Multivariable Mendelian randomization: the use of pleiotropic genetic variants to estimate causal...
Transcript of Multivariable Mendelian randomization: the use of pleiotropic genetic variants to estimate causal...
Multivariable Mendelian randomization: the use ofpleiotropic genetic variants to estimate causal effects
Stephen Burgess ∗ Simon G. Thompson
Department of Public Health and Primary Care,University of Cambridge
December 11, 2014
Running head: Multivariable Mendelian randomization [37 chars].
Key words: instrumental variables, causal inference, Mendelian randomization,
pleiotropy, lipid fractions.
Abbreviations: 2SLS, two-stage least squares; CHD, coronary heart disease;
HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein choles-
terol.
Acknowledgements: Dr. Stephen Burgess is supported by a fellowship from the
Wellcome Trust (100114). The authors have no conflict of interest.
∗Address: Strangeways Research Laboratory, 2 Worts Causeway, Cambridge, CB1 8RN, UK.Telephone: +44 1223 748651. Fax: +44 1223 748658. Correspondence to: [email protected].
1
Abstract
A conventional Mendelian randomization analysis assesses the causal effect of a risk
factor on an outcome by using genetic variants solely associated with the risk factor of
interest as instrumental variables. However, in some cases, such as for triglycerides as
a cardiovascular risk factor, it may be difficult to find a relevant genetic variant that
is not also associated with related risk factors, such as other lipid fractions. Such a
variant is known as pleiotropic. In this paper, we propose an extension to Mendelian
randomization that uses multiple genetic variants associated with several measured
risk factors to simultaneously estimate the causal effects of each of the risk factors
on the outcome. This “multivariable Mendelian randomization” approach is similar
to the simultaneous assessment of several treatments in a factorial randomized trial.
Methods for estimating the causal effects are presented and compared using real and
simulated data, and the assumptions necessary for a valid multivariable Mendelian
randomization analysis are discussed. Subject to these assumptions, we demonstrate
that triglycerides-related pathways have a causal effect on the risk of coronary heart
disease independent of the effects of low-density lipoprotein cholesterol and high-
density lipoprotein cholesterol. (188 words)
2
Mendelian randomization employs genetic variants as instrumental variables to
estimate the causal effect of a risk factor on an outcome using observational data,
even in the presence of unmeasured confounding [1, 2]. A genetic variable is a valid
instrumental variable if:
i. the variant is associated with the risk factor of interest,
ii. the variant is not associated with any confounder of the risk factor–outcome
association,
iii. the variant is conditionally independent of the outcome given the risk factor and
confounders [3, 4].
These assumptions can be illustrated using a causal directed acyclic graph, dis-
playing a causal effect of one variable on another by an arrow and the absence of a
direct causal effect by the lack of an arrow (Figure 1) [5]. Although a genetic variant
need not be causally associated with the risk factor to be a valid instrumental vari-
able, we assume that there exists a causal variant for which the measured variant is
a proxy [6].
[Figure 1 should appear about here]
In order to avoid violations of the second and third instrumental variable assump-
tions, Mendelian randomization experiments have generally relied on genetic variants
which are associated with a single risk factor. In practice, however, many variants
are pleiotropic, that is associated with multiple risk factors. Indeed, in some cases,
there may be no variants which are solely associated with the risk factor of inter-
est, and a Mendelian randomization analysis cannot be performed without consid-
ering pleiotropic variants. In any case, it may desirable to include information on
pleiotropic variants in order to provide a more powerful analysis, provided that this
3
does not prejudice its validity. It may also be that multiple quantitative traits relat-
ing to the same risk factor are of interest; for example in cardiovascular disease, the
concentration of lipoprotein(a) and the size of lipoprotein(a) particles [7]. In this case,
the relative proportions of risk reduction associated with interventions separately tar-
geting lipoprotein(a) concentrations and the size of lipoprotein(a) particles may be of
interest, and the traits may be regarded as independent risk factors, even if the same
genetic variants influence both traits. The possibility of including multiple risk fac-
tors in an instrumental variable analysis is discussed in many econometric textbooks
[8], and applied instrumental variable analyses involving multiple risk factors have
been performed [9, 10], but we are unaware of application of the approach in genetic
epidemiology.
The context of this paper is that there are measurements on multiple genetic
variants and several associated risk factors, the causal effect of at least one of which
on the outcome is of interest. We assume that the genetic variants do not influence
the outcome via any pathway except those fully mediated by one of the measured
risk factors, or by some combination of the measured risk factors. Questions about
variants with potentially unmeasured or unknown pleiotropic associations are reserved
for the discussion. We initially discuss how pleiotropic associations may arise, and
the methods and necessary assumptions for estimating causal effects with several
risk factors. We demonstrate the use of these methods in an applied example, and
then construct a simulation study with parameters chosen to be similar to those in the
example to investigate how the methods perform. Finally, we discuss the application of
the methods in epidemiological practice and the interpretation of the applied example.
4
Methods
Mechanisms for association with multiple risk factors
There are several causal mechanisms by which a genetic variant may be associated
with multiple risk factors [11]. We divide the possible mechanisms into two cases
(Figure 2): (1) vertical pleiotropy, where a variant is associated with multiple risk
factors due to the causal effect of the primary risk factor on a secondary trait; and (2)
functional pleiotropy, where the genetic variant is associated with multiple pathways.
These two cases are not mutually exclusive; it is possible for both these to occur for
the same variant.
[Figure 2 should appear about here]
In the case of vertical pleiotropy, genetic variants associated with the primary risk
factor would be expected to show consistent associations with the secondary trait. For
example, genetic variants associated with higher body mass index would be expected
to show a consistent association with higher blood pressure. In this case, the causal
effect of body mass index on the outcome would include an indirect effect, mediated
through blood pressure, as well as a direct effect comprising all other pathways from
body mass index to the outcome that are not operating via blood pressure. If it is
assumed that the only pathway by which the genetic variants are associated with the
secondary trait is via the primary risk factor, then a simple Mendelian randomization
analysis would consistently estimate the causal effect of the primary risk factor on the
outcome in spite of the apparent pleiotropic association.
In the case of functional pleiotropy, we suppose there are multiple genetic variants
(at least as many variants as there are risk factors) which have different magnitudes
of effect on the risk factors. These genetic variants can be used to estimate the
causal effects of each risk factor even if none of the variants are specifically associated
5
with any one particular risk factor. Since Mendelian randomization is analogous to
a randomized trial [12], the use of genetic variants to assess the causal effects of
multiple risk factors in a single study is analogous to a factorial randomized trial
(Web Figure 1), where multiple randomized interventions are simultaneously assessed
[13]. We refer to such an analysis as “multivariable Mendelian randomization”.
Assumptions
For a multivariable Mendelian randomization analysis to be valid, the genetic variants
must satisfy a similar set of assumptions as a conventional instrumental variable, but
in this case the variants must be exclusively associated not with a single risk factor,
but with a set of measured risk factors. It is not necessary for each variant to be
associated with every risk factor in the set, but a variant cannot have associations
with the outcome except via the risk factors of interest. Specifically, for each variant,
we assume that:
i. the variant is associated with one or more of the risk factors,
ii. the variant is not associated with a confounder of any of the risk factor–outcome
associations,
iii. the variant is conditionally independent of the outcome given the risk factors
and confounders.
In order to define and interpret causal effect estimates, we initially assume that the
effect of each of the risk factors on the outcome is not mediated by another of the
risk factors: each risk factor can be intervened on independently of all the other
risk factors, and an intervention on one risk factor will not influence any other risk
factor. We refer to such risk factors as “causally independent”. A causal directed
acyclic graph corresponding to these assumptions with three genetic variants and two
6
risk factors is presented in Figure 3 (left). We later relax the assumption of causal
independence and allow causal effects between the risk factors, as in Figure 3 (right).
In this paper, we assume that all associations are linear.
[Figure 3 should appear about here]
Individual-level data: two-stage least squares method
If individual-level data are available on the genetic variants, risk factors and outcome,
causal effects of the risk factors on the outcome can be estimated using a two-stage
least squares (2SLS) approach [14]. The risk factors are regressed on the genetic vari-
ants in a multivariate linear regression (first stage; a multivariate multiple regression,
as there are multiple dependent variables and multiple independent variables), and
then the outcome is regressed linearly on the fitted values of each of the risk factors
(second stage; a univariate multiple regression, as there is one dependent variable
and multiple independent variables). An alternative model for the genetic association
with the risk factor could be proposed (such as one including interaction terms), but a
model which is additive and linear in the variants is used here for comparability with
the summarized data methods considered in the next section. Although a sequential
regression approach gives the correct point estimates, the use of 2SLS software (such
as the ivreg2 command in Stata [15]) is recommended for estimation in practice to
give correct standard errors [16]. Estimates from the method are valid even if the
genetic variants are in linkage disequilibrium.
