Stated importance, correlation, regression and key driver analysis

The Leadership Factor discuss the best way to measure what is important to customers and how these measures should be used for deciding how to improve customer satisfaction and loyalty.

There is much debate over the best way to measure what is important to customers and how these measures should be used for deciding how to improve customer satisfaction and loyalty.

First of all, derived importance is not a measure of attribute importance, but of the strength of association between the way each attribute and overall satisfaction are scored. The well-established (though over-stated) tendency for certain requirements to be scored high for stated importance, even though they seem to have little impact (i.e. "givens" or "hygiene factors") reflects two things:

1. Derived importance depends on variation - if performance is consistently good impact coefficients will be low
2. The relationship between performance and overall satisfaction is often nonlinear. In other words derived importance tends to be dynamic, reflecting current performance

Overt ratings (stated importance) are a better guide to attribute importance since they do not reflect current performance. Prioritising based purely on derived importance thus risks diverting attention from key areas, leading to a fall in performance followed by an increase in impact. "Givens" by their very nature are areas in which you cannot afford to make mistakes.

By the same token overt ratings are a better guide to long-term importance. It has been shown Michael Johnson et al at Michigan University that a model based on stated importance is a better predictor of loyalty than (regression-based) derived importance. Nonetheless derived importance is a useful tool, and is used in The Leadership Factor’s analysis as part of a suite of tools designed to select priorities for improvement (PFIs) that will provide the best way to safeguard and improve overall satisfaction in the long term.

We agree, correlation is not causation (though regression can't prove it either!). We prefer correlation to regression as a measure of the strength of association between attributes and overall satisfaction since regression has been shown to be very unstable in the presence of strong multicollinearity (as is normally the case with customer sat data). This can result in a very misleading measure of the relative importance of different attributes, and a much greater risk of mis-prioritising. More robust measures of importance do exist (e.g. Kruskal's measure or "True Driver Analysis"), but they are very complex and hard to explain, so on balance we normally opt for correlation as a simple but effective guide.

Where causal models are required much more sophisticated techniques than regression must be used to perform Key Driver Analysis (KDA) effectively. At minimum Principal Components Regression (or similar) must be used in order to reduce the multicollinearity present in the data. Alternatively techniques such Partial Least Squares (PLS) or Structural Equation Modelling (SEM) could be used.

Information contributed by The Leadership Factor.