Structural Equation Modeling -- The Other 'SEM'
Sam Koslowsky, Harte-Hanks
While the arsenal of tools available to market researchers has certainly grown over the years, perhaps none has been more insightful than what is referred to as structural equation modeling (I call it the “other” SEM, distinct from search engine marketing). SEM provides an ability to perform analysis that frequently market researchers can only dream about. Based on what can be quantified, theories and predictions are developed on behaviors that cannot be quantified.
I am reminded about a stately looking gentleman who, upon emerging with a new Rolex watch, was sure to display his new timepiece in full view of his colleagues. A nearby observer referred to him as a “conspicuous consumer.” Others thought that he was a “quality seeker.” While we may never know who was right, what is important to realize is that both these descriptions were based simply on perceptions. However, SEM is the tool that allows the researcher to analyze these sorts of events. Investigators refer to these as constructs or latent variables.
Market researchers employ questionnaires to delve into latent-type variables such as “conspicuous consumption.” Consider the following true or false statement: “I make sure my neighbors see my new vehicle.” The response to this item is quantifiable. The “conspicuous consumption” is a non-directly measurable theory — a latent construct. Researchers try to indirectly measure the latent variable via numeric data captured from surveys.
Each of us may have assumptions about how a particular market behavior works, but the association between measurable data and latent-unobserved variables is often unclear. How does the investigator deal with such a phenomenon? Enter structural equation modeling.
Marketers assess whether variables are correlated through a set of relationships by evaluating the statistical variability of the variables. Both direct and indirect relationships among observed variables can be detected with SEM. Furthermore, estimates of the indirect effects of the independent variables (predictors) on the dependent variables (that which we are predicting) can be computed. Marketers can assess assumptions and hypotheses by identifying which variables are associated with other variables. The analyst has the option of examining certain sections of the path diagram that may be supported by some preconceived notions, and disregard other paths that may have no basis in reality.
Sam Koslowsky is VP of modeling solutions at Harte-Hanks.