Multidimensional Item Information 1
Running Head: MULTIDIMENSIONAL ITEM INFORMATION
Directional Item Information for the Multidimensional Three-Parameter Logistic Model
Damon U. Bryant
University of Central Florida
Damon U. Bryant, University of Central Florida, Department of Psychology, PH 215
Orlando, FL 32816-1390, E-mail: [email protected] , Phone: (386) 325-0343.
This research was supported in part by an Educational Testing Service (ETS) Harold T.
Gulliksen Psychometric Research Predoctoral Fellowship, awarded to Damon U. Bryant.
Multidimensional Item Information 2
Abstract
The primary purpose of this article was to derive the item information function in a specified
direction for the Multidimensional 3-Parameter Logistic (M3-PL) model. The M3-PL item
information function in a specific direction was shown to be a general case of item information
for a variety of unidimensional and multidimensional measurement models. A necessary
condition for maximizing item information for unidimensional and multidimensional 1-, 2-, and
3-PL models was given. Implications for future research were presented.
Title: “Directional Item Information for the Multidimensional Three-Parameter Logistic Model”
Keywords: Multidimensional Item Information, Compensatory Model, Maximum Information,
3-Parameter Logistic Model.
Multidimensional Item Information 3
Directional Item Information for the Multidimensional Three-Parameter Logistic Model
In item response theory, the item information function (IIF) and the location of maximum
item information play an important role in test development (Birnbaum, 1968; Hambleton,
Swaminathan, & Rogers, 1991; Lord, 1980; Samejima, 1977); this is especially the case when
constructing criterion-referenced and computer-adaptive tests (Weiss, 1982). In criterion-
referenced test design, items are selected that discriminate in the psychometrically desirable way
at a particular point on the latent trait (Lord, 1980). In computer-adaptive test development,
items are first calibrated on a pretest sample and then banked according to the location of
maximum item information or theta maximum, θmax. When examinees are administered an
operational version of a computer-adaptive test (CAT), algorithms are typically employed to
compute a provisional ability estimate and then select the next item with the maximum amount
of information at the provisional estimate of ability (Chang, Qian, & Ying, 2001). It is evident
that research and test development efforts with IRT measurement models require knowledge and
use of the IIF.
There has been a call for research to investigate multidimensional measurement models
(Embretson & Reise, 2000; Hambleton, et al., 1991; Segall, 2001). Moreover, there is also a
practical application of such models when the assumptions of unidimensional IRT are untenable.
Federal legislation (i.e., The No Child Left Behind Act) requires yearly assessments in public
schools for accountability purposes. These high-stakes assessments must provide not only an
overall index of performance, but also diagnostic and interpretative information in specific
content areas. It is likely that multidimensional models can serve this purpose [See Reckase
(1997) for an applied example]. In work organizations where applicants are selected on the basis
of test scores, multidimensional measurement models may facilitate placement of successful job
Multidimensional Item Information 4
applicants with different composites of skills or abilities. These models may also be helpful to
debrief unsuccessful applicants with diagnostic information so that specific knowledge or skills
can be targeted for further improvement.
Formulas for item information and theta maximum (θmax) are well known for
unidimensional 1-, 2-, and 3-Parameter Logistic (PL) models. See Table 1. However, IIFs and
theta maximums are not well known for compensatory, multidimensional models (Reckase &
McKinley, 1991; Segall, 1996). In general, multidimensional extensions of IIFs are represented
as matrices (Segall, 1996), thus selecting items with the maximum information at a point is
relatively difficult as compared to unidimensional approaches. Representing the item information
function as a single value or a scalar may simplify the process of selecting items for criterion-
referenced or computer adaptive multidimensional tests (Segall, 2001). Notwithstanding the
work done by Reckase (Reckase, 1997; Reckase & McKinley, 1991), little research has been
conducted in the area of scalar IIFs for multidimensional models. Reckase and McKinley (1991)
only derived the IIF in a direction for the Multidimensional 2-PL model. So, the purpose of this
article is to expand on previous work and provide explicit formulas for item information in a
direction and theta maximum for the Multidimensional 3-PL (M3-PL) model.
