|
The short answer is simple: It's getting very hard to draw a random sample of people who will vote.
In this context, the concept of "margin of error" is quite misleading, and is not actually a measure of the poll's range of accuracy. M of E only holds meaning if the poll sample is truly randomly drawn from the population to which one wishes to generalize.
The population in Dem-v.-Bush polls everyone is getting excited about lately is hypothetical--it's the totality of people who would vote for President today if given the chance. The M of E as calculated in instances like this is purely a function of sample size. Increase the sample and reduce the M of E. It's actually a square root function, not a linear one, so you hit a point of diminishing returns over a thousand or 2 respondents. But the point remains, no poll is any better than the sampling technique used to select the participants, regardless of the M of E. Increasingly, as people tend to refuse to answer polls or lie to pollsters, or switch entirely to cell phones, or delist their numbers, the obtained samples do not match the population, and the polls are accordingly corrupted. The single exception is exit polls, which remain relatively accurate because the pollster has access to the entire population of interest, i.e. people who just actually voted. However, exit polls don't see very far into the future at all; they at best allow one to project comlete results on the basis of partial returns.
|