Princeton Election Consortium

A first draft of electoral history. Since 2004

Three Easy Pieces

November 1st, 2016, 9:16pm by Sam Wang


Three quick notes.

  1. Updates are now done every hour. I’m not trolling for traffic. I just want to be able to see the effects of new state polls soon after they hit the Huffington Post (their RSS feed is in the left sidebar). Like you, I watch the automated Princeton Election Consortium calculation!
  2. The Upshot, which carries our calculations, asked for individual estimates of Maine and Nebraska’s Congressional districts. We implemented that on Saturday, October 29th, mainly through a change in EV_estimator.m. The update took the Presidential Meta-Margin toward Donald Trump by 0.1%, principally because there is a chance Hillary Clinton will lose Maine-CD2 . The jerseyvotes calculation is also updated: at the moment, turning out one voter for Clinton or Trump is about equally valuable in Maine’s 2nd Congressional District, Pennsylvania, or Virginia.
  3. There seems to be a lot of interest in understanding how and why the PEC calculation differs from other sites. I will address that topic soon.

Now, what are jerseyvotes and why do they matter for get-out-the-vote activity?The basic idea behind the jerseyvotes calculation is that individual voters have very different amounts of power to affect the overall Presidential race. For example, Vermont is a guaranteed Democratic state and Oklahoma is a guaranteed Republican state. Therefore, in those states an individual voter’s preference only matters in primaries and some local races. On the other hand, in a state that could go either way, individual votes matter a lot. And if the state is small, voters are extra-powerful for two reasons: small voting population, and disproportionately low numbers of votes are assigned per electoral vote.

Back in 2008, I set up jerseyvotes as a way to assign a numeric score to a voter’s power. The idea is as follows: how much power does one vote have to swing the entire national race? This is implemented by seeing how much the Democratic (Clinton) or Republican (Trump) win probability is affected by adding one vote in one state. That is then converted to a vote power by scaling it to the most powerful voters – currently New Hampshire.

Right now, the most powerful voters are almost all in Clinton-favoring states. That is because voters are only powerful if the national Presidental race is on a knife-edge. Currently, it’s not – Clinton is favored. So first we shift all polls over toward Trump by an amount equal to the Meta-Margin. That creates a perfect toss-up. Only then do we calculate the voter power. Consequently, the most powerful voters are in states that are between tied and Clinton +6%. Translated into English: To have any shot at winning, Trump partisans have to dive into weak Clinton states and hope that the national race swings enough to make things competitive.

The Jerseyvotes concept is equally applicable to the Senate, where things could genuinely go either way. If you are reading this site, you care a lot about whether a President Hillary Clinton has a friendly or a hostile Senate. This is especially important considering the disintegration of governing traditions in which the President used to be given a free hand in Cabinet and Supreme Court appointments. Senate control can go either way, so it’s the critical leverage point.

For activists, I consider the Senate jerseyvotes calculation to be the most important numbers that this site offers. At the moment that I write this, Senate voter power is as follows:

  1. New Hampshire: for influencing Senate control, voters are the most powerful in the nation. Score: 100.
  2. Nevada: 31, i.e. each voter has about one-third as much power as a New Hampshire voter.
  3. Missouri: 23.
  4. North Carolina: 15.
  5. Indiana: 14.
  6. Pennsylvania: 4.
  7. Every other state: less.

Tags: 2016 Election · President · Senate

75 Comments so far ↓

  • Debbie Lefkowitz

    While I’m glad you are not going to get into a pissing match with the 538 crew, I think it is important to state that 538 is a private company owned by a large corporation and that impacts their work. I don’t think they are meddling with their model but I do think that they are required to drive lots of traffic to their site and that means a more volatile model is of benefit. I’m hoping I’m wrong but the endless round of “what if Trump won the electoral college and Clinton the popular vote” waste of space articles suggests they are driven as much by eyeball requirements as by the desire to analyze the race correctly. This is very disappointing to me and makes me very, very grateful for your work Sam. Having an analysis that is not driven by profit motives is very helpful.

