House Model Behind the Curtain: Potential Error Sources
Just up front, this will be one of the drier writeups but I wanted to get it on the record in a few places that I haven’t already covered that could end up being sources of error in the model. The last two also cover the, “why are his numbers different than Election Day predictions” for when I do the autopsy write-up and report any numbers afterward. I think this helps understand the inner workings of the model better though and will be different from a future writeup about what’s next with upcoming plans for the 2022 version.
Are Previous Open Seat Flip Winners Getting Screwed?
Quick note: In a previous post, I talked about a group of representatives that flipped hard fought seats last cycle but are now not really being challenged, and why there may be error there. Most of those like MI-8 and ME-2 actually had the distinction of beating incumbents and so will not be included in the following.
A lot of seats that Democrats flipped (and Republicans in MN-1 and MN-8) came from open seats, which the incumbent was not in the general because of retirement or being defeated in the primary. The model doesn’t give a huge effect for incumbents and they only receive a 2% bump, and an open seat is a 3% penalty for the party that is defending it. This is in line with an article I wrote in August 2018 (https://centerforpolitics.org/crystalball/articles/exploring-the-incumbency-advantage/) and honestly really helped. It was one reason my model was quick to point out that a lot of incumbents would lose early on before the conventional wisdom picked up on it and would catch massive losses among those who had “local brands.” The basis for this number was originally derived from these charts in the article that I’ve attached below.
You’ll notice that an open seat to an incumbent of a different party bonus was actually 3% but I have it at only 2% to match the one with more data points. But really this was a chart that absolutely could have and should have been updated after 2018 and wasn’t because I missed it until just now. At this point, I’ll accept the backtesting from before this cycle to show that this has worked fine so far but the small n makes me wonder. Additionally, because the model back adjusts for open seat status, it would take away 3 points from past open seats but only give 2 points to the new incumbent of the other party, essentially depriving them of an incumbent bonus. This group of open seat flippers may be systematically underestimated in polling and I’m keeping an eye out for any change after the election but deciding to detail it here for full transparency.
National Environment
I address this in my first update piece but basically my model is currently operating on Biden state polling predictions plus the generic ballot at about a six or seven point win. I have also taken care for each metric to show what a landslide looks like. And lastly it doesn’t matter too much because in the interactive version in the link you’re supposed to put in different presidential numbers and see what happens.
https://1drv.ms/x/s!AvfJM5SfJyk7hVFaln5R43JkWE79?e=9YWMaa
Buuuuuut basically the biggest source of error will just be getting the states or the whole environment wrong. I’d be fine with it, to be perfectly honest. The model has two parts and the first is to take the statewide and determine the presidential lean in that state. That part could and will definitely see work on it but I’m more concerned about getting that lean right than the topline. The second part is predicting the split ticket levels downballot which have far more extensive work on them. If I get a predicted topline of a state wrong or Biden/Trump outperforms every state by 3 points then that’s just not really what the model was built for. Any post-election analysis is going to focus only on the actual statewide numbers and build out from there though I know from a prediction standpoint that can be frustrating but that’s not exactly the main focus of what I’m trying to find out.
Iowa 4th Types and VA-7 This Cycle
Steve King of Iowa’s 4th district made some news after a consecutive string of racist comments and attention to his opponent, JD Scholten soon followed. The model didn’t yet have ideology in it but considering the district lean and Scholten being left wing, it wouldn’t have picked up on it being closer. Going into Election Day it predicted a 10-point win and was fairly confident it was noise. So I was a bit surprised when King only won by a narrow share, leading by less than 4%. What happened? Well, the model moves a lot for spending and since the last reporting period had ended, Scholten had raised and spend $1.5 million dollars which is shown in the chart below and where the model would have knocked it down to a 6-point projected King win.
My next update will include the 10/14 candidate filings and I’ll be keeping up to date on IE spending but it may still miss late spending and then come back around and be right after the election. I know this looks shady so you’ll have to trust me or you can use the money metric I’ll also include in the final update for spending cash on hand. This gets me to Virginia’s 7th, maybe the one I’ve gotten the most comments on.
On the surface, this race should be close. It’s trending blue but slower than a lot of other places, Spanberger spent a lot last cycle only to win by less than 2%, and GOP groups across the spectrum have come together to try to keep up financial pressure this time around. The first model update the seat was at Likely R. Now after some outside spending, it’s Leans R. And next update, it’s going to be a Tossup or Tilts R. This is because Spanberger has been holding her fire, when she had $2.5 million on hand at the end of October. She has been choosing to spend late and the model won’t reflect that until after the spend. The fundamentals on VA-7 not being a walk in my opinion are extremely strong but the late money is moving this left.
If this race or others are slightly off on the model in favor of who had more cash on hand on election day and then I publish the autopsy that shows less error, it will be because of this effect, not any manipulation of the numbers.
