Election models megathread

[Senator Incitatus]

https://twitter.com/NateSilver538/status/1542228065123811329
It's confirmed.

[Senator Incitatus]

It is good that they're taking a more systematic approach to polling error:

We assume errors are more correlated from race to race. In presidential races, forecast and polling errors are highly correlated from state to state. This also holds in congressional races, though to a considerably lesser degree since there are different candidates on the ballot in each state. However, our analysis suggests that polling errors are becoming more correlated in congressional races, too. Therefore, we’ve increased the degree to which race-by-race errors are assumed to be correlated in the model.

Not a single mention of response bias though, which suggests they're going full-speed ahead on pretending the polling industry isn't in flames. Not a surprise, given how much motivated reasoning is necessary to keep people relying on statistical models at this point, but still disappointing. Maybe not quite the place for it.

This is also encouraging:

In evaluating fundraising for congressional candidates, the model now places more emphasis on contributions received within the candidate’s state. We found that a dollar received from within a candidate’s state is about five times as valuable as a dollar received outside of it in predicting the eventual election outcome.1 That’s why the model now applies that ratio, multiplying any funds raised from voters within the candidate’s state by 5, when assessing which candidate has the fundraising edge.

[Senator Incitatus]

Worth noting that right now the Democratic lead is extremely slight (less than 50% of events) in all four close states and only slightly better in NH. So if you believe the (a) polling error is still underestimated theory or (b) waves break late theory, the model is still entirely consistent with a Republican wave.

I tend to believe (a) and am skeptical about (b), which the model appears to already account for. The point is, 538 does serviceable work and it's good to have a model to look at even if it isn't perfect, so long as you recognize its limitations.

[Senator Incitatus]

The Governors forecast seems quite favorable to Dems in some places (i.e 91% Whitmer) but considering the fundamentals of the GOP candidates literally getting arrested, Whitmer's strong 2018 performance, and relatively good approval and poll numbers, I could see why a model would output that.

I think the main issue with Michigan may be that the top fundraiser by far and a few others are out of the race. Fundraising still has prospective value as a measure there, since the remaining candidates will have to make up for Craig's missing money, but it understates the retrospective strength of the remaining candidates, one or more of whom would have been receiving those funds if Craig hadn't been in the race.

I would expect Whitmer's numbers to come back down to earth a bit as the Republican candidate(s) establish themselves.

[Senator Incitatus]

The senate model is clearly underestimating Republicans due to reliance on polls, the GOP has outperformed what the polls said in the Senate 4 cycles in a row, 2014, 2016, 2018 and 2020, in all 4 cycles the GOP won more senate seats than most projected, no reason to believe the polls have corrected for the bias.

I know that’s true for 3 of the elections, but are you sure about 2018? I thought the results matched the polls in the Senate then - it was only Democrat hopium that was defeated.

The 2018 senate polls were biased, look at states like Indiana, Missouri, Florida etc, even in many senate races democrats won, they won by smaller margins than the polls said.

https://projects.fivethirtyeight.com/2018-midterm-election-forecast/senate/

GOP did better than the median 538 senate forecast.

Yes: my big takeaway on election night 2018 is that a bunch more seats could have gone Republican if anyone had realized they were competitive. Models overstated the margins big time in some races Democrats won, implying they were completely safe: Ohio (11.6 vs. 6.80), Michigan (11.3 vs. 6.5), West Virginia (7.5 vs. 3.3). And massively understated Blackburn (5.3 vs. 10.80) and Cramer (4.6 vs. 10.80).

Nevada was obviously the big exception among competitive states, as it usually is. Pennsylvania was also a lowball for Casey, though not too shocking since Barletta completely stopped campaigning. One of the few races the model did really nail, New Jersey, is why I expected polling to be accurate there in 2021 and was totally wrong.

[Senator Incitatus]

Yes: my big takeaway on election night 2018 is that a bunch more seats could have gone Republican if anyone had realized they were competitive. Models overstated the margins big time in some races Democrats won, implying they were completely safe: Ohio (11.6 vs. 6.80), Michigan (11.3 vs. 6.5), West Virginia (7.5 vs. 3.3). And massively understated Blackburn (5.3 vs. 10.80) and Cramer (4.6 vs. 10.80).

I understand the frustration at seeing elections written of as uncompetitive finish off close, but I don’t think you can draw the conclusion that they could have swung even further to a Republican win. Midterm turnout in 2018 was extremely high, more like a Presidential election, which was unprecedented. The surge in invisible Republican voters was already banked in the elections finishing closer than projected - that shows how some people were missed by polls or avoided them. But with polarization and turnout so high, there was no reservoir of untapped voters OR swing voters who would have gone the other way that would have pushed a Renacci over the top. To expect that is to count those invisible Republican voters twice - once when the candidates overperformed the polls, and a second time “if they only knew…” but they can only vote once.

I didn’t mean to imply this was more than frustration; it was my present sense impression of the results as they came in. But yes, they could have expanded the baseline by putting more resources into those races, thus making the surprise result even closer. They also could have spent less in safe races like Tennessee.