NCJ Number
89239
Date Published
Unknown
Length
47 pages
Annotation
Using a plausible conceptual model, this study compares the performance characteristics of the LISREL and partial least-squares (PLS) causal models for the origins of criminality.
Abstract
In addition to examining the substantive interpretations yielded by the results of each analysis, the study compared each model on the performance criteria of (1) stability of parameter estimates, (2) computational time, (3) sensitivity of estimates to non-normally distributed data, and (4) residual distribution characteristics. The study samples were drawn from a true birth cohort, originally studied in an examination of criminality in XYY men. This cohort consisted of all male children born from January 1, 1944, to December 1947 (n=31,436) whose mothers were residents in Copenhagen, Denmark. The data obtained related to parental socioeconomic status, family status, the impact of the 'social strain' of war on later criminality, scholastic aptitude, amount of schooling, respondents' socioeconomic status, and criminality. The testing supported the conceptual viability of the proposed structural model in most regards. Despite obvious violations of multivariate normality distributional assumptions, both LISREL and PLS seem to provide reasonable, replicable results. PLS seems to provide most of the important parameter estimates in considerably less computational time, but with some loss in both capabilities of complex parameterization and in terms of fitting the observed correlation matrix. PLS is clearly superior in the area of computational costs. An area for further examination is the development of a PLS-LISREL interface. In an interface, the PLS estimates could be used as relatively optimal starting values for the LISREL solution. Tabular data and mathematical equations are provided, along with 22 references.