Regression Techniques

There is a huge technical problem in organizational surveys for which there is no ultimate solution. This is the problem of multicolinearity. In the words of the physicist, "Everything is connected to everything else." The reason why this is a major problem for organizational surveys is that, "Everything is connected to everything else, by a lot!" There are a number of statistical techniques for partitioning variance in the context of multicolinearity, and helping us find practical uses and practical explanations for data.

When these variance partitioning techniques are used as data bound statistical black boxes, the risk of capitalizing on error, and letting the data run away with the analysis increases. Let's take regression analysis as an example. If all independent variables under consideration were orthogonal to each other, there is no need to choose from among several operational approaches. These operational techniques are commonly known as forward selection, backward selection, step-wise, a priority, and path analysis (a variation on the regression concept.) To this list, we can include the selection of variable weights by other means. These might be the value of the zero order correlation, the squared value of the same, partial correlation coefficients, or semi partial correlation coefficients.

The question becomes, which is the correct, or at least the best, operational technique to use with organizational survey data. The answer is very simple. There is no correct way, nor is there a best way, except in the case of orthogonality. So what is a poor researcher to do? First, recognize that the test theory foundation of surveys is not a true theory, but a tautology. Thus, the utility of any tautology is based on the usefulness and practicality of outcomes.

There is a technique for partitioning regression in the context of organizational surveys that may yield more useful results than the others. This technique is a variation on backward selection, and focusing on incremental variance in the presence of all independent variables. Its application in research on organizational surveys seems to offer more satisfying explanatory benefits than the other techniques.