Sunday, October 30, 2016

Cautionary Note on the Election's Public Opinion Polls

I don't know?

I remember the arguments in 2012, and even that one website called "unskewed polls," or something like that, that said Mitt Romney was going to win the election. And of course Gallup saw a Romney surge to 51 percent a couple of days before the election.

I've forgotten which ones, but some or the other polls also had Romney winning Ohio. It was looking really good for the GOP.

And then what? I remember the networks calling Obama's reelection by 6:00pm on the West Coast. There was no cliffhanger. No all night of recounts. Nothing. It was over before it had barely begun.

So, I'm not going to over-analyze the polls this year. I expect Trump's doing better than is shown in most polls, but I'm not going to underestimate the enthusiasm for Hillary Clinton. The Democrats have a ground game. Trump's campaign not so much.

In any case, at IBD, "What Explains the Wide Range of Poll Results Between IBD/TIPP and Others?":

IBD/TIPP is fairly transparent. The typical poll is intended to have between 750 and 900 respondents, a random sample of registered voters. Those are then further winnowed by identifying likely voters, as opposed to just registered voters, through both targeted questions and demographics of the respondents.

The Oct. 24 poll is pretty typical: It yielded 815 likely voters with a margin of error of plus-or-minus 3.6 percentage points at a 95% confidence level. That means, based on the sample, there's a 95% certainty that the "true" support levels for the candidates are within 3.6 percentage points of the reported results.

TechnoMetrica, IBD's polling company, conducts the survey by telephone. It uses both landline and cellphones, with about 35% coming from landlines and 65% coming from cellphones. All of the interviews are done live — no "robocalls" or other dodgy techniques that might bias the outcome.

The numbers are not reported raw. They are adjusted to match the presumed registration percentages of the political parties. That way no party is systematically underrepresented. The same is done for race, gender, region, and party affiliation.

This ensures a more accurate end result than simply relying on raw poll responses. On party affiliation, the presumed mix is as follows: Democrats 37% of likely voters; Republicans a bit over 29%; and independents at 34%.

In the end, that latter category may be key. IBD/TIPP in its latest poll has Trump ahead among independents and "other" by 41% to 32%. That's much wider than most other polls, and one possible explanation for why the poll differs from others.

Polls are, by their very nature, approximations. They use a wide variety of means to guess what literally tens of millions will do, based on just a small sample. Sometimes that yields very big differences, as it has this time.

What about political bias, as some darkly allege? Well, no pollster wants to be wrong. If any poll is really out of whack, it's likely because they missed something in their polling — not political bias.