So Gallup’s latest has Romney up six, and I’m scratching my head. I know Gallup just can’t throw out the numbers because they “look weird,” but I kind of doubt that even Mitt Romney believes he’s up six nationally.
But I understand this happens, and happens a lot. So as someone unschooled in statistics, is there any useful information we can get from such results (in general, not necessarily this particular one)? Is it ONLY by comparison that we can figure out which results are likely and which aren’t (I assume so, but as I said, unschooled)?
The Gallup LV screen is generally seen as a pretty tight one, and has produced some odd results in the past. It tends to run a bit older and have a smaller minority representation than some (read: Democrats) would like. So that’s certainly some of it.
It also hasn’t been the most accurate pollster recently, although it’s certainly not actively bad.
Your best bet is to look at an average of pollsters and the trend, and then weight it against state polls. The big question right now is whether the state polls (which have held up a bit better for Obama) or national ones (which have moved pretty strong to Romney, but may have flattened) are more accurate right now.
Strictly speaking, the only way to determine which results are likely is with reference to whatever set of assumptions you’re willing to make.
If you believe that Gallup drew a uniform, random sample; that its adjustments for voter likeliness are sound; and that its questions correctly elicited each respondent’s choice, then the results should provide an unbiased estimate of what would occur should the election be held today. The extent to which these conditions hold is conjecture, and we have to weigh the results against the very reasonable prior that Romney is unlikely to be six points up, so, yeah.
Taking it as a given that the results are wrong to some undetermined degree, do they provide useful information anyway? I suspect so — Gallup surely has some idea of how to run a poll and adjust the results. Examining a range of polls employing different assumptions is probably a good idea (you can be like Nate Silver and use them to build complicated ad hoc election models!), but comparison requires that you make some determination about the distribution of poll results, and we’re back to untestable assumptions.
(There’s only one surefire way to predict the results of an election: rig it.)
Today’s 538.com summary has some very interesting theorizing about this. It’s focused on a set of polls by YouGov that re-contact the same voters to see if anybody has changed their mind. It showed very small movement towards Romney.
The hypothesis, such as it is, is that non-recontact polls (which is almost all of them) over-weight enthusiasm due to their extremely low response rates (10% is really good, 3-5% is typical for robocalls). So bounces get exaggerated because not only did the “bouncer” pick up new voters but his supporters are also much more likely to pick up the phone (and pass the likely voter screen). Recontact polls have their own disadvantage (described in the article), but they do provide an interesting thought.
Certainly gives more reason to cast a skeptical eye toward bounces in general and look at longer-term averages and trendlines. We’ll see if the next week or so moves things back to the long-term equilibrium (roughly O+2).
Considering that it’s at the extreme upper range of the available polling data, I think that’s probably unlikely (just like I think it’s unlikely that Obama has a 3-point lead right now - which appears to be the other extreme from the last couple of weeks).
RealClearPolitics has a piece which explains that Gallup tends to be “bouncy.” I suspect this result is mostly a demonstration of that. But sincerely, bouncing between tied and R+6 is bad news for Obama.
And what about all the non-Gallup polls? What’s the explanation for them?
The simple fact is, outliers happen. If a poll says some number is (say) 53% with a 2.5% margin of error, that doesn’t mean that the true result absolutely must be somewhere in the range of 50.5% to 55.5%. If they’re using standard deviation to define their margin of error, then what it means (approximately) is that 60-something percent of the time, the true number will be within that range (and 30-something percent of the time, it’ll be outside of that range). Nor is there any guarantee further out: A couple of percent of the time, you’ll be twice the margin of error out from the truth. And we’ve had hundreds of polls; it’s only to be expected that, even if the pollsters do everything perfectly, there will be some rather silly outliers.
I wonder if the econometrics professor thinks everyone at Gallup just decided to make up a poll favoring Romney all of a sudden. Seriously, wtf does he think happened? That for decades Gallup has been a reliable polling organization, and then one week in October they suddenly decide to throw it all away on a whim?
Chances are, whichever poll showing the best result for Romney or the best result for Obama, is wrong. Which is why an average of polls is better. I think another pollster has Obama up by 4. Most polls show Romney up by 1 or even with the President.