Hillary Clinton and Donald Trump (along with outside groups) spent a staggering $2.65 billion in their quest to win the White House. The two crisscrossed the country in search of votes and with hopes of energizing partisans. Ads became universal; rallies nightly spectacles mirrored across news stations and social media feeds. But did the billions spent on television ads, long the staple in political marketing, actually boost candidate vote totals? Did the tireless thousands of miles traveled bear fruit on election night?
Probably not, at least when controlling for a state’s partisan tilt and its demographics. In fact, just four variables explain 96 percent of Hillary Clinton’s vote variance across the 50 states and the District of Columbia: A given state’s partisan voter index (PVI) as determined by the Cook Political Report, the state’s GDP per capita, its percent minority, and the percent of the state’s population with at least a bachelor’s degree. These variables track a state’s electoral history and its demography – the measures themselves stand independent of 2016’s torrent of advertisements and other campaign activities.
Regressing Clinton’s vote share on those variables shows that all are significant at the 90 percent confidence level; PVI, percent minority, and percent with at least a bachelor’s degree are all significant at the 99 percent confidence level (see table 1 for complete regression output).
Table 1 shows that as each state’s GDP per capita grows, so does Donald Trump’s vote share (each $1 increase decreases Clinton’s expected vote share by the amount shown, miniscule at the margin but meaningful when speaking about tens of thousands of dollars). Self-evdientily, the more Democratic a state’s been in the past, the higher Clinton’s expected vote share. The database codes Democratic PVIs as negative numbers, so the more the negative number, the more Clinton’s vote share should rise. Also intuitive, the more minorities within a state, the higher Clinton’s vote share — this effect is both statistically significant and profound. Each 1 percentage point increase in a state’s minority population share increases Clinton’s vote share by 0.149 points, around the margin by which she lost Wisconsin, Michigan, and Pennsylvania. Lastly, the regression points to a well-documented shift in electoral coalitions: When controlling for income, each 1 percentage point increase in a state’s college-educated population increased Clinton’s vote share by 0.005 percentage points.
Importantly, swing states do not dominate the residuals, or states whose predicted results deviate from their actual results. In other words, the four aforementioned variables quite accurately explain results in the states bombarded by ads and frequented by the two candidates – had swing states been subject to large errors, it might have implied that ads and candidate visits tinkered at the margins, enough, perhaps, to swing a few percent of voters. Simulating Clinton’s vote share 25,000 times with the above regression formula produces chart 1, which shows that swing state results can be explained largely by underlying forces unaffected by Clinton’s dramatic fundraising and advertising advantage.
The states with the biggest errors are Hawaii and North Dakota, by no means swing states and whose discrepancies can be easily explained. Hawaii, without its native on the ballot, moved towards its historical Democratic mean; North Dakota, an energy-rich state, swarmed to Trump in numbers greater than the simulation predicted because of Trump’s pro-energy, anti-regulation stance. Clinton also underperformed in New Mexico and Colorado, the former because of former New Mexico Gary Johnson’s presence. Colorado is the only swing state error likely attributable to ad and appearance discrepancy (Clinton went off the air in Colorado quite early in the campaign though Trump continued to purchase ad time).
Including ad spending by each candidate and the number of events held in each state by the presidential and vice presidential candidates yields no additional explanatory value. Using Ad Age data on Clinton and Trump spending from April 2015 up through election day, by state and nationally – an admittedly imperfect dataset because it includes primary ad spending and misses some spending from outside groups, though the patterns it shows still offers considerable explanatory value – and the National Popular Vote’s party event tracker (which begins after the conventions), I included ad spending and candidate events in the regression and met statistical insignificance.
Each Democratic dollar spent in a state increased, ever so slightly, Clinton’s vote share; similarly, each event she or Tim Kaine held increased their expected vote share. Republican ad spending and events had the opposite effect. But these findings are not statistically significant, meaning that we cannot say, with a resounding degree of confidence, that the relationship is causal (though it likely is). Table 2 has the complete regression results with ad spending and candidate visits.
Accepting that these relationships are not spurious, or due to chance, the coefficients imply that, all else equal, it would have taken 500 Clinton/Kaine events in a given state to improve their vote share by just 1 percentage point (this assumes Trump/Pence simply ignored that state). Similarly, even if the Trump campaign spent $0 in a certain state, it would still take $100,000,000 worth of ads in a single state to shift expected vote share by a single point.
After $2.6 billion and 399 events, it cannot be said that a year and a half of activity caused the vote to change. What does this mean going forward into a cycle of renewed Democratic activism and looking ahead to 2020, which promises to be no less ferocious or expensive than 2016? Arguably, Donald Trump’s star power and media dominance – his broadcast and cable omnipresence amounted to billions in free media – likely played a role in disseminating his message and creating or expanding the character he has for years cultivated. Perhaps in a more normal year, ads and events would have causally effected voting outcomes. Or perhaps in times of heightened partisanship formed, increasingly, around segregated demographic groups, ads and events might only influence turnout.
One thing seems clear, though: Going forward, campaign practitioners must rethink the accepted playbook. Ads and events might not drive votes, certainly not to the extent portrayed or imagined. Candidates and candidate committees – not to mention the outside groups with wealthy benefactors – must find ways to Make Campaigns Matter Again through useful and effective spending and a willingness to buck the long-ingrained campaign norm.