A Conversation with Nan Laird

Statistic al Scienc e 2015, V ol. 30, No. 4, 582– 596 DOI: 10.1214 /15-STS528 c  Institute of Mathematical Statisti cs , 2015 A Conversation with Nan Laird Louise Ry an Abstr act. Nan McKenzie Laird is the Harve y V. Fineb erg Professor of Biostatistics at the Harv ard T. H. Chan School of Pu blic Health. She has made fund amen tal contributions to statistica l metho ds for longitudinal data analysis, missin g data and meta-analysis. I n addition, she is w idely known for her wo rk in statistical genetics and in statistic al meth- o ds f or p syc hiatric epidemiology . Her 1977 p ap er with Dempster and Rubin on the EM algorithm is among the top 100 most highly cited pap ers in science [ Natur e 524 (2014 ) 550–5 53]. Her applied w ork on medical pr actice errors is widely cited among the medical malpractice comm unity . Nan was b orn in Gainesville, Florida, in 1943. Shortly thereafter, her paren ts Angus McKenzie Laird and Myra Adelia Do yle, mov ed to T a llahassee, Florida, with Nan and her sister Victoria Mell. Nan started college at Rice Un iv ersit y in 196 1, bu t then transferred to the Univ ersit y of Georgia wh ere she receiv ed a B.S. in Statistics in 1969 and wa s elected to Ph i Beta Kappa. After graduation Nan work ed at the Massac husetts Institute of T echnolog y Drap er Lab oratories wh ere sh e w ork ed on Kalman fi ltering for the Ap ollo Man to th e Mo on Program. She enrolled in the Statistics Departmen t at Harv ard Univ ersit y in 1971 and receiv ed h er Ph .D. in 1975. Sh e joined the facult y of Harv ard S c ho ol of Public Health up on receiving her Ph.D., an d remains there as researc h professor, after her retiremen t in 2015. In the 40 y ears that Nan M. Laird sp ent at Harv ard she authored or co-authored o ver 300 pap ers and three b o oks, and mento red n umerous graduate stud en ts, p ostdo ctoral fel- lo w s and junior facult y . According to Go ogle sc holar, h er wo rk has b een cited ov er 111,000 times. Sh e has receiv ed m any a wa rds f or her v aried contributions to statistical science. Some highligh ts include the Samuel S . Wilks Award, her election as F ello w of the Am er ican Asso ciation for the Adv ancement of Science, h er elect ion as F ello w of the American Sta- tistical Asso ciation, the Janet Norwoo d Prize, the F. N. Da vid Aw ard and, most recen tly , the Marvin Zelen Leadership Award in Statisti cal Science. Professor Laird h as served on man y panels and editorial b oards in cluding a National Academ y of Science P anel on Air- liner Cabin Environmen t, whic h led to the eli mination of smoking on airplanes. Professor Laird c h aired the Departmen t of Biostatistics at the Harv ard Sc ho ol of Public Health from 1990 –1999 where she w as the Henry Pic k ering W alcott Professor of Biostatistics. I n this inte rview s he talks ab out her career, including her p assion for mentoring studen ts. She offers s ome helpful advice ab out b alancing work and family life, ac kno wledging the p ow erful encouragement and supp ort that her own family has giv en her ov er the y ears. The in terview w as conducted in Boston, Massac husetts, in Ju ly 2014. A link to Nan’s full CV can b e found at www.hsp h.harv ard.edu /nan-laird/ . Key wor ds and phr ases: EM algorithm, longitudinal d ata, meta-analysis, missing data, statistica l genetics. L ouise Ryan is Distinguishe d Pr ofessor of Statistics at University of T e chnolo gy Sydney, Austr alian R ese ar ch Council Centr e of Exc el lenc e in Mathematic al and Statistic al F r ont iers, UTS Scho ol of Mathematic al and Physic al Scienc es, PO Box 123, Br o adway, NSW 20 07, Austr alia e-mail: louise.m.ryan@uts.e du.au . This is an electronic r eprint o f the original article published by the Institute of Mathematica l Statistics in Statistic al Scienc e , 20 15, V ol. 30, No. 4, 582 –596 . This reprint differ s from the original in pagination and t yp ogr aphic detail. 1 2 L. R Y A N 1. EARL Y LIFE Ryan : T ell us ab ou t y our early life. L air d : I had a conv ent ional southern childhoo d in T al lahassee, Florida. My mother w as a sc ho ol teac her, who stay ed home to care for my sister and m yself. My father was briefly a professor of p olitical science, b ut tur ned to state go vernmen t and p olitics shortly after my sister and I w ere b orn. Mu c h of m y father’s job was stressful and not so inte resting to him; most of his energy w en t to bu ying and im- pro ving small p arcels of rural land. Our life cen tered around the ch ur c h and f amily , gardenin g and my fa- ther’s pro jects. W e liv ed in a small neigh b orho o d in T allahassee, full of kids, and I recall sp en ding our da ys roaming free on the hea vily w o o ded lo cal golf course. W e had a great deal of ind ep endence. My c h ildho o d was a h ap py time and I am still close to man y of my c hildh o o d neighbors and friends. Ryan : W ere y ou in terested in math as a kid? L air d : I alw ays lo ve d math and in h igh school it w as m y fav orite su b ject. My sister and I wen t to the lo cal pub lic high sc h o ol and some of the teac hers w ere ou tstand ing. My c hoice of und ergraduate col- lege, Rice Universit y in Houston, w as a mism atc h. A t th e time it wa s a bit of a bac kw ater; w omen we re not encouraged in mathematics nor in the sciences in general. W omen w ere suc h a m inorit y that I wa s alw a ys the only girl in m y math section. I switc h ed m y ma jor to F rench, b ut then I got married to a f el- lo w Rice studen t, d ropp ed out after m y j u nior yea r, mo v ed to New Y ork Cit y and had m y fir st child, Ric h ard. W e mo ve d to Georgia a few y ears later, and I we nt bac k to und ergraduate studies at the Universit y of Georgia. I wan ted to study something p r actical, so I decided on compu ter science. This wa s in the late sixties, and computers w ere ju st b ecoming main- stream. F ortunately Computer S cience and Statis- tics w ere in the s ame d ep artmen t and the depart- men t c hair, Carl Kossack, told me th at S tatistics w as so m uc h more in teresting, so I studied statis- tics. I to ok a course from Kossac k in S tatistical De- cision Th eory that used Herman Chernoff ’s b o ok. I lov ed this course! Wh at app ealed to me w as th e idea that one could use math form ulas to mak e m un- dane life decisions, lik e whether or not to tak e y our um brella to work. O f course, in retrosp ect I r ealize the course w as really ab out d ecision th eory and not m uc h ab out statistics at all, but still fascinating. It w as s till one of the exp eriences that p ersuaded me to ma jor in statistics. Ryan : What did y ou do at first when yo u gradu- ated from college and what motiv ated you to go to graduate sc ho ol? I und erstand that at s ome p oint y ou had a b r ief stin t as a fashion mo del? L air d : Y es, I did b riefly try mo deling and it w as a lot of fun! But I figur ed it was n ot a go o d long- term career path. After gradu ating from Unive rsity of Georgia w ith a B.S. in statistics, w e m ov ed to Boston and I work ed at MIT with the impr essiv e title of “engineer.” It wa s actually a prett y excit- ing pro j ect at MIT’s Drap er Labs working on tec h - nology to su pp ort the Ap ollo Mo on P r ogram. My role was to deve lop computer programs to test the Kalman filtering u sed for the inertial guidance sys- tem. The main thing I learned w as that to wo rk in a creativ e en vironmen t I needed more education, so after a couple of ye ars, I wen t to grad sc ho ol in statistics at Harv ard. 2. EARL Y DA YS A T HARV ARD Ryan : What was it like b eing a graduate student in the Harv ard statistics departmen t in the 1970s? L air d : I really lo v ed m y y ears in graduate sc ho ol. I had great colleag ues among my fello w stud ents, and I alw a ys felt that I was treated wel l. I did hear stories ab out the lac k of wome n facult y in the de- partmen t and I also heard a r umor that I w as not of- fered fund ing w h en I app lied b ecause it w as though t that I w ould just b ecome a “housewife” after grad- uation. I knew that this (the part ab out my b ecom- ing a h ou s ewife) w as n ot true b ecause I had already tried that. Although I b eliev ed the tru th of the sto- ries, I do not feel that it detracted from m y o v erall p ositiv e exp erience in the department. I felt treate d lik e an ind ividual. That b