Two-way Causation in Life Satisfaction Research: Structural Equation Models with Granger-Causation

Two-way causation issues are the bete noire of life satisfaction research. As acknowledged in several landmark reviews, many variables routinely reported as causes or determinants of life satisfaction could equally well be consequences, or perhaps both causes and consequences (Diener, 1984; Diener, Suh, Lucas and Smith, 1999; Argyle, 2001; Frey and Stutzer, 2002). These variables include one's state of health, social support and participation, exercise, job satisfaction and satisfaction with one's partner and family life. In previous attempts to disentangle two-way causation issues, a wide variety of statistical models have been deployed. Conflicting empirical results have been reported, together with concerns about model 'goodness of fit' and model stability. In this paper we estimate five-wave structural equation models using data from large, nationally representative panel surveys in Australia, Britain and Germany. The models are based on a modified concept of Granger-causation (Granger, 1969). The main intuition behind Granger-causation is that if lagged versions of time-series variable x have statistically significant effects on time-series variable y in equations in which multiple lagged versions of y are also included, then it can be inferred that x is one cause (or 'Granger-cause') of y. We adapt Granger's approach, extending it to encompass two-way causation and panel survey data. It transpires that our Granger-style models have satisfactory fits to the panel data and are stable. Alternative models fit the data much less well. Substantively, we find that two-way causation is pervasive: all of the x variables mentioned above appear to be both causes and consequences of life satisfaction. With minor exceptions, results replicate across all three datasets, despite non-trivial differences between the measures used in the three countries. One implication is that researchers who have assumed one-way causation have seriously over-estimated the effects of x variables on LS. A second implication is that many people apparently experience multi-year gains or losses of life satisfaction rather than just recording short term fluctuations around their own normal level or 'set-point'.