以蒙地卡羅法探討人工對照組之適用性

Abstract

本研究透過蒙地卡羅法之模擬以評量Hsiao et al. (2012) 提出的縱橫資料反事實法(Panel-Data Counterfactual Method)與 Abadie et al. (2010) 提出的合成對照組法(Synthetic Control Method)之人工對照組估計之準確度。在使用主體之選擇由資料產生，與具單位處理效果(Treatment effect)之估計方法中，何者對隨時間變動之處理效果估計較佳為本研究之目標。將此二法，應用於多個實例資料，進行模擬及交叉比對，以評估其可用性。結果顯示當共同因素在時間中的變動與因素負荷在地區間的變動均小時，此二法之估計均良好。但以均方差衡量對人工對照的估計時，縱橫資料反事實法在大多數情況下, 比合成對照組法準確。儘管此二法均需謹慎使用，由蒙地卡羅模擬之結果顯示，縱橫資料反事實法明顯較佳。

The accuracy of the artificial control estimation using the panel-data counterfactual method proposed by Hsiao et al. (2012) and the synthetic control method proposed by Abadie et al. (2010) were evaluated using the Monte Carlo simulations. The aim was to determine which of the methods is superior in studies with time-variant treatment effect, individual treatment effect and data-driven subject selection process. A cross checking process and simulations conducted under various model settings provide guidance on the applicability of these two methods. Both methods perform satisfactory when the variation of common factors in time and factor-loadings across regions are small. In most cases the panel-data counterfactual method is more accurate in artificial control estimation in term of mean-square-deviation criteria than the synthetic control method. Though both methods must be used with caution, the panel-data counterfactual method is clearly the better method suggested by the Monte Carlo results.