Power and type I errors of meta-analysis of gene-level tests in the presence of missing data.

Datasets were simulated according to the genetic and phenotype model described in METHODS. Within the gene region, 20% of the variant sites are deemed causal. Meta-analysis was performed to combine 10 cohorts with 2000 individuals each. For each replicate, summary association statistics were generated, and a certain fraction (10%, 30% or 50%) of the generated summary statistics were masked as missing. Scenarios with different combinations of known variant effect, candidate variant effects and fractions of missingness were considered. To evaluate the power loss due to missing data, we also analyzed the full dataset as a gold standard. Type I errors and power were evaluated for three rare variant tests (simple burden, SKAT and VT) using 1 million replicates under the significance threshold of α = 0.005.