Sunday, February 03, 2008

Divination by Polls

We are drowning in polls this weekend. Each poll is being breathlessly touted over at DailyKos by one side or the other as holy writ.

Polls as Divination
Polls are a method of divination. For thousands of years human societies have sought ways to predict the future. Polls may be more accurate than reading tea leaves or studying entrails (although I have not seen any scientific studies to prove this), but their intent is the same. Scholars use arcane methods to tell the unlettered masses what shall happen in the future.

Polling is a Science
So is astrology. Polling is divination masquerading as science. One of the dirty little secrets of scientific polling is that you can manipulate the results using subtle changes in wording, lead up questions designed to direct the respondent to the desired answer (push polling), or just changing the order of the questions.

Polling Accuracy
In 1936, the Literary Digest ran a postcard poll with a sampling size of 2 million, a margin of error of under 1%. The poll predicted Alf Landon defeating FDR 57% to 43%. The actual result gave Roosevelt a 61% victory.

The most recent Field Poll in California has Clinton (36%), Obama (34%), and undecided (18%) with a 4.5% margin of error. Throwing in the undecideds it is predicting the Clinton vote will be anywhere between 60% and 32%; Obama between 57% and 30%. I am confident this will be true. It is also a meaningless prediction.

This is the range predicted for the Republican California vote.
McCain: 56% to 28%
Romney: 50% to 20%
Huckabee: 39% to 9%
Paul: 25% to 5%
The only thing one can say with certainty is that McCain will beat Paul on Tuesday. While it is most likely the actual results will fall near the middle of those ranges (the Bell Curve principle), it is not guaranteed.

Of course, the range is never included in reporting on polling results. Ranges make for less exciting reportage and diminish the occult status pollsters seek. Pollsters want to appear prescient; publishing the ranges makes them appear what they are, educated guessers. Undecideds are ignored or divided up in some fashion to make the range disappear.

Cell Phones and Caller ID
Pollsters always claim they have a "random sampling." Sometimes they achieve this "randomness" by overweighting some of the results to compensate for "biased samples" (i.e. the lack of randomness in their random sampling). People (mostly young renters) with only cell phones are never polled. People who screen their calls with Caller ID are seldom reached. People, like me, who hang up on pollsters or, if we are feeling puckish, lie to the pollsters can screw up the results.

Compensating Errors and Polls of Polls
Pollsters really count on the accounting principle of compensating errors. They have to believe that the people they miss, or the people who screw with them, are just like the people they reach. In the end, they hope, that all of the things that screw up their random samplings will balance each other out.

The phenomena of compiling polls into some kind of grand total, like at Real Clear Politics, relies on the theory that the inherent errors in random polls are, themselves, random. If you average a collection of flawed polls the resulting total is, magically, not flawed.

Conclusion
I like reading polls. I also like ghost hunting television programs and stories about vampires. It amuses me to believe them but it is foolish to make decisions based upon them.

No comments: