Behind the Controversy: A Primer on U.S. Presidential Exit Polls

By Mark Lindeman and Rick Brady

Part II. Exit Polling Methods and Sources of Error

It has been argued that exit polls are inherently less error-prone than other surveys because they simply ask people to report what they have just done. Indeed, exit polls avoid the inherent uncertainty of preelection polls about who will vote, as well as for whom the actual voters will vote. They are, however, vulnerable to their own errors, which can be mitigated but not reliably eliminated. It is because of these vulnerabilities that the exit poll designers resist projecting outcomes before receiving considerable vote-count data, unless the indications are overwhelming. And it is these vulnerabilities that make exit polls, at least as now designed, “blunt instruments” for detecting fraud in the vote counts, if it exists.

Precinct sampling

To start with, the polltakers must select precinct samples that accurately reflect the political characteristics of each state race. In the 2004 presidential election, the National Election Pool (NEP) attempted interviews in 1,480 precincts nationwide, ranging from 15 precincts in uncompetitive states to 55 in Florida and 50 in many other battleground states. (For instance, in Ohio, NEP interviewed in 50 out of some 11,500 precincts statewide.) The NEP draws a substantially larger research sample well in advance of the election, and then draws exit poll precincts from it closer to election day.

Warren Mitofsky has described the drawing of the research sample as follows (much condensed):

First, precincts within a state are divided into two strata by size, of roughly equal total vote, to ensure that small precincts are adequately represented in the sample. (No interviews are conducted in very small precincts, however, because the results would be subject to large error variance.) Within these strata, precincts are further divided into county-based geographic strata based on salient characteristics, such as region or urbanity.

Finally, a precinct sample is drawn with probability proportionate to size of precinct (PPS), such that the range of partisanship within each stratum is evenly represented, and each voter in the state has a roughly equal probability of being selected. In drawing the final exit poll sample, precincts that are more than 80 percent African-American are oversampled (and subsequently downweighted), because turnout variations in these precincts can have a disproportionate impact on the election results.

A cluster sample of fifty-odd (or many fewer) precincts out of an entire state is inherently error-prone. But in 2004, the precinct sampling seems to have performed reasonably well. While the official vote counts in several state samples did significantly misestimate the final returns, the errors were generally modest and relatively unbiased overall. (The mean error for the exit poll samples was 0.43 percentage points in favor of George W. Bush.)

Absentee and early voters

Paradoxically, one problem confronting exit pollsters is the integration of data on voters who cast their votes before election day. These absentee and early voters—or convenience voters—cast about 12 percent of votes in the 1996 presidential election, but about 19 percent in 2004. The growing trend toward convenience voting poses new challenges for exit pollsters. If convenience voters’ preferences vary sharply from those of election day voters, then failure to estimate either the size or the distribution of the convenience vote may wreak havoc on projections.

VNS first supplemented election day exit poll data with a pre-election day survey of convenience voters in 1996. Daniel M. Merkle and Murray Edelman have concluded that across states, convenience voters are older than election day voters, but other patterns—including party affiliation—vary from state to state. In Florida in 2000, the Voter News Service (VNS) projection model assumed that 7 percent of votes cast would be from convenience voters; the actual figure was 12 percent. Because convenience voters were almost 24 percent more Republican than other voters, the model distinctly underestimated Bush’s support. In 2004, the Florida telephone poll of convenience voters may also have underestimated Bush’s support.

Convenience votes can also complicate postelection analysis of exit polls in jurisdictions where these votes are combined with election day votes by precinct. In this instance, it can be difficult or impossible to compare election day votes with the exit poll results.