eing said, the atmosphere in the statistics department could b e difficult at times. It was a ve ry small department and if fac- ult y w ere feuding, it came across to the students. W e had a great department admin istrator, Louisa V an Baalen, wh o wa s influenti al in k eeping th e at- mosphere p ositiv e. Ryan : T ell us more ab out th e p eople in the depart- men t in those d ays? W ere there an y other w omen? What w as the atmospher e lik e? L air d : When I arriv ed in the Departmen t, I w as one of four in the en tering Ph .D. class (the others w ere George Cobb , Diek Kryut and S haron And er- son). P ersi Diaconis had en tered the previous Jan- uary , so he w as taking some courses with our cohort. The facult y were Bill Co chran, F red Mostelle r, Art A CONV ERSA TION WITH NAN L A IRD 3 Fig. 1. Nan L air d wi th her Ph.D. advisor, Pr ofessor Art hur Dempster, in the Harvar d Stat istics Dep artment, cir c a 1975. Dempster and Paul Holland. W e were informed th at since it was Co c hran ’s last y ear of teac hing at Har- v ard, w e had to take all h is courses s traigh ta wa y . So w e to ok Samp le Su rv eys and Design of Exp eri- men ts. I recall th at the fi rst da y of class in Sample Surveys, Co chran sto o d at the b oard and s aid (in a kind of ind istin ct m um ble) that this was the most b oring topic in the world and h e could n ot imagine wh y any one w ould tak e the class. I also recall that I we nt to his office to ask him questions ab out the class. The first time wa s OK, but on the second o c- casion, h e said to me “yo u are getting to b e a bit of a n uisance.” Ha ving Co chran in the department w as r eally sp ecial and eve ryone treated him w ith the utmost resp ect. P aul Holland w as an other facult y memb er wh o w as v ery inv olv ed in teac h in g and engaged with stu- den ts. I r eally admired his teac hing st yle. F red Mosteller was th ere and I work ed on his grant with Da v e Hoaglin, as I d id not ha v e an y fu nding b e- sides w orking as a teac hing assistan t. F r ed w as in- v olv ed in a lot of activities, national panels, teac h ing and collab orating in other departments and schools. He had so man y in terests and pro jects going on; he sho w ed me that statistics was imp ortan t ev erywhere and statisticians could ha v e a big impact in man y fields. Ryan : Th at m ust ha v e b een a great exp erience w orking with F red. T ell us more. L air d : It wa s wonderful w orking with F r ed . He treated all of his gradu ate students like team mem- b ers. Y ou got to kno w ab out the pro jects he wa s w orking on, and sometimes got to w ork w ith him on them. He w as v ery inv olv ed in a lot of activities and not so a v ailable as other facult y , y et it was a privi- lege to w ork w ith him. Lo oking bac k on it, I realize it w as lik e b eing a part of a statistics “lab.” One imp ortant lesson I learned from F r ed wa s ab out criticism. I su bmitted th e fir st pap er f rom m y th esis to Biometrik a. It came bac k asking for c h anges. I th ough t it wa s a negativ e rejectio n so I ju st put it in a draw er. After a while F red aske d me ab out it and I said it w as a n egativ e review. He ask ed to read th e reviews, and on doing so said, “This is not a bad review, this is a go o d review.” Of course, he was right; it w as easy to get the pap er published after minor c hanges. Y ears late r I sa w genuine bad reviews and I w as grateful to F r ed for teac hing me the difference. Ryan : No w I know that Art Dempster w as your Ph.D. sup ervisor. T ell us ab out that. L air d : Art Dempster was the departmen t c hair at the time and w as basically the m ain “go to” advi- sor for all the studen ts until y ou started y our the- sis. F or m e, he w as a great source of stabilit y in the department. He taugh t quite a few of our b a- sic courses and I found him accessible eve n though some studen ts did not. I wan ted to work w ith Art b ecause I lik ed the w ay he thought and I lik ed his approac h to statistics. He was in terested in mo dels and suggested that w e work on mo dels w ith ran- dom effects and v ariance comp onents. He w as not a v ery “hands on” thesis advisor. I r ecall that close to the end , within a few w eeks of finish ing my thesis, I ask ed him a question ab out the last c h apter. He resp ond ed b y saying that h e really had not follo wed m y w ork v ery closely so he d id not thin k h e could help! I got a b ig kic k out of that. Ryan : How did the EM pap er (Dempster, Laird and Rubin ( 1977 )) come ab out? What was it lik e to w ork with Art Dempster and Don Rub in on th at? Did y ou hav e any idea of just h o w famous that p ap er w ould b ecome? L air d : T he idea of the EM algorithm came ab out during the last year of m y graduate studies. I wa s w orking with Art on a rand om effects approac h to smo oth t w o-w a y con tin gency tables. I h ad a log- linear m o del for the coun ts, then pu t a normal prior distribution on th e log parameters with zero mean 4 L. R Y A N Fig. 2. Junior faculty memb ers fr om the Bi ostatistics Dep artment at the Harvar d Scho ol of Public He al th, cir c a 1984. Back r ow: David Harrington, Ste phen L agakos, Colin Be gg; Se c ond to Back r ow: Richar d Gelb er, David Scho enfeld, Michael F eldstein, Thomas L oui s, R eb e c c a Gelman; F r ont r ow: C hristine W aternaux, Cyrus Mehta, Nan L air d, Henry F eldman. and a single v ariance comp onen t. But I w as strug- gling with the problem of how to estimat e the v ari- ance comp onen t—no w it seems trivial, bu t at the time, computations w ere difficu lt and messy (w e w ere still in the p unch card era). I asked Art about estimating the v ariance comp onen t; he said, “Why don’t y ou just us e that subs titution algorithm?” I had neve r heard of the su b stitution alg orithm, but after reading some pap ers and further discuss ion with Art, I wen t off and work ed on the application of this idea to my random effects prob lem. I disco v ered that the maximum likeli ho o d esti- mate of the v ariance comp onent w as a fixed p oin t of the substitution algorithm, where th e lik eliho o d w as obtained b y in tegrating out the random effects. So I r ep orted this bac k to Art. After a charact eris- A CONV ERSA TION WITH NAN L A IRD 5 Fig. 3. Senior faculty m emb ers fr om the Bi ostatistics Dep artment at the H arvar d Scho ol of Publi c He alth when Nan to ok over as Chair in 1990. Back r ow: David Harrington, Stephen L agakos, James War e, Butch Tsiatis, Mar c el lo Pagano, Bernar d R osner. F r ont r ow: L. J. Wei, Nan L air d, Marvin Zelen. tic short silence, Art said, “W ell, this deriv ation will w ork for any exp onential family likelihoo d with an y kind of missing data.” That w as the origin of the EM for me. A t the time Art w as also w orking on missing data with Don Rubin, who w as at the Educational T esting S ervice (ETS). Don receiv ed h is Ph.D. from Harv ard a couple of y ears b efore I came, so I d id not kno w h im. Art suggested that the thr ee of us write a pap er together. Art and I dro v e to Princeton early on to work with Don. That w as the first time I recall meeting Don. I did realize that the EM was a very im p ortant con tribution, and lik ely to b e a famous pap er. But the EM pap er w as to b e one of my first real pap ers and I didn’t wan t to trade on that m y whole life, so I w as pr ett y in ten t on branc hing out into other areas as wel l. Ryan : I remem b er fr om w hen I started there in 1979 th at yo u were still somewhat inv olv ed in the statistics d epartmen t then, but grad u ally y our activ- ities all mo ve d o v er to the Harv ard Sc ho ol of Public Health (HSPH). Ho w d id that come ab out? L air d : I completed m y Ph.D. in 1975 and to ok a p osition as an assistan t professor in the Biostati s- tics Department at HSPH. I n itially , I w as secondary with s tatistics so I taugh t an d advised student s there for a few y ears, but ev entually found my work at HSPH w as very engag ing. Ryan : Y ou w ere there during a time of ma j or c h ange and transition, not just for the departmen t, but I think for the field of biostatistics as well. Y ou w ould h av e b een there when F red Moste ller b ecame c h air and in those f amous (or should I sa y inf a- mous?) da ys when Marvin Z elen arriv ed en masse from Buffalo. What was the Departmen t lik e in those early