IJE Advance Access originally published online on April 28, 2005
International Journal of Epidemiology 2005 34(4):835-836; doi:10.1093/ije/dyi094
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Commentary |
Commentary: Advancing research into SES mechansisms that affect health
Division of Epidemiology and Community Health, University of Minnesota, Minneapolis, MN, USA. E-mail: oakes{at}epi.umn.edu
Readers should appreciate the importance of Smith and Frank's effort to unpack the micro-level mechanisms that relate socioeconomic status (SES) to health, and link the same to macroeconomic trends.1 Such efforts move us beyond the mere documentation of the SES–health gradient. Much to their credit, Smith and Frank (hereinafter S&F) consider a plausible mechanism, exploit high-quality data, and employ proper analytical techniques. Should it withstand scrutiny, S&F's work holds enormous implications for social epidemiology and public policy more generally, especially since it complicates the 40-year-old ‘returns to schooling’ literature, which has heretofore presumed health monotonically increases with schooling.2
 |
Shortcomings |
|---|
 Top Shortcomings Steps for the futureConclusionsReferences |
|---|
One might carp about the validity of self-reported health proxies or a short (4 year) observation interval. But in an effort to encourage advancement, I discuss three more nuanced shortcomings in S&F's paper: (i) underdeveloped causal reasoning; (ii) trust in exposure measure; and (iii) fixation on theory of psychological distress.
Underdeveloped causal reasoning
S&F do not say much about causality. Indeed, they only mention the word once, during their critique of prior research. The paper would be stronger if they addressed the issue directly.
Consider the hypothetical experiment S&F aim to mimic: simultaneously observe persons in occupations both concordant and discordant with their individual educational attainment. Wait four years and measure health. The mean between-exposure difference is the (technically, ‘an’) effect of SES discordance. Implications follow.
First, since we cannot observe one of the conditions—the counterfactual—we must seek reasonable substitutes. Second, unless we disregard basic ethical principles, we will never be able to conduct a randomized experiment, which yield the best substitutes by balancing confounders, measured and unmeasured, between conditions. We are left to infer effects from observational designs, and so must rely upon strong theory and worry about exchangeability. The key question is whether or not those ‘assigned’ to discordant occupations are sufficient counterfactual substitutes for those assigned to concordant occupations? Were the ‘overqualified’ really so?
I doubt it, despite S&F's valiant efforts to statistically adjust for constructs such as self-esteem, physical activity, and so on. What they needed to have controlled for is ‘aspirations’ or ‘ability’ which is a complex construct.3 Just because someone is trained for a given occupation does not mean they aspire to it.4 Accordingly, I worry that S&F's analysis suffers omitted variable bias, which may wreak havoc on point estimates.5
Trust in exposure measure
Despite S&F's reasonable comments regarding the validity of their exposure measure, I remain worried that misclassification may explain key findings.
With respect to educational attainment, S&F implicitly assume homogeneous school ‘quality’ but this is surely analytical fiction. For example, everyone knows the returns to a University of Toronto education exceed those from other universities—perhaps even one or two American ones! Thus even putting aside the selection bias discussed above, ignoring the approximate rank-ordering of the returns commonly associated with universities may lead to classifying some as qualified for occupations for which they are not, at least given competition for scarce jobs and/or in the eyes of potential employers.3 For better or worse, the prestige or ranking of a university signals expected productivity to potential employers. The same may also be true for lower-levels of attainment.6
With respect to occupations, S&F's reliance on the four skill strata of the National Occupational Classification (NOC) matrix is reasonable but I trust they would concede that such taxonomies are but coarse approximations.7 This seems especially true when it comes to the occupations associated ‘some college’ and ‘university’ training—where S&F observe their primary effect. For example, ‘performing artists’ are classified as needing university education but ‘manufacturing supervisors’ are believed to need only some college education. Omitted from S&F's discussion of the occupational categories is the important qualifier ‘usually’, which the NOC uses routinely.
S&F discount the potential effect of misclassification bias on their findings; they even claim that their results might be even stronger were it not for the presumed attenuation due to misclassification. But the presumption of attenuation is only reasonable when misclassification is non-differential.8 Accentuation bias is possible in situations of differential misclassification, which we should expect since the True table of education and occupation would probably differentially classify high- and low-achievers.
Fixation on status inconsistency and psychological distress
Assuming S&F's empirical results are correct, I remain sceptical of their theory or posited explanatory mechanism. Status inconsistency and Dressler's ‘goal-striving’ stress are not the only ‘theories’ consistent with S&F's results.9
It seems that those in the best jobs (i.e. those with the highest skill requirements) have the best self-reported health. And as one moves down the occupational hierarchy, health diminishes, just as social epidemiologists and economists might predict. Thus, I am not sure why these results are not attributed to income or resource effects. Such materialist explanations have little to do with status inconsistency or psychological distress, but rather the economic wherewithal to purchase salubrious environments and/or better healthcare. Without further analyses to refute these and perhaps other theoretical possibilities, S&F have not fully tested a theory of status inconsistency. Instead they found results supporting their a priori ideas. I hold dear the immortal words of Cochran, who admonished researchers to test alternative theories.10
|
Steps for the future |
|---|
|
TopShortcomingsSteps for the futureConclusionsReferences |
|---|
I hope S&F push their important ideas and analyses further. Here is my wish list: First, although it is not clear what measures might be used, an instrumental variable analysis aiming to minimize omitted variable bias would be major contribution. Second, I would appreciate a sensitivity analysis of exposure misclassification. Third, how about analyses that include income, which is available in the SPHS? What is the relationship between occupational category and income? Can S&F refute my speculation about income vs psychological effects? Finally, is it possible to flesh out the relationships in the high end (e.g. PhD or MD) of attainment?
|
Conclusions |
|---|
|
TopShortcomingsSteps for the futureConclusionsReferences |
|---|
S&F's paper falls short of definitive answers on how SES mechanisms affect health. But their effort is a solid step in the right direction. Let us hope they are relentless.
|
References |
|---|
|
TopShortcomingsSteps for the futureConclusionsReferences |
|---|
1 Smith P, Frank J. When aspirations and achievements don't meet. A longitudinal examination of the differential effect of educational and occupational attainment on declines in self-rated health among canadian labour force participants. Int J Epidemiol 2005;34:827–34.
2 Becker GS. Human Capital. New York: Columbia University Press for the National Bureau of Economic Research, 1964.
3 Arrow K, Bowles S, Gintis H. Meritocracy and Economic Inequality. Princeton, NJ: Princeton University Press, 2000.
4 Jencks C, Perman L, Rainwater L. What is a good job? A new measure of labor-market success. Am J Sociol 1988;93:1322–57.[CrossRef]
5 Yatchew A, Griliches Z. Specification error in probit models. Rev Econ Stat 1984;67:134–39.
6 Betts JR. Does school quality matter? Evidence from the National Longitudinal Survey of Youth. Rev Econ Stat 1995;77:231–50.[CrossRef]
7 Grusky DB, Rompaey SEV. The vertical scaling of occupations: some cautionary comments and reflections. Am J Sociol 1992;97:1712–28.[CrossRef]
8 Gustafson P. Measurement Error and Misclassification in Statistics and Epidemiology. Boca Raton, LA: Chapman & Hall/CRC, 2004.
9 Dressler WW. Social consistency and psychological distress. J Health Soc Behav 1988;29:79–91.[CrossRef][Web of Science][Medline]
10 Cochran WG. Research techniques in the study of human beings. Milbank Mem Fund Q 1955;33:121–36.[Medline]
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||