Thread by @NateSilver538

Basically, the additional uncertainty introduced into an election forecast by COVID-19 falls into two buckets:

1) It means there's a lot of *news* and economic volatility.

2) It could screw with the mechanics of voting, vote-counting or polling in unpredictable ways.

Bucket 1) is easy enough to handle empirically. You can look at whether election cycles with lots of major news developments tend to be associated with more polling volatility and/or polling error, and indeed they do, although note the usual issues with small sample sizes.

Bucket 2) is a lot trickier, I think, since there are no real solid precedents for COVID. (Maybe something like Hurricane Sandy, but that only affected a couple of states.) But there have been too many issues IMO with voting in these late-stage primaries to ignore it.

So we're going to try to account for both types of error. Bucket 2) involves a LOT of guesswork. Technically, it will be based on how much forecast error increases historically when there are big changes in turnout. But to be honest this is just an educated guess.

Because the sample size for drawing this inference is basically n=0, we'll draw this additional error term from a distribution containing rather fat tails. What this means is that *most* of the time, we assume this will have little impact but there's SOME chance it has a BIG one.

We'll also assume that this additional error could either affect all states systematically, or could impact states on a one-off basis.

One upshot is that if I were a campaign, I might play a fairly broad map, in case things get weird in some state that's part of my path to 270.

In terms of the magnitude, we're assuming that the additional error from bucket (2) will, on average, change the outcome in 1-2 states. But again, there will be occasional simulations where it will have a larger effect along with many others where it has little/no effect.

The other philosophical point here is that you have to decide whether your forecast is a conditional forecast (holds true given certain "reasonable" assumptions) or an unconditional one (holds true given "real world" uncertainties and perhaps even some "unknown unknowns").

We try as much as possible to aim for *unconditional* forecasts, but of course there are some limits to this.

We don't account for the chance that one candidate drops out, for instance. Our NBA playoff odds (pre-COVID) didn't account for the season being interrupted by COVID.