Polls-only is now higher than the Nowcast, I believe for the first time.
This happened briefly yesterday too. I was going to note it, but it switched back an hour later. Right now the difference is just noise (0.1%), but I wonder if there’s something that could cause them to truly diverge this close to the election.
It’s an indication that the models have converged. Models measuring the same value will bounce around that value due to noise.
I see Polls-Plus is now between two and three percent of the other two. Also expected behavior.
Not necessarily. It depends on the justifiable level of confidence one has about how the undecided vote will behave and how sure the lead margin is. Problem is by Wang’s stated level of confidence 99% is crazy. 97% maybe.
Here’s the PEC take on undecideds:quote]… Data from SurveyMonkey suggests that Johnson supporters break about evenly between Clinton and Trump, while Stein supporters tilt strongly toward Clinton. This is consistent with many state polls that show Clinton doing the same or slightly better when the matchup is Clinton/Trump compared with Clinton/Trump/Johnson/Stein. So the net expected effect is, on average, slightly toward Clinton*. …
… The overall combined change is a net increase of margin in Clinton’s favor by 0.7 +/- 2.3%. Given current conditions and based on the uncertainty of 2.3%, in 3 out of 100 cases enough voters would switch to Trump to close his current deficit.
[/quote]
So by his analysis the undecideds are more likely to increase Clinton’s margin than to decrease it but in 3% of case they could go the other way enough to close his (then) current deficit.
Silver’s probably right on this. Trump’s chance of winning has long been dependent on some big surprise coming up in his favor. With every day that passes without that surprise, his chances go down. And when a big day like a debate passes without that surprise, they go down a lot.
That said, the possibility of that surprise is still out there. And given the data that we have, I don’t think it’s possible for any honest model to truly be 99% confident of the outcome. We’ve only had about 50 presidential elections, so right there, that’s a sign that we can’t be more than 98% confident. And it’s even less if you set some historical benchmark and count only “modern” elections, on the theory that they used to be too much different from what they are now.
Nate Silver’s most recent little article on his 538 site speaks directly to the OP:
[QUOTE=Nate Silver]
Wait — I hear you saying — didn’t you just tell me that there’s an 87 percent chance Clinton will win? That doesn’t sound like all that much uncertainty.
Well, it depends on your tolerance for risk, I suppose. Donald Trump’s chances in our model are about 1 in 7, and you’d expect a candidate with a 1-in-7 chance to win about one presidential election every 28 years.1 So while it would be a rare occurrence, we’re not quite in once-in-a-lifetime territory.
[/quote]
On the first paragraph I agree that’s what he said and would add specifically the possibility of a collective* poll error which Silver specifically mentioned. That’s an undefined probability that’s pretty much constant till you see if it’s true or not at the election. It’s also tempered though by the parallel discussion of what advantage Clinton has in literal day of the election GOTV** that by definition can’t show up in polls, as opposed to stuff which plausibly should show up but might not.
On second I agree with the general idea but the lack of data to verify particular %'s likelihoods is much worse than that. 50 elections would still be way too few if they all hadn’t ended with 98%/2% predictions. All the ones which ended up 50-50 or 60-40 wouldn’t directly verify the accuracy of the model in the tail of the distribution. It would still be two few to be highly confident of the particular probability predicted even the actual result was 2 out of 50 wins just in cases where a 98% prediction was made.
Trump doesn’t have a very good chance if the herd of consensus pollsters puts him down by 5+ (5.3 including Rasmussen and IBD/TIPP) this close. Beyond that who knows the actual probability? IMO.
*or everyone but Rasmussen and IBD/TIPP; the LA times polls as often discussed has a methodology which arguably puts it in a different category.
**the general effect of a ‘ground game’ should show up before the election, and some votes are cast before the election of course. But the effect of literally offering a ride to the polls election day can’t itself be reflected in a poll the day before. How big is it? The consensus here in a poll was 2 or more % points, AFAIK most pro’s would say 1% or less, and Obama’s 2+% overperformance of RCP avg in 2012 probably wasn’t mostly due to literal GOTV.
And what happens if Trump decides to punish taintstream Republicans for their treachery and disloyalty? What if he instructs his onions to vote for him at the top of the ticket, but ignore the rest of the Pubbies downticket?
He has already made such noises, what if that is the best way to get more attention? Then the Republicans may get a respectable loss on the Prez, but get a jolly good Rogering downticket?
Groovy.
Yeah, even Silver commented on that in the Friday podcast. He didn’t mention the PEC by name, but said that any model that gives that high certainty is overweighting the most recent polls, and not taking a lot of things into account, like polling accuracy.
Which is fair–that’s what the PEC does. It presumes the polls are accurate. What it does is take a current aggregate and then look at how much things have changed throughout the election to predict how much it will change at the end. He then offers two extra maps that say the polling is off by 2% either way.
That said, it sure seems like most people are talking like the likelihood of Trump winning is less than 1 in 7. Including Nate himself.
