Lasso Meta-Regression and Its Application to Psychiatric Genomics

Meta-analysis is an essential technique in psychiatric genomics. Heterogeneity of the effect sizes included in a meta-analysis can reduce its statistical power. Meta-regression provides a way to identify study-level covariates that are associated with heterogeneity. However, in psychiatric genomics, and other new fields, there is little theory to guide the choice of study-level covariates to be included in a meta-regression. We propose the use of shrinkage regression methods, such as the Lasso and the SCAD, as means to select study-level covariates automatically. We used simulated data to compare these methods to simple regression, a current standard approach. The criterion used in the comparison was the precision of the method in selecting covariates. We found that precision was highly dependent on the number of study level covariates included, the number of studies included, and the relative sizes of the included studies. When many studies were included, a bootstrapped version of the Lasso was able to select covariates with adequate precision. In moderate-sized meta-analyses, a bootstrapped version of simple regression performed best. The comparison of methods presented in this work is applicable to the design of new consortia of studies in psychiatric genomics, as well as to the evaluation and comparison of existing consortia.