多组比较的倾向性评分模型构建及匹配法的研究和应用

发布时间：2018-05-08 20:04

本文选题：倾向性评分匹配 + 最邻近匹配法　；参考：《第二军医大学》2014年博士论文

【摘要】：研究背景：随着信息技术的不断发展，观察性研究无论是在数量上还是在研究准确性上都在不断增加和提高。大样本的观察性研究在医学研究当中发挥着越来越重要的作用。但在观察性研究中，由于研究对象所在的组别不是随机分配的，而是自然存在的，因此具有某些特征的研究对象更倾向于进入处理组或对照组，导致不同组间存在混杂偏倚。倾向性评分法(propensity score, PS)是解决观察性研究中存在混杂偏倚的常用研究方法。该方法便于理解、研究步骤标准化程度高，近些年在非随机化大样本的观察性研究当中被广泛应用。倾向性评分法的应用主要包括匹配法、分层法和回归校正法等，以匹配法最具优势，应用范围也最为广泛。倾向性评分匹配法主要包括最邻近匹配法、卡钳匹配法和马氏距离匹配法等几种方法。目前，对于倾向性评分匹配法的应用上还有一些问题尚未得到解决。例如，对于在倾向性评分模型中应放入何种类型的协变量，目前仍存在着争议；何种匹配方法更具优势目前尚未得到定论；另外，目前倾向性评分匹配法主要用于分组因素为二分类的观察性研究资料，很少有研究将其用于分组因素为多分类的观察性研究资料中。研究目的：构建分组因素为有序三分类的倾向性评分匹配方法。通过模拟研究筛选纳入到倾向性评分模型中的协变量，比较多种匹配方法在分组因素为有序三分类情况下优劣，通过调整参数确定不同数据特征下最具优势的匹配方式，同时在分组因素为有序三分类的情况下对不同倾向性评分应用方法进行比较，最后将模拟研究中建立的最优倾向性评分匹配方法应用到实际数据分析中。研究方法：本研究采用蒙特卡洛法模拟数据集。分组因素模拟为有序三分类，并分别调整不同组间的样本量比例为1:1:1、2:3:5、1:2:3和1:4:5。根据协变量与分组因素和结局的关系模拟不同类型的协变量，包括与分组因素和结局均相关联的协变量、与分组因素相关联的协变量、与结局相关联的协变量和与分组因素和结局均不相关联的协变量。通过在倾向性评分模型中纳入不同类型的协变量，确定在分组因素为有序三分类情况下倾向性评分模型中应纳入的协变量类型。根据分组因素为二分类的倾向性评分匹配方法的基本思想，构建分组因素为有序三分类的倾向性评分匹配法，包括最邻近匹配法、卡钳匹配法和马氏距离匹配法，并通过SAS宏程序实现各种匹配方法。在不同匹配方法中设定不同匹配参数，如匹配比例、卡钳值等，通过比较不同匹配方法和设定不同匹配参数确定不同数据特征下最具优势的匹配方式。另外，还将利用模拟数据比较不同倾向性评分应用方法，包括匹配法、分层法、回归校正法和匹配后回归校正法。采用有序logistic回归分析法计算分组因素为有序三分类的研究对象的倾向性评分值。在倾向性评分匹配前后需要对放入倾向性评分模型中的协变量进行均衡性检验。本研究采用标准化差异法(standardized differences, SD)来评价不同组间协变量的均衡性。通过预实验得到，当分组因素为有序三分类时，，不同组间标准化差异的绝对值的最大值大于0.1时，三组间的协变量尚未达到均衡。当完成倾向性评分匹配后，还要对模型的偏性和精度进行评价。本研究采用相对偏倚(relative bias, RB)来评价模型的偏性，RB的绝对值越小，表明模型的偏性就越小；采用平均误差均方(mean squarederror, MSE)来评价模型的精度，MSE越小，表明模型的精度越高。