Commentary

Social epidemiology encompasses the study of relationships between health and a broad range of social factors such as race, social class, gender, social policies, and so on. One could broadly partition the work of social epidemiology into surveillance (ie, descriptive relationships between social factors and health, tracking of health inequalities over time) and etiology (ie, causal effects of social exposures on health).1 Many social epidemiologists believe these twin pursuits should ultimately serve to structure interventions aimed at reducing health-damaging social exposures or increasing exposure to social factors that enhance health. SOCIAL EPIDEMIOLOGY RISING Two decades ago one could have argued that social epidemiology was a fringe subdiscipline; now it is as prominent as any other specialization within epidemiology. Many schools of public health now offer concentrations in social epidemiology, most countries have public health goals aimed specifically at reducing social inequalities in health, and no fewer than three general and two methodological textbooks on social epidemiology are now in circulation. More importantly, social epidemiology now occupies an important place in the realm of public health policy. The World Health Organization’s landmark Commission on the Social Determinants of Health Report,2 published in 2008, attempted to bring worldwide attention to social determinants of health and to affect policy decisions in ways that would reduce social inequalities in health. The Commission’s report would not have been possible without the pathbreaking work on health inequalities that developed in the United Kingdom after the Black Report3 in 1980, which subsequently led to two full-scale reports on health inequalities in the United Kingdom.4,5 Each of these reports has used evidence assembled by social epidemiologists to advocate for policies aimed at reducing health inequalities, and yet, after the most recent report5—which found little evidence that health inequalities in England have decreased—there seems to be considerable consternation about the potential for policies to reduce social inequalities in health. Johan Mackenbach, who has produced as much evidence on health inequalities across the European region as anyone, seemed especially pessimistic: “The main conclusion therefore is that reducing health inequalities is currently beyond our means. That is the sad but inevitable conclusion from the story of the English strategy to reduce health inequalities.”6 Echoing these same concerns, Bambra and colleagues were dismayed by the similarity of the conclusions of each of the English reports more than 30-odd years and questioned the purpose of accumulating evidence from social epidemiology, given the fact that the policy conclusions never seem to change.7 These self-critical remarks among committed researchers are concordant with evidence from policymakers who charge social epidemiologists (and the broader social determinants of health movement) with providing “policy-free evidence” and the “right answers to the wrong questions.”8 CRITIQUES OF THE SOCIAL APPROACH The gap between the questions policymakers want answered and the kinds of questions social epidemiologists have been answering is evident in a series of commentaries on the most recent report9 by Michael Marmot on health inequalities in the United Kingdom. While social epidemiologists largely praised the report (some even chided the UK committee for not being radical enough in their policy recommendations),10 others questioned whether the report had provided sufficient evidence to back its strong policy recommendations.11,12 Critics accused social epidemiologists of being “casual about causality” and argued that there was, in fact, little evidence to suggest causal effects of income or education on health (although there is better evidence for the association between early-life conditions and health in later life). The discrepant readings of this evidence may be a consequence of social epidemiologists generally answering descriptive questions (whether socially disadvantaged individuals have poorer health), but interpreting their findings causally (whether socially disadvantaged individuals would have better health if they were to become advantaged, or vice versa). Such interpretation implies both an intervention and a causal effect of that intervention. Social epidemiologists often present their results in counterfactual terms that imply causation: if clerical workers had the same mortality as administrators,13 if blacks had the same mortality as whites,14 if everyone had the same mortality rates as those with a university education,15 the reduction in disease burden would be substantial. Furthermore, such counterfactuals often do not correspond to any known or feasible interventions. We suggest that the epidemiologic question should be reframed and that changing the question can make an important difference both to the answer one gets about the impact of social conditions on health and to one’s ability to provide useful information to policymakers. Some examples follow. REFRAMING THE QUESTION The UK Whitehall studies have provided numerous analyses demonstrating positive associations between civil service occupational grade and health, but concerns are often raised about the potential for social gradients to arise because of unmeasured confounding or health-related selection. Whitehall investigators have largely attributed their results to social causation, arguing that any social selection based on health is not strong enough to overwhelm the effect of social causation.16,17 These studies typically ask the descriptive question of whether those in higher status occupations have better health (controlling for a range of covariates), rather than the question of whether intervening to change occupational status would lead to improved health outcomes. Case and Paxson,18 however, use the Whitehall data to reframe the question and ask whether changes in occupational grade affect changes in health and, in a parallel fashion, whether changes in health affect changes in occupational grade. Using a fixed-effects design that controls for unmeasured time-invariant individual characteristics, they find no evidence that civil service grade affects health, but some evidence that health affects future employment. They also find that future civil service grade predicts current health, suggesting health-related selection or unmeasured confounding. Although this identification strategy trades off a lot of precision for a reduction in bias,19 it is probably closer to the kind of question that we would ask if we could design a randomized trial to evaluate the effects of occupational grade on health. A second example comes from a randomized trial called the Moving to Opportunity study. The backdrop here is an abundance of observational research showing that neighborhood characteristics such as poverty concentration are associated with a wide range of health and social outcomes.20 In the randomized study, nearly 5000 families living in