538 Democratic primary model is up

[Amenhotep Bakari-Sellers]

Biden should not be the nominee

[GeorgiaModerate]

Nate Silver with an interesting article exploring the model's current results.

https://fivethirtyeight.com/features/biden-is-the-front-runner-but-there-is-no-clear-favorite/

[Mr. Morden]

Nate Silver with an interesting article exploring the model's current results.

https://fivethirtyeight.com/features/biden-is-the-front-runner-but-there-is-no-clear-favorite/

Wow, from that first table he posts, it looks like he's saying that, in all likelihood, for any one of the top 4 candidates, if they win Iowa, then the model will most likely have them as the favorite to win a national delegate plurality (if not majority) the very next day.  (Though I'm unclear on whether the model does anything to account for an Iowa "split decision", where one candidate wins on first preferences and another wins the final state delegate equivalent # post-reallocation.)

That makes me even more eager to see how the model moves later today in the event that the Selzer poll comes out and shows weaker than expected #s for Biden in Iowa.

[Fmr. Gov. NickG]

From this article, it looks like the model is placing way too much importance on who wins Iowa.  Biden is 80% to win an absolute majority if he wins Iowa, but only 20% if he loses.  Sanders is 61% to win an absolute majority if he wins Iowa, but only 8% if he loses.  I really don't think Iowa has been that overwhelmingly pivotal in past primaries, with the exception of 2004.

[Mr. Morden]

From this article, it looks like the model is placing way too much importance on who wins Iowa.  Biden is 80% to win an absolute majority if he wins Iowa, but only 20% if he loses.  Sanders is 61% to win an absolute majority if he wins Iowa, but only 8% if he loses.  I really don't think Iowa has been that overwhelmingly pivotal in past primaries, with the exception of 2004.

It's unclear to me how much is correlation vs. causation.  A world in which Sanders, for example, overperforms in Iowa is also one in which he's likely to overperform nationally.  So how much is it that Iowa would *cause* him to win the nomination vs. it just being the case that him winning Iowa is likely to happen in a world where he's been gaining in polls everywhere between now and a few weeks from now?

[Skye]

From this article, it looks like the model is placing way too much importance on who wins Iowa.  Biden is 80% to win an absolute majority if he wins Iowa, but only 20% if he loses.  Sanders is 61% to win an absolute majority if he wins Iowa, but only 8% if he loses.  I really don't think Iowa has been that overwhelmingly pivotal in past primaries, with the exception of 2004.

Let's be real, if Biden wins IA, it would be hard for him to lose the nomination.

[GeorgiaModerate]

From this article, it looks like the model is placing way too much importance on who wins Iowa.  Biden is 80% to win an absolute majority if he wins Iowa, but only 20% if he loses.  Sanders is 61% to win an absolute majority if he wins Iowa, but only 8% if he loses.  I really don't think Iowa has been that overwhelmingly pivotal in past primaries, with the exception of 2004.

Let's be real, if Biden wins IA, it would be hard for him to lose the nomination.

True, but the reverse is not so clear-cut.  It's still very feasible for him to win even if he loses Iowa, so the model's 20% chance for him in that case seems way too low.

[The Mikado]

From this article, it looks like the model is placing way too much importance on who wins Iowa.  Biden is 80% to win an absolute majority if he wins Iowa, but only 20% if he loses.  Sanders is 61% to win an absolute majority if he wins Iowa, but only 8% if he loses.  I really don't think Iowa has been that overwhelmingly pivotal in past primaries, with the exception of 2004.

Let's be real, if Biden wins IA, it would be hard for him to lose the nomination.

True, but the reverse is not so clear-cut.  It's still very feasible for him to win even if he loses Iowa, so the model's 20% chance for him in that case seems way too low.

If Buttigieg wins Iowa, Biden would still be a heavy favorite. If Sanders wins Iowa, we have a real race on our hands.

[Beefalow and the Consumer]

https://projects.fivethirtyeight.com/2020-primary-forecast/

Current chances of winning >50% of pledged delegates:

Biden 42%
Sanders 22%
No one (contested convention) 13%
Warren 12%
Buttigieg 9%
All others 2%

It also includes individual state forecasts.

13% seems really high for a contested convention.

2% seems really low for a field that includes a slew of potential sleepers.

