A Conversation with Alan Gelfand

Statistic al Scienc e 2015, V ol. 30, No. 3, 413– 422 DOI: 10.1214 /15-STS521 c  Institute of Mathematical Statisti cs , 2015 A Conversation with Alan Gelfand Bradley P . Ca rlin and Amy H. Herring Abstr act. Alan E. Gelfand wa s b orn April 17, 1945 , in the Bronx, New Y ork. He attended pub lic grade sc ho ols and did his undergradu- ate w ork at what w as th en called Cit y College of New Y ork (CCNY, no w C UNY), excelling at mathematics. He then surp rised and sad- dened his mother by going all the wa y across the coun try to Stanford to graduate sc ho ol, wh ere he completed h is dissertation in 1969 und er the direction of Professor Herb ert S olomon, making h im an academic grandson of Herman Rub in and Harold Hotelling. Alan th en accepted a facult y p osition at the Universit y of Connecticut (UConn) where he w as p romoted to ten ur ed associate professor in 1975 and to full professor in 1980. A few yea rs later he b ecame interested in decision theory , then empirical Ba yes, wh ic h ev entually led to the publication of Gelfand and Smith [ J. Amer. Statist. Asso c. 85 (199 0) 398– 409], the pap er that in tr o duced the Gibbs sampler to most sta tisticians and rev olutionized Bay esian computing. In the mid-1990s, Alan’s in terests turned str ongly to sp atial statistics, leading to fund amen tal con tribu- tions in spatially-v arying co efficien t mo dels, coregionalizat ion, an d sp a- tial b oundary analysis (wom bling). He sp en t 33 y ears on the facult y at UCon n , retiring in 2002 to b ecome the James B. Duke Professor of Statistics and Decision Sciences at Duk e Univ ersit y , servin g as c h air from 2007–2012 . A t Duk e, he has con tin ued his wo rk in spatial metho d- ology while increasing his impact in the en vironmenta l sciences. T o date, h e has p ublished o ve r 260 p ap ers and 6 b o oks; he h as also s u - p ervised 36 Ph.D. diss er tations and 10 p ostdo cs. This in terview w as done j u st p rior to a conference of h is family , academic descendan ts, and colleagues to celebrate his 70th birthday and h is con tributions to statistics whic h to ok place on April 19–22, 2015 at Duk e Univ ersit y . Key wor ds and phr ases: Ba yes, CC NY, Duke , Gibbs sampling, music, spatial statistics, Stanford, UConn. Br ad ley P. Carlin is Pr ofessor and He ad of Biostatistics, Division of Biostatistics, Scho ol of Public He alth, University of Minnesota, MMC 303, 420 Delawar e S t. S.E., Minne ap olis, Minnesota 55455, USA e-mail: br ad@biostat.umn.e du . Amy H. Herring is Asso ciate Chair and Pr ofessor, Dep artment of Biostatistics, UNC Gil lings Scho ol of Public He alth, University of North Car olina at Chap el Hil l, 3104-D McGavr an-Gr e enb er g Hal l, 135 Dauer Drive, Campus Box 7420, Chap el Hil l, North Car olina 27599, USA e-mail: aherring@bios. unc.e du . 1. EARL Y YEARS , CITY COLL EGE, AND ST A NF ORD Am y: Thank y ou v ery muc h for your time and letting us talk with y ou to da y . Alan: I am delight ed! This is an electronic reprint of the o riginal article published by the Institute of Mathematica l Statistics in Statistic al S cienc e , 2015, V ol. 30, No. 3, 41 3 –422 . This reprint differ s from the or iginal in pa gination and t yp ogr aphic detail. 1 2 B. P . CARLIN AND A. H. HERRING Fig. 1. Alan, age 2, F al l 1947. Brad: Y ou were b orn in April 1945 ju st as W orld W ar I I w as ending, w en t to the same Bronx, NY junior high sc ho ol as George Casella, and b o wled and play ed bridge at C CNY in the 1960s. T ell us ab out y our p aren ts, y our c hild ho o d, y our life as a CCNY u ndergrad, and y our path to Stanford for graduate school. Alan: I wa s “to o y oung” all the wa y through sc ho ol. A t that time adm inistrators encouraged chil- dren to skip grades, an d I graduated h igh school and w as a fr eshman in college at 16. Because I was t w o y ears younger than all th e f emales when I wen t off to colleg e, I neve r h ad muc h of a s o cial life until I wen t out we st. I w as really lo oking for a new exp e- rience. In my mind, California wa s the land of milk and honey , and it w as as far aw a y from the Bronx as I could get ! I remem b er driving a wa y , and m y mother w as in tears b ecause sh e th ou ght I was go- ing to disapp ear in to the Paci fic and nev er b een seen again! Brad: What did yo ur father d o? Alan: He w as a C P A (Certified Public Accoun- tan t), and h is fondest desire was to op en Gelfand and Gelfand, C P As. It w as n ev er going to happ en. I pla y ed with n umb er s to o, but not the wa y he did . Brad: I und er s tand th at where y ou grew up in the Bronx wa s a n ice Jewish family neighborh o o d. Alan: Y es, I grew u p in a completely Jewish n eigh- b orho o d: m y elementa ry sc ho ol was 95% J ewish, th e Bronx High Sc ho ol of S cience wa s 90% Jewish , and Cit y College w as 90% Jewish. I though t th e w hole w orld was Jewish ! There we re man y smart kids in NYC, and they sta y ed in NYC, w en t to the s p ecial- ized high sc ho ol, and then attended Cit y College. It was just th e w ay it w as b ac k then, and I never actually considered applying anywhere else. Am y: As a math undergraduate ma jor, what m ade y ou c h o ose graduate school in statistics instead of math? Alan: Th is b o ok [the Hogg and C raig text he u sed at S tanford] is wh at op ened the do or for me; I just fell in lo v e with mathematical statistics. I though t it wa s so elegan t, so cool, all th e distribution the- ory , all the basic probabilit y theory , th e formal infer- ence id eas, ev erything ab out it. I to ok mathematical statistics in the b eginning of m y senior y ear and im- mediately decided it was for me. Brad: W as yo ur mother h eartbrok en ab out y our mo v e w est? Alan: She th ough t it was the end of the world, esp ecially since I had full sc h olarships at Y ale and Colum bia. It was my decision to go west, ev en though m y mother tried to