Identifying moderators in latent growth curve models with SEM Boruta

Scientific studies in social and behavioral sciences often explain heterogeneity through the use of moderators. However, the effectiveness of this approach depends on how systematically moderators are identified and addressed and to what degree the default assumption of linearity holds. Challenges such as overfitting and misspecification persist which often results in an increased risk of Type I and Type II errors. This highlights the necessity for using robust and reproducible feature selection techniques with an objective criterion for selection. One such data-driven approach to mitigate heterogeneity is feature selection, often performed using non-linear and non-parametric models like random forests or decision trees. Boruta expands on random-forest-based feature selection to enable the computation of unbiased p-values. Structural Equation Models with Boruta Feature Selection (SEM Boruta) examines a set of predictors and quantifies their value as moderators for an underlying Structural Equation Model. Brandmaier et al. (2013, 2016) introduced Structural Equation Model Trees which combine the flexibility of random forests with the rigor of structural equation modeling to assess variable importance. Building upon this framework, SEM Boruta was developed as a method to identify key variables or moderators by evaluating their rank importance within the model. The robustness of the SEM Boruta algorithm is demonstrated by its ability to identify moderators between groups in a latent growth curve model, as applied to simulated datasets.