Friday, October 28, 2016

Reuters-Ipsos: Ask the Oregonians

The following table shows how each state's percentage of the Reuters-Ipsos' national polling sample running from 7/28 through 10/25 compared to its percentage of the nationwide vote in the 2012 presidential election. In other words, it shows how over- and under-weighted each state is in determining R-I's nationwide polling results. Figures over 100% indicate that a state is being oversampled by R-I; figures under 100% indicate that a state is being undersampled. States in red are those Romney won by 6+ points; blue that Obama won by 6+ points, and purple those with margins of less than 6 points:

StateWeightn =
1) Oregon215.5%683
2) Nevada212.8%389
3) Kansas208.9%435
4) Vermont208.8%112
5) Maine199.3%254
6) Idaho193.2%225
7) Delaware180.0%135
8) North Dakota175.6%101
9) Connecticut172.4%488
10) Arizona170.8%709
11) Oklahoma168.4%413
12) South Dakota168.3%110
13) Nebraska166.7%238
14) Kentucky166.3%540
15) Montana165.7%143
16) Minnesota163.7%862
17) Arkansas163.4%313
18) West Virginia163.4%197
19) Tennessee162.4%723
20) Colorado160.3%739
21) New Hampshire155.6%200
22) Utah153.5%280
23) Wisconsin145.5%810
24) New Jersey145.3%962
25) Hawaii141.2%111
26) Alabama140.8%531
27) Iowa138.0%394
28) Maryland135.1%658
29) Massachusetts134.8%770
30) Louisiana134.3%484
31) Virginia125.3%874
32) Washington123.1%689
33) Missouri118.8%591
34) Indiana118.2%559
35) New Mexico116.6%161
36) South Carolina109.3%389
37) Georgia96.5%683
38) Mississippi90.1%211
39) WY+RI+DC+AK88.6%204
43) Michigan79.5%684
44) Illinois79.4%752
45) Ohio77.9%786
46) North Carolina73.5%601
47) Pennsylvania61.6%642
48) Florida48.4%747
49) New York45.8%587
50) Texas40.1%580
51) California24.0%561

As you might guess just by looking at the results, there's no statistically significant correlation between 2012 results and the apparent R-I sampling discrepancies for 2016. No blatant partisan slant is apparent in the ranking.

You'd be forgiven for assuming that because the big three electorally-safe states of California, Texas, and New York are severely undersampled, R-I is intentionally pulling more heavily from swing states and less so from uncontested ones. But rounding out the bottom and coming in just ahead of these three are Florida, Pennsylvania, North Carolina, and Ohio. These are arguably the four most determinative swing states in the whole election and yet they, too, are substantially undersampled.

The final column in the table includes the actual number of survey responses by state in anticipation of the objection that it's as easy to accurately sample California when n = 1,000 as it is to sample Delaware when n = 1,000.

Lacking said information that would be a reasonable objection. As presented above, however, it's clear that there is no sample size threshold being striven for in the sampling methodology. Four of the states had insufficient samples for R-I to even show individual results, with an average of 51 people polled in each of them during the three month period, while Oregon had 683 and Minnesota 862 sampled over the same time frame.

As noted in a previous post, R-I weights its results based on four characteristics--education, sex, age, and ethnicity. The poll doesn't weight by geographic location.

Yet this isn't the result of random sampling, either, even though there is a moderate correlation (.60) between R-I sample size and a state's total number of voters. There's no way California, with a population 10 times that of Oregon, only gets 561 participants to Oregon's 683. The chances of that happening randomly approach zero.

So that you don't feel as though you've wasted five minutes reading inconclusive blather, here's Z-Man pithily summing up the take-home message regarding the R-I poll in particular and polling on the 2016 presidential race in general:
My theory for a while is that it is not so much chicanery at work as confusion. The old models are not proving useful. Polling outfits cannot admit that as it invalidates their reason to exist. The natural response is to huddle around the coin flip range. They keep tweaking their models and sampling to get closer to that comfortable zone of a tie. All of them seem to be drifting to that happy place. If all the pollsters declare the race too fluid to call, then no one gets blamed.
One caveat: While there's clearly a lot of uncertainty as to what the final results are going to be, polling outfits are not above trying to bring about self-fulfilling prophecies. The needle they're trying to thread is to demoralize the Trump Train as much as possible and then as time runs out shift closer to a coin toss to save face.


Anonymous said...

Notice how the polls are starting to converge towards a tie - at least within the margin of error. CYA!!

Audacious Epigone said...

The ABC tracking poll is especially remarkable. A week ago it was Clinton +12, now it's Clinton +2.

Audacious Epigone said...

This morning the ABC tracking poll moved again and is now just Clinton +1.

Random Dude on the Internet said...

Should be interesting to see how it plays out here on out.

A few days ago, my Democrat friends were saying Arizona and Texas were in play. Today, they are all but ready to admit defeat. Rachel Maddow was crying on television. Donald Trump now has rallies in New Mexico and several planned in Michigan. Hillary Clinton is now spending money on advertising in Wisconsin. This was all pre-FBI.

Will polls continue to normalize or will they keep it in the margin of error as to not wipe out the enthusiasm of the already dragging base? These polling companies can always handwave away a Trump landslide as the emergence of secret racist whites aka the Bradley effect. These people have reputations but the leadership still wants to be invited to the right cocktail parties and that means helping out Democrats whenever they can.

So now we have the investigations, a supposed bombshell from Wikileaks, another Project Veritas video, people finding out that their healthcare premiums are shooting up, and a potential strike in Philadelphia. These last few days are going to be awful for Hillary. If the Clintons have a "break glass in case of emergency" bomb for Trump, they better use it now and it better not just be more women talking about how he groped them. It's going to have to be some dead girl/live boy type of revelation here.

Audacious Epigone said...

Random Dude,

Yes, Trump's scheduled appearance in NM caught my eye. In a traditional campaign, that'd be huge news. It probably still is (though he's held rallies in places like Mississippi during the general election, so who knows).

After the 2005 video came out, I was reading everywhere that this was just the first of several "tapes" the Clinton campaign had on Trump. That's looking like total garbage now, though. We're well into early voting and just over a week away from the general. Some more of these bombs--if there were any--would've come out between then and now.

Also, Hillary's Facebook feed is full of posts attacking Comey and the FBI. It looks like a pretty desperate attempt at damage control.

Dan said...

One reason that the polls may have shifted so much is that they now take into account early voting turnouts. Compared to 2012, turnout among blacks is down and turnout among 18-29 is way down.

Early voting turnouts for 2016 would be new data that they just didn't have until now. No deception on the part of pollsters, they were just using 2012 data. Now they have 2016 data. One clue on enthusiasm was primary turnout.

Another reason for the Trump bump is that the Trump tape release is drifting further into to the past. That tape was a blow for SoCons but now it is old news.