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

Paths of Glory

July 19th, 2016, 10:27pm by Sam Wang

Welcome, New York Times readers! Josh Katz unveils The Upshot’s model for the November election – complete with comparison to other sites: electronic bidders (PredictWise), polls-only (FiveThirtyEight and the Princeton Election Consortium), and professional prognosticators (Charlie Cook, Stuart Rothenberg, and Larry Sabato).

Qualitatively, PEC and The Upshot mostly agree, though PEC shows less confidence in Nevada, Ohio, and Mississippi. That is likely because PEC uses polls only; The Upshot’s calculations include consideration of past elections.

In addition to the usual displays of probabilities and electoral vote totals, there is a cool interactive. For example, with the five assignments listed below, if Donald Trump is to win the Presidency he would need to sweep four states: Florida, Pennsylvania, Ohio, and North Carolina. Currently, Hillary Clinton is slightly favored in all four. Give it a whirl!

Tags: 2016 Election · President

22 Comments so far ↓

  • BillSct


    The networks seem to be using real time tracking polls (MSNBC keeps talking about SurveyMonkey). Could you comment on these as a new wrinkle? In particular, I am wondering if these polls suffer from the problem of the sample being self-selecting.

  • ItsSupercar

    I love that interactive tree of outcomes– they had the same thing in 2012 and I’m so pleased to see it return. Really gives a sense of what the stakes and influence are with each state.

  • A New Jersey Farmer

    Great graphic. The narrative now is that Trump has narrowed the gap from a month ago, but this shows how slim his path actually is.

  • Jay Sheckley

    Spectacular coverage and visuals.
    Congratulations, Sam!
    Oh and though I couldn’t accept it at first, what you said on the Woocast is completely accurate and many pollsters, Jon Stewart and finally me have come around to your conclusion: Despite stunt casting, this election _is_ behaving normally: Quite like 2012 re: demographics etc, but more stable. LOVE the moving NYT graphics which instantly answer all questions. And resemble blood flow :D Now that the Times has this and has totally accepted PEC, I must resubscribe.
    Congratulations again to Princeton election Consortium’s growing recognition as the most accurate, most data-based [ergo least biased] election thermometer, and obviously also the clearest [ie least 50/50] prognosticator.
    To stay calm this season. come back often and check the header. Do it with math! But dont forget to vote.
    Fyi, readers: Far as I know, Dr Wang still needs a Princeton community person to assist him here on PEC and also in the neurology lab. For all he does, let’s help get the word out.

  • Amitabh Lath

    The level of agreement among the predictions is interesting. Everyone starts the wall of red at Georgia. Given the mean values are clustered so closely, I worry about groupthink (aka correlated systematic uncertainties).

    Likely voter screens are one big unknown this year. There are stories on both sides of long-time fence sitters (blue collar whites, Latinos, poor women) who are being activated in “record numbers”, “as never before”.

    I tend to discount stories which do not quote hard numbers but I have to admit there are a lot of them nowadays.

    • Matt McIrvin

      One of my worries is that this election campaign is going to be so repulsive that people tune out in fear or disgust, and it actually turns out to be a very low turnout election (making it possible for strange things to happen).

    • Amitabh Lath

      Matt, yes that’s definitely part of the uncertainty. Subroup A is suppressed due to disillusionment or whatever, while subgroup B is enhanced due to resonant rhetoric, and LV screens largely based on 2012 participation fail to pick up any of this.

      Theoretically poll aggregation should mitigate this problem, but if all the pollsters are constructing similar filters, then maybe not.

      Perhaps the thing to do is to weigh pollsters based on how different their LV filters are from the average filter, to ensure a flat distribution in filter-space.

    • Matt McIrvin

      …That said, my impression of the people who say resigned “both sides are equally bad” kinds of things is that they are predominantly of the Caucasian persuasion, which might upend some assumptions about what happens in low-turnout elections.

