小样本非正态数据结构方程模型估计方法研究与医学应用

发布时间：2018-04-03 08:07

本文选题：结构方程模型　切入点：非正态　出处：《山西医科大学》2007年硕士论文

【摘要】： 当前结构方程模型(structural equation modeling,SEM)已经广泛应用于医学研究中因果关系和先验假设理论的检验,是一种多功能和方便使用的分析技术。通常使用默认的ML估计方法。多元正态分布和大样本是应用ML估计的两个关键假定。然而,应用实践中收集的数据明显地违反了正态分布假定,又没有足够大的样本来使用新发展的任意分布估计方法。迫切需要在小样本且数据分布非正态条件下,仍能够准确地检验模型参数和评价模型拟合的方法。本文介绍两种估计非正态、较小样本SEM的方法:S-B调整方法(Satorra-Bentler scaled,S-B)和自助抽样方法(bootstrap resampling)。可以分别使用SEM估计的软件包EQS和AMOS得到。采用了蒙特卡罗模拟(monte carlo simulation)研究了三个样本大小(100,250,500)和三种多元分布(多元正态,轻度和严重偏离正态类型)条件下,模型正确指定时候这两种方法和常用的ML估计的性能比较。另外,将研究的两种较稳健的估计方法应用到医护人员职业紧张SEM模型估计的实例中,得到的结论与正态理论估计方法的不一致,说明了忽视数据条件而使用常规的正态理论估计方法的偏差。模拟研究表明,对于适当指定的模型,ML只有在多元正态条件下估计准确,随着数据偏离正态,它估计的模型检验统计量偏高,参数的标准误偏低,假设检验的Ⅰ型误差率升高,对模型和参数的评价出现歪曲。S-B调整方法除了在最小样本条件外较稳健,而自助方法在各条件下估计的结果都较理想。自助抽样的次数(B=250和500)对于模型估计没有大的差异。通过当前的研究,使得结构方程模型的应用能在一定程度上突破样本量和数据分布类型的约束,指导医学应用研究者正确使用结构方程模型,并为揭示医学中的因果机制和验证事先架构的理论假设提供依据。
[Abstract]:Structural equation modeling (SEM) has been widely used in the testing of causality and priori hypothesis theory in medical research. It is a multifunctional and convenient analytical technique.The default ML estimation method is usually used.Multivariate normal distribution and large samples are the two key assumptions in the application of ML estimation.However, the data collected in practice obviously violate the assumption of normal distribution, and there are not enough samples to use the newly developed arbitrary distribution estimation method.It is urgent to test the parameters of the model and evaluate the fitting method of the model accurately under the condition of small sample and non-normal data distribution.In this paper, we introduce two methods of estimating non-normal distribution, one is the small sample SEM, the other is the small sample SEM, and the other is the small sample SEM. The two methods are: Satorra-Bentler scaledS-B) and bootstrap sampling method bootstrap amplification.It can be obtained by using EQS and AMOS software packages estimated by SEM, respectively.Monte Carlo simulation of monte carlo is used to study the performance of three sample sizes (100250500) and three multivariate distributions (multivariate normality, mild or severe deviation from normal type). When the model is correctly specified, the performance of these two methods is compared with that of the commonly used ML estimators.In addition, two more robust estimation methods are applied to the SEM model estimation of occupational stress in medical and nursing personnel, and the results are inconsistent with the normal theory estimation method.In this paper, the deviation of the conventional normal theory estimation method is explained by ignoring the data condition.The simulation results show that the model ML can only be estimated accurately under the condition of multivariate normality. With the deviation of the data from the normal state, the estimated statistical quantity of the model test is higher, the standard error of the parameters is lower, and the type I error rate of the hypothesis test increases.The evaluation of the model and parameters is distorted. S-B adjustment method is more robust than the minimum sample condition, while the self-help method estimates the results well under each condition.The frequency of self-help sampling is not significantly different from that of model estimation.Through the current research, the application of structural equation model can break through the constraints of sample size and data distribution to a certain extent, and instruct medical application researchers to use the structural equation model correctly.It also provides the basis for revealing the causality mechanism in medicine and validating the theoretical hypothesis of prior structure.
【学位授予单位】：山西医科大学
【学位级别】：硕士
【学位授予年份】：2007
【分类号】：R311

【引证文献】