非隨機遺漏之結構方程模型估計─潛在變項選擇模型與組型混合模型

• Main Authors: Chung-Ping Cheng, 鄭中平
• Other Authors: Li-Jen Weng
• Format: Others
• Language: zh-TW
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/22957351311811714596

博士 === 國立臺灣大學 === 心理學研究所 === 91 === Structural equation modeling (SEM) is a popular statistical method for social science research. Incomplete data are often encountered when researchers make an effort to collect empirical data for test of their hypothesized models. Most missing data treatment methods in SEM assume that data are missing completely at random or missing at random. These two missing patterns consider the impact of manifest variables on data missingness. Missing data mechanisms of selection model and pattern mixture model also focus on the relationship between missingness and manifest variables. For psychologists interested in latent variables, the missing data mechanisms that ignore the influences of latent variables tend to be insufficient. The research therefore introduced latent variables into selection model and pattern mixture model, referred to as latent variable selection model and latent variable pattern mixture model, respectively. Latent variable selection models assume that missingness is affected by continuous latent variables, as suggested by Muthén et al. (1987). Stochastic EM algorithm with rejection/acceptance sampler was developed to estimate the parameters. The method was called “latent variable selection model maximum likelihood estimation (LVSM-ML)”. Latent variable pattern mixture model assume that missing data patterns reflect latent classes. Categorical latent variables were added to the pattern mixture model. Maximum likelihood estimation using the EM algorithm was developed and referred to as “latent variable pattern mixture model maximum likelihood estimation (LVPM-ML)”. Three Monte Carlo studies were employed to explore the performance of LVSM-ML and LVPM-ML when latent variables affected data missingness. Results of the present study showed that with missing data mechanism of latent variable selection model, LVSM-ML performed better than other missing data treatment methods at high degrees of data missingness and severe departure from missing completely at random. If the missing data mechanism is latent variable pattern mixture model, LVPM-ML also performed better than other methods. The assumptions of the two missing data treatment methods proposed and directions for future research were also discussed.