The use of latent variable mixture models to identify invariant items in test construction

Area of Science:

• Psychometrics
• Statistical Modeling

Background:

• Patient-reported outcome measures (PROMs) are crucial for assessing health outcomes in diverse patient groups.
• Ignoring population heterogeneity (e.g., gender, ethnicity) in PROMs can lead to biased results.
• Latent variable mixture models (LVMMs) offer a method to address heterogeneity and assess measurement invariance (MI).

Purpose of the Study:

• To discuss and demonstrate the application of LVMMs for identifying invariant items in test construction.
• To explore the use of LVMMs within the Draper-Lindely-de Finetti (DLD) framework for latent variable measurement.
• To improve the reliability and validity of PROMs in heterogeneous populations.

Main Methods:

• Utilized LVMMs to compare 1- and 2-class item response theory (IRT) models on a 39-item daily activities measure.
• Employed differential item functioning (DIF) analyses to detect non-invariant items.
• Iteratively removed non-invariant items and re-applied LVMMs and DIF testing until measurement invariance was achieved.

Main Results:

• A 2-class LVMM demonstrated a better fit than a 1-class IRT model, indicating underlying heterogeneity.
• Initial analyses revealed statistically significant bivariate residuals, suggesting local dependence.
• After iterative refinement, nine items were identified as the most invariant, ensuring MI.

Conclusions:

• LVMMs provide a robust framework for identifying measurement invariance in PROMs.
• The DLD framework supports the use of LVMMs for item selection in test development.
• This approach has significant potential for advancing research on PROMs for heterogeneous populations.