Measuring Socioeconomic Status

6. How SES Should Be Measured

I insist that regression adjustment for SES merits careful consideration. Too few appreciate that regression “adjustment” is a form of imputation where the outcome measure being analyzed is altered by the machinery of the model being employed – often linear interpolation. In other words, adjustment for SES may give an analyst comfort in that she enhanced the exchangeability of confounding risk factors. But such comfort comes with the cost of assuming a correct model, including full control of the confounding influence of SES. The trouble is that our models are hardly ever correct: viz, assuming that a poor man has the same access to resources as a rich man once we statistical equate their income is simply absurd. Further, all too often researchers “adjust for SES” across racial groups and conclude that the remaining effect of race is genetically induced (Kaufman, Cooper and McGee 1997).  I refer the reader to the great insights of the late William Cochran (Cochran 1957; Cochran 1963; Cochran 1968) who I dare to summarize as saying that one should use regression adjustment only when it is not needed because when it is needed there is great potential to end up comparing apples to oranges and supporting inferences not with real data but on model-induced “facts.” Over the past few years I have advanced these ideas into the definition of structural confounding,  a term I coined to convey the problem of confounding that cannot be overcome by regression adjustment without heroic modeling assumptions (Oakes, Messer and Mason 2010).

<a href="/longdesc/Oakes_Sec06_Exer4.html">Flash Description</a>

Flash is not available on mobile devices. Please view the Flash Description.

Cochran, W. G. (1957). Analysis of Covariance: Its Nature and Use. Biometrics, 13, 261-280.
Oakes, J. M., Messer, L. C., & Mason, S. (2010). Messer et al. Respond to “Positivity in Practice”. American Journal of Epidemiology.