I have three thoughts to consider on today’s polling which included two polls likely voters on live telephones. One of them, October 17-22, showed Trump leading by 1 points and another, October 20-22, showed Clinton leading by 12 points.

PredictWise20161023b

Margin of Error is Bigger than Reported: I wrote up a recent paper with Sharad Goel in the New York Times a few week ago; we show that the true margin of error for probability-based polling is closer to double the reported margin of error. Thus, the stated 3.5 and 3.6 MOE should be read as 7 each. But, the error on the average of polls is quite small, closer to 2 percentage points on average.

Measurement Error: The IBD poll is clearly asking the questions in a way to get undecideds and thirty party supports to say undecided or other (and/or the ABC poll is forcing people to make a decision). That may be implicit or explicit, but it happening. All of the movement between these polls is between Clinton and undecided. That makes sense, because Clinton is the likely recipient of more undecided and third party support than Trump.

Overall Picture Clear: Here is every poll started after the Access Hollywood tapes in either RealClearPolitics or Huffington Post’s Pollster (I just dropped the LAT/USC poll, because it obviously meaningless). Notice that Trump is very tightly getting between 39.4 and 43.6 percent support, while Clinton is between 45.1 and 51.7 percent. The standard deviation includes the middle 68 percent of polls. Both the IBD and ABC poll had Trump at un-winnable numbers 43 and 41 percent. Clinton is getting variation because measurement error fluctuates the amount of her undecided and thirty party supporters report for her. Trump is getting little variation because he has a core group of 41-42 percent support that is unwavering, but is a relatively hard ceiling that he is not growing.

PredictWise20161023a