Nate Cohn

Nate Cohn @Nate_Cohn

I've been struggling with this question all year, and my inclination, though it may be controversial, is to avoid weighting on self-reported race unless it just falls well beyond plausibility, including in states without self-reported race on the file. — PolitiTweet.org

Nate Cohn @Nate_Cohn

In a state with a lot of Hispanic voters, it might be the exact opposite! If there's acquiescence bias toward Hispanic ID in our surveys, and we weight it *down* to the census, then we'd have too few Hispanics! — PolitiTweet.org

Nate Cohn @Nate_Cohn

Now, extend this problem to the rest of the US, where voters don't say their race when they register. Our voter file based race variables elsewhere are weaker: they're modeled, not self-reported But if we weight to census targets, will we be biased toward Dems--and not know it? — PolitiTweet.org

Nate Cohn @Nate_Cohn

I wasn't clear: it is out of the question. 30% of GA RVs are black when they registered. That's a *fact*, and you can look it up on GA SOS website. If we weighted on the Census, then our RVs would be 32% black when they registered. That means it's wrong https://t.co/BuzzhghCjI — PolitiTweet.org

Nate Cohn @Nate_Cohn

And interestingly, the people who were white on the voter file, but *didn't* say they were white were even *more* for Trump than those who did call themselves white. So that also throws a wrench in the idea that we're getting the wrong people or something — PolitiTweet.org

Nate Cohn @Nate_Cohn

On the question of whether our *respondents* are biased, one test is to look at the respondents who said they were black on their voter registration form, but not in our survey. They still vote overwhelmingly Democratic, which to me says our sample is probably unbiased — PolitiTweet.org

Nate Cohn @Nate_Cohn

I can think of ev for all three, to some extent or another. --We know design/questions effect answers (see voter file v. our poll!) --The Hisp. question could be subject to acquiescence bias --We know response rates vary by race --We know the CPS is high on southern turnout — PolitiTweet.org

Nate Cohn @Nate_Cohn

This raises (at least) a few possibilities: --our poll is biased v. IRL pop. We get too many Hispanics, for ex. --our design/format leads to bias v census. Here, we have right respondents but they respond differently in this setting --CPS turnout way overshoots black turnout/reg — PolitiTweet.org

Nate Cohn @Nate_Cohn

What does this mean in practice? Well, if we weighted on the *census* variable in GA, like most pollsters do or as we often would in states, we would have had Biden ahead. But 32% of our LVs would have been black on the voter file, which is just factually wrong. — PolitiTweet.org

Nate Cohn @Nate_Cohn

This pattern exists everywhere in the South (SC, GA, NC, FL) where we can weight off of voter file-based race: --More self-rpt Hispanic voters than Census (but right amount by voter file var) --Fewer self-rpt Black voters than Census (but right amount by voter file var) — PolitiTweet.org

Nate Cohn @Nate_Cohn

Now let's compare to the CPS voter reg. estimate for RVs in 2018. Here, it's white, non-Hispanic 61, Black 33 (any race), Hisp 3. Here's our RVs by self-report race again, but here dropping DK for a more direct comp: White 60, Black 30 (any race for comp), Hisp 6. — PolitiTweet.org

Nate Cohn @Nate_Cohn

Let's take GA, because it's recent and you can see the data on the GA SOS website yourself. Our RVs by voter file race: W54, B30, H3, O5, Unknown 9 Our RVs by self-report race: W54, B28, H6, O6, DK5 So same people, but answering the question differently w/dif settings — PolitiTweet.org

Nate Cohn @Nate_Cohn

Here, I'm really comparing three things: --the race our respondents gave when they registered to vote --the race our respondents gave us in our survey in a census-like question --race in the census / CPS-based estimates — PolitiTweet.org

Nate Cohn @Nate_Cohn

Weighting on self-reported race is a ... 100% standard methodological decision that's employed by appx. 99% of pollsters. But as I look at our data in NC/SC/FL/GA, where the voter reg form includes a question about race, I'm left with *a lot* of pause about using it everywhere — PolitiTweet.org

