*takes deep breath* this tweet shows simultaneously why 538’s forecast is not informative to the average person and why it’s a statistically sound model https://twitter.com/natesilver538/status/1325135496561954817

Let me be clear. From purely a statistics perspective, it seems to me that both in 2016 and 2020 FiveThirtyEight produced an election forecast model that fairly accurately captured the uncertainty around who would win the presidency.

HOWEVER, it is also true that when you produce a model where the top line numbers are 70/30 or 90/10, even ppl who are educated in statistics & understand what those forecasts are saying cannot parse those numbers to gain any useful insight into the q we want answered

538’s model will never be able to tell you who will win (barring a model that shows 100% certainty). And that’s what I suspect 99% of people who look at their model are trying to figure out

I hope that people who feel “burned” by trusting the model will actually read up and become more educated around stats, because I think 538 is in a unique position to provide this education and have the talent to evangelize this knowledge.

But I also hope that 538 reflects around the nature of their work and its affect not only on the media’s general consensus about the election but on the way that the average person understands their work.

There is a balancing act here. If 538 were a niche stats blog this would matter a lot less. But the reality is that anyone who is deeply engaged in politics is affected by 538’s model to varying degrees. This then filters down through various media as “conventional wisdom”

So, to circle the square as it were, the fact that this election happened to satisfy the outcome of simulation #22176 of 538’s model proves that the model captured the ultimate outcome.

This is, I’m sure, gratifying to Nate Silver and anyone who worked on the model. Unfortunately this is a post hoc validation of a model that, for —I suspect— a majority of people have a different impression.

Telling those people that they are not understanding what the model is doing is both true and useless. Ppl will always try to figure out the outcome of high stakes events like an election (hell they’ll do it for week 9 NFL games).

We need to do a better job of conveying the possibility of outcomes and we need to focus less on granular probabilities that even trained statisticians cannot parse to a useful degree.

Last tweet of this thread. If it’s helpful at all- I went into the election pretty nervous, not because I didn’t understand the polls or misread them, but because I understood that the only practical info I could gather from them was that if Trump won, it would not be a landslide