brib e me with a car to sta y on th e east coast! In the end, I mo v ed we st with t w o other C it y College guys; w e ro omed toge ther, so I wa sn’t totally by m yself. Brad: I kn ow y ou are p assionate ab out cars. What did y ou drive to California? Alan: I drov e an American Motors Rambler. Th is car was so slo w, it would do zero to 60 miles p er hour in t wo minutes . It was p ainful . W e limp ed in to P alo Alto, and I r ememb er crossing the Ba y Bridge for the very first time in my life, and suddenly think- ing, “W o w , San F r an cisco.” I really didn ’t know ho w strong a sc ho ol Stanford was, or anything ab out any of the facult y . Ho wev er, arriving in Pa lo Alto in 1965 w as just one of those serendipitous ev en ts. It was an incredi- ble time in the sen se that a lot of things w ere com- ing together then: the Vietnam W ar, the protests, the rev olution in music, p syc hedelia, an d dr ugs. W e though t we were going to c hange the w orld. It didn’t happ en , but bac k then there w as a sp irit that we ma y nev er captur e again. There w as some inn o cence in the country that p r obably is lost forever. I par- ticularly embraced the music. Y ou cannot imagine ho w m any acts I sa w. I sa w the ve ry first pu b lic p er- formances by b oth Stev e Miller and Sant ana, I sa w Janis Joplin several times, and I s a w Jefferson Air- plane probably a dozen times. It w as w onderful. Brad: Y ou’re making me crazy; I play that stuff with my b and! Alan: The face of music just completely c hanged at that p oint . Before then it w as T op 40 r o c k and 3-min ute songs, and then all of a s udden eve rything A CONV ERSA TION WITH ALAN GELF AND 3 Fig. 2. A lan (right) with f ather Ab e, m other F r anc es, and sister Eli ssa, just after Alan ’s high scho ol gr aduation at age 16, Spring 1961. op ened u p; some p eople claim it was the golden age for r o c k and roll. All I kno w is it w as prett y exciting. Am y: When did you fir s t do statistics on a com- puter? Alan: Me? I’m still wait ing for it to happ en! Th is is an embarrassing story . My Ph.D. thesis w as on seriation metho ds: c h r onologica l sequen cing, partic- ularly driven by arc h aeologi cal data. I prov ed sev- eral theorems ab ou t sequencing data from matrix represent ations. Then I had to d o a real example, and. . . I hired someb o dy! Brad: T ell us ab out the statistics departmen t at Stanford in th e 1960s. Alan: The facult y w as quite prestigious. I hold the record for the most courses anyb o dy has ever tak en from C h arles Stein: 11 quarters. I also to ok the v ery first course that Brad Efron taugh t. He fin ished his Ph.D. in s p ring of 1965 and taugh t that fall. I had the fir st yea r of mathematica l statist ics from him. He wa s ins p irational, and I still ha v e th e notes from that yea r with h im. I recall Kai Lai Chung, who w ould p ound chalk to a fr azzle; he wo uld go through a b o x of chalk in a lecture, in a ro om filled with c halk dust and cigarette s m ok e. His fa vorite expression was, “And w e contin ue to b eat the dead hors e.” Of course, Herb Solomon was m y mento r at Stan- ford, and he was wonderful. He was a pioneer in terms of b ringing external fund ing into the depart- men t. He had connections with all the DOD (US Departmen t of Defense) agencies and with NSF (US National S cience F oundation). He raised so muc h money that he was pro viding summer supp ort for a go o d p ortion of the Stanford facult y . He was not ad- equately appreciated b ecause they did not view him as a th eoretical giant. Ho wev er, he was bringing in money at a time when most statisticians were to o pure to get “dirty” trying to c hase money . After I gradu ated I we nt bac k to Stanford f or t wo decades of summers, p articipating in pro jects with Herb. He was lik e a second father in man y w a ys; he and [his wife] Lottie w ere really v ery go o d to m e. I w as young, and he encour aged me to go to Hillel (a worldwide J ewish campu s organizat ion). I wa s nev er religious, but I w en t to Hillel b ecause of the p ossibilit y of meeting females. Brad: Did it work? Alan: A little bit. Am y: Ho w did you b ecome interested in statistical applications in archaeol ogy and la w? Alan: An arc haeologist at S tanford raised some quan titativ e qu estions with Herb, and the data were in teresting and led to m y thesis. Herb had a r eal passion for la w and jus tice pr oblems, and in the end this area w as muc h , m uch more in teresting to me. At first w e fo cused on j ury decision-making, bu t th en w e explored v arious t yp es of discrimination, jury se- lection problems, and, ev en tually , criminal justice. Later I also did a fair bit of exp ert testimon y , which is a v ery d ifferent game fr om teac hin g and research. Brad: Y ou sound like an applied statistician, y et y ou w ere not doing an y computing! Alan: Life wasn’t predicated on computing. It was a lot of work just to inv ert a 3 × 3 matrix, so y ou just didn’t do those thin gs. I did a lot of analysis with electric calculators. I used to ha v e a Monro e and a F rieden on my desk; these w ere a step b etter th an those mac hines where you turned a crank, a b unch of wheels w ould roll, and you waited f or an answ er to come up. I did n’t really do ve ry muc h programming or wo rking with b ig computing mac h ines. 2. UCONN, BA YES, THE GIBBS SAMPLER, AND BIG DA T A Brad: What led y ou to the Universit y of Connecti- cut (UConn)? Alan: I in terview ed at five p laces: the Stanf ord R e- searc h In stitute, the Universit y of California-Da vis, the Univ ersit y of Maryland, Bell Labs, and UConn. I decided I preferr ed academia. Although UConn w as somewhat sleepy back th en, it was close to my 4 B. P . CARLIN AND A. H. HERRING family in New Y ork, and something ab out New Eng- land was app ealing, so it emerged as the winner. Am y: Based on your C V, y ou wen t u p for ten ure at UConn with just 6 p ap ers: t wo fi rst-authored pa- p ers in the archaeo logical literature, a sole-authored pap er in Communic ations , t w o JA SA pap ers with y our advisor, and a pap er in The Americ an Statis- tician . Ho w confident you w ere feeling ab out th is promotion? Alan: W o w, I really appreciate that question! I think there might ha v e b een a few more pap ers b efore ten ure. I n any ev en t, candid ly , I did n’t ev en kno w what a goo d vita wa s; all I knew wa s that I w as b eing p r o ductive , and it was go o d en ough, but b y tod a y’s standard s it wouldn’t eve n come close to “cutting the mustard.” It was a different time, the bar w as differen t, and the exp ectatio ns just weren’t what they are to da y . I really had somewhat of a w asted y outh. I w as trained to b e a mathematical statistician, but I was nev er mean t to b e a m athematical statisticia n. I tried to prov e th eorems b ecause that’s what y ou do if yo u’re a mathematical statistician, bu t I r eally sp ent a lot of time trying to fi nd my nic he. I w an- dered in to decision theory for a while, whic h led to a tr an s ition to emp irical Ba yes (EB). What even tu- ally emerged w as that I was b orn to b e a stochastic mo deler; it’s just that sto c hastic mo deling and, in particular, hierarchical mo d eling, didn ’t really b los- som until arou n d 1990. I was fortun ate to fi nd the area in wh ic h I could cont ribu te, but for the fir st 20 yea rs of m y career, I w as searc hing. Ho we ve r, for the last 25 y ears it h as b een a wonderful ride, and I feel very fortunate. Brad: Y ou were not “raised” as a Ba ye sian, but y ou b ecame one of the w orld’s b est-kno wn and strongest advocates for the Ba yesian approac h. S o I’m int rigued by y our “con v ersion.” It soun ds lik e it w as not a dramatic “Damascus exp erience” lik e your fello w Stanford grad Jay Kadane, wh o apparently had s uc h an “Oh, w hat a fo ol I ’v e b een” momen t after a few conv ersations with Jimmie S av age . My sense is that your con version w as m uch more lik e an empirical Ba yes-st yle con version, in whic h y ou put y our to e in the water b y writing do wn a mixing dis- tribution, and pr ett y so on y ou find yo urself wish ing y ou could compute p osteriors an d so forth. C an y ou tell us ab out yo ur transition to Bay esian inference? Alan: I was alw ays a lik eliho o dist, and I explored empirical Ba yes b ecause of its connections with deci- sion theory . A t the time I imagined that it w ould b e a nice compr omise. But, of course, it turned out that EB m ade n ob o d y happy: the frequentists didn ’t like it, an d th e Ba ye sians did n ’t either. In EB we sp en t a lot of time trying to figure out ho w to do what Ba yesia ns eve ntuall y could d o without needing the corrections that emp irical Ba ye sians h ad to deve lop in order to capture u ncertain t y . My f ull conv ersion happ en ed in Nottingham. I to ok Adrian S mith’s short course at Bo wlin g Green State Univ ersit y in Oh io, whic h w as orga- nized b y Jim Alb ert. Adrian ga ve a wo nderf ul week of lectures, and at the end of that w eek I ask ed, “An y c hance I could come and sp end a sabb at- ical in Nott ingham?” And he replied, “Oh, sure, come!” He h ad a numerical integrat ion pac k age called Ba yes 4 (Smith et al. ( 1985 )), which could do 6- or 7-dimensional n umerical in tegrations. T hat w as as cutting edge as y ou could p ossibly imagine bac k then: sop h isticated quadrature id eas, p seudo- random in tegration, and a lot of tric ks to address the in tegration problem in Bay esian infer en ce. I wen t there to see if I could us e his soft w are to s olve some empirical Ba y es pr ob lems. It’s a w on d erful story . Adrian p ic k ed m y family up, all four of us, at Gat wic k Airp ort. Adr ian ren ted a rick et y old v an b ecause he nev er owned a car (still do esn’t). The v ery fir s t day in Nott ingham, in the space of 24 hours we mov ed, b ought a car, and w en t to a b arb ecue. Two da ys later I w en t to Nottingham for the fi rst time, and Adrian suggested I r ead T anner and W ong ( 1987 ). W e d ecided to ex- plore v ariations of their metho d. A few w eeks later, Da vid Cla yton, who w as at L eicester at the time, came to Nottingham for a da y , and , in the con text of the T anner and W ong p ap er, he remark ed that w e sh ould read the pap er b y Geman and Geman ( 1984 ) in P AMI (P attern An alysis and Mac hine In - telligence , an IEEE journal). I r emem b er getting a cop y of that p ap er and thinking it was clearly m uc h b etter suited for Ba y esian inference than it was for image reconstruction, which was their conte xt. T he do ors h ad op ened, and we sa w how to go forward. Y ou m ust recall that we were ve ry naiv e bac k then. In those d a ys, only if you were desp erate, as a last resort, w ould yo u use Mon te Carlo metho ds. No w suc h metho ds are often the first to ol, and p eople don’t try to b e analytic very often. Whether that’s go o d or b ad, the landscap e h as certainly c h anged. Brad: A great story . T hough I though t Adrian tossed the Geman and Geman pap er in y our lap, but in fact he p oin ted you to T anner and W ong. A CONV ERSA TION WITH ALAN GELF AND 5 Alan: I t was definitely Da vid Cla yton who con- nected us to Geman and Geman, and Da vid w as underapp reciated in this regard. He had seen that pap er, and the IEEE journ als w ere a literature that few statisticians read bac k then. Also remark able at the time was Mic hael Escobar’s Ph.D. thesis, whic h included what wa s a Gibbs samp ler for imp lemen t- ing Diric hlet pro cess mixing. He had nev er heard of the Gibb s sampler; he just inv ent ed this idea for his particular application. He w as also un derappre- ciated. Am y: On e th ing that’s remark able ab out yo ur tra- jectory is h o w y our pro d u ctivit y and yo ur creativit y ha v e r eally in creased with age. Alan: If y ou lo ok at m y vita, I hav e ab