A partial least squares method is described for estimating parameters of linear structural relation models. This method is an extension of Herman Wold's proposal for estimation parameters without distributional assumptions, using some algorithms worked out by Paul Horst. The method (LISPLS) determines a different number of latent variables from several sets of measured variables simultaneously and fits the relations of the latent variables either to a path model or to a correlation pattern. The name LISPLS indicates LISREL (linear structural relation model) under partial least squares criteria. Applications to three different models are described: canonical correlation analysis, the canonical redundancy model, and the interbattery factor model. These models may be generalized to analyze more blocks and different numbers of factors from each block. The model and its assumptions are described. Applications are shown to a path analysis and a multi-set factor analysis. (CTM)