Commentary

Abstract

The article by Nandi and colleagues1 is the latest addition to research investigating the extent to which behavioral risk factors contribute to socioeconomic disparities in mortality. This topic is “a classic” in social epidemiology, but findings from these studies have substantial policy implications. It is therefore important to ensure that such findings, taken from the ivory tower of the academia to the policy arena, are as free as possible from bias. The considerable interest of the present article as an illustration of contemporary epidemiologic approaches lays in its advanced modeling of longitudinal data, application of a marginal structural model to solve the conundrum of variables that both confound and mediate the associations investigated, and marathon of sensitivity analyses reported on (almost) every measurement or modeling aspects of interest. Could all these efforts, compared with the simpler approaches of previous studies, provide more relevant information for policy? ASSESSMENT OF STUDY VARIABLES: A FIRST LAYER OF SENSITIVITY ANALYSES Nandi and colleagues1 relied on factor analyses to assess adult socioeconomic status as their main exposure. This approach was used to derive a comprehensive measure of socioeconomic status, which, as opposed to previous studies that relied on a unique socioeconomic indicator,2–5 allowed inferences on overall socioeconomic disparities rather than on components such as educational or occupational disparities. The cost, however, was that socioeconomic status was a model-based indicator with associated error rather than a directly observed quantity. To push their Herculean labor of sensitivity analyses a step further, Nandi and colleagues1 could have accounted for the uncertainty inherent in their socioeconomic variable in the modeling of its total and direct effects on mortality. The nature of information on health behavior was crude but comparable with that of previous studies (even if no information was available on diet, as in European work).4,5 The authors attempted to tackle this with a panoply of sensitivity analyses that together questioned the definition of behavioral variables used (with an examination of alternative measures), changes in the definition of behavioral indicators over study waves, and the reliability of the measures (influence of measurement error on the controlled direct effect).6 Regarding the temporal assessment of health behavior, some previous studies3–5 (like the present one) addressed health behaviors as time-varying variables to account for widening socioeconomic differences in behavioral risk factors and thus for their potentially increasing contribution to the socioeconomic mortality gradient. In the United Kingdom, Stringhini and colleagues5 found that a time-varying assessment of health behavior explained a much larger percentage of the socioeconomic gradient in mortality than did a baseline assessment of health behavior—an aspect that was not examined by Nandi and colleagues.1 Nonetheless, Nandi and colleagues1 expanded on previous literature by attempting to account for the accumulation of risk over a portion of respondents’ histories of health behavior. However, averaging behavioral information over successive study waves did not explain a larger share of the socioeconomic gradient in mortality, perhaps because of the relatively short fraction of individual histories that was available. Future work relying on longer follow-up should formally compare baseline versus time-varying, and single time point versus cumulative, assessments of health behaviors. CONFOUNDING AND SELECTION: MORE SENSITIVITY ANALYSES! Previous studies have routinely addressed confounding of the association between socioeconomic status and mortality; however, they have devoted much less attention to confounding of the relationships between health behaviors and mortality. While Nandi and colleagues1 are correct to mention that previous US work on the relationships among socioeconomic status, health behavior, and mortality controlled for health status, this previous work did so for different reasons. For example, one study2 adjusted for health status at baseline to examine whether socioeconomic status generated additional health disparities in later life conditional on health status in midlife. A clear strength of the analytical design by Nandi and colleagues lays in the temporal ordering of study variables: time-varying health status was lagged one study wave prior to health behaviors (which were themselves lagged one wave prior to the assessment of mortality) in an attempt to account for the hypothesized causal effect of health status on subsequent health behaviors that may confound behaviors’ association with mortality. The study provides an interesting example from social epidemiology in which conditioning on a variable through conventional regression adjustment removes confounding from the mediator-outcome association, as desired, but at the same time overadjusts the exposure-outcome association. In addition, this illustrates a marginal structural model that safely handles confounding through weighting (with the additional benefit of addressing potential collider-stratification bias for the exposure-outcome association from conditioning on health behavior as determined by both socioeconomic status and health status). Nandi et al1 suggest that there was “modest evidence of confounding of the relation between health behaviors and mortality by prior health status.” A closer look at the data in eTables 8–10 suggests that the relationships between poor health status and subsequent health-damaging behaviors varied from positive to negative, depending on the health status indicator and the health behavior examined. A variety of potential mechanisms would have to be adjusted for, such as medical guidelines encouraging sick participants to refrain from smoking and drinking alcohol or, in the opposite direction, psychiatric conditions potentially increasing smoking as a coping behavior, or disease-associated fatigue leading to physical