最后，将模拟研究建立的分组因素为有序三分类的倾向性评分匹配方法应用到实例分析中。实例分析部分的数据来源于第二军医大学承担的“中国大陆胃肠道疾病流行病学调查”的数据。本研究利用问卷中调查对象的一般信息、体格检查问卷和SF-36健康调查问卷中的数据，评价腹部肥胖与健康相关的生活质量(health-related quality oflife, HRQOL)之间的关系。人口学信息包括性别、年龄、身高、体重、教育水平、职业和慢性病发病情况等。腹部特征定义为“正常腰围”、“轻度腹部肥胖”和“重度腹部肥胖”三类。健康相关的生活质量采用中文版的健康测量简表(SF-36)进行评价。以腹部特征为分组因素，健康相关的生活质量的各个维度得分为结局，筛选人口学信息中的变量为协变量，构建倾向性评分模型。利用模拟研究建立的倾向性评分匹配方法控制混杂因素对结局的影响，从而评价腹部肥胖对健康相关的生活质量的影响。研究结果：（1）协变量筛选：在分组因素为有序三分类的情况下，当倾向性评分模型中纳入与结局相关联的协变量时，可获得相对较高的匹配比例，并且估计的处理效应的偏性相对最小，精度最高。当逐步从模型中剔除一个协变量后，如果该协变量与分组因素和结局变量均相关联，会极大增加处理效应估计值的偏性，降低其精度，说明与分组因素和结局变量均相关联的协变量需全部纳入，同时再纳入与结局相关联但与分组因素不相关联的协变量可进一步减小处理效应估计的偏性，增大处理效应估计的精度。因此，在分组因素为有序三分类的情况下，倾向性评分模型中需纳入与结局相关联的协变量，无论其是否与分组因素相关联。（2）匹配方法构建和比较：本研究构建了分组因素为有序三分类的倾向性评分匹配方法，包括最邻近匹配法、卡钳匹配法和马氏距离法，并对不同匹配方法进行比较。在不同组间样本量比例下，卡钳匹配法的效果均达到最好。当组间样本量比例为1:1:1时，采用卡钳匹配法（卡钳值设为0.005）进行1:1:1匹配效果最好；当组间样本量比例为2:3:5时，采用卡钳匹配法（卡钳值设为0.01）进行1:1:1匹配效果最好；当组间样本量比例为1:2:3时，采用卡钳匹配法（卡钳值设为0.01）进行1:1:1匹配效果最好；组间样本量比例为1:4:5时，采用卡钳匹配法（卡钳值设为0.01）进行1:2:2匹配效果最好。（3）不同倾向性评分应用方法比较：不同倾向性评分方法均能极大地降低处理效应估计值的偏性，提高处理效应估计值的精度。无论组间样本量比例如何，匹配法和匹配后回归校正法的效果均优于其他方法。当组间样本量比例为1:1:1时，回归校正法优于分层法；当组间样本量的比例逐渐拉大时，分层法优于回归校正法。（4）实例研究：经倾向性评分匹配后，所有与结局相关联的协变量均在不同腹部特征组间达到了均衡，因此可以直接评价腹部肥胖对健康相关的生活质量的作用。结果表明，在体能维度上，重度腹部肥胖组的人群得分均显著低与正常腰围组，而轻度腹部肥胖组的人群得分显著高于正常腰围组。而在社会功能维度上，只有重度腹部肥胖组的人群在得分上显著低于正常腰围组人群，轻度腹部肥胖组人群与正常腰围组人群在得分上无统计学差别。研究结论：在分组因素为有序三分类的情况下，倾向性评分模型中应纳入与结局相关联的协变量。在进行倾向性评分匹配时，采用卡钳匹配法进行匹配效果最好，卡钳值和匹配比例根据组间样本量比例进行调整。在不同倾向性评分应用方法中，以匹配法和匹配后回归校正法的效果最好。与传统多因素统计方法相比，本研究建立的分组因素为有序三分类的倾向性评分匹配方法可通过控制混杂因素定量评价不同组间连续型结局变量的差异。
[Abstract]:Background of Study :

