倾向性评分法和马氏距离法在匹配中的比较与应用

Comparison and Application of Matching in Propensity Score Method and Mahalanobis Distance Method

目的比较倾向性评分法与马氏距离法在匹配中的效果,在医学数据中验证倾向性评分悖论的观点。方法通过最邻近匹配及卡钳匹配选择最佳匹配方法,计算不同卡钳值下删减个体数后样本的不平衡性,比较倾向性评分法与马氏距离法的稳定性。结果对于本研究的数据,倾向性评分法的卡钳匹配是最佳的匹配方法;倾向性评分法在删减个体数达到一定后,继续删减匹配较差个体会增加样本的不平衡性,马氏距离匹配的样本不平衡性随着删减个体数的增加而减少。结论倾向性评分匹配法调整混杂时,不宜删减较多个体寻找更加精确匹配的匹配集。

Objective To compared the effective of matching in propensity score method and mahalanobis distance method, and to verify the viewpoint of propensity score paradox（PSP） in medicine data. Methods Choosing the best matching method through the nearest neighborhood matching and caliper matching of propensity score matching （PSM） and mahalanobis distance matching（MDM）. Calculation the samples imbalance with different caliper values by pruning different numbers of ob- servation, comparing the stability between the PSM and MDM. Results For the existing Right Heart Catheterization dataset, cal- iper matching of PSM is the best matching method;and the level of imbalance is decreased when the numbers of pruned observa- tions reaches a certain, then samples imbalance is increased as more observations are pruned randomly in PSM. By contrast, as pruning the number of observations turns to decrease the samples imbalance in MDM. Conclusion Using propensity score matc- hing to adjust the confounding factors, should not be pruned more observation to look for more precise matching dataset, other- wise propensity score paradox will be appeared.