Several statistical techniques are subsumed within the SEM framework, including t-tests, ANOVA, ANCOVA, MANOVA, multiple regression, factor analysis, path analysis and measurement error models, among others. What makes SEM special is its flexibility for specifying these or other models and relaxing or imposing constraints that align with theoretical notions. For example, one could fit – simultaneously – two repeated measures ANOVA models to two processes that occur over time, facilitating exploration of the processes’ dynamics as they evolve, while relaxing stringent assumptions of the standard repeated measures ANOVA model.
SEM with unobserved, or latent, variables are particularly useful to fields seeking to study intangible concepts; psychology and education apply SEM widely as they focus on personality, attitudes, achievement, cognition and many other latent constructs. However, latent variables are not unique to these fields; indeed, marketing, management and business, for example, can find value in understanding perceptions, satisfaction, innovation or performance, whereas manufacturing and engineering can benefit from modeling quality, energy and other unobservable factors in industrial processes.
With these seemingly advantageous features, why are SEMs not part of everyone’s toolkit? I hypothesize there are three driving factors:
Accessibility. SEM is a technique taught primarily in graduate social sciences’ programs, so it would be unusual for a classically trained statistician or biostatistician to come across SEM or learn the technical details behind it.
Ease of use. Leading SEM software packages have been historically developed by academics with great understanding of the technique but not enough attention to usability. Moreover, great flexibility sometimes brings great complexity; as analysts aim to solve big challenges with SEM, overly complicated models are prone to specification errors, and ultimately estimation errors.
Domain expertise. Analysts’ knowledge allows them to specify models that align with theory and previous knowledge. In many applications, it is not important to understand why or how certain factors predict an outcome. In these cases, unsupervised models, driven entirely by the data, are an adequate choice, and SEM is not.
Luckily, accessibility and ease of use need no longer be barriers to using SEM. JMP Pro recently introduced an SEM platform to its already rich collection of statistical methods, and this means analysts can find lots of resources for learning how to apply SEM to guide decisions that improve their organizations. The SEM platform in JMP Pro inherits all the interactive and dynamic features that JMP is known for, plus ease of use is at the forefront of its development, with ongoing error checking to alert users to potential issues before they arise in estimation. It has never been easier to add SEM to your statistical toolkit. This sophisticated technique might just be the tool that leads your organization to make better data-driven decisions.