US elections: Economy or identity?

’It’s the economy stupid!’ is the familiar catchphrase when people talk about predicting or explaining election outcomes. It is attributed to James Carville who led Bill Clinton’s campaign in 1992. It also describes the kneejerk reaction of economists in 2016 when Donald Trump beat Hillary Clinton. Trump supposedly tapped into the economic hardship of blue collar Americans who lost their jobs to globalization and technology, and were experiencing ever increasing income inequality. Economic narratives can be powerful and internally consistent. However, they may have nothing to do with reality.

While economic inequality has increased and Trump got elected this does not necessarily mean that people voted for Trump because of rising economic inequality. The same applies to the loss in manufacturing jobs or stagnating income. Unfortunately, time series information does not allow us to disentangle these causalities. Instead, in this special report we use cross-section information at the time of the 2016 elections to identify which factors were behind Trump’s victory. If economic inequality was behind it then counties with the largest increase in economic inequality should have shown the strongest support for Trump. We look at a range of economic and other factors and let the data speak through a logit-model.

To test the hypothesis that economics affected the outcome of the election, we investigated the relationship between economic variables and the county-level outcome of the 2016 presidential elections. To do this, a logit regression was used with 2612 observations (US counties)[1], with the dependent variable being whether the county voted majority Republican or not (1 or 0). Several first difference variables were considered, as shown in table 1.

[1] There are 3142 counties and county-equivalents in the US, but only 2612 counties were available with all these variables.

In the first-differences[2] economic regression, only 3 variables proved significant: the change in unemployment rate, change in real personal income, and manufacturing employment growth. As unemployment fell during Obama’s second term, the positive regression coefficient shows that counties with a stronger decline were less likely to vote for Trump. In other words, the voters did reward the Democrats for reducing unemployment. Real personal income increased between 2012 and 2016, so the negative regression coefficient shows that counties with a stronger rise in income were less likely to vote for Trump. This effect also shows that voters were rewarding the Democrats for economic improvement. Manufacturing employment growth was negative, so the positive regression coefficient shows that counties with larger losses of manufacturing jobs were less likely to vote for Trump. In other words, voters did not think that he would bring back the vanished manufacturing jobs.

So we can conclude that the economic factors made counties less likely to vote for Trump. In other words, Trump’s victory in the Electoral College cannot be attributed to economics. (Trump lost the popular vote, but economic factors did not make counties more likely to support Trump.) The economic developments during President Obama’s second term that had a significant effect on voter behavior - the decline in unemployment, the rise in real personal income and the decline in manufacturing employment-, all made voters more likely to vote for Hillary Clinton. So we should look at other factors that motivated voters to support Trump and help him to a majority in the Electoral College. In fact, according to the economic model Hillary Clinton would have won the Electoral College if either unemployment had fallen further, real personal income had risen more, or manufacturing jobs loss had been more severe.

[2] A level regression for economic variables was performed, but the results were generally insignificant, especially when combined with the demographic variables, and thus, scrapped for brevity.

While economists were spinning their self-contained narrative in 2016, other social scientists were pointing at demographic, sociological and even psychological variables to explain the Trump vote. To test this alternative narrative, several county-level variables were chosen to again fit a logit model of the Republican winning the county: median age, percent non-Hispanic white, percent male, percent without a college degree, number of drug deaths per 100,000, and percent of the population which is physically active in their leisure time[3]. That last variable is a little out of left field, but there has been recent interest in the so-called yoga vote. Basically, this is a lifestyle variable with strong implications for political preference that transcends the simple urban-rural divide. On the one hand, urbanites walk and cycle to work at a higher incidence, have more access to gyms and exercise classes, while also being more likely to vote Democrat. On the other hand, places which have more opportunities for outdoor leisure, or places that would appeal to so-called ‘crunchy’[4] types would also be included within this lifestyle variable. ‘Crunchy’ people stereotypically do yoga, live all-natural lifestyles, and purposefully spend time in nature. Because of their attachment to nature, they are frequently found in rural areas. However, they tend to vote for the Democrats. Capturing this demographic group by the urban-rural divide would be misplaced, therefore we directly identify them through their level of physical activity.

