In 1990, ABC, CBS, CNN, and NBC formed a consortium, Voter Research and Surveys (VRS), to control costs and deflate public perception that early projections were the result of network competition. VRS conducted exit polls of the 1990 and 1992 elections. Warren Mitofsky and Murray Edelman noted that the “biggest single change in exit polling since the days when the networks each did their own exit polls” was the leaking of early exit poll data, as estimates not intended for publication found their way on the air and into print.

In 1993, VRS merged operations with the National Election Service to become the Voter News Service (VNS), adding the Associated Press as a partner. Warren Mitofsky, head of VRS, declined to join the new consortium, establishing instead a competing company, Mitofsky International (MI), which conducted a competing exit poll in 1994. That same year, the competition to make early calls was rekindled, as ABC called several key races about an hour before VNS did. Subsequently, the other networks developed their own capability to make calls independent of VNS. (For instance, MI provided projections and analysis for both CNN and CBS beginning in 1996.)

Over time, the exit polls acquired a reputation for phenomenal accuracy. From 1967 to 1988, CBS made only five mistakes in approximately fifteen hundred calls. During the 1990s, VNS had only one error in seven hundred projections. This excellent record derived more from careful interpretation of exit poll data combined with election returns than from any uncanny accuracy in the data themselves, a distinction lost on many casual observers. For instance, the 1988 CBS exit poll interviewed more Dukakis voters than Bush voters, although Bush won the election by about eight percentage points. Here, as elsewhere, the unweighted results were misleading. Much (but not all) of the discrepancy can be attributed to an overrepresentation of black voters, which CBS presumably could detect and compensate for as the results came in—although, seventeen years later, we are unable to reconstruct the election night scene. CBS did incorrectly call one state (Illinois) for Dukakis, although not primarily based on exit poll data.

The presidential exit polls in 1992 and 1996 similarly had distinct biases—on net favoring the Democratic candidate Bill Clinton—that did not affect the election night calls. In 1992, the VRS exit poll overstated Clinton’s vote share by about two points (or his win margin by about four points); the average error within each exit poll precinct was about five points on the margin. (VRS incurred a larger error in the New Hampshire Republican primary, underestimating George H. W. Bush’s margin over Pat Buchanan by about ten points.) In 1996, the VNS exit poll overstated Clinton’s vote share by about one point. No state calls were affected, in part because analysts avoid calling competitive states from exit poll data alone.

In the 2000 election, however, VNS stumbled badly in the decisive state of Florida, where George W. Bush was ultimately declared the winner by just 537 votes out of almost 6 million cast. First VNS called Florida for Al Gore at 7:52 p.m., before polls had even closed in the Florida Panhandle, and was soon followed by the networks. Having reversed the Gore call, the networks (although not VNS) called the state for Bush a bit after 2 a.m., retracting those calls in turn by 4 a.m. Afterward, Congress held hearings on the debacle, while VNS and its subscribers commissioned a variety of reports.

The 2000 miscalls illustrated the complexity of election night estimation. Nationally, the exit poll estimates were even less biased than in 1996, although they leaned perceptibly toward Gore, the Democrat. In Florida, the initial exit poll estimate gave Gore a 7.3-point lead (while the true value was 0.0). As Mitofsky and Edelman reported in 2002, only about 2.6 points of this error came from survey error at the precinct level—that is, random sampling error and nonresponse bias. Most of the error owed to a faulty estimation of the absentee vote and an unlucky selection of the past race used for comparison in the projection model. The model relied on the past race with the best apparent fit, namely, the 1998 governor’s race. The 1996 presidential race, which fit almost as well, would have yielded a much smaller error—and a no-call decision. Mistaken vote-count data also influenced the calls. The Bush calls, in particular, were influenced by an error by Volusia County, which temporarily understated Gore’s tally by about twenty thousand votes.