[GeorgiaModerate]

https://twitter.com/NateSilver538/status/1215784163414020098

[Figueira]

Sanders just pulled ahead in Iowa, Nevada, California, Colorado, Utah, and Washington. Also Warren is now leading in Massachusetts.

[Amenhotep Bakari-Sellers]

As I stated many times on this forum, Bernie is directly tied to Bloomberg, and Bloomberg's ad about going after corruption, with a positive message, is helping Bernie, with Minorities and females. Having Jill Biden sidelined, since the Hunter Biden story broke doesnt help Biden, whom have lost support, like Trump, among women

[redjohn]

Worth noting that Sanders was never ahead (or even close) in 538's model in 2016. At this time in 2016, he was at a 16% chance of winning Iowa (to Clinton's 68%).

[Pouring Rain and Blairing Music]

I will preference this by saying that I am not an expert statistician.

I have dabbled in statistics as part of my job and have some passing familiarity with zero-inflated systems. Reporting the average percentage from this distribution of data is really stupid in my opinion.


Screenshot from 538’s model of Biden’s expected performance in Mississippi with my note.

Now a zero-inflated Poisson model isn’t exactly what you would use here, but it shares the concept of having a large number of zeros in the dataset along with a different distribution of non-zero numbers. Without getting too much into my job, this type of model can be used to model whether a given habitat is favorable, and whether it’s more favorable for single animals or groups of animals of that same species. In this case, we are looking at whether Biden will be in the race; if he is, we then want to know his percentage (this is admittedly different from how I use this concept at work).

Ideally, they would report something like the likelihood of Biden staying in the race in a state as well the average percentage he would receive if he is still in the race. Granted, this is a clunky way of reporting it, but it’s more useful. If Biden is still in the race, we’d expect him to get a given percentage (roughly 55%, eyeballing the chart), and if he’s out, we would expect him to get zero (or very close). Saying that he will get 38.7% is bet-hedging at its finest and produces a relatively meaningless result despite being the average. I would also not use the average, but rather the median of the “Biden is still in” results. It looks like they might have actually used the median, but I can’t tell for sure from the graph.

I really do appreciate 538’s modeling and the effort they’ve put in to it. I just don’t think that reporting the average of all the simulations is a really meaningful number.

[Florida Man for Crime]

I will preference this by saying that I am not an expert statistician.

I have dabbled in statistics as part of my job and have some passing familiarity with zero-inflated systems. Reporting the average percentage from this distribution of data is really stupid in my opinion.


Screenshot from 538’s model of Biden’s expected performance in Mississippi with my note.

Now a zero-inflated Poisson model isn’t exactly what you would use here, but it shares the concept of having a large number of zeros in the dataset along with a different distribution of non-zero numbers. Without getting too much into my job, this type of model can be used to model whether a given habitat is favorable, and whether it’s more favorable for single animals or groups of animals of that same species. In this case, we are looking at whether Biden will be in the race; if he is, we then want to know his percentage (this is admittedly different from how I use this concept at work).

Ideally, they would report something like the likelihood of Biden staying in the race in a state as well the average percentage he would receive if he is still in the race. Granted, this is a clunky way of reporting it, but it’s more useful. If Biden is still in the race, we’d expect him to get a given percentage (roughly 55%, eyeballing the chart), and if he’s out, we would expect him to get zero (or very close). Saying that he will get 38.7% is bet-hedging at its finest and produces a relatively meaningless result despite being the average. I would also not use the average, but rather the median of the “Biden is still in” results. It looks like they might have actually used the median, but I can’t tell for sure from the graph.

I really do appreciate 538’s modeling and the effort they’ve put in to it. I just don’t think that reporting the average of all the simulations is a really meaningful number.

This is a good point, however...

1) By reporting the averages, they get numbers that should add up basically to (or close to) 100% in total for all the candidates. That makes it seem more familiar to people. If you cut off all the 0s and reported just the conditional averages given that candidates had not dropped out of the race, you would get something like Biden 45, Sanders 35, Warren 25 Buttigieg 20, etc... which adds up to 125. This would lead casual observers who would confuse the conditional averages with expected vote shares to expect every candidate to get higher vote shares than they really should.

