He started his professional life as a probabilist and mathematical statistician with Bayesian leanings but became one of the world’s leading applied statisticians and a circumspect frequentist. In his words:

My own experience suggests that neither decision-makers nor their statisticians do in fact have prior probabilities. A large part of Bayesian statistics is about what you would do if you had a prior. For the rest, statisticians make up priors that are mathematically convenient or attractive. Once used, priors become familiar; therefore, they come to be accepted as ‘natural’ and are liable to be used again; such priors may eventually generate their own technical literature… Similarly, a large part of [frequentist] statistics is about what you would do if you had a model; and all of us spend enormous amounts of energy finding out what would happen if the data kept pouring in.

I have draft posts: one is about his book titled as Statistical Models: Theory and Practice and the other is about his article appeared in arXiv:stat not many months ago and now published in the American Statistician (TAS). In my opinion, both would help astronomers lowering the barrier of theoretical statistics, Bayesian and frequentist methods alike. I blame myself for delaying these posts. Carrying on one’s legacy, I believe, is easier while the person is alive.

Terence’s Stuff: A Rose… by Terry Speed

In recent years I have found myself increasingly irked by the way in which computer scientists, physicists and others have systematically renamed concepts from our world, either to mystify or appropriate them — or (which is worse?) because they were unaware that we had already invented and named them ourselves.

From the context above, I don’t think physicists want to applaud at the following example from the article:

…, logs of probabilities are potential, Hamiltonian or energy functions.

Yet, I’m sympathetic to his uncomfortable feeling when reading astronomical literature from time to time (my inexperience cannot match his deep insights, though).

Although a quintessentially statistical notion, my impression is that PCA has always been more popular with non-statisticians. Of course we love to prove its optimality properties in our courses, and at one time the distribution theory of sample covariance matrices was heavily studied.

…but who could not feel suspicious when observing the explosive growth in the use of PCA in the biological and physical sciences and engineering, not to mention economics?…it became the analysis tool of choice of the hordes of former physicists, chemists and mathematicians who unwittingly found themselves having to be statisticians in the computer age.

My initial theory for its popularity was simply that they were in love with the prefix eigen-, and felt that anything involving it acquired the cachet of quantum mechanics, where, you will recall, everything important has that prefix.

He gave the following eigen-’s: eigengenes, eigenarrays, eigenexpression, eigenproteins, eigenprofiles, eigenpathways, eigenSNPs, eigenimages, eigenfaces, eigenpatterns, eigenresult, and even eigenGoogle.

How many miracles must one witness before becoming a convert?…Well, I’ve seen my three miracles of exploratory data analysis, examples where I found I had a problem, and could do something about it using PCA, so now I’m a believer.

No need to mention that astronomers explore data with PCA and utilize eigen- values and vectors to transform raw data into more interpretable ones.

Over the years I’ve had many strong likes and dislikes for the various parts our subject. At different times I have confidently asserted this or that topic to be useless, wrong-headed, stupid, superficial, impossible, inappropriate, irrelevant, phony, boring, or finished. I’ve been in love with sufficiency and hated cluster analysis. I thought the theory of games was elegant, while that of linear models lacked style. Group theory and invariance were fascinating to me, while maximum likelihood seemed mundane. Coordinate-free was the way to go, explicit parameters were to be avoided. Brownian theory was hot, sampling theory was not. The Markov property was natural, the mixing property artificial. Category theory was pure, applied probability wasn’t applied. Rao-Blackwellizing was cool, the delta method left me cold. Exact results were good, approximate ones bad. Scientific applications were beautiful, technological applications were ugly. Frequentist inference was objective, Bayesian inference subjective. And so it went on. My view was that means were to be avoided; extremes were the place to be.

I’ve noticed another trend over my career. For decades I have jealously watched other people work on fascinating, complicated things — data, questions, contexts, models, methods and theory — leading them to fame and fortune, while I have been working on uninteresting, simple things, condemning myself to obscurity and poverty. I hasten to add that my things are always very interesting to me, and sometimes quite complicated too, just not to others. But the strange thing is that as time passed, many of those dimly-recalled, fascinating, complicated things from the past that others worked on, turn out to be just what I needed in order to answer a question at a later date. I’ve been behind the times, but, at least in some cases, I’ve caught up eventually.

[Once I quoted an article by Prof. Terry Speed in IMS Bulletin: Data-Doctors. Reading his columns in the IMS Bulletin provides me an opportunity to reflect who I am as a statistician and some guidance for treating data. Although his ideas were not from astronomy or astronomical data analysis, I often find his thoughts and words can be shared with astronomers.]

“What’s the question (this model is supposed to help us answer)?” I want to shout. More politely, Samuel Karlin once said “The purpose of models is not to fit the data but to sharpen the question”. Or, John Tukey: “Our focus should be on questions, not models…Models can – and will – get us in deep trouble if we expect them to tell us what the unique proper questions are.

We’ve all heard George Box’s quaqua-versal quotation “All models are wrong, some models are useful”, and I agree with the second half. But where do we find out which models are useful and which aren’t, which are appropriate and which aren’t? You’d think there must be lots of examples; do you know one?