980 this had grown tenfold and Pet Plan adopted Llyod’s as

underwriter.

1980

o Janruary - Pet Health Support of Anaheim begins offering pet health insurance with annual

premiums ranging from $23 for cats and $47 for dogs. Effort goes nowhere.

o September – Judi Goose of Santa Ana and Medial Pet Services (MPS) of San Diego will begin

offering pet health insurance. Policies range from $31 for cats to $70 for dogs. Since June MPS has

paid out claims of $10,000. Both fold.

1981

o November - California Veterinary Services (founded in 1980 with funds from 700 to 800

veterinarians in the state) will begin offering pet insurance in California in 1982 (the future parent

of VPI). Rhulen Agency Inc. will do the same in New York.

1982

o February – Veterinary Pet Insurance (VPI), a division of California Veterinary Services will make

policies available in March. Advertisements are posted in 500 of the state’s 1700 veterinary

hospitals, more than 35,000 people have sent for information. Pre-existing conditions, intentional

injuries, and congenital or hereditary defects will not be covered (this has implications for later in

the life of VPI). Neither plan (catastrophe only or catastrophe with sickness and major medical)

cover routine care. The lack of routine coverage is intentional to keep premiums low. VPI is still

awaiting approval from the California Department of Insurance.

o April – Frontier Insurance Company will offer pet insurance in New York as of April 27th (Frontier

is the new name for the efforts of Rhulen Agency Inc.). Mr. Rhulen states that if pet insurance is

successful in New York then Frontier will go national. Policy restrictions are similar to VPI’s

offerings, with the additional condition that all pets of the same species within a household must be

covered.

175

o July – Concern is starting to be expresses that the availability of pet health insurance will lead to

substantial increases in the cost of veterinary care. The National Insurance Consumer Organization

argues that pet medical bills are not the type of expense for which people should buy insurance.

“It’s an absurd expenditure for insurance” and “This is in the junk coverage category” according to

J. Robert Hunter, president of NICO: “If you are really worried that someday you will have a big

veterinary bill, put $50 a year way in a bank account and collect interest on it.” Hunter asserts that

these plans will result in additional and unneeded costs to consumers: “If everybody buys the

insurance we will get CAT scans for cats and dog scans for dogs and all kinds of crazy machines

for pets that nobody would ever have through of using. And pet owners will pay for it.”

1983

o January – Frontier Insurance reports that it has 350 policyholders, roughly half in NYC (premiums

range from $42-$79). Nearly $2,000 in claims has been paid out. VPI has sold 7,000 policies and

paid close to $70,000 in claims (premiums range from $19-$120).

1985

o January – John Robbins (independent insurance agent) begins writing health insurance policies for

pets in Westerville, Ohio. Not clear that his policies are in fact legal, according to Ohio statue.

o November – Alpo plans to begin testing a pet insurance program for consumers in cooperation with

an insurance company on the West Coast (probably VPI). Nothing more of this plan is heard.

1988

o March – Animal Health Insurance Co. began selling policies in Connecticut in 1987, hopes to start

selling in 49 states (Tennessee does not allow pet-care policies). VPI has about 150,000

policyholders at this point and offers plans in 27 states, with plans to expand to 12 more states by

mid-year. According to Rebecca Moore, marketing representative for VPI, veterinary costs rose

183% between 1981 and 1986 (this seems like an exaggeration). Michael Garvey, chairman of the

department of medicine at New York’s Animal Medical Centre, says growth of the pet health

insurance industry has been slow because of a lack of advertising and promotion.

o May- AHIC receives the support of the Massachusetts Society for the Prevention of Cruelty to

Animals and starts offering plans in that state.

