2016 Presidential Statewide Polling—A Substandard Performance: A Proposal and Application for Evaluating Preelection Poll Accuracy

Abstract

This study implements a statistical accuracy (SA) measurement for assessing preelection poll accuracy by comparing Mosteller (1949) Method 5 (absolute difference between poll results and election results) with the poll’s margin of error (MOE) or credibility interval. The expectation is that 95% of poll results would be SA by falling between the poll’s margin of error or credibility interval and the actual margin of victory. The new measurement is described and then applied to the statewide preelection polls from the 2012 Presidential (n = 331) and 2016 Presidential (n = 539) races using n = 182 polling organizations in the last 21 days of each election cycle. This analysis finds statewide preelection polling in 2012 had a 94% SA and was not statistically different from the expected 95%, while the statewide polling in 2016 had a 77% SA and a binomial test found the distribution differs significantly from the expected 95%. There is a significant difference in SA between the two election cycles, χ²(1, N = 870) = 45.24, p < .000. The 2012 biased polls favored the Republican candidate 68% of the time; however, a binomial test found this distribution did not differ significantly from the expected 50/50 distribution, .50, p = .167 (two-tailed), suggesting this was caused by random error. In 2016, biased polls favored the Democratic candidate 90% of the time, a binomial test indicated that the proportion was higher than the expected .50, p < .000 (two-tailed), suggesting a systemic bias.