Princeton Election Consortium

Innovations in democracy since 2004

Nov 03: Biden 342 EV (D+5.3% from toss-up), Senate 53 D, 47 R (D+3.9%), House control D+4.6%
Moneyball states: President AZ NE-2 NV, Senate MT ME AK, Legislatures KS TX NC

What’s more important than a Presidential probability?

August 9th, 2020, 2:21pm by Sam Wang


Presidential models are mainly of interest in close elections, which have been the norm since 2000. This year, Joe Biden occasionally gets within sight of 400 electoral votes, which we haven’t seen since Bill Clinton or George H.W. Bush. In addition, U.S. Senate and state legislative races are of paramount importance. So we’ve spent less time on the Presidential race.

But today, let’s drill into the presidential model a bit. It’s modified to take into account lessons from 2016. In this piece we’ll tell you Joe Biden’s November win probability – but also why we are deemphasizing it.

At its core, the PEC Presidential model is the same as in 2012 and 2016. It starts with a snapshot of current polls, converting them to a virtual lead in terms of the Electoral College. That virtual lead, which corresponds to popular support, is called the Meta-margin. The race would have to swing by that much toward the underdog to create a perfect electoral-vote tossup. The Meta-Margin is a sharp measure that responds promptly and with low noise to political events. An example is Trump’s response to the coronavirus pandemic in mid-April.

To generate a prediction, the Meta-margin is projected forward in time by assuming drift in either direction. That assumption takes the form of a distribution that captures (1) movement in opinion in either direction, (2) a systematic, correlated error in polling across all states, and (3) the likelihood that conditions will revert back to some expected average (the Bayesian “prior”). The Bayesian prior assumes that Trump is likely to perform in the range that he has from February until now, in other words somewhat better than current conditions.

In response to 2016, we’ve changed the drift function. On time scales of greater than two months, the drift is set smaller because volatility in public opinion has gone down so much since the mid-1990s. But more important, in the home stretch of the campaign, the drift function never gets smaller than two percentage points. This takes into account the polling error of several points that occurred in 2016.

Here is what the drift function looked like in 2016 versus now:

As you can see, it now has a pieced-together look. But for what we need, a “hurricane strike zone” in November, it will do.

The red “hurricane strike zone” represents the one-sigma range, where about two-thirds of outcomes will fall. Yellow represents two-sigma, or 95% of outcomes.

The new drift function would have given Hillary Clinton about an 80% win probability in 2016 on the night before the election, as opposed to what we published on the home stretch, 93%.

With the new increased uncertainty, the random drift probability currently shows a high win probability for Biden. If you want to track it, it’s in the data folder. The Bayesian probability, which is higher, is the first value. However, these probabilities do not get into factors that are exogenous to the polls, such as difficulties in voting by mail. They are also not what I consider to be the main product of the model. I strongly advise you to think twice before focusing on it.

As I wrote in the aftermath of the 2016 election and again in the Columbia Journalism Review, probabilities are the wrong unit of measure for 2020. They encourage complacency rather than active citizenship. And more than ever, now is the time for active citizenship.

Instead, we are emphasizing the Meta-margin. At 6.4 points favoring Biden, this means that the margins would have to move, or be wrong, by that much to create tossup conditions between Biden and Trump. That is more than twice as large as the polling error in 2016.

More important than the Meta-margin is voter power. We are emphasizing voter power for three types of important races: the Presidency, U.S. Senate races, and state legislative races. Voter power is defined as the ability of one or a small group of voters to move the probability of one of three important questions:

President: providing tipping-point support to get Biden or Trump above 270 EV.

U.S. Senate: providing tipping-point support to get 50 Democratic plus independent seats.

State legislatures: providing tipping-point support to gain control of one chamber of the state legislature, in states where this would lead to bipartisan control over redistricting.

In all cases, voter power tells you where to put your efforts as a citizen. It’s what we are calling the “Moneyball 2020” approach.

As of today, the states with highest per-voter power are:

President: Nevada, Arizona, North Carolina.

U.S. Senate: Montana, Alaska, South Carolina, Kansas.

State legislatures: Kansas, Texas, North Carolina.

States listed more than once are in italic.

Voters in North Carolina are especially powerful this year. That is no surprise, since North Carolina is such a regular partisan battle ground. But there are also some surprises: Montana, Texas, and most of all Kansas.

In weeks to come, we will dive into some of the states and explain why they are so important.

If you like what we do, you can contribute to the Princeton Gerrymandering Project. If you want to support your favorite candidate, you can give in the right side bar to ActBlue for Democrats (state legislative or U.S. Senate) or WinRed for Republicans (both categories combined).

Tags: 2020 Election · Moneyball · President · Redistricting

20 Comments so far ↓

  • George Bussey

    Thanks for the update, and especially the “money ball” info. With personal ties to NC and SC, I look forward to supporting those two Dem senators and the state legislative races in NC.
    One question – using your “hurricane strike zone analogy,” it seems that it has been relatively, if not totally stationary so far this cycle. Is there a point in time, i.e., “X days before the election” or “Y days before early voting starts in key states” that, just like the hurricane strike zone, it will start to narrow to cover a smaller range of likely outcomes? And if so, when would that be? Thanks.

    • Sam Wang

      The strike zone will narrow gradually from now until the start of October, as the one-sigma drift decreases. It will stay the same size from then until Election Day, because the one-sigma curve is flat. That is what a minimum uncertainty means.

  • Amitabh Lath

    Should the drift magnitude take into account the number of undecided voters? I imagine a race with 20% undecided has a much higher probability to drift than one with 2%.