The Item Response Model
The probability of a correct response for the M3-PL model (Reckase, 1997) is
Pi(θj) = ci + (1 – ci) [1 + Exp(–L)] -1
(1),
where L = D(ai
,θj + di), D is equal to a scaling constant 1.7 or 1, ai is a vector of k discrimination
parameters for item i, [a1i, a2i, …, aki],, k is the number of dimensions, θj is a vector of k ability
Multidimensional Item Information 5
parameters for person j, [θ1j, θ2j,..,θkj],, and di is a scalar related to difficulty. The probability of
an incorrect response is given by Qi(θj) = 1 – Pi(θj) or
Qi(θj) = (1 – ci) [1 + Exp(L)] –1
(2).
The point of steepest slope in the ability space is known as multidimensional
discrimination,
MDISCi = ||ai|| = (ai
,ai)
1/2 (3),
where || || represents the length of a vector that is computed as the square root of the sum of
squared elements of a vector. MDISC is interpreted in the same manner as the discrimination
parameter (ai) in unidimensional IRT. The difficulty of the item is the signed distance from the
origin of the multidimensional space to the point of steepest slope. The formula for
multidimensional difficulty is given by
MDIFFi = – di(||ai||) -1
= –di /MDISCi (4),
and it is interpreted in the same way as the difficulty parameter (bi) in unidimensional IRT.
Reckase (1985) has shown MDIFF to be equal to the unidimensional measure of difficulty (bi)
when there is only one dimension.
Item Information Function
The IIF in a specific direction for the multidimensional logistic model (Reckase, 1997;
Reckase & McKinley, 1991) is
Ii(θj) = [∇ Pi(θj) · ui ]2[Pi(θj)Qi(θj)]
-1 (5).
∇ Pi(θj) · u is the directional derivative, ∇ Pi(θj) is the gradient. The vector of directional cosines,
ui, is [a1i/||ai||, a2i/||ai||, …, aki/||ai||], or [cos α1i, cos α 2i, …, cos α ki]
,, where cos α k is the cosine
Multidimensional Item Information 6
of the angle (α k) from the axis orthogonal to dimension k. It should be noted that ||ui|| = 1. Next
is the derivation of the information function in a specified direction.
The first term in brackets within (5) is the gradient of the function Pi(θj), which can be
written as
∇ Pi(θj) = [∂ Pi(θj)/∂θ1j, ∂ Pi(θj)/∂θ2j , …, ∂ Pi(θj)/∂θkj], (6),
where ∂ Pi(θj)/∂θkj is the first partial derivative of Pi(θj) with respect to θkj. The general
expression for the derivative of the kth
term in the gradient is as follows:
∂ Pi(θj)/∂θkj = ∂{ci + (1 – ci) [1 + Exp(–L)]-1
}/∂θkj = – 1(1 – ci) [1 + Exp(–L)]-2
Exp(–L)–Daki
= Daki(1– ci) [1 + Exp(L)] -1
[1 + Exp(–L) ] -1
(7),
or substituting expression (2) into (7)
∂ Pi(θj)/∂θkj = DakiQi(θj)[1 + Exp(–L)]-1
(8).
With the k elements of the gradient having the general form of (8) and the elements of ui having
the general form of aki/||ai||, the directional derivative of the function Pi(θj) in the direction ui can
be expressed as
∇ Pi(θj) · ui = D(ai
,ui)Qi(θj)[1 + Exp(–L)]
-1 (9).
With the right-hand side of (9) substituted in (5), the IIF for the M3-PL model in a specified
direction is
Ii(θj) = {D(ai
,ui)Qi(θj)[1 + Exp(–L)]
-1}
2[Pi(θj)Qi(θj)]
-1 (10)
or with some algebra, the item information in a direction ui becomes
Ii(θj) = D2(ai
,ui)
2 Qi(θj){Pi(θj)[1 + Exp(–L)]
2}
-1 (11).
Multidimensional Item Information 7
Corollary 1a. If it is assumed that there is no guessing (i.e., ci = 0), the information function in
(11) becomes the IIF in a direction for the M2-PL model,
Ii(θj) = D2(ai
,ui)
2 Pi(θj) Qi(θj) (12),
which is similar to the formula derived by Reckase and McKinley (1991).
