基于混合效应模型的纵向暴露数据方差-协方差结构的模拟研究

Simulation of Longitudinal Exposure Data with Variance-Covariance Structures based on Mixed Models

目的探讨基于混合效应模型对短期纵向数据进行模拟,从而获得长期纵向数据这一方法的效果。方法以金属镉的膳食暴露数据作为应用实例来介绍该方法。首先计算某省成人的镉短期暴露量,在此基础上构建三个不同的模型并计算相应参数,包括固定效应参数、随机效应参数和一阶自相关系数。通过Monte Carlo模拟得到不同模拟人数和天数下人群的每日暴露量及其分布。三个标准用于评价模拟数据与原始数据的相似度:总平均值和标准差、方差成分比例和自相关系数。结果固定效应参数:年份、地区及两者的一级交互作用被纳入混合效应模型;随机效应参数:个体间方差σ^(2)_(b)=0.056,个体内方差σ^(2)_(e)=0.717,一阶自相关系数AR(1)ρ=-0.026;模型选择:在同样的模拟人数和天数下,模拟结果显示含有自相关结构的混合效应模型模拟的数据结构与原始数据结构最接近;考虑自相关的混合效应模型的模拟结果:模拟人数固定为1000人,模拟天数从30天递增至360天,随着时间长度逐渐增加,个体内与个体间方差成分比例以及自相关系数都很稳定,且接近原始数据的结果。结论基于混合效应模型的纵向暴露数据模拟方法可获得与原始数据的方差-协方差、自相关结构一致的长期数据。

Objective To discuss the effect of the method of simulation for obtaining long-term longitudinal data using short-term longitudinal data based on mixed models.Methods Dietary exposure data of cadmium were used as an application instance to introduce this method.The short-term cadmium exposure data of adults in one province were calculated firstly.Three different models were constructed to calculate parameters related to the data structure characteristics,including fixed effect parameters,random effect parameters and lag-one autocorrelation coefficients.Based on the estimated model parameters,long-term daily personal exposure and its distribution were obtained by Monte Carlo simulation with flexible numbers of subjects and days.There are three standards helping examine the similarity between the simulated data and the short-term data:overall mean and standard deviation,variance components percentages,autocorrelation coefficient.Results Fixed effect term:year,region and their first-order interaction were included in the model.Random effect term:the inter-subject and intra-subject variances were 0.056 and 0.717 respectively,and the lag-one autocorrelation coefficient was-0.026.Model selection:in the same condition of simulated number of subjects and days,the results show that the simulated data structure of the mixed model considering autocorrelation is the closest one to the structure of the observed data.Simulation results:when the number of subjects was fixed at 1000 and the number of days ranged from 30 to 360,component percentages of the inter-subject and intra-subject variances,lag-one autocorrelation coefficients were stable as the time length increased,and also close to the results of the observed data.Conclusion This simulation method based on mixed models can generate long-term data which has similar variance-covariance structures and autocorrelation structures compared to the observed longitudinal data.