Monte Carlo simulations for evaluating the performance of modern missing data techniques when estimating structural equation models with latent variables. A systematic analysis of different types of multiple imputation.

在这项研究项目中，将测试各种现代缺失数据技术(MDT)，这些技术可以用于估计具有潜在变量的结构方程模型。这些技术包括多重归因法、期望最大化法和直接最大似然估计法。通过蒙特卡罗模拟，我们将在一定的内生检验标准(如估计偏差、收敛和模型拟合)和不同的外生模型参数(如不同的值分布、不同的测量尺度、不同的缺失值机制、不同的缺失值份额和不同的样本量等)下检查MDTS的性能。与以往评估缺失数据技术优势的蒙特卡罗模拟研究不同，我们将集中于不同的多重填充技术的性能。特别是，我们将调查各种数据条件的相关性，以选择特定的MI技术时，估计具有潜在变量的结构方程模型。此外，我们的蒙特卡罗模拟的结果将使用真实的经验数据集进行演示。这将显示在分析经验数据时没有充分应用特定的MDT(特别是错误类型的多重归因法)的后果。该项目的结果将为社会研究实践提供指导，以便研究人员可以选择最适合他们的经验数据的MDT(包括最优的多重归因类型)。

In this research project, various modern missing data techniques (MDT) will be tested that can be used when estimating structural equation models with latent variables. These techniques include methods of multiple imputation, expectation maximization and Direct-ML estimation. Using Monte Carlo simulations the performance of MDTs will be checked in respect to certain endogenous test criteria (such as estimation bias, convergence and model fit) and under varying exogenous model parameters (such as different value distributions, varying measuring scales, different missing value mechanisms, different shares of missing values and various sample sizes etc.). Unlike former Monte Carlo-simulation studies evaluating the advantages of missing data techniques, we will concentrate on the performance of different techniques of multiple imputation. Especially, we will investigate the relevance of various data conditions for selecting a particular MI-technique when estimating structural equation models with latent variables. In addition, the results of our Monte Carlo-simulations will be demonstrated using real empirical data sets. This will show the consequences of not adequately applying a particular MDT (especially a wrong type of multiple imputation) when analyzing empirical data. The results of this project will generate guidelines for social research practice, so that researches can select the MDT (including the optimal type of multiple imputation) that best fits to their empirical data.