Princeton Election Consortium

A first draft of electoral history. Since 2004

Slow train coming

November 3rd, 2016, 9:26am by Sam Wang


The PEC calculation relies on state-level polls, which take time to come in. But by now the Meta-Analysis is current: of the top ten states listed in The Power Of One Voter, nearly all polls are post-Comey. The exceptions are Iowa and Michigan.

Michigan is vexing because of one oddball survey taken over a two-month period showing Clinton +20%. This is where medians help a lot; that race is more like Clinton +4.5%, which makes it about as competitive as Virginia and North Carolina. I hear Hillary Clinton is going to Michigan. In addition, there are three close House races there. It’s probably a good use of her time.

If you want an estimate that uses national polls, see The Upshot. I’m a little concerned that FiveThirtyEight’s code double-counts (i.e. overcounts) the swings in national and state polls. They’ve been a bit underconfident and volatile.

Of far greater import is the recent shift in Senate polls, which are also post-Comey (though we’re using a longer time window there). As of now, six races show medians (calculated using the PEC rule) of 1 percentage point or less. Entering the sharpest knife-edge zone are two states that have been trending toward Republicans: Indiana and Wisconsin. Democratic candidates Bayh (D-IN) and Feingold (D-WI) may be caught in an undertow caused by a shift from Clinton +6% (median of 19 polls that were in the field on October 24th) to Clinton +2% (median of 7 polls, October 31).

Things aren’t looking great for Bayh – the most recent 3 polls show his opponent Todd Young (R) leading by a median of 4 percentage points. The Senate Meta-Margin is staying steady, in part because Jason Kander (D-MO) is improving.

Tags: 2016 Election · Senate

125 Comments so far ↓

  • Vuk Vukovic

    Hi Sam,
    have you ever tried to include different types of polls in your models? In particular I’m referring to citizen forecaster polls.
    In yesterday’s article at the New Scientist https://www.newscientist.com/article/2111298-as-us-election-looms-time-is-ripe-for-a-new-science-of-polling/ this is the idea – let’s try and find a way to make polls more scientific by experimentally testing what works. In terms of their viability to predict the outcome of course.

  • Roger

    The simple thing I have trouble understanding if this election is really close is the female voter.

    Clinton is supposed to have an abnormally large lead with women more than making up her gap with men.

    Women are the largest demographic group of voters. ~52% or so in recent elections.

    So if Clinton has a commanding lead with a group that should be 52% or more of the electorate and that lead is much more than her gap with the other 48% of the electorate, how can the election be close?

    • Alex P

      There’s a theory that women are saying they’re voting to pollsters for Trump to keep the peace with their husbands. It doesn’t really make all that much sense. The evidence (sorta) is a YouGov poll from early Oct. Married men think their spouses will vote for Trump 42-33 and women say Clinton 41-33. The actual numbers were men: Trump 49-31 and women: Clinton 41-37.

      https://twitter.com/williamjordann/status/785251665230307329/photo/1?ref_src=twsrc%5Etfw

      It doesn’t make all that much sense, since it basically requires husbands to be in the room and a live poll

  • Alan White

    As media access changes, people’s attitudes change with the tech, and I’d think that must affect sample bases of polls. For example, I have a cell phone and never give out my number except to close friends, and never answer numbers I do not recognize, and I have a land-line phone that I only answer by message. As a result, for years I have never participated in polls. My sense is that careful people–some educated (PhD for me)–do not get sampled in random-call polls. But this kind of phenomenon of non-responsiveness to polls I would think stretches across the political spectrum–educated, paranoid, communicatively apathetic but politically participatory, off the grid but voting–and so I wonder how even sophisticated polls take account of this. Do they assume that all these non-responses are a statistical wash? Or is this a variable somehow controlled in some circumstances?

  • Roger NH

    I wonder about the actual random-ness of swing state polls. I’m a “likely voter” in NH (actually a definite voter as I have voted absentee) and have received scores of calls from pollsters since 2015. I used to respond to all of them because I figured if I was an avid poll consumer, then I ought to supply some of the raw material. About a week ago, I stopped responding, in part because it appeared that I was getting called repeatedly by the same few firms but also because I’m just sick of it. For a small state like NH, the number of people *still* willing to answer the phone has got to be vanishingly small. And yet there is this huge demand for up-to-the-minute polling. Perhaps the pressure on research firms to get to a certain “n” is so intense, that the firms are just going back to the same people that answered before….not really a random sample at all.

  • WildIrish

    I believe HRC will win the EV. It would be much better for her if she also gets at least 50% of the popular vote. That would be one less stick the Rs can use to beat her with.

    Thus, I am curious how the EVs translate into actual votes. Even in many of the red states, the difference is just a few percentage points. Even though she will not get the EVs for those states, she *will* get that percentage of the popular vote.

    I would like to see a simple table with the total number of LVs for each state, and then the breakout by D/R/I/G parties, using the current median percentages from PEC.

    I don’t know how to do it, but I’m sure one of the math/stats genies on this site have the ability.

    Any takers?

  • Richard H. Serlin

    “Early this season, I noticed that no matter what happened, opinion didn’t move that much.” — Sam at CNN interview today

    To the extent this is true, it would imply more spending on get out the vote efforts instead of things like TV advertising?

    • Richard H. Serlin

      Of course, at some point there are diminishing returns for get out the vote efforts, but at what point, how fast do the returns diminish, etc., and compared to diminishing returns behavior in TV advertising, and so on. A lot to think about.

