Chunk #50 — Reasons to be Concerned about the Published cGxE Literature — Problems with the Recipe: Statistical Concerns in cGxE Research — Power to detect and characterize different types of interactions
Moreover, even “true” ordinal interactions that are statistically significant can appear to be of a cross-over type (Sher & Steinley, 2013) due to random error. All linear interactions imply a cross over at some point, even if outside the range of observed values. This is not a trivial issue since a typical practice is to plot values +1 standard deviation (SD) and −1 SD above and below the mean of the moderator (Aiken & West, 1991). However, −1 SD can represent values that rarely or never exist in nature for skewed predictors. We note that some authors (Roisman et al., 2012) have recently recommended extending the Aiken & West (1991) guidelines to + 2 SDs, in order to provide 95% coverage of the observed values. Given the highly skewed nature of many environmental exposures, attention to the underlying distribution of all study constructs is necessary so as not to generate misleading regression plots covering regions of sparse or imaginary data.
