Having established these general differences among dispositions, we next set out to explore specific differences within each disposition. Again, these differences were limited to completes that occurred at four or fewer call attempts, since the goal was to assess bias that might occur during a typical preelection poll. Furthermore, we set out to confirm the overall bivariate differences in a multivariate model.

Here is a summary of logistic regression results:

• Fifth+ call complete: Respondents who completed interviews with five or more call attempts were significantly, though moderately, more likely to say they voted for Kerry if they voted for a major candidate, although they were also, overall, more likely to say they voted for a third-party candidate. These respondents were also significantly more likely not to be registered, though those who were registered were more likely to say they voted. Also, there were a number of substantive negative relationships between completing a call at attempt five or greater and a range of demographic variables, including age, number of adults in the household, and number of children in the household. Finally, African Americans and Hispanics were more likely to complete interviews at five attempts or more.

• First call complete: Respondents who completed on the very first call attempt were substantively more likely to have voted for a major-party candidate, by a factor of two to one, compared to all other completed interviews. Also, first-call completes were more likely to be older, married, and white.

• Refusal conversion: Those who completed an interview through a refusal conversion were significantly more likely to say they voted for Kerry rather than Bush. Even more substantively, among registered voters, refusal conversions were far more likely—in fact by a ratio of four to one—to say they had voted in the 2004 election. There were no significant differences by any demographic or sample-level variables.

• Privacy managers: Respondents with privacy managers exhibited the most variation across the independent variables of the analysis. First, those who had call zappers were far less likely to have voted. However, those who did vote were substantively more likely, again at over a four to one margin, to have voted for a third-party candidate. Respondents who completed and had privacy managers were also more likely to be Democrats instead of Republicans and, even more so, to be Independent rather than either Republican or Democrat. Again, no significant differences were found with regard to demographics or sample variables.

• Callbacks: Respondents who completed interviews through callbacks were significantly more likely to have voted for a major-party candidate than a third-party candidate. They were also far more likely to say they voted, at nearly a four to one margin, to other types of respondents. No differences were detected with regard to demographics and sample variables.

• No answer only: There were no significant differences between no-answer respondents and others within the major variables of analysis. However, no-answer-only completes were more likely to be younger, divorced or widowed, white, highly educated, and have many children in their household.

What assessment can we make regarding the quality of preelection polls given these results? We believe the news is both good and bad. First, the bad: The differences reported above show that preelection polls in the 2004 election were likely biased toward President Bush, since first-attempt completes, four or fewer call-attempt completes, and no-answer-only completes all favored Bush by modest to substantive margins. These results suggest that in order to attain more accurate results, preelection polls should work sample more thoroughly, through both more call attempts and more secondary efforts, such as refusal conversions and the addressing of language barriers.

However, the good news is that doing so provides, to some degree, diminishing returns, since a far larger share of respondents in these secondary dispositions and higher call-attempt completes are not registered and less likely to vote even if registered. Furthermore, one must remember that a small number of the overall completes, even in a moderate- to high-response-rate survey such as the study reported here, are attained in the five to fifteen attempt range, and within secondary dispositions. Twenty-eight and a half percent of all the interviews completed for this study were completed on the first attempt, 73 percent in the first four attempts. And, while 45.5 percent of all interviews were completed through a secondary disposition, only 21.4 percent were completed through a secondary disposition within the first four call attempts. So, overall, the differences between five or more attempts and four or fewer attempts—an 8 percent swing— resulted in only a 2 percent impact from four or fewer call attempts and all attempts with regard to vote-share, and 3 percent with regard to voter registration.

Nevertheless, we believe these differences, substantive by disposition but moderate overall, are of significant concern, given the bipolar nature of the present U.S. electorate, where a difference of 2 percent in most preelection polls of the last two presidential elections often spelled the difference between calling the race for one candidate versus attaining a statistical dead heat. In short, a “2 percent bonus” within the last election likely pushed many election polls outside the margin of error and toward a statistical favoring of Bush over Kerry.