In two previous national elections, the exit polls had behaved badly. Premature calls by the networks in Florida led to a congressional investigation in 2000. Two years later, a computer meltdown resulted in no release of data on Election Day.

On Tuesday, new problems surfaced: a 2 1/2-hour data blackout and samples that at one point or another included too many women, too few Westerners, not enough Republicans and a lead for Democratic presidential nominee John F. Kerry in the national survey that persisted until late in the evening.

In two instances on election night -- the results for Virginia and South Carolina -- the networks held off projecting a winner when voting ended because exit polls showed that the races were too close to call, only to see President Bush win easily in both states.

[...]

Some problems are inevitable. A total of 12,047 randomly selected voters were interviewed Tuesday as they left their polling places, and those results were fed into computers. The accumulated results were reported several times over the course of Election Day.

Results based on the first few rounds of interviewing are usually only approximations of the final vote. Printouts warn that estimates of each candidate's support are unreliable and not for on-air use. Those estimates are untrustworthy because people who vote earlier in the day tend to be different from those who vote in the middle of the day or the evening. For instance, the early national sample Tuesday that was 59 percent female probably reflected that more women vote in the day than the evening.

That is why the early leaks anger Lenski. "The basic issue here is the leaking of this information without any sophisticated understanding or analysis, in a way that makes it look inaccurate," he said.

After the survey is completed and the votes are counted, the exit poll results are adjusted to reflect the actual vote, which in theory improves the accuracy of all the exit poll results, including the breakdown of the vote by age, gender and other characteristics.