Election day is in one week, and James Comey has, at the very least, upended Clinton's smooth sail to victory. Assessing the impact, though, is still hard. At the national level, RealClearPolitics has Clinton's lead down to 2.4 points. That's narrow enough that I'm going to have to revisit the "Nate Silver is full of shit" series. Why? If Clinton's lead slips just a bit more, we are seeing a race close enough for an electoral college-popular vote split to be within the realm of plausibility, in which case we need to pay closer attention to those combinatorics games that Silver plays. His math becomes, if not quite right, then at least... less wrong. Yes, if things get any closer, I'm going to tell you to start following Nate Silver's math! It's designed for close elections.
For now, though, assessing where things are depends on what you think of a little group called "Remington Research." Who are these people? Frankly, my dear, I don't give a damn. What I do care about is that they are a Republican polling firm who conducted some Sunday-only polls of battleground states. What does it mean that they are a Republican firm? It means they want to cook the numbers to make them look good for Trump. How? There are two main methods: the "likely voter screen" and the weighting system.
The "likely voter screen" is the method of determining who is likely to vote. Here's a very simple, and very crude example. Assume that someone who voted in the last election will vote this time. Well, which election? The 2014 election was a strongly Republican midterm, and the 2012 election was a Democratically leaning presidential election. Use the one that leaned towards your party. More complicated methods could be based on assumptions about turnout from demographics and the assumption that your party will do better with turnout. Either way, you can cook the numbers if your goal is not to measure electoral preferences but to put out numbers that make things look good for your party.
Then there's the weighting system. Everyone wants a random sample. Personally, I'd love to stick everybody's social security number in a giant system, and pull them out at random, or have The Machine from Person of Interest do it. Then have Reese and Shaw chase people down and force them to answer questions. Alas, my IRB would object. So, nobody ever gets a true random sample. Instead, everybody gets too many of group X, and too few of group Y. Consequently, we give extra weight to group Y, and less to group X in our math. But, by how much? Well, if your goal is not to measure the preferences of the electorate, but to put out a poll that makes things look good for your side, you can add or subtract as much weight as you want! Just assume that your side will have a great year for turnout. How much weight is Remington adding, and to whom? This is where there is a little more transparency. Their formal reports give weighted demographics, which means readers can compare Remington's weighted sample to state demographics and reasonable estimates of the electorate, and there isn't anything obviously way off there. Here, for example, is Ohio. Of course, it doesn't take much to tip things by a couple of percentage points!
All I know is this: on Sunday, Remington Research did polls in a bunch of battleground states. Those polls showed Trump gaining a lot of ground from the Comey announcement. Is that because the announcement did major damage to Clinton, or because Remington is a hack firm run by partisans who are cooking the numbers for partisan purposes? Well, Remington is the only firm so far with a bunch of post-FBI-only polls. I guess we'll see.
You don't get to ignore data just because a firm is partisan. The reason we look at polling averages is that all polls have at least the potential to be screwed up. Hell, I ditched my land line years ago, making me really hard to reach since polling firms have to shell out extra for the right to call cell phones, and response rates plummeted long ago when telemarketing got out of control. Any one poll is probably garbage anyway, intentionally or not. We cannot pretend that these Remington polls don't exist. They do, and they don't have much competition for post-Comey data.