Commentary: Identification, neighbourhoods and families

Abstract

It has been more than 10 years since social epidemiologists focused their attention on estimating the independent effect of neighbourhood contexts on health outcomes. My read of the literature is that the flurry of research, and the corresponding taxpayer investment, is tapering, with fewer and fewer neighbourhood-effect studies being funded or published. After a decade’s worth of ambiguous results, this is probably a good thing. As I wrote almost 10 years ago, without experimental manipulation it is very difficult to determine the extent of a neighbourhood’s influence on health outcomes. Among other issues, selection and structural confounding undermine the identification of neighbourhood effects in all non-experimental designs. Further, the potential for policy change as a result of neighbourhood effects research seems increasingly remote, at least in the USA. The new paper by Sariaslan and colleagues (hereinafter ‘the Authors’) only reinforces my perspective. The paper merits attention because the Authors create and exploit an astonishingly rich data resource that seems available nowhere else, certainly not in the USA,where distrust of government monitoring appears to be growing. For purposes here, the key characteristics of the data are its population scope, the composite deprivation index, the longitudinal information on siblings and the observed residential mobility. The Authors’ careful analysis shows that the independent effect of neighbourhood contexts on violence and substance misuse is negligible. Among other strengths, I am most impressed with the Authors’ efforts to consider alternative explanations for their findings. Social epidemiology needs more of this kind of work. But, as science tends to advance by criticism, I shall offer some. First, I am not convinced by the Authors’ specification of their deprivation index. Do divorce and immigrant status really imply deprivation? Further, if the index was created from the very same data used to assess its relationship to the study’s outcome variables, any association between the index and outcomes capitalizes on chance, rendering P-values too small. It is always best to create an index (i.e. latent variable) in a different data set than the one in which its associations with other variables are evaluated; split-half and independent sample techniques are recommended. That said, it is hard to imagine alternative specifications or procedures would alter the study’s conclusions. I also might quibble with the Authors’ use of the term ‘quasi-experiment’. I reserve this term for designs with an exogenous intervention/treatment that is not randomized by a researcher or by Mother Nature. But again, this has no impact on the findings. Ultimately, my primary concern rests with the Authors’ (in)ability to identify desired casual effects. As the word implies, identification is the process of teasing out empirically defensible causal effects from competing explanations even as sample sizes approach infinity. Identification means that one and only one explanation or model explains the data, that there are no competing explanations for the very same results. Table 2 in the Authors’ paper reveals the expected relationships: as neighbourhood deprivation increases, so too does the prevalence of undesirable outcomes. This is straightforward descriptive epidemiology and tells an important story: the probability of a subject committing a violent crime is 0.008 in an advantaged neighbourhood and 0.055 in a deprived neighbourhood. Although negligible in absolute terms, there is a nearly 7-fold increase in violent crime from the lowest to the highest decile of deprivation. The Authors next ask how much variation is potentially attributable to neighbourhoods themselves. The answer comes in their Table 3, which presents intraclass correlation coefficients (ICCs). Results show that 12% and 4% of violence and substance misuse, respectively, are potentially attributable to neighbourhoods. This is about as much clustering as I would expect with a rare dichotomous outcome in a society less segregated than the USA. Still, when it comes to making claims about potential explanatory power of neighbourhoods, it is important to remember that the key assumption in the calculation of an ICC is that the variance components are independent of one another. To see this, consider a simple linear mixed Published by Oxford University Press on behalf of the International Epidemiological Association