The broken leg problem
Our model is pessimistic for Biden. But the reality is probably worse.
This is the first Model Talk column here at Silver Bulletin. Originally our plan had been to update our presidential election forecast once a week through Labor Day (and more frequently thereafter) for paying subscribers. However, we quickly realized there were two problems: 1) we were going to want to update the forecast a lot more often than once a week; 2) it was going to be awkward to bury the numbers in some sort of newsletter post instead of having a dedicated landing page for them.
So we hope this is a case of having underpromised and overdelivered. However, part of our promise was to provide paid subscribers with exclusive written columns in addition to model updates. And that’s what this is: Model Talk is the brand name we’ll use for a weekly column (twice weekly after Labor Day) about the election as told through the lens of the model. Rest assured that there will be plenty of free columns about the election also — like this one from Wednesday — but Model Talk will often err toward the nerdier side of things.
The clinical psychologist Paul E. Meehl published a journal article in 1957, "When shall we use our heads instead of the formula?”, that still contains some of the best practical advice I’ve seen for making predictions in our messy real world. That’s important, because real life usually does get messy instead of resembling the laboratory conditions of academia. The past two elections, for instance, have been highly unusual. There was COVID in 2020. And this year, one of the candidates — President Biden — is more likely than not to drop out by election Day, at least according to prediction markets, a conclusion I find basically sensible.
If Biden does drop out, the model should work fine: we’ll wait what I’m guessing will be about two weeks until there’s enough polling to benchmark how his replacement is faring, and then turn it back on. (Although I’m sure it would get a lot of clicks, I’d think it would be pretty irresponsible to publish a forecast of an alternative candidate’s prospects before then.1) In some ways, though, I’m less worried about what our model would assume about, say, Kamala Harris as the Democratic nominee — and more worried about whether it’s appropriately handling Biden.
The canonical example in Meehl’s article involves a model designed to forecast whether a certain professor will go to the movies that night — but, the professor has just broken his leg, a circumstance that the model doesn’t account for. I suppose the chances of him going to the movies are not precisely zero — maybe the professor could get surgery immediately and limp into the AMC Arlington/Fort Worth Megaplex 24 on crutches. But it’s clearly a material factor that any reasonable researcher would want to account for, even if it’s not in the training data.
This example of an injury is one that hits home as someone who’s made his fair share of sports bets. In the New York Knicks’ disappointing Game 7 loss to the Indiana Pacers in the Eastern Conference semifinals, for instance, it was immediately clear that OG Anunoby — bravely trying to play after a hamstring injury — was at best a shadow of himself and couldn’t really run or jump, and he was removed from the game after only 5 minutes. To have persisted in applying a model that didn’t account for this would have been exceptionally stubborn. And would have resulted in your punting off money: The Knicks eventually lost by 21 points.
Joe Biden is the equivalent of OG Anunoby in Game 7 — except, obviously, it’s a much higher-stakes problem for the country. Even though Biden’s chances have fallen considerably in our forecast — to 28 percent now from 35 percent before the debate — it’s still probably too optimistic. He likely isn’t capable of providing the sort of performance he needs to fully realize his chances of a comeback.