Give a lot of credit to the University of Southern California: It’s not often you see a major public polling outfit release a lengthy analysis of why its opinion polls might be wrong, but they’ve just done exactly that.

For background, USC has been doing a daily tracking poll of the American Presidential Election. So far, it’s been pretty much in line with the rest of the public polling:

The most recent set of numbers there show Biden with slightly more than a ten point national lead: 53% to 43%, with 4% undecided.

But that’s not the only question they’re asking, as they explained last night in a lengthy blog post titled “are these experimental questions pointing to a Trump victory?”:

In this year’s Daybreak Poll, researchers are asking participants two additional questions that are intended to, as they say, “harvest the wisdom of crowds” to predict the election outcome. The “social-circle question” asks respondents to report the percentage of their social contacts they expect to vote for each of the candidates. The other one, known as the “state winner question,” asks participants who they think will win the election in their state…..

…. in all five of the elections in which we tested this question, the social circle question predicted election outcomes better than traditional questions about voters’ own intentions. These five elections were the 2016 U.S. Presidential election, the 2017 French Presidential election, the 2017 Dutch Parliamentary election, the 2018 Swedish Parliamentary election, and the 2018 U.S. election for House of Representatives….

…. We believe there are three main reasons. By asking people how their social contacts will vote, we’re implicitly gaining a larger and more diverse sample of participants.

Second, it can be less embarrassing or intimidating for someone to tell a pollster that their friends plan to vote for an unpopular candidate than to report their own intention to vote for that candidate.

Third, we are all influenced by our social contacts. Even if we report an intention to vote for a candidate other than the one most of our friends support, there is a chance they will eventually persuade us to vote for their candidate.

That’s a fairly straightforward explanation, and as they note, it’s not really a new technique. Often times, on sensitive issues, pollsters find that they may get a different answer if they ask people “who do you think your neighbour is voting for” than if they ask them outright who they’re voting for themselves. The basic idea is that you might not want to admit that you yourself are voting for Trump, or Brexit, or whatever, but that you’re much happier to say that your neighbours are.

Anyway, enough with the explanations. What do the numbers actually say? Well, Biden’s still ahead, but it’s much, much, much closer:

Trump can’t win if he’s losing the popular vote by ten per cent. But if he’s losing by two or three per cent, like in this poll, then he certainly can win.

Of course, even then, it’s no sure thing. A swing of just 50,000 votes across three states would have handed the White House to Mrs. Clinton in 2016. Trump’s electoral college majority, even if he performed relatively identically to four years ago, could be blown away by a small gust of wind.

And on the pessimistic side, there is lots of evidence for a pretty strong gust of wind. David Wasserman is about as close as you can get to knowing everything there is to know about American politics, and he’s been reading polls that the public generally doesn’t get to see: in individual congressional districts:

Wasserman notes that this year, those district polls are showing something very different:

We’ll know the answers in eight days, give or take. But the very best scenario for Trump is that it’s going to be a close run thing, and the worst scenario is a landslide defeat. Of course, all of the polling could be wrong, but – forgive me for being a pessimist here – it’s telling that even the experimental data can only get him to within two points of Biden.