high-poverty urban neighborhoods in five U.S. cities were randomly assigned one of two treatments, one of which required relocating to a more prosperous neighborhood. The intervention was successful in reducing exposure to concentrated poverty, but intention-to-treat (ITT) estimates on social and health outcomes (in the presence of substantial non-compliance) were largely null.21 Sociologists Clampet-Lundquist and Massey used the same data but analyzed it as an observational study, arguing that the ITT estimate can “measure the effects of the policy initiative, but is not well suited to capturing neighborhood effects.”22 This quote seems to reflect some of the tension mentioned above between the kind of evidence wanted by policymakers (causal effect of the policy) and the kind of evidence being delivered by social epidemiology. The investigators found evidence that increased exposure to low poverty areas is beneficial—although by breaking the randomization they reintroduced precisely the same kinds of selection effects that led to the concerns cited above about being “casual about causality.”23 This does not mean that randomized experiments are the only solution. Sampson and colleagues,24 for example, were able to “recover” the randomization study estimates on verbal IQ by explicitly mimicking the design of a trial in observational data from Chicago neighborhoods. A third example of reframing concerns the effect of education on mortality. Countless social epidemiologic studies have followed persons with high versus low education and found substantially increased mortality among the less educated—an association that is in most cases robust to controls for additional socioeconomic, behavioral, and biologic factors.25 These observational studies ask whether more educated individuals have lower mortality, but generally do not ask what policymakers want to know: what would the health effects be of intervening to increase a person’s amount of education, or their achievement of a certain educational qualification? Most observational study designs cannot answer this question, because it is very difficult to control for the multitude of factors that jointly determine both education and health.26 However, more recent studies have used changes in schooling policies that are more plausibly independent of unobserved individual characteristics and their health outcomes (eg, minimum age required for leaving school) to minimize unmeasured confounding. These studies bring us closer to answering the question of whether intervening to increase educational achievement is likely to have effects on mortality, and the results are decidedly more mixed.25 Several studies provide evidence that exogenous changes in schooling resulting from policies reduce mortality,27–29 but several other studies do not, even for different analyses of the same country.29–32 The contrast between the consistency of studies answering descriptive questions and the mixed evidence for (more approximately) causal questions suggests that this distinction matters. A final example comes from our own work, but again reflects the importance of focusing on a well-specified causal question. A recent article found that U.S. states with medical marijuana laws had higher adolescent marijuana use.33 The authors focused on two comparisons: (1) the difference in marijuana use between states passing laws and states with no laws and no history of such laws and (2) the difference in marijuana use between states that had already passed laws and states that had never passed a law. Neither of these contrasts represent the causal question facing policymakers in a state currently without a medical marijuana law; namely, what is likely to happen to marijuana use among adolescents if we legalize medical marijuana? Yet the data do permit asking precisely this question. Using difference-in-differences estimation, we compared the change in marijuana use among adolescents after a medical marijuana law is passed to corresponding changes over the same period in states that do not pass a law and found little evidence of any effect.34 Although this strategy obviously does not approach randomization to a medical marijuana law, it provides better control for confounding by unmeasured state characteristics and overall secular trends. More importantly, the analytic strategy follows directly from asking a specific question about social policy. CONCLUSIONS The above examples illustrate that the manner in which social epidemiology asks questions has important implications for the evidence it generates. Different questions generate different answers, even with the same data, and if the goal is to inform policy then our questions should be motivated by identifying causal relationships that can be of the most use to policymakers. It goes without saying that in this short essay we have painted social epidemiology with perhaps too broad a brush—many social epidemiologists undoubtedly produce high-quality evidence on the impact of social exposures on health. Yet, our concern here is that much of social epidemiology is providing questionable answers to important questions, rather than focusing on answerable questions. Batty35 recently lamented the lack of interest among the research community in generating experimental or quasi-experimental evidence on social determinants of health. To focus on answerable questions, social epidemiologists may need to sharpen their focus on specific questions that are more narrow in scope, but for which experimental or quasi-experimental data exist. The major strength of experimental and quasi-experimental designs lies in their robustness to unobserved confounding. Such designs also have conceptual strengths that come from asking specific causal questions. Trials have their own limitations and will never answer some social epidemiologic hypotheses, but attempting to mimic the randomized trial one would hypothetically conduct to answer an important question would seem to be a good place to start.36,37 Finally, we would also argue that greater transparency in methods and reporting would increase the value of replication studies38 that can directly address the question of how sensitive research findings in social epidemiology may be to measurement error, model selection, unmeasured confounding, and other potential biases.39,40 Greater attention to both research design and rigorous analysis can help social epidemiologists answer important questions rather than continuing to generate questionable answers. ABOUT THE AUTHORS Sam Harper and Erin Strumpf are Assistant Professors in the Department of Epidemiology, Biostatistics & Occupational Health at McGill University, where Erin Strumpf also holds an appointment in the Department of Economics. Sam Harper works on measuring social inequalities in health and how they change over time. Erin Strumpf's research in health economics focuses on measuring the causal impacts of health and health care policies on spending and health outcomes overall, and in inequalities across groups.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.