Genetic Instrumental Variable (GIV) regression: Explaining socioeconomic and health outcomes in non-experimental data

Estimating causal effects with non-experimental data is of central importance across multiple fields of scientific inquiry. Here, we propose genetic instrumental variables (GIV) regression for this purpose. GIV regression... [ view full abstract ]

Estimating causal effects with non-experimental data is of central importance across multiple fields of scientific inquiry. Here, we propose genetic instrumental variables (GIV) regression for this purpose. GIV regression utilizes large-scale genome-wide association study (GWAS) results that now allow constructing predictive polygenic scores (PGS) for many human traits. Our approach is based on the idea that adding the true PGS for the outcome to a regression model would effectively eliminate bias arising from a genetic correlation between the outcome and an exposure of interest. Without the true PGS in the structural model, using genes as instrumental variables (IVs) as proposed in Mendelian Randomization (MR) is problematic due to pleiotropic effects of genes that invalidate IV regression. However, PGS capture only a small fraction of the heritability of most traits, partly because GWAS are estimated in finite sample sizes that yield noisy estimates of the effects of each genetic marker. We argue and empirically demonstrate that it is possible to correct attenuation bias by splitting the GWAS sample to obtain several PGS (i.e. multiple indicators) in the prediction sample that can be used as instruments for each other in IV regression. Then we extend the approach to the problem of estimating causal effects and gene-environment interactions with non-experimental data. Our approach produces reasonable estimates of the chip heritability of educational attainment (EA) and, unlike the results using MR, GIV-based estimates find that the positive correlation between body height and EA is primarily due to genetic effects.