Techniques and tools
Structured Equation Modelling

Often in market research, we are trying to measure, predict or assess the influence of underlying dimensions in market places. These “latent” dimensions or factors, are not directly observable, though explain a great deal of decision maker behaviour. They can be measured indirectly using a combination of observed variables as proxies. However, while we may have intuitive theories about how a market behaves the precise nature of the relationship between observed and latent variables is often unknown to the researcher in advance.

In order to build a market model, suitable proxy variables are usually included in a questionnaire, based on the researcher’s understanding of the market place or on qualitative research. Examples of latent factors are shown below:

In this example, the three latent dimensions of the product explain the observed variables captured in the questionnaire. The single headed arrows specify the strength and direction of the relationship between each latent factor and each observed variable. So for example, the brand of the previous product is the best indicator of the factor we have called “commitment to brand”, while the number of other products owned by that brand is the poorest indicator. We might also want to assess other dimensions of the product such as “Aesthetics” or “Ease of Use”. For simplicity we will ignore these for now.

While such a model is interesting, it would be more useful if we could determine which factor has the greatest influence on Brand Advocacy (i.e. likelihood of recommending brand to others). Such a relationship is shown in the following diagram. In this example, Brand Advocacy is itself a Latent Factor, measured using two observed questions (likelihood of buying the brand again and likelihood to recommend to others). The weights shown on the single headed arrows represent the impact and direction (positive or negative) that each latent factor has on Brand Advocacy. For examples such as this we can estimate the proportion of variance (movement around the average score) in Brand Advocacy explained by Customer Services Performance, Product Performance and Commitment to brand. We might find, for instance that 70% of its variance can be explained by these factors (a good model by any standards!).

This is a simple illustration of a very powerful statistical technique known as Structured Equation Model (SEM). SEM is a confirmatory (as opposed to an exploratory) model building approach, typically used to validate a “theory” using survey data. It combines several commonly used statistical techniques – Factor Analysis, Multiple Regression and ANOVA (Analysis of Variance) – allowing various models of the market place to be tested and refined. In practice, we can estimate a more complex system of related models which help us understand what drives a market.

Another variant of the SEM technique, known as Partial Least Squares (PLS), can be used when “making predictions” rather than confirming theories is the main goal. This allows us to build systems of models which allow “what-if” scenarios to be run using a simulator. In our example, this will enable the researcher to approximate how much brand-advocacy would change in response to changes in the observed variables on the left hand side of the model.

Structured Equation Models are most effective when applied to very complex systems of relationships where there are many variables influencing many other variables. In the example below, a hypothetical model is specified for television service. In practice we could assess how well this model fits the data and compare it with alternative model specifications.

Contained within this example is an example of a “mediator effect”. This is where the impact of one variable on another variable is mediated by some third variable, which explains the part of the relationship. In this case part of the relationship between perception of Interactive Services and Overall Satisfaction is explained by satisfaction with the TV program choice. Quantifying more complex relationships such as this allows us to obtain a fuller understanding of why certain factors drive certain other factors.

A major application of Structured Equation Modelling which we anticipate emerging in 2006 is the linking of employee satisfaction to customer satisfaction and then the linking of both of these to Financial Performance, Churn, and other business performance metrics. This is most applicable when data is available for a large sample of business units (which would be the unit of measurement). Logit Research has worked on these kinds of application in 2004/5 and has found this a very effective modelling technique for this kind of application.

The latest implementation of SEM includes the use of Bayesian estimation with Structured Equation Models. This allows models to be estimated on much smaller samples and is also more effective at dealing with situations where there is a lot of missing data.

Please contact Gary Bennett for further information (garyb@logitresearch.com).

< Back to Techniques and Tools Home Page