As you work to improve your current products and fill your pipeline with innovations, the first thing you’ll want to know is what’s most important to consumers when they’re considering a product like yours? Teasing out the “must have’s” from the “nice to have’s” can mean the difference between success and failure.
Key driver analysis has been the traditional “go-to” answer to that call, utilizing regression models to deliver a straight-forward hierarchy of what elements are driving product appeal. While a time-honored approach, it is not without its flaws. The good news? There is a better way.
Bayesian networks provide a far more sophisticated, yet still simple-to-explain, solution to uncovering key drivers. The probabilities-based alternative pairs advanced analytic power with an informative, easy-to-digest visual. The networks mitigate many of the shortcomings associated with regression analysis without losing what makes key driver analysis popular in the first place.
The most frequent problem with using a standard regression-based approach to key drivers is the potential for over-simplification of the way product elements relate to each other and to the “target” measure (such as Purchase Intent or Overall Liking). A basic, but necessary, assumption to the regression model is that the measures included as potential drivers are not related to one another. This independence is almost never the case and in many instances, the relationships between attributes are immediately evident and intuitive. For example, consider the researcher asking both “Liking of sugar coating” and “Liking of sweetness” in the context of a product test: both attributes are important to understand, but clearly, they are not independent.
The resulting issue is a high level of “multi-collinearity,” in which the explanatory power of these two attributes overlaps, but where all of that power is attributed to only one of the two elements. Finding that “Liking of sugar coating” is a key driver, but that “Liking of sweetness” is not statistically related to Overall Liking at all yields an awkward interpretation of key driver regression results, and may throw the entire analysis into question.
Solution: Bayesian networks. By creating a structure that considers all attributes of a product in an inter-related network, strong correlations between items become a strength, rather than a weakness of the model. The network structure also visually lays out the interrelationships between metrics. This allows researchers to follow along the entire path (or multiple separate approaches) of all elements to the “target” (e.g., Overall Liking). No longer are the key drivers decided behind a proverbial “black box” linear regression, but instead the full story can be followed like a map. It allows for not just a single, most influential path to product appeal, but for a comprehensive product view from which to build thoughtful business strategy.
In short, Bayesian driver networks are both statistically more robust than traditional regression, and produce a clear and simple output. An elusive “win-win” for researchers and a clearer path to success for your products or services.
Lisa Shaffer is a Senior Marketing Science Specialist with a Bachelor’s Degree in Mathematical Statistics from
Wake Forest University. She has been with RTi research since 2017.