days? L air d : My first f ew y ears in the Biostatistics De- partmen t w ere inte resting. T here actually were quite a few wome n there when I came—Jane W orcester w as c hair and b oth Marge Drolette and Yv onn e 6 L. R Y A N Bishop we re on the facult y . Bob Reed w as th e other senior faculty p erson. I think most p eople like to do something that matters to the live s of others, but p erh ap s w omen more so and this is why w e are dra wn to app lication areas suc h as health. The departmen t w as pr ett y sleep y in those days, with the facult y ha ving heavy teac hing loads or in vo lv e- men t in col lab orativ e p ro jects. When I joined, Jane W orcester was ve ry close to retiremen t and there w as a lot of sp eculation ab out who would b e the next c h air. Ho ward Hiatt wa s a n ew Dean then, and he w as in terested in strengthening Biostatistics. Much to my su rprise, F red Mostelle r took on the job for a few s hort y ears, and r eally did c h ange things. One thing he did wa s to giv e offices to all of the Bio- statistics facult y . An other wa s to hire a lot of junior facult y— J im W are, Larry Thib o deau, T om Louis, Christine W aternaux were some that I recall. This w as great for me b ecause I had colleag ues wh o w ere con temp orary in age and had similar concerns and career goals. O f course, the really big r ecru itmen t w as Marvin Zelen and his team from Buffalo. Ryan : T ell us more ab out that. I und erstand that he brought along quite a few p eople with h im, in- cluding a num b er of y oung u p-and-coming statisti- cians such as Stev e Laga ko s, Ric hard Gelber, Colin Begg and Reb ecca Gelman, to name a few. I n an in terview for the Department newsletter in 1997, y ou said that some referred to “Marvin’s Baseball T eam.” Marvin’s arr iv al w as also the start of a strong conn ection b et we en the departmen t and the Dana-F arb er Cancer In stitute, wh ic h still exists to- da y . I imagi ne this m u st ha v e b een a time of rather mixed feelings—excitemen t at all the new p ossibil- ities, along with the natural anxiet y ab out c hange. Ho w did the dep artment and the sc ho ol as a whole react? Ho w did the department c h an ge as a r esult of these dev elopment s? L air d : W ell, there were many discussions b efore he came, b oth in the Department and within the Sc ho ol. Y es, th ere w as anxiet y ab out their coming. F acult y outside the department w ere worried that Biostatist ics w ould “tak e o ver the school.” F aculty inside the departmen t w ere w orried ab out their fu- tures and that the rew ard system migh t b e differ- en t. But Marvin’s group came, and they assimilated. Some p eople left, but there w ere losses on b oth sides, and ev en tually gains on eac h side. Marvin to ok o ver as d epartmen t c hair a f ew y ears later. I thought he w as a great c h air b ecause he made it clear that the academic program b elonged to the facult y and he exp ected all of us to participate. He set v ery concrete ac h iev ab le goals for the departmen t, lik e strength- ening th e do ctoral program and obtaining training gran ts. T h ese w ere goals th at ev eryo ne could unite b ehind . He also was very pr incipled, esp ecially ab out salary equity . I r eceiv ed big raises after Marvin to ok o ver as c h air, b ecause m y salary wa s far b elo w that of m y colleagues at the same rank. 3. LONGITUDINAL Ryan : I’m inte rested in y our comment ab out most p eople likin g to work on something that matters. F or man y of u s, that translates to working on ap- plications. But I h a ve alw a ys admired your abilit y to take a sligh tly more abstract approac h and come up with to ols and metho ds that ha v e broad applica- tion. I think it tak es confi dence to b eliev e th at one’s metho ds will matter (as opp osed to solving a sp ecific real w orld problem). The EM is a great example, but so is y our work in the early 80s with Jim W are. In particular, your gro wth cur v e p ap er (Laird and W are ( 1982 )) has h ad ma jor impact. T ell u s more ab out that. L air d : Y es, I h a v e alw a ys b een inte rested in lo ok- ing for a general framew ork, r ather than wo rking on a series of smaller “one-off ” p roblems. In a wa y , w orking on some theory is easier b ecause y ou can sa y ho w the metho d will b eha ve un der sp ecific con- ditions. With app lications inv olving data, so often the d ata refuses to coop erate. The answ ers you get can dep end more on c haracteristics of the data and not the statisti cal metho d. I’v e alw a ys like d d ev el- oping method s th at others can apply . In the case of the gro w th curve mo deling, I had started working with Jim b ecause of his int erest in applying rand om effects mo dels in a longitudinal cohort called the “Six Cities Study .” 1 These w ere the da ys when lon- gitudinal d ata analysis wa s just starting to b ecome p opular. The Six Cities Study had a lot of unbal- anced data and some missin g data as w ell, whereas all th e textb o oks w ere provi ding solutions for bal- anced complete data. Using the rand om effect struc- ture w as a natural and it tied in well with the EM algorithm. 1 This was a prospective cohort mortali ty study in volving 8111 randomly selected residen ts of six U.S. cities. The ob jec- tive w as to estimate th e effects of air p ollution on mortalit y and lung function while controlling for other risk factors such as the individ u als’ smoking status and age (Do ck ery et al. ( 1993 )). A CONV ERSA TION WITH NAN L A IRD 7 Fig. 4. Nan L air d with F r e d Mostel ler (l eft) on the awar ding of his Honor ary Do ctor ate at Harvar d, 1991. Jim and I w ork ed on lots of in teresting pr oblems with colleag ues su c h as T om Louis and Ch ristine W aternaux, and w e also co-advised sev eral do ctoral student s (F ong W ang, Nic k Lange, Ch ristl Donnelly , Rob Stiratelli, Masahiro T akeuc hi) on longitudinal metho ds. I think one reason our 1982 pap er was so p opular was b ecause Jim had a really go o d fix on issues that applied statisticia ns w ere grappling with. Jim has alw ays had a great sense of where a fi eld is headed and has an excellen t eye for go o d problems. Christl Donnelly had an in teresting thesis, sh e ap- plied the mixed mo d el to estimate the sp atial dis- tribution of ai r p ollution, using an appr oac h closely related to Kriging. This col lab oration with Jim w as really imp ortant for me. I learned ab out applying for grant s from the National Institutes of Health (NIH). No one had ev er mentioned gran t supp ort to me b efore, let alone encouraged me to get m y o wn. Jim had b een at the NIH b efore coming to HSP H, so he had go o d idea ab out ho w things w ork ed. I learned so muc h by w orking with Jim. Our families also b ecame close and we sp ent a lot of time toge ther outside of w ork. Ryan : I recall r eading the Laird and W are pap er man y ye ars ago and struggling to un d erstand the w a y you w ere applying the EM. In fact, it was only after attending Xiao-Li Meng’s thesis defense that I realized that you we re actually using a tec hnique that hadn’t b een in ve nte d ye t, namely , the ECM (Meng and Rubin ( 1993 )). L air d : Y es, when w e applied the EM strictly in the mixed m o del regression con text, estimating the r egression and v ariance parameters together, the algorithm did not take adv antag e of the we ll- kno wn closed form for th e regression parameters, whic h is a v ailable if the v ariance parameters are kno wn. W e did something that seemed really nat- ural, namely , estimating the regression parameters through weigh ted least squares, assuming the v ari- ance parameters were known, then estimating the v ariance parameters via an EM, but assuming the regression coefficients were kno wn. This mo difica- tion w ork ed w ell, and w as, as y ou sa y , an example of the y et to b e in v en ted ECM. Ryan : Y ou cont inued w orking on longitudin al data analysis for many y ears. L air d : In ad d ition to the w ork on rep eated mea- sures that Jim and I did , I wa s also in terested in mo dels for rep eated categorical data. I had sev eral student s wh o w ork ed in this area. S tu art Lipsitz and Garrett Fitzmaurice b oth w ork ed on parame- terizatio ns for rep eated categorical outcomes, and the int erplay b et we en the mean effects and the co- v ariance parameters, u s ing the so-called m arginal