Summarized data: likelihood-based method
If individual-level data are not available, but rather summarized (aggregated) data
on the beta-coefficients and standard errors for the associations between the genetic
variants and the risk factors and outcome from separate univariate regressions, then
7
the causal effects of the risk factors on the outcome can be estimated using a likelihood-
based method [17]. For example, if there are two risk factors X1 and X2, which each
have no causal effect on the other, a multivariate normal distribution can be assumed
for the beta-coefficients representing the genetic associations with each of the risk
factors X1, X2 and the outcome Y from univariate linear regressions. Specifically, we
assume that the estimate of association of genetic variant j, j = 1, . . . , J with X1 is
X1j with standard error σX1j, and similarly with X2 (X2j, standard error σX2j) and
with Y (Yj, standard error σY j):
X1j
X2j
Yj
∼ N3
ξ1j
ξ2j
β1ξ1j + β2ξ2j
,
σ2X1j ρ12σX1jσX2j ρ1Y σX1jσY j
ρ12σX1jσX2j σ2X2j ρ2Y σX2jσY j
ρ1Y σX1jσY j ρ2Y σX2jσY j σ2Y j
(1)
Estimates of the causal effects of X1 and X2 on Y (β1 and β2) can be obtained
by numerical maximization of this likelihood function, or by Bayesian methods [18].
If there are K risk factors, data are required on K + 1 beta-coefficients for each ge-
netic variant (X1j, . . . , XKj, Yj) and corresponding standard errors; there are K(J+1)
parameters in the model, and equation (1) is a (K + 1)-variate normal distribution.
If the outcome is binary, and the beta-coefficients for the genetic association with
the outcome represent log-relative risks or log-odds ratios, then the causal effect esti-
mates will represent log-relative risks or log-odds ratios respectively. The model for
the genetic associations with the outcome is linear in contributions from the genetic
associations with the risk factors.
The parameters ρ12, ρ1Y , ρ2Y represent the correlations between the beta-coefficients.
These will be non-zero if the beta-coefficients are derived from the same data. Al-
though these correlations can only be estimated from individual-level data, they
8
should be approximately equal to the observational correlations between the vari-
ables X1, X2, and Y . It is advisable to conduct a sensitivity analysis to assess the
impact of these parameter values on the causal estimates. If data on the associations
with the risk factors and outcome are obtained from separate data sources, then the
relevant correlations will be zero.
As the likelihood function comprises contributions from each variant, it is necessary
that the information on the causal parameters provided by each variant is indepen-
dent. Therefore the genetic variants used in a summarized data analysis must be
uncorrelated (not in linkage disequilibrium); otherwise confidence intervals estimated
by the method will be too narrow [17]. If the genetic variants are in linkage disequilib-
rium and the correlations between the variants are known, then these correlations can
be used in a modified likelihood-based model: the correlations between the genetic
variants are the same as the correlations between the beta-coefficients corresponding
to the genetic variants. If all the variants are correlated, then instead of a separate
(K +1)-variate normal distribution for each of the J genetic variants, we can employ
a J(K + 1)-variate normal distribution in (1) for all of the variants together.
Summarized data: regression-based method
A further method which has been proposed for the analysis of summarized data is
a linear regression-based approach, which gives estimates for each of the risk factors
separately [19]. This is performed in two stages. First, the beta-coefficients for the
genetic association with the outcome are regressed on the beta-coefficients for the
competing risk factors. Then, the residuals from the first regression are regressed on
the beta-coefficients for the risk factor of interest.
For example, if there are two risk factors X1 and X2 and we want to estimate the
effect of X1 on Y , the two stages are:
9
1. Regress the beta-coefficients Y1, Y2, . . . , YJ on the beta-coefficientsX21, X22, . . . , X2J
to obtain residuals Y = Y − β2X2.
2. Regress the residuals Y1, Y2, . . . , YJ on the beta-coefficients X11, X12, . . . , X1J .
The regression-based estimate is the regression coefficient for X1.
The intuitive rationale is that these residuals represent any causal effects not explained
by the alternative risk factors, but which are potentially explained by the risk factor
of interest. However, it is an ad hoc approach with no clear theoretical basis, and
which ignores the uncertainty in the beta-coefficients [20].
Example: causal effects of LDL-C, HDL-C and triglyc-
erides on CHD risk
The causal nature of the associations of various lipid fractions, including low-density
lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C) and
triglycerides, with the risk of coronary heart disease (CHD) is an issue with important
consequences for disease prevention and drug development strategies. Observational
studies have shown associations of LDL-C with increased CHD risk, of HDL-C with
decreased CHD risk, and a null association of triglycerides with CHD risk upon adjust-
ment for a number of risk factors including HDL-C and non-HDL-C concentrations,
systolic blood pressure and body mass index [21]. However, a causal interpretation to
these results may be misleading due to unmeasured confounding and the possibility
that some of the covariates adjusted for are on causal pathways, making their inclusion
in a regression model inappropriate. Efforts to elucidate causal relationships using ge-
netic variants in a Mendelian randomization approach have indicated that LDL-C has
a causal role in increasing the risk of CHD [22], and have suggested a null causal
effect of HDL-C on CHD risk [23]. However, the latter estimate had wide confidence
10
intervals, because only a few variants, those not associated with other lipid fractions,
were included in the analysis. The inability to find variants associated with triglyc-
erides and not associated with LDL-C or HDL-C has precluded reliable Mendelian
randomization investigations for triglycerides.
We address the question of the causal effects of LDL-C, HDL-C and triglycerides
on CHD risk by multivariable Mendelian randomization using published data. Wa-
terworth et al. reported genetic associations from univariate regressions for 28 genetic
variants with log-transformed LDL-C, HDL-C and triglycerides, and with the log odds
of CHD [24]. Details of the variants and the beta-coefficients for the associations are
given in Web Table 1. Figures 4 and 5 depict the associations of each of the variants
with the lipid fractions and CHD risk.
We combine these beta-coefficients in a multivariable Mendelian randomization
analysis using the likelihood-based and regression-based methods. Estimates using
the likelihood-based method were obtained in a Bayesian framework using WinBUGS
(http://www.mrc-bsu.cam.ac.uk/bugs). Technical details of the analysis and the code
used are provided in Web Appendix 1. A sensitivity analysis for the values of the
correlation parameters (ρ1Y , ρ2Y , . . .) is given in Web Table 2. Initially, we do not
account for linkage disequilibrium between the genetic variants, so that the analysis
methods can be more directly compared.
Using the likelihood-based method, the multivariable Mendelian randomization
analysis gives causal odds ratios for CHD of 0.50 (95% credible interval 0.40–0.62) per
30% reduction in LDL-C, 1.22 (95% credible interval 0.91–1.63) per 30% reduction in
HDL-C, and 0.77 (95% credible interval 0.68–0.87) per 30% reduction in triglycerides.
This suggests that reductions in LDL-C and in triglycerides are causally protective of
CHD. The causal effect for HDL-C is compatible with the null. The regression-based
method gives rather different results: the corresponding odds ratios with 95% confi-
11
dence intervals are 0.69 (0.51–0.92) for LDL-C, 1.25 (0.90–1.74) for HDL-C, and 0.92
(0.82–1.04) for triglycerides. In particular, the causal odds ratio for triglycerides from
the regression-based method does not reach the conventional threshold for statistical
significance. When linkage disequilibrium between the genetic variants is accounted
for, the likelihood-based method gives odds ratios with 95% credible intervals of 0.52
(0.42–0.65) for LDL-C, 1.28 (0.96–1.71) for HDL-C, and 0.78 (0.70–0.86) for triglyc-
erides.
[Figure 4 should appear about here]
[Figure 5 should appear about here]
Simulation study
In order to assess the statistical properties of the analysis methods used, we perform
a simulation study. The set-up corresponds to the example above.