2) In general, any single point estimate that they could use is going to be pretty damn misleading given the absurdly large range/variance they have in their simulation results. The model is essentially saying at the moment "we have no idea what is going to happen, there are a huge range of things that could plausibly happen." They have Biden ranging in reasonable #s of simulations from like 10% to 90%, for crying out loud. That is just an absurdly huge range. At least at this stage of things, it would save time to simply say "we don't know." Coming up with any single number to represent that huge range is an inherently impossible and misleading task.

[Pouring Rain and Blairing Music]

I will preference this by saying that I am not an expert statistician.

I have dabbled in statistics as part of my job and have some passing familiarity with zero-inflated systems. Reporting the average percentage from this distribution of data is really stupid in my opinion.


Screenshot from 538’s model of Biden’s expected performance in Mississippi with my note.

Now a zero-inflated Poisson model isn’t exactly what you would use here, but it shares the concept of having a large number of zeros in the dataset along with a different distribution of non-zero numbers. Without getting too much into my job, this type of model can be used to model whether a given habitat is favorable, and whether it’s more favorable for single animals or groups of animals of that same species. In this case, we are looking at whether Biden will be in the race; if he is, we then want to know his percentage (this is admittedly different from how I use this concept at work).

Ideally, they would report something like the likelihood of Biden staying in the race in a state as well the average percentage he would receive if he is still in the race. Granted, this is a clunky way of reporting it, but it’s more useful. If Biden is still in the race, we’d expect him to get a given percentage (roughly 55%, eyeballing the chart), and if he’s out, we would expect him to get zero (or very close). Saying that he will get 38.7% is bet-hedging at its finest and produces a relatively meaningless result despite being the average. I would also not use the average, but rather the median of the “Biden is still in” results. It looks like they might have actually used the median, but I can’t tell for sure from the graph.

I really do appreciate 538’s modeling and the effort they’ve put in to it. I just don’t think that reporting the average of all the simulations is a really meaningful number.

This is a good point, however...

1) By reporting the averages, they get numbers that should add up basically to (or close to) 100% in total for all the candidates. That makes it seem more familiar to people. If you cut off all the 0s and reported just the conditional averages given that candidates had not dropped out of the race, you would get something like Biden 45, Sanders 35, Warren 25 Buttigieg 20, etc... which adds up to 125. This would lead casual observers who would confuse the conditional averages with expected vote shares to expect every candidate to get higher vote shares than they really should.

2) In general, any single point estimate that they could use is going to be pretty damn misleading given the absurdly large range/variance they have in their simulation results. The model is essentially saying at the moment "we have no idea what is going to happen, there are a huge range of things that could plausibly happen." They have Biden ranging in reasonable #s of simulations from like 10% to 90%, for crying out loud. That is just an absurdly huge range. At least at this stage of things, it would save time to simply say "we don't know." Coming up with any single number to represent that huge range is an inherently impossible and misleading task.

1. That is a good point and probably why they don’t do it that way. It’s really hard to present this data in a way that the ordinary person can easily understand. Granted their reporting of the 80% credible intervals is not really informative either. 🤷🏻‍♂️

2. Their ranges seem to be really large. I’m not sure if they’re reporting all of the simulations or if they’re only showing the 95% (or whatever %) credible intervals of the distributions. When reporting a set of simulations like this, I would say that reporting the credible intervals is imperative. As you point out and is apparent in the graph, the numbers are really wide, which seems strange to me given that the screenshot is of the Mississippi predictions. That’s a state that I would expect to be a lot more clear cut and tighter around the median (zeros aside).

2. (cont.) There really is a wild amount of variance here. Some of the states are almost meaningless (at least looking at Biden’s vote prediction) because it’s just a flat bar across 10-70%. The more I look at these, the more I think, “why bother?” Like, that’s just a ridiculous amount of variance. Maybe trim down the outliers, report the 85 or 90% credible intervals. pre post edit: Jeez, they’re reporting 80% and it’s still this bad. But yeah, it’s almost worthless, like you mentioned. Pretty much any sensible person who follows politics could tell you that Biden will win Mississippi if he’s still in the race given current polling.

Iowa is a mess in the polling given the 4 way tie, so it’s basically garbage in, garbage out. I’m kinda sad I looked this deeply because it looks like cool data, but it’s effectively useless. Punditry would be more useful than this model.

