| Objective:In exploring the long-term effects of an individual’s food intake on health outcomes,researchers are faced with challenge of the time-varying confounders posed by the variability of an individual’s food intake over time,and with problem of collinearity caused by the compositional nature of individual’s food intake.The key point to choose a proper method to control time-varying confounders in observational studies is to define and identify the different types of time-varying confounding correctly.Meanwhile,the key point to analysis dietary data with compositional nature is conversion from Acheson space to Euclidean space accurately.For controlling the time-varying confounders,and estimating substitution cause effects of dietary exposure with characteristics of composition on health outcome,this study built a statistical strategy to identify different time-varying confounders in observational studies,and using the inverse probability weighted algorithm based on covariance balance and residual confounding balance to control time-varying exposures when time-varying exposures are continuous variables.Combination of the compositional data coordinate transformation method with traditional substitution analysis,then fitting the marginal structural model with transformation coordinates and health outcomes,and estimating the causal effects.Our aim is to provide a statistical analysis strategy and method for causal substitution inference of dietary data with time-varying and compositional characteristics in longitudinal studies of nutrition epidemiology,so as to provide feasible and personalized dietary guidance,and to improve the health level of the population under the reality of changeable dietary behavior.Methods:In the first part,we simulated the different types of time-varying confounders,and compared the estimation accuracy of traditional regression correction approach and marginal structure model based on inverse probability weighting.According to the simulation results,we proposed the necessity of diagnosing time-varying confounders types.Then,we proposed a method for correctly identifying different types of time-varying confounders in observational studies,which was a strategy for further selecting statistical methods to avoid confounding caused by different types of time-varying confounders.According to the rule of exchangeable and the rule of the backdoor criterion,we further expanded the diagnostic methods to identify time-varying confounders and time-varying exposure-confounders feedback with dichotomous exposure,polytomous exposure and continuous exposure by using standardized mean difference,FQ statistics,and weighted correlation based on the generalized propensity score.When the time-varying exposure-confounders feedback was found,we should use G methods to control that confounders.We summarized the existing G methods and their advantages and disadvantages in theoretical methods and applications.Since G methods has rarely been extended to the situation that the time-varying exposure is continuous variable.We expanded the covariance balanced generalized propensity score(CBGPS)methods with marginal structure model(CBGPS-MSM),by constructing CBGPSs in multiple follow-up time points and multiplying the weights based on CBGPS of multiple follow-up time.The weights got from CBGPS were based on covariance balancing,in addition,there was another method like CBGPS,which was balancing the covariance according to the residual confounding(RBW-MSM).We compared these two methods of the estimation accuracy and ability of balancing confounders.In the second part,we introduced the theoretical basis on the substitution analysis,clarified the key problems of the traditional substitution analysis,explained the compositional properties of the diet data.We further proposed to combine the traditional substitution analysis with ilr coordinate transformation from Simplex space to Euclidean space,and presented an approach of compositional transformation substitution analysis(Co TSA).Based on a large prospective cohort study of China Health and Nutrition Survey(CHNS),we used method of traditional substitution analysis and Co TSA to evaluate the effects of animal-based foods substitution on the incidence of hypertension in adult,and further to clarify the advantages of Co TSA in dietary substitution analysis.In the third part,we combined the methods in Part I and Part II to get the causal substitutional effect of foods intake on health outcomes based on the survey of CHNS.After describing trends in animal-based foods intake among the study population,we identified the time-varying confounders,and the time-varying exposure-confounders feedback.Then we used the extended G methods to estimate the causal coefficient between ilr coordinates constructed by multiple independent time-varying continuous exposures,and used the Co TSA analysis to estimate the causal effect between different substitution patterns of animal-based foods and incidence of hypertension.Results:The first part:(1)simulating the different situation with three different types of time-varying confounders respectively.We used the traditional regression correction and marginal structure model(MSM)based on inverse probability weighting(IPW)to estimate the causal effect in different situation with the different sample size,mis-specified model,respectively.Firstly,when the exposure did not change with time but the confounders change with time,we called it "pseudo time-varying confounding",we found that compared with the IPW-MSM,the estimation values obtained by using traditional regression correction method were more accuracy.Secondly,when both exposure and confounding were time-varying,but confounders in the follow-up period was not affected by the previous exposure,which mean that there was no time-varying exposure-confounder feedback,the precision of the traditional method was higher.However,the accuracy of IPW-MSM were higher when time-varying exposure-confounder feedback exist.