I have reflected previously on a definition of public health that I like, suggesting that public health is about assuring the conditions for people to be healthy. I have since been engaged in discussion with some of our faculty about the robustness of this definition, a challenge that had first been taken up in print by our faculty in a provocative paper published in the American Journal of Public Health more than 20 years ago. The central critique of public health’s defined role as that of assuring conditions for people to be healthy is that such a mission is too broad, and that it extends the remit of public health to areas that are more typically the domain of social justice.
In a future Dean’s Note, I will comment a bit further about how I think of the intertwined roles of public health and social justice. Here, however, I wanted to comment briefly on another reason why I like this definition: my concern that population health science simply has no choice but to consider the conditions that shape the health of populations if we are to accurately fulfill our mission of understanding the drivers of population health. This argument rests on recognition of the centrality of—synonymously—“context,” “circumstance,” and “conditions” to population health science, and on recognition that unless context is taken into account, inference we draw from population health science can be wrong, or, worse yet, dangerously misleading.
Context and Population Health Science
Much of population health science rests on the use of studies that are either experimental or observational. Experimental studies attempt to manipulate one particular aspect under study, within a counterfactual paradigm, with the goal of understanding how that one aspect may influence the production of a particular health indicator. Observational studies do much the same, although, absent the capacity to experimentally manipulate any aspect of the study, use methods to take into account other factors that may be confounding, or obscuring, the association between an exposure of interest and a particular health indicator. Methods in observational study for confounding control have improved dramatically over the past few decades. In particular, inverse probability weighting, propensity score methods, and Mendelian randomization have opened the way for more robust causal inference in the face of non-random treatment assignment. In theory, therefore, experimental and observational studies stand to point us in the same direction, and both serve as key methods in our scientific armamentarium.
Context, Experiment, and Observation
The truth, however, is much more complicated, and comparisons of results from observational studies and randomized controlled trials (RCTs) provide an illuminating window into the central role of context in public health and its study. Classic observational studies such as the Harvard Nurses’ Health Study (NHS) have identified treatment effects of critical interest to public health practice, only to be robustly contradicted by data from RCTs. One now well-worn example is that of hormone replacement therapy (HRT) and cardiovascular risk in postmenopausal women. The NHS found in 1985 that HRT (i.e. exogenous estrogen) was protective against coronary heart disease at four-year follow-up [see figure 1, below], later reconfirming this finding at 10-year follow-up, and bolstering the continued proliferation of HRT in postmenopausal women.
In 2002, these observational findings were robustly contradicted by the Women’s Health Initiative (WHI), which found, based on their five-year placebo-controlled RCT, that HRT was associated with greater risk of coronary heart disease [see figure 2, below].
A similar example is that of beta-carotene for prevention of lung cancer. In 1986, a case-control study of beta-carotene and lung cancer found that people with low levels of serum beta-carotene had approximately four times the risk of lung cancer relative to those with high levels. Again, in a striking about turn, the New England Journal of Medicine published a 1996 paper, based on the results of a large five-and-a-half-year placebo-controlled RCT, that found that beta-carotene was associated with a slightly higher risk of lung cancer relative to placebo.
Explaining the Differences Between Experiment and Observation
Why these discrepancies? Technically, much of these discrepancies are due to “unmeasured” confounders, differences between the comparison groups of interest that distinguished the observational and the experimental study. Thinking centrally about conditions, and the role of social context, the NHS was conducted among nurses, a homogenous and healthy group, relative to the general population sample that made up the WHI. The conditions, and context, that shape the lives and circumstances of the nurses in the NHS are different in countless ways from the conditions that shape the lives of the broader population. In addition, RCTs are predicated on the random assignment of active treatment vs. placebo. RCTs are thus more artificial, speaking more to controlled conditions than the real world. Critically, in RCTs, the bind is broken between the choice of treatment and the potential complement of other factors related to study outcome, whereas in observational studies, individuals self-select into treatment condition. One type of bias induced by self-selection in observational studies is referred to as “confounding-by-indication,” so named because those who are treated are often healthier (or sicker) in ways that are correlated with study outcomes. This has long been thought to be the cause of the discrepancy between the NHS and WHI findings. Similarly, this may explain the discrepant findings on beta-carotene and lung cancer, with higher serum levels representing a marker of healthier lifestyle. Another popular potential explanation in the case of HRT is that initiation constitutes a higher risk period for coronary heart disease that levels off with continued use. A recent reanalysis of the NHS treated the observational data as a “sequence” of RCTs of HRT initiation (without random treatment assignment) and arrived at similar effect estimates to those found in the WHI study. The authors suggest that the differences in findings between the initial NHS analysis and the WHI study are largely explained by population differences in time-since-menopause relative to HRT initiation. In particular, those initiating at equal or greater than 10 years post-menopause were at higher risk relative to non-users, while those initiating less than 10 years post-menopause were at lower risk relative to non-users. Their findings also suggest that the study discrepancies cannot be explained by higher risk associated with the treatment initiation period. Taken together, these data suggest that the discrepancies initially noted regarding HRT and coronary heart disease may be rooted not in the contextual features of study design, but rather in the life course context (years since menopause) of the population under study and treatment. On the other hand, the observational study context, and specifically confounding by indication, remains a likely explanation for the divergent findings in studies of beta-carotene and lung cancer.
The Inextricable Role of Context, Or Why Conditions Matter
There has been a large literature that has considered carefully the differences in these observational and experimental studies, much of it technical and important to the advance of population health science. However, at core, the difference rests on differences in context. Ultimately, our science has been about trying to understand how we can isolate particular factors that are causes of population health, with the goal that such isolation may help us intervene. However, as the two examples above and many others illustrate, the role of context is pervasive and inextricable from the observations we draw from our science. Context can well be in the form of life course experience, or in the form of lived circumstances. It can be in the form of inter-individual experiences (social networks, experiences of discrimination), individual resources (poverty, social supports), or shared properties and conditions (social cohesion, inequality). Whatever form context takes, it ineluctably influences the inference we draw from our science, and must be an integral part of population health scholarship.
Does this then justify public health being about the conditions that make people healthy? That, to my mind, is a larger question, and one of relative values and priorities. It does, however, make it clear that the conditions that shape populations are unavoidable if we are interested in generating a science of consequence, and a scholarship that stands the test of time.
I hope everyone has a terrific week. Until next week.
Sandro Galea, MD, DrPH
Dean and Professor, Boston University School of Public Health