Princeton Election Consortium

Innovations in democracy since 2004

Outcome: Biden 306 EV (D+1.2% from toss-up), Senate 50 D (D+1.0%)
Nov 3 polls: Biden 342 EV (D+5.3%), Senate 50-55 D (D+3.9%), House control D+4.6%
Moneyball states: President AZ NE-2 NV, Senate MT ME AK, Legislatures KS TX NC

Sometimes life comes at you fast

September 16th, 2016, 7:15am by Sam Wang

It looks like the Presidential state-poll snapshot is heading for a near-tie. This should become evident in the PEC analysis by the time of the first debate. I believe this will be a temporary situation. It will take at least until after the first debate on September 26th to find out.

National polls currently show Clinton ahead by only 1.0 ± 0.6 % (median ± estimated SEM, 7 pollsters with at least some post-Phlegmghazi respondents). However, our state poll-based analysis moves more slowly. I expect the Meta-Margin to keep on moving toward Trump for at least a week as state poll medians catch up. To get an idea of where the PEC analysis will head in the next few days, see, whose main map displays the most recent single poll for each state and therefore is noisier than my calculation – but more up-to-the-moment.

I estimate that if we had up-to-date data in all states, in an election held today the Presidential outcome would be extremely close, approximately like the map below.

Click the map to create your own at

If the election were today, it would be a nail-biter. But as Glenn Thrush at POLITICO points out, there are lots of ways for Clinton to recover.

Today, I will go over some reasons that may make Clinton supporters (and Trump opponents) feel concerned.

1. National polls have narrowed. Seven polls that include post-September-11 respondents show Clinton ahead by 1.0 ± 0.6 % (median ± SEM, 7 polls), quite close. Compare that to the last week of August, which showed Clinton ahead by 4.0 ± 0.9 % (10 polls).

The narrowing of the race seems to be driven by a break in undecided voters toward Trump. Since early August, Donald Trump has gained about 4 percentage points and undecided voters have decreased by a similar amount. At about 42% support, Trump is nearing his previous highs for the year.

Until now, undecided voters have been running about 4 points ahead of where they were in 2012. Many of those undecideds were probably disaffected Republicans unwilling to support Trump, a radical candidate unlike any that the U.S. has seen in over 100 years. Trump is not saying a lot relative to his usual inflammatory and divisive statements. It appears that this approach of keeping relatively quiet is bringing home the Republican base – and increasing his favorability rating. We will see how long he can keep it up.

If you were to extrapolate this trend linearly, by Election Day, 52 days from now, Donald Trump would win the popular vote by 10 percentage points…and there would be negative 5% undecided voters. This is obviously impossible. But silliness aside, it is within the realm of possibility for Trump to take the lead, at least temporarily.

I still expect Clinton’s lead to increase again, on the grounds that she has led all year. This is basically the concept of regression to the mean. Previously, I noted that the national Clinton-vs.-Trump margin in 2016 has averaged 4.5 percentage points. The standard deviation is 2.2 points, comparable to the four Presidential elections from 2004 to 2012. Such a small standard deviation indicates very high stability, consistent with the intense voter entrenchment of the last 20 years. If conditions swing back toward Clinton, especially after the first debate on September 26th, I can imagine the stories now: “Clinton shows renewed vigor” or something like that.

Despite the lack of big ups and downs this year, it is important to remember that the large number of undecideds and minor-party voters provides a source of uncertainty. We have to wait to see how they shake out. As we see with today’s conditions, recent changes have benefited Trump.

It is usual for support for minor candidates to fade in the home stretch. Johnson and Stein get strong support among young voters, a group that would otherwise be expected to tilt strongly Democratic. A big hope for Hillary Clinton must lie in bringing these voters back. Maybe Clinton should bring out Bernie Sanders, Elizabeth Warren, or some of the entertainment celebrities who showed up at the DNC.

Technical note: The Presidential model’s Bayesian prior does contain the possibility of a big swing in undecided and minor-party voters. First, the prior’s standard deviation was set to 3%, somewhat larger than recent elections. Second, it was set to have “fat tails” by using the t-distribution with 1 degree of freedom instead of the usual bell-shaped distribution. Under the latter assumption, a swing of 2 standard deviations would occur almost 15% of the time. If Trump wins, it was still in the model as an unlikely but not impossible scenario.

