Research Brief - August 2005

If you would like to receive Research Brief via email, please click here.

(click here to download this issue in PDF)

Do Neighborhood Environments Affect Health? Are You Sure?

J. Michael Oakes, Ph.D., assistant professor, Division of Epidemiology and Community Health, University of Minnesota School of Public Health

For the past 5-10 years a great deal of public health research has focused on how neighborhood environments affect health. Although I think this is a good thing, I have instead chosen to focus on methodological issues, specifically whether it is even possible to estimate such effects from observational study designs. This work has led me to two conclusions: (1) the questions are far more complex than many epidemiologists believe and (2) few appreciate the long history of attempted answers. This research brief aims to present these unconventional results in a conversational fashion, accessible to all.

Contemporary investigators aim to disentangle the effects of neighborhood environments from the background characteristics of residents. The goal is to identify the effect of, say, poverty or racial segregation, on health, regardless of one's individual socioeconomic status, genetic predisposition or behavior. The motivation, it seems, is to marshal evidence so as to convince policymakers that public health may be improved by improving neighborhood conditions, with no requirement to change the recalcitrant resident's thoughts or genes.

Of course efforts to disentangle the effects of persons from contexts are not new. One may trace the genesis of such aims to 1897, when French sociologist Emile Durkheim posited that the most self-centered act, suicide, was a predictable function of not individual characteristics but social contexts. A vast amount of research has followed but around 1970, perhaps because of the “War on Poverty,” there was focused attention on whether neighborhood contexts, above and beyond the characteristics of inhabitants, encouraged crime, sense of community, or success. But after years of fits and starts, most researchers gave up, especially after methodologists - scholars who try to determine what conclusions can be legitimately drawn given a set of assumptions - showed that the then-employed statistical techniques could not ever yield a satisfactory answer, at least without randomly assigning people to neighborhood - which has been done. Why then was there, sometime around 1998, an explosion of studies on how neighborhood contexts affect health?

The reason, I submit, is that researchers gained access to a new tool. Contemporary epidemiologists have borrowed from the social sciences a relatively new statistical model (coincidentally developed in biostatistics for analysis of longitudinal data), which some believe permits us to overcome obstacles and disentangle the effects of contexts from background. This so-called “multilevel” model has become de rigueur in social epidemiology; literally hundreds of peer-reviewed papers rely on it to fix fundamental design problems. The trouble is, the model, at least in my opinion, is no panacea for the problems at hand. For better or worse, it is this methodological disagreement with many of my colleagues that has been propelling my most recent work and generating controversy. But make no mistake, I have nothing against the inherent qualities of the multilevel statistical model itself; my objections are with the application, misuse, and misunderstanding of it.

I have published four principal arguments against the application of the multilevel model for estimating neighborhood effects, but will highlight two here. First, the application of the model to the sorts of data most social epidemiologists are analyzing ends up yielding estimates based on inappropriate comparisons. I have shown that the regression model's inherent smoothing over of sparse data cells interpolates data where there can be none. By this I mean that, at least in America, the wealthy will never live in poor neighborhoods and the poor cannot afford to live in rich ones; in other words, social stratification is real and matters. I use the term “structural confounding” to convey this problem; the confounding is structural since it cannot be overcome by better sampling methods or larger sample sizes. The upshot is that unless one imagines a social revolution, it is inappropriate to statistically “equate” people that are quite different in terms of cumulative exposures, opportunity structures, and the like.

Second, my research shows that the conceptual model that many use to motivate their use of the new statistical model is flawed because it fails to recognize where neighborhood “goods” and “bads” come from. Some believe the new “socioecological” theory yields testable hypotheses that, when interrogated by the multilevel model, yield parameters describing how contexts affect individuals. The trouble with this is that while neighborhoods may in fact “make people,” it is inescapably true that people “make neighborhoods.” It is patently obvious that differing levels of economic and political power determine the level of a neighborhood's “good” and “bads.” The Not-In-My-Back-Yard (NIMBY) phenomena is very real; it is not hard to show that there are no toxic dumps in wealthy areas and few, if any, beautiful and safe parks in poor ones. It follows that theories failing to incorporate this so-called micro-to-macro transition are doomed to t fail. Hence, the theoretical motivation for using the multilevel model is without merit. The fact is, individual- and neighborhood-level socioeconomic status are per force highly correlated. Nothing new here either. Overlooked by epidemiologists is the vast economic literature on neighborhood choice, especially the classic 1956 paper by Charles Tiebout outlining how people/households decide where to live based on their interest in, and abilities to afford, various “local public goods.”

Stepping back, my research has revealed that many epidemiologists, even some social epidemiologists, continue to envisage and model Robinson Crusoe: someone affected by only biological organisms and Mother Nature, but not other individuals, strategic interaction, or social institutions such as laws and prices. Even the recent efforts of social epidemiologists seem to treat “society” and “social structure” as island weather, influential but not something to be explained. Such practice defies the fact that epidemiology is a population science and that exposures are not random but socially mediated. People and contexts matter, at least to me. This complaint is not new either. It was voiced by some of the very first epidemiologists, such as Rudolf Virchow, whose slogan was “Medicine is a social science and politics nothing but medicine on a grand scale.” There goes the neighborhood…

For further study:

Oakes, J. M. 2004. “The (mis)estimation of neighborhood effects: causal inference for a practicable social epidemiology.” Social Science & Medicine 58:1929-52. (with discussion)

Blume, Lawrence E. and Steven N. Durlauf. 2006. “The econometrics of social interaction: a review” in Methods in Social Epidemiology, edited by J. M. Oakes and J. S. Kaufman. San Francisco: Jossey-Bass / Wiley.

Hamlin, Christopher. 1998. Public Health and Social Justice in the Age of Chadwick: Britain, 1800-1854. New York: Cambridge.