Sample size and uncertainty when predicting with polls: the shortcomings of confidence intervals

Abstract

The procedure we propose uses polling data to construct a probability model that recreates numerical results from a large number of simulated elections. Probabilistic measures of candidate success have become increasingly common in some areas of election prognosis, moving away from traditional procedures based on confidence intervals. Here we show that, with the same information used to construct a confidence interval, a more precise projection of election results can be calculated demonstrating the probability of a certain candidate winning the election. The procedure can take into account respondent nonresponse of “do not know/refuse to answer” (dk/ref). The ambiguities inherent in confidence intervals and their margins of error are avoided by calculating the probability that one candidate receives more votes. Importantly, throughout the article, we show that our procedure requires a smaller sample size and produces more predictive accuracy.