We generate data for 30 000 individuals indexed by i on three risk factors (X1, X2, X3)
and an outcome (Y ) from the following data-generating model:
x1i =28∑j=1
αG1jgij + αU2u2i + αU3u3i + ϵX1i (2)
x2i =28∑j=1
αG2jgij + αU1u1i + αU3u3i + ϵX2i
x3i =28∑j=1
αG3jgij + αU1u1i + αU2u2i + ϵX3i
yi = βU1u1i + βU2u2i + βU3u3i + β1x1i + β2x2i + β3x3i + ϵY i
gij ∼ Binomial(2, 0.3) independently for each j = 1, . . . , 28
u1i, u2i, u3i ∼ N (0, 1) independently
ϵX1i, ϵX2i, ϵX3i, ϵY i ∼ N (0, 1) independently
12
We set the genetic association parameters αG1j, αG2j, αG3j for j = 1, . . . , 28 to
take the values in Web Table 1 for log-transformed LDL-C, HDL-C and triglycerides
respectively, to be similar to the applied example. The instrumental variables gij
are drawn from binomial distributions, representing independent single nucleotide
polymorphisms with minor allele frequencies of 0.3. The causal effects β1, β2, β3 are
set to 0.3, 0, and −0.1. The variables U1, U2, U3 represent confounders, leading
to correlations between X1, X2, X3 and Y . The parameters βU1, βU2, βU3 are each
fixed at 0.3 throughout, and the parameters αU1, αU2, αU3 are varied to take values
0.3 or −0.3, leading to 8 different scenarios. The mean R2 values, representing the
proportion of variation in each risk factor explained by the 28 instrumental variables
together, for X1, X2, and X3 were 0.6%, 0.5%, and 3.2% respectively, corresponding
to mean F statistics of 6.6, 5.2, and 35.4.
Estimates for 1000 datasets generated in each of the 8 scenarios considered were
derived using the 2SLS, likelihood-based, and regression-based methods. The Monte
Carlo standard errors for the mean estimates were around 0.003 and for the power
around 1%. The likelihood-based method was performed in a Bayesian framework
using WinBUGS; technical details of the analyses are provided in Web Appendix 2.
Table 1 reports for each scenario the mean estimate, the mean standard error, the
standard deviation of estimates, and the power to detect a non-zero effect at a nominal
5% significance level. For β2 = 0, the expected power is 5%. We see that the mean
estimates from the 2SLS and likelihood-based methods are close to the true values,
with some deviation depending on the direction of confounding. This may represent
the effect of weak instrument bias, corresponding to the low F-statistics above for
X1 and X2 [25]. The efficiency of the 2SLS and likelihood-based methods is similar,
despite the reliance of the likelihood-based method on only summarized data.
In contrast, estimates from the regression-based method are biased, although they
13
appear give approximately valid inference for the presence of a causal effect under the
null. However the power, especially to estimate β3, is much lower than from the other
methods. We therefore recommend the likelihood-based method for use in practice
when summarized data are available.
[Table 1 should appear about here]
Causal relationships between risk factors
In order to investigate the performance of the methods when there are causal rela-
tionships between the risk factors, we repeated the simulation but replaced the first
line by:
x1i =28∑j=1
αG1jgij + αU2u2i + αU3u3i + αX2x2i + αX3x3i + ϵX1i
The additional terms αX2x2i and αX3x3i represent causal effects of X2 and X3 (which
were evaluated first) on X1. We set αU1, αU2, αU3 = 0.3 (the first scenario consid-
ered above) and took 9 values of the parameters αX2 and αX3 (Table 2). All other
parameters were taken as in the original simulation study.
[Table 2 should appear about here]
Table 2 shows the mean estimates of the causal parameters from each of the
methods. Aside from the regression-based method, which gives widely varying results,
we see that the estimates do not change substantially as the parameters vary. This
indicates that the 2SLS and likelihood-based methods estimate the direct causal effect
of each risk factor on the outcome, not including paths via the other risk factors.
This can lead to misleading conclusions about the total effects of the variables. For
example, when αX2 = 0, αX3 = 0.5, the total causal effect of X3 on Y is β3+αX1β1 =
−0.1 + 0.5 × 0.3 = 0.05 (including the path via X1). The mean estimates from the
14
2SLS and likelihood-based methods are in the opposite direction to the true total
effect.
The differences between the estimated values of β1, β2, β3 in Table 2 and their
true values can be attributed to weak instrument bias. Weak instrument scenarios
will be common in multivariable Mendelian randomization, as it is necessary to use
multiple instrumental variables to estimate the different causal effects. If the genetic
associations with the risk factors and with the outcome are measured in the same
dataset, this will lead to bias in the direction of the observational association [25],
whereas if the genetic associations with the risk factors and with the outcome come
from different sources (known as a two-sample instrumental variable analysis), the
bias will be in the direction of the null [26]. To demonstrate this, we repeat the
simulation study in Web Appendix 3 with fewer non-weak instrumental variables.
The mean estimates of the causal effect parameters are very close to their true values
(Web Table 3). In this simulation context, we also explore the impact of interactions
between genetic variants in their effects on the risk factors. The likelihood-based
method is robust to these misspecifications of the analysis model (Web Table 4).
We also investigate a modification to the 2SLS method referred to as “sequential
adjustment” by Holmes et al. [27], in which the causal effects of each of the risk
factors are estimated in turn, and alternative risk factors are adjusted for as if they
are confounders. Web Tables 5 and 6 indicate that substantial bias in the sequential
adjustment method is evident even under the null, and its direction depends on the
unknown confounders.
A non-zero causal estimate from a multivariable Mendelian randomization ap-
proach when there are causal relationships between the risk factors implies that the
variable is an independent causal risk factor, in the sense that an intervention on the
variable keeping the other risk factors constant (the controlled direct effect) would
15
affect the outcome. However, the magnitude of the causal estimate may not represent
the total causal effect of the variable on the outcome.
Discussion
In this paper, we have introduced multivariable Mendelian randomization, an impor-
tant and practically relevant extension to the Mendelian randomization paradigm for
the estimation of causal effects using genetic variants associated with more than one
risk factor. For a valid analysis, the variants must satisfy a set of assumptions sim-
ilar to those for an instrumental variable in conventional Mendelian randomization,
but modified to take account of the multiple risk factors. A multivariable Mendelian
randomization analysis may be beneficial where genetic variants are associated with
several related risk factors, such as in the example with lipid fractions. It enables the
causal evaluation of a risk factor even if no variants are uniquely associated with it,
as for triglycerides.
There are several limitations to this approach, many of which are shared with
conventional Mendelian randomization [28, 29]. The specific association of a genetic
variant with a single risk factor may be a reasonable assumption if the function of
the genetic region where the variant is located is well-characterized. The assumption
of an exclusive association between genetic variants and a set of risk factors is un-
likely unless the risk factors have strong biological associations. However, if they are
strongly associated, an assumption that the risk factors are causally independent is
less plausible. Weak instrument bias, a phenomenon by which instrumental variable
estimates using variants not strongly associated with the risk factor of interest are
biased, may be substantial if large numbers of genetic variants are used [30], as may
be necessary in a multivariable Mendelian randomization experiment. The effects of
the risk factors on the outcome are assumed to be linear. While some researchers
16
do not view this as a crucial assumption, citing the interpretation of an instrumental
variable estimate as an average causal effect [6, 31], others have shown that depar-
tures from linearity can affect the findings of an instrumental variable analysis [32].
Our preference is to take a less literal view of causal estimates, and to emphasize the
outcome of a Mendelian randomization analysis as testing a causal effect, rather than
necessarily estimating a causal parameter. From this perspective, while the linearity
assumption is important, it is less important than the other instrumental variable
assumptions.
Although multivariable Mendelian randomization is able to allow for genetic vari-
ants with ‘measured’ pleiotropic associations, under the assumptions discussed in this
paper, it is unable to deal with unmeasured or unknown pleiotropy. If an apparent
causal finding is dependent on the association of a small number of variants with the
outcome, then the result may plausibly be due to pleiotropic variants rather than
being a true causal effect. However, if several variants in different genetic regions
all demonstrate consistent associations with the outcome, then it is perhaps unlikely
that all of these associations reflect pleiotropic mechanisms [33]. In the case of the
applied example in the paper, we constructed a lipid risk score for each variant by
multiplying the genetic associations with each lipid fraction by the estimate of the
lipid fraction’s causal effect on CHD risk; details are given in Web Appendix 4. Web
Figure 2 displays the lipid risk score plotted against the log odds ratio of CHD risk for
each variant. Aside from variant rs2304130, it seems that the estimated causal effect
of the lipid fractions on CHD risk is consistent across variants, and so unmeasured
pleiotropy is unlikely to explain the causal effects.
