Nan Laird is the Harvey V. Fineberg Professor of Biostatistics at Harvard University. During her more than forty years on the faculty, she developed many simple and practical statistical methods for pressing public health and medical problems. Her work on the EM Algorithm, with Art Dempster and Don Rubin, is among the top 100 most cited of all published articles in science. She’s also developed popular and widely used methods for meta-analysis, longitudinal data, and statistical genetics. She has worked in several areas of application including the quantification of adverse events in hospitals, childhood obesity, and genetic studies in Alzheimer’s disease, bipolar disorder, asthma, and lung disease. Laird was awarded the 2021 International Prize in Statistics for, "her work on powerful methods that have made possible the analysis of complex longitudinal studies."
Episode Description
Every two years the International Prize in Statistics is given out to recognize an individual or team for major contributions to the field of statistics particularly those that have practical applications or which lead to breakthroughs in other disciplines. The winner is chosen in a collaboration between the American Statistical Association, the Institute for the Mathematical Sciences, the International Biometric Society, the International Statistical Institute, and the Royal Statistical Society. The 2021 honoree is Nan Laird and her award and career are the focus of this episode of Stats and Stories.
+Full Transcript
Rosemary Pennington
Every two years the International Prize and statistics is given out to recognize an individual or team for major contributions to the field of statistics, particularly those that have practical applications or which lead to breakthroughs and other disciplines. The winner is chosen in a collaboration between the American Statistical Association, the Institute for the mathematical sciences, the International biometric society, the International statistical Institute, and Royal statistical society. The 2021 honoree is man layered and her award and career is the focus of this episode stats and stories where we explore the statistics behind the stories and the stories behind the statistics. I'm Rosemary Pennington. stands in stories is a production of Miami University's Department of Statistics and media journalism and film, as well as the American Statistical Association. Joining me are our regular panelist, John Bailer Chair of Miami statistics department, and Richard Campbell Professor Emeritus in media journalism and film. Our guest today, as I mentioned, is Nan Laird. Laird is the Harvey V. Feinberg, Professor of Biostatistics at Harvard University. During her more than 40 years on the faculty, she's developed many simple and practical statistical methods for pressing public health and medical problems. Her work on the EM algorithm with Art Dempster and Ron Rubin is among the top 100 most cited of all published articles in science. She's also developed popular and widely used methods for meta analysis, longitudinal data and statistical genetics. She's worked in several areas of application, including the quantification of adverse events in hospitals, childhood obesity, and genetic studies and Alzheimer's, bipolar disorder, asthma and lung disease. And as I said, Laird was awarded the 2021 International Prize in statistics for quote, her work on powerful methods that have made possible the analysis of complex longitudinal studies and quote, man, thank you so much for joining us today. And congratulations on the award.
John Bailer
Nan, let me add my congratulations as well, it's just delightful to see you recognized with this award. You know, I mentioned to you previously that I remember seeing this idea of longitudinal data and the way it had to be structured. And in it, there were all these seeming constraints on trying to do this with the traditional methods at the time that I thought Holy cow, no one will be able to do this, no data set will ever conform to it unless I made it up. So I'm trying to, I'd love for you to tell us a little bit about what was the motivation, inspiration for the work that you were doing in longitudinal data methods? And perhaps you might want to start with just saying, What the heck is longitudinal data?
Nan Laird
Oh, yeah. Why don't I start with that? So in a longitudinal study, we have a sample of individuals, and we measure them repeatedly over time on the same variables. And so it's this concept of measuring people repeatedly. That's important, because it means that you can calculate change in the outcome that you're interested in over time. And such studies are often used in public health settings, to look at how people change over time, how do children grow? How do the elderly that climb? And what are some of the factors that you might be able to identify that influence growth, or decline and people. And I'd like to mention, in contrast, another sort of study that people often use, and that's the so-called cross sectional study, and that's the one where, again, you have a sample of individuals, but you only measure on one occasion. And you might, in that setting, find a lot of interesting associations between certain exposures and outcomes. But you, you can't necessarily conclude that if you change a person's exposure from one thing to another, that that will induce a change in the outcome because you haven't measured any changes at all. So this is thought to be one of the most important things. And another feature of the longitudinal study is if you think there are intervening variables that might influence Association, you can hold them constant within a person. So that's another attractive feature of longitudinal study. So you ask how I got interested in this. Well, let me just say one other thing that I think is relevant, I got my PhD in 1975 and Just about the time that I got my PhD was a time that either a lot of longitudinal studies in epidemiology and public health, they were matory. And they had all these data on people that have been followed for a long period of time. And in cases, they didn't really know what to do with it. Because as you said, John, the sort of statistical tools available at the time, were for very tightly structured data. So after I got my PhD, I went to Harvard School of Public Health. And shortly after I was there, I was joined by a colleague, James Ware, who was recruited to the school to work on a study called the Harvard study of air pollution and health. And this study actually was initiated in 75. And one of the initiators of this study confessed recently, that it was initiated without consulting any statisticians whatsoever.
