David M Rothschild on Posted on

It is 1:00 PM ET on Saturday, November 5, 2016 and Democratic nominee Hillary Clinton is 86 percent likely to be our next president (Republican nominee Donald Trump is 14 percent). Clinton is winning, because she is 90 percent or more to win states that equal 274 Electoral Votes, she is about 75 percent for another 4 in New Hampshire, and about 70 percent in both Florida (29 votes) and North Carolina (15 votes). Further, she about 30 percent for 18 votes in OH and about 20 percent for another 11 votes in Arizona. She needs 270 to win.


1) Polling averages do not miss by that much: In a review of 2000, 2004, 2008, 2012 we created a simple average of all reputable polls in the last three week for every presidential race that had at least four polls (i.e., every state that counts). The largest error of a polling average was 5 percentage points. No candidate up by 6 percentage points or more ever lost that state.

2) RealClearPolitics (RCP) is a GOP-leaning polling average: RCP accepts polls that are obviously GOP-leaning. They include polls from GOP organizations that admit their bias and others that show it. These include polls that are landline only (conditional on demographics, voters on landlines are more Republican). And, while RCP just takes a simple average, they pick the cut-off to optimize for the GOP. For example, right now they have Colorado at +2.9 for Clinton and include seven polls. The most recent was 11/3-11/4 and the most ancient 10/28-10/31. Of the seven polls four are right-wing and would be rejected by Huffington Post’s Pollster. At least two of them are landline only and one, Gravis, is run by Breitbart (whose executive chairman is taking a leave to run the Trump campaign!). As far as timing, they do not include a 10/27-11/1 CNN poll in Pennsylvania that would boost Clinton’s average there, even though it was taken on 10/27-11/1, even though that timing working in acceptable Colorado. Anyway, you get the picture.

3) Polling averages are pretty good, and RCP is right-leaning; thus, if Clinton is winning RCP great sign for her. Clinton is, on RCP, +4 or more in 239 Electoral Votes and +2.5 or more in 268. Again 270 needed to win.

4) Polling is really a mixture of two things: voting population and support by voting population. Every poll does two things, it estimates the voting population by asking registered people if they will vote (or some other voter screen) and then it asks them who they will vote for. For this entire campaign I talked about two competing question. (1) Would Clinton’s GOTV help her over-perform the polls? What I am really saying is will her GOTV lead to a more favorable voting population for Clinton than the polls are estimating? (2) Would “shy Trump voters”, hiding as third-party or undecided in polls, push him to over-perform the polls? What I am really saying is will Trump get more favorable support from likely voters than people are saying to the pollsters?

5) We have no evidence of shy Trump voters: Kyle Dropp at Morning Consult confirmed with an experiment what I had already seen with the Pollster data, there is no evidence of shy Trump voters. What we are seeing is Clinton is doing slightly better in live telephone polls than online, but that Trump remains the same in each. This is an artifact of live telephone operators not providing the option of don’t know or undecided, but forcing the respondent to state that if they do not want to name a major party candidate. This is actually good for Clinton, because very few people show up to vote and leave the president line blank!

6) We have evidence that GOTV is working (or, at least, the voter population will be as favorable to the Democrats as 2012): The short of it is that African-American percentage of the voters is down, but registered Democratic Latino voting is way up. Also, with some Democrats avoiding pollsters last weekend due to the Comey/FBI letter, polls were showing a depressed Democratic turnout (what we call partisan non-response), which will make even more likely the voting population is more Democratic than the polls estimated.

7) If we see no evidence of shy Trump voters shifting support his way and we see evidence of turnout looking good for the Democrats, it is increasingly unlikely that Trump will see a (1) massive (2) correlated (3) poll-error in his favor (at least in places with sizable Latino populations): that is why Nevada and Colorado are both swinging hard to Clinton, despite a small, but consistent for polling lead for Clinton in Colorado and downright mixed in Nevada. In Colorado, where Democrats had less early voting, but won by 5.5 percent in 2012, they have a small lead in early voting. In Nevada, where Democrats won by 7 percent in 2012, they are up the same amount in early voting as 2012.

8) FiveThirtyEight does not include early voting, thus their model does not know that turnout is looking for Democrats. I have argued before that it is very unlikely to see polling error that is: (1) massive (2) correlated (3) in Trump’s direction that FiveThirtyEight’s model accounts for. But, this is especially true as turnout in the form of early voting begins to dominate estimated turnout in the form of polling. The irrelevance of the FiveThirtyEight model in light of early voting  was seen clearly yesterday as Democratic vote share of early voting surged, but polling was mixed:


The raw Betfair topline price surged, the PredictWise probability also increased, meanwhile the FiveThirtyEight probability cratered. They are currently showing Trump at 35.3 percent versus PredictWise’s 14 percent.