Forecasting elections is hard. You would think that after 2016, the experts would have some humility and accept that even if Joe Biden is leading in all of the relevant polls, Donald Trump still has a decent chance of being reelected.
Some of the experts have learned that lesson, but others seemingly haven’t. One new analysis from rising-star pollsters Rachel Bitecofer and Sam Epstein calculated that Trump had a 0.5% chance of winning reelection. How did they come up with that?
They used polls and demographic data to model the likelihood of Trump winning a given state. Then, as the authors state, “you can calculate the overall chance of different scenarios occurring by multiplying probabilities.”
To simplify: If Trump, in addition to his shoo-in states, needs to win Texas, Georgia, North Carolina, Ohio, Iowa, Florida, and Wisconsin to win, then you multiply his odds of winning each one to determine his odds of winning all of them. If each of them were a coin flip, a 50/50 state, and Trump had to win all seven of these, he would have a 1-in-128 chance of winning, according to this model.
The biggest problem with this “Multiple Coin Toss” technique, as the authors call it, is that when you are tossing a coin, each toss is independent of the others. When you are running for president, though, not each state moves independently. These days, a candidate who surges or outperforms the polls or models in Iowa is likely to surge or outperform in Ohio and Wisconsin, too.
There are two ways Trump could outperform the model in any given state, and in both cases, he’d be likely to outperform the model in multiple states.
Maybe Trump surges
Maybe Trump is doing poorly now because Main Streets are closed, people are still getting the coronavirus, and unemployment is still 8.4%. Maybe 50 days from now, Main Street will be open, unemployment will be 6.4% and falling, and coronavirus cases will be at a trickle, and a vaccine will be on the way. Maybe Trump dominates the debates, and doesn’t say or do anything horrific. Maybe the foreign policy wins keep coming.
These are not necessarily likely outcomes. Trump would have to do really good work and get lucky for this to happen.
To use a sports analogy, Trump would have to hit five straight free throws.
But the Bitecofer model thinks he has to hit five straight free throws in Texas and then five straight in Florida and so on in every swing state. He doesn’t. Texas, Florida, Wisconsin, Iowa, etc. … are all in the same arena. He’s shooting the free throws for all of them at once.
If Trump surges in one place, he is extremely likely to surge in another place. Trump doesn’t have to make a comeback in every swing state. He has to make a single, national comeback that is big enough to flip most of those swing states.
Maybe the polls/models are wrong
The reason so many of us were wrong in 2016 wasn’t because we said Trump had no chance to win Florida or Pennsylvania or North Carolina or Iowa or Michigan or Wisconsin. We knew the polls could be wrong, and that they were close. But what were the odds that they were wrong enough and in the same direction in all of those states?
Well, they were. And it wasn’t chance. The state polls in 2016 made the same error in all of those states. Namely, they didn’t weight for education.
That is, pollsters don’t just turn their raw results into numbers. If they find Smith beating Jones 50% to 40% among their respondents, but if their sample was 60% female, they will tweak the results. They’ll say: Based on past results, we basically know the electorate will be 52% female, and so we will weight the females in our sample as less and the males as more. If Jones did better among men, then the final published margin might be 48% to 42%.
Well, in 2016, pollsters got too few respondents without college degrees. They didn’t think it mattered, because in past elections, educational attainment wasn’t that big of a predictor of partisan preference. But in 2016, Trump did very well among voters without college degrees, and so that undersample by the pollsters produced an error.
So if, for whatever reason that we can’t ascertain ahead of time, Bitecofer’s model is off in Wisconsin, where she gives Trump a 29% chance of winning, there’s a good chance that her model is off in Iowa and Michigan for the same reasons.
To use another sports analogy, if a baseball team takes the field in the bottom of the ninth, leading by two runs, that team historically has a 92% chance of winning the game.
Bitecofer might argue that the odds of losing a two-run lead two games in a row is 8% of 8%, or less than 1%. But if the closer who blows Game 1 is Edwin Diaz, and he’s the closer in Game 2, then you realize that blowing a save in Game 1 is not independent of blowing a save in Game 2, because you have the same wrench in the machinery in both games.
