Conventional wisdom is that Wednesday was a bad day for Team Blue on the polling front; it started out with a surprising poll for Bloomberg (by Selzer, who are considered the “gold standard” in Iowa but are less proven elsewhere) showing Hillary Clinton trailing by 5 in Ohio. A national poll from Quinnipiac showing Clinton leading by 5 nationally provided some relief, but then that was quickly followed by two state polls from CNN, most notably a second poll showing a 5-point deficit in Ohio, and on top of that one showing Clinton down 3 in Florida.
So, you would expect that to wreak havoc with our predictive model, right? Well, there was a drop, but it wasn’t the crash landing you’d probably expect. Hillary Clinton’s odds of winning the presidential race are currently 73 percent (in other words, she won simulated model runs on Wednesday night at a rate of 73 out of 100 times). That’s down from 77 percent on Wednesday … but actually up from Monday and Tuesday, when she was at 72 percent.
How did that happen? The short answer is: the model was already treating her as having less than 50 percent odds in Ohio and Florida, something that previously raised a lot of questions (i.e. “why do you have Ohio and Florida light red on the map when the poll average has a small Clinton lead in those states?”). It’s a predictive model, not a “here’s what’s happening today” model, and the fundamentals, at this point, are acting as a bit of a weight on Clinton’s poll numbers. As I’ve said before, running for a third term for the party in power, against the backdrop of so-so economic growth, is always a difficult proposition, historically. So the drop was already, to an extent, “priced in.”
One way to think of a presidential election is that the poll numbers are always trailing the fundamentals, trying to catch up with them (with the state numbers, on top of that, usually trailing the national numbers, since there tend to be not as many state polls). And you could think of Wednesday as: The day that the state polls caught up with everything else (not just potentially difficult fundamentals, but also a national race where the average has slowly drifted down to about 4 points … which would point to a nearly tied race in Ohio and Florida, which usually lag the national percentage by a few points).
Another important thing to remember is that the main number that the model is interested in is 270. If Clinton gets to that, it’s a win, no matter what combination of states is involved. So from that standpoint, maybe the most important poll of the day on Wednesday wasn’t one from the major media organizations, which get headlined around the world immediately … but rather one that barely got any attention at all. It was a poll of Pennsylvania from GBA Strategies on behalf of End Citizens United, and it showed Clinton leading by 6, as well as Democratic candidate Katie McGinty leading the Senate race by 5 points.
That’s important because it’s one more brick in the “blue wall.” It kept Pennsylvania’s polling average at around a 6-point lead for Clinton, and reinforced her chances of winning the Keystone State over 80 percent. So long as that holds, along with similar-sized leads in Colorado, New Hampshire, and Virginia (where we got a poll from PPP Tuesday with a 6-point lead in a 4-way contest, which helped keep her Virginia odds up around 90 percent, and which probably contributed to her bump in the overall odds on Tuesday night), that’s a win ... regardless of what happens in Florida, Ohio, or Iowa. That explains why her overall odds have stabilized, for now, in the mid-70s.
McGinty’s lead in that Pennsylvania poll also helped keep the Democratic odds of controlling the Senate pretty stable; they fell from 46 to 45 percent. McGinty’s chances in Pennsylvania are still better above water, which compensates for the further decreased odds in the Ohio and Florida Senate races after Ted Strickland and Patrick Murphy’s poor showing in those CNN polls.
Finally, there’s one reason to be a little skeptical of the Ohio and Florida polls … but it’s probably not what you’re thinking I’m going to say. There was a lot of crosstab-citing of the Bloomberg poll of Ohio (at least until the afternoon, when the CNN poll came along and seemed to confirm the results); the main warning sign for people was that it seemed to point to a 2004-style electorate, one where Republicans significantly outnumbered Democrats.
That’s not, by itself, necessarily a problem, though. Pollsters (at least good public pollsters) don’t use party ID to weight a poll. (If you think back to the Dean Chambers debacle of 2012, that was the very criterion he was using to “unskew” the polls, despite pollsters’ protestations that they weren’t weighting to party ID in the first place.) Pollsters prefer to weight by immutable categories that aren’t subject to change, like race or age. (Ideally, they’d prefer not to weight at all, but people don’t answer the phone at equal rates across all ages, genders, and races.) Party ID, by contrast, isn’t fixed; it’s an identity that people can pick up and put down as they change their minds from election to election, or even within a cycle.
A person who has previously declared himself a Democrat or independent (and is even registered that way) might identify as a Republican on a poll, if he’s planning to vote for the Republican that year. So, it’s possible that Selzer simply ended up with a lot more Republicans than expected, either by making incorrect assumptions about the ages and races of the likely electorate (which would possibly overweight categories who are likelier to vote Republican), or even that they got the assumptions right, but just stumbled into an unusually Republican patch (in terms of not just Trump preference, but also self-described party ID) of a normally modeled electorate.
What could cause that “unusually Republican patch” to suddenly flare up? Two possibilities are that Republican leaners were, compared to everyone else, more fired up, and therefore more willing to stop what they’re doing and tell a pollster about it … or that Democratic leaners were less fired up. What tends to do that is other events in the news. We can surmise that during and after the Democratic convention (especially when compounded with the Khan family disaster), Republicans weren’t really feeling it. And … you can probably guess where I’m going with this … over this weekend, Democrats were possibly too busy being worried or demoralized about the story about Clinton’s health to be enthused about talking to a pollster about it, when these polls were in the field.
One way to describe that problem is “non-response bias;” in other words, the responses of those who choose to respond would be different than those you choose not to respond. It’s a phenomenon that we’ve been aware of for a long time … it may have been the primary culprit in the notoriously disastrous Literary Digest poll that predicted a landslide victory for Alf Landon in the 1936 presidential race … but one that pollsters are just now starting to grapple with.
A more recent case was the polling spike that Mitt Romney received after a poor performance by Barack Obama in the first debate in 2012. Research after the fact, however, suggested that Romney didn’t suddenly get an influx of new backers, as much as Obama’s backers were demoralized and temporarily less willing to talk to pollsters, and Romney was temporarily winning by subtraction, which explained why that debate bump quickly wore off. Pollsters using more advanced techniques … especially Obama’s internal pollsters, who were relying on multiple levels of voter file information to sort voters, instead of just using random-digit dialing and talking to whoever answered … found that there really wasn’t much of a debate effect at all, and the race stayed in pretty much the same narrow band from April on.
And pollsters who are willing to dig a little below the surface (and not interested in feeding a horse race narrative in the media) are finding similar things this year. For instance, check out the graphs in Wednesday’s blog post from Mark Blumenthal, who used to run HuffPo Pollster but now is research head at Survey Monkey, which performs NBC’s tracking poll. There’s remarkable stability in not just the topline numbers, but also the party ID numbers, which they ask each week as well. You can see narrow closings and widenings during periods like the conventions, when one side’s voters are energized and the other side’s voters are not. But for the most part, the lines since May have been almost flat.
Unfortunately, though, we’re going to see lots more zigzagging from day to day, from the media pollsters who use random-digit dialing instead of paying for the more expensive voter file. If the natural state of the race really is around a 3- or 4-point national lead (instead of the gaudy 8-ish we were seeing after the Dem convention), then, as part of a normal distribution, you’re occasionally going to see polls with Team Blue losing, especially in the states that are a click to the right of the national average (like Ohio and Florida); there’s no way around it. I can certainly tell you to stay calm by looking at polling aggregates and predictive models to smooth out the noise, but there’s still going to be a lot of (emotional as well as quantitative) ups and downs as individual polls roll in each day.