538 model & poll tracker thread

[Alben Barkley]

I actually have a theory on why Nate Silver is dragging his feet-

I'm sure his forecast has Biden somewhere around 90%- he doesn't want to release it now, and risk Trump  making a comeback and winning, and then he'd have to deal with all these idiots who will hammer him about being wrong because they don't understand how stuff like that works- thus he'll wait until either Trump narrows the gap or we get to the point where he has zero chance of winning

If Silver is really doing that then he should be ashamed of himself. Taking criticism is part of his field of work.

He did this already in the primaries. After Super Tuesday it took him FOREVER to release the updated model showing Biden with a 99% chance of winning. And I have no doubt it was because he feared the wrath of the bros even though he was right.

[Alben Barkley]

Good find, I expect you're right.

He did this already in the primaries. After Super Tuesday it took him FOREVER to release the updated model showing Biden with a 99% chance of winning. And I have no doubt it was because he feared the wrath of the bros even though he was right.

What malarkey, he updated within 2 days.
Grow up, outside of high school hanging on to grudges won't get you far.

Not holding on to grudges, that’s simply not the way I remember it. If I’m wrong I’ll happily admit it. But I distinctly recall that at SOME point Silver seemed to hold out on releasing a model update for an abnormal length of time, then when it was released it showed a massive lopsided Biden victory. I at least thought it was right after Super Tuesday. And given he immediately released the updated model post-SC (which somehow showed Biden’s chances slightly decreasing, I remember incredulously), those two days would have felt like a while in comparison.

[Alben Barkley]

Wow... the model is garbage.

Take what happens if we lie to our model and tell it that the election is going to be held today. It spits out that Biden has a 93 percent chance of winning

That’s the relevant number then, Nate. It’s impossible to predict how things are going to change in the next three months (never mind voting starts in some key states in just one month), or if they are even going to change at all (race has overall been steady to this point, no reason to assume it will stop), so the only tangible factors to go on are what the data says now.

In this article — partly as a corrective against what I see as overconfident assessments elsewhere — I’m mostly focused on the reasons why Trump’s chances are higher than they might appear.

The purpose of the model shouldn’t be to contrast with other people’s models, and you shouldn’t deliberately rig it to do so.

I’m honestly getting the sense that Nate made this both to be a contrarian and to make as conservative a model as possible as an overcorrective to 2016. In fact I think he waited so long because it took some effort to force the model to look like this. And I think it’s no coincidence the model was released with the same numbers as the final 2016 forecast.

I think Nate has jumped the shark. Lost focus on the data, too focused instead on punditry and concerns about appearances and intangibles. More worried about hedging his bets in case the unexpected happens again so he doesn’t look too wrong rather than just going by the polls and letting the chips fall where they may. Pretty sad.

In 2016, the reason Trump had a pretty decent chance in our final forecast was mostly just because the polls were fairly close (despite the media narrative to the contrary), close enough that even a modest-sized polling error in the right group of states could be enough to give Trump a victory in the Electoral College.

The uncertainty in our current 2020 forecast, conversely, stems mostly from the fact that there’s still a long way to go until the election.

But that alone shouldn’t account for the odds dropping from 93% to 71%. You admit Biden is doing better than Hillary at her peak, yet you had Hillary in a MUCH better position in your 2016 model at her peak. And the Nowcast never differed this dramatically from the other forecasts. The flawed and frankly baseless assumption that Trump is going to close the gap, when again this has been a remarkably steady race so far and there isn’t even that much time before voting starts, should not be weighed this heavily in the model. There is no guarantee of anything like another Comey letter.

[Alben Barkley]

https://twitter.com/jon_m_rob/status/1293525159056486405
https://twitter.com/jon_m_rob/status/1293525620098555906

Nate is way smarter than me I'm sure, but that cube root function for the model's "uncertainty index" seems very arbitrary and not particularly connected to anything empirical?? Why not base it on actual past data of how predictive polls are x days before an election?

This seems to confirm to me that Nate is using this non-empirical conception of “national drift” to arbitrarily inflate Trump’s chances.

“The race could change!”

Yeah, but it also could not. Or change in Biden’s favor. That should be accounted for at least as much.