    • Sam Wang

      Yes. That’s probably where Trump’s current surge is coming from.

    • Richard H. Serlin

      What I mean is, all other things equal, if you’re not going to move public opinion much, then it makes sense to spend relatively less on TV advertising to change minds, etc., and relatively more on get out the vote efforts. And for the longer term, crucially, on referendums and other efforts to get automatic registration and automatic ballot mailing in as many states as possible.

      The same applies to the extent that mean reversion in preferences exists and is strong; if so, spend relatively more on get out the vote efforts.

      But many factors, it may really behoove the Democrats, if they haven’t already, to get some serious academic firepower to study and plan all this, possible including advanced AI programs looking for patterns, and trying to optimize strategy.

  • LJ Breedlove

    Much is being made of early voter turnout in battleground states, and the breakdown I’d by party and by race. I haven’t be able to find much with a gender breakdown. Anyone know what’s going on?

    Here’s why I ask: I belong to a discussion group of about 500,000 Hillary supporters. Anecdotally, I’m seeing a lot of posts from women who are keeping quiet or even lying to spouses and family and then voting for Clinton. Not Trump’s “shy Trump supporter” a concept that makes my eyes cross, but a shy woman voter which makes a lot more sense culturally.

    Politico had an early October story on requests for absentee ballots being up among women, but nothing since then.

    (And then there was that interesting tidbit that said 25 plus percent of Republican voters voted for Clinton, which I was skeptical of until I started thinking about how gender might be factoring in.)

    I figure if any one can shed light on this question, they’d be found here. :)

  • Jason Bennett

    Does anyone know if the HuffPo feed in the bottom left matches the latest polls that PEC has processed, or is PEC more up to date than that? The RSS feed from HP seems to lag several hours behind.

  • Timothy_B

    What? No commenters recognizing Dylan’s relevant lyrics here?

    • Mick

      That’s the first thing that came to mind. That great album and song. But of course, who knows if Sam planned this or not …

    • Michael Hahn

      Not planned by Sam??? Who are we kidding!!! :D

  • Roger

    Would someone define Bayesian Probability in layman’s terms?

    I have tried studying it a few times but the definitions are so dependent on other statistical terms that I cannot even get a basic understanding of it.

    • Lorem

      Well, in general, it’s somewhat complicated.

      However, if you’re referring to the “Bayesian” in the prediction, all that’s really saying is “we’re assuming that there is regression to the mean”, i.e. that if the polls are temporarily far away from where they have been on average, they’re more likely to go back towards the average than to move further away.

      This is contrasted with Random Walk, which assumes that they’re as likely to go up as to go down at any given time, regardless of history.

    • Slartibartfast

      Roger,

      Bayes’ theorem provides a method (called Bayesian inference) for determining how our confidence in a hypothesis should change as a result of new evidence. Let’s say there is an event (like a poll). If we know what the probability of that event occurring given our hypothesis being true, the chance of the event given our hypothesis being false, and our prior confidence in the hypothesis, Bayes’ theorem gives us a formula for the probability that the hypothesis is true given the event.

      Because this formula contains the “prior” estimate, it has a tendency to revert to its previous state in the face of equivocal evidence. Dr. Wang’s model uses this type of calculation to determine the DRIFT of the meta-margin between now and election day. Thus there is an assumption that the most likely thing is for the meta-margin to move towards its long-term mean. That is all.

      I’m not sure where and how this is applied in Sam’s calculations, but Bayes theorem is used in some part of the process of converting polling medians to probabilities for each candidate winning a state and then figuring out the probability of all of the possible combinations. As a result, the model “thinks” that it is more likely that the race moves back towards the mean than becoming more extreme.

      As we get closer to election day, this effect becomes less significant as there is less time for the polls to move. As I understand it, the Bayesian and random drift probabilities will converge on election eve to a single number based on how likely it is for a candidate to win a state given the observed polling median.

      Hope that helps.

    • Simon Fraser

      I was under the impression it was a measure of bias in psychological studies. For example when applied to people who deny climate change, the level of evidence required to shift their view is so great that it is practically impossible to sway their views.

    • Sithi

      Roger, here’s an explanation of bayes theorem that might help
      https://betterexplained.com/articles/an-intuitive-and-short-explanation-of-bayes-theorem

    • Roger

      Thanks for all the explanations. From your help, I understand the concept generally enough top make sense of the two random drift and the Bayesian.

  • weichi

    Note that Sam doesn’t actually publish a “if the election were held today” probability. But if you look at today’s histogram, it looks like the model says a 97.5% (maybe a tiny bit lower) chance of a Clintonvictory if election were held today.

    Since the Bayesian probability for Clinton to win on Nov 8 is 99% (or more), it seems clear that the Bayesian approach has a pretty strong bias that the race is going to shift back to Clinton in the next 5 days.

    • weichi

      “belief” or “expectation” would have been a better choice of words than “bias” above!

  • Alan

    Also note Virginia will need a special election to replace Kaine. If the Senate is tied after Nov 8 this could determine Senate control.

    • SP

      Part of the reason I didn’t want Kaine chosen as VP. There are not a lot of Democrats left in the state with name recognition, whereas there are those like Allen that could easily run and win against a lesser known opponent.

    • Josh

      The governor of VA (Democrat Terry McAuliffe) appoints a replacement should Kaine become VP.