mo dels. Jarek Harezlak w as one of m y students who v ery ind ep endently found his o wn topic on applied functional an alysis in the longitudinal data setting. I think I learned more from him than he did from me! S kip Olsen lo ok ed at adjus tin g for the baseline measure when the interest is c h ange o v er time. 4. MET A -ANAL YSIS Ryan : I enjo ye d r eadin g y our accoun t in the re- cen t CO PSS b o ok (Laird ( 2014 )) of w riting y our pap ers w ith Reb ecca DerSimonian on meta-analysis ( DerSimonian and Laird , 1983 , 1986 ). T ell us a bit more ab out this. L air d : Y es, the p ap er w ith Reb ecca in Clinic al T ri- als is one of my most cited. T hat has b een a big surpr ise for me, in part b ecause it is suc h a sleep er. The citati ons are no w gro wing exp onen tially , but it came ab out something b y acciden t. F red Mosteller ask ed me to ha v e a lo ok at a pap er (Slac k and Porter ( 1980 )) that combined a series of 23 studies on the impact of coac hing to improv e SA T (Sc holastic Ap- titude T est) scores. T hey concluded that coac hin g help ed, thereb y con tradicting the principle that the SA T measured in n ate abilit y , but w ere advised to consult w ith a biostatistic ian. I read the pap er and w as str uc k b y th e amount of v ariabilit y in th e stud- ies as well as ho w muc h the results v aried by study . There wa s also v ariation in study design: some s tu d- ies had no con trol group, some w ere observ ational, a 8 L. R Y A N few w ere rand omized and a few in volv ed m atc hin g. The degree of cont rol in the coac h ed group seemed to b e inv ersely related to the magnitude of effect. F or example, s ome stud ies only ev aluated coac hed student s, comparing th eir scores b efore and after coac hing. So all this got me thinking that it would b e useful to dev elop a formal statistica l framew ork for the analysis. Reb ecca Dersimonian wa s m y graduate student at the time, w orking on rand om effects mo d- eling. I got her w orking on a mo d eling framework f or meta-analysis that allo wed for study -to-study v ari- ation in effect sizes through the in clusion of random effects. W e reanalyzed Slac k and P orter’s data, con- cluding that any effects of coac hing w ere to o small to b e of practical imp ortance. While our pap er in the Harvar d Educ ation R eview n ev er got that many citatio ns in the scien tific literature, it got a lot media atten tion. A few yea rs later, Reb ecca and I publish ed a follo w-up pap er in Contr ol le d Clinic al T rials , apply- ing the metho d to the clinical trial setting and this pap er has b een v ery highly cited. It is sometimes cited as “the s tandard” approac h to m eta-analysis. A nice feature of our approac h w as that it yielded a closed-form estimator and I think this is one of the reasons why the pap er got so m uc h interest. Another feature is that it shows how to do the analysis ju st b y using data su mmaries. Ryan : Bac k in the 90s, I remem b er talking with y ou ab out a study I w as d oing with Lew Holmes from Massac h usetts General Hospital on a meta- analysis of adverse effects asso ciated with c horionic villus s ampling (CVS ). W e w ant ed yo ur advice on the wisdom of going after individu al lev el data ver- sus w orking with s u mmary statistics b ased on p ub- lished pap er s . L air d : W ell, I don’t really remember the d etails of that. Dan yu L in (Lin and Zeng ( 2010 )) sa ys th at it shouldn’t really matter—if the mo dels are the same, then the results should b e the same w h ether y ou use individual lev el data or data summaries. But I find researc h ers often o v eremphasize the imp ortance of individual lev el data. I think that is a “case stud y” rather than a statistical p ersp ectiv e. I n practice, it is probably more imp ortan t to fo cus on what study c h aracteristics help to explain study-to-study v ari- ation in the observ ed effects. In addition, it can b e v ery hard to get individual lev el data. Ryan : Y es, I agree! In the case of our CVS study , w e did manage to get individual lev el data f or a few studies, bu t they tended to b e the studies where nothing at all interesting wa s happ ening . . . . Per- haps an in v erse of pub lication bias? What are y our though ts ab out trying to tak e accoun t of study qual- it y in the con text of meta-analysis? L air d : In m y pap er with R eb ecca, w e originally stratified by study design an d th en lo ok ed at the distribution of effect sizes. Study design generally explains a lot. 5. PSYCHIA TRIC-EPIDEMIOLOGY Ryan : Let’s talk ab out some of your other wo rk from the 1980s an d 1990s. I kn o w yo u con tinued y our wo rk on missing d ata, but y ou also got very in terested in metho d s for applicatio n in psyc hiatric epidemiology . Ho w did that come ab out? L air d : Ch ristine W aternaux mov ed her main ap- p ointmen t to McLean Hospital, one of Harv ard ’s teac hing hospitals sp ecializing in psy chiatric d isor- ders. Christine wa s interested in submitting a tr ain- ing gran t application to the National Institutes of Men tal Health. The Departmen t of Ep idemiology had one for a long time, bu t it h ad lapsed. Chris- tine and I collab orated with the epid emiology de- partmen t, successfully b ringing in a new training gran t from the National Institute of Men tal Health (NIMH) that sup p orted b oth epidemiology and bio- statistics students and p ostdo ctoral fello ws. Ryan : T ell us m ore ab out how the gran t work ed . It can b e a c h allenge to ha v e a grant that crosses t w o departmen ts lik e that. What were some of the in teresting pr oblems that arose for yo u and your s tu- den ts? L air d : The gran t ga ve us the resources and the con tacts to get inv olve d in a num b er of excit- ing pro jects. What to d o with m ultiple informan ts comes up a lot in the p syc hiatric setting, wh ere as- sessmen ts of stu d y participants ma y b e provided by the ind ividual themselv es, their care-giv er or a rela- tiv e. Garrett Fitzmaurice and Nic k Horton b oth h ad a bac kgroun d in psyc hology which made their work on the NIMH training gran t very v aluable. Garrett’s thesis work on rep eated categorical outcomes ex- tended naturally to this m u ltiple informan t setting. The in teresting fact ab out his app roac h w as that it could apply to either the outcome or to the co v ari- ate, or b oth, and this comes in hand y in the multiple informant s setting as w ell. Nic k Horton also w ork ed on multiple informants and the three of u s , p lu s Stuart L ip sitz and Sharon-Lise Normand, deve lop ed a nice set of to ols for h an d ling multiple informan t A CONV ERSA TION WITH NAN L A IRD 9 Fig. 5. Nan L air d with her famil y in 2009. F r om left to right, Nan, Richar d Hughes, Lily Altst ein and Jo el Altstein. Lil y i s a biostatistician at Massachusetts Gener al Hospital and marrie d to Cory Zigler who is an assistant pr ofessor of Biostatistics at the Harvar d T. H. Chan Scho ol of Public He alth. Richar d i s a che f i n Conc or d, Massach usetts, and Jo el is a R e al Estate Develop er in Cambridge, M assachuset ts. problems, esp ecially for applications in psychiat ry and health services researc h. My backg round with metho ds for longitudinal data led to an inv olv emen t in the S tirling Coun t y Study lo oking at depression and anxiet y , with d ata collect ed in three cross-sectional wa v es as w ell as longitudinally . Nic k Horton and Heather Litman b oth wo rked on Stirlin g Count y Study data as a part of their do ctorates. Kr istin Ja v aras and J im Hud- son were b oth great p ostdo cs wh o were later su p- p orted by th e NIMH tr aining gran t. The psyc hiatric- epidemiology w ork led n aturally to lots of in terest- ing prob lems with missing data, esp ecially drop out. Often, the dr op out w as thought to b e informativ e. 6. MISSING D A T A Ryan : Y es, tell us more ab out that. Giv en y our bac kground with the EM, it seems lik e a natural area for y ou. L air d : Y es, missing data is a big issue in longi- tudinal studies, bu t happ en s just as wel l in cross- sectional studies. Certainly the EM algorithm is a natural w ay to handle missing data in lots of settings, but it can b e problematic w h en miss- ingness is in formativ e. Man y p eople at that time w ere in terested in the so-called selectio n mo del ap- proac h, whic h r equired sp ecifying the probability of drop out, conditional on b oth observ ed and unob- serv ed data. The appr oac h is very app ealing b ecause it allo ws y ou to sp ecify familiar mo d els, bu t it is highly p arametric and it is not really p ossible to test it empirically , nor to u nderstand th e limitations of the mo del. S tu art Bak er and I work ed on a selec- tion mo del for categorical data in the logistic re- gression framew ork, and obtained some interesting results ab out mo del id en tifiabilit y . Bob Glynn also w ork ed on nonignorable nonresp onse in the sample surve y setting. Jo e Hogan w orke d on nonignorable drop outs in longitudinal studies, but using a com- pletely d ifferen t mo deling str ategy , one r elated to the pattern-mixture m o dels. The idea w as to re- v erse the conditioning to lo ok at the distrib ution of observ ed data giv en eac h particular pattern of missingness, and then to inte grate o ver the v arious patterns to get the marginal distribution of int erest. With this appr oac h it is easier to exp ose the mo del assumptions. 