The Study Of Model Uncertainty And Its Application In Benchmark Dose

The estimate of Benchmark dose based on the dose-response model is the core part of a quantitative risk assessment.Due to the limited sample data and the limited knowledge on true dose-response function,model uncertainty in dose-response study cannot be omitted,especially when data are collected from population.Recently,Bayesian methods have been paid more attention on uncertainty of benchmark dose estimation.Bayesian model average method takes the uncertainty of independent variables and the model equation into account,and gets the weighted estimation by combining the alternative models,provides a new way to explore the application strategy and performance of BMA compared to classic optimal model method.This study provides methodological support for BMD model choice method and calculation complement by computer.Results:(1)The epidemiological population data analysisa.The best fit model of epidemiological population data of Gansu was log-probit model,and the BMD/BMDL of log-probit model was 3.46/2.69 ug/g cr.BMD/BMDL estimated by BIC approximation based BMA were 2.94/2.14 ug/g cr.The MCMC based BMA method's BMD/BMDL was 2.92/2.07 ug/g Crb.The best fit model of epidemiological population data of Guangzhou was log-probit model,and the BMD/BMDL of log-probit model was 14.24/10.38 ug/g cr.BMD/BMDL estimated by BIC approximation based BMA were 14.37/9.86 ug/g cr.The MCMC based BMA method's BMD/BMDL was 14.39/9.84 ug/g Cr(2)Simulation studya.The percentages that models identified by BIC based on optimal model method were different from different true dose-response models.This percentage was 45.4%± 9.9%and 14%± 1.4%when simulated dose-response curve was set as the log-probit model and the weibull model respectively.b.The relative bias of estimation and simulation's benchmark dose was calculated in this study.In single model simulations,relative bias from the optimal model method was less than Bayesian model average methods.In the mixed model with equal weight simulation,Bayesian model average method was preferable to AIC's method.For the two kinds of Bayesian model average methods,relative bias of MCMC estimation was slightly lower than BIC method in mixed model simulation and slightly higher in single model simulations.c.The ? parameter in Weibull model was use as an indicator to detect the convergence of all model convergence,two statistics was used to show the model convergence performance,the Geweke statistic of greater than 2 and the 97.5%of shrink factor with greater than 1.15 were consider as not convergence.MCMC convergence diagnostic results showed that the model parameters had not reached the convergence when umber of dose group and sample size were small;With the number of dose group and sample size increasing,the chains of parameters could get better convergence.Conclusion:(1)When the dose-response model type of epidemiological population data was one of two-stage model,log-logistic model,Weibull model,Log-probit model or gamma model,true dose-response could not be selected by optimal model selection merely based on AIC.(2)The epidemiological population data contain numerous confounding factors,which could not ignore the model uncertainty.Compared with the optimal model method,the Bayesian model average method taken the model uncertainty into account and got more reliable BMDL.BMA was a better choice for the population with large variation.(3)Bayesian model average method based on MCMC was a better choice if computing burden is not the main concern.