An Analysis of the Accuracy of "Trial Heat" Polls During the 1992 Presidential Election

摘要: There is little systematic research on the multitude of factors that influence accuracy poll results. This article examines six methodological directly under survey researcher's control and two exogenous concerning nature public opinion as sources error. Data for this study come from 56 "trial heat" polls conducted during last month 1992 presidential election. The most important variables influencing were number days a in field, which increased total one-half percentage point per day; conducting interviews only weekdays (and thus evening hours), reduced overall rates by more than 1 point; "tracking" poll, about 1.5 points. Sample size was not related to rates. Results also indicated sampling frames "likely voters" (relative "registered voters") tended overestimate support George Bush underestimate Ross Perot, interviewing led overestimates Bush, strict methods defining respondent "supporting" candidate hurt newcomers, Perot Bill Clinton, Bush. In light these data it recommended common practice reporting "margins error" based solely sample sizes be abandoned misleading replaced empirically justifiable measure response One favorite postelection retrospectives played after every campaign "How well did do predicting levels different candidates?" (see, e.g., Broder 1988; Costikyan Kaplan 1984; Morin 1989; Richburg Roper 1983). That question seems particularly pointed 1993 campaign, with added volatility caused candidacy, controversies surRICHARD R. LAU iS professor political science at Rutgers-The State University. He would like thank Janice Ballou, Gerry Pomper, Cliff Zukin, anonymous reviewers helpful comments earlier drafts article. Public Opinion Quarterly Volume 58:2-20 ? 1994 American Association Research All rights reserved. 0033-362X/94/5801-0003$02.50 content downloaded 157.55.39.123 Mon, 18 Jul 2016 06:00:49 UTC use subject http://about.jstor.org/terms An Analysis "Trial Heat" Polls 3 rounding wildly fluctuating results due, some accounts, poorly explained midcampaign changes methodology (Brennan 1992; Houston Rosenstiel Taylor Krane 1993; but see Hugick, Molyneux, Norman 1993).1 President Bush's late surge polls, noticeably greater vigor his campaigning few weeks general underestimation magnitude vote may have all been result artifacts due shifts ("registered or how "undecided" voters are allocated between candidates. Add overestimation Clinton's victory early exit election day (Public Perspective 1993), we fairly mixed picture polling campaign. Some fluctuation inevitable, course, well-known principles any random distribution: larger sample, smaller error can expected. What much less known other aspects process, including choice frames, rates, instability opinion, so on, affect fact, Buchanan (1986) estimates observed fluctuations vary twice expected alone. Converse Traugott an excellent job listing variety could accuracyincluding size, "house" effects, composition, interviewer training, questionnaire construction, on-but they report simultaneously considered host such factors. Only Crespi (1988, chap. 9) reports multivariate analysis considers impact endogenous (sample importance goal researcher, until election) (primary election, whether incumbent running reelection, turnout margin winning candidate, percent undecided public) factors, where "endogenous" refers control, "exogenous" characteristics environment beyond control. striking aspect Crespi's size-the one factor base "margin calculations reported results-did significantly predict independent equation. attempt contribute line research. I 1. Whether misunderstanding fault researchers media who their another question.