Arnold Zellner, a classic Bayesian

It took me an embarrassingly long time to fully understand the difference between Bayesian and frequentist views of probability.  I don't think I fully appreciated the Bayesian perspective until I got to listen to Arnold Zellner, who spent his emeritus years at my alma mater, UC Berkeley ARE, when I was in graduate school there. 

I only interacted with him personally a couple of times, and then only very briefly, but was often privy to his comments and questions during seminars.  I found it interesting how he combined passionately strong views with an amazingly kind and gentle demeanor.

From Andrew Gelman's blog:

Steve Ziliak reports:

I [Ziliak] am sorry to share this sad news about Arnold Zellner (AEA Distinguished Fellow, 2002, ASA President, 1991, ISBA co-founding president, all around genius and sweet fellow), who died yesterday morning (August 11, 2010). He was a truly great statistician and to me and to many others a generous and wonderful friend, colleague, and hero. I will miss him. His cancer was spreading everywhere though you wouldn't know it as his energy level was Arnold's typical: abnormally high. But then he had a stroke just a few days after an unsuccessful surgery "to help with breathing" the doctor said, and the combination of events weakened him terribly. He was vibrant through June and much of July, and maintained an 8 hour work day at the office. He never lost his sense of humor nor his joy of life. He died at home in hospice care and fortunately he did not suffer long.

From the official announcement from the University of Chicago:

Arnold began his academic career in 1955 at the University of Washington, Seattle, then moved to the University of Wisconsin, Madison. In 1966 he joined the faculty of Chicago Booth and remained on our faculty until his retirement in 1996. 

Arnold pioneered the field of Bayesian econometrics and was highly regarded by colleagues in his field. He founded the International Society of Bayesian Analysis, served in numerous leadership roles of the American Statistical Association, and received several honorary degrees. His teaching also was recognized by the McKinsey Award for Excellence in Teaching. He remained active after his retirement, continuing to do research, publish papers, and serve as a mentor to students.

Arnold was a distinguished researcher, award-winning teacher, and wonderful colleague. His friendly greeting and gracious manner will be missed. We are fortunate to have had someone as remarkable both professionally and personally as Arnold be a member of our community for so many years. His legacy reminds us what makes this institution such a special place.

Zellner was an old-school Bayesian, focusing on statistical models rather than philosophy. He also straddled the fields of statistics and econometrics, which makes me think of some similarities and differences between these sister disciplines.

To a statistician such as myself, econometrics seems to have two different, nearly opposing, personalities. On one side, econometrics is the study of physics-like laws--supply and demand, utility theory, simultaneous equation models, all sorts of attempts to capture economic behavior with mathematical laws. More recently, some of this focus has moved to agent-based modeling, but it's still the same basic idea to me: serious mathematical modeling. The data are there to understand the fundamental underlying economic processes.

But there's another side to economics, a side that I think has become much more prominent, and that's the anti-modeling approach, the distribution-free methods that try to assume as little as possible (replacing distributional assumptions by second-order stationarity, etc.) to be able to make forecasting or causal claims as robustly as possible.
To the extent that economics is a model-centered field, I think it's naturally Bayesian, and Zellner's methods fit in well. To the extent that economists are interested in robust, non-model-based population inference, I think Bayesian methods are also important--nonparametric methods get complicated quickly, and Bayesian inference is a good way to structure that complexity.

Unfortunately, Bayesian methods have a bad name in some quarters of econometrics because they are associated with subjectivity, which goes against both mainstream threads in econometrics. Whether you're doing physics-influenced modeling or statistics-influenced nonparametrics, you want your inferences to be objective as possible. So both kinds of econometricans can agree to disdain Bayes.

What Zellner showed in his work was how Bayesian methods could be objective, and statistically efficient, and solve problems in econometrics. This was, and is, important.
P.S. I only met Zellner a few times and did not know him personally. My only Zellner story comes from the famous conference at Ohio State University in 1991 on Bayesian Computation via Stochastic Simulation. At one point near the end of the meeting, Zellner stood up and said: Hearing all this important work makes me (Zellner) realize we need to start a crash research program on these methods. And you know what they say about crash research programs. It's like trying to create a baby by getting nine women pregnant and waiting one month. (pause) It might not work, but you'll have a hell of a time trying. (followed by complete stunned silence)

The other thing I remember about Zellner was his statistics seminar at the business school, which I attended a few times during my semester visiting the University of Chigago. No matter who the speaker was, Zeller was always interrupting, asking questions, giving his own views. Not in that aggressive econ-seminar style that we all know and hate; rather Zeller always gave the impression of being a participant in his seminar, one among many who just had the privilege of being able to speak whenever he had a thought--which was often. He was lucky to have the chance to express his statistical thoughts in many venues, and we as a field were lucky to be there to hear him.