out 260 pa- p ers now, and ma yb e 200 of them are p ost-1990. Tw o things h app ened. One is I foun d something I w as reasonably go o d at, that created a c hallenge, and it led me to b uild interdisciplinary connections. It just op ened up opp ortunities that w ere not there b efore. Second, as yo u b ecome more senior, y ou are able to bu ild a h ierarc hy in y our researc h team, with p ostdo cs, graduate student s, and more junior colla b- orators. Y ou b ecome more pro d u ctiv e b ecause y ou ha v e more p eople helping y ou to get things done. It’s a different situation from b eing a junior researc h er where y ou’re muc h more fo cu sed; these d a ys I’m guiding 10 to 15 differen t pr o jects. Finding the Gibbs s ampler with Adrian and ha v- ing that su ccessful pap er wa s really go o d fortune. Man y smart p eople work really hard an d d on’t get so lu c ky . I w as fortunate to connect with a sem- inal pap er, and the only thin g I can congratulate m yself for is th e fact that I’v e work ed prett y hard for the subsequent 25 years in taking adv antag e of this win do w of opp ortunit y . I’v e b een able to k eep it growing with studen ts and p ostdo cs and building bridges. I t was suc h a fan tastic opp ortunit y , it wa s suc h a goo d fit with whatev er skill set I hav e, so that really is the b est explanation for the delta in pro du ctivit y . Again, m y eyes really op en ed u p a lot from 1990 forward, and , Brad, y ou we re on the cusp of it. I wa s on sabbatical while y ou w ere finishing y our thesis, and I came b ac k with the Gibbs sam- pler, and y ou lost inte rest in the thesis! Y ou wa nte d to get on b oard with the Gibbs s amp ler as m uch as y ou could. Brad: Do y ou agree with Denn is Lindley’s view that Ba y es is going to take o v er the statistical world, or do y ou th ink the w orld is going to con tinue to b e kind of a Bay es-frequenti st h ybrid, with the choice made out of con venience on a problem-by-problem basis? Alan: I th ink we all know Dennis forecasted a 21st Ba yesian cent ury b ecause he thought th at p eo- ple would ju st ev en tually realize that th e Ba yesian paradigm was most natural for inference in science under uncertain t y . But in fact it emerged b ecause it wa s able to handle p roblems that w ere pr eviously inaccessible. Moreo v er, in m y mind, it’s not in equi- librium y et; we ’re still wa tc hing an increase in the use of Ba ye sian metho ds. It may b e very muc h ac- cording to the type of problem th at you’re fo cusing on; sometimes p eople sa y , “Y es, we need to u se hier- arc h ical mo deling and MCMC for this problem, bu t for that one, no, ma yb e w e don’t.” I think usage hasn’t actually stabilized yet , and no w it’s b ecom- ing more complicated with all the big data and d ata science that’s en tering the picture. Ho w will that in- fluence th e f uture of Ba y esian w ork? Altogether, it really is b ecoming a 21st Ba y esian century , bu t pri- marily for reasons different from what Lindley migh t ha v e liked or en visioned. Brad: Statisticians are still largel y frequen tist in wh at they’re doing. I f y ou s ubmit results of a Phase I I I clinical trial to FD A (the US F o o d and Drug Administration), you still need a significant p -v alue; many things ha v en’t c hanged. Y ou’re righ t that there’s a lot of Ba ye s out there; for instance, when you go to amazon.com to buy an Arnold Sc hw arzenegger mo v ie, y ou also see a link to a Jean- Claude V an Damme m o vie. That’s the result of a Ba yesia n inferen ce engine; it h as inferred th at y ou lik e aging Euro-American action hero es. Alan: Inte restingly , scientists in other fields ha v e no p r oblem thin king in terms of a Ba ye sian paradigm. They’re p erfectly comfortable inferr ing what yo u don’t kno w giv en what yo u’ve seen, instead of try- ing to in fer what you migh t see giv en wh at y ou d on’t kno w, w hic h seems bac kw ards. A lot of the challenge is actually more within the statistical communit y itself, and, to date, only certain t yp es of problems seem to demand Ba yesian inference. Brad: MCMC has certainly made the w orld “safer” for b eing Ba ye sian. But are y ou surp rised that nothing has really replaced it? There w as a time when there w as a different Ba yesia n compu ta- tional paradigm ev ery 10 yea rs or so, b ut we’v e b een prett y stable now for 25 y ears. Is a n ew generation of metho ds going to replace the curr en t generation of MCMC to ols? 6 B. P . CARLIN AND A. H. HERRING Fig. 3. L–R: Nick Polson, Br ad Carli n, John Wakefield, Alan, and Di p ak Dey on the frigid b e ach at Pe ˜ nisc ola, Sp ain during the V alencia 4 me eting, April 1991. Alan: Many sa y that the size of data s ets is go- ing to mak e MCMC unusable. I do think some- thing is going to happ en. Th e candidates ha v en’t en tirely emerged: INLA (based on int egrated nested Laplace appr o ximations) is not completely satisfy- ing, ABC (appro ximate Ba yesia n computation) cer- tainly has limitatio ns, and v ariational Ba y es do esn’t allo w enough inference and is really residing pr imar- ily in the machine learning communit y . I don’t see sequen tial algorithms, particle learnin g, an d p article filters emerging to o verta ke MCMC. S till, as data sets k eep getting b igger and bigger, the d a ys when MCMC can still b e utilized are going to b ecome few er and few er, so. . . Brad: But as computers get faster. . . Alan: But the data sets are getting bigger. T here’s no win in that situation. Brad: Dueling asymptotics! Alan: An other concern is what big data is ab out. I think it’s actually a different ph ilosoph y in man y situations from what statistics is ab out. Most of the w ork in my world is hyp othesis-drive n: I’m think in g ab out a problem, ab out a pro cess, learnin g ab out the b eha vior of the pro cess, and I’m trying to build mo dels to und erstand the pro cess, and to h yp othe- size ab out its b eha vior. But