With the development of information technology , observational studies have been increasing and improving both in quantity and in research accuracy . The observational study of large samples plays a more and more important role in medical research .
What kind of matching method is more advantageous and has not yet been finalized ;
In addition , the current tendency score matching method is mainly used for observational study data of grouping factors into two categories , and few researches have been used in observational study data for grouping factors into multi - classification .

Purpose of study :

In this paper , we construct the matching method of propensity score in order three classification , and compare multiple matching methods under the condition of grouping factor into ordered three classification , and compare the best advantage in different data characteristics by adjusting the parameters , and then compare the application methods of different inclination scores under the condition of grouping factors as ordered three categories , and finally apply the optimal propensity score matching method established in the simulation study to the actual data analysis .

Study method :

In this study , the data set is simulated by Monte Carlo method . The grouping factors are modeled as ordered three categories , and the proportion of sample size between different groups is 1 : 1 : 1 , 2 : 3 : 5 , 1 : 2 : 3 and 1 : 4 : 5 .

By means of sequential logistic regression analysis , we calculated the tendency score value of the grouped factors into the ordered three categories . By pre - experiment , the equilibrium between different groups was evaluated by standardized differences ( SD ) . When grouping factors were ordered three categories , the covariables between the three groups had not yet reached equilibrium . When the tendency score was completed , the bias and accuracy of the model were evaluated . The smaller the absolute value of RB , the smaller the bias of the model was shown .
The smaller the mean squarederror ( MSE ) is used to evaluate the accuracy of the model , the smaller the MSE , the higher the accuracy of the model .

Finally , the relationship between obesity and health - related quality of life ( HRQOL ) was evaluated by using the data from the general information , physical examination questionnaire and SF - 36 health questionnaire . The data from the questionnaire included sex , age , height , weight , education level , occupational and chronic diseases . The health - related quality of life was defined as " normal waist circumference " , " mild abdominal obesity " and " severe abdominal obesity " .

Results of the study :

( 1 ) Covariate screening : In the case of grouping factors into an ordered three classification , a relatively high matching ratio can be obtained when the covariables associated with the outcome are included in the propensity score model , and the accuracy is the highest . If the covariables are associated with both the grouping factor and the outcome variable , the accuracy of the processing effect estimate can be greatly increased , and the covariables associated with the outcome variables and the outcome variables can be further reduced , so that the accuracy of the processing effect estimation is increased . Therefore , in the case of the grouping factors being ordered three categories , the covariables associated with the outcomes need to be included in the propensity score model regardless of whether or not it is associated with the grouping factor .

( 2 ) Construction and comparison of matching method : This study constructed the matching method of propensity score based on grouping factors as ordered three classification , including the most adjacent matching method , the caliper matching method and the Markov distance method . The effect of the caliper matching method is the best when the sample size ratio of the groups is 1 : 1 : 1 . When the sample size ratio is 1 : 1 : 1 , the matching effect of the caliper matching method is 1 : 1 : 1 .
When the sample size ratio of the group is 2 : 3 : 5 , the matching effect of 1 : 1 : 1 is best done by using the caliper matching method ( the caliper value is set to 0.01 ) .
When the sample size ratio of the group is 1 : 2 : 3 , the matching effect of 1 : 1 : 1 is best done by using the caliper matching method ( the caliper value is set to 0.01 ) .
When the sample size ratio between groups is 1 : 4 : 5 , the matching effect of 1 : 2 : 2 is the best by adopting the caliper matching method ( the caliper value is set to 0.01 ) .

( 3 ) Compared with other methods , the method of different propensity score can greatly reduce the deviation of treatment effect estimation value and improve the accuracy of treatment effect estimation value . The regression correction method is superior to other methods , regardless of the proportion of sample size , the matching method and the post - matching regression correction method .
When the proportion of sample size in the group gradually increases , the stratification method is superior to the regression correction method .

( 4 ) Case study : After the matching of the propensity score , all the covariables associated with the outcome were balanced among the different abdominal characteristic groups , so it was possible to directly evaluate the effect of abdominal obesity on the health - related quality of life . The results showed that the scores of the patients with severe abdominal obesity were significantly lower than those in the normal waist group .

Conclusions of the study :

In the case of grouping factors as ordered three classification , the covariables associated with the outcomes should be included in the propensity score model . The best results are compared with the traditional multi - factor statistical methods . The grouping factors established in this study are the best results compared with the traditional multi - factor statistical methods .

【学位授予单位】：第二军医大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：R181.2

【引证文献】