抽樣調查資料因樣本代表性失真而造成對母體推估偏差的補救辦法通常是採用加權處理，基本上以社會人口特徵為依據。然而僅從樣本人口特徵的分布是否和母體相符來判斷樣本的代表性並不能保證樣本在認知、態度與行為等主題變項在分布上的推論就不會有偏差。本文以總統選舉投票行為的調查資料為例，探討依據入選機率的次樣本分組（subclassification on the propensity score）所調整的電話調查結果在投票行為推估方面的成效。首先利用二○○四年台灣地區社會變遷基本調查四期五次公民權組問卷資料，以樣本重抽法（bootstrapping）產生包含20,000 案的擬母體（Pseudo-population），再從擬母體中以簡單隨機抽樣法抽取200 案為參考樣本，另外從一個典型的電話訪問調查資料中以分層隨機法抽出800 案為試驗樣本。兩套樣本組合成為一套1,000 案的新樣本。其次依據Lee (2006) 的入選機率調整法(propensity score adjustment, PSA）進行電訪樣本「投票行為」估計值的調整。整個流程進行2,000 次的模擬分析。評估結果確認藉由次樣本分組所產生的入選機率調整法的確有降低推估偏差的功效。

It is common to use socio-demographic variables for weighting, but this does not mean that other inferences, especially on attitudes and behavior variables, will be free of bias. This article takes voting behavior in a presidential election as an example to examine the effects of adjusting the telephone survey results using subclassifications on the propensity score as suggested by Lee (2006). We first used the Taiwan Social Change Survey (TSCS, 2004) wave 4 module for a citizenship study to produce a pseudo-population dataset based on bootstrapping sampling. A random sample of 200 cases was drawn from this dataset. In addition, a stratified random sample of 800 cases was drawn from a telephone survey. The two samples were combined into a sample of 1,000 cases. Propensity Score Adjustment (PSA) allows us to adjust and evaluate the proper estimation of voting behavior based on the subclassification of propensity score. The procedure was repeated 2,000 times. The results showed that the PSA method does effectively reduce telephone survey estimate bias.