a lot of “big d ata anal- ysis” seems to b e searc hing big data sets for s tr uc- ture; you’re n ot hyp othesizing muc h of anything. If statisticia ns con tin ue to b e interested in h yp othesis dev elopmen t and examination, I’m not sure big data metho ds are alw a ys going to b e the answer. Brad: I agree; hyp othesis inv estigation r equires y ou to ha ve to hav e some idea ab out uncertain ty . Y ou hav e to ha ve some sort of v ariance estimate to test a h yp othesis or form a confidence in terv al, whereas the big data guys seem primarily interested in a p oint estimate or m a yb e a ranking. Alan: Statistics m ust main tain its in tellectual gen- esis, which is inference un der uncertain t y , and con- tin ue to argue that suc h inference is v aluable. W e can’t liv e in a purely deductiv e wo rld, we need a formal inferential world with r andomness. W e ha ve to contin ue to train p eople to think th at wa y ab ou t problems. 3. SP A TIAL, APPLICA TIONS, AND THE MO VE TO DUKE Am y: In the late 1 990s y our inte rests tu rned strongly to s patial statistics. Ho w did you b ecome in terested in spatial statistic s, and ho w has it re- tained your atten tion for so long? Alan: A fello w named Mark Ec k er came to UConn for his Ph .D. after earn in g a master’s degree fr om the Universit y of Rho de Island. He came in to my office one day with that classic spatial data s et on scallop catc hes in the A tlanti c Ocean, and ask ed, “What can I d o w ith this s tuff, and what the h ec k is a v ariogram?” I said, “I h a v e no clue.” I had n ev er seen an y spatial data, bu t I though t it wa s int erest- ing. Mark’s qu estion literally op ened the d o or in the spring of 1994, and 20 y ears later I’m still in terested in spatial statistics. It was just another of those un- exp ected b ut fortunate th ings that happ en ed . A t that time, GIS softw are already p ermitted vi- sual o v erla y of spatial data la y ers for making lo v ely pictures and telling n ice descrip tiv e stories, b ut I w an ted to b e able to add an in feren tial en gine to it. So essen tially , Brad, Sudipto Banerjee, and I set ab out creating a fully Bay esian inference engine for spatial analysis; it’s in the b ook and its revision (Banerjee, Carlin and Gelfand ( 2014 )). S tructured dep end ence r eally excited me; I foun d it elegan t that y ou could u se it to learn ab out the b eha vior of an uncounta ble num b er of random v ariables seeing only a finite n umb er of them. I enjoy ed the c h allenges of lo oking at dep endence in t wo dimens ions v ersus de- p end en ce in one dimension (where there’s ord er), and I realize d that I w as muc h more comfortable with inte rp olation than I was with forecasting. I also realized that there were failures with th e customary asymptotics used w ith time series, w h ere y ou let t go to infi nit y; that is n ot what you w an t to do spatially . I got particularly excited ab out the enormous range A CONV ERSA TION WITH ALAN GELF AND 7 of application that w as av aila ble as p eople starting collect ing more and more s patially referenced data. It seemed natural and imp ortan t to tak e adv antag e of spatial r eferencing in bu ilding mo dels. I hav e b een excited to s ee sp atial analysis mo ving f rom the p e- riphery of statistics into the mainstr eam. Brad: Sometimes in academia, in order to get a significan t raise y ou hav e to threaten to lea v e for another p osition. Did y ou eve r thin k ab out lea vin g the Universit y of Con n ecticut? Y ou were there for essen tially your wh ole career; y ou ha v e h ad a second career at Duke, but y ou had a fu ll career at UC on n . Alan: Defin itely , with 33 yea rs at UC onn, yo u are absolutely righ t. UConn was alwa ys v ery go o d to me, and I felt lo yalt y and affection for UConn . Th ey treated me w ell, and I though t the qu alit y of life in New England was go o d, so, honestly , I never really lo ok ed. Brad: There must ha ve b een attempts to lure you a wa y? Alan: O p p ortun ities started b ecoming serious af- ter 1990; all of a sud den I had invitat ions to b ecome a f u ll professor at a n u m b er of differen t places— three or four universities in the UK, and ma yb e h alf a dozen in th e US. How ev er, my kids were still fin- ishing high sc ho ol, and I wasn’t ready to mov e. Du k e had con tacted me in the mid 1990s and again in the late 1990s; finally , b y 2001 I was ready , and in 2002 I made the mo ve . Brad: Gelfand and Smith ( 1990 ) is clearly y our most famous pap er, but wh at other p ap ers on y our CV do you particularly lik e or feel ma y ha ve b een underapp reciated? Alan: Th at’s a r eally go o d question. I’ve b een prett y lucky , and a lot of pap ers ha ve b een well- cited [ note: Al an ’s h-index at the time of writing is 60 ]. Before the spatial work, I lik e an underap p re- ciated pr ior predictiv e mo deling chec ks pap er w ith Dipak Dey , Pan telis Vlac hos and T im Swartz (Dey et al. ( 1998 )). Although m ost of the comm unit y has ab dicated this to p osterior pr ed ictiv e chec ks (e.g., Gelman, Meng and S tern ( 199 6 )), I think prior p re- dictiv e c hec ks hav e adv anta ges. Po sterior pr edictiv e c h ec ks are not based on the mo del that is presumed to generate the data, and they use the data t wice, making it r eally hard to criticize mo dels. Prior p r e- dictiv e c h ec ks av oid that trap, and I don’t under - stand w h y there isn ’t more int erest. I’m revisiting this cur ren tly in the cont ext of p oint patterns to sho w h o w we can b etter assess pattern mo del ade- quacy . I also lik e our h ierarc hical cente ring wo rk for im- pro ving MCMC con v ergence (Gelfand, S ah u and Carlin, 1995 , 1996 ). W e found a nice analytical solu- tion, at least in Gaussian cases, it wa s a demonstra- bly sensible thing to do, and others con tinued along those lin es, includin g Papaspiliopou los, Rob erts and Sk¨ old ( 2007 ). In a different