inactivity. These opposite patterns of associations may have resulted in confounding of the health behavior-mortality relationship in opposite directions, and thereby to net confounding of relatively small magnitude. An issue is that the relationships reported in eTables 8–10 between prior health status and current health behaviors are unlikely to reflect causal effects because, contrary to what would be expected from the causal graph in eFigure 2, they were not adjusted for prior health behavior. For example, despite the careful temporal lags introduced in the data, the strong increase in smoking prevalence associated with worsening self-rated health (eTable 8) likely expresses a causal effect in the opposite direction than that hypothesized, ie, a causal effect of smoking on self-rated health. (It seems unlikely that poor self-rated health would encourage smoking or discourage quitting smoking.) In this particular case, correcting for crude difference in health status across health behavior groups (as a consequence of behavior, rather than a cause of it) would be an overadjustment of the relationship between health behaviors and mortality. Did the authors attempt to remedy the problem by also taking account of prior health behavior in t – 2 in the inverse probability weights to handle confounding of the relationship between health behavior in t – 1 and mortality in t? Inclusion of health behavior in t – 2 in the inverse probability weights was only briefly alluded to in eAppendix B. For example, due to the likely high temporal autocorrelation in smoking, ensuring exchangeability between smokers and nonsmokers in t – 1 with respect to their smoking status in t – 2 implied a drastic increase in the prevalence of new smokers and new nonsmokers in the reweighted population (reflected in eTable 11 by the large spread of weights to handle confounding of the smoking-mortality relationship). A discussion on the rationale and implication for the controlled direct effect of ensuring exchangeability according to behaviors in t – 2 for estimating the effects of behaviors in t – 1 on mortality in t would have been useful. Another issue is related to the notion of lifestyle as a plausibly stronger source of confounding for the relationships between health behaviors and mortality. People often have a coherent attitude with respect to various health behaviors. This overall attitude, referred to as a healthy or unhealthy lifestyle, results in a sizeable correlation among not only smoking, alcohol consumption, and physical activity, but also dietary habits, engagement in preventive care, compliance with treatments, and so on. In the absence of control for the range of these behaviors, the estimated share of the socioeconomic gradient in mortality that is related to the examined behaviors likely includes a portion of the effect related to unmeasured/unknown mediators. This issue is trivial for epidemiologists who, regrettably, are becoming used to reporting confounded estimates from observational studies, but it is not trivial for policy makers who might take the findings at face value. Finally, the present commentary cannot omit warmly recommending that readers look at the admittedly speculative, but brilliant and insightful, conjectures made to quantify the bias from selective mortality following the principal stratification approach.7,8 PROPORTION OF THE GRADIENT ELIMINATED BY ADDRESSING HEALTH-DAMAGING BEHAVIORS Another strength of the study lays in the methodological rigor of the assessment of the proportion of the socioeconomic mortality gradient eliminated by addressing health-damaging behaviors (fixing behaviors to an absence of health-damaging behaviors).9–11 Nandi and colleagues1 relied on the calculation of the total effect and controlled direct effect to determine the proportion eliminated. They verified and discussed the required assumptions of absence of interaction between socioeconomic status and health behaviors and absence of confounding of the relationships between socioeconomic status and mortality and between health behaviors and mortality, and they compared the findings obtained from a risk ratio model and risk difference model9,10 (although they reported the proportion eliminated from the first but not from the second approach). Most of these aspects have not been covered in previous studies.2,3,5 The authors stated that the risk difference model “informed qualitatively similar conclusions as models on the risk ratio scale.” However, the proportion of effect eliminated by addressing health behaviors was 0.68 when calculated from the risk ratio model ([2.84–1.59]/[2.84–1] from figures in Table 4), while it was 0.51 based on the risk difference model (calculated from figures in eTable 7 as [0.035–0.017]/0.035]), which is a sizeable difference. In addition to the fact that these proportions eliminated were expressed on different scales (respectively, on the excess relative risk and risk difference scales),9,10 another potential source of difference is that, as the authors documented no interaction between socioeconomic status and health behaviors on the risk ratio scale, there were presumably interactions between them on the risk difference scale (unless the analyses were underpowered to detect interactions on one of those scales).12 It was not reported whether interactions were investigated in the risk difference model. When investigating whether intervening on a set of mediators could eliminate a portion of, socioeconomic disparities in health, for example, a legitimate expectation from public health decision makers is that epidemiologists and biostatisticians quantify the unique share of the gradient eliminated by addressing each of the examined health-damaging behaviors. Such a question should receive a simple answer in the absence of interactions between the mediators, while a slightly more complex message would have to be delivered if interactions are present. Information on the independent contribution of each behavior is important for prioritizing policy decisions as additional intervention