Sunday, September 19, 2004

The likelihood of God, the reliability of political polls, and the end of human civilization on Earth
Stephen Unwin's "The Probability of God" begins: "Do you realize there is some probability that before you complete this sentence, you will be hoofed insensible by a wayward, miniature Mediterranean ass?" Likely, you made it to this sentence, as the probability is so perilously close to zero that one can be forgiven for mistaking the two numbers as identical.

Probabilities and the predictions they birth can be tantalizing. Numbers seem to make concrete the unknown, and even the admitedly unknowable, and we eat it up. And Why? I think it's because scientists hoodwink us with their confidence. We rarely are told, or ask, the level of assumption and bias of either the equations/logic generating the conclusions or the scientist who decided on the them. Unwin tries to investigate God's existence outside of the role of faith and uses a mathematical formula--inputing numbers he believes in--to determine there is a 67% likelihood that a personal God exists. Michael Shermer of Skeptic magazine computes the exact same formula using numbers he believes in and concludes the probability at 2%.

Sir Martin Rees the astronomer brings his entire career of knowledge and study to bear in a new book, predicting that the chances of civilization lasting into the 22nd century are only 50-50 due to the rising number of ways developing to bring about the annihilation of humans. While a tad grim there, that prediction also seems a bit lacking in the courage department, assuring as it does that whatever happens, Sir Rees in a hundred years (or at the moment right before we're all wiped out) will be considered essentially to have been about half correct. And, not to mention that 50-50 is also the exact same probability one arrives at from a theoretical point of maximum ignorance. So why bother?

Those numbers and their level of insight is on my mind these days reading the wave of debate and consternation over the state of political polls, which are essentially predictions of how the election would turn out if it were held at that time. We understand that polls are "scientific" because they usually say that somewhere up front, to blunt our immediate reaction that 600 respondents couldn't possibly predict an election of millions. But, really, how do they work? Are they all scientific in the same way? Is there one scientific and one no-scientific way to poll an election?

The truth is they do not all follow the same rules. And, like the difference in variables used by Unwin and Shermer in predicting the existence of God, pollsters make similar decisions that affect the outcome of a poll. Some polls report the results just as the raw data presents, while others weight the results afterward to "more accurately reflect" the makeup of the voting public (usually based on previous election turnout). Some polls weight for party identification, so that if 50% of your respondents claim to be Republicans, their weight will be decreased before the results are announced, because in the last several elections self-identified Republicans have made up a much smaller percentage than 50.

Gallup is under fire, at least in the blog world (major media continue to present their polls as if they are not controversial), for basing the party ID weight on a model that has no reasonable basis, assuming the percentage of Republican voters to be well over where it actually has been in any recent national election. Not surprisingly, Gallup almost alone shows Bush with a healthy lead, along with the NYT poll, which uses a similarly indefensible model. And Zogby is claiming that traditional polling methods are more and more unreliable because they continue to depend on home phone calls. Respondent levels have dropped to nearly 25% from 36% 4 years ago. That means when you hear George Bush has polled at 50%, that's only 50% of the 25% that bothered to answer the phone and the questions. Why aren't polling firms required to report the results like that? I'd like to see it:

Not answering 52%
Not interested 25%
Bush 12%
Kerry 11%

I think it would make us a bit more properly skeptical. The question is, do the people not answering, or refusing to cooperate, represent a group that has characteristics that make them different, politically, than the group that answers? Or, are they essentially the same as the responding group, with the only difference being that they don't answer, or don't have, a home phone? Could that really be the only difference? This problem validates the necessity of weighting, but that also increases the amount of influence the pollster has over the results, depending as it does on the weighting model they decide to use, and the premise that Americans will turn out to vote, essentially, as they did in 2000.

But why would they?

People on all sides tell us this is the most important election of our lifetimes, and this is also the first Presidential election since Sept. 2001. So, if there is a time for voter turnout models to be quite wrong, one way or the other, it would be this year. I don't know which candidate it will benefit. And I don't think anyone does. Non-courageously, I see Kerry's chances at 50-50.

The sad thing, as I've said before, is that poll numbers do influence both turnout and results. I wish the media that report them would take that responsibility seriously. There are people--I don't know if they are more or less likely to be home, or answer their phone--who love jumping on a bandwagon. And the one making the rounds these days is that Bush is on the move.

No comments: