Drive Action with Bayesian Network Modeling
- February 16, 2015
Expert marketers know that to drive sales, you need to enhance brand perceptions and improve the most important aspects of customer experience. But what brand perceptions should you try to enhance first? And which aspects of customer experience most strongly influence your brand’s health?
Brand equities and customer experience features don’t exist in a vacuum: if you change one perception or aspect of the experience, it tends to impact other perceptions. And since you can’t expend resources improving every possible perception of your brand, you need to determine what specific improvements will get you the most bang for your buck overall through positive interaction effects.
Traditional derived importance analysis (or “drivers analysis”) will tell you what brand attributes or product features have the biggest direct impact on the business outcome you desire (like intent to purchase). However, they can’t tell you about how different brand perceptions interact with one another, or simulate the shortest path toward the biggest overall impact with the lowest amount of investment. Using the power of Bayesian network modeling, you can uncover such valuable insights.
Traditionally, divers analysis is done using multiple, or often, bivariate regression. The objective is to rank the relative impact of different independent variables (like the consumer’s perception of the store’s layout, customer service, and product) on a specific dependent variable (such as future purchase likelihood), as in the example below:
|Brand Perceptions||Relative impact on future purchase likelihood (if “average” impact is 100)|
|Store is laid out well||140|
|Customer service is excellent||125|
|Has products I’m interested in||110|
|I feel welcome when I arrive||80|
Looking at the example above, we might feel confident that we should invest the most energy and resources in improving store layout and customer service, and the least in making customers feel welcome. But regression-based drivers analysis has a number of blind spots when used for this purpose. Bayesian Network modeling fills in those blind spots:
- Interaction effects. Assume that customer service and feeling welcome correlate strongly with one another, such that improved perceptions of customer service would lead to an improvement in feeling welcome. In that case, focusing on customer service might be even more important than focusing on store layout, even if customer service has a lower direct impact. This is because customer service has indirect effects on other important perceptions (like feeling welcome) that will get you closer to your overall goal of increasing purchase likelihood when the effects are combined. Both multiple and bivariate regression would miss such interactive effects and thereby undermine effective prioritization.
- Bayesian Network modeling more accurately accounts for these types of “interaction effects” when it builds a model. For example , it might be that improving customer service would improve feeling welcome so much that the two combined will make customers even more likely to purchase than would improving just store layout on its own. Bayesian modeling would make it clear that we should therefore prioritize customer service because of its more positive “total effect;” i.e., direct plus indirect impact.
- The Bottom Line: Regression-based drivers don’t take into account the interaction between different brand perceptions: Bayesian Network drivers do, thus allowing you to more accurately estimate the overall impact of specific potential improvements.
- Multicollinearity. Assume again that “customer service” and “feeling welcome” correlate strongly with each other, and so have a similar, highly correlated impact on purchase likelihood; i.e., they have high multicollinearity. If excellent customer service has just a bit higher correlation with purchase likelihood than does feeling welcome, multiple regression models will indicate that customer service accounts for the lion’s share of the total impact of both. This will make feeling welcome appear significantly less important than customer service, even if the independent impact of each is nearly the same.
- Unlike in the example above, Bayesian Network modeling parses out exactly how much impact customer service and feeling welcome each account for, even if the two perceptions are highly correlated with one another. This means that instead of just learning which attribute is more important than the other, and being led to believe that the less impactive attribute is less important than it actually is, we instead get a detailed picture of just how far each individual brand perception moves the needle on purchase intent.
- The Bottom Line: Highly similar brand perceptions don’t get lost in the statistical shuffle using BayesNets drivers modeling: illuminating clusters and factors of brand perceptions instead of forcing perceptions into a ranking makes sure every attribute gets its due.
Bayesian Network modeling drivers analysis makes the same basic assumption that previous generations of drivers analyses have; that is, that association and correlation with the target outcome will help us better prioritize what “levers to pull.” But BayesNets drivers allow us to prioritize more accurately, taking into account statistical relationships that better resemble the highly interactive, multi-collinear reality in which brand perceptions actually influence brand health outcomes: the real world marketplace.