1989

o May – In the UK, several insurance agencies drop out of the market including: Prudential stops

accepting new customers for it Prupet contract, Vetex withdraws its credit-card type scheme, and

Holman General Facilities ceases operating its Holdfast dog and cat plan. Pet Plan has over

175,000 policyholders (at least 51% of the UK market), policies are underwritten by Lloyd’s.

o June- AHIC is now Animal Health Insurance Agency (AHIA), and is being underwritten by

Llyod’s (which has long history of writing obscure policies).

o October – Two companies enter the Canadian market. Pet Plan will be sold by Reed Stenhouse Ltd

(Canada’s largest insurance brokerage), imitates the UK product. PetSure will be offered by

PetSure Canada Inc., a subsidiary of Aegon Insurance Co. (a large European firm).

o November – The Massachusetts Co. begins offering the Pet Lover’s Visa Card, which offers a 10%

discount on pet health insurance through Lloyd’s of London.

1990

o October – Nichol Insurance Brokers Ltd. becomes the latest Canadian company to offer pet health

insurance, called Medipet Anti-Maux.

1991

o January- Although VPI and AHIA claim to have sold over 250,000 policies between them,

veterinarians say that they have no or only a few clients with coverage. A segment of clients are

dropping the policies after discovering that the insurance does not cover what they expected, or

reimbursements are very slow in coming (or not at all). Veterinaries are expressing the opinion that

not covering routine care is a substantial reason for why they don’t recommend the insurance to

their clients. Prior efforts in the area of insurance have failed because there were too many limits on

coverage.

176

o March- California Insurance Commissioner John Garamendi files charges against VPI for delays in

paying claims.

o May – Current estimate of 100,000 animals covered nationwide by VPI and AHIA. This is a tiny

fraction of the potential market. Pet insurance is available in 49 states, except Tennessee which still

does not allow it.

o October – A South Florida chain of animal clinics is launching a health maintenance organization

which would provide routine care and a discount on surgery for an annual premium.

1992

o February – Estimate of 100 million pets in the USA (meaning that VPI and AHIA cover roughly

0.091% of the market), 52 million dogs and 55 million cats. The Fireman’s Fund Insurance

Company will start promoting the Medipet plan, in partnership with AHIA. They will be airing a

cable television infomercial aimed at reaching 20 million viewers. This follows a direct mail

campaign begun the prior fall through Sears Roebuck that targeted 600,000 households. Consumer

advocates and some economists state that this new effort by a mainstream company is ridiculous.

Mr. Hunter re-asserts the position that insurance in this sector will lead to unnecessary costs and

cause clients to approve treatments that they wouldn’t okay if they were covering the cost directly.

He expressed interest in life insurance for cats: “Imagine a type of life insurance where you don’t

have to pay the first eight times the policyholder dies.”

o July – Jardine, a Birmingham, UK-based insurance brokerage, launches “Moggies and Mongrels”

pet insurance policy. Also offers ‘Paws’ plan.

o Fall – Reed Stenhouse sells Pet Plan (Canada) to H.E.D. Leipsic. PetSure Canada Inc. was almost

closed down, but Pet Plan Group Ltd. of London England took it over. Pet Sure and Pet Plan are

combined into one entity.

1993

o April – Pet Plan in the UK, now is collecting premiums in the tens of millions of pounds.

o August – Pet Plan (Canada) is receiving blowback from canceling policies for higher risk clients.

The Ontario Insurance Commission is investigating claims of false advertising. Llyoyd’s the

underwriter flagged policies it wanted dropped after the transition of Pet Plan and Petsure. Pet

Plan’s 14,000 polices cover far less than 1% of Canadian pets, the company believes it needs to get

5% to stay in business.

1994

o April 1st – AHIA fails after a national effort to promote is unsuccessful.

o May – Pet Plan Canada now has 12,000 policies. Medi-Pet has folded and Pet Sure was merged

with Pet Plan. Pet insurance is now a $150 million dollar industry in the UK.

1995

o August - Current estimates of US pet population: 63 million cats and 54 million dogs, with

Americans spending $17 billion a year on them. Increase in the number of newspaper articles

discussing the changing nature of pets and their inclusion as a family member.

o September – A report from the Animal Hospital Association finds that 70% of the population

considers pets to part of the family. However 69% of pet owners don’t like insurance plans, saying

they are too expensive.

o For the year – In the UK pet owners paid $89 million in premiums to insure 700,000 dogs and cats

(estimated at 5% of the population)

1996

o May – Pet Plan, UK is sold to Cornhill for 16 million pounds.

o November – The Consumer Federation of America provides a list of insurance products that

consumers should avoid, included in the list is pet health insurance. Mr. Hunter, now the director of

insurance for this entity, reiterates the idea that it is better to bank the premiums and highlights the

fact that the policies do not covering pre-existing conditions.

o December – VPI is trying to design a policy that covers condition endemic to certain breeds, such

as hip dysplasia.

