Selecting an Appropriate Technique

Now that we have described the fundamental characteristics of three commonly used analytical techniques that come from the covariance-based family of analyses, the natural next question is which of these is the best? Unfortunately, there is no easy answer, as the choice of the technique depends on both—the analytical objectives as well as the properties of the data available to estimate the relationships. As a very simple illustration, if the research objective is to assess the impact of overall guest satisfaction on the revenue provided by the guest, correlation analysis is unlikely to work. It is very plausible that if guest satisfaction is measured on a 1 to 10 scale, every point improvement is likely to have a revenue impact that is beyond the plus or minus one range presented by correlation analyses. Therefore, in this case, correlation analysis will not be able to answer the research objective. Instead, regression or SEM might be more appropriate.

Overall, while many considerations affect the choice of an appropriate analytical technique, we list some that are important in selecting an optimal tool. One key consideration is the need to perform simulations. As previously discussed, when simulations about likely scenarios are required correlation analyses do not suffice because they have a range restriction on the estimated strength of association. The second key consideration is whether consumer behavior in the industry is consistent with a bivariate technique or a multivariate technique. In other words, would most customers be evaluating various areas of experience one at a time, or would they be thinking of many things all at once. If the customer evaluation of the consumption experience is likely to be multivariate, then regression analysis and SEM will be preferred candidates.

A third key consideration is the volume of available data and the amount of missing data. Since correlation analysis is a bivariate technique that works through pairs of measures, it typically requires fewer data points, than multivariate techniques such as regression and SEM. Of the three analytical techniques discussed, SEM requires the largest sample size. From our experience, SEM models require about 15 to 20 times the number of data points as the number of measures in the model. Therefore, if a model has 20 measures in it, SEM might require 300 to 400 data points for estimation. While considering the available volume of data, it is also very important to estimate the effective base size. Missing data introduces the difference between available data and effective base size, especially for multivariate techniques. In typical multivariate techniques, if any measure that is included in the model has missing data, then the analysis excludes the observation from the computation—unless some form of missing data can be imputed. For example, if a model includes various drivers of satisfaction with a grocery-store visit, and one of the drivers included in the model is a measure with very low incidence, such as the availability of enough wheel chairs, then the model is likely to have lots of missing data for this measure. If the missing values cannot be imputed the effective base size available for analyses will be smaller. Last, but an equally important consideration, is the need to work with latent constructs, and the need to incorporate measurement error. When such a need exists, SEM might be the analytical tool of choice.