Cross-Sectional Model-Building for Research on Subjective Well-Being: Gaining Clarity on Control Variables

Happiness/well-being researchers who use quantitative analysis often do not give persuasive reasons why particular variables should be included as controls in their cross-sectional models. One commonly sees notions of a “standard set” of controls, or the “usual suspects”, etc. These notions are not coherent and can lead to results that are significantly biased with respect to a genuine causal relationship. This article presents some core principles for making more effective decisions of that sort. The contribution is to introduce a framework (the “causal revolution”, e.g. Pearl and Mackenzie 2018) unfamiliar to many social scientists (though well established in epidemiology) and to show how it can be put into practice for empirical analysis of causal questions. In simplified form, the core principles are: control for confounding variables, and do not control for intervening variables or colliders. A more comprehensive approach uses directed acyclic graphs (DAGs) to discern models that meet a minimum/efficient criterion for identification of causal effects. The article demonstrates this mode of analysis via a stylized investigation of the effect of unemployment on happiness. Most researchers would include other determinants of happiness as controls for this purpose. One such determinant is income—but income is an intervening variable in the path from unemployment to happiness, and including it leads to substantial bias. Other commonly-used variables are simply unnecessary, e.g. religiosity and sex. From this perspective, identifying the effect of unemployment on happiness requires controlling only for age and education; a small (parsimonious) model is evidently preferable to a more complex one in this instance.