Corollary 1b. If it is assumed that discrimination parameters on all of k dimensions are fixed at 1
and there is no guessing (i.e., ai = [1, 1, ..., 1], and ci = 0), then (11) reduces to the M1-PL,
Ii(θj) = D2k
Pi(θj) Qi(θj) (13),
where k is equal to the number of dimensions.
Corollary 1c. If there is only one dimension, then (13), (12), and (11) become the item
information functions for the unidimensional 1-, 2-, and 3-parameter logistic models,
respectively.
Theta Maximum for the Multidimensional 3-PL Model
The formula for the location of maximum item information or theta maximum in a
specified direction is derived by setting the directional derivative of the IIF for the M3-PL model
equal to zero,
∇ Ii(θj) · ui = 0 (14),
where ∇ Ii(θj) is defined as [∂ Ii(θj)/∂θ1j, ∂ Ii(θj)/∂θ2j, …, ∂ Ii(θj)/∂θkj ],
and ui was defined
earlier. The information function in (11) is expressed in a different form as
Ii(θj) = D2(ai
,ui)
2(1 – ci)Exp(–L){[1 + ciExp(–L)][1 + Exp(–L)]
2}
-1 (15),
where L is the logit, which is equal to D(ai
,θj + di). The general form of the kth element of the
gradient, ∇ Ii(θj), is derived using the quotient, product, and chain rules of calculus.
Multidimensional Item Information 8
∂ Ii(θj) / ∂θkj = ∂ D2(ai
,ui)
2(1 – ci)Exp(–L){[1 + ciExp(–L)][ 1 + Exp(–L)]
2}
-1 / ∂θkj
= [–Daki D2(ai
,ui)
2(1 – ci)Exp(–L) [1 + ciExp(–L)] [1 + Exp(–L)]
2 –
{2[1 + Exp(–L)] Exp(–L) – Daki[1+ ciExp(–L)] + [1 + Exp(–L)]2ciExp(–L)–Daki } D
2 (ai
,ui)
2(1–
ci)Exp(–L)]{[1 + ciExp(–L)][1 + Exp(–L)]2}
-2
= [–aki D3(ai
,ui)
2 [Exp(–L) – ciExp(–L)] [1 + ciExp(–L)] [1 + Exp(–L)]
2 +
aki2Exp(–L) [1 + Exp(–L)] [1+ ciExp(–L)] D3(ai
,ui)
2[Exp(–L) – ciExp(–L)] +
akiciExp(–L) [1 + Exp(–L)]2 D
3(ai
,ui)
2 [Exp(–L) – ciExp(–L)]]{[1 + ciExp(–L)][1 + Exp(–L)]
2}
-2
= [aki D3(ai
,ui)
2 [Exp(–L) – ciExp(–L)] [1 + Exp(–L)]
2{–1 – ciExp(–L) + ciExp(–L)}+
aki2Exp(–L) [1 + Exp(–L)] [1+ ciExp(–L)] D3(ai
,ui)
2[Exp(–L) – ciExp(–L)]] {[1 + ciExp(–L)][1
+ Exp(–L)]2}
-2
= [aki D3(ai
,ui)
2 [Exp(–L) – ciExp(–L)] [1 + Exp(–L)]
2{–1} + aki2Exp(–L) [1 + Exp(–L)]
[1+ ciExp(–L)] D3(ai
,ui)
2[Exp(–L) – ciExp(–L)]]{[1 + ciExp(–L)][1 + Exp(–L)]
2}
-2
= D3aki(ai
,ui)
2 [Exp(–L) – ciExp(–L)] [1 + Exp(–L)] [{–1}[1 + Exp(–L)] + 2Exp(–L)
[1+ ciExp(–L)] ]{[1 + ciExp(–L)][1 + Exp(–L)]2}
-2
= D3aki(ai
,ui)
2 [Exp(–L) – ciExp(–L)] [1 + Exp(–L)] [{–1}[1 + Exp(–L)] + 2Exp(–L)
[1 + ciExp(–L)] ]{ [1 + ciExp(–L)]2[1 + Exp(–L)]
4}
-1
= D3aki(ai
,ui)
2 [Exp(–L) – ciExp(–L)] [{–1}[1 + Exp(–L)] + 2Exp(–L)
Multidimensional Item Information 9
[1 + ciExp(–L)] ]{[1 + ciExp(–L)]2[1 + Exp(–L)]
3}
-1
= – D3aki(ai
,ui)
2[Exp(–L) – ciExp(–L)] [1 + Exp(–L)]{[1 + ciExp(–L)]
2[1 + Exp(–L)]
3}
-1
+ D3aki(ai
,ui)
2 [Exp(–L) – ciExp(–L)] 2Exp(–L) [1 + ciExp(–L)]{[1 + ciExp(–L)]
2[1 + Exp(–
L)]3}
-1
= –D3aki(ai
,ui)
2 [Exp(–L) – ciExp(–L)]{[1 + ciExp(–L)]
2[1 + Exp(–L)]
2 }
-1 +
D3aki(ai
,ui)
2 [Exp(–L) – ciExp(–L)] 2Exp(–L){[1 + ciExp(–L)] [1 + Exp(–L)]
3}
-1 (16).
