NYT LIVE POLL THREAD:

[Tintrlvr]

Honestly, it seems statistically very unlikely that polls of each of these races would find results within 1-2% of tied. Something odd is going on.

[Tintrlvr]

These polls aren't actually that interesting if they are all just weighted to show up as virtual ties.

That's a good point. It would be pretty suspicious if literally every poll showed a result within 1 or 2%.

Well, they are polling highly competitive districts so far.  It's not that surprising.

Even if every district were in fact tied, sampling error should produce a larger range of results.  If you saw six polls of the same race that were all within the 1% of each other, you would start to suspect herding, often through manipulating the weights to achieve a certain balance in the turnout model.

Exactly. I brought this up earlier. It's suspicious even with just four polls. If each of the races is exactly tied in reality, we should still be seeing some results with sizeable leads for one candidate or another. Instead, we have four polls that are each within 1%. (If you look at the numbers without rounding, all four finished polls show results with a less than 1-point margin.) That's statistically very improbable.

[Tintrlvr]

The way they are wording the description is strange, but I think what they are doing is taking everyone at their word that they definitely will vote but then also counting a lot of people who say they "probably will"  or "might" vote if they have generally voted in the past. You'll notice that the "definitely will vote" electorate is much smaller than their actual total expected votes for every poll.

[Tintrlvr]

Interesting that the Roskam poll looks like it has zero responses from west of Palatine. What's that neighborhood like? It's hard to get a read on it from Google Maps.

It's a very low-density area compared to the rest of the district (especially adjacent areas to the north and east), and a significant portion is the Spring Lake Forest Reserve, so not surprising that they had few contacts there and no completed responses. It's very wealthy and very Republican. It's a small enough part of the district, population-wise, that I wouldn't think a lack of responses skews the result at all.

https://en.wikipedia.org/wiki/Barrington_Hills,_Illinois

[Tintrlvr]

If you are polling from an existing file of registered voters, how do you account for people who haven't registered yet, but will prior to the election day?  Voters in every state still have about a month left to register.  And I have to imagine those voters will heavily favor Democrats.

You ignore those people. They're a small enough portion of the electorate that they wouldn't matter more than a few tenths of a percent at most in margin anyway. Which can certainly make the difference in a very close race, but polling is not precise enough to make distinctions at that level of detail.

[Tintrlvr]

They appear to have finished the poll of TX-23.

That means we can... download the microdata!!!

https://int.nyt.com/newsgraphics/2018/live-polls-2018/races/elections-poll-tx23-1-microdata.csv

They ended up having a sample of 19 18-29 year olds, and the age 18-29 crosstab ended up 65-32 Republican for Hurd.


Now, some details about these 19 people polled, from the microdata:

1)

4 out of the 19 people (21% of them!) self-reported in the poll that they were not actually age 18-29, but were older. 1 of those self-reported that they were in the 35-49 age group, 2 of them reported they were age 50-64, and 1 of them reported they were age 65+.

One possibility here is that they were lying about their age in the poll. An alternative possibility is maybe there were 2 people at the same phone number (e.g. a "Jr." and a "Sr.") that have the same name, but have different ages, and they are actually the Senior, not the Junior.


2)

Of those 19 people, 10 of them said they were voting for Hurd, 7 for Ortiz, and 2 undecided. Without weighting, that would come out to a 53-37 Hurd lead, but the weighting transforms that to the 65-32 lead that is actually shown in the age 18-29 crosstab.

From their methodological description, the weighting sounds like it will be partly based on turnout scores, and partly based on demographics. The issue here is that the turnout scores seem to be particularly heterogeneous for the age 18-29s. There are just a few voters with high turnout scores (80%+) whereas most of the 19 people have turnout scores that are much lower (13 of 19 are below 50%, and 8 of 19 are below 25%).

Now, the issue is that the person with the 2nd highest turnout score (highest is 92.25%, 2nd highest is 85.25%) happens to be that same voter who said that they are not actually age 18-29 at all, but self-reported that they were age 65+ (the voter with the 92.25% turnout score is also voting Republican). So that voter, who may well not in fact be 18-29 at all, seems to be getting disproportionate weight in the 18-29 category. And in turn, the age 18-29 sample gets disproportionate weight in the poll overall due to the horrendously low response rate from age 18-29s.

Also, only 9/19 of the age 18-29s included in the sample said that they were Hispanic.

It’s the black-18-year-old-for-Trump problem all over again that YouGov (?) was getting flak over in 2016 in their panel polling. (Not certain it was YouGov but was someone with an online panel that didn’t change.) They had huge swings to and away from Trump based exclusively on whether that person was responding to the poll that round. It’s a major hazard of modern polling. They should be dropping anyone whose self-reported age is inconsistent with their registration data from the poll.

