傾向分數分層法應用在隨機對照試驗上

Abstract

臨床試驗中研究者有時會進行變數調整以提升研究檢定力。然而這讓研究者可以挑選基線變數進行調整以尋求有利實驗治療的結果。這樣會令人質疑變數調整的客觀性。機率倒數權重法可以改善這問題。但我們發現，當臨床試驗的平均治療結果越遠離0時，機率倒數權重法的表現會越差。我們提出傾向分數法。此法保有機率倒數權重法的客觀性優點，而無其缺點。我們進行電腦模擬探討此法之統計性質。我們認為傾向分數法值得推薦於臨床試驗中使用，如此可保有客觀性，又能提高研究之檢定力。

Researchers sometimes conduct covariate adjustments in clinical trials to improve study power. However, this allows researchers to cherry-pick baseline covariates in their analysis in order to seek favorable results of the experimental treatment. This casts doubts on the objectivity of covariate adjustments. The inverse probability weighting method can mitigate the problem. But we found that when the average treatment results of the trial are more away from zero, the performances of the inverse probability weighting method will become poorer. We propose a propensity score method. This method retains the advantage of objectivity in the inverse probability weighting method but without its shortcoming. We conduct computer simulations to investigate its statistical properties. The propensity score method is worth recommending for use in clinical trials to preserve objectivity and enhance study power.