7. BECOMING CHAIR Ryan : Y ou b ecame Chair of the Harv ard Biostatis- tics Departmen t in 1990. In a tribute to your ac- 10 L. R Y A N complishmen ts during y our nine y ears as c hair, Jim W are noted ho w muc h the d epartmen t grew und er y our leadership. A lot of inno v ations happ ened in that nine y ear p er io d: the Center for Biostatistics in AIDS Researc h (CBAR) was established u nder Stev e Lagak os’ leadership, the De partment’s minor- it y trainin g p rogram got started. T ell us more ab out those y ears, in particular, what y ou feel w ere y our biggest accomplishmen ts. L air d : I w as excited to b e department c hair. Of course, Marvin Z elen w as a v er y difficult act to fol- lo w , but I h ad a very different p ersonalit y and dif- feren t wa ys of d oing things, so I nev er felt th at I wa s trying to replicate what h e did. Marvin had a lot of little “tric ks” that were part of what made him su c h an effectiv e c h air, and I u sed a lot of those myself. One thing he did wa s turn o v er the ru nning of the departmen t to the facult y b y setting up a series of committees that did all the department w ork. W e had the degree program committee that set p olicy ab out requirements for degrees, the curr iculum com- mittee that decided what courses we re taugh t, the student ad v isin g committee that set p olicy ab out student s, lik e stip ends, and assigned advisors . This mean t th at Marvin did not h a ve to mak e all the d a y to day decisions ab ou t operations. But more imp or- tan tly , it ga ve the f acult y p o we r ov er ru n ning the de- partmen t. Marvin rarely inte rfered w ith department business, but if there was an issu e h e cared ab out, he let the facult y vo te; he just s et up his vo tes in adv ance with lobbying so ev ery one knew what he w an ted! I though t these w ere great strategies and did m y b est to apply them as w ell. My goals were broadening our r esearch agenda and strengthening facult y hires, esp ecially since w e w er e expanding at a rapid r ate w ith the initiation of CBAR and the AIDS w ork. I felt we needed to impro v e the envi- ronment for new facult y . O ne thing I in itiated wa s the p olicy that ev ery new jun ior faculty hire h av e supp ort for at least 40% of their time in their fir st 3 yea rs to d ev elop their own researc h agenda. Ov er the years th e 40% has fluctuated, b ut the p rinciple remains. I also increased the Department’s activit y in sta- tistical genetics. This was ab out the time of the com- pletion of the Hum an Genome Pro ject and the field w as highly visible and c hanging r apidly . As d epart- men t c hair I br ough t in a lot of outside sp eak ers in the area, obtained gran t supp ort for genetic meth- o ds and d ev elop ed s everal imp ortant collab orations. Initiating the 40% researc h p olicy for ju nior fac- ult y that allo w ed them to s tr engthen their own re- searc h agenda, and devel oping statistical genetics as a r esearc h area for the departmen t w ere m y t wo main con tributions as departmen t c hair. 8. ST A TISTICAL GENETICS Ryan : Y ou also got very in v olv ed p ersonally in statistica l genetics researc h. T ell m e ab out that, in particular, what motiv ated that mo ve for yo u? L air d : I w as in trigued b y genetics ev er since writ- ing th e EM pap er b ecause man y of th e earliest ap- plications of EM were in genetics. I t is only natur al b ecause, un til relativ ely recent ly , one could not di- rectly observ e DNA. I h ad a sabbatical at the end of m y fi r st fiv e years as department c h air and decided to sp end it learning more ab out problems in genet- ics; this pro ve d to b e very fruitful. O n e problem that caugh t m y eye was using family data to test wh ether or n ot a p articular genetic lo cus w as in vo lv ed in causing disease. A t that time, p eople we re us in g th e TDT (T rans ition Disequilibriu m T est) that required genetic data on paren ts and their d iseased child. But the question arose as to how to generalize this to u s- ing siblings and other family mem b ers when parents w ere n ot av ailable. This w as common in late onset disease such as Alzheimer’s. I collaborated with s ev- eral talente d students, p ostdo ctoral fello ws and f ac- ult y to dev elop a general framewo rk for family de- signs an d develo p ed the FBA T soft w are which has b een quite p opular. Stev e Horv ath, Stev e Lak e and Christoph Lange extended the theory and the pro- gram, add ing a lot of regression ideas for measured disease outcomes. T om Hoffmann extended the wo rk that Stev e Lak e d id on gene–en vironment in terac- tion studies. I had quite a few other students o v er the years working in genetics: Xiaolin W ang, w ho w as my fir st student in th is area, Ronnie Sebr o, who also work ed with Neil Risc h , C y r il Rako vski, Xiao Ding, Gourab De. Cu rrently , Christina McIn tosh is lo oking at design issues in family stud ies with m u l- tiple affected ind ividuals. Christoph Lange and I we re actually inv olv ed in one of the earliest genome-wide studies. W e w ere lo oking at the f amily comp onen t of the F raming- ham Stu dy w ith a view to id en tifying genes asso- ciated with ob esit y . W e had data fr om a chip with 100,00 0 S NPs (single neucleotide p olymorp hisms). Christoph had an id ea for a metho d that exploited the indep endence of within and b et w een family in- formation. W e used the b et we en family information A CONV ERSA TION WITH NAN L A IRD 11 to estimate the p ow er of eac h SNP test, then c hose only the top ten for testing. This enabled us to a void the loss of p o wer asso ciated with the multiple test- ing problem. W e foun d a SNP near the INSIG2 gene asso ciated with ob esit y . The results were contro v er- sial and many in vestiga tors attempted r eplications. Some replicated our resu lts, w hile others did not. A follo w-up meta-analysis sh o w ed that the asso ci- ation is present in studies of general p opulations, but th at the r ev erse asso ciation is seen in studies in v olving p opulations selected to b e health y . I had a somewhat similar exp erience wh en I was in v olv ed in a study th at identified a p oten tial Alzheimer’s gene. Th ere were a num b er of follo w- up studies, s ome that replicated, some that didn ’t. Genetic data are tricky . I t is easy to find lots of as- so ciated SNPs, bu t r eally h ard to pin d o wn causal mec h anisms in the DNA sequence. With complex diseases it is very difficult to attribute ma j or effects to one small disrup tion in the DNA. A f ew dramatic cases (Mend elian disorders) h a v e influenced p eople to think they will find a v ariation in the DNA co d ing that explains what’s going on. In realit y , it’s m ost lik ely a complex interpla y b et we en m ultiple genetic v arian ts and en vironm en tal factors. There is a lot of v ariation in the qu alit y of p ap ers app earing in the genetics and statistical genetics lit- erature and, truth b e told, th er e’s not a lot of qualit y con trol. P eople are under a lot of pressure to pu b- lish, and referees are under pr essure to d o fast r e- views, but there are so man y pitfalls. Genes are v ery complex—y ou can lo ok at gene expression, DNA, exons, introns, meth ylation sites etc.