From (16), the kth
element of the gradient ∇ Ii(θj) can be expressed as
∂ Ii(θj) / ∂θkj = [D3aki(ai
,ui)
2 [Exp(–L) – ciExp(–L)]{[1 + ciExp(–L)] [1 + Exp(–L)]
2}
-1]
{–1[1 + ciExp(–L)] -1
+ 2Exp(–L)[1 + Exp(–L)]-1
} (17).
With the k elements of the gradient, ∇ Ii(θj), having the general form of (17) and the elements of
ui have the general form of aki/||ai||, the directional derivative of the IIF, Ii(θj), in the direction ui
is expressed as
∇ Ii(θj) · ui = [D3(ai
,ui)
3 [Exp(–L) – ciExp(–L)]{[1 + ciExp(–L)] [1 + Exp(–L)]
2}
-1]
{–1[1 + ciExp(–L)]-1
+ 2Exp(-L)[1 + Exp(–L)]-1
}= 0 (18).
From (18), item information is maximized when the following condition is satisfied:
–1[1 + ciExp(–L)]-1
+ 2Exp(-L)[1 + Exp(–L)]-1
= 0 (19),
or Pi(θj) = .5Exp(L) (20).
Expression (20) is a necessary condition for maximizing item information for the
unidimensional and multidimensional 1-, 2-, and 3-PL models. In the unidimensional 1-PL
model, information is maximized when Pi(θj) = Qi(θj) = .5. When Pi(θj) is equal to .5, then the
Multidimensional Item Information 10
natural logarithm of the odds of getting the item correct is 0, i.e., ln[Pi(θj)/Qi(θj)] = 0, which
implies that L in (20) is 0 and Exp(L) = 1. This leaves the well-known condition for maximizing
information in the 1- and 2-PL cases, which is Pi(θj) = Qi(θj) = .5. For the 3-PL unidimensional
and multidimensional models, the right-hand side of (20) and the probability of a correct
response are not equal to .5 when information is maximized, but the equality is satisfied at a
different value, which is primarily a function of guessing. The expression in (20) can also be
written as
2Exp(–L) = [Pi(θj)]-1
(21).
The probability of a correct response can also be written as
Pi(θj) = [Exp(L) + ci][Exp(L) + 1] -1
(22).
With the right-hand side of (22) substituted into (21),
2Exp(–L) = [Exp(L) +1][Exp(L) + ci] -1
(23).
To solve for L, expression (23) is written as
2 + 2ciExp(–L) = Exp(L) +1 (24).
After multiplying both sides by Exp(L),
2Exp(L) + 2ci = Exp(2L) + Exp(L) (25),
which, after subtracting 2Exp(L) from both sides, is equivalent to
2ci = Exp(2L) – Exp(L) (26).
After multiplying both sides by 4, expression (26) becomes
8ci = 4 Exp(2L) – 4Exp(L) (27).
When 1 is added to both sides, then (27) becomes
8ci + 1 = 4 Exp(2L) – 4Exp(L) + 1 (28).
Multidimensional Item Information 11
By way of the binomial theorem, the equality of (28) can be written as
8ci + 1 = [2 Exp(L) – 1]2 (29),
which after taking the square root of both sides becomes
(8ci + 1)1/2
= 2 Exp(L) – 1 (30).
