Bias in Preelection Polls

The primary focus of research regarding preelection polls is, unsurprisingly, the question of accuracy. Pollsters and academics alike are, of course, always striving to make such polls more accurately reflect actual voter sentiment, and indeed, historical research (for instance, a 1979 study by Paul Perry) has found a significant improvement over the years. Perhaps the most thoroughly researched concern is likely-voter models, since selecting who will be considered a voter in any upcoming election has the most obvious impact on getting accurate results. Without the right sampling frame, it is truly hard to get a proper sample. 

Others have studied polling accuracy by days and “daypart.” In 1994, Richard R. Lau, for example, found that every day a poll was in the field tended to increase its accuracy by about half a percentage point. Furthermore, Lau found that it was important to interview on both weekends and weekdays: Polls conducted only on weekdays were around one percentage point less accurate. Other findings included evidence that tracking polls were one and a half percentage points more accurate than cross-sections.

A select few others have looked at the potential bias that may be introduced into a poll based on how sample is worked. In 1981, Warren Mitofsky first noted that respondents who completed an interview on the first attempt differed in candidate preference from those who completed on later calls. Extending this question into specific dispositions in 1991, Richard Bolstein explored potential differences between completed interviews and refused interviews, as well as households that never answered the phone and respondents who were selected but said they were not available during the poll, through validation of actual voter roles. He found that households fitting the latter two definitions, “call rule exhausted” and “unavailable during the poll,” were far less likely actually to have voted.