Could the polls be wrong
Welcome back to Bournbrook’s regular series on the state of the US presidential race, your impartial guide to the campaign and what you can expect as election date draws near.
We are just a few days away from the biggest televised event of the campaign. On Tuesday, Joe Biden and Donald Trump will face off in the first of three presidential debates, with two more scheduled for October 15th and 22nd respectively. A single vice presidential debate, between Kamala Harris and Mike Pence, will also take place on October 7th.
Now, this is the point where I’m supposed to tell you just how much of an impact the debates could have. Yes, voters have seen a lot of Biden, and even more of Trump, but who knows what could happen once the two finally confront each other on the stage? Perhaps one will make a gaffe so embarrassing that the floor will collapse from under their campaign? Perhaps the other will pin them down so hard with Facts and Logic that voters will finally see through their lies and bluster?
At least that is what I would have told you if I worked for one of the TV networks. The reality is that presidential debates have never mattered that much, and tend to matter less and less with every election cycle. I’ll go into a little more detail below. I will also be covering a subject that often dominates discussion on social media, could the polls be wrong again, and should we trust them? But first…
The state of the race
Election day is still 38 days away, but many are already voting. To be more specific, nearly 900,000 absentee or early votes have been cast so far in the 2020 election. This includes over 230,000 in North Carolina and almost 170,000 in Wisconsin, both of them crucial battleground states. It is important to keep these facts in mind when going over the latest polls. With millions of people already in the process of receiving ballots, and hundreds of thousands voting as we speak, every day that a candidate is substantially down in the polls is a day when they are losing the election in a very real sense.
Speaking of which, as of this morning, Joe Biden leads Donald Trump by 7.2% in the FiveThirtyEight national polling average, with the Democrat recovering half a point since last week. President Trump’s approval remains virtually unchanged, at about 43%. It’s not all bad for Trump, however. The polling in Florida (which is, I should note, now the president’s home state), continues to tighten, with Biden now only ahead by 1.7%. That being said, the Democrat has pulled ahead in Ohio, his first lead in the Buckeye State since August.
The overall picture remains essentially where it was last week, or the week before, or any week going back to April. Biden maintains a healthy lead but Trump is still competitive. As I wrote two weeks ago, a key feature of this election is Joe Biden overperforming Clinton and Obama in historically Republican states, while his lead in traditionally blue states (most notably: Michigan, Pennsylvania and Wisconsin) is hardly insurmountable. In recent election years, you would have expected that, if the race is effectively tied in Texas and Georgia, then the Democrat must be running far ahead in traditional battlegrounds. Not quite this time. What this more or less means is that it wouldn’t take that much for the result to shift from a narrow Trump win to a Biden landslide, or vice versa.
Could the polls be wrong?
Biden maintains a solid polling lead, but just how much can we Trust those polls, anyway? Didn’t they say that Trump couldn’t win in 2016? Aren’t there millions of ‘shy’ Trump voters that are embarrassed to tell pollsters how they really feel? Shouldn’t other metrics, such as voter enthusiasm, take priority? Let’s take these one by one.
Were the polls wrong in 2016? Yes and no. The national polls, which showed an average Clinton lead of 3.2% on election day, were almost exactly right (Clinton won the popular vote by 2.1%, which is well within any margin of error). There were notable problems in state polls. However, it is still important to keep those in perspective.
Take the three states which ultimately put Trump in the White House: Michigan, Pennsylvania and Wisconsin. In Michigan, Clinton led the polls by 3.6% on election day, while Trump won by 0.3%. In Pennsylvania, Clinton led by 2.1% and Trump won by 0.7%. In Wisconsin, Clinton led by 6.5% while Trump carried the state by 0.7%. Only in Wisconsin were the polls off by a considerable margin. In Michigan and Pennsylvania, they showed a race too close to call.
Sidenote: it is important to remember that a polling margin or error refers to the support for each candidate, not the gap between them. For example, let’s say a poll shows Candidate A with 53% and Candidate B with 47%, with a margin of error of 4 points. What this means is that, accounting for the margin of error, Candidate A should be between 49-57% and Candidate B between 43-51%. Even with a 6% gap, the race is very much competitive.
The problem in 2016 was less so with the polls and more with how they were interpreted, both by election analysts and the general public. Casual observers tend to assume that the job of an opinion poll is to say which candidate will win, when it is really more about the plausible range of outcomes.
In 2012, the final polling average in Michigan showed Obama ahead of Romney by 4%, whereas he actually won the state by 9.5%. Therefore, even though the polls correctly predicted an Obama victory and missed a Trump upset, the 2016 polls (3.3% off) were actually more accurate than the 2012 polls (5.5% off). Unfortunately, people tend to complain only when the polls miss the winner, when that is not necessarily their job.