1997

177

o June – The pet industry is now a $25 billion dollar market in the USA.

o July – VPI now has an estimated 75,000 policyholders and is available in 43 states. Vets, clients,

and the newspaper reports are increasingly frustrated with the caps on low-cost policies. They

assert that these policies are more akin to a discount on veterinary services, rather than proper

insurance. The American Veterinary Medical Association (AVMA) supports pet insurance calling

such coverage “important to the future of the veterinary profession’s ability to provide high quality

and up-to-date veterinary services.” Mr. Hunter retorts “It’s no accident that most pet insurance was

invented by vets, who are jealous of health insurance for people and of the high-price procedures

that allows.”

o August – Pet insurance in the UK now covers roughly 13% of dogs and 5% of cats, as compared to

Canada and the US were both of these percentages are below 1%.

o September- Pet Assure is established as a pseudo HMO, offering a 25% discount on veterinary care

and services through partner members (launched the prior year). Members choose from a list of

participating vets. Vets in return receive referrals and marketing assistance.

o October – Veterinary Centers of America Inc. invests $6 million in VPI.

1998

o April - The AVMA estimates that Americans spend $6.68 billion a year on pet health care. VPI has

roughly 75,000 policies (estimate of $9 million in premiums). Rewards Plus of America adds pet

health insurance to its offerings. “We were all laughing” says Frank Longwell, vice president of

marketing.

o August – Firms offering pet health insurance as a job benefit, in the increasing aggressive effort to

recruit employees. A study by the American Pet Product Manufactures Association shows that 80%

of pet owners celebrate their pet’s birthday.

o November – Pet Assure moves into group marketing to target employee benefit packages. CEO Jay

Bloom expects hundreds, if not thousands of companies to begin offering pet care as a benefit in

the next few years.

1999

o January – Roughly 16,000 Canadians now have pet insurance.

o March – Pets Health, in Canton, Ohio has sold 4,000 policies in the past 18 months in 40 states.

o June – A bill is introduced to the New York assembly that would permit the establishment of court-

administrated trusts for the care and feedings of dogs and other animals.

o July – American spend ~$27 billion annually on pets, with $12 billion going to healthcare. The pet

pharmaceutical market is estimated at $3 billion and 20% annual growth.

o October – VPI has struggled, but sales rose 90% last year to $26 million (with premiums averaging

$200, implying ~130,000 policies).

o During the Year – Pet Care Insurance Brokers Ltd. is started in Oakville, Ontario.

2000

o February – Los Angeles-based Answer Financial Inc. AFI, an online provider of voluntary benefits

of voluntary benefits ads five pet insurance plans to its portfolio. Joining an increasing group of

providers.

o May – Vancouver City Savings Credit Union begins offering the SafeRate pet insurance policy. “It

all has an oh, what next ring to it” according to Marianne Chatten, sales rep with VanCity

Insurance Services Ltd. PetPlan is still the dominant player in the Candian market, with an

estimated 15,000 clients. Other competitors include Pet Care & Petcetera, a retail pet store that

offers insurance. The Ontario, Alberta, and Canadian Veterinary Medical Associations endorse the

PetCare program exclusively for a two-year period (probably a pay for endorsement deal). Valerie

Goddard, marketing executive for Pet Plan, states “People thought we were this big monopoly

making tons of money when we are not. We have waited for competition for a long time. It brings

credibility.”

o July – VPI annual premium is roughly $265 a year.

o November – Reader’s Digest Association will start marketing Pethealth Inc’s accident and health

insurance policies in the United States and Canada. Pethealth in Oakville, Ontario

178

o December – According to Dr. Stephens, VPI has about 200,000 policies in force. VPI is now

majority owned by Nationwide Insurance (60%) via Scottsdale. Premier, in Wisconsin, is also

insuring several thousand dogs, started 1998 by Tom Kurtz.

