Impact of Model-Form Selection on the Accuracy of Rate Estimation

A key assumption underlying the use of model-based estimates in epidemiology is that the structural-model form is an adequate mathematical description of the dependence of disease occurrence on exposures and covariates (that is, the model form is correctly specified). If this assumption is violated, model-based point estimators and variance estimators may be biased, standard confidence intervals may be invalid, and inferences derived from these estimators may be incorrect. In practice, the true structural-model form is usually unknown, and investigators frequently use their data to help select a model form. We conducted a simulation study to examine the impact of model-form selection on the accuracy of rate estimation in cohort-study situations resembling those found in environmental and occupational epidemiology. For the situations we examined, the increase in variance produced by using model-form selection was often more than offset by the corresponding reduction in bias, sometimes resulting in a dramatic increase in accuracy. Model-form selection was observed to be most beneficial relative to no selection when effects were stronger, the sample size was larger, and the candidate model forms included the true model form or allowed the model to more closely approximate the true model form. It was least beneficial when effects were weak and the sample size was small, even if the candidate model forms included the true model form.