| RCT is optimal method without doubt to estimate average treatment effectbetween groups in a mass of observational data and clinical trial. RCT is not onlygolden standard, but also a base on statistical analysis of clinical trial. It’s invalidto analyze non-randomized data on statistical test, because randomization canbalance the groups to remove selective bias, then a true treatment effects can beestimated. In practice, RCT is limited because of high cost, a long time andethical constraints and so on. In Chinese traditional medicine, which therapy ischosen is according to patients’ conditions, so it’s hard to be randomized.Non-randomized studies were taken frequently in epidemiology, observationalstudies and clinical confirmation trials on medical appliance.In non-randomized studies, there were many confounding factors in groups,so treatment groups are different on observed covariates because of thoseconfounding factors. Therefore, a biased estimate of treatment average effects would exist. To reduce the bias, multiple variable models, Mantel-Haenszelstratification analysis were taken in studies. But these traditional methods are notsuitable for studies with rare events and a large number of covariates.Propensity score method was introduced as an effective method to removeplenty of bias. It was usually used in non-randomized studies and in manymedical fields, and be loved by many investigators. Propensity score, introducedby Rosenbaum and Rubin in1983, is defined as a subject’s probability ofreceiving a specific treatment conditional on a group of all observed covariates.As the representative of many covariates, it is estimated at baseline to controlselection bias and estimate the true treatment effect.At the present time, every subject’s PS was estimated by logistic regressionmodel, then, PSM with caliper was taken to deal with data. Logistic regressionmodel is a non-probability linear model, and have many advantages of estimatethe subject’s PS. PSM with caliper as a simple mature method was often used inmany aspects. The optimal caliper was calculated by Austin. PSM with caliperestimates subject’s PS by point estimate of parameter. If every subject’s PS wasestimated by95%confidence interval of parameter, then according these PSmatching was taken. How was the result? The issues were studies in this paper.The study introduced the concept and idea of PSIM and how to calculatePSIM for the first time. The optimal caliper interval was calculated by MonteCarlo simulations. In this paper, logistic regression model also was used toestimate every subject’s PS. Then PSM was taken by95%confidence interval ofPS. This article discussed the capability of three methods including propensityscore interval matching, propensity score matching with caliper and logisticregression in dealing with the non-randomized data. Feasibility and practicabilityof PSIM was validated. The main comments of this article and results of this study are listed as following:1. Monte Carlo simulations were used to explore and study the propensityscore interval matching, furthermore to compare propensity score intervalmatching methods with different calipers and choose the optimal caliper intervalin the application of propensity score interval matching. When the caliperintervals were70%,80%,85%and90%of overlap with PS interval betweengroups in week correlation model and strong correlation model, statistical power,type I error rate, standardized difference and matching proportion were used toevaluate PS interval matching, PSM with caliper and logistic regression model.The result suggests when caliper interval was80%, the above four index responsewell.2. This article discussed the capacity of balancing the covariates among thePS interval matching, PS matching with caliper and logistic regression model bystatistical power, type I error rate, standardized difference and matchingproportion. The results suggest that these three methods balance the covariatesbetween groups well. In the week correlation model, PS interval matching andPSM with caliper can balance all the covariates which are related to resultvariable. Standardized difference can show this result. But in strong correlationmodel, only PS interval matching can balance all the covariates. In matchingproportion, PS interval matching has done a little better than PSM with caliper.3. This article discussed the capacity of balancing the covariates which aretri-categorical variable data and continuous variable between PS intervalmatching and PSM with caliper by Monte Carlo simulation. The simulationsresults show that there is no difference between PS interval matching and PSMwith caliper to balance tri-categorical and continuous covariates. And they bothhave done well. 4. According to Efficacy and Safety of Firebird Cobalt-Chromium AlloyedSirolimus-Eluting Stent in Treatment of Complex Coronary Lesions in Diabetesclinical trial, PS interval matching was used to match Firebird I clinical trial dataand Firebird II clinical trial data, then estimate the difference between Firebird Iand Firebird II. Eight variables were included the models to estimate the PS. Theresult shows: eight variables included in the models were balanced by PS intervalmatching. And the true average treatment effect was estimated reliably. Thisresult also confirmed that PS interval matching is a valid and feasibility methodwhich deals with the non-randomized data.The advantages of this method study consist in the following three points:First of all, PSIM was produced and the optimal caliper interval-80%wasintroduced by simulations with four conditions. Secondly, four indexes includingstatistical power, type I error rate, standardized difference and matchingproportion were applied to evaluate three methods with different data. Finally,Efficacy and Safety of Firebird Cobalt-Chromium Alloyed Sirolimus-ElutingStent in Treatment of Complex Coronary Lesions in Diabetes clinical trial wasadopted to evaluate the PS interval matching method and has validated thefeasibility and practicability of PSIM. |
PDF Full Text Request