2001

o January – Vetinsurance, underwritten by Allianz Insurance Company of Canada (a subsidiary of

Allianz AG) is introduced in Canada. Lincoln General Insurance Co. agrees to underwrite Pethealth

Inc. entry into the United States.

o July – The American Pet Product Manufactures Association estimates the pet market at $28.5

billion. A study by the AAHA reports that 75% of survey participants would be willing to go into

debt in order to pay for veterinary care. AIG introduces Health Pet. VPI estimates that revenue will

reach $55 million (roughly 250,000 policies). Dr. Stevens estimates that his four competitors have

at best 25,000 to 30,000 policies combined.

2002

o April – Royal & Sun Alliance (RSA) will start fielding a 26 person team of pet bereavement

counselors in an effort to increase the attractiveness of its UK policy offering. The pet insurance

industry in the UK is estimated at 160 million punds, with 12% of dog owners and 7% of cat

owners. There are now roughly 60 entities offering policies, Pet Plan is still the dominant player

with 40%, Tesco Personal Finance has captured 20%.

o June – Rescue shelters and adoption groups have begun offering two-months of pet health

insurance coverage as part of the adoption package. VPI reports an eightfold increase in revenue

over five years, nearly $72 million. The American Kennel Club enters a partnership with a British

firm to start offering insurance. Pethealth Inc. (Candaian) forms an alliance with Petco to sell

policies in the US. Current renewal rates at VPI average 82%. Dr Stephens (VPI) reveals that his

company lost money every year between 1982 and 2000. Estimates profit of $2,000,000 on 250,000

policies in 2001.

o November – Owing to economic troubles firms are cutting back on employee benefits, but pet

health has remained relatively sticky (the number of companies offering this perk doubled in the

last decade). This is mostly due to the voluntary nature of the offering, and the fact that few

employers cover part of the cost. Complications that these benefits might fall under the purview of

ERISA slow down adoption.

o December – A study from AAHA finds that 47% of owners would spend any amount to save a

pet’s life. A report that rescue dogs from the World Trade Center are receiving better care than

human workers leads to investigations and complaints. VPI has donated lifetime medical policies to

every rescue dog.

2003

o January – Major players in the US market include VPI, Petcare (Candian), and Petshealth Care Plan

(which recently absorbed Premium Pet Insurance after it went bust and lost its underwriter).

o June- The USA pet market is projected to grow 10% this year to $31 billion. Pharmaceutical

growth is enormous, for example pain control is up 275% in six years to more than $150 million.

o August – Petshealth sells 11,736 new core policies during the second quarter of 2003 and 42,761

ShelterCare policies (a cross promotion with petfinder.com). Sales are up 130% over the prior year.

However payouts are up by 23%, and the business is not profitable owing to the high rate of claims.

Packaged Facts, a division of MarketResearch.com, estimates that spending on Pet Insurance in the

US climbed 342% from 1998 to 2002.

o September – Estimate that VPI now has 340,000 policies and pays out 35,000 claims a month.

Pethealth has now sold 21,723 new core policies for the first six months of 2003, a 76% increase

over the 12,323 sold in the same period in 2002. Laura Bennet, plans to launch Embrace Pet

Insurance in the US the following year.

2004

179

o March – Pethealth Inc’s fourth quarter loss is $349,701 for the three months ending Dec. 31, 2003,

compared with a loss of $1.14 million during the same period a year earlier. Quarterly revenue

grew to $2.04 million from $1.21 million. Chris Ashton starts Fetch Pet Insurance, which will later

launch Petplan insurance in the USA (August, 28, 2006), underwritten by American National

Property and Casualty Company. Fetch holds an exclusive license with Petplan Limited, a wholly

owned subsidiary of Allianz Cornhill Insurance PLC.

o June – The AVMA estimates that 58% of US households have at least one pet, with an estimated

68.9 million cats and 66.6 million dogs. HSBC joins an alliance with VPI to sell policies to its bank

clients.

o Novemeber – Petshealth reports a third quarter loss of about $50,000 with revenue of $2.99 million.