We have performed the applied analysis for the causal effects of lipid fractions
on CHD risk in this paper at face value, assuming that the instrumental variable
assumptions are satisfied. In reality, the assumptions of the division into only three
17
lipid categories, and the effects of the genetic variants being restricted to these lipid
fractions are over-simplifications. Some lipid fractions (for example, intermediate-
density lipoprotein cholesterol) are omitted in the analysis, and the variability of
particle size within the categories is ignored [34]. An assumption that the causal effect
of triglycerides on CHD risk is independent of those of LDL-C and HDL-C may not be
satisfied, particularly as evidenced by the attenuation of the observational association
of triglycerides with CHD risk upon adjustment for HDL-C and non-HDL-C [21] and
studies of the APOA5 gene [35]. Our simulations above have shown that, in the case
of causal effects between risk factors, estimates represent the direct causal effect of
each risk factor on the outcome by a pathway that is not operating via the other risk
factors. This may not equal the total causal effect of the risk factor, but it provides
important evidence on the independent causal effect of the risk factor. Finally, the
estimated causal odds ratio for a 30% decrease in LDL-C reported in this manuscript is
surprisingly large, compared to both estimates of the effect of statin usage, which also
reduces LDL-C by around 30% [22], and to a Mendelian randomization analysis using
variants solely associated with LDL-C [17]. A nuanced interpretation of Mendelian
randomization estimates, and of multivariable Mendelian randomization estimates in
particular, is required in the light of the uncertainty of the underlying assumptions in
any applied analysis. These aspects are considered in more detail elsewhere [20].
In conclusion, the findings of this manuscript provide some evidence of a causal
effect of triglyceride-related pathways on CHD risk independent of the effects of LDL-
C and HDL-C, but the weight of evidence attributed to the findings is a matter
of interpretation depending on the degree of validity attributed to the instrumental
variable assumptions.
18
References
[1] Davey Smith G, Ebrahim S. ‘Mendelian randomization’: can genetic epidemiol-
ogy contribute to understanding environmental determinants of disease? Inter-
national Journal of Epidemiology 2003;32(1):1–22.
[2] Lawlor D, Harbord R, Sterne J, Timpson N, Davey Smith G. Mendelian random-
ization: using genes as instruments for making causal inferences in epidemiology.
Statistics in Medicine 2008;27(8):1133–1163.
[3] Greenland S. An introduction to instrumental variables for epidemiologists. In-
ternational Journal of Epidemiology 2000;29(4):722–729.
[4] Martens E, Pestman W, de Boer A, Belitser S, Klungel O. Instrumental variables:
application and limitations. Epidemiology 2006;17(3):260–267.
[5] Didelez V, Sheehan N. Mendelian randomization as an instrumental variable
approach to causal inference. Statistical Methods in Medical Research 2007;
16(4):309–330.
[6] Hernan M, Robins J. Instruments for causal inference: an epidemiologist’s
dream? Epidemiology 2006;17(4):360–372.
[7] Clarke R, Peden J, Hopewell J, Kyriakou T, Goel A, Heath S, et al. Genetic vari-
ants associated with Lp(a) lipoprotein level and coronary disease. New England
Journal of Medicine 2009;361(26):2518–2528.
[8] Wooldridge J. Introductory econometrics: A modern approach. South-Western,
Nashville, TN, 2009.
19
[9] Angrist JD. Instrumental variables methods in experimental criminological re-
search: what, why and how. Journal of Experimental Criminology 2006;2(1):23–
44.
[10] Ludwig J, Kling JR. Is crime contagious? Journal of Law and Economics 2007;
50(3):491–518.
[11] Hodgkin J. Seven types of pleiotropy. International Journal of Developmental
Biology 1998;42(3):501–505.
[12] Hingorani A, Humphries S. Nature’s randomised trials. The Lancet 2005;
366(9501):1906–1908.
[13] Stampfer M, Buring J, Willett W, Rosner B, Eberlein K, Hennekens C. The 2
× 2 factorial design: Its application to a randomized trial of aspirin and U.S.
physicians. Statistics in Medicine 1985;4(2):111–116.
[14] Baum C, Schaffer M, Stillman S. Instrumental variables and GMM: Estimation
and testing. Stata Journal 2003;3(1):1–31.
[15] StataCorp. Stata Statistical Software: Release 12. College Station, TX, 2011.
[16] Angrist J, Pischke J. Mostly harmless econometrics: an empiricist’s companion.
Chapter 4: Instrumental variables in action: sometimes you get what you need.
Princeton University Press, 2009.
[17] Burgess S, Butterworth A, Thompson S. Mendelian randomization analysis with
multiple genetic variants using summarized data. Genetic Epidemiology 2013;
37(7):658–665.
20
[18] Thompson J, Minelli C, Abrams K, Tobin M, Riley R. Meta-analysis of genetic
studies using Mendelian randomization – a multivariate approach. Statistics in
Medicine 2005;24(14):2241–2254.
[19] Do R, Willer CJ, Schmidt EM, Sengupta S, Gao C, Peloso GM, et al. Common
variants associated with plasma triglycerides and risk for coronary artery disease.
Nature Genetics 2013;45:1345–1352.
[20] Burgess S, Freitag D, Khan H, Gorman D, Thompson S. Using multivariable
Mendelian randomization to disentangle the causal effects of lipid fractions. PLoS
One 2014;available online before print.
[21] Di Angelantonio E, Sarwar N, Perry P, Kaptoge S, Ray KK, Thompson A, et al.
Major lipids, apolipoproteins, and risk of vascular disease. Journal of the Amer-
ican Medical Association 2009;302(18):1993–2000.
[22] Burgess S, Butterworth A, Malarstig A, Thompson S. Use of Mendelian randomi-
sation to assess potential benefit of clinical intervention. British Medical Journal
2012;345:e7325.
[23] Voight B, Peloso G, Orho-Melander M, Frikke-Schmidt R, Barbalic M, Jensen
M, et al. Plasma HDL cholesterol and risk of myocardial infarction: a mendelian
randomisation study. The Lancet 2012;380(9841):572–580.
[24] Waterworth D, Ricketts S, Song K, Chen L, Zhao J, Ripatti S, et al. Genetic
variants influencing circulating lipid levels and risk of coronary artery disease.
Arteriosclerosis, Thrombosis, and Vascular Biology 2010;30(11):2264–2276.
[25] Burgess S, Thompson S. Bias in causal estimates from Mendelian randomization
studies with weak instruments. Statistics in Medicine 2011;30(11):1312–1323.
21
[26] Pierce B, Burgess S. Efficient design for Mendelian randomization studies: sub-
sample and two-sample instrumental variable estimators. American Journal of
Epidemiology 2013;178(7):1177–1184.
[27] Holmes MV, Asselbergs FW, Palmer TM, Drenos F, Lanktree MB, Nelson CP,
et al. Mendelian randomization of blood lipids for coronary heart disease. Eu-
ropean Heart Journal 2014;[Available online ahead of print, January 27, 2014]
DOI: 10.1093/eurheartj/eht571.
[28] Davey Smith G, Ebrahim S. Mendelian randomization: prospects, potentials,
and limitations. International Journal of Epidemiology 2004;33(1):30–42.
[29] Schatzkin A, Abnet C, Cross A, Gunter M, Pfeiffer R, Gail M, et al. Mendelian
randomization: how it can – and cannot – help confirm causal relations between
nutrition and cancer. Cancer Prevention Research 2009;2(2):104–113.
[30] Burgess S, Thompson S, CRP CHD Genetics Collaboration. Avoiding bias from
weak instruments in Mendelian randomization studies. International Journal of
Epidemiology 2011;40(3):755–764.
[31] Angrist J, Pischke JS. The credibility revolution in empirical economics: How
better research design is taking the con out of econometrics. The Journal of
Economic Perspectives 2010;24(2):3–30.
[32] Mogstad M, Wiswall M. Linearity in instrumental variables estimation: Problems
and solutions. Tech. rep., Forschungsinstitut zur Zukunft der Arbeit. Bonn,
Germany., 2010.
[33] Davey Smith G. Random allocation in observational data: how small but ro-
bust effects could facilitate hypothesis-free causal inference. Epidemiology 2011;
22(4):460–463.
22
[34] Wurtz P, Kangas AJ, Soininen P, Lehtimaki T, Kahonen M, Viikari JS, et al.
