Satiscan and Regression Analysis: A Comparison

The comparison shows the advantages of SatiscanTM, an analytical method from TRC, over regression in identifying the correct and cost efficient action steps.

 

Satiscan™ is an artificial intelligence based method that was developed to conduct key driver analysis in a way that regression is not capable of doing. Traditional regression analysis considers the direct impact of each independent variable on the dependent variable, but has no way of identifying relationships among independent variables. As a consequence, if there are complex relationships between independent variables, the inability to specify them will cause incorrect estimation of the relative importance of each independent variable.

For example, independent variable A may be important both because of its direct impact on the dependent variable X and because of its impact on another independent variable B that in turn affects the dependent variable. In this situation, traditional regression analysis will underestimate the total importance of A.

The practical effects of this can be quite important. Most often, the results of key driver analysis are used to make decisions about where to allocate service improvement resources. Ordinarily, only some of the independent variables in a key driver analysis are subject to direct manipulation – perhaps only a few of them. Consequently, failure to appreciate the true importance of a driver can result in lost opportunities and/or serious misallocation of service improvement resources.

In contrast to traditional regression analysis, Satiscan™ uses artificial intelligence techniques to model all relationships among variables. It systematically searches all of the possible path models to find the one that is most consistent with the data.

The next two pages show an actual example, a case where the same data was analyzed using traditional regression analysis on one hand and Satiscan™ on the other. The first page shows the traditional regression analysis, with the impact of each independent variable expressed as a Beta weight. (Beta weights measure the importance of each variable in determining another variable. In this case, for example, “called back when rep said” is only about a fifth as important as “satisfaction with service rep” in determining “satisfaction with the call experience.”)

In some ways, the Satiscan™ model of the same data, shown on the next page, is similar. “Satisfaction with service rep” is still the most important direct determinant of “satisfaction with call experience,” with exactly the same Beta weight or measured importance. “Had to call only once” is still the second most important independent variable, in terms of its direct effect, and its importance is almost the same.

But “had to wait when first called,” the third most important independent variable in the traditional regression analysis, has no impact at all in the Satiscan™ model. As a potential independent variable, it doesn’t affect either the dependent variable or any of the other independent variables.

Figure 1

Figure 2

One can well imagine the apparent importance of “had to wait when first called” in the traditional regression analysis having real practical consequences. Of the top three variables in the traditional regression analysis, it is probably the one most easily acted upon. After all, increasing “satisfaction with service rep” is difficult at best, and “had to call only once” may be impossible in many service situations or with many systems complexities. But to improve “had to wait when first called” one need only increase staffing at the call center. Doing so may be expensive, but at least it is possible.

In contrast to the regression analysis, Satiscan™ shows that increasing call center staffing to reduce initial waiting times, by itself, would be waste of money.

By specifying chains of causation, Satiscan™ paints a picture that can be acted upon. “Rep listened,” “rep took responsibility,” and “rep was knowledgeable” form such a chain, with “rep listened” having an important impact on “rep took responsibility,” which in turn is an important determinant of “rep was knowledgeable.” Both “rep took responsibility” and “rep was knowledgeable” directly affect ultimate satisfaction. Arguably, however, increasing either responsibility or knowledge is quite difficult, while one might more easily get reps to listen more carefully, or at least to make sure callers know that they are listening. By identifying the relationships among the three variables, Satiscan™ shows that getting reps to listen more carefully will be fruitful. Thus, it identifies a practical leverage point.

As these examples make clear, the critical difference between the two models occurs when the direct and indirect effects of the predictor variables are examined in Satiscan™something traditional regression analysis simply cannot do. In Satiscan™the whole impact of a variable can be calculated as a “total effect,” and shown on a “Total Effects Table.”

The total effect of a variable is the sum of:

• Its direct effect on the dependent variable, plus
• Its direct effect on another independent variable multiplied by that independent variable’s direct effect on the dependent variable, plus

 

Etc., until all of its effects have been exhausted.

Thus, in the Satiscan™ model, the Total Effect of “had to call only once” is:

 

Thus, “had to call only once” is more than 50% more important than the regression analysis said it was.

 

This content was provided by TRC. Visit their website at www.trchome.com.