Using Canadian and Korean data, I assess the degree to which standard estimates of the importance of income and education for promoting life satisfaction may be flawed due to response biases in respondents.
Background:
Discrete response scales used ubiquitously for the reporting of life satisfaction pose cognitive challenges to survey respondents, so differing cognitive abilities result in different uses of the scale, and thus potential bias in statistical inference. An enormous “happiness”, or life satisfaction, literature in economics makes use of Likert-like scales in assessing survey respondents' cognitive evaluations of their lives, overall. These measures are being used to estimate economic benefits in every empirical field of economics.
Typically, analysis of these data have shown remarkably low direct returns of education for improving subjective well-being. In addition, arguably, the inferred impact of material wealth and income using this method is also unexpectedly low as compared with other, social factors, and as compared with economists' prior expectations which underlie, in some sense, support for using GDP as a proxy for more general quality of life goals.
Methods:
“Life satisfaction” data are now most standardly assessed on an eleven-point (zero to ten) scale, in accordance with recommendations by the OECD, US National Academies, and others. Typically, reduced form models of these subjective but quantitative data treat the integer responses as continuous variables and often even as interpersonally comparable, although using ordinal models with weaker assumptions makes little difference in practice. However, an often-overlooked feature of the distribution of responses to life satisfaction questions is that they exhibit certain enhancements at focal values, in particular at 0, 5, and 10 on the eleven-point scale. In this paper, I investigate the reasons for, and implications of, these response patterns.
Data come from the Canadian Community Health Survey and the Korean Labour and Income Panel Survey, both of which include a life satisfaction question along with other demographic, economic, and social variables.
Results:
I offer three quantitative analytic contributions to evaluate the bias in standard applications of life satisfaction data. These include a model to account for the focal-value behavior using a latent variable approach to capture the "internal" cognitive evaluation before it is translated to the discrete scale of a survey question. In this way I quantify the bias which the focal-value behavior introduces, showing that, as a result, it is possible for reported average life satisfaction scores to decrease with increasing education.