July 6, 2016 by Sam Wang

We are repairing a problem in our HuffPollster feed. We were not scraping states with few polls, and in those cases were using 2012 election results. This is now corrected. Current polls make the following states closer by at least three percentage points compared with 2012: AR, CO, CT, KS, ME, TN, TX, and WA. In addition, Nevada is now flipped to Trump +7% (one poll only). As a result, the EV estimator takes a jump today, as well as the Meta-Margin (from Clinton +3.9% to Clinton +3.3%). This is not a real jump, but a correction to make sure that all polls are included. I apologize for the error. (Update, July 8th: the history has been recalculated using the correct data. The jump is now gone.)

For those who are interested, code is here and daily output files are stored here. Documentation coming soon.

These corrections reflect a general phenomenon in which both parties’ candidates are underperforming compared to 2012, perhaps because of undecided voters. In early July 2012, approximately 90.7 +/- 1.1% (mean +/- SEM) of voters reported a preference for Obama or Romney. In 2016, total Clinton+Trump support is currently 85.2 +/- 2.0%. This year’s reduction in decided voters may reflect reduced enthusiasm for the candidates, particularly Donald Trump, who is notably weak in strongly Republican states. More on this later.


Matt McIrvin says:

That incidentally makes the Meta-Margin extremely consistent with Clinton’s lead in national polls (earlier I’d forgotten that there is a factor of 2 to figure into the definitions of these numbers).

Sam Wang says:

Actually, there’s no factor of 2 involved. The Meta-Margin ought to correspond to the Clinton-Trump margin (currently about Clinton +5%) directly. It does not, probably because of Trump’s weakness in strongly Republican states.

Matt McIrvin says:

I see… I think I was misreading the opening paragraphs of the old “About the Meta-Margin” page.

Kevin says:

Or, it may reflect a willingness to sit out the election, or to vote 3rd party. In particular, it may well stay low all season.

Perry Dane says:

Is there any sound, rigorous, way to predict how undecided voters or late deciders will ultimately vote? Do they tend to break more or less in proportion to other voters, or is there some other discernable pattern?

Kevin says:

Is there a need to retroactively correct the meta margin history in order to run the Bayesian uncertainty model?

Partha Neogy says:

Slightly off-topic. Pollster’s model estimates ( show that Clinton’s numbers have hardly budged from about 45% since last October. Trump’s numbers have been a lot more volatile. This is expected since Clinton is more of a “known quantity.” It is always difficult to guess how Trump will be perceived by his supporters with time. But should we expect Trump’s numbers to settle down after the convention?

Argusdusty says:

I noticed that your new Nevada +7% Trump poll (NBC/WSJ/Marist) is actually an Indiana poll. It appears to be a mistake in the Pollster API that incorrectly labeled the poll (See: ).
I have my own Pollster API scraper that loads all polls and scans for a question name of the form “2016 President: Trump vs. Clinton” or “2016 Presidential GE”, which seems to work well, and for NV it just finds a single +3% Trump Morning Consult poll from Nov, while correctly listing the +7% Trump poll for Indiana.

Sam Wang says:

Excellent catch. I’ve passed it on. They are reachable as – there is a Google group as well.
What is your scraper written in? Care to pass it on? I am sswang at princeton dot edu.

Leave a Reply

Your email address will not be published. Required fields are marked *