A Study Of The Caliper Matching Method For Three-group Data

It is helpful to evaluate treatment outcomes in non-randomized research using data from non-randomized clinical trials.Selectivity bias,however,might result in issues with covariate distribution that are uneven across groups,which can generate significant errors in determining treatment effects.The propensity score method,which is mostly used with treatment dichotomous data,is a superior way to control for selectivity bias.The challenges of caliper value selection and equilibrium test in the caliper matching method are not properly solved in the case of three-group data,making it impossible to achieve improved matching quality.This will serve as the foundation for this study’s investigation of the issue of choosing the best caliper values for sorted and unordered three-partitioned data.The specific research work is as follows:(1)Investigate the best caliper values for ordered three-group data.It is assumed that the sample ratios for the various groups are 1:1:1,3:3:4,1:2:7,1:2:3,2:3:5,1:4:5,and 1:1:2,respectively.Set the step size to 2.5%or 5%and the search range of the caliper values to be 5%～95%of the total standard deviation of the propensity scores after logit transformation.According to the simulation and case study results,in the scenario where the grouping variables are ordered trichotomous data,the optimal caliper value for the caliper matching method should be set to 22.5 percent of the combined standard deviation of the propensity scores after logit transformation.(2)Optimize to get the best caliper values for unordered three-group data.The sample ratios between groups are assumed to be 1:1:1,1:2:3,1:2:7,2:3:5,and 1:4:5,respectively.The search range of the caliper value is set to be 5%sim95%of the combined standard deviation of the propensity score after logit transformation,and the step size is 5%.In the case of unordered trichotomous data,the results of the simulation and case study indicated that the optimal caliper value for the caliper matching method should be set to 25%of the combined standard deviation of propensity scores transformed by logit.This improved matching quality over the previous method,which was set to 20%of the combined standard deviation of propensity scores transformed by logit.