How much can we learn from happiness data?

Survey data on happiness, mental health, job satisfaction, and wellbeing are now widely used in economic research. They are also increasingly collected by government statistical agencies and bodies like the OECD. However, human feelings are measured on ordinal rather than cardinal scales. Hence, it has always been known that such data have to be handled with care. Some recent work has argued that, at least in principle, certain results in the research literature might be capable of being ‘reversed’ – turning estimated positive effects into negatives, and vice versa. How important are such theoretical possibilities? We show that self-reported wellbeing data can in most relevant circumstances be used safely. First, using several large-scale datasets from the US, Germany, and the Netherlands, we find that, in empirical practice, effect reversals are rare or even impossible for a number of core socio-economic variables, including people’s incomes and employment status. Second, we demonstrate that respondents would have to answer (numerical) happiness questions in a strongly non-linear fashion for reversals to actually appear. Yet, the evidence suggests approximately linear response behaviour. Third, as a more methodological contribution, we derive a simple and general non-reversal condition for regressions of ordinal data, and we derive bounds for ratios of coefficients. We finish with a set of suggested recommendations for appropriate practice in empirical research and sketch avenues for future research.