    • Josh Soffer

      Its not just the volatility of the 538 model. Silver’s current electoral college numbers depart significantly from PEC’s. I think Nate Silver takes a lot of pride in getting it right, and the volatility of his model may owe more to (over?)compensating for his misstep in forecasting the GOP primary than to ratings pressure. Plus, would followers of 538, PEC or Real Clear Politics rather see Clinton at a 65% probability or a 99% probability?

  • Michael Smith

    Why HuffPo and not (for example) RealClearPolitics or other poll aggregators? They seem to disagree on which polls to include, and that could introduce a bias, couldn’t it?

    • William

      If I had to guess, it’s probably because HuffPo provides easier to work with data :)

      RCP doesn’t like to give out the internal data that they work with, presumably because of their business model. So Sam would have to scrape the site – not fun.

    • Brooklyn23

      I’ve been wondering this too. Some states have a big difference in polling averages between RealClear and HuffPo (Colorado is a good example). If PEC started referring to RealClear, do you think it would cause a major shift in the meta-margin?

  • truedson

    Does this report mean anything new…or is it baked in the polling cake…

    http://nymag.com/daily/intelligencer/2016/11/black-turnout-down-in-early-voting-in-good-sign-for-trump.html

  • John Parenteau

    Regarding: “There seems to be a lot of interest in understanding how and why the PEC calculation differs from other sites. I will address that topic soon.”

    I am awaiting the election in eager anticipation of seeing the results in Maine (CD 2) and Nebraska (CD 2), where, according to the NYT Upshot, PEC’s probabilities are very dramatically different from the rest of the field’s. Everyone else (NYT, 538, Huffpo, Predictwise, DailyKos, etc) essentially has them as tossups, while PEC has each at > 90% for the favorite.

    Any comment on those particular predictions, Sam?

    • Matt McIrvin

      There aren’t a lot of polls in those districts, are there? Sample differences could account for a lot.

  • David

    Today most (all) of the other aggregators are going down for Clinton, but PEC goes up – from 97/99 to 98/99. This makes me happy, but it also makes me long for a narrative explanation. Would love to read one. Thanks!

    • Marc

      Yeah, I second that. My skeptic bell keeps ringing when I see 99% Clinton. It would be naïve to ignore 1) that she’s under federal criminal investigation (a first for a presidential nominee); 2) the intense anger much of the population has towards government and the direction of the country; 3) much lower enthusiasm among African-Americans; and 4) that populist nationalists like Trump tend to be very underpolled – even Ed Rendell thinks there’s a big hidden Trump vote in PA.

      I mean, these things have to count for something? I just fear that now (more than any prior point in the campaign) there’s a very real chance Trump can actually pull this off.

    • Roger

      The last two mornings it has been at 98/99 then went to 97/99 by mid/late morning to stay there the rest of the day.

      Is this a real measurement or some quirk?

    • David

      That’s certainly what I’m wondering – when there’s fluctuation like that, especially in the direction of Clinton, even if minimal, what does it mean, when all the other aggregators seem to be going the other way?

    • TeddyVienna

      HuffPo is at 98.3. And they have a persuasive piece pointing out that Clinton’s numbers right now are better than Obama’s at this point in 2012, yet 538′s probability was much higher in 2012: http://www.huffingtonpost.com/entry/state-polls-clinton-leading_us_581a8098e4b08f9841ad3a03

  • Matt McIrvin

    There’s been lots of attention paid this time around to “phantom swings” caused by non-response bias–which implies that the volatility of recent presidential races is actually even lower than we think it is.

    It makes me wonder to what extent the high-volatility environment that existed before 1992 was even real. Surely opinion polling was less sophisticated back then (and there were fewer polls). Were there just massive swings happening in response bias?

    Obviously, there was a lot more variation in voting patterns from year to year back before the Reagan era, so I wouldn’t be surprised to find that some of the volatility within an election season was real too. But I wonder.

    • Sam Wang

      No, but if you look at polling data from Wlezien and Erikson, the time trends were fairly slow, not at all the kind of one-week moves that we are able to observe today. Carter/Reagan/Anderson in 1980 is a great example, though a bit extreme.

      Basically, I think the big advance in polling is our ability to see with better time resolution and better state-by-state resolution.

  • Shma

    The NYT has reported an ME2 and NE2 number for PEC for a while now and those numbers were different than the respective numbers for the full states. What was being done to calculate them before the change to ev estimator?

