@article{article_1813806, title={An Empirical Demonstration of Selecting Predictors for Multilevel Models}, journal={Journal of Measurement and Evaluation in Education and Psychology}, volume={17}, pages={42–62}, year={2026}, url={https://izlik.org/JA39HY39MA}, author={Mumba, Brian and Aydın, Burak}, keywords={predictor selection strategies, multilevel modelling, AIC, BIC, EGMA}, abstract={This paper presents a comparative demonstration of variable selection based on the Akaike Information Criterion (AIC), the Bayesian Information Criterion (BIC), and deviance for identifying the best-fitting two-level model using Early Grade Mathematics Assessment (EGMA) data collected in Zambia. A model that included all available predictor variables as fixed effects with random intercepts was run in R using the lme4 package, followed by an input model for comparison using a custom R function introduced by Nimon (2018). The analysis generated 108 models: 99 valid and nine invalid. The study determined a final model as the best fitting based on the principle of parsimony. The final model revealed that fixed effects of students’ group-mean-centered reading ability, home reading status, gender, school reading average score, and number of pupils, and the random effect of group-mean-centered reading ability predicted the early-grade mathematics ability. The results from the retained model, compared with the null model, showed substantial improvements in model fit indices with a pseudo-R² value of 0.309. Overall, this study provides an empirical demonstration of selecting predictors among numerous variables for multilevel models, a crucial practical issue that is common in educational research due to the increasing availability of large datasets.}, number={1}