Discussion - Edward I. George

Outline:

1. Why I Think This Paper Is Wonderful
2. Some Critical Remarks
3. A Crucial Next Step

Why I Think This Paper Is Wonderful

• It pursues the approach of Blattberg & George (1991)

• But goes much further and succeeds! How?
  • It uses a more elaborate model

  • It takes advantage of the fully Bayesian formulation to answer some real questions of interest

When I first read the paper I felt bad about it. We had made a stab at hierarchical modeling of supermarket scanner data and my interest there was mostly in trying to get approved estimators. I spent some time working on theoretical properties of shrinkage estimators and I really wasn't sure how practical or feasible this stuff would be.

What Alan has really done is gone much further and shown just what the potential of hierarchical modeling would be, or illustrate it in this context. What Alan is doing is using a much more elaborate model than we used. I will go through that--it takes time to absorb. It takes advantage of the fully Bayesian formulation to answer real questions of interest. What's nice if you're Bayesian is that you build a comprehensive stochastic model for everything and now you can answer very easily far more questions than you could in certain cumbersome frequentist contexts.

Back