Polling

This is a response to Josh Marshall’s latest blog on his indispensable news site Talking Points Memo about the state of the race. I’ve been reading TPM since 2004, almost took a Fellowship gig there about a decade ago, but regrettably, only became a member recently. Don’t make my mistake, you should join now.

Anyway, as usual, Marshall writes an astute, pragmatic take on the State of the Race given his sense of how the vibes around the Harris campaign have translated into polling. It’s hard to argue with the overall conclusion, that Harris is doing much better than Biden but could very possibly lose. That’s true, and we all need to work extremely hard to make sure she wins.

I want to offer a subtly, but meaningfully, different, data-driven take on what we're seeing in the polls that comes to a similar conclusion as Marshall on where the race stands but from a bit more of an optimistic place. The basic take is this: Biden was never really "behind," per se, but his weaker supporters (those less politically engaged folks you mention) were skeptical enough about him to express their doubt by polling as undecided or third party, a trend which accelerated in the weeks after the debate. We can see evidence of this in RFK's support reaching more or less its high point in the polling averages the week following the debate, while Biden's support fell off and Trump's support remained pretty steady:

What Kamala did, astonishingly fast and well, was successfully consolidate those weaker Biden supporters into stronger Harris supporters, which we see evidence of in RFK's fall from 10% after the debate to below 5% in the averages the day he dropped out (and, of course, the shocking rise in her approval polls):

To Marshall’s point about convention bounces, that's what conventions are normally supposed to do: bring any stragglers in from the cold by pumping up the candidate and celebrating party unity so that the campaign can turn its focus to undecideds in the fall. Kamala had already accomplished that by the time the convention started. I suspect that's why none of these "bounces" have yet been ephemeral, and also why Marshall has been writing that he feels, as I do, that something different is going on in this campaign that the conventional press hasn't grasped yet: she hasn't been getting bounces at all, she's been expertly consolidating latent support that's been there all along for the right candidate, at a much faster pace than usual given her talent and the extraordinary nature of this cycle.

That's also why the thematic focus of the convention was cross-ideological stuff like patriotism and American values, not Democratic unity. If Biden had stayed in the race, the convention would have been about consolidating those straying supporters. It's likely many of them would have ultimately returned to Biden, maybe even enough to squeak out a narrow win via the blue wall states. But he would have had to simultaneously fight to keep the coalition together and fight for undecided voters all through November, the political equivalent of pursuing a war on two fronts. By consolidating that weaker support so quickly and decisively, Kamala now gets to fight on only one front, to convince undecideds. That's an enormous advantage, especially against an opponent who has their own consolidation problems.

This analysis all comes from an understanding of polling based on a very technical, scientific view of their purpose that's been afforded to me by my strange career path in academic publishing tech, and heavily borrowed from a scientist name Carl Allen, who makes a unique, and convincing, argument about how polls ought to be applied to understanding elections. Statistically, polls aren't meant to predict the future in any way. They're meant to predict the current state of an entire population by querying a small, representative sample of it. An election poll's margin of error isn't the margin of error between what it captured and how the electorate will vote on election day, nor is it the margin between what it captured and how the electorate would vote today. It's the potential for error between how the sample population the poll reached responded and how the full population the poll is trying to predict would have responded if the poll had somehow been able to talk to everyone in the entire population. Properly conducted polls are, basically by definition, very good at this particular prediction. Averaging polls can make these predictions stronger and reduce margin of error, but that still doesn't predict the future in any scientific or technical way.

This may seem like a subtle point, but I think it's significant for understanding where we are, because it helps to think of polls not as predictors of future results, but as snapshots in time you can play forwards against various assumptions and evidence. So, when Marshall writes it's quite plausible Harris could lose, even without any polling error, he’s right, and for a specific reason. As of today, Harris is at 47.3% in the averages and Trump is at 43.7%. That means about 9% of potential voters are currently undecided or planning to vote for third parties, which is a bigger source of uncertainty than a typical poll's +/- 3-4% (so, 6-8%) margin of error. If all of these currently non-Harris/Trump voters broke for Trump, Harris would lose ~53-47, way outside the margin of error, and the polls would have been accurate. If all of them broke for Harris, she would win ~56-44, way outside the margin of error, and the polls would still have been accurate. Of course, what will actually happen is some people will stay home, some will vote third party, and the rest will break decisively, but not entirely, in one direction.

And that's where I think it's justified to be a little more optimistic than Marshall — though again, with this understanding, anyone can play the prediction game with their own assumptions and evidence. For one thing, I suspect the gap between Harris's numbers and the relatively better Democratic Senate numbers Marshall mentions is a piece of evidence that may predict how undecideds will break. As he wrote, Senate elections are different because they're more local, and so there's less national polarization at play. For the same reason, however, it also makes less sense to vote third party or to stay home for a race that low-engagement voters perceive as impacting them more personally than the Presidential. Obviously, some will split their ballots, but I also think it's fair to assume that as those voters start to tune in in the fall, their Senate choices will be a decent predictor of how they'll break.

We saw this play out in reverse 2016 and 2020, where GOP Senate candidates in swing states tended to poll better than Trump, and that loosely correlated with the way undecideds broke. Speaking of 2016 and 2020, I think their polling stories also offer a case for optimism, for the simple reasons that there's far less uncertainty today than there was in 2016, and we're approaching levels of support for Harris that, like in 2020, can't be overcome by undecideds breaking decisively. At her polling peak on October 28, 2016, the day of the Comey Letter, Clinton polled at 46% in the 538 average to Trump's 39.6%, meaning a whopping 17% of voters were third party or undecided. It's a lot easier to climb back into contention when you have that big of a pool to draw on. 

Where there's a real case for worry is a universe where undecideds again break decisively for Trump, which is perfectly plausible. In 2020, Biden hovered around 49-50% in the averages basically from April 2020 until the very end, when he got a late boost, with undecideds and third parties hovering around 5-7%. I wish Kamala was already there, too. But this also isn't 2016. It's harder to make up ground when only 9% of the electorate is undecided or third party, which is historically pretty normal. Even if we assume no third-party voting and a 6-3 break for Trump, it results in a narrow Trump popular vote loss given today's averages. But since RFK dropped out, and especially as pollsters apply likely voter screens, we're seeing more 50%+ polling for Harris, indicating that of that remaining 5% of Kennedy supporters, about 2% were leaning towards her. Accordingly, I bet we'll see Harris take more like 49-44 leads in the averages in the coming weeks, meaning that even to get close enough to win narrowly in the Electoral College (a gap which appears significantly smaller this cycle), Trump would have to win nearly all third party voters and undecideds.

All to say, Marshall is right that this will be a slog. And yes, we all need to buckle down and work hard. But I think there's a real case for being a bit more optimistic about where we stand, and operating from a few assumptions about how and why Harris is likely to win this. The great news is that Harris has basically no consolidation to do. The bad news is that Harris still has to convince enough undecided voters to break for her (or stay home) to eliminate the risk of tipping the thing to Trump. So that's her, and our, task: to get out there and do that convincing. Luckily, that task is much, much easier than Clinton's was in 2016, and is approaching Biden's in 2020.