Linear latent variable models and covariance structures (Q915307)

This paper treats a generalization of factor analysis of observational vectors \(\underset{\tilde{}} X\) in terms of the unobservable latent vectors \(\underset{\tilde{}} Z_ k\) as \[ \underset{\tilde{}} X=\mu +\sum^{K}_{k=0}\Lambda_ k(\lambda)\underset{\tilde{}} Z_ k,\text{ with } E(\underset{\tilde{}} Z_ k\underset{\tilde{}} Z_ k')=\Phi_ k. \] The factor loading matrices \(\Lambda_ k\) and \(\Phi_ 0\) are assumed to be functions of parameter vectors \(\lambda\) and \(\tau\), respectively, and the covariance \(\Sigma\) of X a function of the vector \(\theta '=[\lambda ',\tau ',\Phi_ k'].\) Extending the earlier results obtained under the normality assumption it is shown that the maximum likelihood estimator of \(\lambda\) will be asymptotically normal if the \(\underset{\tilde{}} Z_ k\) are independent with finite second order moments and one of the Z's is nonstochastic, thus avoiding the normality assumption. The likelihood ratio criterion also tests goodness of fit of \(\Sigma\) under non- normality. Alternative goodness of fit tests also share this property. Examples of applications for testing independence of subvectors of \(\underset{\tilde{}} X\), multiple battery factor analysis, and panel studies in autoregressive processes are given.