Using Machine Learning to Explore the Efficacy of Administrative Variables in Prediction of Subjective-Wellbeing Outcomes in New Zealand.

Abstract

The growing acknowledgment of population wellbeing as a key indicator of societal prosperity has propelled governments worldwide to devise policies aimed at improving their citizens' overall wellbeing. In New Zealand, the General Social Survey provides wellbeing metrics for a representative subset of the population (~ 10,000 individuals). However, this sample size only provides a surface-level understanding of the country's wellbeing landscape, limiting our ability to comprehensively assess the impacts of governmental policies, particularly on smaller subgroups who may be of high policy interest. To overcome this challenge, comprehensive population-level wellbeing data is imperative. Leveraging New Zealand's Integrated Data Infrastructure, this study developed and validated the efficacy of three predictive models-Stepwise Linear Regression, Elastic Net Regression, and Random Forest-for predicting subjective wellbeing outcomes (life satisfaction, life worthwhileness, family wellbeing, and mental wellbeing) using census-level administrative variables as predictors. Our results demonstrated the Random Forest model's effectiveness in predicting subjective wellbeing, reflected in low RMSE values (~ 1.5). Nonetheless, the models exhibited low R2 values, suggesting limited explanatory capacity for the nuanced variability in outcome variables. While achieving reasonable predictive accuracy, our findings underscore the necessity for further model refinements to enhance the prediction of subjective wellbeing outcomes.