[Alben Barkley]

This is dumb.  Nate got triggered too hard by the dopey Twitter extremists who decided to attack him for giving Clinton a 70% chance when she lost.  So many aspects of this presentation seem designed to just cover his ass and make sure nobody can say the same thing again if he's "wrong" this year.  Right down to the fact that it says "Biden wins 71 out of 100" instead of "71%".  He's also obviously including a bunch of nonsense just to try to add more uncertainty so Trump has a better chance, because that's what his gut tells him and that's what he "got wrong" in 2016.

Nate needs to realize that his core audience is not idiots on Twitter.  His core audience is people who look to him to give us a realistic portrayal of the state of the race based on empirical data.  He is supposed to be the counterpoint to all the pundits on TV spouting off about unquantifiable nonsense like "enthusiasm" or "crowd sizes" or "unskewing the polls" or things like "no president has ever won/lost re-election when XYZ is true (a good economy, low approvals, etc.)"  This was illustrated perfectly in 2012 when all the Republicans were working overtime to convince themselves that Romney was actually winning in spite of the polls, and Nate's data-based model got every single state right and predicted the Obama victory pretty much exactly.  Now where do we turn for a purely data-based model?

Furthermore, Nate has really leaned hard into this idea of "uncertainty."  That was the main problem with his primary model as well.  Nate, we just want a snapshot of the race as it appears today based on the hard data available.  If Biden is leading by 20, you shouldn't say that he only has a 65% chance of winning because "anything could happen, there's so much uncertainty."  The model used to just say "if the election was held today, this is what would happen", which is actually useful.

This should realistically be the end of Nate Silver as the premier election forecaster.  His website has been totally silly ever since the ESPN buy-out, and he and his crew spend their entire lives on Twitter.  Ah, for the days when he was "the baseball guy running election simulations for the New York Times."

100000% agree.

I guess The Economist model which Nate attacked so hard is the best thing we’ve got now.

[Alben Barkley]

I do think that uncertainty is especially high at the moment because we're in a polling drought. I'd expect uncertainty to shrink quite a bit if there's a big wave of polls this weekend (which I think is likely as people want to get their pre-convention snapshots to compare with post-convention).

I do really miss the density of state polls we had in the 00s. I remember in the late 2000s, SUSA literally did a "Bush approval in all 50 states" monthly tracker for a while. No pollster would dare touch something like that now. If we had more actual polls of, say, Mississippi, it'd help 538's model get to 99% for Trump and eliminate some of the uncertainty.

Morning Consult at least was doing a Trump approval tracker in all 50 states, but I’m not sure if they’ve updated it in a while.

[Alben Barkley]

If I could offer a counter-point to the "538 overestimates Trump" line, consider: What % probability would you give to the Trump administration pulling something out of a hat like the Comey letter in 2016? For example, Barr arresting Hunter Biden for espionage 10 days before the election (of course, only to release him later). That uncertainty would seem to model such a scenario, even if Nate claims this doesn't model election interference. The venn diagram of interference and black swan events would seem to overlap quite a bit.

Such a thing COULD happen but:

1. That doesn’t mean it’s probable.

2. There is no guarantee voters will react the same way they did in 2016, or if that even would be enough to swing the election given that Biden’s starting from a higher point and that many people in key states will have already voted before any last-minute stunts due to the increased prevalence of early voting this year.

3. Such hypothetical possibilities should not have such an outsized effect on the model, especially when the “uncertainty” cuts both ways AND when Nate claims election interference doesn’t affect the model as you say.

It’s just absolutely absurd that they acknowledge that Biden is doing better than Hillary at her absolute peak, yet are giving him the same odds as Hillary at her low point after the Comey letter. Period. There’s no getting around it.

[Alben Barkley]

I don't mind the model as much as I mind the god awful appearance of it as a whole.

I actually like the winding path of electoral votes that visualises the idea of the tipping point state, and it probably is more effective at showing where the race stands than a map where the two sides are at roughly similar land area even if they are at very different levels of electoral votes.

I like that too. They had something like that in their 2016 model I believe. But as a whole, I just find this model's look to be a little oversaturated. I'm also not a big fan of having to continuously scroll down to view different aspects of the model, I prefer tabs to take you to different pages instead (like in their 2016 one.) 

The 2016 model was better in literally every conceivable way. Most importantly, the structure of the model itself was superior. But also the UI was superior, the overall presentation was superior, and the fact that it had options for "polls-only" and "nowcast" was vastly superior. If they had those now I could just look at them and ignore Silver's Howard Hughes level insanity about measuring New York Times headlines to predict uncertainty.