When it comes to how the 2016 polling was (mis) interpreted by professional election analysts, the issue came from assuming polling errors in different states to be independent from each other. Let’s say that Trump had a 30% chance of winning each of Michigan, Pennsylvania and Wisconsin. Now, the odds of three 30% events occurring back to back are fairly slim, which is why you had some forecasts showing Clinton with a 98% chance of winning. However, polling errors are likely to be replicated in demographically similar states, such as the three above (some forecasters, such as FiveThirtyEight, recognised this and gave Trump much better odds).
That’s not to say that there weren’t significant issues with the polls themselves. Most notably, a failure to weight samples by education. Many state polls in 2016 had disproportionate amounts of graduates, which were (and still are) far less likely to vote Republican. This year, far more pollsters are weighting by education. Furthermore, there was a serious lack of high quality polling in states like Wisconsin, which were mistaken to be fairly safe for Clinton. That is certainty not the case this time around.
But what about the so-called ‘shy’ Trump voter effect? Could it be that many Republicans are afraid to tell pollsters their true opinions as a result of a ‘social desirability bias’? In short, there is little to no evidence to support this theory. If shy Trump voters did exist in anything close to significant numbers, we would expect to see two things happen:
Trump performing notably better in online as opposed to live telephone surveys, where they do not have to speak to another person.
Trump overperforming the polling in Democratic states, where ‘social desirability bias’ would be at its most strongest.
An in-depth study by the American Association for Public Opinion Research (AAPOR), found no notable difference between polls with or without a live interviewer during the 2016 election. They also found no evidence that ‘pro-Trump areas were systematically under-represented in national or state polls’. Meanwhile, Trump tended to underperform his polling in heavily Democratic states such as California.
Furthermore, a simple glance at a precinct-level map of the 2016 election results will tell you that most Trump voters live in places full of other Trump voters. In fact, since Democrats are so heavily concentrated in urban areas, we might as well instead be talking about a shy Biden effect. Indeed, the following statements, endlessly regurgitated on social media, are wholly contradictory:
Trump voters are shy to admit their true voting intention
The polls must be wrong because my area is absolutely swimming in Trump/Pence signs.
Finally, what about the so-called ‘enthusiasm gap’? It is true that Trump voters are more excited about their candidate than their Democratic counterparts. However, enthusiasm about a given candidate does not necessarily translate into higher likelihood to vote.
This is especially the case in re-election years, when the race tends to become a referendum on the incumbent. To that effect, while Democrats might not be too enthusiastic about voting for Biden, they are very enthusiastic about voting against Trump. If the ‘enthusiasm gap’ meant much, we would see Trump performing much better among likely voters. At this late point in the race, there isn’t much evidence of that.
Opinion polling is imperfect, polls have been wrong before and will be wrong again. It is entirely possible, if unlikely, that state polling will be off by a similar margin to 2016. It is also just as likely that they will underestimate Biden, as they did Obama in 2012. After all, every poll comes with an indicated margin of error for a reason. However, regardless of their imperfections, they remain the best method for assessing the state of the race at any given time. There is no getting around the fact that, most of the time, the candidate who is ahead in the polls wins.
What of the debates?
To quote John Sides, associate professor of political science at George Washington University... ‘That presidential debates can be “game changers” is a belief almost universally held by political pundits and strategists. Political scientists, however, aren’t so sure’.
Despite the huge amount of preparation candidates do for the debates, and how much the events are hyped up by TV networks (who obviously have a monetary incentive to do so), there is little evidence that they have ever changed the outcome of a presidential race. That is not to say that debates have no impact on the polling. They frequently do. However, post-debate shifts are often a mirage.
In fact, if the previous two elections are of any indication, the most predictive week in presidential polling is now, just before the first presidential debate. In 2016, Clinton led nationally by 2.3% on the eve of the debates. The final polling average had her ahead by 3.2% and she won the popular vote by 2.1%. In 2012, Obama led by 3.1% before the first debate. His lead dropped to 0.7% by election day and he ultimately won by 3.9% nationally. In an average election year, the race will ultimately revert to where it was just before the first debate.
Debates are likely to have a greater impact when more voters are undecided. To that effect, the 2020 election has among the fewest undecided voters ever at this point in the race. Meanwhile, factors that typically reduce uncertainly, such as high polarization, will likewise reduce the effect which the debates could have. When most voters are as committed to their candidate as they are now, and as averse to their opponent, there is little either can say or do to change their mind.
Put it another way, an environment in which the debates can create a meaningful shift in the race is one where the polling is relatively close, many voters remain undecided and polarization is low. In every case, the opposite is now true. If Donald Trump is able to win re-election, it probably won’t be because of what goes down on Tuesday.
