Generalized Linear Mixed Models by penalized Lasso in modelling the scores of Indonesian students

Abstract

The Generalized linear mixed model (GLMM) is an extension of the generalized linear model by adding random effects to linear predictors to accommodate clustered or over dispersion. Severe computational problems in the GLMM modelling cause its use restricted for only a few predictors. When many predictors are available, the estimators become very unstable. Therefore, the procedure for selecting relevant variables is essential in modelling. The use of penalty techniques for selecting variables in mixed models is still rarely applied. In this article, the penalized Lasso approach proposed to handle these kinds of problems. The proposed methods select variables and estimate coefficients simultaneously in GLMM. Based on the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and standard error criteria, it was found that glmmLasso has a better performance than GLMM. For the factors affecting Indonesian’s student scores, where glmmLasso produces three significant covariates for the GLMM model while GLMM without penalized Lasso has five covariates, which means that the GLMM model is more complicated than glmmLasso. Gender, school quality based on National Examination (UN) scores and the opportunity for students to investigate to test their ideas are essential covariates as factors that influence the rating of Indonesian students.