aa r X i v : . [ s t a t . O T ] S e p Statistical Science (cid:13)
Institute of Mathematical Statistics, 2015
A Conversation with Alan Gelfand
Bradley P. Carlin and Amy H. Herring
Abstract.
Alan E. Gelfand was born April 17, 1945, in the Bronx,New York. He attended public grade schools and did his undergradu-ate work at what was then called City College of New York (CCNY,now CUNY), excelling at mathematics. He then surprised and sad-dened his mother by going all the way across the country to Stanfordto graduate school, where he completed his dissertation in 1969 underthe direction of Professor Herbert Solomon, making him an academicgrandson of Herman Rubin and Harold Hotelling. Alan then accepteda faculty position at the University of Connecticut (UConn) wherehe was promoted to tenured associate professor in 1975 and to fullprofessor in 1980. A few years later he became interested in decisiontheory, then empirical Bayes, which eventually led to the publicationof Gelfand and Smith [
J. Amer. Statist. Assoc. (1990) 398–409],the paper that introduced the Gibbs sampler to most statisticians andrevolutionized Bayesian computing. In the mid-1990s, Alan’s intereststurned strongly to spatial statistics, leading to fundamental contribu-tions in spatially-varying coefficient models, coregionalization, and spa-tial boundary analysis (wombling). He spent 33 years on the facultyat UConn, retiring in 2002 to become the James B. Duke Professorof Statistics and Decision Sciences at Duke University, serving as chairfrom 2007–2012. At Duke, he has continued his work in spatial method-ology while increasing his impact in the environmental sciences. Todate, he has published over 260 papers and 6 books; he has also su-pervised 36 Ph.D. dissertations and 10 postdocs. This interview wasdone just prior to a conference of his family, academic descendants,and colleagues to celebrate his 70th birthday and his contributions tostatistics which took place on April 19–22, 2015 at Duke University. Key words and phrases:
Bayes, CCNY, Duke, Gibbs sampling, music,spatial statistics, Stanford, UConn.
Bradley P. Carlin is Professor and Head ofBiostatistics, Division of Biostatistics, School of PublicHealth, University of Minnesota, MMC 303, 420Delaware St. S.E., Minneapolis, Minnesota 55455, USAe-mail: [email protected]. Amy H. Herring isAssociate Chair and Professor, Department ofBiostatistics, UNC Gillings School of Public Health,University of North Carolina at Chapel Hill, 3104-DMcGavran-Greenberg Hall, 135 Dauer Drive, CampusBox 7420, Chapel Hill, North Carolina 27599, USAe-mail: [email protected].
1. EARLY YEARS, CITY COLLEGE, ANDSTANFORD
Amy:
Thank you very much for your time andletting us talk with you today.
Alan:
I am delighted!
This is an electronic reprint of the original articlepublished by the Institute of Mathematical Statistics in
Statistical Science , 2015, Vol. 30, No. 3, 413–422. Thisreprint differs from the original in pagination andtypographic detail. B. P. CARLIN AND A. H. HERRING
Fig. 1.
Alan, age 2, Fall 1947.
Brad:
You were born in April 1945 just as WorldWar II was ending, went to the same Bronx, NYjunior high school as George Casella, and bowledand played bridge at CCNY in the 1960s. Tell usabout your parents, your childhood, your life as aCCNY undergrad, and your path to Stanford forgraduate school.
Alan:
I was “too young” all the way throughschool. At that time administrators encouraged chil-dren to skip grades, and I graduated high school andwas a freshman in college at 16. Because I was twoyears younger than all the females when I went offto college, I never had much of a social life untilI went out west. I was really looking for a new expe-rience. In my mind, California was the land of milkand honey, and it was as far away from the Bronxas I could get! I remember driving away, and mymother was in tears because she thought I was go-ing to disappear into the Pacific and never been seenagain!
Brad:
What did your father do?
Alan:
He was a CPA (Certified Public Accoun-tant), and his fondest desire was to open Gelfandand Gelfand, CPAs. It was never going to happen.I played with numbers too, but not the way he did.
Brad:
I understand that where you grew up in theBronx was a nice Jewish family neighborhood.
Alan:
Yes, I grew up in a completely Jewish neigh-borhood: my elementary school was 95% Jewish, theBronx High School of Science was 90% Jewish, andCity College was 90% Jewish. I thought the wholeworld was Jewish! There were many smart kids inNYC, and they stayed in NYC, went to the special-ized high school, and then attended City College. It was just the way it was back then, and I neveractually considered applying anywhere else.
Amy:
As a math undergraduate major, what madeyou choose graduate school in statistics instead ofmath?
Alan:
This book [the Hogg and Craig text he usedat Stanford] is what opened the door for me; I justfell in love with mathematical statistics. I thoughtit was so elegant, so cool, all the distribution the-ory, all the basic probability theory, the formal infer-ence ideas, everything about it. I took mathematicalstatistics in the beginning of my senior year and im-mediately decided it was for me.
Brad:
Was your mother heartbroken about yourmove west?