2005

o February – The pet industry has doubled in size in the last ten years to $34 billion. VPI has sold

360,000 policies during 2004, as compared to 157,000 in 2000. About 1,100 US companies offer

VPI as an employee benefit.

o May – Don Cherry, a famous Canadian hockey commentator, joins a partnership with Petheath Inc.

to create an insurance program for dogs and Cats (named CherryBlue)

o November – A study from the AKC reports that on average dog owners will incur $2,127 in one-

time expenses and $2,489 in annual expenses.

o December – In the UK open heart surgery is performed successfully on a cat for the first time.

2006

o May – Estimate that VPI now has 369,000 active policies (roughly 80% of the market). Less than

1% of the estimated 90.5 million cats and 73.9 million dogs. Roughly $110 million in premiums/

o August – PetPlan (Canada) reports that the company is growing 35-40% a year (yet less than 1% of

owners in Canada have coverage)

o July – Fetch Inc. begins selling policies under the PetPlan license from Cornhill Allianz. Chris

Ashton pursues the PetPlan partnership in order to get the brand name cache and access to the

actuarial data held by the parent company.

o September – Petsecure (from Australia and Canada) plans to launch operations in the US market.

Hollard insurance in the planned underwriter, with Petsecure providing all back office

administration and risk profiling.

o October – The ASPCA beings offering a plan in connection with Hartville Group.

o November – PetPlan (Canada) insures 42,000 pets. SecuriCan General Insurance which

underwrites and administrates PetPlan, also underwrites programs for PC Financial, the Canadian

Automobile Association, and Overwaites Foods.

2007

o January – Sales of pet insurance in the US topped $160 million in 2005, up nearly 25% from 2004.

Packaged Facts and Consumer Reports estimate that Americans spent $230 million on pet health

insurance in 2006.

o February – The pet pharmaceutical industry is $5 billion, growing 14% annually. There are now

90,000 pet policies on the books in Canada. VPI now covers 415,000 policies. An estimated 1,600

US firms now offer pet insurance as an employment benefit.

o July – Nestle Purina PetCare Co. launches PurinaCare Pet Health Insurance in Canada,

underwritten by SecuriCan General Insurance Co. PurinaCare, unlike most insurance, covers

routine examinations.

o August – Canadian pet insurance now covers roughly 110,000 pets.

o October – Vsurance begins offering life insurance policies for dogs in the US. Coverage is

available for up to $10,800. Eli Lily comes out with a canine version of Prozac.

2008

o February – Fetch Inc. has grown to 11 employees, with expectations to expand to 100 in the next

three years. VPI has 400 employees and $150 million in annual premium sales. Datamonitor

forecasts the UK pet health insurance market to grow to $1.17 billion in 2011 from nearly $740

million in 2006.

180

o March – AVMA reports that the national average for a veterinarian visit in 2006 was $135 for dogs

and $112 for cats.

o April – AVMA reports that $24.5 billion on health care for all pets in 2006. Pet population

estimated at 81.7 million cats and 72 million dogs. There are 83,730 veterinarians in the nation.

o August – VPI estimated at 450,000 policies, double six years earlier.

o October – VPI now at 465,000 policies. Sales are remaining resilient in light of economic

conditions.

2009

o August – The pet industry has grown to $46 billion, from $17 billion in 1994. The American Pet

Products Association estimates that there are 93.6 million cats and 77.5 million dogs. Hurricane

Katrina and other events are pushing legislatures and institutions (such as the Red Cross) to

reexamine the classification of pets as property.

o November – Central States Indemnity, a subsidiary of Berkshire Hathaway, wins a contract to

underwrite PurinaCare in the US. Veterinary spending is expected to increase to $12.2 billion.

