Using Bayes Networks to Simplify Statistical Complexity

On September 28, the 4th Annual BayesiaLab Conference will kick off in Nashville, gathering together some of the greatest minds using Bayes Networks, a statistical framework that helps unravel extremely complex problems by making visible relationships in data that can be overlooked. This year’s program will explore everything from modeling Olympic athlete training plans to demystifying consumer purchase journeys. Our own Charles Hammerslough, Ph.D., VSA’s Director of Data Science, will take the stage to explain how his team is applying Bayes Networks to better measure return on marketing investment. Here, we sit down with him to learn a bit more about his tactics and why Bayes Networks could be the key to unlocking vexing business issues.

Q: Tell us more about your background in Bayes Networks. How do they help in your work with data?  

I started reading about Bayesian analysis about four years ago, when I started at VSA. I knew Bayes Networks could be helpful in resolving certain kinds of problems our clients bring to us. Classic inferential statistics—the model I was taught—analyzes the world through very disciplined sampling. Take a basic example: determining height. You may measure the height of students in one classroom as a way of inferring the height of all the students in that grade. You don’t have to measure everybody to draw an informed conclusion.

This works for lots of problems, but I knew that Bayes Networks provide specific advantages for better understanding cause and effect relationships and managing missing data. We often deal with both in marketing strategy, so I kept this method in the back of my mind until last year, when a client’s problem seemed like a good opportunity to apply Bayes Networks to unpack their return on marketing investment (ROMI).

Q: So in layman’s terms, what can a Bayes Network reveal that a traditional data model can’t do?

A Bayes Network takes a table full of data and translates it into a visual network that’s easier to read than a spreadsheet. In a table, each row is an observation, and each column records the variables that have been observed. This makes it easy to understand how tables can grow quite large and hard to read.

In a Bayes Network, each data variable is represented by a node. Arcs connect nodes that are correlated for specific reasons. Since everything isn’t correlated to everything else, you end up with a visually compact network that reveals strong, non-linear relationships. The marketing world is full of these kinds of relationships—things that are true but aren’t always obvious.

Bayes Networks help make these relationships visible, so they’re great tools for simplifying business complexities. This includes things like influences on sales, which lies at the heart of measuring ROMI.

Q: Can you give us the high-level overview of how your team is using Bayes Networks to better understand ROMI for VSA’s clients?

We are currently putting Bayes Networks into practice for one of the world’s largest B2B2C companies. We’re using them to help our client assess which advertising channels lead to the best sales, which of course helps us determine ROMI. Our client gave us weekly sales and advertising figures, and we also collected a large array of non-advertising factors that could also potentially drive sales. Then we ran two models. The first used batch inference to estimate weekly store sales that could have occurred without advertising. The second estimated weekly store sales that could have directly occurred due to advertising (controlling for store sales in the absence of advertising.)

The results estimated ROMIs for both scenarios and pointed out different relationships between sales and advertising that weren’t so obvious before. With this knowledge, our client can be confident that their advertising dollars will be well spent, driving sales through solid marketing spend allocations.  

Q: Business aside, what are you looking forward to learning at this year’s BayesiaLab Conference?

There are still relatively few practitioners of Bayes Network analysis, so I’m excited to meet other people using this technique. Bayes Networks are applied across a wide spectrum of problems to answer many different statistical questions. And even though we’re all working on different problems, the conference gives us a space where we can speak a common language and compare notes on using the technique.

I’ll be especially pleased if I learn more about using Bayes Networks to better understand Halo and Cannibalization effects, which examine how specific ads can affect sales of other related products. All in all, I’m looking forward to seeing what advances and discoveries in the field this year has in store.

Charles Hammerslough, Ph.D., is Associate Partner and Director of Data Science at VSA. He holds a Ph.D. in Sociology and Demography from Princeton University and is a well-known expert in the field of market potential analysis, predictive analytics for marketing, and direct marketing. More information on his BayesiaLab Conference talk is available here