—and there is h uge p otentia l for false p ositiv es. The problem is ex- acerbated b y the fact that environmen tal exp osures are n otoriously h ard to measure accurately , even when we know what the r elev ant ones are. I think it is imp ortan t to bu ild genetic inv estigations around a theory , as we co mmonly do in epidemiology . Ryan : It seems these days that b ecause of tec h- nologies like GW AS, p eople d on’t wan t to pin do wn their h yp otheses in adv ance any more. L air d : I agree—the general null h yp othesis of no genetic v ariant affects the outcome is not so useful. There are alwa ys adjustments su c h as Bonferr oni or FDR appr oac hes, split sample etc., but p o we r is often low. P eople are also no w starting to impute SNPs in an effort to b o ost p o w er b y increasing co v- erage. I th ink these issues n eed a lot more statistical atten tion. Fig. 6. F ong Wang-Clow with Nan at her r etir ement c ele- br ation, May 2015. But I think a b igger problem is that p eople don’t tak e accoun t of design. P eople ha v e d one a lot of con v enience sampling, obtaining b iological samples from pr eexisting studies designed for different pur- p oses. 9. WHA T’ S NEXT W ORKWISE? L air d : Lo oking bac k, I think that v en turin g into the genetics area to ok ov er my career. But I enjo y ed w orking in the area. I t is exciting when y ou can us e kno wledge of ho w the b iology of inheritance works to driv e the analysis. My other work hasn’t h ad those strong biologica l und erpinnings. Ryan : So do you think you’ll k eep working in ge- netics? L air d : I’m lo oking forw ard to working on whatev er I lik e. Th at ma y or ma y not include genetics. Ryan : Y es, I had a coup le more questions ab out that. Random effects mo deling f eatures very promi- nen tly in your work. W ere you ev er tempted to b e- come a Ba y esian? L air d : No, I was n ot. The business of the p rior alw a ys seemed to get in the w a y . Ryan : Ev en w orking with Ar t Dempster? L air d : Is Art a Ba y esian? Certainly when I w orke d with him I w ould not h a v e called him strictly Ba y es. I w ould say he emphasized the like liho o d , but if he used priors, it was usually flat prior and often empirical Ba yes. My early wo rk was alwa y s ab out lik eliho o d-based m etho ds. I su pp ose a lot of my w ork could b e describ ed as empir ical Ba yes b ecause w e w ere dealing with hierarc hical mo dels. But we 12 L. R Y A N nev er sp ecified hyp er p riors, just u s ed the data to es- timate unkno wn parameters such as v ariance com- p onents. My work in genetics, ho w ev er, has b een largely score-based tests. In this setting we could obtain the distribution of the data under H0 using Mendel’s la ws, so it w as a natural approac h . Ryan : What ab out su r viv al analysis? L air d : Y es, I did one pap er with Don Olivier (Laird and Olivier ( 1981 )). W e formulate d the prob- lem as piecewise exp onen tial and made a link to log- linear mo dels. It was a fun pap er and provided a wa y to fi t mo dels to surviv al d ata in the da ys w hen gen- eral surviv al soft wa re w asn’t a v ailable. It also pro- vided a simp le w ay to test for nonprop ortional h az- ards. Ryan : Y es, that pap er w as very u seful. I remem b er reading it and using some of the ideas with one of m y stud en ts who wa s working on th e analysis of carcinogenicit y data. On e of the m an y things I ha ve admired ab out you o ver the yea rs Nan is that you don’t “t weak”—y our pap ers all ha ve real impact. L air d : Tweaking? Oh, I’ve t wea k ed, Louise! When- ev er I felt that I w as starting to t we ak, that’s when I mo v ed to a new area. It’s part of the reason wh y I think it’s time for me to start to wind do wn, re- searc h wise. But really , I’d like to explore other wa ys to sta y engaged p rofessionally . Ryan : T ell us more. L air d : W ell I’v e al wa ys really enjoy ed teac hing and wo rking with studen ts. I lik e collaborating as w ell. I h av e also greatly enjo y ed m y collaborations. I will con tinue to w ork with colleagues at the Chan- ning Labs (Harv ard Medical Sc ho ol) where we are lo oking at the genetics of Ch ronic Ob s tructiv e L ung Disease. In the early 1980s I started collaborating w ith col- leagues at Harv ard ’s S PH, Medical School, School of Go vernmen t and the Law Sc ho ol. New Y ork State w an ted an empir ical basis for establishing reform of their la ws on medical malpractice , and w e u nderto ok a samp le sur v ey of hospitalizations in New Y ork to determine the r ate of adverse eve nts, the p ercen t due to negligence, and the p ercen tage of adv erse ev en ts that are neve r rep orted. It was a land mark stu dy and the firs t rigorous one of its kind. It s till serv es as a mo del for tort r eform. Russ Lo calio, a former student now at U. P enn, and I were the statistici ans in v olv ed. I w ould lik e to get inv olv ed again w ith some Na- tional Academ y or Ins titute of Medicine p anels. One panel I w orke d w ith in 1986 w as The Airliner Cabin En vironment, headed b y T om Chalmers. I remem- b er this as one of the most fun and rewarding panels I w as on. T he senate had ask ed for an in v estigation of the qualit y of air on planes. It w as the era of smoking sections in the bac k, and , d ue to econom y measures, airlines were starting to use recirculated air in the cabins. I recall thin king righ t at th e b e- ginning that the simplest solution w as to eliminate smoking, b ut it took slogging th rough a lot of data to get to that recommendation. I recall Dimitri T ri- c h op oulos (then future c hair of HSPH Ep idemiology Departmen t) talking about h is wo rk on the v ery small b ut imp ortant risks of side s tr eam cigarette smok e, and Brian MacMa hon’s (then current c hair of HSPH Epidemiology Department) b elief that an y o dds ratio less than 2 w as to o sm all to b e of any relev ance. In the end, all of u s on the committee w ere con vinced that a smoking ban w as in order. The senate w as also con vinced, and it was enacted in to law shortly thereafter. Th is w as certainly one of my most s ignal public health con tributions. Al- though w e did not do a formal meta-analysis of the health effects of s id e-stream smok e, this exp erience influenced b oth me and T om Chalmers ab out the imp ortance of meta-analysis f or public health. 10. BOOKS Ryan : Y ou’v e co- authored sev eral already , includ - ing one on longitudinal data (Fitzmaurice, Laird and W are ( 2004 )) and one on statistical geneti cs (Laird and Lange ( 2011 )). Are there like ly to b e more? L air d : I am very p r oud of the b o oks and enjo y ed those collab orations a great deal, bu t writing a b o ok is v ery demandin g. Th ere is also an IMS Monograph on Longitudinal and clustered data. Th e b o ok with Christoph is ab out statistica l genetics, but written from the p ersp ectiv e of a statistician. It la ys out some of the fund amen tal id eas that drov e statisti- cal genetics, starting w ith link age, then moving in to family-based designs and GW AS, b ut stopping sh ort of s equ ence analysis. It’s written for the graduate lev el, but is fairly applied and has b een used by a lot of p eople for teac hing. Ryan : T he b o ok on longitudinal is m ore classically statistica l, righ t? How did that one come ab out? L air d : After wo rking w ith Jim for a few yea rs, b oth on r esearc h and on teac h ing short courses mostly outside of Harv ard , I decided to teac h a reg- ular course for the graduate s tudent s here in th e d e- partmen t on longitudinal data analysis. The course A CONV ERSA TION WITH NAN L A IRD 13 w as r igorous, bu t pitc h ed at a lev el accessible to b oth biostati stics and epidemiology stud en ts. After I’d taught it a couple of times, Jim and then Gar- rett to ok o ver teac hing the cour se. The b o ok evo lve d from the course notes and it is very p opular. I think Garrett d eserv es a lot of credit for the b o ok’s suc- cess. He is a wonderful co-author. He’s a v ery goo d writer and put in a huge amoun t of wo rk on the b o ok, b uilding in lots of d etailed examples as we ll as SAS and R co de. He alw a ys to ok J im’s and my in- put constru ctiv ely . The b o ok is m uc h br oader than gro w th curve mod eling, including GEEs as w ell as clustered data and m ultiple informan ts. T h e b o ok is v ery p edagogic al and comprehensiv e. It captures a lot of my professional life and I su pp