There are three potential solutions to this bias. The first would be to allow an election poll to run the course of at least a full week, with moderate to vigorous refusal conversion attempts and callbacks. We immediately discount this possibility, since in today’s media climate, it is just not reasonable to expect the press to be interested in results more than three or four days old.

A second possibility is to conduct one such poll just before the final push, say, two months out from election day, to attain estimates of the bias particular to that potential election and allow compensating factors to be built into the preelection polls conducted from that point forward. Again, we believe this option is not optimal, since such bias could change over time, and because, generally speaking, most preelection pollsters have realized the best polls come through minimal intrusion of such factors (for the same reason, for example, nearly all pollsters refuse to weight by party identification unless such numbers tend to be dramatically off-target from expectations).

A final option, and one that we believe to be the most methodologically sound, is to employ a tightly controlled, rolling cross-sectional study rather than a fast-field-period, cross-sectional design. This design has been found in the past by at least one researcher to have increased accuracy over cross-sectional designs. Our findings provide suggestive support of this position. In a properly designed rolling cross-sectional study, a small amount of sample is released each day and is worked either until a completion or other final disposition is reached, or until it is a set number of days old—for example, ten to fifteen days. After being in the field for one week, the overall amount of sample being worked is a respectable mix of old and new, and since each sample piece is available for up to fifteen days, research firms are able to put significant efforts toward secondary dispositions.

Further, the design does not violate the constraints placed by media organizations, since on any given day the completes from the past three or four (or whatever is desired) number of days can be compiled for that day’s poll. With the rolling cross-sectional design, in short, each day becomes a random event, and the methodology of a more in-depth, high-response-rate study can be attained at the same time.

Many polling firms and their media partners already have such designs in place. The next step, for this analysis, will be to explore whether in 2004 there was a significant difference in the results of polls conducted with this methodology as compared to basic cross-sectional polls.

Overall, with today’s bipolar electorate, pollsters have to choose between risking erroneous results and paying more attention to accuracy. In a world where two percentage points has been making all the difference, we recommend strongly that pollsters conduct analyses similar to ours, where possible, to measure the bias potentially residing in their own surveys, and consider the use of the rolling cross-sectional design over weekly cross-sections.

We have been faced with the challenge of continuing to provide accurate and telling results in a tough media environment and a split electorate. Pollsters must respond by becoming more reflexive of their methods and of the potential impact of their methods if they hope to maintain high reputations and relevance into the future.

David Dutwin is a senior researcher and Melissa Herrmann is vice president of the Opinion, Policy, and Evaluation Group at ICR/International Communications Research.