Lipoprotein subclass profiling reveals pleiotropy in the genetic variants of lipid
risk factors for coronary heart disease: a note on Mendelian randomization stud-
ies. Journal of the American College of Cardiology 2013;62(20):1906–1908.
[35] Triglyceride Coronary Disease Genetics Consortium, Emerging Risk Factors Col-
laboration. Triglyceride-mediated pathways and coronary disease: collaborative
analysis of 101 studies. The Lancet 2010;375(9726):1634–1639.
23
Tab
le1:
Resultsfrom
simulation
studywithou
tcausalrelation
shipsbetweenrisk
factors
Estim
ates
ofβ1=
0.3
Two-stageleastsquares
(2SLS)
Likelihood-basedmethod
Regression-basedmethod
αU1
αU2
αU3
Mean
SE
SD
Pow
er(%
)Mean
SE
SD
Pow
er(%
)Mean
SE
SD
Pow
er(%
)
0.3
0.3
0.3
0.318
0.090
0.091
93.0%
0.317
0.092
0.091
91.2%
0.224
0.078
0.070
85.1%
0.3
0.3
-0.3
0.290
0.090
0.091
88.8%
0.290
0.092
0.092
86.9%
0.208
0.078
0.070
77.7%
0.3
-0.3
0.3
0.296
0.090
0.089
90.2%
0.296
0.092
0.090
89.3%
0.211
0.078
0.069
81.7%
0.3
-0.3
-0.3
0.271
0.090
0.089
86.7%
0.272
0.092
0.090
85.3%
0.193
0.077
0.067
74.9%
-0.3
0.3
0.3
0.329
0.091
0.092
93.1%
0.328
0.092
0.093
91.7%
0.234
0.079
0.073
85.9%
-0.3
0.3
-0.3
0.305
0.090
0.090
92.1%
0.304
0.092
0.091
91.0%
0.220
0.078
0.069
85.0%
-0.3
-0.3
0.3
0.309
0.090
0.086
93.6%
0.309
0.092
0.086
92.1%
0.221
0.079
0.068
84.4%
-0.3
-0.3
-0.3
0.283
0.090
0.088
88.1%
0.283
0.092
0.089
86.7%
0.204
0.078
0.068
77.9%
Estim
ates
ofβ2=
0
0.3
0.3
0.3
0.055
0.110
0.111
7.7
0.054
0.113
0.111
7.8
0.035
0.085
0.069
4.2
0.3
0.3
-0.3
0.010
0.111
0.113
5.9
0.010
0.114
0.114
6.1
0.007
0.085
0.069
2.2
0.3
-0.3
0.3
0.040
0.111
0.112
6.2
0.040
0.113
0.113
6.4
0.027
0.085
0.070
2.8
0.3
-0.3
-0.3
0.001
0.111
0.111
4.1
0.001
0.114
0.112
4.8
0.002
0.085
0.069
2.2
-0.3
0.3
0.3
0.001
0.111
0.112
3.9
0.001
0.114
0.113
4.7
-0.001
0.085
0.069
1.7
-0.3
0.3
-0.3
-0.050
0.111
0.107
6.6
-0.049
0.113
0.108
6.8
-0.033
0.085
0.067
2.8
-0.3
-0.3
0.3
-0.008
0.111
0.113
5.6
-0.007
0.114
0.115
6.0
-0.006
0.086
0.071
2.6
-0.3
-0.3
-0.3
-0.045
0.110
0.111
7.6
-0.045
0.112
0.112
7.1
-0.031
0.085
0.070
2.9
Estim
ates
ofβ3=
−0.1
0.3
0.3
0.3
-0.087
0.047
0.045
47.4
-0.087
0.047
0.045
46.4
-0.039
0.033
0.023
13.0
0.3
0.3
-0.3
-0.090
0.047
0.049
49.3
-0.090
0.048
0.049
49.8
-0.041
0.033
0.024
16.7
0.3
-0.3
0.3
-0.090
0.047
0.045
49.2
-0.090
0.047
0.046
47.8
-0.041
0.033
0.023
14.3
0.3
-0.3
-0.3
-0.094
0.047
0.047
52.0
-0.094
0.048
0.047
49.7
-0.043
0.033
0.023
16.9
-0.3
0.3
0.3
-0.106
0.047
0.048
61.2
-0.106
0.048
0.048
58.6
-0.049
0.033
0.025
25.8
-0.3
0.3
-0.3
-0.111
0.047
0.046
66.2
-0.110
0.048
0.047
62.6
-0.052
0.033
0.024
28.2
-0.3
-0.3
0.3
-0.105
0.047
0.044
63.5
-0.105
0.048
0.045
59.7
-0.049
0.034
0.023
24.8
-0.3
-0.3
-0.3
-0.111
0.047
0.046
66.4
-0.110
0.047
0.047
63.1
-0.052
0.033
0.024
28.9
Pow
eristheem
pirical
pow
erto
detectacausaleff
ectat
anom
inal
5%sign
ificance
level.Results(m
eanestimates,meanstan
dard
errors,stan
dard
deviationsof
estimates,an
dpow
er)aregiven
usingthreean
alysismethods(two-stag
eleastsquares,likelihood-based
,an
dregression-basedmethods)
toestimatecausaleff
ects
ofX
1on
Y(β
1=
0.3),ofX
2on
Y(β
2=
0),an
dof
X3on
Y(β
3=
0.1)
Abbreviation
s:SD,stan
darddeviation
;SE,stan
darderror
25
Table 2: Results from simulation study with causal relationships between risk factorsTwo-stage least squares Likelihood-based Regression-based
αX2 αX3 β1 β2 β3 β1 β2 β3 β1 β2 β30 0 0.318 0.055 -0.087 0.317 0.054 -0.087 0.224 0.035 -0.0390.5 0 0.322 0.038 -0.090 0.321 0.038 -0.090 0.229 0.020 -0.041−0.5 0 0.318 0.059 -0.089 0.318 0.058 -0.089 0.164 0.034 -0.0410 0.5 0.317 0.046 -0.097 0.316 0.045 -0.097 0.064 0.030 -0.0160 −0.5 0.321 0.048 -0.079 0.320 0.048 -0.079 0.192 0.030 -0.0370.5 0.5 0.316 0.041 -0.097 0.315 0.041 -0.097 0.077 0.021 -0.016−0.5 0.5 0.318 0.060 -0.098 0.318 0.060 -0.098 0.049 0.036 -0.0160.5 −0.5 0.318 0.042 -0.080 0.317 0.042 -0.081 0.125 0.022 -0.038−0.5 −0.5 0.317 0.057 -0.079 0.316 0.056 -0.080 0.240 0.034 -0.037
26
Figure captions
Figure 1: Causal directed acyclic graph of Mendelian randomization assumptions for
variant G with risk factor X in confounded association with outcome Y . Confounders
represented by U are assumed to be unknown.
Figure 2: Causal directed acyclic graphs of associations between variant G, risk
factors X1, X2, and outcome Y illustrating A) vertical and B) functional pleiotropy.
Figure 3: Causal directed acyclic graph of associations between variantsG1, G2, G3,
risk factors X1, X2, and outcome Y illustrating multivariable Mendelian randomiza-
tion. Confounders U1 and U2 are assumed to be unknown. A) Risk factors are causally
independent (no causal effects between X1 and X2); B) Risk factors are causally de-
pendent (X1 has a causal effect on X2).
Figure 4: Association of coronary heart disease (CHD) risk-increasing alleles of
28 genetic variants with all possible pairings of low-density lipoprotein cholesterol
(LDL-C), high-density lipoprotein cholesterol (HDL-C), and triglycerides (color and
size of points: darker points correspond to stronger associations with CHD risk, larger
points correspond to more precise estimates). Note that some points are overlapping.
Figure 5: Association of coronary heart disease (CHD) risk-increasing alleles
of 28 genetic variants with odds of CHD, and with each of low-density lipoprotein
cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C) and triglycerides
in turn (color and size of points: darker points correspond to stronger associations
with CHD risk, larger points correspond to more precise estimates). Note that some
points are overlapping.
27
Figures
G X
U
YFigure 1: Causal directed acyclic graph of Mendelian randomization assumptions forvariant G with risk factor X in confounded association with outcome Y . Confoundersrepresented by U are assumed to be unknown.
X1G Y
G X1 X2 Y1)
2)
X2
Figure 2: Causal directed acyclic graphs of associations between variant G, risk factorsX1, X2, and outcome Y illustrating A) vertical and B) functional pleiotropy.
