Why statistical adjustment for social and economic background fails and what to do about it

Abstract

Scholars often hope to filter class, clout, poverty, deprivation, inequity and lack of opportunity from their analyses: factors that affect both how we act and what we achieve, and thereby muddle other associations in the data. In practice, adjustment relies on self- reported low-fidelity proxies such as occupation and education that capture but a fraction of the variance due to social and economic background and inflate the estimated effects of any correlated variable. When background is measured in a way that matches the research at hand, included in the analysis multidimensionally and not as a composite, and corrected for measurement error, it is not rare to see regression coefficients cut in half. We illustrate how poor measurement and poor models lead to residual confounding by social and economic background, review reliable alternatives to generic measures of socioeconomic status and demonstrate the effectiveness of statistical corrections for any error that remains.