Is the potential outcomes approach to social epidemiology satisfactory?
Essay 5 - Part II of Social Epi
In this essay, I will argue that the olive branch extended by the thinkers of the potential outcomes framework (POF) to social epidemiologists is unsatisfactory because POF’s proposal does not adequately accommodate the long-standing concerns in social epidemiology. First, I will demonstrate how POF conceptually falls short of understanding or appreciating the complexity of social exposures. Next, I will argue why the ‘spectrum’ as a tool fails to accommodate some of the most important aspects of social epidemiology. Taken together, these arguments will depict that POF is not yet ready to take on social epidemiology. This essay should be read as a critique of and response to certain arguments made by Galea and Hernan 2019 (G&H from here on).
G&H posit that the tension between the social and causal inference subdisciplines in epidemiology arises from certain misconceptions. They argue that considering social exposures qualitatively different from other (e.g., clinical, pharmacological, biological, etc.) exposures is a misconception potentially arising from the fact that the “academically young” social epidemiology is still figuring out ways to define social exposures as causes. G&H instead propose that social exposures can be treated similarly to other exposures under POF. I will argue below why this proposal is misguided. In the previous essay on ‘social factors as causes’, I noted that social exposures have a group-level attribute (or characteristic as social epidemiologists call it) and individual-level manifestations that may be different than the group-level attribute and that vary across individuals based on other characteristics of the individuals. Put simply, social exposures can be multilevel. The multilevel nature makes social exposures different from biological, pharmacological, clinical, or other single-level exposures that motivated the POF’s development and that have become standard examples to demonstrate the framework’s usefulness. Hence, the POF, at least the specific version that requires exposures to be well-defined humanly feasible interventions (i.e., the restricted POF as Vandenbroucke et al. 2016 called it), does not have a way to accommodate multilevel social exposures. As pointed out by Robinson and Bailey 2019 (R&B from here on), G&H prescribe a reductive “individual-focused” view of social exposures that principally does not align with social epidemiologists’ view – as G&H seem to leave out the group-level characteristic. Vanderweele and Robinson 2014 (V&R from here on) suggested that social exposures can be conceptualized as composites in which multiple inter-dependent exposures are bundled together. Each of the exposures within the composite may (or may not) have a well-defined (hypothetical) intervention under certain assumptions. For now, I won’t focus on whether conceptualizing social variables as composites is valid and useful for causal inference or not. However, my point is that such a conception was not needed for routine POF exposures. The need for the conception of composite exposures provides another instance as to why social exposures are not the same as other exposures usually considered in POF. Hence, it is not social epidemiologists' misconception but their real concern that POF thinkers do not understand how the social exposures are complex (multilevel and composites of inter-dependent exposures) and hence, different from biological, pharmacological, or clinical exposures.
G&H’s proposal of organizing the social exposures on a spectrum from no to perfect consensus over imagining well-defined intervention for the exposure and focusing on exposures that perhaps have a better consensus (i.e., right-of-the-spectrum) is also flawed for several reasons. Some of these flaws arise from their lack of appreciation for the complexity of the social exposures as noted above. For instance, the spectrum does not consider the composite nature of exposure such as race – as conceptualized by V&R. If one follows V&R’s conceptualization of race as a composite (though other conceptualizations are possible depending on the research question and context), one would need to order (or rank) physical phenotype (P), parental physical phenotype (PP), genetic background (G), and cultural context (C) on the spectrum. However, their ordering is complicated by at least two issues. First, some of these are individual-level (P, PP, and G) while others are not (C). G&H’s spectrum does not seem to have a way of distinguishing between the two or answering whether such a distinction matters or not. Second, these are interdependent exposures. Per V&R one path could be: G → P ← PP → C. The linear spectrum that goes from ‘no’ to ‘perfect’ consensus has no way of expressing such interdependence. Suppose the purpose of introducing the spectrum is to have a (conceptual) tool that can progress the use of POF for social exposures. In that case, the tool should be able to resolve these complexities inherent to social exposures. The spectrum’s inability to do so makes it limited if not futile for its intended purpose.
