Andrew:

For me, “doesn’t mean anything” would be enough to break the tie. But your call…

]]>Dalton:

I agree with you 0.1% is consequential. But it’s not really measurable.

]]>0.1% is something like 140,000+ people. Less then total number of votes Trump won Wisconsin, Pennsylvania and Michigan by in 2016 and less then the total number of people who died from COVID-19 in the United States. I think the precision is warranted.

]]>Bjs2:

It’s a tough call. On one hand, 52.6% is too much precision in the sense that this last decimal place doesn’t mean anything. On the other hand, if we just say 53%, then you’ll get this big apparent jump from 53% to 52% if the forecast drops from 52.6 to 52.4. So I could go either way on this one.

]]>Andrew:

Just because you *want* more precision doesn’t mean you *have* more precision.

Don’t be the guy who digs his heels in (the guy you often lament on this blog). If there is a scientific reason why you think you have 1 decimal of precision for a vote share 2 months from now, then make that case. If there isn’t, then you might want to consider updating your display for the forecast.

]]>So insightful. I can only say for myself that you, M and G have a ~tremendous~ amount to offer the public domain on these topics and I think it would benefit public discourse for it to get out. Working with a journalist helps to amplify. So does joining podcasts and other venues. I really hope you, G, and M do more work getting the word out on your perspectives as I think it’s refreshingly transparent and honest in a way that 538 has not been…

]]>Nate Silver actually had a nice little twitter thread about this issue. Basically the idea is that you have to consider both the higher rejection rate, but also the potentially higher turnout with mail in ballots. The idea is that it’s a lot more convenient to turn a mail-in ballot, then it is to go to the polls. So you also have to consider the “rejection rate” of people who decide to go the polls on election day, but run out of gas, or forget, or are too hungover, etc.

]]>I’d add that I for example have little knowledge of American Politics. I for example learned that DC has electoral college votes…

]]>Anon:

I guess it’s ok. I’ve been enjoying working with Elliott and Merlin. It’s kinda Nate’s call on how he wants to interact with us. I was disappointed in the whole Carmelo Anthony thing.

The more general question, maybe, is how journalists and academics can interact. A traditional model is that the academic does the research and the journalist writes about it. Or the academic does the work and the journalists writes about it with a critical eye, Felix Salmon style. A different model is that the journalist and the researcher are the same person: that’s what Nate is doing. Maybe a better way to put this is that the “journalist” and “academic” roles have been erased and replaced by the analyst, who does both. Bill James was a pioneer in this. Finally, there’s the model in which the academics and journalists collaborate, which is what Merlin and I are doing with Elliott. At this point, you might ask, why do Merlin and I need Elliott at all: why would a forecast by two political scientists be improved by a journalist? The immediate answer is that the Economist forecast is Elliott’s baby: he came to us to ask for help. The longer answer is that 3 people are better than 2, and the distinction between academic and journalist is not always so clear. I do a lot of writing, Elliott does a lot of programming, and we both have thought a lot about politics. I’ve found that collaboration almost always makes things better, as long as the collaborators can get along.

Anyway, Nate seems pretty set in his go-it-alone, don’t involve academic researchers approach, and I really like to collaborate, so maybe that’s one reason we’re having difficulty communicating.

Also, unrelatedly, Nate is a public figure and so he suffers from what I’ve called the David Brooks or Paul Krugman problem: he gets so much low-quality criticism from randos on the internet, that he’s developed a way of pattern of ignoring or firing back at criticism, rather than engaging with it directly. It can be hard to have a conversation, public or private, with someone who’s gotten into the habit of considering outside criticism as a nuisance rather than a source of valuable input.

]]>David:

We don’t consider any elections as an outlier, not even 1968 where a third-party candidate won 5 states. We’re just working with the two-party vote.

]]>Has there been studies about whether it is right to treat 1980 as an outlier due to the presence of John Andersen’s 3rd party campaign that fell apart largely right before the election??? I you’re going to include a lot more years, it seems it would be right to look for and assess outliers or to trim the final results some…

]]>Really is a shame for so many reasons that it hasn’t happened, esp. since you’ve written papers together.

]]>I hate twitter.

]]>Anon:

I’m happy to do so but I don’t know how likely this is to happen, as I’ve invited Nate to comment on the blog or to elaborate on statements such as calling MRP the Carmelo Anthony of election forecasting methods, and he hasn’t really followed up on it.

]]>Km:

In this post I wasn’t looking at changes in the forecasts; I was just looking at the forecasts themselves. The Fivethirtyeight forecast interval for Florida has been very wide the whole time, I think.

]]>Bjs12:

Ordinarily I round to the nearest percentage point but I wanted more precision here because I was comparing the two forecasts.

]]>My ‘theory’ is that Trump believes his chances are better if more people vote by mail than by showing up at the polls, if only because the mail ballot introduces multiple possible errors, while most people who show up managed to fill out the little circles. I assume he’d then rather Democratic efforts be aimed at mail, reducing their energy for actual turnout.

]]>“Making predictions, seeing where they look implausible, and using this to improve our modeling”

So, I think about my models in this way as well. But, where is the line between “implausible” and tweaking the model to make it what you already believe to be correct? Like, what about 43% or 44% or 45% or 46% for Biden? Where does implausible stop?

This isn’t really like “implausible” in the sense of the prior predictive checks example of air particulate matter in the visualization paper where implausible was actually technically impossible (air as dense as a neutron star; if I am remembering the example in the paper correctly).

Assuming the model’s prediction for both happening is small (and I imagine it is), that’s a 1-in-8 chance of an outcome which I think most political observers would rate at more like <1-in-100, like the Economist model does. And then 10% is a high enough number that you could go to modern presidential elections and ask how often something like that has happened. It then gets subjective, of course (does Nixon winning 49 states count, or were people expecting that?) but it gives you a comparison point.

]]>Thanks, I’d been meaning to ask what margin the Economist prior is showing.

Also, I agree that it is questionable that raising N by brining in really old data. Assimilating more data only reduces uncertainty if the underlying data you are adding is really representative. The world was pretty different 30 years ago, let alone 100. I think there is just no way around the small N problem (in terms of final outcomes, at least).

]]>Justin — Good point. Our national prior is Biden +6-7 today. Nate says his is much closer to 50/50. I do wonder what it would be if he included presidential approval in his fundamentals model, like we do, or if he didn’t go all the way back to the 1800s with it (which seems irrelevant, a move looking for more data when you could just acknowledge the small-n problems).

]]>The numbers you derived here seem consistent with 538 probabilities. A two-party vote share of 51% and a standard deviation of 3.9% yields a win probability of 60% for Biden (assuming normal distribution). 538 is reporting 62%.

Economist has a point prediction of 52.6% and 2.8% standard deviation, which implies an 82% chance. The Economist is reporting 79%, so again it is consistent.

But what if we apply the point predictions and standard deviations differences one at a time?

If we used 538’s point prediction and Econ’s standard deviation, the probability only increases from 60% to 64%.

If we use Econ’s point prediction with 538’s standard deviation, we 60%->74%.

So while the 40% increase in standard deviation sounds large, I think it is more the differences in the point projections that is the primary driver for the probability difference.

]]>