2010

o April – Fetch Inc. now employs 40 people. Their product PetPlan is rated tops by the Humane

Society of the United States

o June – Canada now at 140,000 pets covered of an estimated 14.4 million cats and dogs. SecuriCan

General is now Western Financial Insurance Co. For 2009 revenue from policies was $32 million,

up from $2.4 million in 2004. Net profit of $3.8 million, with roughly half of premium revenue

paid out in claims.

o August – VPI awards a bronze trophy to Ellie, a Labrador, for most unusual pet health insurance

claim, after she ate a beehive containing pesticides and thousands of dead bees.

2011

o April – VPI now at 485,000 policies, up from 195,000 in 2001. The American Pet Products

Association expects Americans to spend slightly less than $400 million on pet health insurance in

2011.

o May – In the UK, RSA joins Tesco in a joint effort to sell Pet Health Insurance.

o August- - Fetch, Inc. (i.e. PetPlan) reports a 2,207 % growth rate over three years. Revenue grew to

$18.7 million in 2010 from $812,000 in 2007. Firm has roughly 100,000 policyholders.

o October – Capital Blue Cross, in Harrisburg, starts insuring pets under a program managed by

Petplan.

2012

o March – USA Today reports that there are now 11 companies offering pet insurance in the US

market. Revenue in the sector was $303 million in 2009. Americans spent an estimated $14.1

billion in veterinary care in 2011 according to the American Pet Products Association.

181

Appendix C: SRMR & pseudo-SRMR (pSRMR)

SRMR is a global fit measure of how closely the model-estimated correlation matrix

differs from the correlation matrix of the observed data and is one of the most commonly used

measures of fit (Hu & Bentler 1999). In order to calculate SRMR it is necessary to generate a

matrix of the residuals (denoted 𝑹𝑅) between the original correlation matrix of the manifest

variables (denoted RD) and the correlation matrix implied by the model parameters denoted

(denoted RM). Some software can output this implied correlation matrix, otherwise it can be

calculated from the factor analysis model.

𝑹𝑅 = 𝑹𝐷 − 𝑹𝑀 (A1)

𝑹𝑅 = 𝑹𝐷 − (𝚲𝚽𝚲′ + 𝚯𝛿) (A2)

Assuming p observed variables and m latent factors, 𝚲 is a (p x m) matrix of factor

loadings, 𝚽 is a (m x m) correlation matrix between the latent factors, and 𝚯𝛿 is a (p x p) matrix

with unique variances on the diagonal and correlations between observed variable unique

variances (UVs) on the off-diagonal. 𝑹𝑅 will thus contain p(p+1)/2 unique elements. Note that

we include the diagonal in the calculation of SRMR, although in ML based factor analysis,

accounting for rounding errors, the residuals on the diagonal will always equal zero. Following

the Mplus Technical Appendix 5 SRMR is calculated as:

𝑆𝑅𝑀𝑅 = √(∑ ∑ 𝑟𝑗𝑘2

𝑘≤𝑗𝑗 )/ (𝑝(𝑝+1)

2) (A3)

182

In Equation A3 𝑟𝑗𝑘 is defined as:

𝑟𝑗𝑘 =𝑠𝑗𝑘

√𝑠𝑗𝑗√𝑠𝑘𝑘−

�̂�𝑗𝑘

√�̂�𝑗𝑗√�̂�𝑘𝑘

(A4)

In Equation A4 𝑠𝑗𝑘 represents a covariance between two observed measures, 𝑠𝑗𝑗 is the

variance of observed measure j, 𝑠𝑘𝑘 is the variance of observed measure k, �̂�𝑗𝑘is the model-

estimated covariance between measures j and k, �̂�𝑗𝑗 is the model-estimated variance for measure j,

and �̂�𝑘𝑘 is the model-estimated variance for measure k. It should be noted that in Equations A1

and A2 the off-diagonal elements of 𝑹𝑅 correspond to 𝑟𝑗𝑘.