Alan:
She thought it was the end of the world,especially since I had full scholarships at Yale andColumbia. It was my decision to go west, eventhough my mother tried to bribe me with a car tostay on the east coast! In the end, I moved west withtwo other City College guys; we roomed together, soI wasn’t totally by myself.
Brad:
I know you are passionate about cars. Whatdid you drive to California?
Alan:
I drove an American Motors Rambler. Thiscar was so slow, it would do zero to 60 miles perhour in two minutes . It was painful . We limped intoPalo Alto, and I remember crossing the Bay Bridgefor the very first time in my life, and suddenly think-ing, “Wow, San Francisco.” I really didn’t know howstrong a school Stanford was, or anything about anyof the faculty.However, arriving in Palo Alto in 1965 was justone of those serendipitous events. It was an incredi-ble time in the sense that a lot of things were com-ing together then: the Vietnam War, the protests,the revolution in music, psychedelia, and drugs. Wethought we were going to change the world. It didn’thappen, but back then there was a spirit that wemay never capture again. There was some innocencein the country that probably is lost forever. I par-ticularly embraced the music. You cannot imaginehow many acts I saw. I saw the very first public per-formances by both Steve Miller and Santana, I sawJanis Joplin several times, and I saw Jefferson Air-plane probably a dozen times. It was wonderful.
Brad:
You’re making me crazy; I play that stuffwith my band!
Alan:
The face of music just completely changedat that point. Before then it was Top 40 rock and3-minute songs, and then all of a sudden everything
CONVERSATION WITH ALAN GELFAND Fig. 2.
Alan (right) with father Abe, mother Frances, andsister Elissa, just after Alan’s high school graduation at age16, Spring 1961. opened up; some people claim it was the golden agefor rock and roll. All I know is it was pretty exciting.
Amy:
When did you first do statistics on a com-puter?
Alan:
Me? I’m still waiting for it to happen! Thisis an embarrassing story. My Ph.D. thesis was onseriation methods: chronological sequencing, partic-ularly driven by archaeological data. I proved sev-eral theorems about sequencing data from matrixrepresentations. Then I had to do a real example,and. . . I hired somebody!
Brad:
Tell us about the statistics department atStanford in the 1960s.
Alan:
The faculty was quite prestigious. I hold therecord for the most courses anybody has ever takenfrom Charles Stein: 11 quarters. I also took the veryfirst course that Brad Efron taught. He finished hisPh.D. in spring of 1965 and taught that fall. I hadthe first year of mathematical statistics from him.He was inspirational, and I still have the notes fromthat year with him.I recall Kai Lai Chung, who would pound chalkto a frazzle; he would go through a box of chalkin a lecture, in a room filled with chalk dust andcigarette smoke. His favorite expression was, “Andwe continue to beat the dead horse.”Of course, Herb Solomon was my mentor at Stan-ford, and he was wonderful. He was a pioneer interms of bringing external funding into the depart-ment. He had connections with all the DOD (USDepartment of Defense) agencies and with NSF (US National Science Foundation). He raised so muchmoney that he was providing summer support for agood portion of the Stanford faculty. He was not ad-equately appreciated because they did not view himas a theoretical giant. However, he was bringing inmoney at a time when most statisticians were toopure to get “dirty” trying to chase money.After I graduated I went back to Stanford for twodecades of summers, participating in projects withHerb. He was like a second father in many ways; heand [his wife] Lottie were really very good to me.I was young, and he encouraged me to go to Hillel(a worldwide Jewish campus organization). I wasnever religious, but I went to Hillel because of thepossibility of meeting females.
Brad:
Did it work?
Alan:
A little bit.
Amy:
How did you become interested in statisticalapplications in archaeology and law?
Alan:
An archaeologist at Stanford raised somequantitative questions with Herb, and the data wereinteresting and led to my thesis. Herb had a realpassion for law and justice problems, and in the endthis area was much, much more interesting to me. Atfirst we focused on jury decision-making, but thenwe explored various types of discrimination, jury se-lection problems, and, eventually, criminal justice.Later I also did a fair bit of expert testimony, whichis a very different game from teaching and research.
Brad:
You sound like an applied statistician, yetyou were not doing any computing!
Alan:
Life wasn’t predicated on computing. It wasa lot of work just to invert a 3 ×
2. UCONN, BAYES, THE GIBBS SAMPLER,AND BIG DATA
Brad:
What led you to the University of Connecti-cut (UConn)?
Alan:
I interviewed at five places: the Stanford Re-search Institute, the University of California-Davis,the University of Maryland, Bell Labs, and UConn.I decided I preferred academia. Although UConnwas somewhat sleepy back then, it was close to my
B. P. CARLIN AND A. H. HERRING family in New York, and something about New Eng-land was appealing, so it emerged as the winner.
Amy:
Based on your CV, you went up for tenureat UConn with just 6 papers: two first-authored pa-pers in the archaeological literature, a sole-authoredpaper in
Communications , two
JASA papers withyour advisor, and a paper in
The American Statis-tician . How confident you were feeling about thispromotion?