    • Sam Wang

      Those probabilities were hard-coded using polling data available at the time, assuming constant district-by-district offsets estimated from 2008/2012 elections. I basically sent them fixed numbers for display. Now it’s using the same rule, but updated dynamically to follow current polls.

  • Ed Wittens Cat

    From Drew Linzers TL– another perspective is conditional probability– of course state to state covariance plays into the calculation– but this still looks like 1-2 % from my POV
    https://twitter.com/DrewLinzer/status/793526542446030849

  • N75

    I usually lurk, but… is the title of this post a nod to Stravinsky, or Buffalo Tom?

    Thanks for a great site, Dr. Wang.

    • Sam Wang

      Actually, Richard Feynman, who I believe was nodding to Jack Nicholson/Bob Rafelson/Adrien Joyce. Thanks for reading.

    • Matt McIrvin

      The title was probably picked by the editor or publisher–that book was compiled from material from the Lectures on Physics well after Feynman’s death.

    • 538 Refugee

      With Feynman, “easy”, is relative. (That looks like a pun now that I typed it. ;) ) I won’t tell you how many times I started “QED” and had to stop before I finally made it through. And that was supposed to be easy version for the masses? I read a some questions asked by interviewers for Google. One was, “Tell me everything you know about QED in 10 minutes.” I could have at least bluffed 2 or 3.

  • Richard H. Serlin

    A comment I made in response to Kevin Drum’s thoughts on mean reversion today:

    Honestly, at this point it’s hard for me to even watch (but I have to look at the basic facts each day, for responsibility reasons). But I will say this about “reversion to the mean”. This is a pretty high-tech statistics concept. And it’s often hotly debated whether it exists in a given situation. For example, those who argue that financial markets are extremely efficient, even perfectly efficient (I think almost always for ideological reasons, and/or because for that kind of theory they’ve had years of toil to get expert in, and so they don’t want its value and prestige decreased), argue that real asset prices follow a (perfect, or near perfect) random walk, which means no reversion to the mean. Future events are independent of past ones, like coin flips, or dice rolls; same exact random process no matter if you previously flipped 10 heads in a row. It’s still a 50% probability of a tail on the next flip.

    But my big political point here is that if there really is mean reversion – substantial mean reversion – and there really could be – these are human beings not coins or dice, then a political party could potentially gain a substantial advantage by hiring serious PhD statisticians to study and model it, to see when and how it occurs, and thus where and when to better deploy resources.

    Even further, if there really are pretty predictable patterns of this type; like mean reversion, but others as well, then sophisticated AI programs can study the data to find them, and perhaps point to substantially improved strategies. Of course, the party that actually believes in science and reason, and thinking beyond only simple-minded soundbite reasoning and dogma, would be more likely and able to take advantage of this kind of thing, so it might be something for the Democrats to look into, if they haven’t already, with some serious academic firepower.

    Interesting for economics and finance research: If the reasoning behind EMH is true, then not only should financial markets follow a random walk, but so should political candidate support. If human beings all possess all the public information, and the expertise to analyze it instantly and well, and they’re 100% forward looking, then poll numbers would only change due to unexpected events, which would be random, independent of all previous information. So, political poll numbers would follow a random walk.

    If you can show through econometrics that these poll numbers do not follow a random walk, very substantially, then that would be (or should be) strong evidence against the EMH.

  • Norbert Hirschhorn

    I have a small experience with statistics from my medical research but a better understanding of epidemiology. I wonder about polls, of course. Polling in the past few elections has come close to predicting the actual outcome, but each election has to be an N=1, this one in particular, for all sound and fury, whose influence is unknown. We assume sexual predation and careless use of an email server shouldn’t be equivalent, but we don’t know that.

    I think of other unknown unknowns.

    1. As polling proceeds, each person polled has to be aware of past polls. If a large number of cold calls are made to derive an acceptable number of respondents, who are the ones excluded? That is, the sample cannot be a true reflection of the universe (+/- x % points). This works both ways for Republican and Democrat voters, but may not be a wash.

    2. Who are the people doing the actual polling on the phone (or however)? If minimum wage workers, what’s their background, voice, accent?
    Again, may not be a wash.