[3] Religiousness could not be investigated due to lack of recent data.

[4] Crunchy: informal North American adjective meaning politically liberal and environmentally aware (Oxford English Dictionary).

Here, average age, drug deaths, and no college degree are statistically insignificant. As the white population or male population increases, the probability of the Republican candidate being elected increases. This is confirmed by 2016 exit polls for those demographics: 52% of men and 57% of white voters voted for Trump. From that same data, 89% of black voters and 66% of Latino voters voted for Clinton, confirming the fact that an increase in black population also decreases the probability of a Republican winning that county. The insignificance of drug deaths is contrary to the narrative that communities in despair voted against the incumbent party. Additionally, the insignificance of not attaining a college degree is counter to the idea that the poorly-educated voted for Donald Trump. It seems that the significant variables already capture all of the voter behavior. The most interesting variable is the percentage of the physically active population—the more a population exercises in their leisure time, the less likely they are to vote Republican.

Using only demographic variables has a more statistically-significant regression than using only economic variables. The demographic model predicts more counties correctly than the economic model. Knowing the percentages of physically inactive white men living in a county is a better predictor of the Trump vote than changes in the unemployment rate, real personal income and manufacturing employment.

While comparing the fit of the two models reveals the explanatory power of economic versus demographic factors, it may be more insightful to compare the separate variables in a single encompassing model to see which variables are the most significant. Demographics change slowly and linearly. Economic variables capture the relatively rapid changes experienced by electorate. What is the optimal mix of economic and demographic factors in explaining the Trump vote?

Since both economic and demographic variables appear to explain the behavior of US voters, we combine them in an encompassing model. To do this, we look at all first difference economic variables and demographic variables.

For the demographic variables, the addition of economic variables does not change the conclusion: the percentages of white voters, male voters and physically active voters remain the only significant demographic variables. Even the estimates do not change very much.

In contrast, once we account for demographics, a slightly different set of economic variables is significant. Unemployment change and manufacturing employment growth remain significant, but real personal income is replaced by the change in the poverty rate. During the second term of President Obama the poverty rate rose slightly, but this reduced voter support for Donald Trump. Counties with a larger rise in poverty were more likely to vote for Hillary Clinton. So still the economic factors favored Hillary Clinton.

If we select the statistically significant economic and demographic variables we get the model shown in table 5. This model gets more counties right than either the purely economic model or the purely demographic model. Here, 2327 counties out of 2612 possible counties were correctly predicted—89% correct. It also shows that there are both economic and demographic factors that play a statistically significant role. What’s more, the developments in the economic variables during Obama’s second term were actually helpful for Hillary Clinton. In contrast, demographic factors boosted Trump’s vote count, not enough to win the popular vote, but sufficient to win the Electoral College.

After November 2016, dozens of news articles[5] decried the economic dissatisfaction of working-class whites as the reason for the unexpected loss of Hillary Clinton. Dubbed the revenge of the working-class whites, news outlets wagged their collective finger at the Democrat party, chastising them for forgetting their roots, not focusing on economic issues affecting the blue collared, and causing them to lose what was supposed to be an easy win.

In contrast, our results suggest that economic developments were actually benefitting Hillary Clinton (although not enough to get a majority in the Electoral College), while demographic factors explained the Trump vote: he received support from white voters, male voters and physically inactive voters. These demographic results remained after including economic variables. In other words, being white, being a man, being physically inactive, in itself were ‘reasons’ to vote for Trump. While our county-level data have the benefit of covering the entire electorate, they lack the information on what makes people vote on an individual level. Therefore, in order to interpret the results that we found on county-level data, we take a look at several studies that are based on surveys among individual voters.

