Estimation of Average Treatment Effect Using Bayesian Additive Regression Tree in Observational Study / 中国卫生统计

ABSTRACT

Objective To explore the statistical performance and applicable conditions of Bayesian additive regression tree(BART)for estimating average treatment effect in observational studies.Methods The difference of estimates between BART and multivariate regression,propensity score matching,and inverse probability weighting through simulations and actual epidemiological data was compared.Results The results of these simulations showed that under the linear assumption,the performance of BART was close to that of the commonly used methods;when the relationship among variables in the data was complex and non-linear,BART performed markedly better than the others.When the ignorability assumption was not satisfied and there was unobserved confounding,all methods performed worse,but BART was still significantly better than the others and relatively robust.In the actual epidemiological data,this method was used to estimate the average treatment effect of smoking cessation on weight change.Conclusion In most observational studies,outcomes are influenced by multiple factors,making it difficult for researchers to properly specify relationships between variables.It is difficult to identify all these variables or determine the relationship between them.In terms of model fitting and result accuracy,BART is worth recommending.