Predictive vs. Structural Modeling: PLS vs. ML

Abstract

The Latent Variables Path (LVP) model and also the first viable estimation technique were developed by Karl Jöreskog (from 1973 onwards). Right from the beginning his LISREL method was of the utmost elegance and perfection, as regards the deductive properties of the model. The LISREL program has since been gradually improved (LISREL III: Jöreskog & van Thillo 1973; LISREL VII: Jöreskog & Sörbom 1987), as regards convenience for the user, additional estimation criteria and model diagnostics. However, one inconvenient side of LISREL, well known to all users, is incurable: LISREL tends to produce improper solutions characterized by negative variance estimates and LV correlations greater than unity (van Driel 1978, Jöreskog 1981, Fornell & Bookstein 1982, Rindskopf 1983, 1984).