For example, Hooghe & Dassonneville (2018) looked at a sample of 1000 voters and found that Republican partisanship, conservative ideology, anti-immigrant sentiments, and racist resentments explain the Trump vote. In fact, after controlling for these underlying variables, demographic variables are insignificant. In contrast, they also found that economic evaluation boosted the Clinton vote. Finally, they found that the Trump vote was not a protest vote caused by a lack of political trust or dissatisfaction with democracy.

As another example, Mutz (2018) followed a panel of 1200 voters between 2012 and 2016 and also found that economic hardship did not make people more likely to vote for Trump. Instead, she found that status threat explained the Trump vote. Interestingly, the status threat is twofold. Domestically, the dominant white group feels threatened by other ethnic groups and immigrants. Internationally, this same group feels the global dominance of the US is threatened, in particular by China. These results show that BLM and China are likely to be important factors causing white identity voters to support Trump in 2020. More long term, demographic trends toward majority-minority will continue to fuel this sentiment.

If we go back to our own county-level results, the fact that after taking into account economic factors, white ethnicity and male gender played a decisive role in favor of Trump is in line with the sociological hypothesis tested by Mutz (2018) that white men felt threatened in their social status. The significance of the physical activity variable supports the view that lifestyle plays a role. Taken together, it confirms the view that political preference has become part of a broader sense of identity that includes a certain lifestyle.

If we combine the empirical results from our own county-level analysis with the survey data studies we can conclude that demographic factors are better at explaining the Trump vote than economic factors because identity trumps economy. Trump voters did not vote for him because of the state of the economy, but because of their status in this economy and society in general, and the threat to it.

While our model is helpful in understanding the reasons why people voted for Trump or Clinton in 2016, using it to forecast the 2020 elections would be misleading as this is a very different race.[6] However, the established fact that identity trumps economy does predict that voter support for Trump is more robust than Democrats are hoping. Especially, when it comes to making the effort to vote – as opposed to indicating a preference in an opinion poll – a perceived threat to your identity is a powerful incentive.

[5] However, other articles pointed to identity politics - Americans were not motivated to vote by their pocketbook, but a perceived threat to their dominance.

[6] First, the coefficients for a Trump-Biden race are likely to differ from a Trump-Clinton race. For example, there is statistical evidence that the gender of the candidates did play a role in 2016: see Schaffner, B., MacWilliams, M. & Nteta, T. (2018). Understanding white polarization in the 2016 vote for president: the sobering role of racism and sexism. Political Science Quarterly, vol. 133, no. 01. Also, Biden and Clinton score very differently in terms of personal characteristics as perceived by the electorate. Second, it is unclear which values for the economic explanatory variables we should use. The damage done to the economy by the outbreak of Covid-19 can hardly be fully attributed to the incumbent. Also, the attribution is likely to differ between Republicans and Democrats.

While both narratives appear plausible, the data provide more support for identity than for the economy as the decisive factor in explaining the Trump vote in 2016. The ongoing social unrest of 2020 suggests that four years later identity remains a decisive factor. America has not been this polarized since the nineteenth century. The sorting of the two political parties since the Civil Rights Act that we discussed in United States: Civil unrest explains why identity has become such an important factor in the US elections. In fact, both Democrats and Republicans are playing the identity card, albeit under different names. This also means that economists are misguided if they continue to look at the elections through an economic lens. If identity trumps the economy, economic policies will not help the Democrats against Trumpism. In light of the demographic trends, the polarization in American politics will only increase. Biden’s economic recipe will do little to change that.

Mutz, D. C. (2018). Status threat, not economic hardship, explains the 2016 presidential vote. Proceedings of the National Academy of Sciences of the United States of America, E4330 - E4339.

Hooghe, M., & Dassonneville, R. (2018). Explaining the Trump vote: the effect of racist resentment and anti-immigrant sentiments. American Political Science Association, vol. 51, no. 03. 528–534

Co-author: Olivia Coulson