v ein, I thin k coregionalization is r e- ally a lov ely idea. I couldn ’t und erstand why n ob o dy had adop ted it as a general strategy for building m ultiv ariate s patial mo dels. I thought, wh at could b e easier or more n atural than taking linear trans- formations of ind ep endent pr o cesses to create d e- p end ent pr o cesses? The distribution theory w orks out v ery well, and the implemen tations are also easy (Gelfand et al., 2004 ). This id ea is no w at the foun- dation of a lot of spBayes co de. The sp atially-v arying co efficien ts p ap er (Gelfand et al. ( 2003 )) d iscusses the r emark able idea th at, within the Ba yesian framewo rk, you can learn ab out spatially-v arying in tercepts and sp atially-v arying slop es as pr o cesses w ithout ev er actually observ- ing these pro cesses. Other pap ers I really lik e in- clude th e spatial gradien ts work I did with Su dipto (e.g., Banerjee, Gelfand and S irmans, 2003 ) and the w om bling pap ers that subsequ en tly emerged. Am y: What are y ou r f a v orite pap ers fo cused on applications? Alan: I’m particularly proud of the sp ecies dis- tribution mod eling w ork that I did w ith John Si- lander and his group at UConn . W e pr esen ted it at Carn egie Mellon Univ ers it y at a Ba y esian Case Studies meeting. A version of it is in the very first issue of Bayesian Ana lysis (Gelfand et al. ( 2006 )), and a more tec hn ical version (Gelfand et al. ( 2005 )) w as the m ost cited JRSS-C p ap er of th e fi rst decade of the 2000 s. It seems a lot of p eople f rom ecology and biological sciences found it int eresting. A t th at time, I was going to South Africa regularly to col- lab orate. Researc hers were using simple logistic re- gressions for presence/absence, wh ic h was the state of the art in the field then . W e used a hierarc hical mo del to in d uce pro cess features that inv olve trans- formation of land s cap e, suitabilit y of en vironment s, and a v ailabilit y of environmen ts. This allo we d us to explain not only what you did see but what y ou might see, with implications for conserv ation and managemen t. It resonated w ell, and I am still w ork- ing on these problems. Recen tly I ha ve gotten in to demography , whic h led to some nice material with int egral pro jection 8 B. P . CARLIN AND A. H. HERRING mo dels (IPMs), particularly arguin g to emplo y them on the right p opulation scale and again in a fully hierarc hical w a y . Brad: Is this ho w you b egan collab orating with Jim Clark? Alan: Y es, and that’s another interesting story . When I came to inte rview at Duk e, I w en t to talk to Jim ab out collab orations in Duk e’s Nic holas Sc ho ol for th e En vironment. I had a simply wo nderf ul t wo hours with him. He is a real statistic ian with a com- pletely appropr iate secondary app oin tment in our departmen t here at Duke . He imagines and fits more sophisticated h ierarc h ical mo dels than most statis- ticians ev er will. Brad: So you met him the da y you inte rviewed there! Alan: Y es, I thin k w e’v e now reac hed 40 p ap ers and a b o ok together, so it’s b een a wonderful, w on- derful time, and our partnersh ip cont inues to flour- ish. Am y: Y ou h a v e r aised an abs olutely incredib le generation of researc h statisticians. Do you hav e a strategy for identifying the brighte st or m ost promising student s? What is your men toring ph i- losoph y? Alan: I ha v e nev er actually recruited students; I ha v e alwa ys j ust wa ited for studen ts to come to me to express interest in wo rking with me. I’v e gotte n a lot of go o d student s, and my list of “c h ildren” is really prett y strong I think. My primary motiv ation has b een training students for an academic career. I think 2/3 to 3/4 of my students are in academia in some fashion. Not ev eryb o dy trains in th at fash- ion, but probably it just reflects the fact that an academic lifest yle is the b est lifest yle I can imagine. As far as dev eloping stu dent s, an imp ortan t asp ect is appreciation of the many wa ys a mo d ern statisti- cian can con tribute. Y ou can do theory , m etho dol- ogy , mo deling, computation, data analysis, and visu- alizati on. Y ou can contribute on many dim en sions, and in fact w e try to train across them all. The crit- ical th ing I try to emp hasize to students is to fi n d what y ou can really do well and what’s going to re- w ard you b est. On e size do esn ’t fi t all, and we can’t ha v e th e same exp ectations for eve ry student. I also think it’s im p ortan t to encourage fire, p as- sion, and en thusiasm. W e don’t do this simply as a 9 to 5 lifest yle, we do this b ecause we get a lot of satisfaction out of our work. If you’re going to com- mit a 40-y ear career to something lik e this, yo u’ve got to really b e in lo ve with it; yo u don’t just do this to pay the bills. I try to foster a fair bit of in dep end ence in s tu - den ts b ecause I think it’s critical that they learn to generate p r oblems and b u ild their o wn research agenda. I do this esp ecially with p ostdo cs, b ecause they h a ve a t w o y ear w in do w and , when they enter the job mark et, they need to h a v e a firm sense of what they are going to do after they get the job. Also, my st yle has alw ays b een ab ou t a v ailabilit y . A lot of facult y are very stru ctured in the wa y they in teract with their stud en ts, but I’ve b een very fl ex- ible. I sometimes meet with stu den ts at 8 pm ju s t b ecause that’s a go o d time for me, there’s nothing else obligating me, and student s often ha ve “wo rking in the evening” lifest y les. I f a student is str uggling to do something, I lik e to talk ab out it no w in stead of ha ving the stu den t wait for a we ekly time slot. Brad: I rememb er when I wa s at UConn, y ou once said, “Brad, y ou hav e to d ecide what league yo u w an t to p la y in.” The implication was, your w ork do esn’t h a v e to lo ok exactly like mine, or stress mathematics or computing or an y particular to ol. Y ou just h av e to b e in a w ork en vironment where y ou’re going to b e pr o ductiv e and w h ere yo u’re go- ing to b e a s olid “pla ye r” in that “league.” Alan: That’s true, an d there are more leagues a v ailable no w, and m ore wa ys to con trib u te. I th ink that is w hat’s wonderful ab out our field. 