基于设计与模型的总体参数估计及其在抽样调查中的应用  

Parameter estimation of design-based and model-based methods and their application in sampling survey

作　　者：孙霖[1] 王欣[1] 武留信[2] 徐勇勇[1] 

机构地区：[1]第四军医大学卫生统计学教研室,陕西西安710032 [2]中国人民解放军空军航空医学研究所

出　　处：《实用预防医学》2015年第3期257-261,共5页Practical Preventive Medicine

基　　金："十二五"国家科技支撑计划重点项目(2013BAI04B01)

摘　　要：目的对两种常用的统计推断体系应用于同一问题的结果进行比较,同时反映陕西省年人均住院医疗费用支出情况。方法第五次全国卫生服务调查采用分层多阶段πPS抽样方法,将两种统计推断体系分别应用于陕西省调查数据。结果基于设计的统计推断方法的标准误显著大于基于模型的推断方法,对经过对数变换之后的总体均值的估计二者分别为(4.060840±0.008588)和(4.060051±0.004072)。可见,基于设计的统计推断方法标准误显著大于基于模型的方法。结论两种统计推断体系应用于复杂抽样的大样本数据中各有优缺点,基于设计的统计推断要求一定的抽样比例以保证样本的代表性,且要求除样本数据外的大量辅助信息计算抽样权重;基于模型的统计推断对因变量的总体分布较为敏感,在总体呈现偏态时要进行转换,增加了模型的拟合难度。对于分类自变量要生成哑变量纳入模型,增加了模型的复杂程度。因此要针对两种推断体系各自的优势与不足,以及自身需要选择最适宜的统计推断方法。Objective To compare the results of two kinds of commonly-used statistical inference systems applied to the same problem,and to reflect the average annual medical expense of hospitalization expenditure in Shaanxi Province. Methods Stratified multi-stage πPS sampling method was used in the Fifth National Health Service Survey,and the two kinds of statistical inference systems were applied to the survey data of Shaanxi Province. Results The standard error of design-based statistical inference method was significantly larger than that of the model-based method. The estimations of population mean after logarithmic transformation were（ 4. 060840 ± 0. 008588） and（ 4. 060051 ± 0. 004072）,respectively. It was obvious that the standard error of design-based statistical inference method was significantly larger than that of the model-based method. Conclusions There are advantages and disadvantages in the process of two kinds of statistical inference systems applied in the complex sampling of large-scale data. The design-based statistical inference method requires relatively larger sampling ratio to ensure the representativeness of the sample,and also requires a lot of auxiliary information to calculate sampling weight. The model-based statistical inference method is more sensitive to the overall distribution of variables and transformation is needed when the distribution of population is skewed,which increases the difficulty of model fitting. For the classification variables,it needs to generate dummy variables into the model,which increases the complexity of the model. Therefore,we should aim at the advantages and disadvantages of the two kinds of statistical inference systems as well as our own needs to select the most suitable statistical inference method.

关 键 词：统计推断 复杂抽样 住院费用 基于模型 基于设计 

分 类 号：R195.1[医药卫生—卫生统计学]