Undecideds, "Undecideds," and Shy Voters

Michigan is the fourth state in a row Trump has underperformed public polling.

Trump and the GOP are weak, not strong. They've been having performance problems since Dobbs, and we are seeing these problems, their struggle, in early 2024.

More 👇https://t.co/6fa2YKK1V1

— Simon Rosenberg (@SimonWDC) February 28, 2024

Nate Cohn addressed this question last week. Longtime readers we be shocked to learn I didn't find his arguments all that convincing.

One simple explanation is that undecided voters ultimately backed Ms. Haley, the former South Carolina governor.

This is plausible. Mr. Trump is a well-known candidate — even a de facto incumbent. If you’re a Republican who at this point doesn’t know if you support Mr. Trump, you’re probably just not especially inclined toward the former president. It’s easy to see how you might end up supporting his challenger.

Yes, people who are not inclined to support Trump are less likely to support Trump, but in effectively a two person race, this still begs the question of why, if respondents are being honest and these people really are going into the polls not knowing who they'll vote for, they almost all make the same decision. Even if it's off-balance, how often can that coin keep coming up heads?

If I were picking plausible options instead of just criticizing the NYT, one of them might involve an ironic twist on the Bradley effect, where some saner Republicans are embarrassed to admit their doubts about Trump over the phone but are uncomfortable enough to pull the other lever in the privacy of the booth when things actually matter. I'm not sure how well the data fits this hypothesis, but when's the last time political analysts let that stop them?

 To be fair, Cohn's explanation worked reasonably well in Iowa, where a voter could be not Trump and still legitimately undecided between the other candidates, but in the case of a two candidate race, polling should give us a clean read.

Cohn continues [Emphasis added.]

It’s also a theory with some support in the polling patterns. Other than Mr. DeSantis dropping out of the race, which led that voting group to shift toward Mr. Trump, Mr. Trump’s support in the early states was flat over the month or so before these elections. Over the same period, Ms. Haley tended to make gains — gains most easily attributed to undecided voters coalescing behind her.

"[M]ost easily attributed" needs some supporting arguments -- I can think of plenty of other potential causes -- but putting that aside, we have here another case of Cohn ignoring data that doesn't fit the narrative. As far as I can tell there's nothing about this theory that limits it to early states, but when we pull back...

While we can't know who went where, it was Haley who had the immediate bump in the national polls when DeSantis dropped out while it was Trump who climbed steadily. More importantly, Trump's gains over that period were larger than DeSantis's total support was when he dropped out. On a national level, this suggests it is Trump rather than Haley who is pulling in undecideds.

This doesn't mean that the theory was wrong, but the "some" in "some support in the polling patterns" is doing a lot of heavy lifting.

Which takes us to theory 2. I'm going to tread carefully here. On one hand, this part looks questionable, but it is out of my field of expertise so I'm just going quote Cohn then follow up with a couple of data points from CNN and the Washington Post which I believe were available when Cohn wrote his piece.

Another possibility is that the polls simply got the makeup of the electorate wrong. In this theory, pollsters did a good job of measuring the people they intended to measure, but they were measuring the wrong electorate. In particular, they did not include enough of the Democratic-leaning voters who turned out to support Ms. Haley.

It’s impossible to prove, but I think this is probably a major factor. It’s always relatively hard to predict the makeup of the electorate in a presidential primary, but the large number of Democratic-leaning voters motivated to defeat Mr. Trump is a particularly great challenge this cycle. For the first time since 2012, there’s no competitive Democratic presidential primary to draw Democratic-leaning independents away, and the Republican runner-up is a relative moderate who may be palatable to many Democratic-leaning voters.

We don’t yet have turnout data on how many Democratic-leaning voters actually participated in these primaries, but there’s good reason to believe this is part of what’s going on.

...

How much of a problem for pollsters is this? It could be a big one. The pre-election turnout estimates we used for our election night live model — you may know it simply as the Needle — supposed that 8 percent of the Republican primary electorate would be composed of former Democratic primary voters who hadn’t previously voted in a Republican primary, those who wouldn’t be eligible for the Monmouth/Washington Post poll. That group seems likely to have backed Ms. Haley.

That might seem like a lot of Democrats, but the final results suggest it might have actually been too low. In fact, these same pre-election turnout estimates unequivocally underestimated the turnout in Democratic-leaning areas relative to Republican-leaning areas, suggesting that the turnout from Democratic-leaning voters was even more vigorous than projected.

 Finally we get to the self-selection bias theory.

In this theory, the polls did well in modeling the electorate while undecided voters split between the candidates, but anti-Trump voters simply weren’t as likely to take surveys as pro-Trump voters. If this theory were true, then the general election polls might be underestimating Mr. Biden by just as much as they’ve underestimated Ms. Haley. [A bit of a misrepresentation of this position. Having someone who won't vote for Trump in the general is not the same as having someone who will vote for Biden. -- MP]

...

The absence of evidence for nonresponse bias doesn’t disprove it. Far from it. But in this case, the turnout and undecided voter theories are credible enough that there isn’t reason to assume any nonresponse bias either.

I don't have a dog in this fight. At this point, I'm not going to even hazard a guess as to what theory explains the disconnect (including the bizarro Bradley effect). That said, it's worth keeping in mind that Cohn's argument for dismissing the non-response theory depend on us finding his arguments for the previous two theories so convincing as to leave no room for another explanation. As far as I can tell, we have three theories without compelling evidence or even strong arguments to back them up.