4. TRA VEL S TORIES, HOOPS, M USIC, A ND FUTURE PLANS Am y: Y ou’re also famous for your academic trav- els. Are there one or t wo particularly memorable tra v el stories you’d lik e to share? Alan: Obviously the V alencia meetings ha v e al- w a ys b een a highlight, and I wa s f ortu nate to make 7 of the 9 V alencia meetings, and th ere are to o man y stories from th ose to tell. But I w ould like to remi- nisce ab out one of th e earliest professional meetings I w en t to in Europ e. It was when I wa s ju st t wo y ears out of my thesis, 1971, and it was an archae- ology meeting organized by Da v id K endall in Ma- maia, Romania on the Blac k Sea. It wa s a meeting of statisticia ns, applied mathematicians, and archaeo l- ogists. I’d b een to Eu rop e b efore, but I ’d neve r b een to a comm un ist country . There were several remark- able things that happ ened d uring this meeting that will cause it to live in my mind f orever. After arriving in Romania, I lost m y return plane tic ket. A t that time there was no in ternet. I sent a A CONV ERSA TION WITH ALAN GELF AND 9 Fig. 4. A lan (yel low hat) and some of his “desc endants” and other guests at the “G70” (Gelfand 70th Bi rthday Confer enc e) p oster session, Duke University, Durham, NC, April 20, 2015. telegram to my tra ve l agen t bac k in S torrs, CT to see if they could help me get a new tic ke t. Two d a ys later I receiv ed the follo wing telegram: “No infor- mation ab out passenger Blefarx.” B-L-E-F-A-R-X is h ow “Gelfand” was con verted! So I receiv ed no help at all fr om the trav el agency . A t the meeting, they to ok up a collectio n for me to pay for my tic ket . I arriv ed at the airp ort in Buc harest to return to the States, and at the tic k et coun ter the agen t said, “W e ha v e your tic k et.” It apparen tly had b een foun d on the flo or of the terminal in Buc harest airp ort and placed at the c hec k-in counter, waiti ng for me. Iron- ically , C . R. Rao had also lost his tic k et, and he and I were b oth part of the collection tak en up at the meeting. It w as the first time I had ev er met Profes- sor Rao. A t this same meeting there was a famous Stan- ford ecologist-sta tistician named Lu igi Lu ca Ca v alli- Sforza. Luigi Luca came to the meeting with his wife, who wa s of noble Italian birth. Th e outing for the conference was to tak e a t w o-hour b us ride up to the Dan ub e d elta, where w e wo uld get on a b oat to trav el do wn the rive r. The bu s w as lea ving at 8 o’clo ck in the morning. I arriv ed at r oughly 7:55 and s aid to the bus dr iv er, “I d idn’t hav e time to eat an ything; can I r un in and grab something?” He said, “No problem, no problem.” But when I came bac k out, the bu s wa s gone. I t tur ned out that also missing the bus were Luigi Luca and his wife, who w ere complaining bitterly . Th irt y minutes later a Mercedes 600 stretc h limousine sho ws up w ith Ro- manian fl ags on all 4 f enders. Luigi Lu ca and his wife climb in. . . and I clim b in with them; we are going to catc h up! W e w en t barr eling along on these small roads at 120 km/hour, and after a bit we wen t righ t past th e bu s w e w ere supp osed to b e on. W e w ound up at this cafe near our destination, ab out 45 min utes ahead of the bus. I will n ev er forget that outrageous ride in a Mercedes limousine on th e b ac k roads of rural Romania. Brad: P erhaps n o s tatisticia n in th e U.S . is b etter equipp ed to answ er this one: Whic h college b ask et- ball pr ogram is b etter: UConn or Duk e? Alan: When I came to Duk e to in terview, one of the fir s t things the Dean said to me w as, “I’ll giv e y ou a nice parking space, but don ’t ask for men’s bask etball tick ets.” I said, “Fine with me!” I ha v e gone to man y Duk e wo men’s basket ball games b e- cause I really lik e the w omen’s game. But after 33 y ears at UConn, I am afraid I’m alwa ys going to b e a UConn b ask etball f an. Brad: Y ou r music collect ion is qu ite famous in some circles; yo u once had something lik e 8000 vinyl record albums. I r emem b er find ing an original V erve pressing of “F r eak Out” b y F r ank Zappa an d the Mothers, an d man y other rare or obscur e albums in y our collect ion. Can yo u tell us more ab out that and y our other passions outside of s tatistics? Alan: I started w ith m usic b ac k in 1955–1956 , with those old, small 42 rp m r ecords with the fat holes; they had n o fi d elit y wh atsoever. I listened to the b eginn in gs of ro ck and r oll—Bill Haley and the Comets, the early Elvis Pr esley stuff, etc. In the early 1970s, I un derwen t a life-c h anging ev en t when I started collec ting j azz. I collecte d j azz for proba- bly 25 yea rs. I h ad gotten to 6500 vin yl jazz albums 10 B. P . CARLIN AND A. H. HERRING Fig. 5. A drian Smith and Alan, pr e-G70 dinner, Durham, NC, April 18, 2015. comprising a fairly v aluable colle ction, rough ly 8000 pieces of vinyl altoge ther. T hen, when I w as coming to North C arolina I had to mak e a d ecision: I was collect ing CD’s by that p oin t, what w as I going to do with all the vinyl? If I b o xed it up, I’d h a ve to fin d a p lace to put it when I got to Duk e, and I didn’t ha v e a p lace. I was afraid if I put it in storage up north, it might sit there forev er; my k id s w ere n ever going to b e in terested in acquiring 6500 pieces of vin yl jazz. So, I sold it to a collecto r in Green w ich Village, New Y ork Cit y . He came up to Connecti- cut, pac k ed the colle ction in to 80 b o xes, put it in the back of a big panel truck, and d r o v e off down the drive wa y . Before the