The ordering used by G&H has no basis. “Why does race have no consensus while income has greater consensus than that for race or residential segregation?” is not backed by empirical evidence or expert (social epidemiologist) reasoning. As noted by Glymour and Glymour 2014 (G&G from here on) who critiqued V&R’s conceptualization of race as a composite exposure1 - “… it seems somewhat arbitrary that race should be disregarded as a cause, whereas other variables retain that status.” This critique applies equally well or perhaps better to G&H’s ordering of social exposures. A piece of evidence for the applicability of this critique to G&H’s example is given by the discrepancy in their discussion around income. G&H noted income to be a right-of-the-spectrum social exposure for which there is enough consensus over imagining well-defined interventions for changing income such as income can be studied under POF for estimating its causal effects on a health outcome such as cardiovascular mortality. However, they then went on to note - “We could increase the income of randomly selected households in any number of ways…” [I added the emphasis here to draw attention]. This means that there are multiple versions of the intervention possible, which violates the assumption of having a consensus on a well-defined intervention. Previously, in the discussion around SUTVA, this issue was labeled as ‘multiple versions of the treatment’ that violates SUTVA or the consistency assumption – one of the identifiability assumptions for estimating causal effects within POF. In that discussion, Hernan and Taubman 2008 provided a detailed argument as to why obesity cannot be considered a cause within POF because it is an ill-defined exposure. Their argument used the example that the change in obesity brought out by exercise vs. that brought out by diet are two different interventions. I believe that a change in income then brought out “any number of ways” should result in the same problems as those faced by obesity. R&B also provided examples of how the multiple versions of the treatment would look like for income change (absolute income bumps vs. salary bumps based on some percentage of the base incomes, money handed over to male vs. female heads in the household, etc.) Hence, G&H contradicted themselves in assuming that income is a right-of-the-spectrum social exposure while simultaneously noting that it can be manipulated in different ways.
While I understand that the specific social exposures and the particular ordering in G&H were simply for an illustrative purpose (and the above-noted issues can be perhaps resolved by an expert social epidemiologist), it is noteworthy that the spectrum as a tool is agnostic and does not guide its users on the correctness of any such ordering. A good tool that prides itself on trying to bridge causal inference and social epidemiology would at least assist or nudge its users to think about ordering more carefully if not guide that process.
Finally, I would challenge the use of the word “consensus” for defining the spectrum. The use of consensus is related to G&H’s assumption: “Social epidemiology, a relatively young academic field, is still grappling with what constitutes a social exposure and how to think about its causal effects.” When G&H place race on the ‘no consensus’ end of the spectrum, essentially, they are implying that social epidemiologists need to think about race more carefully and whether or not it should be studied as a ‘cause’ in epidemiology. However, as R&B and several other social epidemiologists would concur, it is not the lack of thinking or consensus on the part of social epidemiologists that prevents conceptualizing well-defined interventions but the appreciation for the complexity of these exposures that the social epidemiologists have imbibed as a result of thinking ‘hard and long’ about these exposures. Hence, the spectrum is not from ‘low to high’ consensus but from ‘less to more’ complex. However, more importantly, the complexity cannot be thought of as a uni-dimensional spectrum from “low to high”. As discussed above, one can have a spectrum for individual-level social exposures and perhaps another spectrum for group-level exposures. The variables along and across the spectra may be inter-dependent and such spectra will have to be embedded in the context, space, and time. The variables, levels, and context would be dictated by the research question at hand and can change based on even subtle modifications to the question’s framing2. All that to say that G&H’s arguments for studying social exposures as any other exposures under POF are misguided due to their lack of appreciation for the complexity of social exposures and their proposal of the spectrum as a tool for the way forward is flawed as it may not be ready for use for the questions that matter to social epidemiology.
In 2014, Robinson and Vanderweele seem to hold the view that race is particularly challenging variable to deal with as a cause within POF (see pg 473, V&R: “There are, however, no reasonable hypothetical interventions on race when race itself is the exposure.”). Vanderweele seems to continue to hold this view in his 2019 response to G&H (see pg 175, Vanderweele 2019: “I fully embrace the spectrum that Galea and Hernán describe. It is easier to imagine hypothetical interventions for some exposures than for others…”). While there is no particular warrant, my reading based on the lack of explicit warrant in R&B is that Robinson may not hold such a view in 2019 or at least that it has evolved or softened to become more hopeful about conceptualizing causal effects of race.
I am leaving out feedback loops among variables that are common in social epidemiology, that the spectrum conceptualized by G&H does not accommodate. This is because I do not think that it is a particular critique of the spectrum. No accommodation for feedback loops among variables, according to my limited understanding, is a limitation of the entire POF and DAG-based conception of causation. In fact, A in DAG stands for ‘acyclic’ implying no loops.