      The special election wouldn’t be held until the fall of 2017.

  • Nathan

    Those recent tied polls in Colorado and the past three NH polls (I know, the one is from a highly unreliable pollster) are making me feel physically ill.

    • Xavier

      Clinton losing NH seems to be already factored in to this calculation – she dropped from 317 EV to 313 today. The overall win probability seems to be steady though.

    • Anthony

      I am still dumbfounded how Trump is doing so well regardless of the polarization issue. I thought hardcore partisans were maybe 38% of the vote but at this point I would be surprised if Trump gets lower than 45% of the vote.

    • llywrch

      Anthony, I agree with you: where are these new Trump supporters coming from? (Unless it’s simply Republicans lying about voting for Trump when they’re actually voting for Clinton.)

      Further, since there is no simply explanation for them, I suspect their rising numbers will continue to panic all of the anti-Trump people. At least until next Tuesday when it will all be over.

    • Sridhar

      If you are worried the best way to track concerns in the campaigns is to see the upcoming schedule of visits to these states. Mrs.Clinton has no visits scheduled there till the election.Chelsea Clinton and president Obama are there with the President visiting on election eve.President Bill Clinton is the only surrogate visiting Colorado so far.Of course all of this could change very quickly but these states are not the focus of the Clinton campaigns indicating a relative level of comfort about the state of the race

    • Josh

      This.

      The Clinton campaign has WAY more riding on this election than you do. If their pollsters say Colorado and New Hampshire are not up for grabs, they’re probably right. Here are the states where Hillary Clinton’s team thinks resources are best allocated over the remaining 100 hours of the race:

      Ohio
      North Carolina
      Iowa
      Arizona
      Michigan
      Florida

      What do all of these states have in common? With the possible exception of Michigan, Hillary is probably within +/- 3 points of Trump in all of these states.

  • Deep breaths, deep breaths

    OK what the heck is going on? The direction the snapshot is going is NOT doing my blood pressure any good! And frankly I just don’t see how it has dropped that much, and that fast. Someone please explain using small, non-statistical words.

  • Brett

    People, don’t waste your time and energy worrying about 538. That is a for-profit media empire (Disney/ESPN)! They want you to keep visiting the site again and again. PEC is a much better model.

    • josh f

      I disagree that just because something makes money it’s implicitly bad. Often it works the other way.

      Also ostensibly 538 is using almost the same model they ran in 2008 and 2012, when they were very accurate. There is some aspect to this year that appears to bring out some unexplainable aspects to their model… Drew Linzer points it out here using Silver’s own article (Silver’s response on Twitter was a non-response):

      https://twitter.com/DrewLinzer/status/794696566695002112

      This other thing that 2016 has pointed out is that Silver isn’t always his own best critic. Since he’s on top of data journalism world right now he may benefit from open-sourcing more of his data and methods and using his reach to build a community that will drive even more traffic to ESPN- sort of the venture capital approach to marketing.

      This year Silver frequently takes the attitude that people that read 538 but disagree with his analysis don’t get it. In fairness to him you can’t do what he did without having some integrity and honesty so expect they’ll make it work better next cycle. Not quite there yet though.

    • Sam Wang

      I wonder if the model is the same, but the source of polling data has changed in quantity or quality.

  • Larry Guy

    Seems like PEC’s 99% Clinton win probability is based on the expectation that the recent relentless pro Trump daily trend will halt or even revert. If that happens, all hail to the model. If it does not, the model goes to the dustbin and it’s back to the drawing board!

    • Slartibartfast

      That is simply not true. There is a 99% chance (based on the assumption of Bayesian drift), that the meta-margin will not decline below zero. If the current decline continues, be somewhere around Clinton +2% on election day. If the actual meta-margin (as opposed to the prediction from the polling) for the election is greater than zero, that means that, by definition, Hillary wins.

    • Shawn Huckaby

      On what do you base this assumption? For your statement to be true you have to assume that the public opinion margin continues to drop at the same or greater rate in nearly every battleground state all the way to election day, and that somehow polling doesn’t capture it.
      Neither of these seems to be especially likely.

      Don’t forget that the probability listed is if the election were to be held today. If in fact there is an unprecedented precipitous drop over the next 5 days the prediction percentage wouldn’t stay at it’s current value.

    • Mark R

      Nope — the definition of the meta margin is different. It’s how much the polls would have to swing to bring the median of the predicted electoral college distribution to a tie. If on election day the meta margin were close to zero but still pro Clinton that certainly would not indicate a certain (e.g. >99%) Clinton win. The probability would be much lower; more of the distribution of electoral college outcomes would be in Trump winning area.

    • Larry Guy

      The “instant” polls are shifting towards Trump very quickly, in concert across many battleground states. Taken individually, each poll “blip” does not mean much, and PEC’s methodology tends to discount it as “noise”. But the scale of what we’re seeing seems unusual, particularly this close to Election Day.

    • weichi

      Shawn,

      “Don’t forget that the probability listed is if the election were to be held today.”

      No, “the banner reports an estimated probability of a win on Election Day, in November”, see http://election.princeton.edu/todays-electoral-vote-histogram/.

      “If in fact there is an unprecedented precipitous drop over the next 5 days the prediction percentage wouldn’t stay at it’s current value.”

      Agreed! Eyeballing the graphs, moving to an electoral college toss-up (or better) for Trump is a 2.5% event. If that actually happens, the banner prediction would drop to 50 % or worse.