28
G1
X1
U1
YX2
G2
G3 U2
G1
X1
U1
YX2
G2
G3 U2
Figure 3: Causal directed acyclic graph of associations between variants G1, G2, G3,risk factors X1, X2, and outcome Y illustrating multivariable Mendelian randomiza-tion. Confounders U1 and U2 are assumed to be unknown. A) Risk factors are causallyindependent (no causal effects between X1 and X2); B) Risk factors are causally de-pendent (X1 has a causal effect on X2).
29
Percentage change in LDL−C per allele
Per
cent
age
chan
ge in
trig
lyer
ides
per
alle
le
−5%
0+
5%+
10%
+15
%
−2% 0 +2% +4% +6%
Percentage change in HDL−C per allele
Per
cent
age
chan
ge in
trig
lyer
ides
per
alle
le
−5%
0+
5%+
10%
+15
%
−4% −2% 0 +2%
Percentage change in LDL−C per allele
Per
cent
age
chan
ge in
HD
L−C
per
alle
le
−2% 0 +2% +4% +6%
−4%
−2%
0+
2%
Figure 4: Association of coronary heart disease (CHD) risk-increasing alleles of 28 ge-netic variants with all possible pairings of low-density lipoprotein cholesterol (LDL-C),high-density lipoprotein cholesterol (HDL-C), and triglycerides (brightness and sizeof points: brighter points correspond to stronger associations with CHD risk, largerpoints correspond to more precise estimates). Note that some points are overlapping.
30
Percentage change in LDL−C per allele
Log
odds
of C
HD
per
alle
le
1.0
1.05
1.1
1.15
1.2
−2% 0 +2% +4% +6%
Percentage change in HDL−C per allele
Log
odds
of C
HD
per
alle
le
−4% −2% 0 +2%
1.0
1.05
1.1
1.15
1.2
Percentage change in triglyerides per allele
Log
odds
of C
HD
per
alle
le
−5% 0 +5% +10% +15%
1.0
1.05
1.1
1.15
1.2
Figure 5: Association of coronary heart disease (CHD) risk-increasing alleles of 28genetic variants with odds of CHD, and with each of low-density lipoprotein choles-terol (LDL-C), high-density lipoprotein cholesterol (HDL-C) and triglycerides in turn(brightness and size of points: brighter points correspond to stronger associationswith CHD risk, larger points correspond to more precise estimates). Note that somepoints are overlapping.
31
���������
������ �������
������ �������
������ ���������
������ �������
�����������������
���������
�����������
����������
��������������������
����������
������� ����
�����������
���
��������� �������
������ ����
�������������
��������� ������� ��
������������� �������
��
����
� ! "�#�� �����
������$
����
� ! "� ���%
%�&%��#��
���&%���%�&%��$
�������
''
��''
� ! "� ���&%���
%�&%��#�����%
%�&%��$
��''
� ! "� ���%
%�&%��#�����%
%�&%��$
Web Figure 1: Analogy between a factorial randomized trial with two treatments (Aand B) and multivariable Mendelian randomization with two genetic variants (1 and2). The minor allele of genetic variant 1 (a) has a large effect on LDL-C and a modesteffect on triglycerides; the minor allele of genetic variant 2 (b) has a large effect ontriglycerides and a modest effect on LDL-C.
Abbreviations: LDL-C = low-density lipoprotein cholesterol, TG = triglycerides.
1
Web Appendix 1
Supplementary methods for applied example
Web Table 1 gives the estimates of the associations between each of 28 genetic variants
and low-density lipoprotein cholesterol (LDL-C), triglycerides, high-density lipopro-
tein cholesterol (HDL-C), and the log odds ratio of coronary heart disease (CHD)
reported by Waterworth et al. [1].
Estimates for the analysis from the likelihood-based method were obtained in a
Bayesian framework using WinBUGS (http://www.mrc-bsu.cam.ac.uk/bugs). Direct
maximization of the likelihood is impractical in this case, as there are 28× 3+3 = 87
parameters to optimize over. In the likelihood function, the correlation between the
beta-coefficients for LDL-C and HDL-C was taken as −0.1, for LDL-C and triglyc-
erides 0.2, for LDL-C and CHD risk 0.1, for HDL-C and triglycerides −0.1, for HDL-C
and CHD risk −0.1, and for triglycerides and CHD risk 0.1. A sensitivity analysis for
these values is given in Web Table 2. Normal priors with mean 0 and variance 102 are
used for each of the unknown parameters. The WinBUGS code used to implement
the likelihood-based analysis is:
model {
beta1 ~ dnorm(0, 0.01) # prior for causal effect of LDL-C
beta2 ~ dnorm(0, 0.01) # prior for causal effect of HDL-C
beta3 ~ dnorm(0, 0.01) # prior for causal effect of triglycerides
for (i in 1:28) { # i indexes the 28 variants
Tau0[i,1:4,1:4] <- inverse(Sigma0[i,1:4,1:4])
# Sigma0 is the variance-covariance matrix for the multivariate
# distribution of the beta-coefficients
2
x[i,1:4] ~ dmnorm(xi[i,1:4], Tau0[i,1:4,1:4])
# x[i, 1:4] are the beta-coefficients for genetic variant i
# with LDL-C, HDL-C, triglycerides, and CHD risk
xi[i,1] ~ dnorm(0, 0.01)
# prior for mean association of genetic variant i with LDL-C
xi[i,2] ~ dnorm(0, 0.01)
# prior for mean association of genetic variant i with HDL-C
xi[i,3] ~ dnorm(0, 0.01)
# prior for mean association of genetic variant i with triglycerides
xi[i,4] <- beta1*xi[i,1] + beta2*xi[i,2] + beta3*xi[i,3]
# mean association of genetic variant i with CHD risk
}
}
The R code for the regression-based based analysis is:
for (j in 1:28) {
bx1 = lm(x1~gene[,j])$coef[2]
bx2 = lm(x2~gene[,j])$coef[2]
bx3 = lm(x3~gene[,j])$coef[2]
by = lm(y ~gene[,j])$coef[2]
bx1se = summary(lm(x1~gene[,j]))$coef[2,2]
bx2se = summary(lm(x2~gene[,j]))$coef[2,2]
bx3se = summary(lm(x3~gene[,j]))$coef[2,2]
byse = summary(lm(y ~gene[,j]))$coef[2,2]
}
beta1 = lm(lm(by~bx2+bx3)$res~bx1)$coef[2]
beta2 = lm(lm(by~bx1+bx3)$res~bx2)$coef[2]
beta3 = lm(lm(by~bx1+bx2)$res~bx3)$coef[2]
beta1se = summary(lm(lm(by~bx2+bx3)$res~bx1))$coef[2,2]
beta2se = summary(lm(lm(by~bx1+bx3)$res~bx2))$coef[2,2]
beta3se = summary(lm(lm(by~bx1+bx2)$res~bx3))$coef[2,2]
Estimates of the correlations between genetic variants were taken from the 1000
Genomes Pilot 1 dataset and obtained from the SNP Annotation and Proxy Search
(SNAP; http://www.broadinstitute.org/mpg/snap/ldsearchpw.php).
3
Sensitivity analysis for correlation parameters in likelihood-
based method
In order to assess the sensitivity of the causal estimates in the applied example to
the choice of correlation parameters, we performed a sensitivity analysis, estimating
the causal effects of LDL-C, HDL-C and triglycerides on CHD risk for a number of
different choices of the parameters. Web Table 2 displays the causal estimates for the
parameter values used in the main analysis (shown in italics), as well as for values 2,
1.5, 0.5, 0, and −1 times these values. We see that the causal estimates and standard
errors are fairly robust to different values of these parameters.
