The Consistency Assumption for Causal Inference in Social Epidemiology: When a Rose Is Not a Rose

Abstract

The assumption that exposures as measured in observational settings have clear and specific definitions underpins epidemiologic research and allows us to use observational data to predict outcomes in interventions. This leap between exposures as measured and exposures as intervened upon is typically supported by the consistency assumption. The consistency assumption has received extensive attention in risk factor epidemiology but relatively little emphasis in social epidemiology. However, violations of the consistency assumption may be especially important to consider when understanding how social and economic exposures influence health. Efforts to clarify the definitions of our exposures, thus bolstering the consistency assumption, will help guide interventions to improve population health and reduce health disparities. This article focuses on the consistency assumption as considered within social epidemiology. We explain how this assumption is articulated in the causal inference literature and give examples of how it might be violated for three common exposure in social epidemiology research: income, education and neighborhood characteristics. We conclude that there is good reason to worry about consistency assumption violations in much of social epidemiology research. Theoretically motivated explorations of mechanisms along with empirical comparisons of research findings under alternative operationalizations of exposure can help identify consistency violations. We recommend that future social epidemiology studies be more explicit to name and discuss the consistency assumption when describing the exposure of interest, including reconciling disparate results in the literature.

Validation of a theoretically motivated approach to measuring childhood socioeconomic circumstances in the Health and Retirement Study

“…This residual measurement error can be particularly problematic when investigating the role of cSES in racial or geographic disparities, potentially leading to an underestimation of the contribution of cSES in such disparities [ 15 , 17 ]. Second, a well-defined treatment is one of the assumptions of causal inference [ 18 ]; without a well-defined exposure variable in observational studies, policy interventions to ameliorate the impacts of childhood socioeconomic disadvantage on later health outcomes are unclear [ 19 ]. For example, should more resources be allocated towards helping parents pursue higher education (e.g.…”

Birth weight, verbal cognition in early adolescence, and lexical and reading skills in late adolescence: a formal mediation analysis using a potential outcomes approach

“…To evaluate the indirect effects of birth weight on oral and written language in late adolescence (lexical knowledge and reading comprehension) via verbal cognitive skills in early adolescence, we used the potential outcomes method proposed by Pearl (), Robins and Greenland () and further elaborated on by Preacher, Rucker, and Hayes (). A key assumption of the potential outcomes framework and of formal mediation modeling is to ensure that the exposure to outcome, exposure to mediator, and mediator to outcome pathways have exchangeability, positivity, and consistency; otherwise, the causal result will be biased (for further details about the assumptions, see Cole and Frangakis (); Hernán and Robins (); Hernán and Taubman () and Rehkopf, Glymour, and Osypuk ()). The potential outcomes approach was implemented in Mplus version 8.0 (Muthén & Muthén, ), which was the software used to run all the analyses.…”

The overall effect of parental supply of alcohol across adolescence on alcohol‐related harms in early adulthood—a prospective cohort study

“…That is, while there is a possibility that participants may be in the same peer group, the assumption of no interference would only be violated if participants both had peers in the study and also affected the role of parental supply in those peers. The second assumption is consistency, which assumes that the counterfactual outcome associated with a given exposure, and the actual outcome observed if that exposure occurs, are the same [41]. This assumption can be violated if, for example, the exposure is defined ambiguously.…”