Evaluation of analyses of univariate discrete twin data

Akiake's Information Criterion (AIC) is commonly used in univariate twin modeling of a discrete trait to prune a full model into a more parsimonious submodel. It is possible that this practice could introduce bias and inaccuracy, and we could identify no prior systematic study of these issues. Thus, we used simulation to investigate the performance of AIC-guided modeling across a broad range of parameters. Our simulations indicated that the use of the AIC to determine the "best" univariate model for a discrete trait tended to yield the incorrect model rather frequently. Moreover the parameter estimates of the "best" model by AIC were biased sharply upward as were the associated 95% confidence intervals. These results suggest that the use of AIC to guide twin modeling for univariate discrete traits should either be abandoned or used with great caution.