    • Shawn Huckaby

      Thanks, weichi. After reading back I was wishing for an ‘edit’ button! What I was trying to convey was the 97/99% probabilities are a snapshot of the currently available polling, and so if the election happened today you would expect a certain win. Or am I still mangling it?

      in any event, thanks for clarifying!

    • weichi

      Shawn,

      The banner probabilities are a snapshot in the sense that they include knowledge we have *today* given available polls. But they are still two different predictions for what will happen on Nov 8, not what would happen today.

      Having said that, it occurs to me that the “random drift” probability is likely a very good proxy for the “if the election were held today” probability. As I understand it, that shows the probability of a clinton victory assuming that a drift in either direction is equally likely. Whereas the bayesian estimate assumes that a drift towards values that have been typical in the race so far are more likely. Since Clinton is currently a bit below that long-run average, it thinks she is going to move back up.

  • Davey

    Setting aside the partisan element for a moment, I also have to raise an eyebrow at 538′s model performance this election.

    As Dr. Wang has written, this has been a remarkably stable election (polls, not demeanors). Yet their model has swung between a 50/50 toss up and a 90% almost-sure-thing three times. The race has shifted at times, but it hasn’t seemed that volatile.

    It makes me wonder if we used the model to analyze 2000 if we’d see lines zooming back and forth off the chart all over the place.

    • Charles

      Vox wrote an article defending 538 today and claimed it was “lazy” to suggest their model is intentionally volatile to attract page views. But I don’t understand why this is a lazy accusation. Compare the variability of their model, which pays their bills, to models like PEC and PollyVote, which are done by folks in their spare time. It’s naive to suggest that 538 is more variable just because reasons.

      Another thing – Vox pointed out that 538′s model in 2012 was very stable. But go back and check HuffPo’s 2012 polling data. The race between Obama and Romney tightened just like this down the stretch but the average then was only Obama +1. Clearly 538 have tweaked their methodology.

    • TeddyVienna

      I don’t think trolling for page views explains everything at 538. I do think they’ve tweaked the model and added a lot of variability, probably as a reaction to 2014, Brexit and the unusual candidacy of Trump.

      That said, there’s a bit of math on Trump that doesn’t add up. He’s surely not getting a lot of people who voted Obama in 2012 — I’d have to think virtually none. And he has clearly lost a lot of traditional Republicans. So how is he still in the race at all? Is it an influx of “new” voters for Trump? Or is it that a lot of anti-Trump voters are still sticking with Johnson and Stein?

    • Arun

      Continually tweaking a model would mean it does an excellent job predicting the past – given enough degrees of freedom, one can model an elephant! If indeed 538 tweaked their model after 2014, then what predictive value does it have today?

    • Seth Gordon

      We see *now* that polling in the election has been very stable, but if the 538 model was built with the assumption that the election *could be* more volatile, it makes sense for that model to become more pessimistic every time the Meta-Margin approaches two points.

  • Tim L

    Anyone else find it odd that 538′s tracker seems to very closely follow the betting markets (like if you look at the Derived Bookies section of PredictWise for instance)? This seems very odd to me since the betting markets are established in order to generate equal action on either side – their expressed intent is not to predict an outcome but generate interest today. It has me feeling that the 538 model is really not a prediction at all, but rather the state of the race at this exact moment, which is expressly NOT what they claim to be.

    I appreciate the work here, Sam. I know we are not trying to drive a wedge between the two modeling techniques, but honestly, ever since the cake finished baking, the most interesting part of the PRESIDENTIAL election, to me, is how the models work and how they are different. I just really want to understand it more, and not just hear on the 9th how everyone was “right”.

  • Bill Dawers

    I think FiveThirtyEight has overplayed the importance of the trend line adjustment. If polling is simply returning to a long-term trend, then it seems inappropriate to project that the most recent changes will continue in subsequent polling.

    For example, right now in FiveThirtyEight’s polls only projection for Ohio, Trump is likely to win 48.2 to 45.8. But Trump’s projection benefits from a +2.1 percent adjustment because of the polls trending his way.

    • gedawei

      I think you’ve got it right. My reading of the “trend” is that we’ve hit Peak Trump in the post-Comey period, and the polls will NOT continue to show the tightening trend by this weekend. So the 538 assumption (if that’s what they’re doing) that the trend of the past few days will continue through to Election Day does in fact overplay what just happened post-Comey – if my reading is correct.

    • DaveM

      FWIW, the MM is up a tick this morning.

  • DCJ

    Sam,

    Just wanted to drop in the comments that I am a long time lurker and appreciate the effort you put into this website. Your level headed analysis without all the clickbait nonsense is VERY much appreciated.

  • Anthony

    Hi Sam, we have about 5 days left before voting happens. For the past two weeks the MM as ticked down over 2%. At what point should we start to panic? My friend texted me today and asked me when we should panic. I pointed him to your site, however it seems like MM is still on the decline.

    • Adam

      I take solace that even a 0.1% win is a win. I am a bit nervous from New Hampshire results, but looking at those two polls, they show a heavier GOP vote. I think that’s from weighting based on likely voter enthusiasm. Not 100% sure, and I would rather see leads by 20 points, but that may give you peace.

    • Adam Finch

      Look at t he history of the meta-margin here:

      http://election.princeton.edu/history-of-meta-analysis/

      If it drops below it’s all time low of ~1.8, and hits something like 1.5, I might begin to worry about the presidential election.