[Alben Barkley]

Some of y'all are acting like Biden has a 15 point lead and that the election is tomorrow. The topline numbers of the model seem perfectly rational, and uncertainly inherently decreases the closer you get to the election.

1. Biden does have a 15 point lead, at least according to some polls. Fact is his overall lead is massive, historically speaking. Elections like 1984 are comparable. FiveThirtyEight themselves have reported on this before. Assuming this lead will shrink enough for Trump to win is not rational, when it could just as easily expand or stay steady. And there is historical precedent for ALL of those outcomes. Giving undue weight to what right now looks like the LEAST likely outcome -- a large Trump comeback -- is preposterous.

2. The election may not be tomorrow, but it's not in a year either. Allowing "uncertainty" to influence your model to THIS degree when people will start voting in about a month is insane. Never mind that both the polls of the election and of the incumbent's approval rating have been incredibly stable and there is little reason to believe that anything will magically change that all of a sudden. Trump is not a normal incumbent and never has been; people made up their minds about him a long time ago, one way or the other, and nothing is going to change that at the last minute. There are a lot fewer undecideds this year for a reason.

3. There is absolutely NOTHING "rational" about giving Biden the same odds now as you did Hillary on Election Day 2016, after the Comey letter. Biden is performing better than Hillary at her peak, as they also have commented on before. So why are they giving him the odds they gave her at her low point? It makes absolutely no sense whatsoever and nothing anyone can say can make me change my mind about that. I don't care what kind of convoluted, ridiculous, post hoc "reasoning" (apparently involving measuring NYT headlines) Silver used to come up with these numbers -- it is beyond obvious that he deliberately rigged this model to way overcorrect for 2016 and be as conservative as possible.

It is not a coincidence, by the way, that the numbers on this model's release are the exact same as they were on Election Day 2016. It is clear that Silver went out of his way to force the model to spit out those particular numbers, and that may explain why it took so long to release it. Clearly it took quite some time and effort to convince a mathematical model to tell you what is obviously, objectively not true.

[Alben Barkley]

I don’t see how Silver can give Trump an almost 2/3rds chance to win Texas when the polling looks as bad for him as it does. Feels like he’s relying *way* too much on past fundamentals there.

It's actually over 2/3rds. It's the inverse of the overall number -- 71 Trump, 29 Biden.

This is obviously insane. Silver is saying a race that the polls clearly and consistently show is a complete toss-up is as safe for one candidate as the overall race is for the other candidate who has clearly and consistently had much bigger leads in the national polls.

The page says they actually account MORE for factors other than polls in coming to these numbers, which is a horrendous mistake. Texas is a state that's been trending D rapidly. Weighing how it voted in f--king 2012 heavier than you do the polls there today is utterly ludicrous.

[Alben Barkley]

The race is not tightening. A 1 point shift over a month or two, mostly due to outliers like the CNN poll and a series of subpar internet polls, does not change the fact that Biden’s strong leads remain consistent in the highest quality polls. It’s clearly just noise.

Sad to see Nate may have totally lost his touch and become just another hack political pundit trying desperately to concoct a horse race where one does not exist, even if he has to deny the obvious reality right before all of our eyes to do so.

His rambling article published earlier attempting to respond to criticisms of the “uncertainty” in the model did nothing for me except confirm the man is deep down a slide into Howard Hughes la la land. Totally detached from reality, going on about Michael Dukakis and convention bumps and debate pageantry and s—t in a year like this, when COVID is all that matters, the electorate is extremely stable and polarized, and none of that other crap has even really mattered in a while anyway.

As for his not including Trump’s incredibly stable approval ratings as part of the model, a guy in the comments said it better than I could:

The fundamental flaw with the model is that Nate Silvers is considerading the uncertainty "normal" in this election year. That fact alone should be exploding a whole bunch of logic bombs that destroy this model. It just isn't true. There is no reasonable level of uncertainty that exists that could stand to benefit Trump in a significant enough way to make this anything close to a 70-30 race. There just isn't a universe where that can possibly be true.

