Monday, September 29, 2008

Polls are useless

We are getting poll results every day on the presidential election. Even at other times, polls are used on a regular basis to elicit which way the public leans on policy matters. Is it a good idea to do so?

John Morgan and Phillip Stocken would probably say polls are of very limited use. There are several issues at hand. First, there is the sample size, with many polls aggregating the opinions of 1000 people or less, the statistical significance of the result may be very low, especially for close outcomes. With small samples, polls work reliably only when the population is relatively homogeneous. But then you do not need polls.

Second, there is strategic behavior, especially in larger polls, where responses tend to follow ideology instead of truth-telling. Inferring outcomes from polls without taking into account strategic behavior can be very misleading. Luckily the authors provide estimators that correct for such biases. But this raises the question: If the polity knows such estimators are used, would it not adjust its strategy accordingly?

Third, poll results differ from referendum or election results. The problem is that in a limited sample poll, it is impossible for voters to convey credible information in equilibrium. We need referendums, as I called for the other day.

The critical point here is that the polled ones have information that the pollster does not have, and that it is costless to convey information in a poll. As an example, the article takes the case of Oregon, where it was contemplated whether to lower the minimum wage for tipped workers. To gain information about the economic effects of such a policy change, restaurant owners were polled. It is clear that it was in their best interest to act strategically instead of revealing what they really know about the industry. They faced no negative consequences if lying, and they knew they could influence policy their way by lying.

1 comment:

VIlfredo said...

That polls have little predictive power is not new. The various election markets that let participants bet some real money on outcomes have shown much higher accuracy.