[Mr. Morden]

I might have missed this, but in the explanation of how the model is built, is it explained exactly what they're putting into the demographic regression to predict who's leading in states that haven't been polled?  Are they using the topline #s from existing state polls (e.g., Biden leads by this much in the Deep South, which has a lot of black voters, so must be doing this well with black voters overall, etc.), or are they using the demographic crosstabs of polls as well (which means national polls could also be included)?

[Sorenroy]

I know this is no longer the map, but, to the same basic point of Fubart/DCS, it feels like 538's strive for an average leads to a scenario that has a zero percent chance of happening. Is there really a world in which, after winning 44 of the 46 state races, Biden isn't even hitting 40% in any of the final states (barely in DC)? The strive for "he's either dropped out or is at 70%" means that we get a projection of a future that has no chance of happening. Maybe the model clears up after a few of the early races, but it feels unusable at this point for specific numbers.

[junior chįmp]

More junk from 538.

[jaichind]

I am surprised given the high upward, even though very unlikely, potential for a Bloomberg surge Super Tuesday followed by sweeping other states that Bloomberg is nowhere in this model.

[ElectionsGuy]

I know this is no longer the map, but, to the same basic point of Fubart/DCS, it feels like 538's strive for an average leads to a scenario that has a zero percent chance of happening. Is there really a world in which, after winning 44 of the 46 state races, Biden isn't even hitting 40% in any of the final states (barely in DC)? The strive for "he's either dropped out or is at 70%" means that we get a projection of a future that has no chance of happening. Maybe the model clears up after a few of the early races, but it feels unusable at this point for specific numbers.

Essentially, it's only usable for the next immediate [set of] primaries. The farther away from the contest, the increasingly useless the model is. Given how much the model adjusted for just one poll out of Iowa, this thing is going to have massive adjustments to actual results.

[Mr. Morden]

I was bored, so I spent the last ~45 minutes or so doing something silly, just “for fun”.  I made a spreadsheet of the current version of the 538 model, and its “average predicted popular vote %” for each of these candidates: Biden, Sanders, Warren, Buttigieg, Bloomberg, Klobuchar, Steyer, Yang, and Gabbard (the other candidates are projected to get such a small vote share that they don’t matter), but just for the 10 states with the highest 2016 Dem. primary vote count #s.  (Doing all 50 states would be too time consuming, and restricting it to the top 10 states does get you ~60% of the vote or so.)

If we assume that the 2020 Dem. primary vote turnout share from each state is the same as it was in 2016 for these 10 states, and that the other 40 states have basically the same vote split, then this is what the current 538 model predicts as the national popular vote in the "average" case:

Biden 32%
Sanders 24%
Warren 18%
Buttigieg 13%
Bloomberg 5%
Klobuchar 3%
Steyer 2%
Yang 2%
Gabbard 1%

And I guess that translates into something like this for pledged delegates:

Biden 39%
Sanders 27%
Warren 17%
Buttigieg 13%
everyone else combined 4%

That’s probably a little optimistic for Biden because California will probably account for a larger vote share this time since it’s on Super Tuesday instead of June, and California is one of Biden’s weaker states in the model, but close enough.

I guess this result makes sense, given how we’re told the model works: The popular vote share vaguely resembles what we see in the national polls, except with a bit of a boost for Biden, Sanders, and Buttigieg, since those three are currently the top 3 in IA and NH, and the model assumes that whoever wins those early states will get a momentum boost going forward.

And then in terms of translating into delegates, Bloomberg and other single digit candidates aren’t projected to get as many as their vote share would indicate, since they’ll be below 15% in most congressional districts.  Whereas Biden would be above 15% basically everywhere, so his delegate share would beat his popular vote share.

[jaichind]

Given where Bloomberg is in the polls this model should really include him if they are going to include Buttigieg.

[Lord Halifax]

Given where Bloomberg is in the polls this model should really include him if they are going to include Buttigieg.

Yes, he has a more realistic/less unrealistic path to the nomination than Buttigieg.

[Vaccinated Russian Bear]

Given where Bloomberg is in the polls this model should really include him if they are going to include Buttigieg.

He is included. The model just thinks he has no path.



All others
<1 in 100
(0.5%)