Solving for L in (30) gives
L = ln{.5 [1 + (8ci + 1)1/2
]} (31),
which becomes
D(ai
,θj + di) = ln{.5 [1 + (8ci + 1)
1/2 ]} (32).
Now the objective is to solve for the vector θ that maximizes the information function of the M3-
PL model in the direction ui. To that end,
ai
,θ = ln{.5 [1 + (8ci + 1)
1/2 ]}D
-1 – di (33).
Because ui
,= ai
,/||ai||, both sides are divided by ||ai|| or MDISCi that results in
ui
,θ = ln{.5 [1 + (8ci + 1)
1/2 ]}(D||ai||)
-1 – di(||ai||)
-1 (34),
Now both sides are pre-multiplied by ui,
ui ui
,
θ = ui[ln{.5 [1 + (8ci + 1)1/2
]}(D||ai||)-1
– di(||ai||) -1] (35).
Because the associative law holds for multiplication of matrices, the left-hand side of (35) can be
written as (ui ui
,)θ. The product of (ui ui
,) is a k by k matrix, U, which has a few properties that
should be noted: 1. It is symmetric, thus U = U,, 2. The determinant of the matrix U is zero (0),
thus it is singular, 3. U2 = U
,U = UU = U, thus U is idempotent, 4. The main diagonal elements
of U are cos2α ki, thus the trace of U [i.e., tr(U)] is equal to one (1), and 5. The rank of an
Multidimensional Item Information 12
idempotent matrix is equal to its trace, thus the rank of U is equal to 1. With U substituted for ui
ui
,
, expression (35) is written as
Uθ = ui[ln{.5 [1 + (8ci + 1)1/2
]}(D||ai||)-1
– di(||ai||) -1] (36).
Pre-multiply both sides by U yields
UUθ = Uui[ln{.5 [1 + (8ci + 1)1/2
]}(D||ai||)-1
– di(||ai||) -1] (37).
By the third property mentioned above for U, (37) can be written as
Uθ = Uui[ln{.5 [1 + (8ci + 1)1/2
]}(D||ai||)-1
– di(||ai||) -1] (38).
Expression (38) is written as
Uθ – Uui[ln{.5 [1 + (8ci + 1)1/2
]}(D||ai||)-1
– di(||ai||) -1] = 0 (39),
where 0 is the same size as ui or by the distributive law for matrices,
U{θ – ui[ln{.5 [1 + (8ci + 1)1/2
]}(D||ai||)-1
– di(||ai||) -1]} = 0 (40).
Solutions that satisfy (40) are when U is equal to a null matrix (O) or when
θ = ui[ln{.5 [1 + (8ci + 1)1/2
]}(D||ai||)-1
– di(||ai||) -1] (41).
Therefore, theta maximum in a specified direction for the M3-PL model is
θmax = ui[ln{.5 [1 + (8ci + 1)1/2
]}(D||ai||)-1
– di(||ai||) -1] (42)
or substituting (3) and (4) into (42)
θmax = ui[ln{.5 [1 + (8ci + 1)1/2
]}(D.MDISCi )-1
+ MDIFFi] (43).
Several corollaries follow from this result in (43).
Corollary 2a. The location on the kth
dimension where information is maximized is given by
Multidimensional Item Information 13
θmax k = [ln{.5 [1 + (8ci + 1)1/2
]}(D.MDISCi )-1
+ MDIFFi] cos αki (44).
Corollary 2b. If it is assumed that there is no guessing, i.e., ci = 0, expression (43) reduces to
θmax = MDIFFi ui (45).
Expression (45) is the same as that implied by Reckase and McKinley (1991) for the location of
maximum item information for the multidimensional 2-PL model.
Corollary 2c. If it is assumed that there is only one dimension, then the expression in (43)
reduces to the well-known formula for theta maximum derived by Birnbaum (1968),
θmax = ln{.5 [1 + (8ci + 1)1/2
]}(Dai )-1
+ bi (46).
IIFs and formulas for location of maximum item information for the multidimensional 1-, 2-, and
3-PL models are listed in Table 2.
