Every pollster knows what happened in 2016 and 2020 (especially the Senate races) and is scared witless it could happen again in 2022, causing the public to reject all future polls, thus destroying the industry. Hence there is a spate of stories in the media now with interviews of pollsters talking about the problem.
The causes are sort of understood now. First, some Trump supporters told the pollsters they were undecided when they were actually dead set on voting for Trump. Second, some Trump supporters hung up when the caller was identified as a pollster. The first problem can be dealt with. The second is much harder.
If all the Trump voters were in the first category (i.e., the Trumpsters answered the phone, but just lied about being undecided), then the sample would be valid. Then in any race where the Democrat was above 50% (or 49% in most cases), one could validly infer that the Democrat will win because even if all the undecideds went Republican, that wouldn't be enough. If many of the Trump supporters refused to be polled, the sample would not be valid. That's where the problem lies.
All pollsters are trying to weight the Republicans in their sample to be equal to the number of Republicans in the electorate, but they won't know how many Republicans plan to vote if they refuse to be polled. Marquette Law School pollster Charles Franklin said what all pollsters are thinking: "The troubling part is how much of that is unique to when Donald Trump is on the ballot, versus midterms when he is not on the ballot." If the "won't-talk-to-a-pollster" holds only when Trump is actually on the ballot, then the polls will be in good shape. But if his supporters now will never talk to pollsters, even when he is not on the ballot, they have a huge problem and know it.
As The New York Times has noted some of the states where Democrats are apparently doing very well now in the polls are precisely the states where the polling misses were biggest in the past. This raises the specter of the Democrats' apparent successes simply being due to Republicans refusing to talk to the pollsters and thus be undercounted.
Doug Schwartz, director of the Quinnipiac University poll, said he has changed the procedure to separate the people who say they are undecided from those who simply refuse to answer the horse race question. The former might actually be undecided but the latter are very likely to be hard-core Republicans. But this procedure still doesn't account for Republicans who hang up on them or won't answer the phone in the first place. Since only 5-10% of the calls they make result in contact, they can't assume all these people are Republicans. Many are women putting their kids to bed, people are in meetings, folks are afraid of strangers, and more of the like. Very little can be concluded about their demographics since they refuse to talk at all.
The director of the Marist poll, Lee Miringoff, said they have expanded how they contact people, using calls to landlines and cell phones, text messages, and online interviews. That may help some, but still doesn't deal with the problem of Trumpy voters who simply do not want to be polled, no matter how they are contacted. Nevertheless, Miringoff is optimistic that with Trump not on the ballot, the problem won't be as big—this year, at least.
The director of the Sienna College poll, Don Levy, said he is being "as careful as can be" both in the sampling (who gets called) and the weighting (how much each response counts for to fix underrepresentation). He said that just calling more Republicans doesn't help because it is a specific kind of Republican who is missing. Asking 1,000 Lisa Murkowski-type Republicans doesn't solve the problem. Levy: "It's not partisan nonresponse. It's hardened Trump-backer nonresponse."
Patrick Murray, director of polling at Monmouth University, is trying a different tack. Instead of asking "horse-race questions," he is asking how much the respondent supports each candidate, rather than pitting the candidates against each other. This deals with the problem of people claiming they are undecided, but still doesn't deal with die-hard Republicans who hang up the moment they are told they are talking to a pollster—or who don't pick up in the first place.
One thing Franklin of Marquette Law School is looking at is using statistics to deal with the problem. Specifically, he is looking at nonresponse rates by county to see if they correlate with the Trump vote in 2020. If the nonresponse rate is [x]% higher in Trumpy counties than in non-Trumpy counties, he might be able to infer that those [x[% are Trump supporters who won't talk to him. If something like that works, he might be onto something. (V)