Application Of Propensity Score In Medical Research

BackgroundRCT (randomized clinical trial) is one of the most scientific study methods in clinical study. But sometimes because of the ethical and technical reasons, we can't perform the RCT. In this case, the selection of subjects may be biased due to certain covariates. As the development of statistical analysis, more and more new study methods involve and explore this problem. Among them, Propensity score analysis method is one of analysis method that being discussed a lot.Objective1. Using propensity score to analyse the data from clinical study and epidemiological study, and comparing with multivariable logistic regression analysis.2. Comparing propensity score analysis with multivariate logistic in terms of theory, conditions of application, odds ratios. Analysing the reasons of the differences and evaluating the strengths and limitationsMethodsUsing datasets of clinical study and epidemiology as an example, we compare and analyze the results of the propensity score and multivariable logisitc regression.Clinical follow up study: A follow up study about patients who accept treatment of ESWL in Shenzhen was iniatiated. We determine the depth of therapy as the independent factor and efficacy of ESWL as the dependent factor. Gender, age, period of disease, size of concretion, with or without prethrapy and so on were regarded as covariates. We performed multivariable logistic regression to get the odds ratios and P values of corresponding covariates. The method of propensity score was further used. First of all, with the depth of therapy to be dependent variable ,we put all the known variable into the propensity score model, using logistic regression to get the propensity score of each patients. And then we sorted out the propensity scores from lowest to highest, dividing them into five strata and calculating each odds ratio in each strata. We finally used it to get overall odds ratio and corresoponding 95% confidence intervals.Epidemiological study: Based on the collected materials and data about disposition of patients and efficacy of ESWL from certain community, we determine the intensity of smoking to be the independent factor; risk of gastric cancer to be the dependent factor; gender, age, education background , exercise habit and alcohol drinking to be the covariates. Then employing multivariate logistic regression to statistic,with the low intensity of smoking to be reference, analyze and get the odds ratios and P values of asscoiated covariates. Then employing propensity score to analyze the same materials and data. First of all , get the propensity score of every study patients. and then sort the propensity score from small to large, divide it into five strata.calculate each odds ratio of every strata, mean it to get overall odds ratio and associated 95% confidence intervals.ResultsIn clinical study of depth of thrapy and efficacy of ESWL. The results of multivariate logistic regression are as follows: the odds ratio of medium depth vs low depth is 1.50, the associated confidence interval is (0.54,4.17). while the odds ratio of high depth vs low depth is 0.54, the associated confidence interval is (0.13,2.27). The results of propensity score are as follows: the odds ratio of medium depth vs low depth is 1.398 the associated confidence interval is (0.60,3.24). while the odds ratio of high depth vs low depth is 0.686, the associated confidence interval is (0.17,2.82).In epidemiology of intensity of smoking and the risk of gastric cancer. The results of multivariate logistic regression are as follows: the odds ratio of medium intensity of smoking vs low is 1.42, the associated confidence interval is (0.54,4.17). while the odds ratio of high intensity vs low depth is 1.50, the associated confidence interval is (0.13,2.27). The results of propensity score are as follows: the odds ratio of medium depth vs low depth is 1.38 the associated confidence interval is (0.65,5.21). while the odds ratio of high depth vs low depth is 0.69, the associated confidence interval is (0.14,2.32)。DisscusionsWhen there are serious collinearities in the model, or the number of events per variable is less than seven. Propensity score will produce relatively accurate results. Otherwise, because of complex of propensity score calculation. I recommend multivariate logistic regression to be a regular method to deal with binary outcome.