Alan:
Wow, I really appreciate that question!I think there might have been a few more papersbefore tenure. In any event, candidly, I didn’t evenknow what a good vita was; all I knew was thatI was being productive, and it was good enough, butby today’s standards it wouldn’t even come close to“cutting the mustard.” It was a different time, thebar was different, and the expectations just weren’twhat they are today.I really had somewhat of a wasted youth. I wastrained to be a mathematical statistician, but I wasnever meant to be a mathematical statistician.I tried to prove theorems because that’s what youdo if you’re a mathematical statistician, but I reallyspent a lot of time trying to find my niche. I wan-dered into decision theory for a while, which led toa transition to empirical Bayes (EB). What eventu-ally emerged was that I was born to be a stochasticmodeler; it’s just that stochastic modeling and, inparticular, hierarchical modeling, didn’t really blos-som until around 1990. I was fortunate to find thearea in which I could contribute, but for the first20 years of my career, I was searching. However, forthe last 25 years it has been a wonderful ride, andI feel very fortunate.
Brad:
You were not “raised” as a Bayesian, butyou became one of the world’s best-known andstrongest advocates for the Bayesian approach. SoI’m intrigued by your “conversion.” It sounds like itwas not a dramatic “Damascus experience” like yourfellow Stanford grad Jay Kadane, who apparentlyhad such an “Oh, what a fool I’ve been” momentafter a few conversations with Jimmie Savage. Mysense is that your conversion was much more like anempirical Bayes-style conversion, in which you putyour toe in the water by writing down a mixing dis-tribution, and pretty soon you find yourself wishingyou could compute posteriors and so forth. Can youtell us about your transition to Bayesian inference?
Alan:
I was always a likelihoodist, and I exploredempirical Bayes because of its connections with deci-sion theory. At the time I imagined that it would be a nice compromise. But, of course, it turned out thatEB made nobody happy: the frequentists didn’t likeit, and the Bayesians didn’t either. In EB we spenta lot of time trying to figure out how to do whatBayesians eventually could do without needing thecorrections that empirical Bayesians had to developin order to capture uncertainty.My full conversion happened in Nottingham.I took Adrian Smith’s short course at BowlingGreen State University in Ohio, which was orga-nized by Jim Albert. Adrian gave a wonderful weekof lectures, and at the end of that week I asked,“Any chance I could come and spend a sabbat-ical in Nottingham?” And he replied, “Oh, sure,come!” He had a numerical integration packagecalled Bayes 4 (Smith et al. (1985)), which coulddo 6- or 7-dimensional numerical integrations. Thatwas as cutting edge as you could possibly imagineback then: sophisticated quadrature ideas, pseudo-random integration, and a lot of tricks to address theintegration problem in Bayesian inference. I wentthere to see if I could use his software to solve someempirical Bayes problems.It’s a wonderful story. Adrian picked my familyup, all four of us, at Gatwick Airport. Adrian renteda rickety old van because he never owned a car(still doesn’t). The very first day in Nottingham,in the space of 24 hours we moved, bought a car,and went to a barbecue. Two days later I went toNottingham for the first time, and Adrian suggestedI read Tanner and Wong (1987). We decided to ex-plore variations of their method. A few weeks later,David Clayton, who was at Leicester at the time,came to Nottingham for a day, and, in the contextof the Tanner and Wong paper, he remarked thatwe should read the paper by Geman and Geman(1984) in PAMI (Pattern Analysis and Machine In-telligence, an IEEE journal). I remember getting acopy of that paper and thinking it was clearly muchbetter suited for Bayesian inference than it was forimage reconstruction, which was their context. Thedoors had opened, and we saw how to go forward.You must recall that we were very naive back then.In those days, only if you were desperate, as a lastresort, would you use Monte Carlo methods. Nowsuch methods are often the first tool, and peopledon’t try to be analytic very often. Whether that’sgood or bad, the landscape has certainly changed.
Brad:
A great story. Though I thought Adriantossed the Geman and Geman paper in your lap,but in fact he pointed you to Tanner and Wong.
CONVERSATION WITH ALAN GELFAND Alan:
It was definitely David Clayton who con-nected us to Geman and Geman, and David wasunderappreciated in this regard. He had seen thatpaper, and the IEEE journals were a literature thatfew statisticians read back then. Also remarkable atthe time was Michael Escobar’s Ph.D. thesis, whichincluded what was a Gibbs sampler for implement-ing Dirichlet process mixing. He had never heardof the Gibbs sampler; he just invented this idea forhis particular application. He was also underappre-ciated.
Amy:
One thing that’s remarkable about your tra-jectory is how your productivity and your creativityhave really increased with age.
