Study And Application Of Longitudinal Causal Mediation And Sensitivity Analysis Within Bayesian

With the development of the information age and advances in data science technology,the inference of causal relationships and influences among variables in data has been widely applied in fields such as biology,economics,and social sciences.Causal mediation analysis,based on causal inference research,revealed causal mediation mechanisms and promoted more effective intervention measures and policy-making.Longitudinal data causal mediation analysis had more application scenarios and greater adaptability compared to cross-sectional studies,providing more reliable and useful information for revealing causal mediation mechanisms among variables.The current research has addressed the issues of confounding variables and interactions in longitudinal causal mediation analysis.However,existing longitudinal causal mediation analysis methods still have some limitations:(1)different types of longitudinal causal mediation effects are incomplete,including direct,indirect,interaction and total effect;(2)The four types of unmeasured confounding assumptions in longitudinal causal mediation analysis are not fully considered,including the exposure-mediation relationship,exposure-outcome relationship,and mediation-outcome relationship affected by exposure variables.To address these issues,this paper proposes a method for estimation of longitudinal causal mediation effects and a sensitivity analysis method.The relevant content of the paper is summarized as follows:1)A Bayesian-based method for estimating longitudinal causal mediation effects was proposed to estimate different types of longitudinal causal mediation effects.First,a directed acyclic graph model for longitudinal causal mediation was built,.which combines the observed data values with the prior information of the model parameters to obtain the posterior distribution of the model parameters.Then,the maximum a posteriori method is used to estimate the model parameters from the posterior distribution.Finally,the estimation of different longitudinal causal mediation effects were obtained based on the counterfactual definition.Simulation results demonstrates that when the sample size is600,the bias obtained from the Bayesian regression model is smaller than the other three machine learning models,with the minimum bias degree being 0.001.The proposed longitudinal causal mediation analysis method in this paper is superior to the other three methods,and the accuracy can be improved by up to 24.4%.2)A Bayesian-based sensitivity analysis method was proposed to simultaneously consider the influence of four types of unobserved confounding assumptions by setting the prior distribution of bias parameters.First,a directed acyclic graph model about unmeasured time-varying confounding variables was constructed,and the joint probability density function was defined.Biased parameters were set to follow a uniform distribution.Then,the model parameters were estimated using Monte Carlo simulation methods.Finally,Finally,by using the directed acyclic graph model and simulating sampling from the original dataset,a copy of the dataset was generated.The longitudinal causal mediation effect estimation method proposed in this paper was used to obtain the estimation and confidence intervals for direct,indirect,interactive,and total effects.Simulation experiment results demonstrates that the confidence intervals obtained by this method can cover the true value,and the estimateion are closer to the true value.This method was applied to real data,revealing that the direct effect of obesity on blood glucose accounts for 95.7%,the indirect and interaction effects may be more susceptible to the influence of age and triglycerides.