IA-Selzer: Trump +7

[RI]

That Selzer was right and so many others were wrong may suggest education weighting is not the panacea so many thought. Selzer doesn't do it.

[RI]

Something I posted a few weeks back:

We've been over this.

The problem with 2016 was that pollsters were not controlling for education because historically, education level was not a reliable predictor of voting patterns. That has changed, and now the more educated you are the more likely you are to vote Democrat.

That gap is especially true among white voters.

Pollsters are now controlling for education.

So unless there's something else they're missing, they're probably not overestimating Biden's white support.

Here's the thing: weighting by education only solves your problem if the small sample of non-college voters you do successfully get is conditionally representative of the non-responsive non-college population.

As I doubt pollsters' ability to reach non-college whites has increased since 2016, this means inflating the importance of the few non-college respondents you do have, which is a group which tends to be disproportionately elderly. If, for example, elderly non-college voters have moved left since 2016 while others have not (or even moved right), education weighting may actually make polls less accurate than without education weighting.

[RI]

Do we know much about Selzers methodology here? I remember hearing a podcast about how she did polling of Iowa causes, but I can't imagine that she does the same thing here?

IIRC, Selzer has a very laissez-faire approach to polling: use very little weighing (only age, sex, and congressional district), make few prior assumptions and let the data tell you what it tells you. She doesn't take past voting history into account. She doesn't make efforts to re-sample missed voters.

The only thing somewhat notable is that she records demographics of non-voters as well as voters in order to weigh the entire sample to the overall population using Census data.