Robustness of Individual Score Methods against Model Misspecification in Autoregressive Panel Models

Abstract

Different methods to obtain individual scores from multiple item latent variable models exist, but their performance under realistic conditions is currently underresearched. We investigate the performance of the regression method, the Bartlett method, the Kalman filter, and the mean score under misspecification in autoregressive panel models. Results from three simulations show different patterns of findings for the mean absolute error, for the correlations between individual scores and the true scores (correlation criterion), and for the coverage in our settings: a) all individual score methods are generally quite robust against the chosen misspecification in the loadings, b) all methods are similarly sensitive to positively skewed as well as leptokurtic response distributions with regard to the correlation criterion, c) only the mean score is not robust against an integrated trend component, and d) coverage for the mean score is consistently below the nominal value.