In a recent blog post, Hershbein & company of The Hamilton Project (Brookings Institute) describe a simulation using Current Population Survey (CPS) data assigning a random ten percent sample of non-college educated men aged 25-64 an income (wages, salary, and self-employment) randomly drawn from the college educated (or higher) population to examine its influence on income inequality (Gini and Theil) in 2013. They propose this as a model for a policy intervention wherein a random (but not small) subset of individuals who do not actually have college or advanced diplomas are, counter to fact, instantaneous granted one. Always on the lookout for the ways people conceptualize the effects of education (arguably the "strongest," most "fundamental" social determinant of health), two aspects of the post stood out to me:
First, their approach -- In their write-up, they helpfully identified two limitations: a more plausible intervention would be one aimed at those newly entering the workforce (i.e. younger individuals) and their intervention allowed diplomas to be granted to those who weren't even close to obtaining a one (e.g. individuals with less than high school education). These concerns address some of the challenge of finding appropriate counterfactuals to identify "instant diploma" effects, but don't go far enough. While its generally understood that non-college educated men and college educated men are not the same, even the process of matching individuals (e.g. through propensity scores) to new incomes based on numerous characteristics available through CPS data might not be so straightforward.
That's because it's unclear what spontaneously obtaining a piece of paper (diploma, transcript, or augmented résumé) really means. Can individuals suddenly "download" all of the information, training, experiences, and social connections of several years of college (à la the Matrix)? This might be an intuitive, albeit improbable, mechanism. In such cases, a random assignment of new college-graduated identities may not too far off, on average. Alternatively, perhaps the credentials are counterfeited (à la Catch Me If You Can)? If so, it might be important to consider how individual incomes might differ just from a boost in prestige and whether the group of college grads can make a fair comparison. Beyond the context of this simulation, what the diploma "stands for" has material consequences for good counterfactual match(es) between someone with and without a degree and whether the replaced outcome fairly represents the imagined effect. A different way to think about the issue might be to alternately consider what would happen if a random set of those with college degrees had their diplomas revoked. If it the prestige of the college diploma is most relevant, it might be reasonable to compare diploma holders to those without any college education at all. On the other hand, if the experience or learning is important in and of itself, years or type of schooling should be used for contrasts.
A motivating assumption for this simulation might have been that these "instant diploma" scenarios (or response types) are randomly distributed amongst both the college grad and no college groups. Given the numerous potential common causes for individual education and income, there is nothing to suggest that would be the case. An alternative solution, perhaps, is that the "counterfactual" represents an entire world where ten percent of non-college grads are replaced by college grads. In this case, you would not have to worry about defining counterfactuals (separately or in aggregate), but the interpretation is quite useless. The fact that the authors allude to assigning individuals higher education suggests they don't actually mean the latter. This challenge of defining the effect of education is well-recognized, including in the health context. A recent issue of Social Science & Medicine tackles several approaches of identifying effects of educational attainment on adult health. These include mandatory schooling policies where the counterfactual is slightly better defined, e.g. being forced to stay in school for an extra year when you otherwise would have left. Such designs are favorable specifically because they define the exposure and treatment groups more discretely.
Even more interestingly, I was intrigued by their approach to addressing wage response to the increasing in proportion of college degree holders. A spillover effect (or violations in SUTVA for the technically minded) is the when the effect of an intervention isn't limited to the individual being targeted. A classic example is the protective effect of vaccines, though this is clearly quite relevant to conceptualizing interventions on educational and income inequalities as well. How do we estimate how much one would earn in a world where everyone has a bachelor's degree? What if no one had one? Some places are clearly closer to these extremes than others. Some study designs controlling for educational status imply this very thing. Do we have sufficient data to be able to estimate what incomes would be like, for example, if everyone had an advanced degree? Does the changing income potential for a certain level of education help explain trends in associations with health? If relative social status (education or income) is of interest, implementing a small bias factor, based on some empirical anchor, for the wage response to the target change in education could be applied in many analytic strategies to test this. It would be good for epidemiologists to give this some more rigorous thought.
The second thing that stood out to me was their take away message -- Ultmately, they find their simulated intervention somewhat increases incomes for the lower quantiles of the distribution, but do little to reduce overall income inequality, especially relative to their much lower levels in 1979 (from which they also drew income data). As one comment postulates, this is perhaps unsurprisingly, as extreme incomes that drive inequality may be unrelated to college-degree attainment. The authors themselves note that measures to tackle overall growth in income inequality should be separate and in addition to improvements in educational attainment and incomes for the less well off. Even taking their findings with a heavy grain of salt, this point is worth reiterating. In the health equity world, the distinction between the unfairness of social inequalities and suboptimal health due to certain adverse conditions remains a major point of contention. It can never be stated too often that it need not be an both/neither proposition: It is possible that we may find interventions to reduce unequal health outcomes without great overall reductions in socioeconomic differences. In contrast, there may also be moral reasons for certain interventions to be unacceptable (intense, targeted medical screening and treatments, for example) or for individuals to act collectively to reduce differences in socioeconomic circumstance because of the putative opportunities they afford (for good health or otherwise). Towards the latter, etiologic studies can help provide support for, but not generate new, moral concerns.