    • Sam Wang

      tl;dr: probably won’t do that.

      In the past, undecideds have broken close enough to equally that the two-candidate margin approximates the final result pretty well.

      We would need some way of assessing whether undecideds would break evenly this year. Cognitively, “undecided” just means “I am unable to verbalize.” Undecided voters are pretty committed, even if they don’t know it.

      I suppose we could take a totally different approach. Presidents usually get a vote share within 0-3 points above their approval rating. Trump’s approval (39-42%) suggests that he would get 39-45% of the vote. That implies up to a 22-point loss to Biden in terms of the popular vote. The last time anything like that occurred as Nixon v. McGovern 1972 (520-17 in the Electoral College), and before that Johnson v. Goldwater 1964 (486-52 in the Electoral College). For that kind of election, what you really need is to stop watching the presidential race and go work on Senate and state legislative races.

  • Frances Smith

    Yesterday I sent money to Bullock, McGrath, and Harrison. I’m very pleased to see that I sent my money to the right places. Kansas contribution is on my to do list.

    I go to your website every day. Thank you for the much needed information you send our way..

    • Sam Wang

      Thanks. That’s great to hear.

      To be honest, I think McGrath has become a somewhat worse investment over the last month. She looked like a contender for a while, but after the primary the race swung toward McConnell a fair bit. This happened before. In 2014 Grimes looked strong against McConnell, but ended up underperforming polls and losing decisively. As of today, the median margin is McConnell +5%. Worth watching, but not the most leveraged use of money at the moment.

      Compared with Kentucky, I would be more tempted by the South Carolina, Kansas, and Iowa races at this particular moment.

      The ActBlue thermometer at right reflects current optimization. An exception is Sara Gideon (D-ME), currently not the highest-return choice. But the software is now allowing us to remove it. You can de-select her manually.

  • George Bussey

    Thanks for the strike zone explanation. Did my contributions to MT Senate and NC legislature. And just got done reading about how AZ might be a legislative flip. Did you not list AZ b/c 1) you don’t think it will flip, 2) it is a slam dunk to flip, or 3) the $ required for the desired outcome doesn’t fit the moneyball strategy? FYI, like others, I get more and more obsessive about checking in as the days get closer.

    • Sam Wang

      Arizona has an independent redistricting commission, so the legislature has no say. I agree that the Arizona House and Senate could go either way.

      Other legislative chambers that could potentially switch control, but without affecting redistricting: Iowa House, Michigan House, and Pennsylvania House and Senate.

  • MD

    Update: I see this answer in another post of yours: “red zone (one-sigma, about two-thirds of outcomes) and a yellow zone (two sigma, about 95% of outcomes)”.

    And I take it the asymmetry is due to the Bayesian prior: Trump is likely to drift back toward his better performance earlier in the year.

  • Sherean

    I’m in GA. Better to concentrate my efforts on Ossoff Senate race or local State House and Senate? I have limited time and funds and want them to go where they matter most. One of the state House seats is currently R, but we only lost by 700 votes last time. Same D candidate in an area (ATL northern suburbs) that has been turning. Relatedly, should I sign up for phone banking for Biden or is my time better spent on those other races?

    • Sam Wang

      Logically, between those two choices you should focus on local House/Senate races since every vote you turn out at that level is also a likely vote at the U.S. Senate level.

      Phone banking for President: choose a state in the banner at the top of the page. Currently Arkansas, Alaska, or…Georgia again.

  • Stephen Huegel

    New to this site. Appreciate all the hard work and thought from PEC. Esp useful are the links so that I can contribute with little effort (!) and be effective. Will support PEC from here on out

    Sherean asks good question Research shows how difficult changing anyone’s mind just on the basis of facts.

  • Matt McIrvin

    I kind of think you shouldn’t even have bothered to calculate a probability for this year. The Meta-Margin is where it’s at; leave others to imagine what they think the correlated error distribution is, because everyone’s gonna have a different idea and you can’t really calculate an answer from first principles.

  • Brian Foster

    Sam, quick question on the Voter Power and Moneyball concepts. If it is true, as most commentators seem to agree, that very few voters split their ticket, than wouldn’t the highest leverage contribution always be to to your presidential candidate of choice? In other words, wouldn’t that contribution have a ripple benefit for all same-party senate and house candidates down-ticket because of the reality that most voters will stick to one party?

    • Sam Wang

      No, that is exactly backwards. When you have a choice, always give at the local level.

      For example, if you want to influence the North Carolina Presidential or Senate race, it is always better to push for votes in specific state N.C. legislative districts that are important for determining control of that chamber, and therefore matter for redistricting. Those votes count for legislative control- and also for the federal races.

      Another example is Kansas. If you are a Kansas Democrat interested in influencing both the U.S. Senate race and state legislative control, you are better off giving to Johnson County Democrats than to Barbara Bollier. The votes you mobilize still influence the statewide question.

  • Kevin B.

    How do you model Trump’s chances of winning the electoral college while losing the popular vote? Does the Meta Margin reflect the difference from a national vote of, say, Biden +2, at which point we assume an electoral college tossup?

    • Sam Wang

      The Meta-Margin is relative to an electoral tossup, i.e. half of outcomes are >269 EV and half are <269 EV. This year, that looks like it will happen for a national popular-vote margin of Biden+2.5%

      So, what's the likelihood of a PV-EV split? Not that great, but yes, if Trump wins the Electoral College it will almost certainly be another popular-vote loss.

  • Sam Wang

    Everyone: If the post says there’s something more important than probability, then questions about probability are low priority. Did you get the memo? There are other priorities this year!

Leave a Comment