4
Genetic variant XL (SE) XT (SE) XH (SE) Y (SE)
rs11206510 0.026 (0.004) 0.016 (0.008) 0.002 (0.004) 0.068 (0.029)rs660240 -0.044 (0.004) -0.004 (0.008) 0.005 (0.004) -0.162 (0.030)rs515135 -0.038 (0.004) -0.009 (0.008) 0.003 (0.004) -0.105 (0.031)rs12916 -0.023 (0.003) 0.003 (0.006) 0.001 (0.003) -0.062 (0.024)rs2954021 -0.017 (0.003) -0.039 (0.006) 0.011 (0.003) -0.083 (0.022)rs1558861 -0.031 (0.006) -0.142 (0.012) 0.031 (0.006) -0.128 (0.067)rs2738459 -0.018 (0.004) 0.007 (0.007) -0.003 (0.004) -0.041 (0.037)rs10401969 0.046 (0.007) 0.095 (0.013) -0.007 (0.006) 0.077 (0.054)rs4420638 0.059 (0.004) 0.042 (0.008) -0.021 (0.004) 0.157 (0.031)rs10489615 0.004 (0.003) -0.023 (0.006) 0.018 (0.003) -0.030 (0.024)rs11902417 0.011 (0.004) 0.036 (0.007) -0.017 (0.003) 0.010 (0.028)
rs325 -0.005 (0.005) 0.097 (0.010) -0.047 (0.005) 0.182 (0.040)rs3890182 0.004 (0.005) 0.013 (0.009) 0.022 (0.004) -0.041 (0.034)rs964184 0.022 (0.005) 0.142 (0.009) -0.029 (0.004) 0.199 (0.034)rs9943753 -0.005 (0.004) -0.005 (0.007) 0.016 (0.003) 0.010 (0.038)rs261334 -0.002 (0.004) 0.019 (0.007) 0.034 (0.004) 0.049 (0.029)rs9989419 -0.002 (0.003) 0.003 (0.006) 0.035 (0.003) 0.010 (0.025)rs12449157 0.004 (0.004) -0.018 (0.008) 0.019 (0.004) -0.041 (0.032)rs2156552 0.011 (0.004) -0.020 (0.008) 0.028 (0.004) -0.030 (0.032)rs1168013 0.009 (0.003) 0.035 (0.007) 0.001 (0.003) -0.041 (0.024)rs6544366 -0.011 (0.004) -0.036 (0.007) 0.016 (0.003) -0.020 (0.028)rs1260333 -0.003 (0.003) -0.054 (0.006) 0.005 (0.003) -0.062 (0.022)rs1178979 -0.012 (0.004) 0.054 (0.008) -0.010 (0.004) 0.030 (0.030)rs10105606 0.003 (0.003) 0.067 (0.006) -0.023 (0.003) 0.068 (0.024)rs2954029 -0.015 (0.003) -0.040 (0.006) 0.012 (0.003) -0.073 (0.022)rs4938303 -0.008 (0.004) -0.067 (0.007) 0.018 (0.003) -0.073 (0.025)rs16965220 0.009 (0.003) 0.028 (0.006) -0.006 (0.003) 0.000 (0.026)rs2304130 -0.036 (0.007) -0.070 (0.013) 0.004 (0.006) 0.020 (0.065)
Web Table 1: Association of 28 genetic variants with log-transformed low-densitylipoprotein cholesterol (XL), log-transformed triglycerides (XT ), log-transformedhigh-density lipoprotein cholesterol (XH) and log odds ratio of coronary heart dis-ease (Y ) with standard errors (SE) taken from Waterworth et al. [1]
5
ρLH ρLT ρLY ρHT ρHY ρTY βLDL−C (SE) βHDL−C (SE) βTG (SE)
-0.2 0.4 0.2 -0.2 -0.2 0.2 1.96 (0.31) -0.53 (0.39) 0.74 (0.16)-0.15 0.3 0.15 -0.15 -0.15 0.15 1.96 (0.31) -0.55 (0.40) 0.74 (0.17)-0.1 0.2 0.1 -0.1 -0.1 0.1 1.96 (0.32) -0.56 (0.41) 0.73 (0.17)-0.05 0.1 0.05 -0.05 -0.05 0.05 1.97 (0.32) -0.57 (0.42) 0.73 (0.17)0 0 0 0 0 0 1.97 (0.33) -0.58 (0.42) 0.73 (0.18)0.1 -0.2 -0.1 0.1 0.1 -0.1 1.97 (0.34) -0.59 (0.44) 0.73 (0.18)
Web Table 2: Sensitivity analysis for the correlation parameters between beta-coefficients for genetic associations with low-density lipoprotein cholesterol (L), high-density lipoprotein cholesterol (H), triglycerides (T ) and CHD risk (Y ) in likelihood-based method for multivariable Mendelian randomization (original analysis shown initalics)
6
Web Appendix 2
Supplementary methods for simulation study
For the simulation studies, in the likelihood-based method, we regard the mean and
standard deviation of the posterior distribution as the ‘estimate’ and ‘standard error’,
and the 2.5th to 97.5th percentile range as the ‘95% confidence interval’. The obser-
vational correlations between the variables estimated in the individual-level data were
used for the correlation parameters. Normal priors with mean zero and variance 102
were used for each of the unknown parameters.
7
Web Appendix 3
Additional simulation studies with non-weak instrumental vari-
ables
Using non-weak instrumental variables
In order to show that the differences between estimates in Tables 1 and 2 from the
main text and the true values of the parameters for the two-stage least squares (2SLS)
and likelihood-based methods are due to weak instruments, we repeat the simulation
but using only 5 instrumental variables having strong associations with the risk factor.
The data-generating model is:
x1i =5∑
j=1
αG1jgij + αU2u2i + αU3u3i + αX2x2i + αX3x3i + ϵX1i (3)
x2i =5∑
j=1
αG2jgij + αU1u1i + αU3u3i + ϵX2i
x3i =5∑
j=1
αG3jgij + αU1u1i + αU2u2i + ϵX3i
yi = βU1u1i + βU2u2i + βU3u3i + β1x1i + β2x2i + β3x3i + ϵY i
gij ∼ Binomial(2, 0.3) independently for each j = 1, . . . , 5
u1i, u2i, u3i ∼ N (0, 1) independently
ϵX1i, ϵX2i, ϵX3i, ϵY i ∼ N (0, 1) independently
The coefficients αGkj for risk factor k and genetic variant j are taken as:
8
Genetic variant (j) αG1j αG2j αG3j
1 0.1 0.3 0.5
2 0.2 0.4 0.1
3 0.3 0.5 0.2
4 0.4 0.1 0.3
5 0.5 0.2 0.4
We set αU1, αU2, αU3 = 0.3 and take 9 values of the parameters αX2 and αX3 as
in the second set of simulations in the main paper (the section ‘Causal relationships
between risk factors’). All other parameters are taken as in the main paper. The
average F statistics for the associations of the genetic variants with each of risk factors
are around 1180.
Results are displayed in Web Table 3. We see that there is much less difference
between the estimated and true values of the causal effect parameters for the two-stage
least squares and likelihood-based methods with non-weak instrumental variables.
Two-stage least squares Likelihood-based
αX2 αX3 β1 β2 β3 β1 β2 β30 0 0.299 0.000 -0.099 0.299 0.000 -0.0990.5 0 0.300 0.000 -0.099 0.300 0.000 -0.100−0.5 0 0.300 0.000 -0.100 0.300 0.000 -0.1000 0.5 0.299 0.001 -0.100 0.299 0.001 -0.1000 −0.5 0.301 0.000 -0.102 0.301 0.000 -0.1020.5 0.5 0.301 -0.001 -0.100 0.301 -0.001 -0.100−0.5 0.5 0.301 0.000 -0.100 0.301 0.000 -0.1000.5 −0.5 0.300 0.000 -0.100 0.300 0.000 -0.100−0.5 −0.5 0.299 0.000 -0.100 0.299 0.000 -0.100
Web Table 3: Mean estimates of the causal effects of X1 (true direct effect β1 = 0.3),X2 (β2 = 0), and X3 (β3 = −0.1) using two-stage least squares and likelihood-basedmethods from simulation study of multivariable Mendelian randomization with causalrelationships between the risk factors and non-weak genetic instrumental variables
9
Allowing interaction terms in the genetic associations with the risk factors
To investigate the impact of interactions between the effects of genetic variants on
estimates from the two-stage least squares and likelihood-based methods, we extend
the above simulation (3) by introducing interaction terms in the genetic associations
with the risk factors into the data-generating model:
x1i =5∑
j=1
αG1jgij +∑j1>j2
αG1j1j2gij1gij2 + αU2u2i + αU3u3i + ϵX1i
x2i =5∑
j=1
αG2jgij +∑j1>j2
αG2j1j2gij1gij2 + αU1u1i + αU3u3i + ϵX2i
x3i =5∑
j=1
αG3jgij +∑j1>j2
αG3j1j2gij1gij2 + αU1u1i + αU2u2i + ϵX3i
Effects between the risk factors (αX2, αX3) are set to zero so that the risk factors are
causally independent. The interaction terms αG1j1j2 , αG2j1j2 , αG3j1j2 are drawn from
a normal distribution with mean αG× and variance σαG× . We take σαG× = 0.1 and
consider four different values of αG× = 0.1, 0.05,−0.05,−0.1. We analyze the data
using the two-stage least squares and likelihood-based methods as previously, not
including the interaction terms in the analysis model.