Nate Cohn @Nate_Cohn

We're really locked into like 90% of our design, but at this stage there are two things that I really agonize over: --weighting on self-reported race --everything having to do with early voting data — PolitiTweet.org

Nate Cohn @Nate_Cohn

If PA wasn't the most important state, we wouldn't have to think too hard about all of this and we could just split the difference and declare, Biden+9 nationally and so on. But it is that important, so parsing out just how good those PA polls really are for Biden is important — PolitiTweet.org

Nate Cohn @Nate_Cohn

And if you look everywhere else, you've got mix of great-for-Biden-IA-like polls (Q in TX, Fox MI, Civiqs MN/NV) and murkier-PA-like number (Fox and LD in WI, DFP in MI, maybe the Ipsos AZ number), and some in between (like every FL poll) — PolitiTweet.org

Nate Cohn @Nate_Cohn

And if you look everywhere else, you've got mix of great-for-Biden-IA-like polls (Q in TX, Fox MI, Civiqs MN/NV) than murkier-PA-like number (Fox and LD in WI, DFP in MI, maybe the Ipsos AZ number), and some in between (like every FL poll) — PolitiTweet.org

Nate Cohn @Nate_Cohn

And if you look everywhere else, you've got more of great-for-Biden-IA-like polls (Q in TX, Fox MI, Civiqs MN/NV) than murkier-PA-like number (Fox and LD in WI, DFP in MI, maybe the Ipsos AZ number) — PolitiTweet.org

Nate Cohn @Nate_Cohn

Today's IA phone polls, in contrast, are pretty clear. You've got Biden ahead in a state Clinton lost by 10. You've got a shift toward Biden, on average. And you have firms with relatively limited house effects. — PolitiTweet.org

Nate Cohn @Nate_Cohn

Today's PA phone polls are an odd bunch. On the one hand, they're Biden+7 and fairly consistent. OTOH, three of the four have had a Dem house effect, the fourth doesn't weight on education, and two show a notable shift toward Trump (the other two have no trendline) — PolitiTweet.org

Nate Cohn @Nate_Cohn

One thing I continue to be fascinated about in Fox polling: they just don't show the same white education gap as other pollsters. It's not obvious to me that it works out to anyone's advantage, but it is weird. — PolitiTweet.org

Nate Cohn @Nate_Cohn

Deleted a tweet about the Fox poll that was backward. These are good numbers for Biden, but represent meaningful gains for Trump in WI/PA/OH https://t.co/ATHQyA9kgr — PolitiTweet.org

Nate Cohn @Nate_Cohn

Biden ahead, though the WI/PA/OH polls all represent notable gains for Biden since their last poll https://t.co/ATHQyA9kgr — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @MULawPoll: Pssst! New @MULawPoll will be released Wed. Oct. 28, 12:15 CT, details posted to web by 12:45-1:00CT, depending on how long… — PolitiTweet.org

Nate Cohn @Nate_Cohn

The Florida poll also has an LV/RV gap favoring Democrats, with Biden up 2 pts among RVs https://t.co/Bvrxwm22YW — PolitiTweet.org

Nate Cohn @Nate_Cohn

Yeah, but it's a stretch for there to be a 5 pt LV/RV gap in a high turnout election, let alone one favoring Democrats. https://t.co/BsGvIrWrJU — PolitiTweet.org

Nate Cohn @Nate_Cohn

An important caveat on this CNN/SSRS poll in PA: it's only Biden+5 among registered voters https://t.co/ew6144UrJv — PolitiTweet.org

Nate Cohn @Nate_Cohn

CNN/SSRS Pennsylvania, Biden+10, 53/43 Florida, Biden+4, 50/46 — PolitiTweet.org

Nate Cohn @Nate_Cohn

Biden by 10 in Pennsylvania and 4 in Florida, per new CNN/SSRS poll on TV — PolitiTweet.org