Derivations have been shown for the IIF and theta maximum in a specified direction. Now
an example item is given that uses the derived results. The following are hypothetical
parameters for the M3-PL model: ai = [a1i = 1.5, a2i = .5],, ci = .2, di = .5, and D =1.7. A
graphical illustration of the IIF is shown in Figure 1. It appears that this item gives more
information along the first dimension; this is supported by a higher discrimination parameter on
θ1 (1.5) relative to θ2 (.5). From expression (43), theta maximum is θmax = [-.206, -.069],. This is
verified by the condition satisfied by (20), i.e., Pi(θmax) = .5Exp[1.7(ai
,θmax + .5)] = .653. The
item can be considered as moderately easy with maximum information provided at θ1 = -.206
and θ2 = -.069. Multidimensional difficulty (MDIFFi) is -.316, which is lower than either element
in θmax. In the presence of guessing, this is consistent with values of theta maximum and
Multidimensional Item Information 14
difficulty in the unidimensional case (Hambleton, et al, 1991). By (43) and (11), the maximum
amount of information in direction ui, [.949 .316],, at the location θmax is 1.23.
Conclusion
The purpose of this article was to present the IIF and theta maximum in a specific
direction for the M3-PL model. The M3-PL IIF was shown to be a general case of item
information for unidimensional 1-, 2-, and 3-PL models. A sufficient condition for maximizing
item information for unidimensional and multidimensional 1-, 2-, and 3-parameter logistic
models was given as Pi(θj) = .5Exp(L).
Researchers have shown that developing tests using the IIF is more efficient as compared
to other methods (Hambleton & de Gruijter, 1983; Hambleton, et al., 1991). Some evidence
suggests that using the location of maximum item information is the most effective for
unidimensional models (Gierl, Henderson, Jodoin, & Klinger, 2001). Future research may
explore the extent to which this generalizes to the multidimensional case. In research related to
computer-adaptive testing, algorithms for selecting items that fit the M3-PL model can be further
explored. Evidence suggests that item selection algorithms for computer-adaptive tests based on
multidimensional item response theory may be more effective than conventional unidimensional
IRT algorithms (Segall, 2001). In multidimensional adaptive testing (Segall, 1996), items are
selected to maximize some function of Fisher’s information matrix. For some testing purposes,
e.g., selection, this may be inefficient when decisions are made at a particular point (Samejima,
1977). Alternatively, the cut-off in a multidimensional, criterion-referenced test may be set at a
location given by [θ1, θ2, ..,θk]. Algorithms can select items that satisfy (20) and then compute
the value of information in the direction corresponding to the cut-off. Thus, the IIF presented in
this study may be useful for tests of this type. It is the intent of this article to provide some tools
Multidimensional Item Information 15
[(11), (20), and (43)] that will give a better understanding of the measurement process in
compensatory measurement models.
Multidimensional Item Information 16
References
Birnbaum, A. (1968). Some latent trait models and their use in inferring an examinee's ability. In
F.M. Lord and M.R. Novick, Statistical Theories of Mental Test Scores. Reading, MA:
Addison-Wesley.
Chang, H., Qian, J., & Ying, Z. (2001). a-stratified multistage computerized adaptive testing with
b blocking. Applied Psychological Measurement, 25(4), 333-341.
Embretson, S. E., & Reise, S. P. (2000). Item response theory for psychologists. Manwah, NJ:
Lawrence Erlbaum Associates.
Gierl, M.J., Henderson, D., Jodoin, M., & Klinger, D. (2001). Minimizing the influence of item
parameter estimation errors in test development: A comparison of three selection
procedures. The Journal of Experimental Education, 69(3), 261-279
Hambleton, R. K., & de Gruiter, D. N. (1983). Application of item response models to criterion-
referenced test item selection. Journal of Educational Measurement. 20(4), 355-367
Hambleton, R.K., Swaminathan, H. & Rogers, H.J. (1991). Fundamentals of Item Response
Theory. Newbury Park CA: Sage.
Lord, F.M. (1980). Application of Item Response Theory to Practical Testing Problems.
Hillsdale, NJ: Lawrence Erlbaum Associates.
Lord, F.M., & Novick, M. (1968). Statistical Theories of Mental Test Scores. Reading, MA:
Addison-Wesley.
