Red-Blooded Risk: The Secret History of Wall Street

Polling

This is not a book on statistical practice, but I'm conscious that everything so far is pretty abstract. So let's take one example: political polling. I picked this because it is an important practical application of statistics, and one that seems to be pure objective probability, with no reference to money. In Bernoulli's coin-flip world, political polling is the same problem as drawing marbles from an urn. You sample some random voters and use their replies to set a confidence interval on the election results.

In the real world, the problem is much harder. It is difficult to get a truly random sample, and people don't always tell the truth about whether they will vote or how they will vote. Or they can tell the truth at the time but later change their minds. And it turns out the best sample is not a purely random one; it makes sense to concentrate your subjects in swing vote groups, and to weight each response by the number of similar voters in the population. With all of these difficulties and adjustments, no one would take seriously a theoretical computation of error rate. The only reason to believe a poll result is if the researcher or organization that administered it has been right in the past. Of course, many polls do not even try to be correct, but I'm talking about the ones that do try. Also, some polls are for very specific purposes, for example to direct campaign spending to these most advantageous places. These have to be judged for their success at their purpose. ...