      (Spoiler alert barring Hillary being revealed as a literal demon, it won’t, because as Sam has noted many times this election [and they era we are in in general] is way too polarized).

      It IS time to worry about the Senate races though. Focus on the Senate races in the right column!

    • Matt McIrvin

      What good does panic do?

    • Elian Gonzalez

      As Mr. Wang himself has said, “the cake is already baked.”

      Tens of millions of people have already voted; this election will not be decided solely on breathlessly reported exit polls next Tuesday night. Panicking does not seem to be either helpful or a strategy to possess. (I’m saying that respectfully, not critically.)

    • anonymous

      Look at the slope of the MM line. Try to extrapolate the drop till the line touches Nov 8. Do you get below 1? Instead of panicking, devote some energy to urging everyone you know to vote, even if it is only on social media. Hillary has this in the bag, but the Senate candidates in states like New Hampshire, Florida, Indiana, and Missouri can use your financial help.

    • Anthony

      anonymous, I live in CA and I only know of one person who was voting for Trump and I already convinced him months ago to either vote Gary Johnson or don’t work for president. I have been all over social media for the past 6 months and on reddit trying to convince Jill Stein voters/Bernie or Busters to pull the lever for HRC. Debating trump voters has been a waste of time so I stopped bothering.

    • Ravilyn Sanders

      It is not our panic. The question is when the Republican traditional leadership is going to panic.

      All the Republican primary candidates were thinking, “no way Trump is going to win. Let me position myself to pick their supporters, when he collapses”. He did not collapse until it was too late for them.

      The current Republican leadership is thinking, “Trump is not going to win, let us position ourselves to hold the House and make a play for the Senate.”

      At some point, if they panic, they might finally stop playing with fire.

      Well, … that is wistful thinking. As long as Trump has the support of the vocal base, those guys will never grow a backbone.

    • TeddyVienna

      I don’t know about panic, but I’m unclear as to how the Meta-Margin keeps plummeting while the 97-99% odds remain intact. If it keeps dropping 0.3-0.5 per day, then yes, it could be below 1 next week.

    • Michael Coppola

      What good is panic? I don’t know, maybe catharsis?

      At any rate, what is the evidence that the MM won’t continue to drop? What’s the evidence that the trend won’t accelerate? We’re already in a realm where systematic polling error could mean that the race is a toss up.

      It won’t do any good to panic when a giant asteroids five days from hiiting the earth either, but I suspect that most people will panic anyway.

      It’s been said here before, even by Sam… even a 1% chance of Armageddon is still a chance of Armageddon.

    • anonymous

      @Anthony

      I sympathize with your conundrum. If you have exhausted other options, you could try swapping your vote with a third party voter in a swing state (TrumpTraders.com or some such service). I think it mostly works on the honor system.

    • Charles

      TeddyVienna,

      “I don’t know about panic, but I’m unclear as to how the Meta-Margin keeps plummeting while the 97-99% odds remain intact.”

      The win probability assumes regression to the mean, i.e. it won’t plummet forever.

      Sam has gone over this countless times. Increased partisanship has led to more and more stable polling. This is the most stable polling in history, I think. Hillary has led literally the entire time by an average of about 3.5%. People aren’t changing their minds.

    • TeddyVienna

      Thanks Charles — that makes sense.

      And from a non-statistical point of view, I’d imagine the Comey damage is done, and polling will eventually reflect the backlash against the leaks, along with a flurry of anti-Trump media such as the 300-plus economists warning against him. So getting back to 3.5% or so does seem likely even without the math — on Election Day if not before.

    • Clint

      @Anthony – Panic when you see a clear and realistic path to 270 for Trump. Right now, if I give him all the states where Clinton has less than a 3% lead, he still has only 251 electoral votes. He would then have to flip a couple more states like North Carolina + Minnesota, Pennsylvania alone would do it, basically he would need an additional 20 electoral votes, but he would have to flip states that he has a very slim chance of flipping. We knew the race was going to tighten at the end, it always does. What is of most concern are races in Congress, I think HRC has the big chair locked down.

  • charlie bird

    I forecast stuff for a living, have since at least 98. Not electoral stuff, but have been using multi-model forecasts since I built demand forecasting system for major internet retailer in 98 and have been using aggregation rather than selection since 2006 for major internet search engeine co. Key point is we let individual model wander and when they wander too far from trend, they exclude themselves. With an alg like this 538 would exclude itself from aggregation for being too far from rest of the models. Just thought you might like to know how one forecaster would deal with 538.

  • Scott

    I did some research:

    Since the aggregate of state polls continues to show a victory by Trump remains unlikely, I became curious about what might happen after Clinton became President. Here is what I learned:

    IF CLINTON WINS AND IS INDICTED BEFORE INAUGURATION:

    An indictment does not disqualify a person from being eligible for the presidency. Because Clinton would have to be indicted, undergo trial and be convicted prior to her inauguration in order to be deemed incapacitated, it is most likely that she would be inaugurated on Jan 20 while under indictment.

    IF CLINTON WINS AND IS INDICTED, TRIED AND CONVICTED PRIOR TO INAUGURATION:

    This scenario is highly unlikely, considering an indictment, trial and conviction prior to inauguration day (Jan 20) is virtually impossible.
    However, in the unlikely case this happened, Section 3 of the 20th Amendment kicks in and Vice President-Elect Tim Kaine, would become President.

    IF CLINTON WINS AND IS NOT INDICTED BEFORE BEING INAUGURATED, COULD SHE STILL BE INDICTED?

    The Department of Justice addressed this in 1973 and 2000, determining that a President is immune from indictment and prosecution for the duration of their time in office. So, in this scenario, Clinton would most likely be immune to indictment as of inauguration day (Jan 20) – at least until her term ended.

    IF CLINTON WINS AND IS INAUGURATED AFTER BEING INDICTED, COULD SHE STILL BE TRIED?

    No. See above. Immunity would kick in upon her inauguration.

    COULD CLINTON BE IMPEACHED AFTER BEING INAUGURATED?

    The law is clear, here. Presidents cannot be impeached for offenses they committed before they took office. Since the conduct took place prior to her becoming President, she couldn’t be impeached for it.

    So, no matter how you slice it, if Republicans want to avoid having a Clinton in the White House for the next 4 years, the only way to achieve it (aside from the fastest indictment, trial and conviction in history) is to win on Nov. 8

  • Some Body

    MM is +2.7 right now. A systematic error in the polls of a 3 pts or so magnitude is nothing unusual. I frankly don’t buy the claim that this MM corresponds to a 97% probability. Something has to be wrong.

  • J

    I manually input all the state probabilities from 538 and ran a simulation to see Clinton’s chance of getting 270 electoral votes. The probability came out to be 77.2%. So it looks like they did double count the uncertainty. I might have made some errors when I typed in the probabilities, so it would be nice if someone else could repeat the exercise.

    Without knowing the details, I also found their model to be overly complicated (the trend adjustments for example). Problem with presidential prediction is that there are probably 4 good data points for training the model. How can you have meaningful validation of a complicated model with such a small dataset?

    • BW

      What do you mean “ran a simulation”. Some details would be nice if you’re asking someone to reproduce the exercise.

  • Bryan Muenzer

    Dr. Wang,

    Does the PEC factor in early voting results in states like North Carolina and Nevada that seem to be well documented and available? Do you believe the reported lower black vote will play a major role in this election as is being hyped by the media? Should I stop watching CNN 24 hrs a day and maybe go outside?

  • DP

    I’m curious about your comment on 538 potentially double-counting swings, because I’ve noted something odd each of the last two days. Clinton’s chance of victory has dropped over 1% each time a single poll from Missouri shows up with a 1 point increase in Trump’s margin here. That’s a pretty large movement for a single poll in a state that has never been part of any Clinton victory strategy (i.e., always assumed to go to Trump).

    Any thoughts?

    • Nathan

      I believe that they highly correlate state results, so I’d assume that even though Clinton has no chance in MO, the model says that since Trump is improving there, he must be also improving elsewhere.

    • Doctor J

      Yes, it seems clear, and Silver is quoted as saying that they believe state polls to be highly correlated with national movement.

      This assumption means that outlier results in non-swing states would shift the 538 model quite heavily, as the model views those numbers as potential canaries in the coal mine for a broader national shift. Median-based and swing-state focused models should be less reactive.

    • jon stein

      I have seen the same thing – they also seem to be using national polls too much at this point – they really should be focusing on state polling at this point, since its EV’s that elect president. Its almost like they are panicking over Trump gaining nationally.

    • Matt McIrvin

      I suspect that’s a sign that 538′s model is smeared out way too much in both directions. In other words, previously, a significant part of its Clinton win probability actually came from Missouri.

    • Kenny Johnson

      I think it has to do with how 538′s model correlates results to other states. So if Trump is doing better than expected in a state like Missouri then that increases his chances in other states as well. “Polls Only” is a bit of a misnomer.

      Likewise, when Clinton polled well in Texas, it increased her odds in other states as well.

      Both of those cause the overall probability to shift. But it does seem to be WAYYYY too swingy this year. Keep in mind that Obama’s margin was smaller at this point in 2012, but his 538 probability was about 20% higher. Nate attributes this to Clinton’s overall lower numbers (e.g. mid 40s instead of high 40s) and large undecideds and 3rd party voters adding more uncertainty. I’m not sure I agree with his approach.

    • Nick B

      I too also found this a bit strange given that Missouri is almost certain to go to Trump. I think the issue here is the over reliance on national polls by 538. The very volatile national polls over the past week seems to correspond with increased uncertainty in the model.

      Case in point: 2012 had very tight polls throughout the race (average Obama +1), but 538 predicted a ~91% Obama win. In 2016, the current average of national polls is about Clinton +3 – 4, but with only a ~65% Clinton win.

    • The Vindicator

      Unfortunately, 538′s model(s) this year more resemble those of betting sites. In fact, a market-based site called latestpollresults.com is mirroring Nate Silver’s site. Clinton’s chances go down on 538, latestpollresults quickly follows. Many bettors (or “investors,” heh, heh, they call them on Predictit) are using 538 to make their wages.

      To quote Sam Wang: “Betting markets reflect the conventional wisdom. In the case of political races, the conventional wisdom is basically the true probability calculated from polls, watered down towards 50-50, especially if the race is within 5 percentage points. That condition applies here.”

    • JH

      I too am curious about this 538 comment. I would like to know about the specifics of how it is doublecounting/overcounting.

      It seems in general, compared to the other models, that 538 is more competitive in its outlook. Some of this is the “secret sauce” they apply, but I’m curious how they arrived at a such a different result.

      In general, as an IT guy, I like open source, so the open source aggregation model I read about recently on Slate seems like a step in the right direction.

    • George H

      “there’s been a strong correlation between the number of undecided and third-party voters, and polling volatility” as per

      http://fivethirtyeight.com/features/election-update-why-our-model-is-more-bullish-than-others-on-trump/

    • KLB

      I believe Sam’s policy is not to comment on other models. Personally, I think that the 538 model is so overwhelmed with data that it’s kicking back junk – think of it like a phone camera at 8x zoom on a cheap tripod.

    • AAF

      If there are other states with similar demographics to Missouri that haven’t been polled as recently, he will impute to them the same movement as seen in Missouri when it is polled.

      So if, for example, North Carolina is similar to Missouri and hadn’t been polled in a week, and Missouri gets polled and has moved 1% in Trump’s favor, he will adjust NC’s numbers by a little less than 1% in Trump’s favor as compared to NC’s numbers as they stood before.

    • Alex P

      AAF, I know you’re right about the adjustment, but how similar is NC to MO really? The education levels are pretty similar sure, but 63.8% are non-Hispanic white and 81%, respectively. MO has never been in serious contention. NC moved .5% more than the national average between 2008 and 2012. MO moved 8% more.

      This is getting a bit more into punditry than a statistical site should, but Clinton is partly running on Obama’s record. Missouri isn’t going to move towards to her that much, no matter how many times Trump brags about sexual assault.

  • Bob McConnaughey

    well..my last sizable donation (to races) just went to the ActBlue link on the top left of this page. I VERY much appreciate having a rational means of allocating $$ available.
    Is it also worth giving some to general “GOTV” groups/group? MoveOn …(Acorn -;)

  • Josh

    Somewhat unrelated: has anyone seen Alan Abramowitz’s prediction for the House?

    http://www.centerforpolitics.org/crystalball/articles/incumbency-not-republican-gerrymandering-is-the-main-obstacle-to-a-democratic-house-majority/

    He believes that GOP gerrymandering isn’t *that* big a deal, and that if the Democrats win the generic ballot by more than 5-6 points they’ll likely take back the house. Seems quite at odds with what pretty much everyone else is saying (best case, Dems would need an 8-point win to take the House, worst case it’s basically mathematically impossible).

    • BrianTH

      My understanding is you would really need to know the distributional pattern of relative support before you could pin down the margin needed to flip the House. And that might not stay constant from election to election as coalitions shift, people migrate, and so on.

    • Roger

      I always thought Gerrymander actually weakened districts for the party doing it.

      From my understanding, Gerrymander is spreading your party’s state vote over more districts to gain more districts than normal drawn districts would produce. That means they are spreading their vote thins in more district.

      To me that makes them strong during normal elections but if a large wave election happened, then all those districts would probably fall because their margins are not that high (i.e. spread thin.)

    • Davey

      I enjoyed Mr Abramowitz’s mathematical analysis, but I found his conclusion to be fallacious and illogical. (That winning or losing a majority is all about incumbency and, as he says, has “nothing” to do with gerrymandering).

      His thesis is a false dilemma. Incumbency and gerrymandering play a role, it’s not one or the other.

    • Ravilyn Sanders

      @Roger // Nov 3, 2016 at 12:03 pm
      You are correct in saying, when the wave overtops the levee, all the districts will get flooded. But how high should the wave be?

      Currently the Redmap project created 60R-40D districts. House district level polling is scant, expensive and not public. Sam has calculated based on correlations, a 6% Dem advantage in national polls in the generic congressional ballot will be needed to put the House in play. 8% will likely get Dem house.

      If the control comes to the Dems, it could be in a landslide. But as of now, there is no polling data to say Dems could get it.

  • ADC Wonk

    The news on the Senate is very disheartening.

    I did read (from YouGov, perhaps) in answer to a Nate Silver question about why their polls were so steady, the following: “Fewer voters were changing their minds than were changing their inclination to respond to surveys.”

    Might that explain the dip in Senate chances, and, perhaps, it’ll be better for the Dems than currently predicted?

    • Charles

      Someone else (maybe Sam, actually) touched on this same phenomenon the last time Trump apparently “closed” the gap. When a candidate has a bad news cycle, their supporters tend to not response to surveys, which wrongly suggests a swing toward their opponent. Then the news fades and the margin normalizes again.

  • Mike

    Would it be fair to say that the PEC model’s use of medians without taking into account national polls harmed it’s senate predictions? They have seemed very wobbly. But Dems 50 seems like its probably been the best prediction the whole time.

  • StevenRN

    Im so anxious about this election and only your prediction/analysis/model calms me. Don’t let me down!

  • Gil

    Guys, I’m really puzzled at HuffPo. As much as I’d love to believe the numbers there, I’m not sure I can.

    Click on Ohio. Pretty much all the recent polls listed there show Trump with sizeable leads, yet HuffPo still has Ohio as Hillary with a small lead. It just doesn’t make sense, especially when everybody else is putting Ohio as lean R, or solid R.

    Also, their Senate estimate as 94% D seems incredible. If Bayh is trending down (and he has for the entire cycle), plus considering Indiana as solid Trump, there’s no way he’s going to pull it off. I don’t see Kander (MO) winning either on a red state. Considering that the races in blue states (NV, NH and now even WI) are essentially inexplicably tied (voters voting Hillary but R for Senate?!), I just don’t see how we can have D control anywhere higher than 50% at this point.

    Thoughts anyone?

