工具變數為處理非隨機試驗所面臨問題的方法之一，近來廣泛應用於計量經濟及流行病學領域；其主要目的在於控制不可觀測的干擾因素，使資料經過調整後「近似」於隨機試驗所得的資料，進而求出處理效果的一致估計值。由於先前研究大多探討連續型變數的情形，本篇論文將透過模擬與實證分析，針對二元之工具變數、反應變數及處理變數，比較一階段廣義線性估計量，two-stage predictor substitution (2SPS)，two-stage residual inclusion (2SRI)，及two-stage residual inclusion-Taylor expansion (2SRI-T) 這四種估計方法。 模擬結果顯示，當偏誤為主要考量時，2SPS與2SRI有較好的表現；然而，同時考慮偏誤及變異的情況下，2SRI-T則為較適合的估計方法。值得注意的是，模擬試驗所得出的結果與Terza等(2008)不同，2SRI並未優於2SPS。另外，將此四種方法套用至探討有小孩與否對生活的滿意度的影響之實際資料，其表現結果與模擬試驗結果一致。 / Instrumental variable (IV) analysis, one of the techniques to solve problems generated from non-random experiments, has been increasingly applied in many fields such as econometrics and epidemiology. Its utility stems from the belief that IV, if correctly selected, can potentially mimic randomization by adjusting for unmeasured confounders. However, because of less concern about IV analysis on categorical data, we center our discussion on binary outcome, treatment, and IV in this study. Four methods are compared: the one-stage generalized linear model (GLM), two-stage predictor substitution (2SPS), two-stage residual inclusion (2SRI), and two-stage residual inclusion considering Taylor expansion (2SRI-T). We conduct both the simulation and the empirical study to evaluate the performances of these four estimators. The simulation results indicate that, while 2SPS and 2SRI have better performances than the other two estimators with respect to the bias, they suffer from larger variability. On the other hand, 2SRI-T generally has smaller standard error than 2SPS and 2SRI, and hence might be preferred if MSE is the main concern. Noticeably, it also suggests that 2SRI does not outperform 2SPS which was inversely shown in Terza et al. (2008). The same conclusion is also found when implementing these methods on a real dataset to investigate whether having children has significant effect on one’s life satisfaction.