Paul Kelleher caught some flack for a blog post last week in which he approvingly cited a 2003 study by Angus Deaton and Darren Lubotsky (DL) that supposedly refutes the idea that income inequality causes poor health. I was curious. Paul sent me the original paper, a 2009 published critique by Ash and Robinson (AR), a reply by DL, and a longer article by Deaton.
Even though the original DL paper is getting old, and the field has progressed substantially since 2003, there is still an important take home message from this debate. Stated simply: income inequality may be one cause of poor health, but income inequality is also likely to be a mediator of other upstream causes of mortality.
The Original Finding and Interpretation
In the original DL paper the authors regress cross-sectional age-adjusted mortality data in U.S. states and Metropolitan Statistical Areas (MSAs) on the region’s measured income inequality. (They focus on the Gini coefficient, but find that the results hold across many summary statistics.) They then take an extra step of introducing a control for the fraction of the MSA that is black during the time period. They summarize their findings in the abstract:
“Across states and Metropolitan Statistical Areas (MSAs), the fraction of the population that is black is positively correlated with average white incomes, and negatively correlated with average black incomes. Between-group income inequality is therefore higher where the fraction black is higher, as is income inequality in general. Conditional on the fraction black, neither city nor state mortality rates are correlated with income inequality. Mortality rates are higher where the fraction black is higher, not only because of the mechanical effect of higher black mortality rates and lower black incomes, but because white mortality rates are higher in places where the fraction black is higher.”
Below is the key table from their study. As the table shows, in the pooled data, the Gini coefficient is a risk factor for mortality in regressions that do not control for fraction black. Once this is controlled for the effect is actually reversed (for example, going from .55 to -.38 for all males. The same basic result holds for the pooled data for females.
One puzzling finding of the DL study is why controlling for the share of blacks would also attenuate the associations between Gini and mortality within whites shown in the righthand panel. If the inequality-mortality relationship were merely an artifact of racial composition, there is no clear reason why the racial composition effect would persist after stratifying on race. Here is what DL say in the conclusion of their paper:
“The fraction of the population that is black is a significant risk-factor for mortality, not only for the population as a whole—which would follow mechanically from the fact that blacks have higher mortality rates than whites—but for both blacks and whites separately. Our empirical results provide no evidence that the association between the fraction black and white mortality is the result of confounding. The effect is robust to conditioning on education, it is present for all age-groups except boys aged 1 to 9, and it is present within geographical regions of the country.”
Is Racial Composition Conceptually Distinct from Income Inequality?
In a reply article Ash and Robinson challenge not only the original analysis but also the substantive interpretation DL offer of their results. First, they argue that the finding that fraction black attenuates the association of the Gini on mortality within whites should cause us to question the underlying social mechanisms in these regions: “Death by human agency is rare; the increased black presence is not ‘‘causing’’ white deaths in any direct way. The presence of blacks in a metropolitan area reflects the presence of social determinants that are detrimental to population health, and it is these factors that warrant identification.”
Second, and relatedly, they argue that “the ratio of blacks to whites in a given locality… captures something about the dynamics of political and economic power in that locality” pointing to historical policies that have contributed to the dominance of blacks by whites, including slavery, Jim Crow segregation, and residential polarization. In the south, where most American blacks live, many political scientists argue that efforts to enact progressive social policies have been stymied by aversion by whites of all income groups against redistribution to blacks, this political dynamic prevents lower-class whites and blacks from banding together to advocate for policies that would jointly promote their wellbeing and improve their health.
I agree with Ash and Robinson, but I think they could have made their point more clearly. I think what we should say is that racial composition is a legacy of particular economic and political processes that are themselves conceptually distinct but related to many of the processes that generate economic inequality. In some settings, and in some regions, they are likely to be more salient than others, and indeed the data from both the DL and AR results shows some variation in the effect of racial composition in different regions of the United States.
DL are absolutely wrong to call racial composition a confounder in their 2003 paper however, and I think their 2009 reply paper is much more careful. To call an omitted variable a confounder is to claim that the explanatory variable (in this case income inequality) is caused by the omitted variable and the omitted variable also causes the outcome (mortality), but that the explanatory variable does not cause the outcome. We should always include confounders in the regression equation because otherwise we will obtain biased estimates of the relationship of the explanatory variable on the outcome.
If Ash and Robinson are right, we should think of income inequality as a mediator of racial composition and mortality, that is racial composition is a causally antecedent variable (at least in some settings), and its effect is transmitted through income inequality. Whether to control directly for either racial composition or income inequality in a regression is therefore a non-obvious question that depends on the type of inference the investigator is interested in drawing. One should only include both of them in the regression when they are both intrinsically of interest, but even so it is generally not possible to identify the independent effect of each with observational data. Better yet, we would hope to find some instrumental variable that allows us to separate the exogenous effect of racial composition from income inequality. But to get to this point we need to have a much deeper and richer understanding of what processes historically have regulated racial politics and resource redistribution in American regions.