STATGEN 2024
Pleiotropy-robust methods for high-dimensional multivariable Mendelian randomization (HDMR)
Nathan LaPierre presenting, co-authors: Matthew Stephens, Xin He
In HDMR, we have many genetically correlated exposures, which may be explained by unobserved shared factors. These can be inferred by factor analysis.
Flexible, modular framework: Factor-Augmented MR
1. Factor Analysis
2. Regression/Variable Selection
#Genetics #StatisticalGenetics #MendelianRandomization #STATGEN2024
Variable selection challenges
1. Highly-correlated exposures
2. Instrument selection
Susie-FAMR
Use SuSiE-RSS for Bayesian variable selection of exposures -> sets of highly correlated variables.
A flexible two-stage framework.
Question: How do you distinguish between confounding and mediating?
Answer: Sometimes you get a factor that is very similar to an exposure, so we filter those out. Usually the factors don't take away too much of the signal.
Question: Correlations among exposures?
Answer: This is handled by the SuSiE model, which takes into account correlation among the exposures. SuSiE fits causal effects one-by-one - fits one and models the rest as being correlated with that exposure. Finds a set of highly correlated exposures within which one is causal.
Daniel E. Weeks