After Clinton clinched the Democrat nomination, there was considerable concern that many of Sanders’ supporters would stay at home or even vote for Trump rather than cast a vote for Clinton. Indeed, one of the main narratives of the election results is that Clinton paid the price for being an ‘establishment’ candidate at a time when the electorate was demanding radical change. Post-election, Bernie Bros have popped up explaining why they stayed home. But what do the data say?
I use data 2,045 counties from the 2012 and 2016 Presidential Elections results and the 2016 Democrat primaries to investigate trends in voter turnout and voting behaviour. The data is compiled from Dave Leip’s excellent US Election Atlas. A couple of caveats are worth making about the analysis at the outset:
- Even though counties provide a very fine-grained analysis of voting trends, they are still not individual-level data and we hence need to be very careful in imputing individual behaviour to trends in aggregated data. This is known as the ecological fallacy; and
- The analysis is limited only to those states that hold primaries rather than caucuses. This limits the scope of the analysis to 36 states and Washington DC. These states come from across the traditional political spectrum – from Democrat strongholds like California and DC to the Rust Belt swing states like Pennsylvania and Ohio to traditional republican strongholds like Kentucky and West Virginia. Given this, there is no reason to think the patterns here are not skewed by this bias. Because third party candidates play an important role in the analysis, it is worth noting that Utah is not included in the sample so the effects identified here are not skewed by the particular impact that McMullin had in that state.
Did Bernie’s supporters stay home?
Certainly, it is clear that turnout in general suffered in Democrat-strong counties (see Figure 1); counties that had been more strongly Democrat in 2012 saw a greater average drop in votes cast in 2016. Indicatively, the average drop in turnout in counties that had voted more than 75% Democrat in 2012 was 6.1%. In contrast, counties that had voted more than 75% Republican in 2012 saw an average increase in the total number of votes cast of 4.9%.
But no such relationship exists if we focus down on the level of support Sanders’ received in the primaries and the change in voter turnout. In fact, there is a small (although still significant) positive relationship between the level of support for Sanders and the change in the total number of votes cast at the county level. Indeed if we include state-level fixed effects – focusing on the extent to which level of support for Sanders affected the difference in turnout among different counties within states – the effect is even more strongly positive (see Table 1). The interpretation of this estimation is that within a given state, for every extra 10% Sanders received in a county’s primary vote, the number of votes cast at the Presidential election in that county increased by 2.6%.
In interpreting this, though, we do need to be careful of the ecological fallacy. This does not show for sure that a higher proportion of Sanders’ primary supporters voted than the rest of the population. It would be better to draw a regional inference: areas from which Sanders drew more support in the primaries were also areas in which a relatively greater number of people turned out to vote. But it certainly seems safe to reject the alternative negative hypothesis that the Democrats feared: Sanders’ supporters didn’t stay home.
How did Bernie’s supporters vote?
If Sanders’ supporters did go out to vote, how did they vote? In order to look at this, we need to compare county-level votes in 2016 with previous voting levels in 2012. Overall, levels of support for Sanders in the primaries is correlated with a significant decline in the Democrat share of the vote and a significant increase in the share of votes for both Republican and other candidates (including write-ins), but with a noticeable geographical split. In the key Rust Belt states that swung the election, the effect on the major parties was reversed, with levels of support for Sanders positively correlated with an increase in the Democrat vote and negatively correlated with an increase in the Republican vote.
In both broad geographic regions, however, the correlation between support for Sanders in the primary and the change in vote for ‘Other’ candidates remains strong and positive. Indeed, this is one of the clearest patterns to emerge from the data. Figure 2 shows the relationship between Sanders’ primary vote and the absolute share of vote for ‘other’ candidates in the Presidential election, excluding those counties where only Clinton and Trump were on the ballot paper.
Put together, these features suggest a ‘bloody nose’ hypothesis: many of Sanders supporters were motivated to cast a vote for Other candidates or even for Trump to give the Clinton campaign a bloody nose, but only in areas where there was less risk of a Trump victory. If this were indeed the case, we might expect to see different patterns of voting behaviour in areas with different track records in the 2012 election.
To investigate this further, we can run a more sophisticated multivariate analysis in which we look at how support for Sanders in the primaries correlates with change in the number of votes cast for the different parties (including ‘Others’ as a pseudo-party) in areas with different political track records. Table 3 presents two models of this.
