Presented by Mark Garratt and Ross Link at the Institute for International Research's Marketing Mix Modeling Conference, March 2001, New York

Beer is one of the most heavily advertised categories. Several firms, including Anheuser-Busch, Miller, Coors, Heineken and Corona concentrate multiple messages against the same 21-34 male target week-in and week-out. The relatively high responsiveness that occurs in infrequently advertised categories, or those with strong informational content (e.g. a new product) is missing. This situation of high clutter generates considerable uncertainty. Attempts to diagnose the uncertainty usually proceed by fractioning the available data into multiple components: by market, campaign, promotional content, etc. These "drill-down" methods usually fall apart because a weak signal for advertising is degraded by reducing the sample size. Attempts to solve this problem by building interactions into the model can lead to overfitting and bad holdout performance.

We present the (disguised) results of random effects and hierarchical Bayes models that allow for market, campaign and market by campaign differentiation of ad response. These models work by shrinking response parameters that have high variance to a higher level estimate that is more reliable. They allow the detail necessary to diagnose ad response (what's working and where amongst the clutter) by constraining those responses to belong to distributions that are estimated at the same time.