On the convergence of the partial least squares path modeling algorithm (Q626259): Difference between revisions

Latent variable path modeling is considered with partial least squares (PLS) estimation of the model parameters. The author discusses artificial data on which the Lohmöller's PLS algorithm does not converge. All the datasets are obtained from a model with three latent variables \(\xi_i\), where \(\xi_1,\xi_2\sim N(0,1)\) are exogenous and \(\xi_3=0.3\xi_1+0.5\xi_2+0.3\xi_1\xi_2+\zeta\), \(\zeta\sim N(0,0.57)\). All the manifest variables \(x_{jh}\), \(j=1,2,3\), \(h=1,2\) are generated as \(x_{jh}=\sqrt{0.5}\xi_j+\varepsilon_{jh}\), \(\varepsilon_{jh}\sim N(0,0.5)\). The path model for the analysis neglected the interaction effect \(\xi_1\xi_2\) in the description of \(\xi_1\). In 3 000 simulated samples of different size, six samples on which PLS does not converge are found. The behaviour of the stopping criterion on these sets is discussed.