Promoting Measured Genes and Measured Environments: On the Importance of Careful Statistical Analyses and Biological Relevance—Reply

In reply

We appreciate the opportunity to respond to the letter by Drs Joober, Sengupta, and Schmitz. We were attempting to replicate a study conducted by Caspi et al¹ and we therefore sought to replicate the methods and measures they used insofar as we could. We used logistic regression to model risk for conduct disorder because Caspi et al used logistic regression to model risk for conduct disorder. Logistic regression also has the advantage of being relatively robust to scaling artifacts (heteroscedasticity), which can be an issue when modeling interactions. We created a simple count variable of each measured adversity to obtain a quantitative measure of severity of exposure to family adversity. In practice, the architecture of familial adversity is likely to be complex and we would welcome efforts to address this issue in larger samples with greater power. The issue of statistical power is not a trivial one and our sample size was relatively modest (n = 514 male twins). We therefore published our data in a transparent way to permit evaluation and reanalysis. The reported interaction was not estimated on the basis of a single data point and 1-sided significance tests should be conducted when there is a clear directional hypothesis based on prior work. Ultimately, functional studies will be required to validate observed statistical associations, including statistical interactions, consistent with the presence of genotype × environment interactions. Until then, meta-analysis of published studies provides one way of attempting to summarize the strength of statistical evidence across studies of variable size.²

Correspondence: Dr Foley, University of Melbourne, ORYGEN Research Centre, Applied Genetics Unit, 35 Poplar Rd, Parkville, Victoria 3052, Australia (dfoley@unimelb.edu.au).

Financial Disclosure: None reported.