Evaluating Household Consumption Patterns: Comparative Analysis Using Ordinary Least Squares and Random Forest Regression Models

Abstract

This research aims to decompose the contribution of socioeconomic factors towards household consumption expenditure using a regression approach, with log per capita expenditure as the dependent variable. Our study stands out as the first to utilise SHAP analysis and Machine Learning models to analyse household consumption expenditure. We select both OLS (linear) and Random Forest (nonlinear) models to compare how they estimate consumption expenditure differently. Both models explain about 85% of the variation in log per capita expenditure. The SHAP analysis reveals the nonlinear relationships inside the Random Forest model. Several insightful findings were suggested that can be integrated into current policy-making. The results are as follows: (1) Both models agree that income, household size, and educational level are major factors in the purchasing power of household heads. (2) The Random Forest model demonstrated a nonlinear contribution of age and household size towards log per capita expenditure, contrasting with previous studies that treated them as linear. (3) Household heads with a higher income and educational level tend to spend more. (4) Current policy should consider focusing on households with larger sizes and lower incomes, who tend to spend more despite earning less, primarily by assisting them with non-cash transfers and subsidies. Doi: 10.28991/HIJ-2024-05-02-019 Full Text: PDF

Keywords

• household consumption
• machine learning
• linear regression
• random forest
• shapley value.