Linkage analysis and driver analysis

One of the more interesting tools to arrive in recent years has been linkage analysis. One way to describe linkage analysis is through examples:

• One item of an employee survey is known to be highly correlated with employee retention. Issues which seem to predict retention are statistically identified.
• An employee survey from 1998 is combined with customer survey data from 1999. Using the same methods, we can try to find the employee survey items (or categories) which best predict customer satisfaction or loyalty.
• Using an enrollment database and student survey data, along with any of the techniques listed above, predicting retention and academic achievement based on student satisfaction, SAT scores, high school GPA, age, and other factors.

Statistically, there are many important issues to consider with regard to linkage analysis. Most of the time, people use cross-sectional data - that is, we look at different locations or people at the same time. Generally, I have tried to have the predictors from one year and the outcomes from a later year, but this is not always possible, and is only a partial answer in any case. We are making many assumptions about causality. (There are other statistical assumptions, and many writers have described them in detail. I would like to refer you to StatSoft's site, since they have an on-line textbook that covers much of this).

Note: I am not making a distinction between driver analysis and linkage analysis. Driver analysis can refer to the attempt to find drivers of one factor, such as intention to leave the company, within a single data source, such as an employee survey; linkage analysis is similar but uses more than one data source (hence "linkage").

Some of the methods used are:

• LISREL - essentially, models are statistically tested to see how well they fit the data. One could generate models using theory or other statistical techniques and check them for goodness-of-fit using LISREL. This tests the entire model rather than a single relationship. (A model could be as simple as employee satisfaction increases customer satisfaction which increases revenue).
• CHAID, CART, and many other "tree-building" systems - using a wide variety of techniques, the data are analyzed to see which splits can best predict the outcomes. For example, the decision to buy an Acme widget can best be predicted first by a customer's prior widget purchase, then (for purchasers) by their satisfaction and (for nonpurchasers) by their position in an organization. Again, I refer you to StatSoft's site.
• MULTIPLE REGRESSION -Regression is the most accessible method, as well as one of the most robust. In most cases, regression will be enough to show a relationship.
• DISCRIMINANT ANALYSIS - This tends to have similar results as regression, but is used to "work backwards."
• PATH ANALYSIS - Not covered by this site.

One problem which is generally not covered refers to the real life use of linkage. Often, corporations and other organizations use the "linkage map" to focus their actions on issues which will have the highest impact on some outcome - customer service, profit, turnover, etc. If this is the end goal, then most methods, which are concerned with the "truth" (or goodness of fit) between the whole model and the data, may not be appropriate, because you really want to know what you can do to get the most change in the outcome. Sometimes, it pays to ignore the issue of shared variance and work on the major contributors to an outcome individually.