[Tintrlvr]

I was looking at the response rates and asking myself if there were any interesting patterns:

Minnesota 8:       2.70%
Minnesota 3:       2.38%
Illinois 12:           2.36%
Wisconsin 1:       2.15% (preliminary)
Virginia 7:           2.01% (preliminary)
West Virginia 3:  1.96%
Kentucky 6:        1.69%
California 48:     1.44%
Illinois 6:            1.41%
Texas 23:           1.38%

So might there be patterns like rural midwesterners being more likely to respond and suburban sunbelters less likely? Or is there too much noise coming from the polling firm calling people at different times of the day and other factors?

There's a strong correlation with typical voter turnout by district so I think this is significant but not especially surprising.

[Tintrlvr]

Looks like they're also doing FL-26. Will be interesting to see if it mirrors TX-23.

The dynamics of the two races are quite different to the point of not really being comparable. Latinos in South Florida vote very differently from (and more frequently than) Latinos in South Texas and are much easier to contact. That said, a similar result on the surface is not out of the question.

[Tintrlvr]

It appears that only Republicans are answering their phones this afternoon.

I've actually noticed on these polls that the Republican candidates seem to do better in the afternoon/early evening while the Democratic candidates seem to do better later in the evening. I don't really have a good explanation for why that might be, and the data is hardly definitive just from these polls.

[Tintrlvr]

It's fair to say that the ME-02 and VA-07 polls should be questioned. Are they one data point? Absolutely. NYT/Siena is doing a great job. But it's not "unskewing" to purely say that the samples are off ("Nonwhites" going to Republicans in ME-02 by like 20%?) The samples are so small that the entire poll is not 100% indicative. Same with VA-07 and young people.

FL-26 is surprising, more so because Curbelo right now has such an amazing favorable rating and the D is unknown. However, it's a Clinton +16 district, so it's also not entirely surprising to see partisanship win and have the D up.

ME-02 has a de minimus amount of nonwhites to start with though, studying inaccuracies in polling such a small subgroup is a triviality.

I agree with that. The trouble is with groups that are undersampled and then get weighted up heavily. That is mainly a problem with young voters, who are very hard for pollsters to reach and so often get weighted up on extremely small sample sizes to be a substantial portion of the overall poll result. If that tiny sample size is unrepresentative, which it often is, it can throw the entire poll way off. The added bonus here is that Siena is apparently ignoring that some people who it thought were young voters before they were contacted are self-reporting as being much older, which means that even their paltry sample of young voters is sometimes not only unrepresentative but also consists heavily of voters who aren't even young to begin with.

Non-whites in ME-02 are not being weighted up because they are already a tiny portion of the population there, so having a 3% subsample be probably way off is not really a big deal.

[Tintrlvr]

I was actually referring to the minority vote and young vote, which had a low response rate, and were over represented. If 2016 margins were applied using the sames weighing as the pollster, Golden would be down by 2. I did the math for this one.

Aside from the problems with this that everyone else has already mentioned, there is a tendency for non-white voters in areas where there are very few non-white voters to be less Democratic than non-white voters in which there are large concentrations of other non-white voters.

There are systematic socioeconomic/demographic differences between, for example, African Americans in Maine and the average nationwide African American.

So it is absolutely a mistake to plug in the assumption that Dems will do as well with non-whites in ME-02 as they generally do in other places across the country.

This is more true for some groups than others, in particular Asian Americans. Blacks are pretty uniformly the same level of Democratic-voting everywhere. Maine might be slightly different because a large portion of blacks in Maine are Somali refugees/asylees rather than "African American" blacks (descended from slaves), but even then probably not by a big margin. This is not really an explanation for the figures. The real explanation is that the sample size is tiny and therefore not really worth looking at, and, since it's not being weighted up, is not really affecting the poll results much (even if non-whites were actually 90% Golden and are being reported at 55% Poliquin, that would still only be a net swing of about 1% from Poliquin to Golden as compared to the poll result).

[Tintrlvr]

I guess it could be the case that partisan gravity is just too much for Curbelo to overcome in FL-26. Too bad RINO Tom can't enjoy his Cub's game, now that Curbelo's down by 10.

I wonder what the sub-composition of their Hispanic sample is in terms of Cuban/Non-Cuban. Since apparently they are having trouble getting responses from Hispanics there, it is also possible the problem could be focused particularly with Cubans or Non-Cubans, which would make a difference given that Non-Cubans are significantly more Dem.

They still have Hispanics favoring Curbelo by 7 points (Mucarsel-Powell's margin is entirely from blacks and whites), so I don't think they're undersampling Cubans.

[Tintrlvr]

I think it will be more helpful to look at these polls as a group once they're completed. 500 is not a great sample size, but if one looks at all the polls in the end, one can get a good idea of which party is likely to control the House.

Anyway, I'm kind of shocked to see Culberson up by that much, but it's still early.

The TX-07 poll seems odd in that the margin seems to be generated in significant part by their weighting (which isn't as true for other polls). Notably, NYT presents four weighting alternatives. For TX-07, those alternatives have truly enormous implications:

Weight using Census data: Fletcher +7
Don't weight by primary vote: Culberson +5
Don't weight by education: Culberson +16
NYT/Siena weighting: Culberson +17

None of the other polls so far show anywhere close to such a difference - it's usually no more than 2-3 points either way.