And, their desire was to see if air pollution levels affected growth in children. They were interested not only in certain respiratory outcomes, like cough and respiratory illness, but also measures of lung function. And so they recruited children from six different cities, which had varying levels of air pollution. So they chose six cities, I don't remember all the six, two of them were chosen to have to be fairly dirty cities, two were chosen to be pretty Clean Cities. And then there were two in the middle. And they recruited some 10,000 children through the schools to participate in this study. And they had baseline measurements of all kinds of things. And then they measure people repeatedly for up to six years, in the end annual visits to the school, on intervening respiratory illness, and on lung function and other kinds of pertinent variables. And then, when Jim came, which was probably around in the 1980s, there, we're getting ready to analyze this data. And as John alluded to, really, all the early work on repeated measures, which included measures that were taken over time, under different circumstances, as well as ones where you really wanted to measure growth and change in an outcome. I believe those statistical methods had arisen in the context of agricultural field trials, where you could really control things very carefully. So if you do an experiment, you put a great deal of effort into controlling your variables. And when you take the measurements and how many measurements and make sure everything is the same age and taken under the same circumstance, and that the treatment effect doesn't change over time. And, under those circumstances, you could get some nice results over how to analyze this data. But when you talk about these big epidemiological studies, involving 1000s of individuals over time, even if you were set out to measure children having the exact same set of ages, at the exact same time, things happen. Things this schedule goes out, children drop out, children don't show up for school that day. So you get designs, which can be very highly unbalanced. And even in the case where the longitudinal studies are prospective, and if you get studies that rely on drawing past histories out of clinical records, they can be highly unbalanced sorts of data. And although I hadn't worked in the area of longitudinal data, or repeated measures, I had done some work on various components. And there that was part of my thesis work in a paper that I wrote with Art Dempster and Don Rubin, on the EM algorithm that showed how you could put the concept of incomplete or missing data into this umbrella. And I provided a general framework for how to think about using maximum likelihood. So Jim and I worked on that approach.
In the context of longitudinal data analysis, and we developed what I thought at the time was well obvious, simple, straightforward approach, that's going to work pretty much for anybody who had collected data over time with the objective of studying how people grew, and whether or not that growth was affected by various covariates. And then there might be a particular exposure she wanted to look at. And that was an attractive feature of our work is that it could accommodate virtually any kind of design that you did. And I think another feature of our work was very attractive as we use very ordinary regression models. So if you've only had a course in standard regression, and your first statistics course, you've seen those betas and those x's and those y's and you know how to interpret them. So we put our models into that framework and showed how you could use this very familiar type of framework to express hypotheses of interest even when the data were pretty complicated.
John Bailer
Well, it was brilliant, thank you for making it easier to teach this.
Richard Campbell
Nan, I have a question about your kind of support of applied work and practical work over time. And I was teaching in Michigan at a time when I knew theoretical work was much more valued over applied work. And I'm at Harvard, and I'm imagining the same thing. True. Did you face any pushback or any setbacks in your career for your interest in applying things to, you know, life problems that were practical?
Nan Laird
Well, I, I'm glad you asked that question. Because sometimes I get criticized for just the opposite thing. And that, and this was particularly true in the setting of some work I did on LSAT scores was, I was sort of accused of being more interested in the methods than in the results. And that was a bad thing. But, you know, I guess, I see myself as a person who was always motivated by trying to solve a problem. And I was motivated by how problem a actually was very similar to problem B, even though they might look quite different. They were all the same. But I have to say, I don't think of myself as being a totally applied statistician, because I never spent, I tended to be a colleague that the applied statistician consulted with, and I fell off and I didn't get into the subject matter as much as I could have, or should have or wanted to. But I think your question is quite relevant. People often feel that Applied Statistics is not as important. But for me, I never I never would have succeeded in an institution without being motivated by applications. That's clear.
Rosemary Pennington
You're listening to stats and stories. And today we're talking to Harvard University's Nan Laird, the 2021 International Prize in statistics awardee.
John Bailer
So that's a follow up on that. I'd like to have you talk a little bit about work that you had done associated with cabin air quality I. I was picturing this trip that I took in the early 90s overseas. And I was sitting in the non-smoking part of coach, which was ahead of the smoking part of the coach and behind the smoking part of first class. I quickly realized that smoke did not stay in the seating sections. And I certainly was not a happy guy about that. But I was delighted once when this was removed from air travel. So can you talk a little bit about getting involved in this work, cabin air quality?