    Nate Silver’s analysis and cautions (see today’s) scare me, but are based on the correlation of national polls closeness and electoral outcomes, where states are not independent actors.

  • counsellorben

    Sam,

    When you address the hows and whys of the PEC calculation (for what seems like the nth time), would you please address the factors which have caused the Presidential Meta-Margin this year to be generally a much smaller percentage than the leads in the various aggregations of national polling (such as Pollster’s)?

    If I recall correctly, in 2008 and 2012, the Presidential Meta-Margin tracked the national lead as reported by Pollster much more closely.

    Intuitively, I would not expect such a wide deviation between the Meta-Margin and the national polls, since the possibility of the winner of the national popular vote losing the Electoral College seems to require a popular vote margin of <1.5% or so, while the gap between the Meta-Margin and national polling appears to be at least 2.5%.

    • Josh

      The likeliest explanation is that Clinton has extra support in redder states. This isn’t hard to see–states like Texas, Alaska, Arizona, South Carolina, Utah and Georgia have all trended lighter shades of red than in recent cycles.

    • weichi

      I agree that this has often seemed to be the case (meta-margin smaller than national polling average). But at the moment they are identical!

    • weichi

      I should clarify that they are identical if you measure the national margin using the same approach that Sam uses to determine margins in the individual states: median of polls completed in the last 7 days. This gives Clinton +3.0 per the current set of polls on pollster. The actual average on pollster has it at Clinton + 5.9, which doesn’t seem reasonable, considering that only 2 of the last 10 polls have Clinton higher than +3.

  • Pat L

    I took your advice and went to ActBlue and tossed some funds to the PA Senate race. In my state, IL, the icing is already on the cake for the Senate I believe. Thanks for your site and your continuing updates.

  • GM

    I always imagine Sam hitting the ‘Post’ button with a single-malt in his other hand, given that each new PEC item brings with it variations on the same eleven questions (and occasionally other, more interesting questions).

    These are the kinds of tasks undergrad volunteers were put on this earth for, Sam. :)

  • David G

    Hi Sam,

    Your website and your whole approach have been a welcome source of calm during this crazy election. My one worry (okay, truth be told, one of many worries), is how your model can factor in a really large jolt right before the election – which is exactly what has happened now.
    How are you so confident about Clinton’s ability to withstand that?
    Or to put it another way, I need more convincing that – to use your words – the cake is already baked.

  • Tony

    Sam, I am wondering about the implicit correlation assumption in your polynomial expansion approach. I am thinking, for example, about the reduced case in which all electoral entities (in the practical case, states) are perfectly correlated. In that case, what is the meaning of the estimate you would get from the coefficients of your polynomial expression?

    • Sam Wang

      If I understand you correctly, it’s like this: on the one hand, the polynomial calculation implicitly assumes independence…but it’s just a snapshot.

      On the other hand, the actual November prediction relies on the assumption that the Meta-Margin moves by some amount. That amount contains within it both correlated and uncorrelated movement – and is furthermore something we can observe empirically, just by watching the Meta-Margin move up and down. Its peak-to-peak movement is five percentage points, very close to the range of the national Clinton-Trump margin. It makes me think that nearly all state-poll movement is correlated nationally. However, the way we model it, it doesn’t matter – we’re covered either way.

    • Tony

      Thanks for the reply, Sam. Yes, you do understand me: I thought the polynomial snapshot implicitly assumes independence, in which case it’s not a good estimate of the probability of the outcome. OTOH, I did not understand that your 97% and 99% estimates include an (empirically derived) estimate of state-to-state correlation. Therefore, the difference between your forecasted probability and that of, e.g., 538′s, has to boil down to a difference in that embedded correlation estimate.

    • Michael Ralston

      Tony: It could also just be a variance estimate – Sam’s model is essentially saying “if the race moves by 3.2 points towards Trump, it’s a 50-50 tossup, and also, the odds of Hillary winning are 98-99%”.

      From this, we can trivially derive that the model thinks the odds of the race moving 3.2 points towards Trump are what, 2-3%?