resources should be targeted at the health-damaging behavior whose elimination would most reduce socioeconomic disparities in health. Nandi and colleagues1 investigated each of the three health behaviors separately and modeled them concurrently but did not attempt to weigh the respective contribution of each mediator. The sum of the proportions of the gradient eliminated by addressing each health behavior separately was greater than the overall proportion of the gradient eliminated when all three health behaviors were modeled simultaneously. As an explanation, these researchers referred primarily to the “three-way interactions among the health behaviors on mortality.” While this is certainly plausible, a more straightforward explanation based on the notion of healthy lifestyle discussed above is that each mediator modeled separately also captures a share of the effect of the other two omitted mediators. INTERPRETING EVIDENCE ON SOCIOECONOMIC STATUS, BEHAVIOR, AND MORTALITY: A SENSITIVE SUBJECT As a first point, it is widely accepted within social epidemiology that health-damaging or health-enhancing behaviors are not under the full control of individuals as free agents and that individuals cannot be held fully responsible for their behaviors. However, it is useful to keep in mind that an active stream of research in health economics conceptualizes healthy behaviors as related to personal “effort” under individuals’ responsibility for which individuals have to be “rewarded.” These studies attempt to estimate the share of inequalities in health or mortality that is attributable to health behaviors, described as “legitimate” disparities.13,14 To tackle this interpretation (of which more15 or less16 naive versions exist) that overlooks fundamentals in the determination of phenomena, it is useful to further develop work that sheds light on the social and psychological forces that drive unhealthy behaviors among disadvantaged populations. Second, an immediate implication of studies such as the one reported here is that a relevant strategy to reduce socioeconomic disparities in health is to tackle disparities in unhealthy behaviors. An exclusive focus on interventions addressing the excess prevalence of health-damaging behaviors in disadvantaged populations is based on the hope that social disparities in health can be addressed without addressing social disparities themselves. Even if policies to address unhealthy behaviors among disadvantaged populations are critical, such an orientation may be seen as a conservative policy approach that seeks to address social disparities in health but only within the limiting context of existing socioeconomic disparities. Finally, policies focusing on behavioral mediators without addressing socioeconomic disparities flow from a theoretical fallacy and might, as such, be partly ineffective. For example, smoking may be seen as a way to cope with the stress and adversity associated with socioeconomic disadvantage. Addressing smoking without its socioeconomic determinants would mean leaving people in poverty without one of the strategies to cope with it. This line of thought can be connected to the notion of the “counterfactual fallacy”17(p 122) that emphasizes that an exposure, like smoking, cannot simply be “magicked away.” For example, by virtue of the forces through which low socioeconomic status encourages health-damaging behavior, “something must replace the exposure in question”17 (eg, another coping strategy) if socioeconomic disadvantage itself is not addressed. CONCLUSIONS The article by Nandi et al,1 with its suite of sensitivity analyses, is one of the most remarkable examples in social epidemiology to date, illustrating the value in applying contemporary advances in epidemiology to an old question. Where do Nandi and colleagues bring this research topic, and where should the literature go from here? There is a mixture of data collection/measurement challenges and modeling challenges to address. The fact that a large number of sensitivity analyses were needed to overcome limitations inherent in the data emphasizes the need for better data. It will be a key to obtain data with a longer follow-up and with reliable information on health behaviors. (The studies analyzed often relied on very crude dietary data4,5 and had no data on other behaviors, such as nonuse of preventive care, that potentially contribute to the mortality gradient and confound the effects of the behaviors examined.) On the modeling side, strategies to account for the accumulation of behavioral risk over time will have to be compared. Studies will also have to provide information on the independent contribution of each behavior (or on the joint contribution of two behaviors in case of interaction) to help prioritize interventions. Sensitivity analyses could evaluate the biasing effect of omitted lifestyle behaviors. Finally, the risk difference and relative risk scales provide complementary information for the assessment of socioeconomic disparities in health. (The former is directly related to the excess number of cases, while the second quantifies relative inequalities independent of the prevalence of the phenomena.) Future research will have to clarify the respective usefulness, drawbacks, and interpretability of the proportion of the socioeconomic gradient eliminated by addressing intermediates when calculated on these different scales, especially in the presence of interactions on one of them. ABOUT THE AUTHORS BASILE CHAIX is a director of research at the French Medical Research Institute. He aims to improve measurement (through sensor technologies) and modeling in neighborhood and health research. DAVID EVANS is an epidemiologist in the pharmaceutical industry and has a strong interest in using modern methods for analyzing observational data to inform better clinical-care and health-policy decisions. ETSUJI SUZUKI, an assistant professor of epidemiology at Okayama University, is interested in contributing to the advancement of epidemiologic methods to make sense of causality and mechanism.