2. State polls are still narrowing. The advantage of national polls is that they provide information quickly, within days. But national polls are less accurate than state polls, and worse, do not reflect Electoral College mechanisms.

This is where the Meta-Margin comes in. The Meta-Margin acts like a regular Clinton-versus-Trump margin, but it is based on all state polls. It measures the amount of across-the-board swing in opinion that is needed to tie up the Electoral College. However, it is dependent on the availability of state polls, which dribble in slowly.

The coming 1-2 weeks will probably feature a steady drip, drip of state polls favorable to Trump. Likewise, the Meta-Margin will move toward zero over the coming week, and could even favor Trump at some point. Do not interpret that as new information. That is just the same information that we already know now, expressed at the level of individual states.

3. Are Trump supporters undercounted in polls? In a phenomenon called the Bradley effect, it used to be that poll respondents tended to vote for minority-group candidates a bit less than their answers would indicate. This idea has been used to suggest that additional Trump supporters are lurking in the wings, but are unwilling to admit their possible racism to a human interviewer.

I doubt that there is any hidden bonus for Trump. The Bradley effect was never more than about 2 percentage points, and in any event, it disappeared a few decades ago. A research finding from Morning Consult earlier this year suggested that Trump’s support was lower in live-interviewer polls than in computer-conducted “robopolls” has been the same – but later analysis has not confirmed this result. That original report probably arose by chance or had deficient statistical methods.

Finally, I note that Trump’s support in the primaries was very close to what polls indicated. In all probability, polls will do fine this year.

Having said that, because state polls have not yet caught up with national polls, it is useful to see what would happen if one candidate got a boost in state polls. Over in the right sidebar are links you can click if you want to see what would happen if Trump had 2% more support than state polls currently indicate. That probably gives a good picture of the true state of the race today. Conversely, once the Meta-Margin gets closer to zero, the regression-to-the-mean concept suggests that a Clinton recovery would become likely. At that point you can click on the Clinton +2% link to see where things might head.

4. What about downticket races? Four indicators seem to move roughly in unison: the Presidential race, the Senate, the House generic, and the current President’s approval ratings. This is consistent with the nationalization of politics that has emerged in the last 20 years. After the national party conventions, all four measures moved toward Democrats. Now all four measures are moving toward Republicans.

This coordinated movement does not seem entirely consistent with the idea that undecided voters who moved to Trump were disaffected Republicans. Shouldn’t they have shown up all along in polls as supporting downticket Republicans? This could be a quantitative issue of how much each indicator moves – for example, the Senate aggregate has not moved as much. Or there’s some other mechanism at work, such as differential rate of survey response.


The close race that we see today is not likely to last. Odds favor a swing back toward Democrats, but this is not certain by ant means. It will take a little while to find out – maybe a week, maybe longer. A good number to watch is the median of national polls.

Many of you might want to have an impact on what happens in November. The most effective use of your time and money is to work on races that are closest to the edge, i.e. in the 20-80% probability range. In the right-hand sidebar, The Power Of One Vote lists states where individual voters have the greatest potential to influence (a) Senate control, and (b) the Electoral College outcome (click through for that). Those are good places to campaign. Depending on your political preference, you can also contribute using the ActBlue/NRSC links on the left.

Tags: 2016 Election · President

146 Comments so far ↓

  • Kevin King

    I notice that the meta margin has ticked up a little bit over the weekend, even while the overall Senate race has continued to tighten a little bit. Is perhaps the range of this election pretty robust after all?

    • Adam

      HRC got a nice poll out of Pennsylvania, which helped stop the bleeding a little bit. Holding MI/WI/CO/VA and then winning PA and NH gives HRC the presidency, most likely. So this result is very positive for her, at the moment.

  • Jeremiah

    Hi Sam, I think the House Generic Preference chart is not updating correctly. HuffPost has the last 3 polls as +5,+6,+2. Therefore the median should be +5 but the chart shows +2.

    • Froggy

      Jeremiah, the calculation uses “the last 3 polls, or all polls ending within the previous 21 days, whichever is greater.” There’s way more than three polls being used at this point.

    • Jeremiah

      @Froggy +2 Still seems incorrect. The polls within the last 21 days have been:


      The median of that lot is +3. Maybe the chart only gets updated once per day?

    • Jeremiah

      Looking at the Matlab code Sam filters results for pollster ID so only one result from each pollster is returned. I can’t make out what all the code is doing it’s a bit hard to know because of the run-time resolution of type I guess I could try to run it.

