VQMS: Eric Bradlow (Wharton)

Описание: A Bayesian Methodological “Assault” on Common Problems in Marketing Science

In this talk, I will discuss four current papers, each of which addresses and questions a different (seemingly unrelated but actually are!) practical and methodological problem in Bayesian inference. In the first paper, we address a problem that EVERY researcher faces when building a statistical model: “At What Level of Data Granularity Should One Analyze the Data?”. Extant research has shown that selecting different granularities matters, but NOT how to select it. A false common belief, that we “assault”, is that the most granular data predicts the best. In the second paper, we “assault” the premise that hidden Markov models (HMMs) can forecast well, despite their (sometimes) excellent in-sample properties. That is, they can “backcast well but typically not forecast well unless you get lucky”! Since HMMs are a workhorse of both statistical and reinforcement learning models, this has practical implications for researchers. In the third paper, we “assault” the most commonly used (Normal-Normal) Bayesian learning model. An unfortunate (but never pointed out) property of this model is that posterior variance is ALWAYS lower than the prior variance (i.e. learning happens). Unfortunately, there are many practical situations where this isn’t true which we demonstrate through a series of lab experiments. Finally, we address the question of whether conjoint analysis partworths can be “shared” across product categories (i.e. two-for-one conjoint). That is, we “assault” the premise that utilities are category specific.