67: A Holistic Framework for Assessing Latent Variable Model Fit

Music City Center 

Latent variable models (LVMs) are used across disciplines to investigate theories about the underlying structure of observed variables. Key to building LVMs is model fit evaluation. Fit is measured by how well the model reproduces the empirical covariance matrix (e.g., RMSEA) or by comparing the fit of multiple models (e.g., BIC). These traditional techniques are limited, ignoring the possibility of model misspecification, local misfit, concordance with prior research, and propensity of some models to fit most datasets well. This presentation offers a framework for assessing all aspects of LVM fit and synthesizing them into a holistic description to help researchers understand the relative appropriateness of different LVMs. This framework includes traditional global model fit (RMSEA), model fit comparison (BIC), as well as sensitivity analysis for model misspecification (ant colony optimization), local model fit, concordance with prior research, and fit propensity. Results across methods are synthesized, providing researchers with a more nuanced and potentially generalizable sense for how well a model fits the data. The presentation includes a complete example of the approach.

Model fit assessment

Latent variable models

Sensitivity analysis

Replication 

Social Statistics Section