Factors Affecting Political Poll Accuracy: An Analysis of Undecided Respondents

Recently, pollsters have experienced discrepancies between their poll predictions and subsequent election outcomes. One possible explanation of these divergencies is the growing proportion of undecided respondents in poll samples. This paper discusses three methods of reducing the proportion of undecided respondents: the use of contextual questions; a secret ballot device; and the use of a relative, rather than absolute, behavioural question. It concludes that, although it is possible to completely eliminate this group from poll samples, reductions in the proportion of undecided respondents are not necessarily matched by an increase in the accuracy of the poll estimates. Indeed, this study found that the sub-sample with the highest proportion of undecided respondents also produced the most accurate estimates. The paper concludes that direct intention questions, together with reduced contextual questions, and allocation of undecided respondents in proportion to party support levels, seem likely to produce the most accurate poll estimates.