多组资料倾向性评分加权法模型构建与比较

Construction and comparison of propensity score weighting models for multiple treatments

目的 比较处理多组资料数据时三种倾向性评分加权方法在不同样本量的条件下均衡协变量的能力和估计处理效应的优劣。方法 采用Monte Carlo模拟的方法生成数据集,比较三种倾向性评分加权方法Logistic-IPTW、Logistic-OW和GBM-OW法均衡协变量和估计处理效应的优劣。协变量均衡水平指标为绝对标准均值差。处理效应评价指标包括处理效应的点估计值、均方根误差、可信区间覆盖率。结果 在5种协变量与处理因素、结局变量有不同复杂程度的相关关系的场景下,相比于Logistic-IPTW法和Logistic-OW法,GBMOW法在效应估计方面更优,同时拥有更小的均方根误差;在协变量均衡性方面,三种方法效果都比较好,GBMOW方法在多组资料倾向性评分分布重叠度较低,且在协变量与分组变量、结局变量有越来越复杂的非线性关系时表现变好。结论 在处理多组资料时,GBM-OW法相对另外两种方法,在协变量与分组变量、结局变量之间存在非线性或(和)交互关系时具有优势。运用此方法后效应估计更加接近真实值,为较优的选择。

Objective To compare the ability of three propensity score weighting methods to balance the covariates and the advantages and disadvantages to estimate the treatment effects when dealing with multiple treatment data under different sample sizes.Methods Monte Carlo simulation was used to generate data sets and the advantages and disadvantages of balancing covariates and estimating the treatment effects of three propensity score weighting methods,Logistic-IPTW,Logistic-OW and GBM-OW were compared.The evaluation index of covariate equilibrium level was the absolute standard mean difference.The evaluation indexes of effect estimation included the point estimate of treatment effect,root mean square error and confidence interval coverage.Results Compared with Logistic-IPTW and Logistic-OW,GBM-OW was better in effect estimation and had a smaller root mean square error in five scenarios where covariates were related to treatment factors and outcome variables with different varying degrees of complexity.In terms of covariate equilibrium,all three methods had good effects.GBM-OW method performed better when the overlap of propensity score distribution of multiple treatment data was relatively low and covariables had increasingly complex nonlinear relationships with treatment factors and outcome variables.Conclusion When dealing with multiple treatment data,GBM-OW method has advantages over the other two methods when there is nonlinearity and/or interaction between covariates and treatment factors and outcome variables.Using this method,the effect estimation is closer to the real value,which is a better choice.