Predicting the Likelihood to Vote in Pre-Election Polls

A problem of particular interest in pre-election polls is to predict the likelihood that a sampled individual, whether respondent or non-respondent, will vote. This can be especially difficult in state and local elections since voter turnout is low compared to national elections. We study this problem by means of validated pre-election polls of registered voters in a US city which were conducted for the 1988 Presidential and 1989 Virginia State Gubernatorial elections. Results indicate that respondents and 'refusals' are significantly more likely to vote than non-respondents who cannot be contacted, and that a respondent's self-described likelihood to vote is the best indicator that he will. A logit model is used to predict the likelihood that a registered voter will vote in the State election based on voting status in the Presidential election, age and sex. The presence of the variable sex in the model is explained by the high profile of the abortion issue in the campaign. 1 Background A problem of particular interest in pre-election polls is to predict the likelihood that a sampled individual will vote. Answers to questions aimed at assessing a respondent's likelihood to vote are used to determine what respondent base to use to predict the election outcome. Perry (1973, 1979) concluded that the answer to the question 'are you registered' is the best predictor followed by the respondent's self-described likelihood to vote. Bolstein (1989) obtained results consistent with this in a study restricted to registered voters. It is also of interest to study the likelihood of non-respondents to vote, both by type of non-respondent and in comparison with respondents. This can be important in projecting election results if the non-respondents have different voting patterns. In his study, Bolstein (1989) found no difference in the likelihood to vote among the respondents and those who refused to be interviewed, and that the likelihood to vote was significantly greater in these 'contact' groups than in the 'call-rule exhausted' group, and significantly greater in the combined 'refusal' and 'call-rule exhausted' groups than in the 'unavailable at time of interview' group. The problem of predicting whether an individual will vote can be especially difficult in state and local elections since voter turnout is low compared with national elections. The work of Traugott & Tucker (1984), in which they concluded that past voting behavior is a strong predictor of whether or not an individual will vote, suggests that modeling current voting status (will or will not vote in upcoming election) in terms of past voting status might be especially useful in predicting whether or not non-respondents will vote. In this paper we use a logit model to represent voting status in the 1989 Virginia Gubernatorial election as a function of voting status in the 1988 Presidential election, sex and age. The latter two variables are available on voter registration lists and hence are known for non- respondents as well as respondents. It is no surprise that age is a factor in voting status since it is well-known that voting is a habitual characteristic that requires time to take hold. Thus, Bolstein (1989) found that registered voters under 30 were significantly less likely to vote than those 30 or older in the 1988 election. In a study of spousal voter turnout in the 1984 election based on Current Population Survey data, Straits (1990) also found that the odds of voting increases with age. On the other hand, sex is a factor in our