You could still look at the margin in close races to calibrate the models. If a model predicts that A will win by 5, and A instead wins by 15, that’s as meaningful as if the model predicts A by 10, and the result is B in a squeaker.
Indeed, Sam Wang says that the reason he doesn’t give the equivalent of the “Now Cast” is that on any given day, the chances that one particular candidate will win is always greater than 99%. It’s whichever candidate is ahead in the polls on that day. He assumes, but doesn’t seem to state explicitly, that the polls are always accurate in aggregate.
Can you point to where he says this? It’s… hard to believe.
Here:
http://election.princeton.edu/faq/
Specifically, this section:
My mistake, his figure is 95%, not 99%,
I think the portion being thought of is here -
A related bit is in a post about why he handles covariance much more simply than 538’s
Not exactly the “99%” recalled above but he does state that the if-today probability would be “typically greater than 95%”.
Wang’s take on pollster accuracy is that inaccuracies of individual houses cancel each other out given a large number of polls. (He does not state this but such a concept is used to great utility in brain sciences and is the basic principle that allows for things like auditory evoked responses - send a click via the ear and list for the brain response and you see anything for all the background noise but repeat it 500 times and average the responses and you see the true signal as all the noise cancels out with signal averaging.)
His take on past polls is that they form the best basis for the Bayesian prior for where future polls may go but not for the what would happen if the election was held today - most recent polls do that.
ETA crossposted! But the big differences between the two sites is that Wang tries to keep it as simple as possible and Silver adds complexity even if it adds little to performance. The net result is that Silver’s predictions are much more volatile and his certainty lower.
The most likely outcome in points is again looking at the middle of a distribution. I think it might be feasible to verify the general shape of the middle of the distribution if you really had 50 trials*, but not the tails IMO, which is what it’s about when people are arguing about 17% or 2% likelihood Trump wins now.
*of similar elections in a similar country, the real total is way less than that, the past is a different country and local or state elections are different than national. There’s more like half a dozen previous examples which really look much like today’s politics. Even the swings that happened during Reagan/Carter or Bush/Dukakis may just not happen anymore.
Bottom line is that it is impossible to avoid having some punditry when one determines the uncertainty distribution especially in predicting not what would happen today but at some time in the future. How certain should we be that say the last six presidential elections, provide conclusive evidence of how much final polling will vary from current polling and how much do we say that it informs not at all? Where in between? The answers cannot be completely data-based even if the arguments for accepting one or the other include data.
The answers are also unavoidably chosen with what your goals are in mind. Wang attempts to create a model that is transparent and as stable as possible. He’d rather miss catching a trend early than overreact to noise. Silver does not want to miss a trend and does not mind potentially overreacting to noise. As an illustration Wang decides to accept that undecideds will behave as they have in the past several elections and calculates with that in mind and Silver assumes there is a greater chance that they may do something that they have not done before.
Wang’s lack of uncertainty is, I agree, excessive (and Silvers high level of uncertanty is as well IMHO) but to his credit he is clear as to where it comes from, based on making certain explicit assumptions for the sake of creating the model. If we as consumers believe that the assumptions are less than that which can be completely assumed then we can and should add some uncertainty to his numbers as a result
Meanwhile we can at least comment on this n of one as to which Bayesian prior was the better anchor. Wang’s Bayesian prior was his longer term meta-margin of about Clinton +4 (and it is now 4.8%), hence his Bayesian prediction never moved very Trumpward. Silver’s prior (PollsPlus)was mostly based on the econometrics that would have predicted a close election, thus he never moved as far away from 50-50 share of the vote and when polls moved closer his prediction moreso became nearer to 50-50.
In this one case, to date at least, at least econometrics provided a less accurate anchor than did long term polling data.
I disagree. The event we’re looking for is “The polls are off by at least n points”. The best predictor we have of how likely that is to happen, is the number of times it’s happened in past elections. Or, if it’s never happened in past elections, then the best estimate we have is that it’s somewhere in between 0 and 1 in <number of elections>.
Posted in past half-hour:
Election Update: Trump May Depress Republican Turnout, Spelling Disaster For The GOP
Obama Romney polls had maybe an 8 point range through October with most within a 4 point range. Obama McCain polls had maybe a 9 point range with most clustered within 3 of each other. Bush Kerry similar. (All by RCP as source.)
The current 14 point range (using RCP but labelled 17 points by 538 above), and the lack of as much clustering (two each in the last week of Clinton +12 and Trump +2 in the RCP 4-way), are to me evidence of how the polling houses are struggling with how to apply LV screens and how in this cycle different LV screen emphasis leads to different results.
Stated intent to vote and past voting for example are each used as parts of LV screen, and this cycle may be more in conflict with each other than in past years.
To me this is the biggest potential source of uncertainty this time around and something that I see neither 538 or PEC paying enough attention to.
Karl Rove said on Fox News Sunday this morning that the math doesn’t work for Trump to win. More of this talk and GOP turnout could be really weak.