Analyzing longitudinal data using machine learning with mixed-effects models

Abstract

The study evaluates the performances of two mixed-effect machine learning methods (RE-EM trees and MERF) and mixed effects penalized methods, three classical machine learning methods (random forest, gradient boosting trees, support vector machine), penalized and shrinkage models (Ridge, LASSO, ALASSO, SCAD, shrinkage) in the prediction of high dimensional longitudinal data. It aims to compare mixed effect models and traditional models’ prediction accuracy using correlated data. Two real data applications are used aiming to predict the COVID-19 number of cases and pandemic-related number of deaths. The models are evaluated according to the RMSE values of the predictors.