Life Satisfaction and Age: Getting Control Variables under Control

Researchers investigating the relationship between age and life satisfaction have produced conflicting answers, mainly via disputes over whether to include individual-level control variables in regression models.  Most well-being/happiness scholars believe there is a “U-shaped” relationship, with life satisfaction falling towards middle age and then subsequently rising.  This position emerges mainly in research that uses control variables for e.g. income and marital status.

This approach is incorrect: variables such as these would intervene in the relationship between age and life-satisfaction and should therefore not be controlled.  Regression models should control only for “confounding” variables, i.e., variables that are causally prior to the dependent variable and the core independent variable of interest.  Other individual-level variables cannot determine one’s age; they are not confounders and should not be controlled.  The only potential confounders in this context are cohort and period (and country, if the data are from a multi-country study). 

This paper demonstrates these points via “directed acyclic graphs” (DAGs) and applies them to an analysis of data from the World Values Survey.  A key finding is that there is only a negligible post-middle-age rise in life satisfaction – and the important implication is that there cannot then be a U-shaped relationship between age and life satisfaction.