Partial Measurement Invariance Biases Observed Composite Scores

• In Event: 44.078 - Advanced Issues in Structural Equation Modeling

Abstract

Measurement invariance is implicitly assumed whenever measurement models are used to answer questions about differences in a construct across time or between groups. There has been an abundance of methodological research on how to detect measurement non-invariance. However, there has been surprisingly little research no how to proceed if one finds partial measurement invariance. Our study focuses on how partial measurement invariance influences observed composite scores and downstream statistical tests conducted on them. We simulated a 2-group factor model, with categorical indicators, indicative of Likert-type items. We find that even with a small number of non-invariance thresholds, and a small magnitude of non-invariance, type I error rates of statistical tests are inflated and group means are biased.

Authors

• Jessica Kay Flake, York University

• David B. Flora, York University

• Szu-Han (Steven) Chen, York University