Results are displayed in Web Table 4. Estimates do not change substantially
despite large interaction terms (compared with the main effect terms) resulting in
misspecification of the analysis models. We see that the two-stage least squares and
likelihood-based methods are robust to interactions between the genetic variants in
their associations with the risk factors, even if these interactions are ignored in the
analysis model.
10
Two-stage least squares Likelihood-based
αG× β1 β2 β3 β1 β2 β30.1 0.300 0.000 -0.100 0.301 0.000 -0.1000.05 0.298 0.000 -0.098 0.299 0.000 -0.098−0.05 0.300 0.002 -0.097 0.300 0.002 -0.0970.1 0.302 0.002 -0.100 0.302 0.002 -0.101
Web Table 4: Mean estimates of the causal effects of X1 (true direct effect β1 = 0.3),X2 (β2 = 0), and X3 (β3 = −0.1) using two-stage least squares and likelihood-basedmethods from simulation study of multivariable Mendelian randomization with non-weak genetic instrumental variables and interactions in the data-generating model forthe genetic associations with the risk factors
11
Sequential adjustment method
We additionally consider a method for dealing with multiple risk factors referred to
by Holmes et al. [2] as a sequential adjustment method. This is a 2SLS instrumental
variable analysis for each of the risk factors in turn with adjustment for the other
risk factors in the regression models. We perform two versions of this method. In
the first version of the method, adjustment for the other risk factors is made only in
the second-stage regression model; in the second version, adjustment is made in both
regression stages. For example, in the second version of the method, if there are three
risk factors X1, X2 and X3, then the estimate for X1 is obtained by first regressing
X1 on the IV(s), X2 and X3, and then regressing Y on the fitted values of X1 from
the first-stage regression, X2 and X3. Both versions of the analysis are performed as
it is not clear which version of the method was chosen by the original investigators
in Holmes et al. [2]. To more closely mimic the method performed by the original
authors, we also performed the methods using a weighted allele score in place of the
individual genetic variants [3]. Weights were taken as the true coefficients for the risk
factor under analysis in the data-generating model.
We consider 8 sets of parameter values for αU1, αU2, αU3 = ±0.3 and set αX2 = 0
and αX3 = 0 as in the initial set of simulations in the main paper. All other parameters
are taken as in the main paper.
Results are given in Web Table 5 for the sequential adjustment methods using the
individual variants as instrumental variables, and in Web Table 6 for the sequential
adjustment methods using the allele scores as instrumental variables, with estimates
from the 2SLS method using all of the risk factors as in the main body of the paper
presented in each case for comparison. We see that, while the 2SLS method seems
to give unbiased estimates of the causal parameters, that the estimates from the
sequential adjustment method are biased, and not consistent in direction even under
12
the null.
Two-stage least squares Sequential adjustment (1) Sequential adjustment (2)αU1 αU2 αU3 β1 β2 β3 β1 β2 β3 β1 β2 β3
0.3 0.3 0.3 0.299 0.000 -0.099 0.070 -0.181 -0.290 0.094 -0.179 -0.3060.3 0.3 -0.3 0.300 0.000 -0.100 0.270 -0.090 -0.198 0.300 -0.089 -0.2130.3 -0.3 0.3 0.300 0.000 -0.100 0.179 0.000 -0.181 0.207 0.000 -0.1930.3 -0.3 -0.3 0.299 0.001 -0.100 0.379 0.091 -0.090 0.414 0.090 -0.100-0.3 0.3 0.3 0.301 0.000 -0.102 0.161 -0.091 -0.091 0.186 -0.090 -0.101-0.3 0.3 -0.3 0.301 -0.001 -0.100 0.362 0.000 0.002 0.393 0.000 -0.007-0.3 -0.3 0.3 0.301 -0.001 -0.100 0.270 0.091 0.019 0.300 0.089 0.014-0.3 -0.3 -0.3 0.300 0.000 -0.100 0.470 0.181 0.110 0.506 0.178 0.106
Web Table 5: Mean estimates of the causal effects of X1 (true effect β1 = 0.3), X2
(β2 = 0), and X3 (β3 = −0.1) using three different methods from simulation studyusing individual genetic variants as instrumental variables
Two-stage least squares Sequential adjustment (1) Sequential adjustment (2)αU1 αU2 αU3 β1 β2 β3 β1 β2 β3 β1 β2 β3
0.3 0.3 0.3 0.299 0.000 -0.099 -0.039 -0.346 -0.406 -0.046 -0.432 -0.5140.3 0.3 -0.3 0.300 0.000 -0.100 0.251 -0.199 -0.205 0.300 -0.248 -0.2590.3 -0.3 0.3 0.300 0.000 -0.100 0.139 0.000 -0.279 0.166 0.000 -0.3530.3 -0.3 -0.3 0.299 0.001 -0.100 0.429 0.147 -0.078 0.514 0.183 -0.099-0.3 0.3 0.3 0.301 0.000 -0.102 0.071 -0.148 -0.080 0.085 -0.184 -0.101-0.3 0.3 -0.3 0.301 -0.001 -0.100 0.363 0.001 0.121 0.435 0.001 0.153-0.3 -0.3 0.3 0.301 -0.001 -0.100 0.251 0.200 0.047 0.300 0.249 0.060-0.3 -0.3 -0.3 0.300 0.000 -0.100 0.541 0.346 0.247 0.647 0.431 0.313
Web Table 6: Mean estimates of the causal effects of X1 (true effect β1 = 0.3), X2
(β2 = 0), and X3 (β3 = −0.1) using three different methods from simulation studyusing allele scores constructed from the genetic variants as instrumental variables
13
Web Appendix 4
Consistency of association of genetic variants with lipid frac-
tions and CHD risk
In order to assess the consistency of the association of genetic variants with lipid
fractions and CHD risk, we constructed a lipid risk score for each variant and assessed
the association of this score with CHD risk. This score was calculated using the causal
estimates βLDL−C , βHDL−C and βTG from Web Table 2 and the genetic associations
XL, XH and XT from Web Table 1 for each genetic variant j = 1, . . . , 28 as:
βLDL−CXLj + βHDL−CXHj + βTGXTj (4)
The association between the lipid risk score and the log odds of CHD is plotted in Web
Figure 2. If the associations of each lipid fraction are homogeneous across variants, we
expect the points to lie on a straight line through the origin. With the exception of the
point on the far left representing genetic variant rs2304130, this is approximately true.
This variant shows an association with LDL-C and triglycerides, but no association
with CHD risk and a point estimate in the opposite direction from that which would
be expected based on its associations with LDL-C, HDL-C and triglycerides alone.
Further investigation is needed to show whether the result for variant rs2304130, is
a chance result or reflects a pleiotropic association of this variant with another risk
factor.
The causal odds ratios for CHD based on 27 variants excluding rs2304130 are 0.48
(95% credible interval 0.39–0.60) per 30% reduction in LDL-C, 1.19 (95% credible
interval 0.89–1.58) per 30% reduction in HDL-C, and 0.77 (95% credible interval
0.68–0.87) per 30% reduction in triglycerides.
14
Lipid risk score per allele
Log
odds
of C
HD
per
alle
le
−10% −5% 0 +5% +10% +15%
1.0
1.05
1.1
1.15
1.2
Web Figure 2: Association of coronary heart disease (CHD) risk-increasing allelesof 28 genetic variants with lipid risk score and odds of CHD (brightness and sizeof points: brighter points correspond to stronger associations with CHD risk, largerpoints correspond to more precise estimates). Note that some points are overlapping.
15
References
[1] Waterworth D, Ricketts S, Song K, Chen L, Zhao J, Ripatti S, et al. Genetic
variants influencing circulating lipid levels and risk of coronary artery disease.
Arteriosclerosis, Thrombosis, and Vascular Biology 2010;30(11):2264–2276.
[2] Holmes MV, Asselbergs FW, Palmer TM, Drenos F, Lanktree MB, Nelson CP,
et al. Mendelian randomization of blood lipids for coronary heart disease. Euro-
pean Heart Journal 2014;[Available online ahead of print, January 27, 2014] DOI:
10.1093/eurheartj/eht571.
[3] Burgess S, Thompson S. Use of allele scores as instrumental variables
for Mendelian randomization. International Journal of Epidemiology 2013;
42(4):1134–1144.
16