ose that’s another reason why I feel v ery proud of the b o ok. Ryan : So many of y our stu den ts ha v e done s p ec- tacularly well . Y ou m ust f eel proud. I recall a sp eec h that Garrett ga ve a f ew years back at a dinner in y our honor, where he talk ed ab out what a fantastic men tor yo u are. What do y ou think it tak es to b e a goo d men tor? Did y ou hav e go o d men tors? What are some of yo ur s ecrets? L air d : Y es, my students ha v e done b eautifully , some sp ectacularly , and I am very proud of them all. I s upp ose I alw a ys try to b e s tr aigh t with m y student s. When they are fi rst starting out, I try to giv e them a lot of h elp if they need it, but gradually I step back, letting them tak e more of a lead. I think that’s imp ortan t if they are to b ecome in d ep end ent researc h ers ev entual ly . I alw a ys tried to put aside an h our a we ek to talk with students and I like to mak e them feel connected to the researc h area. So I stress reading pap ers not only clearly related, bu t p eriph erally relat ed also, or going to seminars. Men toring stu den ts has alwa ys b een one of the fa- v orite parts of my w ork. This is one reason that I c hose to ha ve m y r ecen t retirement celebration fo cused en tirely on m y stud en ts. It was organized b y Garrett Fitz maurice and Ch r istoph Lange, b oth of whom sp ok e, as did Reb ecca DerSimonian, Jo e Hogan and Kristin Jav aras. The h ighligh t w as F ong W ang-Clo w, wh o with her husband Eric, d onated $1,000 ,000 to the department to establish the F ong Clo w Do ctoral F ello wship in m y h onor. I was v ery surpr ised and h onored! The Biostatistics and Epi- demiology Psychiatric Semin ar Group also h eld a sp ecial semin ar h onoring my retirement and S tev e Lak e sp ok e there. In terms of my o w n ment ors, I’v e certainly b een influenced by some p o w erful senior statistici ans, suc h as F red Mostelle r, Art Dempster and Marvin Zelen. Bu t other than th e EM with Ar t, and one meta-analysis review article with F red, th ey we re not collab orators. I think I lea rned differen t sorts of things from eac h one of them. Strictly sp eaking, J im W are was a p eer rather than a men tor, bu t I learned a lot from working closely with him o v er the y ears. While of course neither of them w as my stud en t, b oth m y daughter Lily and her husband Cory Zigler ha v e their P h.D.’s in Biostatistics (from UCLA). While I don’t feel I can tak e muc h credit f or their professional develo pment, I am v ery p roud and happy to hav e t wo other b iostatisticians in the fam- ily! 11. BALANCING W ORK AND F AMIL Y LIFE Ryan : Man y of y our colleagues, for example, J im W are, T om Louis and Butc h Tsiatis, w ere quite ac- tiv e in the pr ofessional societies, for example, serv- ing as President of ENAR or the ASA. Y ou never seemed to do muc h of that. Ho w come? L air d : I had to mak e c hoices early in my career. My first husband and I separated wh en I en tered graduate school, and, as a sin gle mother, I j ealously guarded m y time for the sak e of m y s on . Th is feel- ing sta yed with me throughout m y career and after I remarried. I felt my family deserved m y atten tion and I tried to b e a 9 to 5 p erson. I was not very suc- cessful with th is, esp ecially when I was c hair, but I though t it w as imp ortan t to try . Ryan : I’m impressed and even a little surp rised since I’ve neve r heard y ou express those thoughts, ev en after all these y ears that we’v e kno wn eac h other! So do y ou think it is p ossible to b alance the demands of career and family? L air d : I hav e alwa ys b een a bit priv ate ab out p er- sonal things at w ork. I f elt that I came to the of- fice to w ork, n ot so cialize. It was alwa ys a surp rise to me that man y men do s o cialize at work. If y ou ha v e small c hildr en , you ha v e a big resp onsibilit y and I f elt it was imp ortan t to clarify the b oun d- aries b etw een w ork and family . Lo oking around at m y colleagues o v er the yea rs, most of them were out of to wn a lot, sometimes wo rking, b ut often giv- ing seminars and pr esen ting at conferences. Th rough F red Mosteller, I w as in v olv ed in sev er al National Researc h Council Panels. This take s a lot of time in terms of tra v el and p reparation. I had to make h ard c h oices, and , und er s tandably , I decided to eliminate activitie s that did n ot lead to academic publications 14 L. R Y A N or teac h ing success. All it tak es is s a y in g “No” a few times, but I do regret that I wa sn’t able to get more in v olv ed in p r o-b ono statistics wo rk . . . . I’d lik e to do more of that now. I understand that no w, in order to b e an elected fello w of the ASA, you need to hav e a trac k record of service to the ASA—I would nev er qualify no w, but I do feel I hav e made con tributions to our profession. Ryan : W ell, of course, it helps when y ou are really smart and when yo ur w ork is so outstanding that it still gets recognized, even without y our d oing the kinds of pr omotional th ings (talks, conferences etc.) that most of us need to do to get ou r n ames ou t there! But seriously , I think your story is a touc h- ing one and p oin ts to the real c hallenges that face w omen wh o hav e family resp onsibilities, y et wa nt to pursu e a career as w ell. A ttend in g conferences and giving talks are goo d examples of some of the tradi- tional mark ers of su ccess that are very “male.” L air d : Y es, I agree. I fin d it hard to b e comfortable with self-promotion, but sometimes y ou just h a v e to do it. I t can b e w orse to b e silen t w h en y ou are passed o v er. I rememb er early on in my career I had applied f or a grant from the National Science F ou n - dation. Nancy Flournoy called me up to s ay that I hadn’t su ggested any suitable r eview er s . It didn’t ev en o ccur to me that one could do th at! So I really appreciated Nancy’s call. And I really admire ho w forthrigh t Nancy is. I think also that there are different ph ases of yo ur career. Careers are long, and last m u c h longer than a c hildho o d. Y ou need to pic k and c ho ose wh at you w an t an d when y ou wan t it. When I tell this to grad- uate stud en ts to d a y , they can b e horrified that y ou migh t need to cut b ac k on certain activitie s for 10– 15 y ears if y ou wan t to ha v e kids. That seems imp os- sibly long for a young p erson starting out in their career, but of course, it is really sh ort. But I think it do es p oint to the d ifficult y with the tenure clock coinciding with the b iologica l one. I also think the situation is c hanging a lot for men as w ell. T hey are exp ected to b e equal partners in paren ting, and this will c hange their attitude to ward careers. I was luc ky to h a v e a h u sband w ho to ok on a v ery large sh are of paren ting that allo wed me to pursu e a career. And h e w as prou d of my career. I think this is more common no w than it was f orty y ears ago. Ryan : Did yo u try to connect with other w omen scien tists o v er th e y ears? W as this something imp or- tan t for yo u? L air d : Y es, bu t it just happ ened naturally w ith- out m y trying h ard to mak e it happ en. I w as go o d friends with several female faculty m em b ers here at HS P H. There are times y ou n eed to talk with another w oman, though there weren’t that many around, and certainly not man y with children. It wa s prett y difficult, esp ecially as a sin gle mother when I w as in grad s chool. Ryan : Y es, I can’t b egin to imagine. Y our story will b e inspiring for a lot of young wo men starting out in their careers. I thin k y our advice ab out j u st sa ying “No” is really p o w erful. I need to remem b er that m yself ! I h ad plann ed to ask yo u wh y y ou had sta yed at Harv ard all these yea rs instead of mo v- ing around like s o man y p eople do. W as that also influenced by your family circumstances? L air d : Y es, a lot of the reason for sta ying at Har- v ard w as p ersonal. I never felt it w as fair to disr upt the family . On the other hand , Harv ard is a prett y amazing en vironm en t. I alw a ys felt a lot of freedom to explore new id eas. Plus, my family has alwa ys b een a tremendously p ositiv e source of supp ort in m y career. When I got ten ure and wa s interview ed for th e HSPH newsletter, it was really imp ortant to me to ac kno wledge the role of my family . Th e editor questioned p utting this in the newsletter, I think b e- cause men d id not usually do this sort of thing, but I insisted. I rememb er b eing struc k b y the commen ts from a y oung woman faculty mem b er y ears ago at a panel d iscussion on the c hallenges of balancing ca- reer and family . Th is yo un g w oman said that to h er, it wa sn’t a challe nge, bu t rather a real plus since her family ga v e her th e supp ort and encouragement she needed to su cceed. I’ve alw a ys felt that w a y to o. It probably help ed, though, that Jo el was in a differ- en t fi eld. I know there are lot s of couples w ho are in similar fields and that can complicat e things. 