Variable selection methods for descriptive modeling

Variable selection methods are widely used in observational studies. While many penalty-based statistical methods introduced in recent decades have primarily focused on prediction, classical statistical methods remain the standard approach in applied research and education. In this study, we evaluated the variable selection performance of several widely used classical and modern methods for descriptive modeling, using both simulated and real data. A novel aspect of our research is the incorporation of a statistical approach inspired by the supersaturated design-based factor screening method in an observational setting. The methods were evaluated based on Type I and Type II error rates, the average number of predictors selected, variable inclusion frequency, absolute bias, and root mean square error. The detailed results of these evaluations are presented, and the methods' performance is discussed across various simulation scenarios and in application to real data.