An Empirical Evidence of Higher Model Efficiency of Machine Learning Model than A Classical Multiple Linear Regression Model

Abstract

Psychological health of young adults has always been an area of concern for the stakeholders, partially due to the fact that the psychological problems might remain unnoticed unless sever enough. One of the ways of extracting the information about the general psychological health is through questionnaires administered on groups of individuals. Choice of an appropriate statistical technique in analysing this information is as important as the selection of the questionnaire. If more than one technique fit the situation then a natural course of action is to compare the two in terms of some standard parameters. In the present study, data was collected on students enrolled in various colleges affiliated to University of Delhi, India. Information was sought on their psychological wellness and some stresses viz. the academic stress, social stress and financial stress, along with some demographic information such as the year of course and gender.  However, no significant effect of gender and the year of course was found. A classical multiple linear regression (MLR) model and machine learning multiple linear regression (ML-MLR) models with varying train data sizes were applied to the data to estimate/predict wellness scores on the basis of the independent predictors. The three stress factors were all significant. The R2 of the MLR model was 0.5207 and the model was significant with p-value of the F-statistic < 0.001. Both the techniques produced models with estimating/predictive efficiency at least 82%. However, the ML-MLR models had the predictive efficiency at least equal to or more than the estimating efficiency of the classical MLR model. The best results were obtained for train data size 70%, a prevalent size selection in machine learning models. The study reinforced the strength of machine learning models over the classical methods.