American presidential elections are likely influenced by economic25 and social trends,26 which can drive swing voters from one party to another3 and might increase voter turnout. Not surprisingly, we found that counties with no or slower gains in income were significantly more likely to vote for President Trump in 2016 than Senator McCain in 2008. President Trump’s emphasis on non-Hispanic white voters was reflected by his better performance in counties with smaller increases in ethnic and racial diversity.
Mortality rates among lower-income, rural, non-Hispanic white Americans have been rising, even as they are declining in blacks and Hispanics.27 By the 2016 election, counties won by President Trump had a 7.4% higher age-adjusted death rate than counties won by Secretary Clinton; however, counties in which President Trump’s percentage vote in 2016 was higher than Senator McCain’s percentage vote in 2008 had a 15% higher age-adjusted death rate. Age-adjusted death rates remained a significant correlate of a county’s net Republican percentage vote gain between 2008 and 2016 even when adjusting for the county population, rural-urban status, median family income, educational levels, percentage of population that is non-white, and rates of unemployment and health insurance coverage.
Our findings are generally consistent with recent analyses by Bor8 and by Wasfy and colleagues7 but expand them in important ways. Wasfy et al. reported that voting changes between the 2012 and 2016 US presidential elections were strongly and independently correlated with 2014 county-level health (as measured by a composite of age-adjusted death rate, teen birth rate, violent crime rate, primary care physician/100,000 people, and self-reported survey data on average health care costs; physically unhealthy or mentally unhealthy days in the past 30 days; percent overweight or obese; percent diabetic; and percent with food insecurity) after adjusting for essentially the same socioeconomic and demographic covariates we used.7 Rather than focus on static health predictors, we focused on changes over a 15-year time span. Bor reported that the magnitude of improvement in life expectancy between 1980 and 2014 was inversely correlated with a county’s voting share for President Trump in 2016 compared with Senator McCain in 2008 but that this relationship became non-significant after adjusting for state-wide effects as well as a county’s 2014 urban-rural status, economic measures, educational level, and racial/ethnic composition. He concluded that these geographic and socioeconomic measures are driving changes in both voting patterns and life expectancy. However, he considered these state-level measures at one point in time, not their changes over time. Bilal et al. noted that counties with increasing all-cause mortality in persons age 45–54 years were more likely to vote Democratic in 2008 and 2012 but Republican in 2016, with each 15.2/100,000 increase associated with a 1% swing.9 Our analysis therefore adds important nuance to these excellent prior reports. We show that changes over time in the age-adjusted death rates correlate independently with changes in county-level voting after adjusting for changes in other socioeconomic and demographic measures.
By focusing on county-level changes, we eliminated several potential biases. For example, President Trump targeted populations at higher risk of mortality, but we showed that net declines in important socioeconomic indicators, and not just the levels of the indicators themselves, correlate with mortality and voting behavior. Counties with lower baseline income or higher baseline mortality would likely continue to rank low on these measures over time, and controlling for these factors, as Bor did, could cause their effect to disappear. Our data emphasize the importance of relative improvement or decline in county-level factors on voting behavior. Our emphasis on changes in voting, in death rates, and in all other predictive variables is consistent with a substantial literature showing that a person’s happiness is less dependent on their current status than on recent changes in their status—whether economic, social, or health-related—because people tend to adapt, at least in part, to their new status over time by recalibrating their aspirations and expectations.28, 29
In the USA, life expectancy declined from 78.9 years in 2012 to 78.6 years in 2016. Declines among younger, lower-income, non-Hispanic whites, especially those without a bachelor’s degree, more than offset gains among African-Americans.6, 10, 12, 30, 31 Increases in deaths of despair, which approximately doubled between 2000 and 2015, contributed to this decline in life expectancy.11, 32,33,34 However, a variety of socioeconomic, ethnic, behavioral, and metabolic factors are important,35 as are wealth36, 37 and education.38 Although death rates are not the only marker of health and well-being, they might be a marker of relative despair. If so, it is not surprising that deaths related to alcohol, drugs, and suicides rose by 2.5 times as much in counties with a Republican percentage gain compared with counties with a Democratic percentage gain since 2000.
Health insurance coverage, which might save lives39 but cannot fully offset the effects of lifestyle and poverty, increased in counties that became more Democratic and that had greater reductions in death rates.40 Conversely, counties with higher death rates voted for a Republican presidential candidate whose party promised to repeal the Affordable Care Act. These counter-intuitive findings have been broadly discussed in the sociological literature and may be due to political messaging.
Our analyses have several limitations. First, county-level correlations do not guarantee that individuals with these characteristics changed their votes or preferentially voted in one election or another. Second, available data sources cannot prove a causal chain of events, and voting correlations do not equate with causality.
Rates of “deaths of despair,” which are related to alcohol, drugs, and suicide, are increasing at a time when the sum of all other causes of death are declining. Although President Trump over-performed in counties with higher rates of deaths of despair,41 such deaths represent less than 5% of age-adjusted death rates and, at least in our analyses, do not, in and of themselves, explain a substantial proportion of the voting change. It is possible, however, that a more expansive consideration of all deaths related to drugs, alcohol, depression, and anxiety might find a stronger independent relationship. Unfortunately, the notorious unreliability of death certificates42 makes such estimates problematic.
Our primary analysis, as well as three alternative sensitivity analyses, found that relatively modest incremental reductions in age-adjusted, county-level death rates could hypothetically have swung Michigan, Pennsylvania, and Wisconsin and, hence, the election to Secretary Clinton. Even in our most conservative, primary analysis, these age-adjusted death rate reductions were plausible and potentially obtainable—18/100,000 lower in Michigan (to a rate between Pennsylvania and Montana), 44/100,000 lower in Pennsylvania (to a rate between Oregon and Iowa), and 67/100,000 lower in Wisconsin (to a rate between Minnesota and New York). These findings are consistent with a recent county-level analysis suggesting that Secretary Clinton hypothetically could have carried Michigan if its prevalence of diabetes was 7% lower, Pennsylvania if an additional 8% of its residents engaged in regular physical exercise, and Wisconsin if its rate of heavy drinking was 5% lower.5 Death rates may be important markers of the dissatisfaction, discouragement, hopelessness, and fear of cultural displacement that contributed to President Trump’s appeal, especially to the non-urban, white working class.43