The problems with current analytical procedures

Abstract

By the early 1980s three main analytical procedures had been developed for the examination of path models or networks of influence in educational research such as the one presented in Figure 2.1 in the previous chapter. These procedures are: 1. Ordinary least squares path analysis (OLS) (Peaker, 1971; Keeves, 1972). 2. Partial least squares path analysis (PLS) (Wold, 1982; Noonan & Weld, 1985; Sellin. 1986). 3. Linear structural relations analyses (LISREL) (Munck, 1979; Joreskog & Sorbom, 1989). It should be noted that all three analytical procedures had been developed in such a way that they could be employed in the analysis of data collected in large cross-sectional surveys of educational achievement. Nevertheless, while each of these procedures has certain advantages in the analysis of data collected in educational research, their use has helped to expose certain problems which arise in the analysis of cross-sectional survey data in the field of education. The basic problem which confronts analysts of large bodies of data is that of selecting or combining together data on many variables which are considered to be relevant within a simple and coherent model that is capable of being tested. Partial least squares analysis and linear structural relations analysis employ constructs, regarded as latent variables, for which observed or manifest variables act as multiple indicators. Observed variables can be combined together by principal components analysis, by canonical analysis, or by rosettes as carriers of regression (an idea suggested by Tukey), but such an approach requires the preselection of variables for inclusion in a model. Analyses that compare the use of these methods (see Keeves, 1986) would appear to indicate that partial least squares path analysis has more flexibility, although linear structural relations analysis provides a more rigorous testing of the model. In this sense partial least squares analysis can be regarded as exploratory and linear structural relations analyses as confirmatory. One apparent advantage that both ordinary least squares and partial least squares analysis have over linear structural relations analysis is that the two least squares