By Jeff Cartwright-Smith, Ph.D., Senior Vice President
and Andrew Douglas, Vice President

One of the most common requests we hear from clients is to determine, from among many plausible criteria, which ones drive prescribing. Answering this question drives us, in turn, to dig deeply into the marketing science tool kit. We’d like to share some of the techniques used to address this challenge.

The challenge may be presented in the context of a survey of physician perceptions of drug attributes: Of perhaps 20 brand characteristics rated by active prescribers, on which attributes should our client focus limited marketing resources? Or it may come up in the context of a sales force effectiveness survey such as our Detail Tracker or Detail Watch, or a corporate communications program. Or it may be asked in a message effectiveness survey: Which messages, or rep characteristics or sales tactics are effective and which are not? The intended outcome measure is usually an increase in prescribing, measured either in self-reported brand share or secondary script data (NRx). Other intended outcomes are sometimes used, such as satisfaction, prescribing intention, net appeal of a drug or overall detail effectiveness.

One caveat: We all recognize that without the ability to randomly assign respondents to conditions, one cannot expect true scientific proof of causality. Nevertheless, our clients need to know which of the product attributes (or messages, or sales characteristics or tactics) are associated with and are likely to influence a favorable outcome, and which are not.

It is tempting to look simply at the correlations between each attribute and the dependent variable, an approach sometimes called Derived Importance. Instead of only one or two key drivers, there will usually be many drivers almost equally related to the outcome. (See correlations in a contrived but typical example, Table 1). It is not helpful to report that all the attributes are approximately equal drivers! We also know that we will always find serious evidence of what we call multicollinearity: Our candidate drivers are always highly correlated with each other. That is to say if one drug is rated highly on an attribute, it is likely to be rated highly on all attributes, and the opposite will be true for other drugs. If it is a sales rep or a message being evaluated, the same halo effect is observed. (This may be a consequence of physicians being cognitively overwhelmed with a confusing plethora of products and needing to simplify their assessment processes; or, it may simply reflect a lack of discrimination. We observe product halo effects more often among primary care physicians than specialists.) In any event, halo effects and multicollinearity bedevil our search for key drivers of prescribing. Simple correlations between attributes and prescribing are rarely satisfactory.

Table 1

Another sensible approach might be to use multiple regression to select the attributes driving prescribing. You may recall that using multiple regression is discouraged in the presence of correlated or multicollinear attributes. Table 1 shows why. The coefficients shown in an OLS regression will be extremely unstable and quite misleading if the attributes are strongly correlated, as they inevitably are. It is possible for attributes correlating 0.9 with each other to have dramatically different beta weights—even of different signs!

Principal Components Regression

Perhaps you have anticipated a solution here. We might analyze these redundant attributes in order to deal with their underlying principal components or factors, which will be uncorrelated in a traditional orthogonal, Varimax-rotated principal components analysis. The attributes in Table 1 have been colored to illustrate three likely components: efficacy, cost and safety/side effects. We might then regress these components on the outcome and find efficacy has a beta weight of 0.5, cost of 0.1 and safety/side effects of 0.4.

This approach is called principal components regression. We would now have the answer sought by our client if components (“efficacy,”“safety/side effects”) were acceptable as answers. Typically clients need more granular insight than this. We might then weight the attributes by the contribution of the attribute to the underlying component, and then by the degree to which the component influences prescribing, as seen in Table 2.

Table 2

Now we are closer to a solution. The two important component drivers are shown below, along with some of the attributes that in turn drive them.

Of course, this example is idealized for the purpose of this article. In the real world, alas, things seldom so easily yield to such straightforward analysis. Coming in the March issue of Pipeline, we will dig deeper into the subject to explain why new approaches are needed to overcome the limitations discussed here. In addition, we will explore Shapley Value Regression and Partial Least Squares Analysis, two new approaches pioneered by GfK that can at last crack the old problem of multicollinearity. Stay tuned!

Want to learn more on this topic? Please contact: