FiveThirtyEight generally takes an inclusive attitude towards polls. Our forecast models include polls from pollsters who use traditional methods, i.e., live interviewers. And we include surveys conducted with less tested techniques, such as interactive voice response (or “robopolls”) and online panels. We don’t treat all polls equally — our models account for the methodological quality and past accuracy of each pollster — but we’ll take all the data we can get.

This split, however, between live-interview polls and everything else, is something we keep our eye on. When we launched our general election forecasts in late June, there wasn’t a big difference in the results we were getting from polls using traditional methodologies and polls using newer techniques. Now, it’s pretty clear that Hillary Clinton’s lead over Donald Trump is wider in live-telephone surveys than it is in nonlive surveys.

We don’t know exactly why live-interview polls are getting different results than other types of surveys; there are a lot of potential causes and it’s something we’ll be digging into. But it’s harder to measure this gap in the first place than you might think; pollsters make a lot of choices — whether to use a registered or likely voter sample, for example, or whether to poll Gary Johnson and Jill Stein — that could account for the differences between live and nonlive surveys. But one method to get at the difference is to use our forecast models, which account for these things.

As of Tuesday morning, Clinton led Trump by 6 percentage points and had a 79 percent chance of winning, according to our polls-only forecast. But running our polls-only model using only live-interview surveys, Clinton leads Trump by 7 points and has an 86 percent chance of winning. Running it with only nonlive-interview polls, Clinton leads Trump by 5 points and has a 71 percent chance of winning.