You Can't Talk About IVR Polling Without Talking About Rasmussen

It's often a major error to analyze some shift in American politics, especially in the demographics thereof, without asking first whether it's really a nationwide phenomenon you're observing or just a really strong Southern phenomenon. For instance, I recently saw some numbers showing that essentially all of Obama's national deficit with white working-class voters comes from the South. Surprise! Southern whites, including the not-terribly-rich ones, don't like Barack Obama. Who'd've thought?

I think this study suggesting that IVR polls, commonly known as "robo-polls" where there's no live human interviewer, just prerecorded questions that one answers by pressing buttons, are "copying" live-interviewer polls to get better accuracy struggles from somewhat of a similar situation. Their evidence is that IVR polls have really lousy accuracy when they're conducted in a race that no standard-methodology pollster has yet touched, but once a race has seen a standard pollster check it out, IVR polling is just as accurate as live-interviewer thereafter. The people at PublicPolicyPolling, an intensely prolific and typically really accurate IVR polling firm, are naturally not thrilled with this. But I think the real problem with the survey is that, just as one errs by treating all of America, South and otherwise, as being one integral nation, one errs by treating all IVR pollsters alike.

Specifically, of course, I'm talking about Rasmussen Reports. Because here's the thing: we liberal polling junkies have been noticing for several years now that there's a distinctive pattern to Rasmussen's polling. Very early in a cycle, their polls look ridiculous. Specifically, they're ridiculously skewed toward the Republican candidates. Later in the cycle, if they continue to poll at all, their polls become much more sensible, and end up being quite accurate. What it looks like, given that it's pretty well-known that Rasmussen is a deeply partisan firm with a pretty substantial history of hackishness, is that early in cycles Rasmussen is engaged in polling-as-messaging, trying to create a buzz around Republican candidates that, perhaps, no one's ever given much heed to. That happened with Scott Brown, for instance. Then, since polling firms are basically only judged on the last batch of polls they release, later in the cycle Rasmussen starts using more sensible turnout models, etc., perhaps "copying" from the consensus of the polling to get good results and maintain their reputation, such as it is.

The result is that the largest chunk of robo-polling is conducted by a firm that is deliberately being inaccurate early in the cycle, when few of the traditional pollsters are yet at work, and deliberately making itself more accurate at crunch time, when most contests have already received a whole bunch o' polling. That would easily be enough to produce the kind of data that study found, I'm guessing, so I'd like to see the study run again excluding Rasmussen. I'd be pretty surprised if the result didn't change.