The Effect of Cluster Size for Model Performance in High-Dimensional Longitudinal Studies: A Simulation Study

This is the authors' abstract. We don't add key points for this paper.

Objective: In order to prevent model estimation errors and deviations in high-dimensional longitudinal studies, risk models are established through penalized methods. The aim of this study is to examine the effect of small cluster effects on the generalized estimating equations (GEE) and penalized GEE (PGEE) model performances in high-dimensional longitudinal data. Material and Methods: A simulation study was designed to compare the GEE and PGEE model performances, Type I error rates, and power in two-period longitudinal data structures with different cluster sizes (n=20, 30, 50, 100, 200), different numbers of predictors (p=10, 20, 50) and different correlation levels between predictors (r=0.20, 0.50, 0.80). Results: It was observed that the GEE coefficient estimates were misleading and inconsistent, the Type I error rates were high, and the power of the test was weak at insufficient cluster sizes and high correlations between predictors. Even when the number of predictors and cluster size were in the balance (p=10, n=100, 200), Type I error rates were obtanied high for GEE. Increasing the cluster size was not enough to reduce the Type I error rate of GEE. The PGEE produced more successful results than GEE in all conditions. The power of PGEE increased to over 80% in all scenarios. Conclusion: The PGEE yielded more consistent results by controlling the relationships both within the cluster and between the predictors. In highdimensional longitudinal studies, it was observed that the use of PGEE is more effective than GEE.