Tuesday, September 2, 2014

Political polls….can you trust ‘em?


The short answer is: No.

Here's the scoop: at the time the survey is done, there’s no way to calculate how accurate it is.

Doubtless you’ve heard about some polls that went awry in spectacular ways. For instance, in late 2012, the internal polls for the Romney campaign were predicting a clear Republican win. You know how that turned out.

Of course, some polls done by and for political candidates have been right on the money.

But here’s one sticky point: inevitably, with so many surveys being done, a few will be really cockeyed, and a few will be spot on, and most of them will be sort of close to the final outcome, more or less.

But predictive accuracy is what everyone wants.

And that’s what modern pollsters can’t provide. Basically, they’re providing pretty good guesses.

A recent piece on WashingtonPost.com tried to show that current polling generally produces “good quality survey estimates.” It says that the divergence of poll results and actual election outcomes can be measured in single digit percentages. It also explains that surveyors routinely “weight” their findings—a polite way of saying “cook the numbers”—because they can’t reach a satisfactory random sample of respondents.


That’s another sticky point: even a highly respected polling organization like the Pew Research Center reports that, in recent polls, it actually interviewed only 9% of its targeted sample of adults across America. That is, Pew failed to complete an interview with 91% of the people it tried to reach. You know, people just don’t answer their phones any more….

The WashingtonPost.com piece fails to acknowledge another sticky point. With abysmally low completion rates, today’s pollsters are refusing to face up to the statistical 800-lb gorilla in the room:



The mathematical underpinnings of statistical reliability in a survey are contingent on having a "true random sample" of the population being surveyed. That is, every member of the entire population must have an equal chance to be selected for the survey.



Manifestly, this does not occur in every political poll done today, viz.  response ("completion") rates as low as 9%. (Here’s a frame of reference: when I started doing public polling in the late 1970s, with door-to-door interviewers selecting households and respondents at random, we had completion rates in the 75%-80% range).

If only 9% (or 23%, or whatever low percentage) of the targeted respondents actually complete the interview, there is no way to meaningfully calculate "accuracy" (i.e. statistical reliability, error range) for a survey. No way.

Every survey published today is, de facto, not reliable, even if the published result happens to be close to the real outcome. Of course, experienced pollsters can weight the data to try to estimate the "real" result, but in terms of reliably predicting outcomes (e.g. election results) in advance, especially in close races, modern surveys are close to useless.





No comments:

Post a Comment