    • Sam Wang

      If multiple results are reported by a pollster simultaneously (for example 2-way vs. 4-way), the code takes whichever is reported first. The idea is that this represents the pollster’s judgment about which result is more accurate. Since most supporters of minor-party candidates come home to a major candidate, a midpoint between 2-way and 4-way results seems appropriate.

  • George

    And today we see what happens to the meta-margin when an organization drops 50 state polls of unknown size and unknown quality with a 3 week window into the mix – a drop of 1.1 points…. The model is only as good as the data that goes into it…..

    • Greg Gross

      Sam – A drop of 7 points in a day on the Bayesian? Can you explain? Bayesian is supposed to crawl in either direction, no?

      And George – Is Huffpost the organization dropping the polls?

    • ideo

      I think the polls were dropped by Ipsos/Reuters. Just wondered if it’s just a crosstab or dedicated polls.

    • George

      Yes – he pulls from, which I think is from HuffPost.

    • Jeremiah

      Ipsos is the organization publishing the polls. They admit that their confidence in the numbers is “Low” or “Moderate” as noted here:
      The polls are internet based.

    • Jeremiah

      Looking at the state by state numbers from that Ipsos poll the samples are laughably low. Not only that but the sample sizes (and their own confidence levels) seem suspiciously low for all the swing states, e.g. Florida 148, New Mexico 224, Wisconsin 83!

    • George

      Jeremiah, you are identifying the problems with those Ipsos numbers. The “good” ones are as suspicious as the “bad” ones wherever the sample size is low. I don’t think I’ve ever seen a state poll of repute with less than about 400 voters in it. So samples like FL, NM and WI must have incredible margins of error, even assuming the sample was representative.

    • Jeremiah

      George, yes wherever the sample size is decent, e.g: North Carolina with 917 and Ohio with 625 their confidence is “Low” indicating that the sample distribution is probably not representative for key demographics.

    • Adam

      Yeah the Ipsos polls are insanely suspect. I am actually pleased with RealClearPolitics leaving them out. Now, some are favorable to Hillary – some are favorable to Trump.

      I think we all know that Trump folks are incredibly active online – it’s his main advertising medium. So that would make a lot of sense why some of them seem off.

      The pollster I don’t know is Emerson. Most seem very pro-Trump.

      And then, I continually get frustrated with the weighting by 538. It seems to get worse every cycle.

  • Froggy

    Any plans to start considering Maine CD2 separately? It’s looking rather Trumpish these days:

  • Rex

    Hi Sam, it looks like todays drop in the meta-margine and win probability were what you predicted last week in this post. Is it fair to guess with that the inlux of national polls trending upward, we can likewise expect an uptick in the meta-margine next week?

    • Sam Wang

      That is a good point about national polls. At a time of change like this, it is a good idea to keep an eye on them. Could be an uptick, not totally certain yet.

  • Shoreline view

    Question….how regularly is the meta margin updated? I ask because I wake up this morning, see a couple good Hillary polls, Monmouth with that five point lead in Florida for example, then flip over here to see that the Hillary meta-margin has dropped again and Florida is marked as leaning Republican. On the other hand you have Ohio as all tied up again, which I’m not quite seeing yet.

  • Marco

    I follow PEC for the analysis, and Pollster for the raw polling and Electoral-vote for the write ups (chuckle here and there …). Pollster has shown, since yesterday, a small uptick for HRC both in the two-horse raceand in the 4 way (HRC up by 3.1%, was less than 3). It seems that the cause of the sudden drop in metamargin are polls from Ipsos. These polls do not appear to be the “gold standard”. Hopefully the metamargin will return to the mean quickly enough, otherwise, Canada here I come.

    • George

      Marco – if you look at the feed, it was all of those junk Ipsos polls. Kinda like the elephant in the snake, they are gonna have to work themselves through the system. If they are all wet, they will be replaced with better numbers. If accurate … plan your packing? ;-)

    • DaveM

      I looked at HuffPost’s national presidential chart, and the median of polls ending within the last week is HRC +2. I then went back another week, and the median is again HRC +2. A number of state charts show a similar stability.

      Of course, every-state polls like today’s Ipsos/Reuters lollapalooza have the effect of introducing a lot of noise—many states moving in contrary directions—which makes it harder to use state polling means to assess the overall race. (If Clinton is up 3 in Ohio and down 3 in Colorado, which way is the national wind blowing?)

