Critical Review of Gene-Environment Interactions in Psychiatry: Evidence Consistent with the Possibility of Publication Bias, Low Power, Multiple Testing, and Type I Errors in the Literature

摘要

Introduction: Measured gene-by-environment interaction studies have typically been conducted in a candidate GxE (cGxE) fashion, analogous to the candidate gene association studies that were used to search for genetic main effects. Such cGxE research in psychiatry has received widespread attention and acclaim, yet cGxE findings are also controversial. We were interested in determining whether cGxE findings were robust and might help to explain some of the missing heritability in psychiatric genetics or if, in aggregate, cGxE findings were consistent with the existence of publication bias, low power, multiple testing, and type I errors in the cGxE literature in psychiatry. Method: We applied modified meta-analytic procedures to all published studies (to our knowledge) from the first decade of cGxE research in psychiatry, 2000-2009, collapsing across reported interactions in order to identify prevailing trends. Results: Most novel cGxE studies were significant (96%), but only a minority of replication attempts were significant (32%). These findings are consistent with the existence of publication bias among novel GxE studies. There may also be publication bias among replication attempts because significant replication attempts had smaller sample sizes, on average, than null replication attempts. Furthermore, rates of positive replications, observed sample sizes, and power calculations suggested that studies were underpowered. Additionally, patterns of expanding and branching hypotheses have been reported across time, and could be partially due to multiple testing and publication bias. Finally, through simulations we show that low power biases the observed form of interactions (i.e., `crossoveru27 versus `non-crossoveru27). Conclusion: These results are consistent with the hypothesis that published studies provide a biased representation of all cGxE tests that have been conducted and also suggest that many reported positive findings may be type I errors.