    • LBS

      Typically very few people split tickets and I have been suspicious that the actual vote is going to represent the degree of vote splitting that the polls suggest. However, it could be that this year is different. Pollsters are reporting that college educated women, who went 6% for Romney are -27 % Trump. That matches my anecdotal experience. Almost every Republican woman I know (all are college educated) is voting for Hillary or not voting for President. They are supporting Republicans down ballot, however. I am wondering if that explains the polling discrepancy.

    • Matt McIrvin

      Believe their poll collections, not necessarily their model averages. I’ve always thought of them as kind of wacky, and they seem to tweak them a lot.

    • Brian

      If you click ‘Customize This Chart’ at HuffPo from the Ohio page, and choose Less Smoothing, it becomes something closer to Tanenbaum’s result at electoral-vote.com, with Trump having around a 4% lead. According to the link below, their smoothing uses a “trend line estimate” which I assume is heavily influenced by the earlier polls which we were better for HRC:
      http://www.huffingtonpost.com/entry/huffpost-pollster-poll-averages-methodology_us_57d1a3b2e4b06a74c9f361cb

    • Matt McIrvin

      Though in some cases, “Less Smoothing” produces a result that is comically dominated by the very last poll they’ve got.

    • SJWangsness

      Blunt is barely holding on in a state that’s fairly strong for Trump. Also, has issues re: “losing touch” with Missouri (e.g., expensive house in DC). Kander in contrast is fresh, has an excellent chance of winning.

    • TeddyVienna

      LBS raises an interesting point. Trump is doing SO much worse than Romney across so many demographics. How in the world is this race less than 10 points? Is Trump doing that much better with the shrinking demographic of white men without college degrees?

  • FayeL

    A question – I haven’t seen this addressed anywhere – how does early voting affect polls? I would think that people who have already voted are less likely to respond to polls, which means that they are unrepresented in results. Thanks!

    • JB

      I’ve been wondering the exact same thing recently, curious if this is possibly related to the pro-Trump swing since the early voting data in most states suggests pro-Clinton turnout.

    • Jim

      Why would early voters be less likely to respond to polls? The only change for early voters is that they are no longer screened by likely voter questions, they are simply put in the ‘highly likely to vote’ category.

  • Roger

    Does your model take the individual polls listed in Huffington Post or does it take the Huffington Post summation for each state?

    I am wondering because Huffington Post seems to have some equation for averaging polls that alters the weighting. Many of it state’s have overall odds very different from a simple average of the polls it list – some dramatically different. For example, its Texas summation is 47% to 43%, yet most of the Texas polls it lists show Trump +13 type numbers with lows of +6 and +7 for the more recent polls. Only much older polls show Trump only +3 and +2. Yet the overall Texas number is only Trump +3.

    • Brian

      Sam’s model uses the individual polls so he can take the median.

    • Matt McIrvin

      Individual polls. Huffington Post’s smoothing is weird.

    • Brooklyn23

      Thanks for clarifying this. Ive been wondering too. Does anybody know why HuffPo was chosen over RealClearPolitics? If you had to take a guess, would choosing one site over the other cause a major shift in the meta margin?

    • Sam Wang

      In the past, Huffington Post has been run by professionals who apply AAPOR standards for screening polls. There’s a lot of transparency there. Also, as stated already, it’s not possible to get a feed from RealClearPolitics.

    • Brooklyn23

      Thank you for the explanation. I’m new to the world of polling and learning as I go :)

  • Brian Westley

    Just wondering if you or any commenters would like to estimate — if the Senate comes down to one swing vote either way, how likely is it to “turn” a Senator to switch parties and flip the Senate? Who on each side is most likely to turn?

    • Froggy

      I’d rate it as extremely unlikely. Nonetheless, if I had to pick the senators most likely to switch parties, I’d go with Joe Machin (D-WV) and Susan Collins (R-ME). Jeff Flake (R-AZ) seems pretty unhappy with his party lately.

    • FlyingSquirrel

      I think Flake is pretty conservative even if he’s not happy with the party right now – I don’t see him becoming a Democrat. I’d agree that Manchin and Collins are the most likely candidates to switch.

    • Amitabh Lath

      No need to turn, just vote with the other party. See the NY state senate where 5 democrats do just that and effectively give power to the republicans who are technically in the minority.

    • BW

      Collins would never switch. Ever. She has a national media presence as “moderate”, but she has supported Paul LePage, and looked for any chance to support Trump. She’s not leaving the party. She loves the party.

  • Josh

    Just wanted to point out Russ Feingold is D-WI not D-MO.

  • Mark R

    Hi Sam — looking forward to your explanation on how PEC’s calculation differs from others’ predictions. Especially eager to hear your thoughts on the argument put forward by one prominent aggregator with a much lower probability of Clinton winning (ha!) which holds that errors in state polling are not independent of each other but highly correlated — more so than other aggregators are assuming. Thanks!

    • Charles

      I think maybe he covered this once? There were basically no new PA polls immediately post-RNC, but 538 modeled the race there as breaking toward Trump, since that’s what the polling showed in Ohio. But that assumption turned out to be inaccurate, and Clinton has only consolidated her lead in PA.

      Vox described this aspect of 538′s methodology as a “cool feature,” but my highly uneducated opinion is that it’s primarily responsible for their relatively unique projections. I don’t see any evidence that the natural variation in state polling margins suggests real nationwide shifts in opinion.