Fletcher was also leading big last night, and the current result is not only outside the MOE of where the poll was last night but really, really far outside the MOE, to the point where the odds of both results having happened should be vanishingly small. Of course, this doesn't tell us which is inaccurate but does suggest that they are having trouble polling and/or weighting the race accurately.

[Tintrlvr]

TX-07 has now tightened up substantially, to Culberson +6 from Culberson +17 at his peak.

One of the other things that stands out about TX-07 so far aside from the weighting is the extremely low rate of undecided voters - just 2% currently and for a while was at 1%. Other districts have had undecideds hovering around 8-12% or so, with the lowest thus far among completed polls at 6% (in WI-01) and highest at 13% (in MN-08).

[Tintrlvr]

Hope they poll
TX-32
WA-5
NC-13
PA-1
KS-3
MN-2
NE-2
IL-14
NJ-3
NJ-7
NY-22
CA-25
GA-7
MT-AL
MI-8

Based on how quickly they're moving through districts/polls, I'm sure we'll get all of these by election day. Unless NYT pulls the plug on the process for some reason.

[Tintrlvr]

They are polling Kavanaugh in NJ-07. That may actually be more interesting than the topline itself. It will be interesting how much he overperforms or underperforms Lance.

LOL the 1st respondent is a 65+ year old white male Republican who supports Kavanaugh but is voting Dem. Ha!

And approves of Trump. Bizarre that this guy would be voting D. I've always wondered about the rate of error from underpaid call center workers...

[Tintrlvr]

Great response rate thus far in IA-01. Already passed NJ-07 in responses with only about 3/4 of the calls.

[Tintrlvr]

No wonder Blum was triaged. He’s getting Finkenauer’d six ways from Sunday.

[Tintrlvr]

IA-01 now has as many responses as CA-25 with half the number of calls. Although, looking back on them, KS-02 had slightly a better response rate overall. IA-01 is running ahead of MN-08 (and way ahead of WI-01 and MN-03) on response rate, though.

[Tintrlvr]

Why is Finkenauer's name in bold? Same thing happened to Phillips and Crow in their polls. Is it down to the size of her lead?

I think they bold the name when the lead is outside of the margin of error.

[Tintrlvr]

They apparently started polling CA-49 last night but just put it up. Levin (D) has an enormous lead (58-37), but there are only 40 respondents thus far.

[Tintrlvr]

I never said that Hill is finished. Quite the contrary, I said that she could easily make up a 2-point deficit by November (if the poll is accurate.) I was just disappointed to see the poll trend so sharply away from her after she was up by 7 with more than half of all responses in. One thing I am curious about is the implications of IA-01 being so lopsided. While it is one poll, if Blum really is losing by this much, could similar districts in the Midwest end up being more favorable for the Democrats than we're expecting? This races don't all happen in a vacuum.

It is certainly true that each race doesn't happen in a vacuum and to some degree you can infer things about similar districts from a result in one district, but Blum is also an especially weak incumbent who is a very poor fit for his district.

[Tintrlvr]

I never said that Hill is finished. Quite the contrary, I said that she could easily make up a 2-point deficit by November (if the poll is accurate.) I was just disappointed to see the poll trend so sharply away from her after she was up by 7 with more than half of all responses in. One thing I am curious about is the implications of IA-01 being so lopsided. While it is one poll, if Blum really is losing by this much, could similar districts in the Midwest end up being more favorable for the Democrats than we're expecting? This races don't all happen in a vacuum.

It is certainly true that each race doesn't happen in a vacuum and to some degree you can infer things about similar districts from a result in one district, but Blum is also an especially weak incumbent who is a very poor fit for his district.

What basis do you have to say that? Blum outran Trump substantially in 2016... he won by 8 while Trump won his district by 4. And Trump did fantastically well in Iowa, so those are some darn good numbers for Blum.

Blum's opponent in 2016 was a perpetual loser and party-switcher who was not exactly the strongest candidate the Democrats could have put up.

[Tintrlvr]

TX-32 is still very early (88 responses) but this is an interesting divergence:

Sessions 50
Allred 41

O'Rourke 55
Cruz 43

O'Rourke's numbers so far (right now, 57-42), indicate around a 3-5 point lead statewide if applied statewide. It does suggest, when compared to the congressional numbers, O'Rourke is at least expanding the "Clinton coalition" in Texas that does explain where he gets the voters to be competitive statewide, though still, I think, probably not to win.

(Also worth pointing out that this race currently has only 3 18-29s responding, two of whom are voting R and the third undecided, and they're being weighted up heavily, so I would not take Sessions' current lead very seriously.)

[Tintrlvr]

Polling NJ-07 during Yom Kippur might not have been the best idea.

There aren't that many Jews in that part of the state, barely more than other random wealthy suburbs elsewhere in the country. Would be different if they were polling NJ-05.