Alan:
If you look at my vita, I have about 260 pa-pers now, and maybe 200 of them are post-1990.Two things happened. One is I found somethingI was reasonably good at, that created a challenge,and it led me to build interdisciplinary connections.It just opened up opportunities that were not therebefore. Second, as you become more senior, you areable to build a hierarchy in your research team, withpostdocs, graduate students, and more junior collab-orators. You become more productive because youhave more people helping you to get things done. It’sa different situation from being a junior researcherwhere you’re much more focused; these days I’mguiding 10 to 15 different projects.Finding the Gibbs sampler with Adrian and hav-ing that successful paper was really good fortune.Many smart people work really hard and don’t getso lucky. I was fortunate to connect with a sem-inal paper, and the only thing I can congratulatemyself for is the fact that I’ve worked pretty hardfor the subsequent 25 years in taking advantage ofthis window of opportunity. I’ve been able to keepit growing with students and postdocs and buildingbridges. It was such a fantastic opportunity, it wassuch a good fit with whatever skill set I have, sothat really is the best explanation for the delta inproductivity. Again, my eyes really opened up a lotfrom 1990 forward, and, Brad, you were on the cuspof it. I was on sabbatical while you were finishingyour thesis, and I came back with the Gibbs sam-pler, and you lost interest in the thesis! You wantedto get on board with the Gibbs sampler as much asyou could.
Brad:
Do you agree with Dennis Lindley’s viewthat Bayes is going to take over the statistical world,or do you think the world is going to continue to bekind of a Bayes-frequentist hybrid, with the choice made out of convenience on a problem-by-problembasis?
Alan:
I think we all know Dennis forecasted a21st Bayesian century because he thought that peo-ple would just eventually realize that the Bayesianparadigm was most natural for inference in scienceunder uncertainty. But in fact it emerged becauseit was able to handle problems that were previouslyinaccessible. Moreover, in my mind, it’s not in equi-librium yet; we’re still watching an increase in theuse of Bayesian methods. It may be very much ac-cording to the type of problem that you’re focusingon; sometimes people say, “Yes, we need to use hier-archical modeling and MCMC for this problem, butfor that one, no, maybe we don’t.” I think usagehasn’t actually stabilized yet, and now it’s becom-ing more complicated with all the big data and datascience that’s entering the picture. How will that in-fluence the future of Bayesian work? Altogether, itreally is becoming a 21st Bayesian century, but pri-marily for reasons different from what Lindley mighthave liked or envisioned.
Brad:
Statisticians are still largely frequentistin what they’re doing. If you submit results of aPhase III clinical trial to FDA (the US Food andDrug Administration), you still need a significant p -value; many things haven’t changed. You’re rightthat there’s a lot of Bayes out there; for instance,when you go to amazon.com to buy an ArnoldSchwarzenegger movie, you also see a link to a Jean-Claude Van Damme movie. That’s the result of aBayesian inference engine; it has inferred that youlike aging Euro-American action heroes. Alan:
Interestingly, scientists in other fields haveno problem thinking in terms of a Bayesian paradigm.They’re perfectly comfortable inferring what youdon’t know given what you’ve seen, instead of try-ing to infer what you might see given what you don’tknow, which seems backwards. A lot of the challengeis actually more within the statistical communityitself, and, to date, only certain types of problemsseem to demand
Bayesian inference.
Brad:
MCMC has certainly made the world“safer” for being Bayesian. But are you surprisedthat nothing has really replaced it? There was atime when there was a different Bayesian computa-tional paradigm every 10 years or so, but we’ve beenpretty stable now for 25 years. Is a new generationof methods going to replace the current generationof MCMC tools?
B. P. CARLIN AND A. H. HERRING
Fig. 3.
L–R: Nick Polson, Brad Carlin, John Wakefield,Alan, and Dipak Dey on the frigid beach at Pe˜niscola, Spainduring the Valencia 4 meeting, April 1991.
Alan:
Many say that the size of data sets is go-ing to make MCMC unusable. I do think some-thing is going to happen. The candidates haven’tentirely emerged: INLA (based on integrated nestedLaplace approximations) is not completely satisfy-ing, ABC (approximate Bayesian computation) cer-tainly has limitations, and variational Bayes doesn’tallow enough inference and is really residing primar-ily in the machine learning community. I don’t seesequential algorithms, particle learning, and particlefilters emerging to overtake MCMC. Still, as datasets keep getting bigger and bigger, the days whenMCMC can still be utilized are going to becomefewer and fewer, so. . .
Brad:
But as computers get faster. . .
Alan:
But the data sets are getting bigger. There’sno win in that situation.
Brad:
Dueling asymptotics!
Alan:
Another concern is what big data is about.I think it’s actually a different philosophy in manysituations from what statistics is about. Most of thework in my world is hypothesis-driven: I’m thinkingabout a problem, about a process, learning aboutthe behavior of the process, and I’m trying to buildmodels to understand the process, and to hypothe-size about its behavior. But a lot of “big data anal-ysis” seems to be searching big data sets for struc-ture; you’re not hypothesizing much of anything. Ifstatisticians continue to be interested in hypothesisdevelopment and examination, I’m not sure big datamethods are always going to be the answer.
