Bias in the study of prediction of change: A Monte Carlo simulation study of the effects of selective attrition and inappropriate modeling of regression toward the mean

Abstract

Background: Medical researchers often use longitudinal observational studies to examine how risk factors predict change in health over time. Selective attrition and inappropriate modeling of regression toward the mean (RTM) are two potential sources of bias in such studies. Method: The current study used Monte Carlo simulations to examine bias related to selective attrition and inappropriate modeling of RTM in the study of prediction of change. This was done for multiple regression (MR) and change score analysis. Results: MR provided biased results when attrition was dependent on follow-up and baseline variables to quite substantial degrees, while results from change score analysis were biased when attrition was more strongly dependent on variables at one time point than the other. A positive association between the predictor and change in the health variable was underestimated in MR and overestimated in change score analysis due to selective attrition. Inappropriate modeling of RTM, on the other hand, lead to overestimation of this association in MR and underestimation in change score analysis. Hence, selective attrition and inappropriate modeling of RTM biased the results in opposite directions. Conclusion: MR and change score analysis are both quite robust against selective attrition. The interplay between selective attrition and inappropriate modeling of RTM emphasizes that it is not an easy task to assess the degree to which obtained results from empirical studies are over- versus underestimated due to attrition or RTM. Researchers should therefore use modern techniques for handling missing data and be careful to model RTM appropriately.