A Multigroup Factor Analysis Approach to Analyzing Simple Matrix Sampling Planned Missing Data: (When) Does It Work?

• In Event: 18.018 - Advances in Structural Equation Modeling

Abstract

Certain planned-missing designs (e.g., simple-matrix sampling) cause zero covariances between variables not jointly observed, making it impossible to do analyses beyond mean estimations without specialized analyses. We tested a multigroup confirmatory factor analysis (CFA) approach by Cudeck (2000), which obtains a model-estimated variance-covariance matrix to resolve the zero-covariance issue. With a Monte Carlo simulation in 18 conditions (2 sample sizes × 3 inter-item correlations × 3 numbers of anchor items), we found that in most scenarios tested the multigroup CFA approach successfully performed model estimation that was highly comparable to the models of complete data and resolved the zero-covariance problem. We caution using this approach when inter-item correlation or the number anchor items is low.

Authors

• Ting Dai, University of Illinois at Chicago

• Yang Du, University of Illinois at Urbana-Champaign

• Jennifer G. Cromley, University of Illinois at Urbana-Champaign

• Tia M. Fechter, Office of People Analytics

• Frank Nelson, Temple University