Brad:
I agree; hypothesis investigation requiresyou to have to have some idea about uncertainty. You have to have some sort of variance estimateto test a hypothesis or form a confidence interval,whereas the big data guys seem primarily interestedin a point estimate or maybe a ranking.
Alan:
Statistics must maintain its intellectual gen-esis, which is inference under uncertainty, and con-tinue to argue that such inference is valuable. Wecan’t live in a purely deductive world, we need aformal inferential world with randomness. We haveto continue to train people to think that way aboutproblems.
3. SPATIAL, APPLICATIONS, AND THEMOVE TO DUKE
Amy:
In the late 1990s your interests turnedstrongly to spatial statistics. How did you becomeinterested in spatial statistics, and how has it re-tained your attention for so long?
Alan:
A fellow named Mark Ecker came to UConnfor his Ph.D. after earning a master’s degree fromthe University of Rhode Island. He came into myoffice one day with that classic spatial data set onscallop catches in the Atlantic Ocean, and asked,“What can I do with this stuff, and what the heckis a variogram?” I said, “I have no clue.” I had neverseen any spatial data, but I thought it was interest-ing. Mark’s question literally opened the door in thespring of 1994, and 20 years later I’m still interestedin spatial statistics. It was just another of those un-expected but fortunate things that happened.At that time, GIS software already permitted vi-sual overlay of spatial data layers for making lovelypictures and telling nice descriptive stories, butI wanted to be able to add an inferential engineto it. So essentially, Brad, Sudipto Banerjee, andI set about creating a fully Bayesian inference enginefor spatial analysis; it’s in the book and its revision(Banerjee, Carlin and Gelfand (2014)). Structureddependence really excited me; I found it elegant thatyou could use it to learn about the behavior of anuncountable number of random variables seeing onlya finite number of them. I enjoyed the challenges oflooking at dependence in two dimensions versus de-pendence in one dimension (where there’s order),and I realized that I was much more comfortablewith interpolation than I was with forecasting. I alsorealized that there were failures with the customaryasymptotics used with time series, where you let t goto infinity; that is not what you want to do spatially.I got particularly excited about the enormous range CONVERSATION WITH ALAN GELFAND of application that was available as people startingcollecting more and more spatially referenced data.It seemed natural and important to take advantageof spatial referencing in building models. I have beenexcited to see spatial analysis moving from the pe-riphery of statistics into the mainstream. Brad:
Sometimes in academia, in order to get asignificant raise you have to threaten to leave foranother position. Did you ever think about leavingthe University of Connecticut? You were there foressentially your whole career; you have had a secondcareer at Duke, but you had a full career at UConn.
Alan:
Definitely, with 33 years at UConn, you areabsolutely right. UConn was always very good tome, and I felt loyalty and affection for UConn. Theytreated me well, and I thought the quality of life inNew England was good, so, honestly, I never reallylooked.
Brad:
There must have been attempts to lure youaway?
Alan:
Opportunities started becoming serious af-ter 1990; all of a sudden I had invitations to becomea full professor at a number of different places—three or four universities in the UK, and maybe halfa dozen in the US. However, my kids were still fin-ishing high school, and I wasn’t ready to move. Dukehad contacted me in the mid 1990s and again in thelate 1990s; finally, by 2001 I was ready, and in 2002I made the move.
Brad:
Gelfand and Smith (1990) is clearly yourmost famous paper, but what other papers on yourCV do you particularly like or feel may have beenunderappreciated?
Alan:
That’s a really good question. I’ve beenpretty lucky, and a lot of papers have been well-cited [ note: Alan’s h-index at the time of writing is60 ]. Before the spatial work, I like an underappre-ciated prior predictive modeling checks paper withDipak Dey, Pantelis Vlachos and Tim Swartz (Deyet al. (1998)). Although most of the community hasabdicated this to posterior predictive checks (e.g.,Gelman, Meng and Stern (1996)), I think prior pre-dictive checks have advantages. Posterior predictivechecks are not based on the model that is presumedto generate the data, and they use the data twice,making it really hard to criticize models. Prior pre-dictive checks avoid that trap, and I don’t under-stand why there isn’t more interest. I’m revisitingthis currently in the context of point patterns toshow how we can better assess pattern model ade-quacy. I also like our hierarchical centering work for im-proving MCMC convergence (Gelfand, Sahu andCarlin, 1995, 1996). We found a nice analytical solu-tion, at least in Gaussian cases, it was a demonstra-bly sensible thing to do, and others continued alongthose lines, including Papaspiliopoulos, Roberts andSk¨old (2007).In a different vein, I think coregionalization is re-ally a lovely idea. I couldn’t understand why nobodyhad adopted it as a general strategy for buildingmultivariate spatial models. I thought, what couldbe easier or more natural than taking linear trans-formations of independent processes to create de-pendent processes? The distribution theory worksout very well, and the implementations are also easy(Gelfand et al., 2004). This idea is now at the foun-dation of a lot of spBayes code.The spatially-varying coefficients paper (Gelfandet al. (2003)) discusses the remarkable idea that,within the Bayesian framework, you can learn aboutspatially-varying intercepts and spatially-varyingslopes as processes without ever actually observ-ing these processes. Other papers I really like in-clude the spatial gradients work I did with Sudipto(e.g., Banerjee, Gelfand and Sirmans, 2003) and thewombling papers that subsequently emerged.
