Causal inference, mediation analysis and racial inequities

Causal inference, mediation, race, ethnicity, inequity

The presence of large and persistent inequities in health outcomes by race underlines the importance of research aimed at determining modifiable drivers or maintainers of these inequities to be addressed. Mediation analysis informed by causal inference theory can be a valuable tool in this regard, such as the analysis by Sung et al.¹ However, the application of these methods to racial inequities is not without challenges, and researchers must take care that their approaches and interpretation of effects do not result in harm, particularly to racialized groups.^2–5

One of the greatest challenges in the appropriate use of causal inference and counterfactual theory for racial equity research is defining the concept of race.^2,4 Causal inference requires a clear causal question that includes a well-defined and specific exposure.⁶ However, in reality race is a composite variable that includes a multitude of components, many of which are not well defined by researchers.^7,8 Moreover, the meaning of race and its respective components are context-, place-, land- and time-specific⁹ and, by extension, the observed differences in a range of social, health and wellbeing indicators by race are rarely due to biology, but are a result of the historical and ongoing effects of racism and settler colonialism across times, places, lands and contexts.^3,5,10 This challenge is key for researchers to address, as how we define race determines the racial effect that we observe in our analysis, the generation of knowledge(s) and how information is interpreted and acted upon. Therefore, understanding what our race variable is representing, and having sufficient data on the components to tease apart what is driving positive and negative effects, is crucial—particularly if we want to avoid harm from overly simplistic interpretations of inequities driven by complex social forces.

From an epidemiologist’s perspective, there are two broad approaches that can be used to tackle this challenge. The first is to incorporate within our definition of race the complex contextual factors that underpin the process of becoming a member of a racialized group.³ As such, any interpretation of the effect of race acknowledges this context and the process of racism and racialization that has led to both the formation of a racialized identity and inequitable health outcomes. Mediation analysis in this context tries to identify modifiable intermediate factors that exist within this context and that could be intervened on to remove part of the inequity. Importantly, the interpretation of the ‘leftover unmediated effect’ must take these contextual factors into account, given that they are part of the exposure variable of interest. This helps to avoid the use of simplistic explanations based on unexplained biological, cultural or lifestyle factors.

Alternatively, the concept of race could be broken down into its components or ‘bundle of sticks’ to drive understanding of how each of these components relate to health inequities.^7,11 This approach can have the advantage of well-defined exposures rather than a single poorly defined composite race variable. It can also help us to understand the pathways through which various race-based component exposures work to generate disadvantage or promote resilience. This approach also lends itself to mediation analysis, as some of the components of race can be examined as potential intermediate factors contributing to health inequities. Given that socioeconomic inequalities, structural racism and racialized identity are inextricably linked within social contexts, the examination of specific elements frequently linked to racialized identity, such as neighbourhood deprivation or acculturation, could be considered as an example of this approach. It also allows examination of positive components that may improve health outcomes, such as connection to country, engagement with culture, use of language and a strong sense of kinship and solidarity.^12,13 It is important to recognize that residual effects unexplained by the elements of race included in the analysis are due to the effect of the unmeasured elements of race and/or explicitly state that the direct effects are due to unobserved mediating pathways not included in the analysis.

It is the latter approach that has been used by Sung et al.¹ The authors’ present evidence of mediation of racial inequalities in cardiovascular mortality in cancer survivors by neighbourhood, socioeconomic status and insurance status. The remaining direct effects of race on mortality inequities are rightly interpreted as those being driven by unobserved mediating pathways not included in the analysis. However, in order for mediation analysis to provide actionable information and meet the consistency assumption, the potential intervention that could be used to modify the mediator should be well defined. The Sung et al. paper is limited in this regard; for example, neighbourhood socioeconomic status (SES) is a multifaceted exposure that likely includes multiple different elements that could be intervened on to improve racial inequities in health outcomes.

Linked to the complexity of defining race is how we depict our assumptions about the complex pathways involved in the creation of racial inequities in directed acyclic graphs (DAGs). Observed differences by racialized identity can include socioeconomic deprivation, health inequities and others.^2,3 These are intertwined and evolve across generations, as well as over an individual’s and a population’s respective life courses. Many of the contextual factors driving these inequities are difficult to measure and usually missing from research studies. Moreover, the social exposures involved in the formation of inequities interact and reciprocally affect each other over time, and as a result cannot be explained by single static exposure measures. Whilst these reciprocal types of effects can indeed be summarized in DAGs with the use of time-varying exposures, and have been referred to as ‘spirals’ of acyclic, temporally-ordered, bidirectional causal effects,¹⁴ multiple measures over time are rarely available to researchers.

Despite the frequent lack of measurement of complex contextual factors and how they change over time, a key principle in the construction of DAGs is the inclusion of all common causes of variables on the graph which are likely to bias findings.¹⁵ These common causes should be included regardless of whether they have been measured and are available to researchers or not. This principle of DAG construction is essential to the conduct of research on racialized health inequities. First, it continues to highlight the importance of structural factors as the underlying drivers of inequities, rather than their downstream and more individual-based effects. Second, the explicit definition of unmeasured factors to the DAG makes it clear that they are an essential part of interpreting the leftover or direct race effect when using mediation analysis. Whether researchers conceptualize unmeasured contextual factors as part of the race variable itself, or treat them as separate mediating factors, they are an important part of interpreting the observed effects.

Sung et al. have included a DAG in their paper but it is simplistic and does not clearly articulate or identify the unmeasured confounders. Moreover, they do not include confounding pathways between race and the measured mediators, or between the mediators and cardiovascular disease (CVD) death, even though these are likely present. The presence of confounding between mediators and CVD death is particularly important as its presence could mean that the observed mediation relationships are being driven by unblocked confounding pathways.⁸ As a result, any intervention targeting such mediators will likely be ineffective in reducing CVD mortality inequalities by race. The authors have tried to quantify the potential of unmeasured confounding by estimating E-values. However, many of the E-values produced are small enough that confounding is plausible as an explanation. In addition, even the slightly larger E-values (2.02–2.79) for the mediators of neighbourhood socioeconomic status (SES) and insurance could be easily explained by the presence of several unmeasured confounders that together would result in substantial effect sizes.¹⁶ A key confounder that has not been measured and that could confound the association between these mediators and the outcome is individual measures of SES. A second limitation of the DAG is that there is little consideration given to exposures that occur across different time points, including the possibility that some mediators (e.g. reduced access to clinical care) are preceded by other mediators (e.g. neighbourhood SES).

Causal inference and mediation analysis can be of substantial value for understanding the potential effect of addressing downstream mediator(s) of health inequities, when taking context and past exposures into account. The challenge is how to ensure that we appropriately interpret these downstream mediators and their impacts without losing sight of the context within which they were created. Whereas well-defined downstream exposures are easier to measure, align better with causal inference theory and are perhaps more readily amenable to interventions, we should not neglect the more challenging and less easily measured upstream drivers of these structural inequalities and inequities which may be less amenable to theorizing.^2,4 Epidemiological approaches that address upstream and downstream causes of inequities are essential. However, clear and consistent guidance on the best, contextually-specific, methodological approaches to do this are required.

Author contributions

F.S. and A.D. developed the initial list of ideas to discuss. F.S. drafted the first version of the commentary. A.D. and R.M. reviewed and edited subsequent versions. All authors contributed important intellectual expertise to the work and approved the final version.

Funding

A.D. was supported by a University of Queensland Faculty of Medicine Fellowship.

Conflict of interest

None declared.