    • RoF

      Silver is posting on twitter that this is looking like a disaster, and that only his model is correct.

    • Sam Wang

      What’s a disaster – the Presidential race? I am perplexed – it’s just state polls, leaking out more slowly with the same information as what national polls were telling us last week.

      The other piece of evidence is national polls. In the last seven surveys, Clinton leads in six and one is tied. The median is Clinton +3%. That’s an uptick from a few days ago. Doesn’t it seem like state polls ought to come back up to follow?

  • Michael

    What, if any, meaning can be taken from the Clinton campaign publicizing what they call their “internal” polls that they say favor her? Can a campaign’s internals be said to be better, worse, no different than the public ones? Romney allegedly went into Election Day in 2012 convinced he was going to win.

  • TDubs

    Am I just not finding the right link? Is there a place on this website where I can see what polls are included in the calculation? Right now, it says there are 135 polls, which ones?

  • Remi

    I know this isn’t a statistical comment, so may be dropped, but just as an ordinary consumer, it appears to me that polls are either mis-advertised or misunderstood. They change. Day to day.

    To my non-numerical mind, that means none of them is making a prediction about November. All of them are saying “these are the probabilities if the election were held today”.

    If you tell me in January there’s an 80% chance X will happen in Nov. but tell me in August there’s a 20% chance X will happen in Nov. why should I give any credence to the January prediction?

  • Marco

    I understand the difference between national polls and state polls. However a link exists between the two. In the last two days (even before the metamargin cratered), national polls showed upticks for Clinton. Today a national poll by NBC/WSJ shoved Clinton with a 6-point lead in a head-to-head, while 7 if the other two candidates were included. Unless their voter models are screwy, things are bound to go back to “normal”. Crossing fingers, toes and whatever else that can be crossed.

  • Anthony Shanks

    Does anyone notice that even Sam’s model is starting to converge in line with other prediction sites? Before PEC stood out on its own with its very conservative prior, but even with that into account, it’s starting move into a far uncertain outcome.

    Forget being a little nervous, I’m scared now that all prediction models give Trump a good chance.

  • Allan

    Hi Sam,

    Lately Nate Silver’s been tweeting how certain “otherwise smart people” seem to be in “denial” about Trump’s chances.

    That they seem to be making dangerous assumptions that the race will shift back to Clinton at some point.

    What are your thoughts on this? Do you believe that the PEC might fall victim to this type of thing?

    • Sam Wang

      Dog whistle!!! Joking aside, I am certain he doesn’t mean me in particular. There is not a shortage of people who think this.

      I think your criticism is mistimed. I set the algorithm a while ago and was transparent about it. It is not rocket science: snapshot of polls today (slightly favors Clinton) plus regression to the mean (favors her more). You had plenty of time to ask about it.

      Conversely, I think it is a little silly to have a “forecast” that just reflects current conditions. That is not a real forecast, but just a blurred-out snapshot.

  • Allan

    Hi again Sam,

    Thanks for answering my last comment. I just want to clarify that it was not intended to come off as hostile or overtly critical in any way.

    I was curious about your thoughts on Nate’s latest comments. I had a feeling, just as you said, that he wasn’t mentioning you specifically, but I still wanted your input.

    Love your work and all!

    • Sam Wang

      I went and looked. He seems to be complaining about David Rothschild in that one. And then there is this one, which feels like a dog whistle about me. Maybe I’m just sensitive.

      Anyway…I really don’t have much to say here. I have spelled out the assumptions of the PEC calculation. I could do it again, but it gets repetitive. Basically, I think that if anyone ends up with a probability for Trump much larger than what we report, their assumptions implicitly contain far more uncertainty than has occurred in any race since 1952. I guess that is possible, but it seems unwarranted by any data that I am aware of.

    • Kai

      Sam, thank you for explaining again. This article in Daily Kos
      seems to argue that the main difference in forecast models is due to the assumption on how states are correlated. Basically, it says high correlation results in lower probability. Also, models with low correlation will have sharper peaked histograms than models with high correlation. So from this I take for PEC: low correlation, peaky histogram, high Clinton probability. What I would like to understand is how the SD discussion plays into this. Are these two sides of the same coin?

    • Sam Wang

      The correlated-states thing is probably a red herring…people have been talking about it for years. A more important difference is how much volatility we each expect between now and November 8th.

Leave a Comment