Amy:
What are your favorite papers focused onapplications?
Alan:
I’m particularly proud of the species dis-tribution modeling work that I did with John Si-lander and his group at UConn. We presented itat Carnegie Mellon University at a Bayesian CaseStudies meeting. A version of it is in the very firstissue of
Bayesian Analysis (Gelfand et al. (2006)),and a more technical version (Gelfand et al. (2005))was the most cited
JRSS-C paper of the first decadeof the 2000s. It seems a lot of people from ecologyand biological sciences found it interesting. At thattime, I was going to South Africa regularly to col-laborate. Researchers were using simple logistic re-gressions for presence/absence, which was the stateof the art in the field then. We used a hierarchicalmodel to induce process features that involve trans-formation of landscape, suitability of environments,and availability of environments. This allowed us toexplain not only what you did see but what you might see, with implications for conservation andmanagement. It resonated well, and I am still work-ing on these problems.Recently I have gotten into demography, whichled to some nice material with integral projection
B. P. CARLIN AND A. H. HERRING models (IPMs), particularly arguing to employ themon the right population scale and again in a fullyhierarchical way.
Brad:
Is this how you began collaborating withJim Clark?
Alan:
Yes, and that’s another interesting story.When I came to interview at Duke, I went to talk toJim about collaborations in Duke’s Nicholas Schoolfor the Environment. I had a simply wonderful twohours with him. He is a real statistician with a com-pletely appropriate secondary appointment in ourdepartment here at Duke. He imagines and fits moresophisticated hierarchical models than most statis-ticians ever will.
Brad:
So you met him the day you interviewedthere!
Alan:
Yes, I think we’ve now reached 40 papersand a book together, so it’s been a wonderful, won-derful time, and our partnership continues to flour-ish.
Amy:
You have raised an absolutely incrediblegeneration of research statisticians. Do you havea strategy for identifying the brightest or mostpromising students? What is your mentoring phi-losophy?
Alan:
I have never actually recruited students;I have always just waited for students to come to meto express interest in working with me. I’ve gottena lot of good students, and my list of “children” isreally pretty strong I think. My primary motivationhas been training students for an academic career.I think 2/3 to 3/4 of my students are in academiain some fashion. Not everybody trains in that fash-ion, but probably it just reflects the fact that anacademic lifestyle is the best lifestyle I can imagine.As far as developing students, an important aspectis appreciation of the many ways a modern statisti-cian can contribute. You can do theory, methodol-ogy, modeling, computation, data analysis, and visu-alization. You can contribute on many dimensions,and in fact we try to train across them all. The crit-ical thing I try to emphasize to students is to findwhat you can really do well and what’s going to re-ward you best. One size doesn’t fit all, and we can’thave the same expectations for every student.I also think it’s important to encourage fire, pas-sion, and enthusiasm. We don’t do this simply as a9 to 5 lifestyle, we do this because we get a lot ofsatisfaction out of our work. If you’re going to com-mit a 40-year career to something like this, you’ve got to really be in love with it; you don’t just do thisto pay the bills.I try to foster a fair bit of independence in stu-dents because I think it’s critical that they learnto generate problems and build their own researchagenda. I do this especially with postdocs, becausethey have a two year window and, when they enterthe job market, they need to have a firm sense ofwhat they are going to do after they get the job.Also, my style has always been about availability.A lot of faculty are very structured in the way theyinteract with their students, but I’ve been very flex-ible. I sometimes meet with students at 8 pm justbecause that’s a good time for me, there’s nothingelse obligating me, and students often have “workingin the evening” lifestyles. If a student is strugglingto do something, I like to talk about it now insteadof having the student wait for a weekly time slot.
Brad:
I remember when I was at UConn, you oncesaid, “Brad, you have to decide what league youwant to play in.” The implication was, your workdoesn’t have to look exactly like mine, or stressmathematics or computing or any particular tool.You just have to be in a work environment whereyou’re going to be productive and where you’re go-ing to be a solid “player” in that “league.”
Alan:
That’s true, and there are more leaguesavailable now, and more ways to contribute. I thinkthat is what’s wonderful about our field.
4. TRAVEL STORIES, HOOPS, MUSIC, ANDFUTURE PLANS
Amy:
You’re also famous for your academic trav-els. Are there one or two particularly memorabletravel stories you’d like to share?
Alan:
Obviously the Valencia meetings have al-ways been a highlight, and I was fortunate to make7 of the 9 Valencia meetings, and there are too manystories from those to tell. But I would like to remi-nisce about one of the earliest professional meetingsI went to in Europe. It was when I was just twoyears out of my thesis, 1971, and it was an archae-ology meeting organized by David Kendall in Ma-maia, Romania on the Black Sea. It was a meeting ofstatisticians, applied mathematicians, and archaeol-ogists. I’d been to Europe before, but I’d never beento a communist country. There were several remark-able things that happened during this meeting thatwill cause it to live in my mind forever.After arriving in Romania, I lost my return planeticket. At that time there was no internet. I sent a
CONVERSATION WITH ALAN GELFAND Fig. 4.