sale I had pulled roughly 500 pieces of vinyl that I thought migh t neve r b e a v ailable on CD, an d of course that was completely incorrect: no w vir tu ally ev erything is a v ailable on- line. I cont inued to collect CDs, so now I’v e got close to 7000 of those dinosaurs . I probably sh ould hav e k ept the vinyl b ecause vinyl is coming bac k, increas- ing in v alue, whereas CDs may nev er come bac k. Ob viously a second passion for me is h o ops; I h av e alw a ys lo v ed bask etball. I ha ve nev er had an y in ter- est in American fo otball, and baseball is a bit on the b oring side ev en though it’s qu ite statistical. I lik e so ccer a lot, b ut b ask etball is the b est game for me. A third passion for me is cars. I’v e alwa ys flir ted with cars, and I read a couple of car magazines ev- ery month. A fast, h igh p erformance car is probably p olitically incorrect, but I still lik e a qu ic k, go o d - handling vehicle . Brad: No w that you are en tering your eighth decade on the planet, what d o es the future hold for y ou? Do y ou ha ve an y ma jor b o oks or other pro jects under w ay? Will you finally op en that used vin yl and CD store? Alan: I’m lo oking forward to selling m y CD col- lection, but nob o dy op en s a storefron t to do th is an ymore; I wa nt to see ho w w ell I can do online using a w ebsite called Discogs, or p erhaps Am azon. I do hav e three imp ortan t futu r e commitments. One is that I will b e an editor for another hand- b o ok with T aylor and F rancis/Chapman an d Hall, a Handb o ok of Envir onmental Statistics . A second thing I’m going to pursue is a pro ject I’ve dev el- op ed called ENMIEP , whic h is the Eur op ean Net- w ork for Mo del-Driv en Inv estigation of Environmen- tal Pro cesses. I ha v e a team throughout all of Eu - rop e, including Italy , Portugal , the UK, Germany , and S pain, and we are trying to find common in ter- ests in en vironment al researc h problems. That will b e imp ortan t b ecause I’m going to b e s p ending a lot of time in Spain (with my wife, Mariasun Beamon te) and elsewhere in Europ e. I n eed to hav e thin gs to do; I am still curious. Ho wev er, after turning 70, and after 46 y ears in the game, m a y b e it’s time to slo w do wn a bit. My third commitmen t is to sp end as m uc h time as I can with the lo v e of m y life. Sh e is simply w onderful, we w ant to b e together, and that has b ecome a priority that is muc h m ore imp ortant than pub lishing a few m ore p ap ers. W e already ha v e tra v el plans for Vienna, Prague, Budap est, China, and Africa. That’s really th e fu ture. Am y and Brad : Alan, thank y ou so muc h f or shar- ing all of this with us to d a y! Happy 70th birthday! Alan: I h a v e thoroughly enjoy ed it. Thank yo u! REFERENCES Banerjee , S. , Carlin, B. P. and Gelf and, A. E. (2014). Hier ar chic al Mo deling and Analysis for Sp atial Data , 2nd ed. Chapman & Hall/CR C Press, Boca Raton, FL. Banerjee , S. , Gelf and, A. E. and Sirmans, C . F. ( 2003). Directional rates of change under spatial pro cess mod els. J. Amer. Statist. Asso c. 98 946–954. MR2041483 Carlin, B. P. , Gelf an d, A . E. an d Smith, A. F. M. (1992). Hierarc hical Bay esian analysis of changepoint problems. Applie d Statistics 41 389–405. Cla yton, D. G. (1991). A Monte Carlo metho d for Ba yesi an inference in frailty mo dels. Biometrics 47 467–485. Dey, D. K. , Gelf and, A. E. , Sw ar tz, T. B. and Vla- chos, P . K. (1998). A sim ulation-intensiv e ap p roac h for chec kin g hierarchical mo dels. TEST 7 325–346. Gelf and, A. E. , Sahu, S. K. and C arlin, B. P. (1995). Efficien t parameterisations for normal linear mixed mo dels. Biometrika 82 479–488. MR1366275 Gelf and, A. E. , S ahu, S. K. and Carlin, B. P. (1996). Ef- ficient parametrizations for generalized linear mixed mo d- els. In Bayesian Statistics, 5 (Alic ante, 1994) ( J. M . Bernardo , J. O. Berger , A. P. Da wid and A. F. M. A CONV ERSA TION WITH ALAN GELF AND 11 Smith , eds.). O xf or d Sci. Publ. 165–180. Oxford U niv. Press, N ew Y ork. MR1425405 Gelf and, A. E. and S mith, A. F. M. (1990). Sampling- based approaches to calculating m arginal den sities. J. Amer . Statist. Asso c. 85 398–409. MR1141740 Gelf and, A. E. , Kim, H.-J. , Sirmans, C. F. and Ba ner- jee, S. (2003). Sp atial mo deling with spatially v arying co- efficien t pro cesses. J. Amer. Statist. Asso c. 98 387–396. MR1995715 Gelf and, A. E. , Schmidt, A. M. , Banerj ee, S. and Sirmans, C. F. (2004). N onstationary multiv ariate pro- cess mo deling through spatially va rying coregionalization. TEST 13 263–312. MR2154003 Gelf and, A. E. , Schmidt, A. M. , Wu, S. , Silande r, J. A. Jr. , La timer, A. and Rebelo, A. G . (2005). Mod - elling sp ecies diversit y through sp ecies level hierarc hical mod elling. J. R. Stat. So c. Ser. C. Appl. Stat. 54 1–20. MR2134594 Gelf and, A. E. , Silander, J. A. Jr. , Wu , S. , La timer, A. , Lewis, P. O. , Rebelo, A. G. and Holder, M. ( 2006). Explaining sp ecies distribution patterns through hierarchi- cal mo deling. Bayesian Ana l. 1 41–92 . MR222736 2 Gelman, A. , Meng, X.-L. and Stern, H. (1996). Pos terior predictive assessmen t of model fitness via realized discrep- ancies. Statist. Sinic a 6 733–807. MR1422404 Geman, S. and Geman, D. (1984). Sto chasti c relaxation, Gibbs distributions and the Bay esian restoration of images . IEEE T r ans. on Pattern A nalysis and Machine I ntel li genc e 6 721–741. P ap aspiliopoulos, O. , R ober ts, G. O. and Sk ¨ old, M. (2007). A general framew ork for th e parametrization of hi- erarc hical mo d els. Statist. Sci. 22 59–73. MR2408661 Smith, A. F. M. , S kene, A. M. , Sh a w, J. E. H. , Na y- lor, J. C. and Dransfie ld, M. (1985). The implemen- tation of the Ba yesian paradigm. Comm. Statist. The ory Metho ds 14 1079–1102. MR0797634 T anner, M. A. and Wong, W. H. (1987). The calculation of posterior distributions by d ata augmentatio n. J. Amer. Statist. Asso c. 82 528–550. MR0898357