IJE Advance Access originally published online on June 3, 2005
International Journal of Epidemiology 2005 34(4):896-897; doi:10.1093/ije/dyi116
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Commentary |
Commentary: Income inequality and reproductive outcomes—that model is best which models the least
Department of Epidemiology, UNC School of Public Health, 2104C McGavran-Greenberg Hall, Pittsboro Road, CB#7435, Chapel Hill, NC 27599-7435, USA. E-mail: Jay_Kaufman{at}unc.edu
One thing that most of us have figured out by now is that when it comes to health outcomes, it is not good to be poor. Furthermore, in a racially stratified society, it is generally not good for one's health to be a member of the racial group that isn't advantaged in the social hierarchy. All this has been known for a long time, but what social epidemiologists have managed to do quite successfully of late is to suggest new ways of classifying places that are independently predictive of risk for adverse outcomes, conditional on those previously defined individual-level classifications. One of the most successful of these new constructs is income inequality, which can be operationalized with any of a variety of metrics for the spread of individual incomes in a defined geographic area. The association between income inequality and various health outcomes, adjusted for absolute income, has been a dominant trope in social epidemiology for over a decade now,1 so it is natural that authors would begin to extend this to more and more health outcomes, including reproductive outcomes like preterm birth and post-neonatal mortality.2 Since the perinatal period is presumably a sensitive one with respect to environmental and psychosocial influences, it makes perfect sense that we should hypothesize a patterning of birth outcomes along this dimension, even net individual-level status or average income of the clusters.
But what are we to do with such patterns? It is well known that places with a lot of income inequality have a lot of other characteristics that make them noxious.3 Furthermore, we know that these places didn't come to have their observed level of inequality through sheer luck, but rather through local histories of migration and development, and (in the case of units with more autonomous governance, like counties or states) through the pursuit of specific economic and social policies.4 One of the most famous examples of a low income and low inequality setting with surprisingly favorable health indices is the Indian state of Kerala,5 but what is rarely discussed is that its low level of economic inequality was the result of a concerted political struggle, culminating in 1957 with the first democratically elected communist government anywhere in the world. Given that unique history, is it reasonable to put Kerala on one end of a regression line and some place like Mississippi on the other, exponentiate that slope, and claim that we know the effect of something called ‘inequality’?
The statistical analysis reported here by Huynh and colleagues evidences both technical and substantive expertise, and the trends for both reproductive outcomes cross-classified simultaneously on material education and income tertile are impressive in their magnitude and consistency. At the purely descriptive level represented in Tables 2 and 4, therefore, I think that the authors have helped us to learn a great deal about how the world really looks in relation to these quantities. When it comes to the set of adjusted estimates in Tables 3 and 5, however, I find myself at a loss to know what to make of it all. One way to view these numbers is to conclude that the result is the proof of the invalidity of the method, i.e. the regression model makes the statistical assumption of independence within cluster (county), that the women do not ‘interfere’ with one another. But the whole inequality hypothesis suggests that this is false—the outcome for a given woman is indeed affected by how others around her are faring.
The authors begin the paper by situating their research in terms of two contrasting hypothesized mechanisms by which inequality may exert an effect: a neo-material model and a psychosocial model. Their interpretation of Tables 3 and 5 is to suggest that ‘[t]his analysis may provide some evidence in support of an association between income inequality and preterm birth mediated by psychosocial factors,’ a mechanistic inference that I find rather baffling. They seem to base this inference on the observation that an association with inequality tertile ‘remains’ (in the sense of P < 0.05) even after adding individual material education and county-level mean income to the regression models. But this inference rests on decomposition assumptions that are not assessed here, e.g. that there is no additive-scale interaction between inequality and educational attainment.6 Moreover, the enduring ‘significance’ of the association is quite beside the point, especially so given the power to detect minute effects in this large dataset. For the total study population in Table 3, for example, adjustment for a few coarsely measured material factors reduces the odds ratio estimate from a crude value of 1.23 to an adjusted value of 1.06, that is, by roughly 75%. If any mechanistic inference could be drawn from this (and I am not sure that one can), I would tend to see it as evidence in favor of the neo-material hypothesis, rather than the psychosocial alternative endorsed by the authors. In any case, I see no reason why both mechanistic pathways might not be true to some extent, and so the proposition that we will somehow choose between them on the basis of this analysis seems to me a flawed premise from the start.
Huynh and colleagues have demonstrated important relations that exist in the real world, and furthermore they have done so in a dataset that is large enough to support stratified results for a number of racial/ethnic groups. This is quite helpful for showing a remarkable degree of consistency between most of the groups, as well as for being able to observe the rather striking single exception of the Asian/Pacific Islanders, for whom the dose–response relation is opposite of the others. All of this constitutes an important contribution to the social epidemiology of birth outcomes. The paper only falters when it overreaches for causal or mechanistic inferences that go beyond what the statistical analysis can sensibly provide. Regression models are ideal for smoothing over multiple dimensions in order to depict nature in ways that allow us to recognize patterns. They often fail us, however, when we treat them as experiments.7 I do believe that mechanistic understanding will eventually emerge from good social epidemiologic analysis, but not simply by adjusting for an intermediate to see if the exposure coefficient is still ‘significant’ in the adjusted model. Rather, when we begin to understand what the world looks like through careful description in many dimensions, we will be able to use subject-matter knowledge to formulate theories that allow us to predict what we should see in other circumstances. The confirmation of these predictions will then drive the further development of the theory until we have an increasingly elaborate description of the natural world. A theory that grows in this fashion and continues to demonstrate predictive accuracy becomes increasingly credible. This is how epidemiology works.
 |
References |
|---|
 Top References |
|---|
1 Lynch J, Smith GD, Harper S et al. Is income inequality a determinant of population health? Part 1. A systematic review. Milbank Q. 2004;82:5–99.[CrossRef][Web of Science][Medline]
2 Huynh M, Parker JD, Harper S, Pamuk E, Schoendorf KC. Contextual effect of income inequality on birth outcomes. Int J Epidemiol 2005;34:888–95.
3 Kaplan GA, Pamuk ER, Lynch JW, Cohen RD, Balfour JL. Inequality in income and mortality in the United States: analysis of mortality and potential pathways. BMJ 1996;312:999–1003.
4 Thomas JC, Thomas KK. Things ain't what they ought to be: social forces underlying racial disparities in rates of sexually transmitted diseases in a rural North Carolina county. Soc Sci Med 1999;49:1075–84.[Medline]
5 Lynch JW, Smith GD, Kaplan GA, House JS. Income inequality and mortality: importance to health of individual income, psychosocial environment, or material conditions. BMJ 2000;320:1200–04.
6 Kaufman JS, Maclehose RF, Kaufman S. A further critique of the analytic strategy of adjusting for covariates to identify biologic mediation. Epidemiol Perspect Innov 2004;1:4. (http://www.epi-perspectives.com/content/1/1/4, 23 May 2005, date last accessed)[CrossRef][Medline]
7 Berk RA. Regression Analysis: A Constructive Critique. Newbury Park: Sage Publications, 2003.
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||