Lies, Damned Lies, and Statistics

The election is over, but analysts across the globe are still trying to account for Donald Trump’s victory over Hillary Clinton.

As a health policy institute, one explanation in particular caught our attention. A recent analysis from The Economist identified local health outcomes as the top predictor of whether a county voted for Trump or Clinton — stronger even than the presence of white residents without a college degree. So strong, the authors argue, that Michigan would have gone blue if its diabetes prevalence were just seven percent lower.

Here at CHI, this conclusion struck us as odd.

We trust The Economist’s math and know they controlled for the effects of age, sex, race, marital status, income, employment, education and immigration — areas where Trump support and poor health outcomes often overlap. We just couldn’t understand how poor health would increase support for Trump. All else being equal, why would diabetics find his message more appealing?

We don’t think there is a compelling reason to think that the election would have turned out differently simply because there were fewer diabetics. When a relationship is statistically significant but hard to explain, it’s risky to say one thing causes the other. For example, the divorce rate in Maine is strongly correlated with per capita margarine consumption. But since we don’t know how one trend would affect the other, it’s probably coincidence or an overlooked variable that affects both marriage and margarine.

The claim that health status determines election outcomes may not pass CHI’s sniff test, but the study still underscores the importance and complexity of social determinants of health. Social determinants of health, defined by the World Health Organization as the conditions in which people are born, grow, live, work and age, widely influence a community’s health outcomes.

The model tries to address many of these determinants, but their impact is so far-reaching that others were likely left out. Some of these omitted variables could account for the association between health and voting — for example, high rates of violent crime can impact both life expectancy (one of the outcomes modeled in this study) and responsiveness to Trump, who made crime a central part of his message

To learn more about health and policy in the wake of the election, check out CHI’s ongoing analysis of the impact on Colorado.