Alan (yellow hat) and some of his “descendants” and other guests at the “G70” (Gelfand 70th Birthday Conference)poster session, Duke University, Durham, NC, April 20, 2015. telegram to my travel agent back in Storrs, CT tosee if they could help me get a new ticket. Two dayslater I received the following telegram: “No infor-mation about passenger Blefarx.” B-L-E-F-A-R-Xis how “Gelfand” was converted! So I received nohelp at all from the travel agency. At the meeting,they took up a collection for me to pay for my ticket.I arrived at the airport in Bucharest to return to theStates, and at the ticket counter the agent said, “Wehave your ticket.” It apparently had been found onthe floor of the terminal in Bucharest airport andplaced at the check-in counter, waiting for me. Iron-ically, C. R. Rao had also lost his ticket, and he andI were both part of the collection taken up at themeeting. It was the first time I had ever met Profes-sor Rao.At this same meeting there was a famous Stan-ford ecologist-statistician named Luigi Luca Cavalli-Sforza. Luigi Luca came to the meeting with hiswife, who was of noble Italian birth. The outing forthe conference was to take a two-hour bus ride upto the Danube delta, where we would get on a boatto travel down the river. The bus was leaving at8 o’clock in the morning. I arrived at roughly 7:55and said to the bus driver, “I didn’t have time toeat anything; can I run in and grab something?” Hesaid, “No problem, no problem.” But when I cameback out, the bus was gone. It turned out that alsomissing the bus were Luigi Luca and his wife, whowere complaining bitterly. Thirty minutes later aMercedes 600 stretch limousine shows up with Ro-manian flags on all 4 fenders. Luigi Luca and hiswife climb in. . . and I climb in with them; we are going to catch up! We went barreling along on thesesmall roads at 120 km/hour, and after a bit we wentright past the bus we were supposed to be on. Wewound up at this cafe near our destination, about45 minutes ahead of the bus. I will never forget thatoutrageous ride in a Mercedes limousine on the backroads of rural Romania.
Brad:
Perhaps no statistician in the U.S. is betterequipped to answer this one: Which college basket-ball program is better: UConn or Duke?
Alan:
When I came to Duke to interview, one ofthe first things the Dean said to me was, “I’ll giveyou a nice parking space, but don’t ask for men’sbasketball tickets.” I said, “Fine with me!” I havegone to many Duke women’s basketball games be-cause I really like the women’s game. But after 33years at UConn, I am afraid I’m always going to bea UConn basketball fan.
Brad:
Your music collection is quite famous insome circles; you once had something like 8000 vinylrecord albums. I remember finding an original Vervepressing of “Freak Out” by Frank Zappa and theMothers, and many other rare or obscure albums inyour collection. Can you tell us more about that andyour other passions outside of statistics?
Alan:
I started with music back in 1955–1956,with those old, small 42 rpm records with the fatholes; they had no fidelity whatsoever. I listened tothe beginnings of rock and roll—Bill Haley and theComets, the early Elvis Presley stuff, etc. In theearly 1970s, I underwent a life-changing event whenI started collecting jazz. I collected jazz for proba-bly 25 years. I had gotten to 6500 vinyl jazz albums B. P. CARLIN AND A. H. HERRING
Fig. 5.
Adrian Smith and Alan, pre-G70 dinner, Durham,NC, April 18, 2015. comprising a fairly valuable collection, roughly 8000pieces of vinyl altogether. Then, when I was comingto North Carolina I had to make a decision: I wascollecting CD’s by that point, what was I going to dowith all the vinyl? If I boxed it up, I’d have to finda place to put it when I got to Duke, and I didn’thave a place. I was afraid if I put it in storage upnorth, it might sit there forever; my kids were nevergoing to be interested in acquiring 6500 pieces ofvinyl jazz. So, I sold it to a collector in GreenwichVillage, New York City. He came up to Connecti-cut, packed the collection into 80 boxes, put it inthe back of a big panel truck, and drove off downthe driveway. Before the sale I had pulled roughly500 pieces of vinyl that I thought might never beavailable on CD, and of course that was completelyincorrect: now virtually everything is available on-line. I continued to collect CDs, so now I’ve got closeto 7000 of those dinosaurs. I probably should havekept the vinyl because vinyl is coming back, increas-ing in value, whereas CDs may never come back.Obviously a second passion for me is hoops; I havealways loved basketball. I have never had any inter-est in American football, and baseball is a bit on theboring side even though it’s quite statistical. I likesoccer a lot, but basketball is the best game for me.A third passion for me is cars. I’ve always flirtedwith cars, and I read a couple of car magazines ev-ery month. A fast, high performance car is probablypolitically incorrect, but I still like a quick, good-handling vehicle.