12. WRAPPING UP Ryan : So, let’s mov e to w ard wrapp ing up no w. T ell u s ab out some more of the things y ou lik e to do outside of wo rk. L air d : I lo ve b eing with friends and family , esp e- cially my t wo granddaughte rs, Margaret and Gene- viev e. I lo ok forw ard to sp endin g more time with them and watc hing th em gro w. It is amazing ho w eac h c hild is so differen t. I can see wh y studies on the effect of en vironmental infl uences on c hildr en ’s dev elopmen t are so hard to do. I am a p assionate gardener and I h a v e though t ab out learning more ab out landscap e design. I w an t A CONV ERSA TION WITH NAN L A IRD 15 to tra ve l and see some places for pleasure rather than just for work I lik e to s ew, esp ecially quilts. Quilt design is v ery app ealing to me b ecause it is so geometric and colorful. Ryan : Let’s finish u p sp ending just a bit of time talking ab out where y ou see our field h eading. A lot seems to b e h app enin g th ese d a ys and I ’ve heard man y p eople sa y that statisticians are getting left b ehind by data miners, data scien tists and th e lik e. What are yo ur th ough ts? L air d : O ne of th e things I see, and this is esp e- cially so for s tatistical genetics and bioinformatics, is that p eople from fields like p h ysics and co mpu ter science are taking o ver the compu tations. S tatisti- cians can sometimes b e hamp ered by their desire to get metho d s exactly r igh t in terms of their p r op- erties. In cont rast, p eople in other fi elds are happy to get an appr oximate answer. W e statisticians ha v e suc h a tradition of b asin g our pu blications on s trong theory . I am impressed with the rigor, but the result is that m uc h of our w ork is not immediately prac- tical. W e are also slo w. As a result, w e are getting left b ehin d. Ryan : Are there any s olutions? L air d : W ell, certainly w e need to mak e sure that an y metho d s we pu blish are accompanied b y pub lic- access soft w are that can b e used to r eplicate the results and apply the metho ds easily in other set- tings. S oft w are also needs to b e w ell d o cumente d. Some p eople d o this r eally wel l, b ut most don’t. O f course, it is a dou b le-edged s w ord since there is not a lot of q u alit y control. Also, it is difficult and time consuming to write go o d soft w are. Ryan : Is there an ything we can do to make things easier? L air d : It w ould b e go o d to close the gap b et ween classical academic success and con tr ib utions more broadly . Rigorous work in statistics is exp ected for promotion in most academic departments, b ut that w ork ma y h a v e little relev ance outside of academia. W e need to broaden our criteria for success so that p eople whose wo rk has b een in leadership in apply- ing s tatistics at a br oader lev el can b ecome inv olv ed in academia, and vice-v ersa. Ryan : But isn’t that a b it tric ky ? P eople can gro w a v ery long C V compr ising second and higher au- thorship pap ers, but without having the kind of leadership I think you are talking ab out. L air d : W e sh ould b e training our students to b e leaders and encouraging them to think ab out first authorship pap ers in sub ject matter areas. Christl Donnelly , S tev e Horv ath, Nic k Horton, Stev e Lak e, F ong W ang-Clo w and Bob Glynn are just a few ex- amples of m y former student s that come to mind who ha v e h ad creativ e, out-of-the-b o x careers. Ryan : So h o w do w e do that? S hould w e b e train- ing p eople differently? L air d : It is hard to train stud ents b r oadly and deeply at the same time. Th ey need so muc h , so it’s a question of trade-offs. But something n eeds to giv e b ecause the w a y w e are teac h ing no w, our students get channeled into a v ery narro w base. T hat is OK when career lifetimes are long and w e ha v e opp or- tunit y for growth and change after the P h .D. But in academia w e pu sh p eople so strongly to do great w ork in only a f ew y ears. It do es not promote d iv er- sit y of th inking. I also think computing is really im- p ortant. It is a rare p erson who can do b oth comput- ing and statistics, but we sh ould b e at least trying to pro du ce such p eople. W e need to train our stu- den ts to think algorithmically: ho w w ould I compute that? They also n eed to u n derstand data and data structures. P erh ap s first and foremost, they need to kno w w ho will b e r eading their p ap ers and wh y , and who will b e usin g their metho ds . Ryan : I lo ve that phrase, “think algorithmically .” I think y ou are right on target here, Nan. Y ou’ve set a great example, w ith so m uch of y our work either directly fo cu s ed on alg orithms or pr oviding analysis strategies that lend themselves to effect ive compu- tation. In fact, y ou’v e had a s tellar career, seemingly to imagine th e futu re and pu sh statistics in the right directions. Ho w did you manage that? Do y ou ha ve an y final though ts or advice for y oun g p eople start- ing out? L air d : I think it is r eally imp ortant to lik e w hat y ou do, and enjo y your w ork. If y our job requires y ou to d o things that do not bring you p leasure, y ou are n ot going to b e su ccessful. If y ou wa nt to write researc h pap ers, y ou n eed to enj o y d oing the researc h and writing the pap ers. If yo u w ant to b e a successful teac her, you hav e to enjo y w orking with the student s. So many young p eople ask me what y ou n eed to do to succeed. The really hard question is how to fi nd something that y ou lo v e doing. Th at y ou hav e to answer for y ou r self, but it is what yo u need to do if you are to succeed. Ryan : Nan, it’s b een a pleasure and a privilege to hav e this con ve rsation with yo u. Thanks for y our time. 16 L. R Y A N REFERENCES Dempster, A. P. , Laird, N. M. and Rubin, D. B. (197 7). Maxim um likelihood from incomplete data v ia th e EM al- gorithm. J. R oy. Statist. So c. Ser . B 39 1–38. MR0501537 DerSimonian, R. and Laird, N. M. (1983). Ev aluating th e effect of coaching on S A T scores: A meta-analysis.” Har- var d Educ ational R eview 53 1–15. DerSimonian, R. and Laird, N . M. (1986). Meta-analysis in clinical trials. Contr ol le d C linic al T rials 7 177–188. Docker y , D. W. , Po pe, C. A. , Xu , X. , Spen gler, J. D. , W are, J. H. , F a y, M. E. , Ferris, B . G. Jr. and Speize r, F. E. (1993). An association b etw een air pollu- tion and mortalit y in six U.S. cities. New England Jour nal of Me dicine 329 1753–1759. Fitzmaurice, G. M. , Laird , N. M. and W are, J. H. (2004). Applie d L ongitudinal Analysis . Wiley , Hob oken, NJ. MR2063401 Laird, N. M. (2014). Meta-analyses: H eterogeneit y can b e a goo d thing. In Past, Pr esent, and F utur e of St atistic al Scienc e 381–389. Laird, N. M. and Lange, C. (2011). The F undamen- tals of Mo dern Statistic al Genetics . Springer, New Y ork. MR2762582 Laird, N. and Olivier, D. ( 1981). Co vari ance analysis of censored surviv al data using log-linear analysis tec hniq u es. J. Amer. Statist. Asso c. 76 231–24 0. MR0624329 Laird, N. M. and W are, J. H . (1982). R andom-effects mod- els for longitudinal data. Biometrics 38 963–974. Lin, D. Y. and Zeng, D. (2010). Meta-analysis of genome- wide association studies: No efficiency gain in u sing indi- vidual participant d ata. Genet. Epidemiol. 34 60–66. Meng, X.-L. and Rubin, D. B. (1993). Maximum likelihood estimation via the ECM algorithm: A general framewo rk. Biometrika 80 267–278. MR1243503 Slack , W. V. and Por ter, D. (1980). The Scholastic Ap ti- tude T est: A critical appraisal. Harvar d Educ ational R eview 50 154–175. V an Noorden, R. , M aher, B. and Nuzzo, R. (2014). The top 100 pap ers. Natur e 524 550–553.