  • Josh

    Makes me wonder if this year selecting Vilsack over Kaine really might make a difference. Vilsack remains very popular in Iowa, which is a lot closer than VA… I know historically it hasn’t mattered much, but Vilsack isn’t just “a guy who happens to be from that state.” Well-regarded on both sides of the aisle, which is what’s needed in a purple state…

    • Josh

      Oops correction… Amazingly if she wins PA and VA on that chart, she can lose all 8 other swing states and still win.

  • Cervantes

    Matt McIrvin:

    One of my worries is that this election campaign is going to be so repulsive that people tune out in fear or disgust, and it actually turns out to be a very low turnout election (making it possible for strange things to happen).

    For comparison: As a fraction of voting-age population, turn-out in presidential elections over the last forty years (1972-2012) was noticeably lowest in 1996 (49.0%), 1988 (50.2%), and 2000 (51.2%).

    Otherwise in that same forty-year period, turn-out ranged from 52.6% (1980) to 58.2% (2008).

  • BillSct

    Why is the Times histogram for Electoral College outcomes so different from PEC’s?

    • Tony

      PEC’s is if the election were held today, NYT’s is what it will be in November.

  • Kevin

    Re: “The narrative now is that Trump has narrowed the gap from a month ago, this shows how slim his path actually is.”

    This is deceptive. Since the results in these states are correlated, running the table in several swing states at once is not the difficult parlay it appears to be in the chart. If the overall race swings four points in Trump’s direction, the path will seem less narrow, although Clinton will still have chances.

    • Sam Wang

      Yes, but there is still the question of the probability of an overall swing, which would have to be 3.1% (exactly equal to the Meta-Margin). From Eisenhower to now, such a swing would not be out of place. But in the last three elections, the SD of the Meta-Margin has been less than 3 percentage points. Possible, but a tough lift. Which is why PEC has him at an 18% win probability (the nonrounded number is visible over at The Upshot’s scraping of our site).

  • Mark F.

    Trump needs Romney states plus FL, OH and PA. Agreed. Tough.

  • Turgid Jacobian

    I see that there are 15 or 16 political units not included in your state poll csv.
    I keep thinking I remembered that you used the previous election’s margin for states that are as yet unpolled.
    Is that right? I don’t seem to be finding that statement anywhere.

  • SouthJersey

    Has anyone done/come across any actual analysis of the impact that changed voting laws have had or could have (post Shelby County v. Holder) on actual voting vs. poll responses (where RVs or LVs are kept from voting). In this cycle, this is a nagging concern of mine with polls-only analyses.

  • Steve Scarborough

    I find the NYT/Upshot article to be fascinating. Here are my thoughts, as a long time statistician and follower of the PEC:

    1. For my own modeling, which has some similarities to the PEC approach, my results are comparable to the latest NYT/Upshot data, state by state.
    2. Since the article first came out, there is an update as of today. I note that the 538 numbers (win probabilities) for each state have generally gone down 2 to 4 percentage points. I find this to be interesting in that there have really been not all that many polls since the article first came out.
    3. As I write this, the 538 model shows a probability of Clinton win at 58.7% with 289.1 ECVs. Those numbers are down from where they were just a few days ago. I find that interesting.
    4. If you copy in the win percentages for NYT/Upshot, 538, and PEC — state by state — and declare Clinton/Trump gets all the ECVs if one or the other is leading in a particular state, you get (assuming I copied correctly) the following national totals for Clinton:

    Upshot 345 ECVs
    538 330
    PEC 339

    FYI, my modeling gets 347.

    Personal comment: in my eyes all of these numbers are reasonably close to each other.

    5. In going through the 538 modeling methodology, I realize that they have polls only, and polls plus methods. But, I find it a bit difficult to reconcile today’s web site ECVs for Clinton at 289, when a simple win/lose ECV assessment by state shows the 538 model at 330. I know all about the next day versus future, etc., etc.. But do any of you find this discrepancy to be hard to understand?

    6. In line with that, it appears that 538 is attempting to describe the election as a close toss-up. Hmmm. I very much understand and agree with the notion that even with 80/20 odds, an election can be close. But, when the 538 model comes up with ECVs that are essentially a close of the 2012 results, it does give me pause to reflect.



Leave a Comment