Basic Principles and Practices of Structural Equation Modeling in Criminal Justice and Criminology Research

Abstract

Structural equation modeling (SEM) has undergone rapid advances in recent years that have made this technique useful to social scientists studying a variety of topics. The hypothesis‐testing and model‐evaluation capacity of SEM makes this modeling strategy useful for research in criminal justice and criminology. The flexibility of this strategy and the availability of user‐friendly software programs add to the appeal of SEM. This article is a basic introduction to the use of SEM in criminal justice and criminology. An overview of SEM, popular computer programs, and the sequence of steps in building and evaluating models is provided with the intent of offering readers a general roadmap of the use of SEM. The article concludes with a list of recommended readings for those wishing to study this topic in detail.

Notes

1. It is worth noting that SEM programs are not necessarily the best software for all types of analyses—depending on the nature of the data and the research question under investigation, it may be more appropriate to use other programs that offer similar benefits (e.g., HLM or SAS may be preferable to an SEM program for latent growth or multilevel modeling in some circumstances).

2. Mplus, EQS, and AMOS do not have data manipulation or storage capabilities, but LISREL offers PRELIS for this purposes.

3. Free parameters include all paths and covariances (except those that are fixed), one variance estimate for each error or disturbance term that accompanies every dependent variable in the analysis (both manifest and latent), a variance estimate for each independent variable, and one path from each error or disturbance term to its corresponding observed or latent variable.

4. Researchers who wish to drop more than one path from the full model to create the restricted model can do so, either one path at a time or all paths at once. Dropping multiple paths simultaneously, however, lends uncertainty to the interpretation of χ ² _D if it emerges as statistically significant because there is no way to tell which of the dropped paths produced a meaningful reduction in fit. Dropping paths one by one solves this problem because the changes can be evaluated individually.