In Model 1, we look at the impact of relationship between Sanders’ support and the margin of Democrat victory or loss in 2012 on voting outcomes. The interaction term is the crucial factor here, picking up whether there is a stronger relationship between Sanders’ support and voting outcomes in more heavily traditionally Democrat counties. In this analysis, neither Sanders support nor the 2012 Democrat margin had a significant impact on the Republican vote, but they did have a significant impact on both the Democrat and the Others vote. For both categories, the interaction term is positive and the direct effect of the Democrat margin is insignificant, and the direct effect of Sanders’ primary share is significant but in opposite directions. The interpretation of this is that strong support for Sanders in the primary correlated with a reduced vote for the Democrats, but that this effect was mitigated in more strongly Democrat areas, while support for Sanders correlated with an increase in vote for Other candidates which was even stronger in more strongly Democrat areas.
Model 2 provides an alternative specification that instead of looking at prior performance on a continuous scale divides counties into three categories: Republican Strongholds, with a Republican margin of at least 20% over the Democrats in 2012; Democrat Strongholds, with the converse 20% or greater Democrat margin in 2012; and Marginal Seats, where the victory margin of either party was less than 20% in 2012. We then simply map the impact of Sanders’ primary support on each party’s vote in each category of county.
Here, a much clearer pattern emerges. In Republican Strongholds, support for Sanders correlates with a declining vote for the Democrats and an increasing vote for Republicans, albeit at a much lower level. One interpretation of this, that admittedly risks an ecological fallacy, is that in Republican Strongholds, some Sanders supporters did indeed vote Trump but many more simply stayed at home. In Marginal Counties, there was also a significant decline in Democrat votes correlated with the level of support for Sanders and a concomitant significant increase in support for Trump, but a far larger increase in support for Other candidates. Finally, in Democrat Strongholds, the level of support for Sanders did not affect the Democrat or Republican vote at all but is correlated with a very substantial increase in support for Other candidates.
These results are depicted graphically in Figure 3, keeping turnout at its mean across all counties and focusing on the respective party’s ‘Stronghold’ counties. The dashed blue and red line show the estimated impact of Sanders’ primary votes on the relevant party’s vote in Democrat Strongholds, and it is clear from this that the overall impact was negligible. Both lines remain very close to their 2012 level (ratio=1.0), irrespective of support for Sanders. In contrast, there is a small but marked diversion in votes relative to 2012 between the two major parties in Republican Stronghold counties, with the Democrat vote dropping away from its 2012 level as support for Sanders increases and the Republican vote increasing in its 2012 level, albeit by a smaller amount. Clearly, however, the strongest effect is on the change in votes that went to Others in both categories of county, but particularly in Democrat Stronghold counties. It should be remembered that this substantial effect is of course from a much smaller vote base in 2012.
What can we conclude from this analysis? Being careful to avoid ecological fallacies, it certainly seems clear that the story that Sanders’ supporters stayed at home is at best partially true. While there is some evidence that they may have stayed at home in Republican-dominated counties, this does not hold across the board. Perhaps the safest conclusion to draw is that areas that provided strong support for Sanders in the Democrat primary also provided strong support for anti-Establishment candidates in the Presidential election and that in more traditionally Republican areas this took the form of a boost in votes for Trump and third party candidates, and in more traditionally Democrat areas this took the form of votes for third party candidates in particular.
Did this affect the overall outcome of the election? This is clearly rank speculation in that a Sanders nomination may have brought these voters to the Democrat side but there is no way to estimate how many votes he might have lost the party. Moreover, the effects on major party vote we are talking about here are quite small, if statistically significant. But, as Nate Silver points out, a two percentage point difference in support for Clinton across the board would have seen her take 307 Electoral College votes. Moreover, in several of the key states in the Rust Belt that swung the election – Michigan, Pennsylvania, Wisconsin – Trump won a plurality of the votes but without an absolute majority. Had the Sanders supporters in these key states broken for Clinton instead of third party candidates, she may have carried the day.
 The states included in the analysis are: AK, AL, AZ, CA, CT, DC, DE, FL, GA, IL, IN, KY, LA, MA, MD, MI, MO, MS, MT, NC, NH, NJ, NM, NY, OH, OK, OR, PA, RI, SC, SD, TN, TX, VA, VT, WI, and WV