The exit polls are a nice example of collider bias, an important and ubiquitous but not very intuitive problem in statistics.

Collider bias means this: If a and b both are causes of c, then conditioning on c (including it as an independent variable, sampling based on it, etc.) will create a correlation between a and b, even though there is no causal link from either one to the other.

(Collider bias is one reason you should never just throw a bunch of variables into your regression as "controls", without thinking carefully about what causal links you think exist between them.)

In this case, we know that Trump voters are less likely to respond to polls. We don't know exactly why, but the fact that Trump consistently outperforms his polls means that this must be the case.

Where collider bias comes in, is that since Trump voters are less likely to respond to polls, and poll results are necessarily conditioned on responding, *anything* that affects response rates will correlate with Trump votes in poll data, whether or not it does in actual voting.

Specifically, any characteristic that makes people less likely to respond to polls with show a spurious negative correlation with Trump voting, while any characteristic that makes people more likely to respond to polls will show a spurious positive correlation.

So for example people are making a big deal about exit polls finding a higher proportion of black voters supported Trump than in 2016. But that could just as well be explained by black voters being relatively more likely to respond to polls this year than in 2016, for any reason.

To be clear, there are lots of other potential problems with exit poll data - the point of this thread is to explain what collider bias is, not to make any substantive claim about the election. That said, given very low response rates seems like this bias could be quite large.