The general guideline for the use of SRMR in ML estimation is that SRMR should be <

0.08 and ideally less than 0.05 (Hu & Bentler 1999). However, in a Bayesian setting SRMR has

its own distribution rather than being a point estimate (Levy 2011). At this time Mplus and other

programs do not include a Bayesian implementation for SRMR, which would require calculating

SRMR at each iteration of the MCMC chain. Until such features are available we propose as an

alternative, the psuedo-SRMR (pSRMR), which can be used for comparison purposes. This

measure is calculated like SRMR but is derived by comparing the original correlation matrix of

manifest indicators to a recreated correlation matrix using parameter values from the Bayesian

analysis. We elected to calculate pSRMR using the median values from the posterior of each

parameter in 𝚲, 𝚽, and 𝚯𝛿. Within a reasonable range of tolerance the pSRMR value derived

from the median point estimates of the model parameters should closely approximate the true

median value of the actual SRMR distribution.

183

Appendix D: 𝚯δ Matrix Estimation

When conducting the initial groundwork for this article, two of the co-authors had rather

divergent views on the potential utility of estimating the entire 𝚯𝛿 matrix. One felt that there was

the potential for resolving issues related to model fit that were arising from modeling noise,

unique sample characteristics, method factors, etc. The other was much more dubious that the

technique offered any theoretical value and rather served as a means to sweep all of these issues

under the carpet. In order to come to the consensus offered in this article, the more optimistic

author decided to run a simple experiment utilizing 𝚯𝛿 estimation. While the results here are not

definitive and do not carry the robustness of a simulation study, they certainly gave this

individual cause to pause.

The experiment started with the simple premise: can this technique solve model misfit

issues for models that are not theoretically justified? The hope was to find that the technique

would fail in these situations, in which case it could be inferred that estimation of 𝚯𝛿 provided a

means to handle modeling noise (rather than being a Band-Aid for model misspecification).

Along with the ESE scale demonstrated in the methodology section, the survey included

many other scales including career commitment (Blau, 1985), career insight & career identity

(Noe et al., 1990), general self-efficacy, (Schwarzer et al., 1997) and a multi-dimensional career

motivations scale (DeMartino et al., 2006). In total these scales are represented by 55 observed

items. In order to generate a nonsensical model 18 items were randomly selected and used to

create a measurement model with three factors connected to six manifest items each. As expected

the estimation of this PCS model using ML indicated unacceptable fit (RMSEA= 0.139; CFI=

184

0.535; SRMR=0.116). Further indicative of poor modeling are a wide range of factor loadings: -

0.267 to 0.745. Likewise Bayesian estimation of the same PCS model with diffuse priors on the

complexity one loadings, and degenerate priors on all cross-loadings produces a model with

unacceptable fit (PPC= 1259-1365; BIC= 24417; DIC= 21818).

At this point a three-factor model with no theoretical meaning and poor fit had been

developed. The question remained: what would estimating 𝚯𝛿 accomplish with this same model

and would such a model even converge? After specifying an inverse-Wishart prior for 𝚯𝛿,

estimation was attempted in MPlus. Although the model took many iterations to reach

convergence, estimation was successful. According to the PPC= -52.1 – 73.2 criteria the model is

able to faithfully replicate the underlying data correlation matrix. Upon further thought this is not

surprising; estimating the off-diagonal entries in 𝚯𝛿 creates a model that has the potential to

explain nearly all linear relationships in the underlying data. In a comparative sense it is clear

from the BIC= 24607 and DIC= 22327 that this more complex model would not be accepted

relative to the less complex PCS model, but without this comparison we would have no clear

grounds for rejecting the complex model, at least on the grounds of its ability to reproduce the

sample data.

While this model is contrived (a researcher would be likely to reject the model based on

poor factor loadings and other attributes), it does highlight a concern with indiscriminant

estimation of 𝚯𝛿. If estimation of 𝚯𝛿 allows a grossly misspecified model to recreate the

underlying data, it is almost certain that this approach will obscure important theoretical model

misspecifications. Certainly this is not the last word on the veracity of this technique, but it does

highlight the cautions we have provided in this manuscript. Incidentally, the co-author who was

relatively gung-ho about this technique at first is now inclined to take a cautious, pessimistic view

going forward.