Thursday, November 01, 2012

A Better Polling Question: Who Do You Think Will Win?

NYT:
But another kind of polling question, which received far less attention, produced a clearer result: Regardless of whom they supported, which candidate did people expect to win? Americans consistently, and correctly, said that they thought Mr. Bush would.
A version of that question has produced similarly telling results throughout much of modern polling history, according to a new academic study. Over the last 60 years, poll questions that asked people which candidate they expected to win have been a better guide to the outcome of the presidential race than questions asking people whom they planned to vote for, the study found.
Holy crap that's interesting.

3 Comments:

Blogger Pete Mack said...

Broken link, so I dont have context.
But this is hardly surprising, given previous wisdom of crowds analyses, and the general accuracy of political futures gambling odds.

-mac

2:47 AM  
Blogger Winston Smith said...

Fixed. Sorry, I'm link-challenged.

Yeah, I agree. This coheres with other stuff we now have good evidence for.

8:12 AM  
Blogger Pete Mack said...

I thought about this further and realized it's an even stupider argument than I thought.
Let us postulate that the average poll respondent is capable of understanding polls to within their cumulative standard of error. (That is, assume that the average respondents has an intuitive grasp of the mean value theorem.)
This means that a poll on prediction will be based on interpretation of information already available, including other primary resources like previous electoral polls.

The null hypothesis says that predictive results using existing data will be more powerful (by SQRT(n)) than any one poll by itself.

But if there are no basic popularity polls, people won't be able to make accurate 'gut' decisions in the first place!

A more meaningful question is whether the wisdom of crowds is more accurate than
•a really simple statistical model
• a sophisticated statistical model like Nate Silver's 538.

This is not to say that Silver's predictions are all that accurate in detail: he is clearly missing a trick on the Indiana Senate race. (Hint: Mourdock is going down.)
The issue is whether generic crowd wisdom is better than
• a fairly simple statistical model such as Silver's
• an well-informed crowd's wisdom such as shows in the odds of political gambling markets.

Bottom line: it is really easy to make a prediction more accurate than the results of a single poll.

I do recommend Silver's book (Signal and Noise) as a discussion of this topic.

-mac

4:32 AM  

Post a Comment

Subscribe to Post Comments [Atom]

<< Home