      Looking at 538′s model, while the model’s estimate of how much things have to move to become a 50-50 is not clear, it’s gotta be somewhere between 2 and 4%, so the models roughly agree on where the race currently is… which means 538′s model thinks the race is WAY more likely to move by 3 points in Trump’s favor than the PEC model does.

      (given that their forecast is now >25% trump victory, it seems like if the model thinks moving in either direction is equally likely, that it is ALMOST CERTAIN to move about 3 points – since you’d get 25% if it was guaranteed to move by enough to go to a 50-50 if it goes pro-trump, since then it’s effectively two coin flips: one for things moving his way, then him winning the knife-edge race.)

      Intuitively, a 3% chance of moving by 3 percentage points seems a little low, but 538′s “it’s basically guaranteed to move that far” seems … also unlikely to be accurate?

      (I guess there’s the argument about polling error – maybe they have a “it’s probably off by 3%, we just don’t know which way” baked in…?)

    • James McDonald

      538 seems to assume a much higher likelihood of several states making large correlated swings together. That’s plausibly warranted if you think the past 10 or so elections is not enough data to rule out an October surprise of unprecedented magnitude. (For random examples, what if Trump’s tax returns were leaked and showed him broke with a billion in debt to a friend of Putin, or an email turned up with Clinton directing some felony. Those could have major impacts even given the current state of polarization.)

      I think Nate even mentioned somewhere that if he assumed independence Trumps odds dropped to less than 1%.

      PEC’s estimate is more of an “everything behaving similarly to what we’ve already seen” approach, while Nate’s is more paranoid. I think Nate goes a bit overboard, but does put a bit of an outer limit on how freaked out you should be.

    • Tony

      Well, if you assume independence, you don’t need a model to see that Trump has little chance of winning. He has to win FL, NC, and PA. Suppose he has a 30% chance of winning each one. Then if they’re independent, his chances of winning all are .3^3, or 2.7%. But if they’re perfectly correlated, his chances are 30%. So the essence of modeling the election comes down to the co-movement among states.

    • Kevin

      Aren’t there two scenarios that could benefit Trump? The race could move in his direction, or there could be systemic polling error in Clinton’s favor. I sometimes think Nate is adding these together.

    • Matt McIrvin

      Yeah, the only plausible doom scenarios I can come up with where Trump wins are the ones where both a significant further move in his direction AND a large polling miss against him are happening at the same time.

      Now it’s intuitively compelling to make a “momentum” argument that since Hillary’s numbers have slid this far, they’re going to continue sliding at the same rate all the way to Nov. 8, but I don’t know if it ought to be regarded as nearly certain.

    • Michael Ralston

      I think momentum arguments are silly, with the single exception of the time frame shortly after a major event (and we are no longer “shortly” after Comey’s little bomb, in this context) wherein some polls with mixed pre-event and post-event tracking show a moderate effect that is plausibly attributable to the event – in which case, you should expect however much additional change as we get polls that are purely post-event.

      But otherwise, a momentum argument seems to need to argue for a reason why either people would respond to support changes by changing their support (note: there is one for primaries and other multi-way races, where it would be rational to abandon a first-choice candidate in favor of a second or third-choice candidate, and some people probably DO that) or why we’d expect the fallout of whatever event to be significantly delayed.

      Certainly I’m not seeing any reason to expect “momentum” to apply here.

  • Roger

    Know anything about Targetsmart? Its CEO was just on MSNBC, his firm is releasing a Florida poll tomorrow showing CLinton 48 Trump 40. As part of Targetsmart’s poll, it polled early voters. The poll found that 28% of Florida early voting Republicans voted for Clinton. He talked about 15 minutes and said his poll has many controls, and they are very confident in its methodology.

    Anyone have any dope on the poll specifically and Targetsmart in general?

    • Di Zhu

      A lot of people are skeptical; so am I. In a state like FL I don’t think any candidate could beat the other by 8 points. And I’m not so confident his sub sample of early voters really represented this group of voters. I watched live too, and I could see his confidence, so I look forward to more explanation from him.

      Having said that, it also showed Rubio 49-43 leading, which is a very reasonable result.

    • Bela Lubkin

      This poll may have an accurate read of already committed votes, but those votes are not necessarily indicative of the final outcome.

