复杂调查数据似然估计的权数检验及改进

Weight Tests and Improvement of Likelihood Estimation for Complex Survey Data

微观调查数据是微观数据的重要组成,微观调查数据的权数在最小二乘估计下的统计推断,已有多篇文献提出了不同的检验方法,但在最大似然估计下的统计推断,权数的检验还无明确的方法.依据常见的三种似然检验方法,本文给出了Wald权数检验、似然比权数检验和拉格朗日乘子权数检验,理论与模拟的结果发现,这三种似然权数检验均具有较高的犯第一类错误概率,本文利用Bartlett-type correction,获得修正得分权数检验,并从理论和模拟角度说明了获得的修正得分权数检验的优势,同时将这四种权数检验方法应用到中国家庭追踪调查(CFPS)数据,最后对该方法进行了总结和展望.

Survey data is an important component of micro-data.There have been many papers that have proposed different test methods for the statistical inference of survey weight under the least squares estimation.However,there is no clear method for the test of survey weight under the maximum likelihood estimation.We give the Wald test,likelihood ratio test and Lagrange multiplier test for survey weight,basis of the three common likelihood test method.The theoretical and simulation results show that the three likelihood weight test has a higher probability of the Type I error.We use Bartlett-type correction to obtain the modified score weight test,and explain the advantages of the obtained modified score weight test from the theoretical and simulation perspectives.The weight test method is applied to the China Family Panel Studies(CFPS)data,and finally the method is summarized and prospected.