What Are the Most Important Predictors of Subjective Well-Being? Insights From Machine Learning and Linear Regression Approaches on the MIDUS Datasets
What are the most important predictors of subjective well-being? Using a nationally representative publicly available dataset from the Midlife in the United States project (N = 4,378), we applied linear regression, which often relies on assumptions of linearity and a priori interactions, and advanced machine learning approaches, which maximize prediction by thoroughly exploring nonlinear effects and higher-order interactions, to determine the ordering and characteristics of predictors of well-being. Advanced machine learning models generally did not predict well-being more accurately than did regression models, suggesting that many predictors of well-being may be linear and non-interactive. Consistent with this implication, the introduction of product and squared terms in regression models improved prediction, but only nominally. Our findings replicated previous research, with sociability, physical health, disengagement from goals, sex life quality, wealth, and religious activity emerging as the strongest predictors of well-being, and demographic factors emerging as relatively weak predictors. Furthermore, self-reported “aches” (the strongest “objective” predictor of well-being), stress reactivity, and disengagement negatively predicted well-being, reinforcing the role of stress in psychological maladjustment. Finally, unlike prior research, control over one’s life—and control over financial and work matters in particular—strongly predicted well-being.