Prediction of Exit Polls for the Presidential Election in 2007 and 2013 by Using Nonresponse Models

Abstract

Estimation of nonresponses in exit polls frequently produces inaccurate election results. Since estimates of nonresponses differ depending on the assumed mechanism, nonresponse mechanism should be determined prior to the analysis. However, nonresponse mechanism is not testable. From this problem, estimates from all possible models need to be provided. Along with the selection of nonresponse mechanism, ML (maximum likelihood) estimates under NMAR (not missing at random) have a boundary solution problem which is that cell expectations of more than one columns are estimated as zeros. To overcome the boundary solution problem, various Bayesian models have been suggested. In this paper, ML estimation under MAR (missing at random) and NMAR as well as Bayesian estimation under NMAR are applied to the two exit polls for 2007 and 2013 Korean presidential elections with 53 contingency tables. In comparing ML and Bayesian estimates, we also consider weighting methods to reduce biases. As results, ML estimates under MAR produce more accurate election results and better than the Bayesian approaches under NMAR.