复杂抽样Poisson回归分析方法及应用

摘要

Objective To examine the association of Body Mass Index (BMI) with the risk of fall injuries in subjects aged 45 years and over using 2010 Texas Behavioral Risk Factor Surveillance System (BRFSS) data,and to deliberately illustrate the necessity applying complex sample Poisson regression model to the survey data. Methods Association of BMI with the risk of fall injuries was examined using the 2010 Texas BRFSS data after adjustment for confounding factors. Results Findings from conventional Poisson regression significantly overestimated asso ciations of obesity, cardiovascular disease, and annual income less than $ 25000 with the risk of fall injuries; underestimate associations of general health fair to poor, unemployment, unmarried, and low education and the risk of fall injuries, their statistical significances changed lightly, but still kept significant; the association of age with the risk of fall injuries remained unchanged either in the association and its significant test compared to the NB model. When complex sample negative binomial regression was applied to the data,dispersion parameterwas 8. 15 ,standard error 1. 68,and 95％ CI for = 5. 44 - 12. 22, indicating that the variance of the observed count of fall injuries was roughly 8. 15 times the mean, and negative binomial model was more appropriate to the Texas BRFSS data. Conclusion Conventional Poisson regression model resulted in underestimation of variances and standard errors of survey estimates of model parameters, leading to spurious associations of obesity, cardiovascular disease, and annual income less than $ 25000 with the outcome; complex sample NB model adjusted parameters' estimates of the models and was more appropriate to multistage sampling design, such as Texas BRFSS data.%目的 本文通过检验45岁及以上的人群体重指数(BMI)与跌倒性伤害发生的联系,详尽地阐述了复杂抽样Poisson回归分析方法及应用的必要性.方法 本报告应用2010年美国德克萨斯州BRFSS数据,分析了跌倒性伤害与体重指数的联系强度.结果 普通Poisson回归分析高估了肥胖、患有心血管疾病、小于25000美元的年收入与跌倒性伤害的联系;低估了中等和差的健康状况、未被雇用、未婚和受教育程度低与跌倒性伤害的联系,但假设检验结果没有变;与复杂抽样负二项回归(NB)分析结果相比,年龄与跌倒性伤害的联系强度没有改变,其统计学检验也保持未变.数据拟合复杂抽样NB回归,离散参数α=8.15,标准误1.68,α的95％ CI=5.44-12.22,表明跌倒性伤害数据的方差是其均数的8.15倍,因此复杂抽样NB回归模型更适合BRFSS数据.结论 普通Poisson回归分析低估了参数估计的方差和标准误,造成了肥胖、患有心血管疾病、小于25000美元的年收入与跌倒性伤害的假关联关系;复杂抽样NB回归调整了模型的参数估计,适合多阶段抽样设计的数据分析.