Lifestyle Risk Score: handling missingness of individual lifestyle components in meta-analysis of gene-by-lifestyle interactions

Recent studies consider lifestyle risk score (LRS), an aggregation of multiple lifestyle exposures, in identifying association of gene-lifestyle interaction with disease traits. However, not all cohorts have data on all lifestyle factors, leading to increased heterogeneity in the environmental exposure in collaborative meta-analyses. We compared and evaluated four approaches (Naïve, Safe, Complete and Moderator Approaches) to handle the missingness in LRS-stratified meta-analyses under various scenarios. Compared to "benchmark" results with all lifestyle factors available for all cohorts, the Complete Approach, which included only cohorts with all lifestyle components, was underpowered due to lower sample size, and the Naïve Approach, which utilized all available data and ignored the missingness, was slightly inflated. The Safe Approach, which used all data in LRS-exposed group and only included cohorts with all lifestyle factors available in the LRS-unexposed group, and the Moderator Approach, which handled missingness via moderator meta-regression, were both slightly conservative and yielded almost identical p values. We also evaluated the performance of the Safe Approach under different scenarios. We observed that the larger the proportion of cohorts without missingness included, the more accurate the results compared to "benchmark" results. In conclusion, we generally recommend the Safe Approach, a straightforward and non-inflated approach, to handle heterogeneity among cohorts in the LRS based genome-wide interaction meta-analyses.