多水平模型在土壤镉污染与人群尿镉水平关系研究中的应用

Application of multi-level model in study of relationship between soil cadmium contamination and people＇s urinary cadmium level

目的 基于层次结构数据的特点,运用多水平模型分析中部某县人群尿镉水平可能的影响因素,探讨多水平模型实际应用中处理层次结构数据的优势.方法 2013年5月,以整群抽样的方法在中部某县12个行政村抽取年龄≥20岁居民1 460名,进行尿镉水平可能影响因素调查和尿镉水平检测,并对调查区农田土壤特征污染物镉含量进行测定.完成调查并符合纳入标准的对象共计1 410名,调查区农田土壤检测样品共计318份.根据数据特点,将个体看作水平1单位,调查村看作水平2单位,运用SAS 9.3软件中处理层次结构数据的MIXED过程对数据进行多水平分析.在不考虑数据层次结构的情况下,运用SAS 9.3软件拟合一般线性模型,比较多水平模型及一般线性模型的拟合效果.结果 本研究最终纳入1 410名研究对象,年龄（55.2±11.1）岁,其中男性645名（45.74％）,女性765名（54.26％）.家庭人均每年食用大米量（100.9 ±40.3） kg.有从事采选矿工作经历的占18.65％（262/1 410）.尿镉水平为（9.39±2.16） μg/g肌酐.土壤镉普遍超标.一般线性模型对数据拟合结果显示,除了是否从事采选矿工作差异无统计学意义（χ2=1.05,P=0.305）以外,村土壤镉水平、年龄、食用自产大米量、性别差异均有统计学意义（χ2分别为401.39、34.9、4.16、86.15,P值分别为＜0.01、＜0.01、0.041、＜0.01）.空模型拟合结果显示本研究ICC =0.435 5,尿镉在村水平上存在聚集性.多水平模型全模型分析结果显示解释变量村土壤镉水平、年龄、食用自产大米量、性别差异有统计学意义（Wald χ2值分别为2.55、6.34、2.37、10.32,P值分别为0.029、＜0.01、0.018、＜0.01）,是否从事采选矿工作差异无统计学意义（χ2=0.78,P=0.438）.用于进行模型比较的拟合优度指标,多水平模型拟合结果均小于一般线性模型.水平2解释变量村土壤镉水平的回归系数为0.84,可解释的变异约占总变异的35.26％.结论 多水平模型分析层次结构数据较一般线性模型更加合理;村土壤镉水平对人群尿镉水平影响较大.

Objective Based on the characteristics of hierarchical data, a multilevel model was used to analysis possible influencing factors of urinary cadmium levels in one county population, and to discuss the advantages of multilevel model for processing hierarchical data in practical problems. Methods In May 2013, 1 460 participants aged 20 and above in 12 administrative villages in one county in central China were recruited by cluster sampling. Urinary cadmium level and its possible influencing factors were investigated, and cadmium level in farmland soil of survey area was also tested. A total of 1 410 participants completed the survey and met the inclusion criterion. 318 farmland samples in survey area were detected. According to the data, individuals were set as the level one unit, and the village was set as level two unit. the data were analyzed by MIXED procedure for hierarchical data of SAS 9. 3 software. In the case of not considering the hierarchy of data, the general linear model was fitted by SAS 9. 3 software, and the fitting results of the two models were compared. Results A total of 1 410 participants were included finally, the age was （ 55.2±11. 1 ） years. 645 （45.74%） were males and 765 （54. 26% ） were females. The amount of household per capita consumption of rice was （ 100. 9 --40. 3） kg/y. All 18.65% （262/1 410） of the participants had mining and mineral separation work experience. The urinary cadmium level was （9.39 ± 2. 16） μg/g Cr. Most of the soil cadmium levels in villages were greater than tolerance value. The fitting resuhs of general linear model suggested that whether doing mining and mineral separation work does not have significant dilterence （ χ2 = 1.05, P = 0. 305）. There was significant difference in the village soil cadmium levels, age, the amount of household per capita consumption of rice, and gender （ χ2 = 401.39, 34. 9, 4. 16 and 86. 15, respectively, P 〈0. 01, 〈0. 01, 0. 041, 〈0. 01, respectively）. The fitting result of empty model showed the ICC was 0. 435 5, the urinary cadmium had clustering at village level. The results of multilevel mode] showed that the explanatory variables of the village soil cadmium levels, age, the amount of househoht per capita consumption of rice and gender had significant difference （ Wald χ2 values 2.55,6. 34,2.37 and 10. 32, respectively, P =0. 029, 〈0. 01, =0. 018 and 〈0. 01 ）, while whether doing mining and mineral separation work had no significant difference （χ2 = 0. 78, P = 0. 438）. To the fitting optimization index using for the comparison of models, the results of multilevel model were less than that of general linear model. The regression coefficient of level-2 explanatol7 variable （the village soil cadmium levels） was 0. 84, which could explain the 35.26% of the total variance. Conclusion Multilevel model could analyze hierarchical data more reasonably than general linear model. Urinarv cadmium levels is highly influenced by the village soil cadmium levels.