Jonathan Chavez

2/17 Polling analysis and aggregators general weakness right now is that it’s doing a pretty terrible job of dealing with age because it has been modeling (not all, but many) as a standard demographic variable when it isn’t.

3/17 Other demographic variables are stable: there are predictable changes in the overall electorate but individuals don’t morph race/gender/education level (this can change but it’s pretty static)

4/17 For Bayesian-based analysis, it has historically been okay to treat age that way because age gaps, while present, have been relatively small, and there have been predictable patterns of evolution of individuals over time.

5/17 This started to change in 2004 and really in 2008. As I’ve said a number of times, the age gap in the 2000 election was non-existent: While Gore won young voters it was only by a few points. Bush won old voters by a similar margin.

6/17 Shit gets real (and really complicated) when you start seeing birth year cohort as an identity, which the evidence at this point overwhelmingly points towards it being (in terms of political views and engagement).

7/17 The best example of this is that older Millennials (myself included) are not changing politically. When people talked about young voters in 2004 they were talking about voters born after about 1978. That cohort has an identity.

8/17 This ladders up to me being somewhat skeptical of what I’ve seen in a lot of public polls. Pollsters are too aggressively muxing with age adjustments. This isn’t just with 18-29 year old voters, it’s with 30-40 year old voters too.

9/17 In 2016, my general feeling was IF structural error existed, it was going to be with White Working Class voters(https://twitter.com/jschavez/status/793954586126061568?s=21 …). My general sense now is that if there is a source of error it’s with voters under 40.

10/17 Upshot/Sienna have done a great public service with their polling, but it’s gotten to the point that they have an outsized role to play. They are leading the herd and numbers are herding towards them.

11/17 That’s not their fault at all, but there’s real pressure on pollsters for numbers to be “right.” And because of their prominence they are viewed as the default case by pollsters in this election. Stray from them at your own peril.

12/17 I understand the argument for making adjustments for past levels of truthfulness, but I’ve got a whole slew of reasons why I think they have treated this specific adjustment incorrectly, and so far turnout seems to be tilting in my direction on this specific point.

13/17 I wish I could get into more specifics on that, but it takes too much data, some of which is not really releasable. I do not believe that if 65+ will increase their share of the electorate this time around. I don’t find the data I have looked at on that point convincing.

14/17 Add this up, and the conclusion from the data is that if a structural error exists, it’s coming from incorrect treatment of birth year cohorts in age-based truthfulness adjustments and unintended herding that flows from that.

15/17 I don’t know for a fact that will happen, but it seems to be where the largest potential source of systematic error is coming from, along with a high level of unpredictability of 18-19 YOs this year vs past years. They look VERY different in the data compared to the past.

16/17 Hispanic attitudes towards answering the phone have also potentially changed in the past 2 years, introducing some area for potential error as well.