October 27, 2016

Here's why I don't trust the polls

Being objective this (or any) election cycle is difficult.  There's a partisan bias inside everyone, and that's the starting point for all of us (even for those who are supposedly still undecided).  It only gets more challenging from there. Rhetoric is heated and on both sides and often off-the-scale inflammatory in nature.  When it comes to the polls, it's concerning that those conducting the polls share these biases and display proclivities based on those biases or at a minimum based on assumptions that may be guided, at least in part, subconsciously by these biases.

Let me cite one specific example.  The latest ABC News poll  has Clinton leading Trump 48% to 42%, a full 6 points ahead.  Taking that at face value, with only two weeks to go to the election, it seems impossible for Donald Trump to make up enough ground nationally, to win enough states to win the election.  That is, until you drill down just a little bit.
This ABC News/Washington Post poll was conducted by landline and cellular telephone Oct. 22-25, 2016, in English and Spanish, among a random national sample of 1,135 likely voters. Results have a margin of sampling error of 3 points, including the design effect. Partisan divisions are 36-29-29 percent, Democrats-Republicans-independents.
Take a look at the partisan affiliation split. 36-29-29 Democrats-Republicans-Independents. Seems plausible on the surface - Democrats +7. Except if you look at the latest Gallup survey on partisan split (admittedly from mid-September, but affiliation does not change that quickly) The split is quite different. Gallup has the split 32-27-40. Democrats only +5 and many more Independents which are leaning Trump.

There were 1135 surveyed. Based on the poll's internals, that would mean 409 Democrats, 329 Republicans and 329 Independents. That would leave 68 who were polled who probably refused to state their affiliation. If those numbers were reflective of Gallup's polled mix, there would be 363 Democrats, 306 Republicans and 454 Independents, leaving 12 undeclared.

Interestingly the poll does not give support among the candidates by party affiliation. But if we assume 89% of Democrats will vote for Clinton, 88% of Republicans will vote for Trump and 53% of independents will vote for Trump and most of the remainder will not divulge their vote (these percentages result in a 48% to 42% result per ABC, but this is just to establish a baseline comparison, I'm not trying to correct the poll here, just make a point) then something interesting can be derived.

Simply by adjusting to reflect the Gallup partisan breakdown the Clinton lead drops from 6% to 3%.

That's just one aspect of polls that can easily be skewed by re-weighting only seemingly slightly can dramatically affect the outcome of polling.  A lot of people conflate oversampling with weighting. The Atlantic calls Trump supporters idiots for not getting this - meanwhile they conveniently have glossed over the weighting issue.

The latest ABC poll is just one example.  Weighting this way is best-guessing and other polls could easily have erred similarly in Trump's favor.  The truth is we just don't know.  In fact - maybe Gallup is wrong and the ABC weighting is correct. The point is the polls are all over the place and subject to bias, intentional or not based on how that weighting is applied.  The only way you can account for this is the aggregation of polls and looking at not specifically the margins but the trends.  Polls at least should be internally consistent.  If polls are consistently among themselves showing Clinton pulling ahead, or Trump narrowing the gap, then that's reasonable to believe they are directionally indicative of a trend.  And that's only if the bias is not intentional, as an intentionally skewed poll would probably be adjusted to reflect the desired outcome. Individual polls don't mean a lot trends do.

No comments:

Post a Comment

Disagreement is always welcome. Please remain civil. Vulgar or disrespectful comments towards anyone will be removed.

Related Posts Plugin for WordPress, Blogger...

Share This