Author Description

As the international large-scale assessments (ILSAs) become more popular, policy makers and education practitioners are interested in collecting as much student background information as possible to better understand the learning context of their students. To collect such abundant information, administrators need to develop a lot of questions. However, assigning each question to every participant would likely exhaust participants. This dissertation is dedicated to balancing the conflict between the substantial information to be collected and the limitedly available logistics and resources in practical operations of large-scale educational assessments. This dissertation developed and compared various multiple matrix sampling (MMS) designs and missing data methods to plan and handle missing context responses. Based on multiple matrix sampling designs, I divided a long context questionnaire (CQ) into multiple short blocks that were not overlapped, and then developed a variety of forms with each form consisting of several blocks. Each participant was randomly assigned a form. Afterward, I adopted four missing data methods, including dummy coding, multiple imputation with Markov chain Monte Carlo (MCMC) algorithm, multiple imputation with predictive mean matching (PMM) algorithm, and regularized iterative principal component analysis (iPCA) method, to complete the missing context responses planned by each MMS CQ design.I conducted two simulation studies and one empirical study to investigate the performance of the various CQ designs and missing data methods on recovering the true or empirical values of the student population/subpopulation plausible values (PVs), Cronbach's alpha coefficients of the constructs, root mean square error of approximation (RMSEA) values of the confirmatory factor analysis (CFA) models for the context constructs. I also examined the CQ designs and missing data methods' performance on recovering the true/empirical values of the correlations acro