Brad:
Now that you are entering your eighthdecade on the planet, what does the future hold foryou? Do you have any major books or other projectsunder way? Will you finally open that used vinyl andCD store?
Alan:
I’m looking forward to selling my CD col-lection, but nobody opens a storefront to do thisanymore; I want to see how well I can do onlineusing a website called Discogs, or perhaps Amazon.I do have three important future commitments.One is that I will be an editor for another hand-book with Taylor and Francis/Chapman and Hall,a
Handbook of Environmental Statistics . A secondthing I’m going to pursue is a project I’ve devel-oped called ENMIEP, which is the European Net-work for Model-Driven Investigation of Environmen-tal Processes. I have a team throughout all of Eu-rope, including Italy, Portugal, the UK, Germany,and Spain, and we are trying to find common inter-ests in environmental research problems. That willbe important because I’m going to be spending a lotof time in Spain (with my wife, Mariasun Beamonte)and elsewhere in Europe. I need to have things todo; I am still curious. However, after turning 70, andafter 46 years in the game, maybe it’s time to slowdown a bit. My third commitment is to spend asmuch time as I can with the love of my life. She issimply wonderful, we want to be together, and thathas become a priority that is much more importantthan publishing a few more papers. We already havetravel plans for Vienna, Prague, Budapest, China,and Africa. That’s really the future.
Amy and
Brad:
Alan, thank you so much for shar-ing all of this with us today! Happy 70th birthday!
Alan:
I have thoroughly enjoyed it. Thank you!REFERENCES
Banerjee, S. , Carlin, B. P. and
Gelfand, A. E. (2014).
Hierarchical Modeling and Analysis for Spatial Data , 2nded. Chapman & Hall/CRC Press, Boca Raton, FL.
Banerjee, S. , Gelfand, A. E. and
Sirmans, C. F. (2003).Directional rates of change under spatial process models.
J. Amer. Statist. Assoc. Carlin, B. P. , Gelfand, A. E. and
Smith, A. F. M. (1992).Hierarchical Bayesian analysis of changepoint problems.
Applied Statistics Clayton, D. G. (1991). A Monte Carlo method for Bayesianinference in frailty models.
Biometrics Dey, D. K. , Gelfand, A. E. , Swartz, T. B. and
Vla-chos, P. K. (1998). A simulation-intensive approach forchecking hierarchical models.
TEST Gelfand, A. E. , Sahu, S. K. and
Carlin, B. P. (1995).Efficient parameterisations for normal linear mixed models.
Biometrika Gelfand, A. E. , Sahu, S. K. and
Carlin, B. P. (1996). Ef-ficient parametrizations for generalized linear mixed mod-els. In
Bayesian Statistics, 5 (Alicante, 1994) ( J. M.Bernardo , J. O. Berger , A. P. Dawid and
A. F. M.
CONVERSATION WITH ALAN GELFAND Smith , eds.).
Oxford Sci. Publ.
Gelfand, A. E. and
Smith, A. F. M. (1990). Sampling-based approaches to calculating marginal densities.
J.Amer. Statist. Assoc. Gelfand, A. E. , Kim, H.-J. , Sirmans, C. F. and
Baner-jee, S. (2003). Spatial modeling with spatially varying co-efficient processes.
J. Amer. Statist. Assoc. Gelfand, A. E. , Schmidt, A. M. , Banerjee, S. and
Sirmans, C. F. (2004). Nonstationary multivariate pro-cess modeling through spatially varying coregionalization.
TEST Gelfand, A. E. , Schmidt, A. M. , Wu, S. , Silander, J. A.Jr. , Latimer, A. and
Rebelo, A. G. (2005). Mod-elling species diversity through species level hierarchicalmodelling.
J. R. Stat. Soc. Ser. C. Appl. Stat. Gelfand, A. E. , Silander, J. A. Jr. , Wu, S. , Latimer, A. , Lewis, P. O. , Rebelo, A. G. and
Holder, M. (2006). Explaining species distribution patterns through hierarchi-cal modeling.
Bayesian Anal. Gelman, A. , Meng, X.-L. and
Stern, H. (1996). Posteriorpredictive assessment of model fitness via realized discrep-ancies.
Statist. Sinica Geman, S. and
Geman, D. (1984). Stochastic relaxation,Gibbs distributions and the Bayesian restoration of images.
IEEE Trans. on Pattern Analysis and Machine Intelligence Papaspiliopoulos, O. , Roberts, G. O. and
Sk¨old, M. (2007). A general framework for the parametrization of hi-erarchical models.
Statist. Sci. Smith, A. F. M. , Skene, A. M. , Shaw, J. E. H. , Nay-lor, J. C. and
Dransfield, M. (1985). The implemen-tation of the Bayesian paradigm.
Comm. Statist. TheoryMethods Tanner, M. A. and
Wong, W. H. (1987). The calculationof posterior distributions by data augmentation.
J. Amer.Statist. Assoc.82