世論調査と出口調査データから推定した傾向スコアを用いた選挙予測について

摘要

A newspaper company usually conducts two kinds of surveys concerning an election. One is a pre-election survey done a week before the election for the purpose of predicting election results. Another is an exit poll done just after the election for the purpose of obtaining information about the relationship between voters' demographic characteristics, their political attitudes, and their electoral decisions. It is known that percentages of answers to the question, "Which candidate will you vote for?" are further away from the actual election outcome than those obtained through exit polls. In this paper, we estimated propensity scores through both kinds of survey data. The estimation was performed by using covariates such as sex, age, residential area of the voter, political party identification and criterion for selecting the candidate. Next, three adjustment methods were applied to improve predictions using pre-election survey data: (A) weighted averaging by splitting estimated propensity scores into strata, (B) a modified method of (A), and (C) propensity weighting. As a result, forecasted shares of the vote were improved for 70 percent of the candidates in the sample using method (B).%報道機関では，選挙に関する調査を主に2種類行っている．1つは投票日の1週間前に行う 情勢調査で，もう1つは投票日当日に行う出口調査である．これらの調査での「誰に投票する か（したか）」の回答を見ると，情勢調査に比べて出口調査の結果の方が実際の開票結果に近 い．本稿では，この2種類の調査の回答データから傾向スコアを推定する．推定には，性別， 年齢，投票区の市郡別，政党支持，投票の基準の回答データを共変量として使った．そして推 定した傾向スコアを使って3つの方法（A）傾向スコアで層別して加重平均，（B）（A）の係 数を修正したもの，（C）傾向スコアを使った重み付け，によって情勢調査からの予測の改善 を試みた．その結果，（B）を使った方法において，有力候補の7割の予測値に改善が見られ た．