      Republican voters who are crossing over to vote Clinton may fear Trump more than Republican voters who are staying with Trump like Trump or fear Clinton. This fear-driven urgency would tend to bring R-crossover voters to the polls early, leaving a larger proportion of staying-R voters to vote on Election Day.

      Voting early helps ensure an individual’s vote since he becomes immune to personal emergencies (etc.) preventing him from actualizing his intent to vote on Election Day. This has become more important in recent election cycles in which the voter can anticipate crowded, slow polling places and possibly surprising ID requirements (which, if encountered days before the election, he might have a chance to correct).

      If the Election Day mix of voters is significantly reduced of D and R-crossover voters, the voters suffering the most from slow polling places may be the residual R voters. Some of them could become discouraged and give up.

      In that case suppression efforts would have had the dual effects of hastening D voters to vote early (thus banking more assured votes), while impeding R voters on Election Day. (Tactics which make it harder to register in the first place are a different matter, probably less likely to directly backfire. They backfire more slowly by driving up the determination of the suppressed populations, and eventually misfire by being ruled unconstitutional or at least illegal. I hope :)

    • Di Zhu

      Bela, that poll showed Clinton 42-43 with people who haven’t voted yet.

    • Matt McIrvin

      There are always outliers. There’s a SurveyUSA poll out today showing Trump way up and above 50% in North Carolina. Better to just throw it into the aggregate along with all the others and see what happens.

    • Matt McIrvin

      …and I heard about one a couple of days ago showing Hillary leading in Alaska. I don’t necessarily believe that either.

  • Joe Schmoe

    Regarding: “There seems to be a lot of interest in understanding how and why the PEC calculation differs from other sites. I will address that topic soon.”

    I think it would also be interesting for you to lay out what to look for in the actual results next week that would support PEC’s model over other sites such as 538. E.g., Presumably they have a ‘much’ wider dispersion of result probabilities than PEC, where the PEC 99% probability for expected outcome is only one slice of many other possible outcomes plotted by these other sites. Would this mean that PEC should have a much more accurate prediction on the specific outcome and margin of victory in state by state and national results?

    Maybe I’m not capturing this well, but I’m just curious if their are specific election result details (that you could state in advance) that would demonstrate the superiority of the PEC model. Something more than just … “See, HRC won. Q.E.D.”

    • Sam Wang

      That is not really what I am going to do at all. There won’t be any kind of compare-the-models smackdown. I am trying to get past the adolescent phase of this activity.

      I am going to explain how I generate a low-noise model of this type, show why it shows less uncertainty, and highlight the assumptions that make it so. I think my approach is best for the purpose of resource allocation, which is what campaigns and activists do.

    • Josh

      I’m actually surprised that Sam is addressing the “PEC difference” at all, as I thought he has already addressed that multiple times.
      On Nov 8th EVERY site will be predicting a Clinton victory, so that won’t tell you anything. His prediction that Trump would be the nominee long before anyone else thought so, or his prediction prior to the first debate that “no matter who performs better at the debate, polls will move toward Clinton due to regression to the mean” were far more interesting validations of his methodology than the final outcome of the presidential race will be.

    • Joe Schmoe

      Sam,

      Fair enough.

      The ‘why it shows less uncertainty’ is what I’m most interested in, so I look forward to your writeup.

      I really didn’t intend to set this up as a ‘smackdown’ type of question (though it came out that way). I was more interested in the process of evaluating & adjusting (or validating) the model and what you look for in results that might suggest future adjustment.

      But I suppose that’s for another day …

      Thanks for swift reply.

    • Matt McIrvin

      “On Nov 8th EVERY site will be predicting a Clinton victory, so that won’t tell you anything.”

      I’m guessing that 538 will still be giving Trump a substantial win probability, though, probably similar to today. We’re into the zone where correlated uncertainty in the final result is getting to be more important than any motion in the next 7 days.

    • Donna Gresh

      Four years ago I did a little analysis of 538 post election-day, specifically looking at the published uncertainties of results from the day before the election. I used the individual probabilities for each state and binned them roughly. That is, if a published win probability for a state is roughly 80%, then roughly one out of five of those “80%” states should have been called incorrectly. For those with an probability close to 50%, roughly half should have been called wrong. You can plot probability vs percent wrong for each “bin” and if your estimate of your own uncertainty is accurate it should be (roughly) a straight line. I no longer have the result, but 538′s uncertainties just before the election were significantly higher than their “incorrectness.” That is, their uncertainties were too high.