538 missed by far the most important factor for the uncertainty index. In fact, if applied, it would go a long way to removing a huge amount of uncertainthy. Namely, the Standard diviation (and low baseline) of Trump's approval ratings. Yes, I understood Nate's criticism of using approval ratings, however, there is no President with anything remotely similar in U.S. history. In other words, we have almost four years of "polling" that doesn't move an inch, that's totally unresponsive, even when juxtoposed against historically important breaking news. While Carter didn't lose by 20 points (per Nate's example), Carter also (at one point) had an approval rating of almost 75%. There is a calculus possible even with Carter that doesn't exist with Trump. There was evidence that every other president "could" become more popular when things change. There is no evidence to suggest that Trump could get more popular if news changes. I don't know how you can be a quantitatively inclined model and just ignore that. Who seriously thinks Trump is going to get a convention bounce that brings the race closer than 5 points, even if he gives the best speach of his life? It's perfectly predictable based on approval ratings that he won't get a significant bounce. A model that says he might is just clearly not looking at the world with reasonable glasses.

The point is that every other President (or major nominee) was, at least to some extent, objectively "presidential"... Donald Trump isn't. Don't take that as a partisan statement, just look at your very convenient approval rating charts that you posted. For huge swaths of America, Trump is completely disqualified from holding the office. Trump's approval rating at the highest point in his first term can reasonably be seen as a proxy for how many people are at least open to voting for him again. For Trump, that number is a shocking low 45.5%... let's look as some other presidents. Obama 65.1%,George Bush 81.1%, Clinton 60.9%, H.w. Bush 85.8%, Reagan 68%, Carter 74.9% Ford 71%, Nixon 66%... Trump won the first time because Hillary Clinton's approval rating was basically the same as his.

This is what is wrong with your model. Literally every other president in history at least stood a chance of convincing a majority of voters that they deserved relection since every other president in history had already convinced a majority of voters (at some point) that they had done a good job for at least a portion of the their time in office. 538 completely failed in capturing this reasonably obvious fact.

[Alben Barkley]

https://twitter.com/gelliottmorris/status/1298797791523868673?s=20

Certainly it’s not entirely reconcilable, but one factor to remember is that the conventions were a full month earlier in 2016, which would probably imply less uncertainty about the state of the race on this date in 2016.  But it does seem like Silver is just arbitrarily building more uncertainty into his model this year.

The conventions don’t f—king matter this year. No one’s even watching them. Expecting them to have a real significant impact on the election outcome is dubious even in a normal year, let alone this year. Silver is an idiot if he thinks otherwise. It’s obvious all this “UnCeRtAiNty” BS was just his way of forcing the model to not show what the data clearly shows: Biden is the clear and overwhelming favorite to win this election, far more than Hillary ever was. Period. End of story. The discrepancy between the models then and now is absolutely irreconcilable. Nate may act like a robot sometimes, but he is very human, and therefore very much prone to both flaws and emotional bias. Which it is clear, sadly, that he has allowed to influence how he interprets the data: He doesn’t want to be accused of being “wrong” again if another 2016 takes him by surprise, so he’s hedging his bets against the data. Downright cowardly, if you ask me. And the ongoing petty Twitter feud with a man who has made much more logical, less personal arguments and has a much sounder model is not helping his case at all.

As Charlie Cook said earlier this week, the emperor has no clothes:

https://cookpolitical.com/analysis/national/national-politics/many-are-afraid-say-it-not-close-race

But sadly, Silver is among those pundits too afraid to admit it. And yeah, that’s pretty clearly what he is now: Full-time pundit. The objective statistician is long dead. Silver’s  reputation has been immensely and irreparably damaged in my eyes at least over the course of just a few weeks. I will never take anything Silver says or any projection he makes as seriously as I once did again.

[Alben Barkley]

https://twitter.com/gelliottmorris/status/1298797791523868673?s=20

The model isn't based off national polls, Nate Silver has pointed this out, the model works off state polling, the national polls are only used to apply a trend adjustment so it doesn't matter what the national poll say, the 538 model just operates based off state polls.

For example on September 7 2016, Clinton was projected to win PA by 4% and had a 72% chance of doing so, today Biden is projected to win PA by 4% and has a 71% chance of doing so, if you go state by state, the margins and probability from 2016 are the same in 2020, the only difference is compared to this time in 2016 Trump is doing much worse nationally but not worse in a lot of states.

Biden is polling better than Clinton was at this time in most swing states as well.

Also, even if there wasn’t a big difference in the state polls and even if the national polls don’t affect the model, why would there be such a large discrepancy between the overall odds for who is projected to win?