Doubly Robust Estimation In Dynamic Treatment Rules

Dynamic treatment rule is a customized diagnosis and treatment decision based on individual heterogeneity data of patients.There is currently a large amount of scientific research exploring this field,with the vast majority focused on causal inference and optimization of diagnosis and treatment processes.Many decisionmaking optimization explorations focus on a single process of diagnosis and treatment.Alternatively,the overall optimization can be achieved by algebraic summation of feedback values from various stages throughout the dynamic diagnosis and treatment process.At the same time,most of the methods used for estimating the results are based on inverse probability models,but the estimated results are very sensitive to the results of the inverse probability model.Misassumptions in the inverse probability model can affect the consistency and asymptotic variance of the estimates.By adding an additional term that combines both the result model and the propensity score function model,a doubly robust estimation model is constructed.In the doubly robust model of dynamic diagnosis and treatment,the entire diagnosis and treatment process is regarded as various diagnosis and treatment stages divided by time stages.And incorporate the early diagnosis and treatment feedback results into the consideration of variables.When either the outcome model or the propensity score function model is correctly specified,the estimated results of the estimation model are consistent with the results under the correct setting of the original inverse probability model,and have a small variance.When the propensity score function model is correct,the variance of the estimation results of the doubly robust estimation model reaches the minimum under this estimation type.In numerical simulation experiments,it can be seen that the dual robust estimation model has better estimation performance than the inverse probability model.In empirical analysis,the estimation results of a doubly robust estimation model divided into multiple stages and incorporating early feedback results are superior to both the inverse probability model and the single process doubly robust estimation model that does not include early feedback results.