(2)simulating the different situation with three different types of time-varying confounders,as well as the situation that covariate caused feedback was unmeasured.We introduced two categories of diagnostic approaches to identify the different kinds of confounders in prospective cohort study.With the different settings of the sample size,right censoring,we found that the first diagnostic method can effectively identify time-varying confounding.The second diagnostic method can be further used to identify time-varying expose-confounder feedback.The weighted Polyseria correlation coefficient was further used to identify time-varying confounder and feedback when exposure is continuous variable.When there were endogenous censoring variables which were relative to exposure and covariances,the correct identification of time-varying confounders and feedback will be affected.(3)we compared the CBGPS-MSM method with the RBW-MSM method in different settings of simulation.The estimation accuracy and the ability to balance confounders of the two methods were compared under the conditions of different confounder degrees,the model mis-specific and the different sample sizes.The accuracy of RBW-MSM was as accurate as the CBGPS-MSM method.The ability of CBGPS-MSM method to balance confounders was better than that of RBW-MSM method.The second part:This part explored the relationship between the different animal-based foods substitution models the risk of hypertension and compared the results of Co TSA and TSA.Compared with TSA,Co TSA could avoid the problem of multicollinearity caused by correlations,and is not limited to one-to-one substitution.Co TSA estimates substitution effects based on different substitution patterns,and provides flexible and personalized substitution strategies tailored to individual food intake patterns.According to the COTSA analysis,the substitution of red and processed meat for other all animal-based foods was significantly associated with an increased risk of hypertension(HR=1.062,95%CI:1.050-1.074).The third part:Based on the prospective cohort study of CHNS,we found that the red and processed meat consumption was increased in 2011 compared to the baseline survey in2004-2009.Of the 5,394 patients,1,267 were new hypertension cases,so that the incidence of hypertension was 25.8% in the study population.We diagnosed time-varying confounders and time-varying exposure-confounder feedback in the covariances,and we found that BMI,time spending on sitting and sleeping,and total energy intake were time-varying confounders,but none of them formed the time-varying exposure-confounder feedback.Therefore,both of traditional regression and RBW-MSM methods can be used to estimate the causal effect of ilr coordinates on risk of hypertension(βreg=0.0166 vs.βRBW=0.0196),then we used the causal coefficients to get causal substitutional effect by method of Co TSA.We found the consistent conclusion that intaking red and processed meat instead of other all animal-based food can increase risk of hypertension(HR=1.017,95%CI : 1.002-1.032 vs.HR=1.020,95%CI :1.005-1.035).Conclusion:This paper defined the different types of time-varying confounders,and discussed the efficiency of controlling the different time-varying confounders with the traditional regression method and G method.We emphasized that there was no need to using G method in the prospective cohort study because the G method was mainly to avoid overcorrection and selection bias caused by time-varying exposure-confounders feedback.Further,based on the Interchangeability condition and backdoor criterion,we improved the method which was used to evaluate covariance balance and expanded this method to identify the time-varying confounders and time-varying exposure-confounder feedback within different kind of exposure.In order to provide a reference for researchers to select appropriate statistical method for controlling confounders in prospective cohort study.When the time-varying exposure was continuous variance,we used methods of CBGPS-MSM and RBW-MSM to controlling the time-varying exposure-confounder feedback.In addition,we proposed the Co TSA method,which combined the ilr coordinate transformation of compositional data with the concept of traditional substitution analysis,to realize the substitution analysis of dietary data with compositional property.We combined the strategies and methods for analyzing time-varying confounders in prospective cohort study and the Co TSA method,to explore the causal substitution relationship between the animal-based foods and the risk of hypertension.In conclusion,it is recommended to reduce the intake of red and processed meat with other animal-based food in order to prevent hypertension in our study population. |
PDF Full Text Request