Reckase, M. D. (1985). The difficulty of test items that measure more than one ability. Applied
Psychological Measurement, 9(4), 401-412.
Multidimensional Item Information 17
Reckase, M. D., & McKinley, R. L. (1991). The discriminating power of items that measure
more than one dimension. Applied Psychological Measurement, 15(4), 361-373.
Reckase, M. D. (1997). A linear logistic multidimensional model for dichotomous item response
data. In W. J. van der Linden & R. K. Hambleton (Eds.), Handbook of modern item
response theory (pp. 271–286). New York: Springer-Verlag.
Samejima, F. (1977). A use of the information function in tailored testing. Applied Psychological
Measurement, 1, 233-247.
Segall, D. O. (1996). Multidimensional adaptive testing. Psychometrika, 61(2), 331-354.
Segall, D. O. (2001). General ability measurement: An application of multidimensional item
response theory. Psychometrika, 66(1), 79-97.
Weiss, D. J. (1982). Improving measurement quality and efficiency with adaptive testing.
Applied Psychological Measurement, 6, 473-492.
Multidimensional Item Information 18
Table 1.
Item Response Functions, Item Information Functions, and Theta Maximums for Dichotomous,
Unidimensional 1-, 2-, and 3-Parameter Logistic Models
Model (Unidimensional)
Item Information Function
Theta Maximum
1- Parameter Logistic P(θj) = {1 + Exp[–D(θj – bi)]}
-1
Ii(θ) = D
2Pi(θ)Qi(θ)
θmax = bi
2- Parameter Logistic P(θj) = [1 + Exp(–L)]
-1
Ii(θ) = D
2ai
2Pi(θ)Qi(θ)
θmax = bi
3- Parameter Logistic P(θj) = ci + (1-ci)[1 + Exp(–L)]
-1
Ii(θ) = D
2ai
2Qi(θ){Pi(θ)[1 + Exp(-L)]
2}
-1
θmax = ln{.5[1 + (1 + 8ci)
1/2 ]}(Dai)
-1 + bi
Note: D is equal to a scaling constant 1.7 or 1, ai = discrimination of item i, bi = difficulty of item
i, ci = guessing of item i, Exp(-L) = 2.71828-L
, L = Dai(θ – bi), and Qi(θ)= 1– Pi(θ).
Multidimensional Item Information 19
Table 2. Item Response Functions, Item Information Functions, and Theta Maximums for Dichotomous,
Multidimensional 1-, 2-, and 3-Parameter Logistic Models
Model (Multidimensional)
Item Information Function
Theta Maximum
1- Parameter Logistic P(θj)= {1 + Exp[–D(1
,θj + di)]}
-1
Ii(θ) = D
2kPi(θ)Qi(θ)
θmax =[MDIFFi /(k)
1/2, …, MDIFFi /(k)
1/2],
2- Parameter Logistic Pi(θj)= [1 + Exp(–L)]
-1
Ii(θ) = D
2(ai
,ui)
2Pi(θ)Qi(θ)
θmax =[MDIFFi cos α1i, …, MDIFFi cos α ki]
,
3- Parameter Logistic Pi(θj)=ci + (1 – ci)[1 + Exp(–L)]
-1
Ii(θ) = D
2(ai
,ui)
2 Qi(θ){Pi(θ) [1 + Exp(–L)]
2}
-1
θmax=[ln{.5 [1 + (8ci + 1)1/2
]}(D.MDISCi )-1
+ MDIFFi]ui
Note: k =number of dimensions, 1 is a k x 1 vector of ones. D is equal to a scaling constant 1.7
or 1, ai is a k x 1 vector of discrimination parameters for item i, [a1i, a2i, …, aki],, θ is a vector of
k ability parameters, [θ1j, θ2j,..,θkj],, di is a scalar related to difficulty, L = D(ai
,θ + di), MDISCi
= ||ai||, MDIFFi = -di /||ai||, and u is a vector of directional cosines, ai /||ai|| or [cos α1i, cos α 2i,
…, cos α ki],.
Multidimensional Item Information 20
Figure Captions
Figure 1. Item Information Function for the 3-Parameter Logistic Multidimensional Model
Top Related