因果效应推断下倾向性评分与潜变量联合模型的构建及应用研究

To evaluate the treatment effectiveness of certain therapy, non-randomized studies inevitably confront the influence of confounding bias. In fact, the interest treatment outcomes for many studies belongs to latent variables, which cannot be directly observed; often, the latent outcome cannot be comprehensively reflected by single observed indicator, which should be replaced by a set of observed indicators. Measurement errors should be reduced in order to estimate the latent treatment outcome accurately. In observational studies, in order to accurately evaluate the effect of certain therapy on latent outcome, our study aim to combine the propensity score method with latent variable model to develop the joint model, which can be used to reduce or control both the confounding bias and measurement errors simultaneously. We will propose the weighted model averaged method to reduce the adverse influence on estimation of treatment effect, induced by the uncertainty in propensity model selection. Approximate Bayesian estimation method combined with Markov chain monte carlo (MCMC) algorithm could be taken into account for estimating model parameters. Simulation study will be conducted to evaluate the accuracy and precision of joint model estimation of treatment effectiveness. Based on clinical practice data, our study will adopt the constructed joint model to assess the effectiveness of TCM dialectical therapy in treating COPD patients. Our study will provide a reasonable and valid statistical method for comprehensive assessment of treatment effect of certain therapy in observation studies, also, provide essential reference information for decision in optimal treatment plan.

评价非随机对照研究中治疗方式的有效性，常面临不可忽视的混杂偏倚的影响。实际中，很多研究关注的治疗结局是不可直接观测的变量，无法用单一指标进行全面评价，需利用多个指标反映；为准确估计潜在治疗结局，需有效控制测量误差的影响。为有效估计非随机对照研究治疗方式对潜在结局变量的效应，本研究拟将倾向性评分法与潜变量模型结合，构建能够同时控制混杂偏倚与测量误差的联合模型；基于模型加权平均法优化联合模型，解决倾向性评分子模型设定的不确定性对效应估计的影响，采用近似贝叶斯估计结合马尔科夫链蒙特卡洛算法估计模型参数，进行模拟实验评价建立的联合模型估计处理效应的准确性及精确性。基于临床实践数据，利用构建的联合模型评价中医辨证治疗对慢性阻塞性肺疾病的综合疗效。本研究将为非随机对照研究治疗方式的综合疗效评价提供合理、有效的方法，为患者最佳治疗方案的决策提供重要的信息参考。

评价非随机对照研究中处理因素的效应，常面临混杂偏倚的影响。很多研究关注的治疗结局是不可直接观测的变量，需利用多维指标进行反映。为综合评价治疗结局，应有效控制测量误差的影响。本研究将倾向性评分法与广义潜变量模型相结合，以同时控制测量误差与混杂偏倚。模型加权平均法被用于降低倾向性评分选择的不确定性对处理效应估计的影响。采用贝叶斯估计结合马尔科夫链蒙特卡洛（MCMC）算法估计模型参数。从每一候选的模型估计得到的倾向性评分的后验预测分布中进行抽样，基于每一次倾向性评分估计值，执行倾向性评分实施。基于倾向性评分哑变量向量的每个模拟值，估计处理效应，基于各候选模型的后验概率，计算得到模型加权平均后的处理效应估计的后验分布。进行模拟实验评价倾向性评分与潜变量联合模型估计处理因素效应的准确性及精确性，并比较不同模型选择方法估计处理效应的性能差异。进行模拟实验评价倾向性评分与潜变量联合模型估计处理因素效应的准确性及精确性，并比较不同模型选择方法估计处理效应的性能差异。模拟实验结果显示，倾向性评分联合广义潜变量模型估计处理因素效应的准确性较好;随样本含量的增加，该模型估计处理效应的RMSE呈减小趋势。当处理因素的效应真值设定为0.5时，贝叶斯模型加权平均结合广义潜变量估计处理效应的准确性优于单一倾向性模型结合广义潜变量模型。应用构建的贝叶斯倾向性评分与广义潜变量联合模型评价西医规范治疗组与中医辨证治疗组治疗慢阻肺患者的多指标综合疗效差异，单一倾向性评分模型结合广义潜变量模型、贝叶斯模型加权平均结合广义潜变量模型均显示中医辨证治疗的综合疗效相较于西医规范治疗组更优。本研究可为非随机对照研究处理因素的综合疗效评价提供合理、有效的方法参考。