The Study Of Propensity Score Matching For Multiple Treatment Groups

RCT is the golden standard when comparing treatment effects. But it does not satisfy us for taking a long time and high cost to generate answers, confining by ethnics factors. Non-randomized studies occur frequently in epidemiology, post-marketing surveillance and medical device clinical studies.In non-randomized studies, investigators have no control over the treatment assignment. Therefore, large differences on observed covariates in the two groups may exist and could lead to biased estimates of treatment effects. The methods commonly used to reduce bias in non-randomized studies include regression model, matching and stratification. These methods may not be suitable for studies with a large number of covariates and rare events.Propensity score methods have attracted researchers'interests and are increasingly being used to reduce bias in the estimation of treatment effects in non-randomized and observational studies. The propensity score, introduced by Rosenbaum and Rubin in 1983, is defined as a subject's probability of receiving a specific treatment conditional on a group of observed covariates. As the representative of many covariates, it is estimated at baseline to control selection bias.Propensity score methods are mainly applied to two treatment groups rather than multiple treatment groups, because some key issues limited its application to multiple treatment groups remain unsolved, such as the choice of optimal caliper width, the assessment of balance in baseline variables, and sensitivity analysis.The primary objective of this study was to extend propensity score matching to multiple treatment groups. This article discussed and compared power and Type I error rates between propensity score and logistic regression, chosen the suitable method of balance test and discussed the necessity of balance test, further more compared propensity score matching methods using different calipers and chosen the optimal caliper width in the application of multiple treatment groups.The main tasks and results of this study are listed as following:1. This article discussed and compared power and Type I error rates between propensity score and logistic regression. Monte Carlo simulations were employed to compare the propensity score matching and logistic regression when they are used to analyze nominal data. Propensity score matching and logistic regression both can controll Type I error rates well. The nearest available neighbor matching propensity score has higher power than other methods. Propensity score matching should be recommended in practice of non-randomized studies.2. This article chosen suitable method of balance test and discussed the necessity of balance test. Monte Carlo simulations were employed to compare standardized difference and hypothesis test which were used to assess balance of baseline variables. In small size randomized trials, one is more likely to observe systematic differences among baseline variables between groups than in large scale randomized trials. Standardized difference should be better to be recommended in practice.3. This article compared propensity score matching methods using different calipers and chose the optimal caliper width for multiple treatment groups propensity score matching. The balance in baseline variables was assessed by standardized difference. The matching ratio, relative bias, and mean squared error (MSE) of the estimates between groups in different propensity score-matched samples were also reported. The results of Monte Carlo simulations indicate that matching using a caliper width of 0.2 of the pooled standard deviation of the logit of the propensity score affords superior performance in the estimation of treatment effects. This study provides some theoretical basis for the application of propensity score matching of three treatment groups.The achievements of this study consist in the following three points: Firstly, the article compared power and Type I error rates between propensity score and logistic regression and proved propensity score matching should be recommended in practice of non-randomized studies. Secondly, the article defined matching distance and extended standardized difference for multiple treatment groups. Finally, the results of Monte Carlo simulations indicate that matching using a caliper width of 0.2 of the pooled standard deviation of the logit of the propensity score affords superior performance in the estimation of treatment effects.