Saturday, October 22, 2016

An "extremely expensive election to poll accurately"

Gregory Cochran descended to say a few things about political polling and elections outcomes. It is by default worth reading because it's written by him, but commenter pyrrhus' dissenting remark is what really caught my eye:
[Representative samples have] become increasingly difficult to obtain, with many voters having only cellphones, widespread disinterest in answering polls, and candidates challenging the 2 party status quo. Furthermore, when people know that they might be fired for supporting the “wrong” candidate, they are not going to respond….truthfully. That’s how Bernie was 22 points behind in [Michigan] the day before the election and won.
In the primaries we saw polling in the earlier state contests regularly overestimating Trump's performance while polling in later states consistently underestimated it.

A cynic might say it's because the Cathedral wanted the Trump fireworks to go on for awhile for entertainment and pied piper value but later began fearing, once he began to look like the clear GOP favorite, that the Trump forest fire was burning out of control and had to be contained. There are real reasons to suspect that some intentional polling 'irregularities' have occurred.

That's what those of us on the Trump train hope is to be the case, anyway.

Pyrrus concludes thus:
My wife, an expert in this area, comments only that this would be an extremely expensive election to poll accurately.
Let's take a look at the RCP average at the same point at the same point in the 2012 election cycle as we're in now:


To avoid redundancies we're considering a wider time frame in 2012 than we are when we look at RCP's most recent average for 2016. Across 14 polls and spanning nearly three weeks, in 2012 we span from Romney +3 to Obama +3, a range of 6 points.

Compare that to the current RCP average for 2016:



Across 10 polls spanning 11 days we span from Clinton +12 to Trump +2, a range of 14 points.

The polling this time around, at least some of it, is in fact quite inaccurate and will turn out to miss the final mark badly, even with the benefit of margin of error taken into consideration. Including reported MoEs extends our 2016 range from Clinton +15.6 to Trump +5.6--a difference of more than 21 points!--over a sampling period of less than two weeks, all of which was conducted less than a month out from election day.

6 comments:

Random Dude on the Internet said...

These polls primarily act as smelling salts whenever Clinton supporters get too depressed about her failing campaign (low rally attendance, Wikileaks, her obviously poor health, etc.) Then the next step is to have "aggregators" like Nate Silver and RCP present a rosy picture for Democrats. They already have issues getting supporters to the polls. If they found the real numbers or distorted the numbers for Trump the way they did for Hillary, they would just sit out the election in general.

I heard speculation that pollsters will readjust the numbers in the final couple of weeks and I don't think that is true. They will be trying to distort the picture up until election day. The pollsters can just blame it on racist white people for lying to the poor innocent polling companies who have been significantly altering the sample to provide optimism for a dead weight campaign.

If pollsters did sample accurately, it probably would depress Democratic turnout who would prefer to wait for their 2020 non-white savior like Cory Booker or Eric Garcetti instead of eight years of Hillary Clinton. The Clinton wing of the Democrats represent the remaining old and white faction of the party. The Obama wing is non-white, more progressive, and considered the future of the party. They will accept a Hillary victory but more than a few undoubtedly want to flush the Clintons out from the party much like they did to blue dog democrats like Jim Webb a few years ago. Democrats can afford to play the long con with demographics; Republicans don't have that luxury.

Cicatrizatic said...

Note that, even with the low volatility of the last election (compared to this one), the RCP average still missed by more than the MOE. It was off by 3.2 pts, and the typical margin of error for the final polls in 2012 was about 2.5.

The 2014 Senate battleground polls missed wildly. The VA polling average was off by 11.

The Brexit polling average missed by 4 points, also outside the MOE.

I think the RCP average will miss by about 4-5 points this year.

Audacious Epigone said...

Random Dude,

There is a concerted effort across multiple mediums to create the appearance of Trump as a dead-candidate-walking. The same thing happened in the primaries. I specifically recall the word "zombie" being employed. That was ahead of the Northeast and Indiana. The objective was to demoralize Trump supporters. That's what's going on now, in addition to providing cover for voter and electoral fraud.

Cicatrizatic,

If things like yard signs, rally attendance, and enthusiasm are 'intangible' indicators, Trump should be the one who outperforms polling. Four points won't be enough at the moment, but he has a lot of money to burn in these last couple of weeks and the WikiLeaks' slow bleed will continue to weaken the Hilldabeest.

SeanF said...

Cochrane has become a pompous ass who manages the curious trick of being obscure at the same time. He is, in fact, seldom worth reading these days.

Stilicho said...

Gallup's tracking of likely voter party affiliation shows a democrat advantage swinging between 0 and 5 points since 2008. Significantly less than what polls showing a Hillary lead are using to weight their polls (latest ABC poll is weighted 9 points in favor of dems over repubs and significantly underweighs independents if Gallup is close to accurate this cycle). As for predictive power of party affiliation, when was the last time a dem candidate won the popular vote by a 9% margin? Contrast this with Gallup's number of 5% in November 08 and '12. Also, for the 2014 midterms that were such a big Repub victory, the November Gallup affiliation numbers were a dead heat at 28 for both Dems and repubs (independents at 40)... I'd like to see a current poll using that weighting.

Cicatrizatic said...

Stilicho:

You nailed it. For the last 3 presidential elections, the Party ID breakdown on election day has almost precisely matched the average Gallup Party ID margin for that election year.

In 2012, the average Gallup Party ID margin was D+6, and that is exactly what the party breakdown was on election day (D38, R32). The Romney camp was wrong, the polls were not over-sampling Democrats. In fact some of the final polls were using a D+4 sample and thus under-sampling them.

Based on the 2016 Party ID surveys, the margin should be around D+3 to D+4. Lately, in the wake of the various Trump scandals, the Party ID margin in the polls (not in Gallup) has increased to D+7 to D+11. Much of this is likely due to response rate variation - for this, see the "Mythical Swing Voter" paper (link below). Basically, Republicans are less inclined to answer polling questions when their candidate is mired in scandal.

http://www.stat.columbia.edu/~gelman/research/unpublished/swing_voters.pdf

CNN has a new national poll out today with Clinton +5. The Party ID margin is D+7. Not great, but it completely eviscerates the Clinton +12 propaganda that ABC put out over the weekend.

Unless the media launches another coordinated bombshell on Trump, my guess is that in the final week, you will see the Party ID margin tighten for many of the polls. With D+3 samples, they could shrink to Clinton +2 or so. Still, I expect some of the media polls - NBC in particular - to persist in showing a Clinton landslide.

The two historical models (Lichtman, Norpoth) say Trump will win.

LA Times, IBD/TIPP, Rasmussen, and People's Pundit Daily have it tied.

The media and university polls say its going to be a Clinton landslide.

Someone's got to be wrong.