Comparison Of Propensity Score Technique And Applied In Pharmacoeconomics

ObjectiveTo achieve comparable balance between groups,reduce selection bias,improve the reliability of statistical inference,and provide a new way of thinking and methodological support for statistical analysis of pharmacoeconomics evaluation by comparing non-randomized data with propensity score method based on logistic regression and generalized boosted modeling(GBM)and their empirical research in the field of pharmacoeconomics.MethodsWe constructed a selection on the observable model and compared the propensity score techniques which the propensity scores are estimatied by logistic regression model and GBM.Monte Carlo study was used to estimate the average treatment effect.The results were evaluated by the absolute standard deviation mean difference,matching sample proportion,estimation bias,relative bias and mean square error.At the same time,we applied the propensity score technique to the small-scale demonstration in the field of pharmacoeconomics.ResultsThe data under different sample sizes(N=150,300,600,1200,2400,4800)was simulated with multiple matching methods based on Logistic regression and GBM.The results are as follows:(1)When all covariates were observed and the regression equation was correct,the overall performance was the best with the nearest neighbor matching method within 0.25 ? caliper after using Logistic regression;(2)The overall balance of GBM was better than that of Logistic regression in the case of missing important covariates.The author suggested that the nearest neighbor matching method within 0.25 ? caliper of GBM should be preferred.(2)when the important covariates were omitted,the results would produce large deviations,for example,when the value of presupposed intervention effect was 0.25,the relative bias of each model was almost between 160~180%.(3)In the medication data of type 2 diabetes,GBM was better than Logistic regression in estimating the propensity score,The results of the net benefit regression showed that when the patient's maximum willingness to pay was more than 16947.5 CNY,it was suitable to choose the A scheme(insulin analogues);when the patient's maximum willingness to pay was less than 16947.5 CNY,it was suitable to choose the B scheme(human insulin).Conclusions(1)Logistic regression and GBM have their own advantages and disadvantages in estimating the propensity score.The "best" result only depends on the degree of fit between the assumptions contained in the model and the data generation process.In practical applications,We should choose specific methods based on different needs and the characteristics of data.(2)The application of propensity score method to evaluation the evaluation of pharmacoeconomics,to a certain extent,broadens the available data sources of the research on pharmacoeconomics,which can provide a new way of thinking and method for non-randomized evaluation of pharmacoeconomic research.