    • Matt McIrvin

      @Donna Gresh You’re right. Their published uncertainties have always been high–this isn’t just a consequence of them going to ESPN.

      I think Nate Silver explicitly justifies this on the grounds that the low volatility and low errors of state aggregation from 1992 on might just be a fluke, and we could return to the mid-20th-century situation of much higher uncertainty at any time. Whereas Sam is taking as his prior that we’re still in the low-volatility regime.

    • Michael Ralston

      @Donna Gresh One caveat of a “binning” analysis like that is correlation.

      Imagine that someone lined up the states, and said “So, here’s a distribution of how the results will be.”, and from that derived a distribution of the results that gives 50% for one state, 60% for another (where it’s 100% in the 50% of the previous, plus 20% of the remaining 50%…), etc – if that lining is exactly correct, and the distribution is also correctly confident, you’ll still end up with an analysis that looks “underconfident” a lot, and “overconfident” a few other times.

      The only way to judge that is to repeat it over a long stretch of uncorrelated results, since the correlated results are, well, correlated – you can’t tease out “do they seem underconfident because they’re actually underconfident, or because they had a concealed correlation that happened to go the most likely way this time?”

  • Ryan

    Thanks for being you Sam.

  • Ed Wittens Cat

    ur not that different
    i see a definite clustering of superior methodologies
    https://twitter.com/jshkatz/status/793553664724140032

  • SK

    Hi Sam,

    Curious what you make of the betting markets that give Clinton quite a low probability of winning compared to the poll aggregators (Betfair ~72%; PredictIt ~68%).

    Thanks.

    • Sam Wang

      I wrote about this in 2012. 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.

    • SK

      So essentially it’s “favorite-longshot bias” (i.e. favorites underestimated, longshots overestimated).

    • Kevin

      Also I suspect 538 gets too much credit in the betting markets. Nate crows that his average is closest to the betting markets, but he is also the most famous poll aggregator that bettors are referring to.

  • Clark

    What do you make of the North Carolina SurveyUSA poll? What does that do the overall scope of this race? The SurveyMonkey poll was up for Clinton, but the SurveyUSA poll is way up for Trump. Considering that some of the early voting for NC is down for certain parts of the population this year, does this invoke a possible trend in your statistics?

    • Sam Wang

      I don’t answer questions about single polls unless there is some particular angle that is interesting. The whole point of the PEC approach is to avoid poll-sniffing.

    • Geoff

      Hi Dr. Wang,

      What do you make of claims that there are more undecided voters than usual this year, and therefore that the election is more unpredictable? My completely unscientific impression is that many supposedly undecided voters in fact have quietly decided. “Joe the Plumber”, supposedly undecided, didn’t seem likely to decide in favor of Obama. Several of the supposedly-undecided questioners at the town-hall debates this year and in 2012 gave me the impression that they’d already decided, too.

      So I’m skeptical that there are really so many undecideds, which would seem to be consistent with your polarization thesis. But I am not a political scientist, and I have no data to support my supposition. What do you think?

      Thanks for a great site, by the way. I wish I’d found it sooner.

    • Clark

      Okay, if I may ask a more broad question. It seems to me that the Senate is now tilting more towards the Democrats. Is there a correlation, historical or otherwise, that this is an early indicator of how the Presidential race will eventually go?
      Also, I know that Nate doesn’t like to use early voting numbers (although, funny, his partner Harry Enten just wrote how those numbers are helping in Nevada), does your methodology include these numbers at all? If so, why?

    • Sam Wang

      Early voting numbers are super-interesting, but no, I don’t use them. That is a totally different stream of evidence. It is worth watching, especially if an analyst puts them in perspective by making comparisons to 2012.

    • Trey

      Sam,

      Check out Michael McDonald’s twitter and their Elect Project website (http://www.electproject.org/) if you’re interested. They’ve been doing a lot of early vote comparisons between 2012 and 2016. Sometimes even using 2008 if they have the data.