不同混杂结构条件下各倾向性评分方法的模拟比较研究

Comparison of Propensity Score Methods Under Different Confounding Structures:A Simulation Study

目的通过构建不同混杂结构的处理因素模型和结局模型、不同相关性的协变量,比较多种倾向性评分方法在结局模型为线性回归模型的情况下估计处理效应的优劣。方法采用Monte Carlo模拟方法,通过构建四种由简单到复杂的不同结构的混杂模型,生成相应的数据集,再分别应用倾向性评分匹配、回归调整、加权以及分层的方法估计处理效应并进行比较。评价指标包括点估计、标准误、相对偏倚、均方误差。结果在结局模型为线性回归模型情况下,倾向性评分回归调整法估计的相对偏倚最小,稳定性也最好。匹配法卡钳值取0.02较卡钳值取倾向性评分标准差的0.2倍估计的相对偏倚更小。当处理因素模型中含有非线性效应时,用逆概率加权法估计的偏倚较大,并且加权法估计的标准误也最大。倾向性评分分层法在各种情况下估计的相对偏倚都较大。结论倾向性评分回归调整法能够较好地估计处理效应,并且在各种情况下估计都较为稳健。建议当协变量与处理因素和结局变量的关系无法确定时,这四种方法中可以考虑优先使用回归调整法。

Objective The performance of propensity score（PS）methods were compared through constructing different confounding structures and generating covariates with different correlations when the outcome model was linear. Methods Monte Carlo method was used to simulate the datasets by constructing four confounding structures from simple to complex. Then four PS - based methods including PS matching, covariate adjustment, inverse probability of weighting （IPW） and stratification were applied to estimate the treatment effect. The results were compared from different aspects including the point estimate, standard error,relative bias and mean square error. Results When the outcome model was linear, covariate adjustment showed the least biased and stable estimates among the four methods. PS matching with caliper 0. 02 performed better than the other matching methods when the caliper is 0. 2 of the standard deviation of the PS value. When there were nonlinear relationship in the treatment model, IPW showed biased results and largest standard error. PS stratification resulted in biased estimates in all settings. Conclusion Covariate adjustment by PS is robust to complex confounding structure and achieved the least biased estimates. We propose that when the relationships between confounding factors and treatment or outcome variable cannot be confirmed, using PS covariate adjustment seems a better choice.