Bayesian methods for instrumental variable analysis with
genetic instruments ("Mendelian randomization")
Paul McKeigue, Felix Agakov, Dominik Glodzik
The
"Mendelian randomization" approach exploits genotype as an
"instrumental variable" that perturbs a variable of interest (usually a
phenotypic biomarker) to learn about the effects of this intermediate
variable on outcome. We are developing this general approach in
two ways. The first approach develops Bayesian methods for
problems in which the standard instrumental variable argument can be
used, because we can assume that the genotypes do not directly affect
the outcome except through their effects on the biomarkers. In
the context of genetic instruments, this is denoted the "no pleiotropy"
assumption. This assumption restricts applications to studies of
a few biomarkers and genes whose effects are relatively well
understood.
The second more general approach relaxes the no
pleiotropy assumption, and can be applied to genome-wide datasets with
multiple phenotypic biomarkers
Slide presentation based on two-day course given at Erasmus Medical Centre, May 2010
Bayesian methods for instrumental variable analysis of genetic effects with no pleiotropy
We have developed a Bayesian approach to inference for this type of problem, implemented using the JAGS
package. Some advantages of this approach are that it can be
applied just as easily with binary outcomes as with continuous
variables, it does not rely on asymptotic approximations which do
not work well with weak instruments, and it allows formal hypothesis
tests to be constructed based on comparing the (marginal) likelihoods
of causal and non-causal explanations, given the data.
Zipped tutorial with example dataset to be run using JAGS
This example is based on the dataset described in McKeigue et al 2010
Sparse
Bayesian instrumental variable analysis for inference of causal effects
with genome-wide genotype data and multiple phenotypic biomarkers
To
exploit genotype-biomarker associations more generally for instrumental
variable analysis, we have to relax the assumption of no pleiotropy.
In principle this is possible, if we have multiple genetic
instruments.
With more than one instrumental variable, causal
explanations of a biomarker-outcome association can be distinguished
from explanations based on confounding and pleiotropic genetic effects.
We are developing an approach known as sparse Bayesian
instrumental variable analysis as a general framework for inference in
datasets with genotypes, phenotypic biomarkers and continuous or binary
outcome variables. This can be applied to genome-wide datasets
with many biomarkers and multiple outcome variables. One
application is to inference of causal relationships between biomarkers
and outcomes. Where the biomarkers are gene transcript levels,
the method can also be used for fine mapping of genotypic effects on
the outcome
This approach is implemented in the SPIV
package for sparse Bayesian instrumental variable analysis.
Tutorials and downloads are in preparation (May 2010).