Nan Laird
Well, yes, I have to say this is one of the things that people often congratulate me on. So I was asked to be on a committee of the National Academy of Sciences and it's usual for Congress when they think there's a debate about some health effects. To set up a committee convene a panel of experts, and they're going to study this. So the background on the air cabin quality was, you know, in the early days of flying, and I'm not sure I remember that far back, airplanes used to take in fresh air. And that way, they kept everybody happy. But fresh air that's very expensive to use, because you have to intake Way up high where the air is very cold, then you have to heat it up. And it gets so hot that you have then cooled down again, so that you can use it in the cabin. So it was a very expensive proposition. And when fuel got to be, to the ouch factor for the airlines, they redesigned the planes. Basically, they redesigned the planes to work on filtered, or recirculated air. And of course that recirculated air is filtered. And probably filtered much better than it was back in those days. But it was filtered. But it didn't really do a good job of filtering out cigarette smoke. So that's when the airlines created the smoking sections, which were located near the galleys where the airline stewards were. And so it really arose as a problem in occupational health, because people in the airline stewards union were very concerned about the level of exposure to environmental tobacco smoke. And you know, so they had some really great advocates in the US union, and they also had the ear of many senators, particularly senators from Hawaii, who had many long trips on airplanes. So it was an ideal setup. And after intense lobbying for a number of years, they got Congress to set up this panel to consider it. And so, you know, honestly, I don't know how I was on this panel. I was the only statistician. There was an epidemiologist, though. And so right from the beginning, it became clear what we had to do there, there was an awful lot of research into a lot of different things. And it was clear then that viruses could be transmitted through the ventilation system, and that had to be fixed. But one of the things that was really pretty clear was that they weren't going to get rid of environmental tobacco smoke. And I think we estimated that being an airline steward was about equivalent to living with us a pack a day smoker, on a regular basis. So it was a non trivial amount of smoking,
John Bailer
This was before the secondary smoke, all the work on secondary smoke was done.
Nan Laird
It was just I think, at the point where it was beginning. No, there weren't many studies that were specific to the air cabin, and the stewards environment. But there were these studies that were being done. Very interesting studies where they would take a pack a day of smokers and their spouses who were not smoking, and they would compare death rates in these couples. And there was very little but statistically significant evidence present in many of these studies suggesting that the sidestream smoke was harmful to your health. And I remember we took public testimony, and the chairman of the epidemiology department at the School of Public Health, so I knew him pretty well. He gave testimony that said, a risk factor is ridiculous, it doesn't mean a thing. It's way too low, you'd have to have a risk factor five before we believe it. And I just remembered that coming into my mind so clearly, because now I've been recently involved in genetic studies, you know, and the elevation of Rick's risk factors is about 1.02 and people get excited about it. And the other little funny anecdotal thing about the air cabin quality study was one of the most persuasive and best done studies on the effects of sidestream. Smoke was done by a man from Greece who later became the chair of the department at Harvard epidemiology so they were kind of contradicting each other's work. There. But anyway, what was just abundantly clear was this problem wasn't going to go away and less, we ban smoking on airlines. So we did, we recommended a ban on smoking, and airplanes. And I remember there were a number of people on a committee who were initially reluctant. They thought, Well, if we ban smoking on airplanes, we're going to put Congress in a really bad position, because there's the tobacco lobby, and smokers lobby, and we just really shouldn't put them in such a bad position. But some of us on the committee felt that quite clearly that we just should ban it, and we recommended that and a few months later, they banned it. And that was that I don't remember a lot of feedback from the tobacco industry. And what I read lightly was, the anti sentiment smoking was so strong, that they just couldn't get traction. So that was a satisfying result.
Richard Campbell
Can I ask you? This is a two part question because John likes to pack question. When I asked them. How have you felt about the way journalists have covered your work? I mean, have you been? Have there been times when you felt annoyed? The second part is, what can journalists do better to cover the work of statisticians?
Nan Laird
Okay, so Well, let me say, first of all, journalists rarely cover my work louder than they do. One thing I can think of, was when I was a consultant in a court case, and this was a case involving excedrin versus Tylenol, and you know, they're identical, except excedrin has caffeine and otherwise, they're identical. And the, you know, these pharmaceutical companies, then they really have to work for a living, they had all these studies of their kind of like longitudinal studies, but is that in the same individual is given different compounds at different points in time? So I think the one study under question was, these are people who suffered from chronic migraines. So whenever they had a migraine, they were randomized to one of three treatments and say, like Tylenol, excedrin, or placebo, and everything was a disguise, so they didn't know which one they were getting. But in this one particular instance, the investigators thought, well, they had to include placebo, because the FDA made you include placebo. And they thought, however, randomizing subjects in the context, if you didn't promise them that well, maybe there's a chance that they'll get both of these compounds, and that they're not going to get on placebo, was a lot better. So they conceived a design, which was called a balanced, incomplete block design. And in this balanced and complete block design, you only got two concepts. So there were a bunch of people that got a and b, and there are a bunch of people that got a and placebo and a bunch of people that got B and placebo. And how do you analyze that? Well, traditionally, from our agricultural experiments, we were taught that you use the within block information. So in other words, to compare excedrin and Tylenol, you just threw away all of those people that had placebo, and you just looked at the difference between the within person difference. And that's what had been done by the investigators for this study. And if you did that, excedrin was looking better than Tylenol. So the manufacturers of Tylenol said that can't be right. Because caffeine is not a painkiller. It's known as a stimulant, but it's not a painkiller. So how can that be? Right? So I looked at the data and I said, Well, why this recommendation that you only use the within person comparisons, that's only good. If the correlation between the measurements on people is kind of high, or maybe low, I forgot which way it works now, but in this case, the correlation was sort of right at the cut point of where you recommend using one approach or the other. The other approach involves actually estimating a variance component. So it's a much more complicated thing that people don't like to do. But we did this other approach. We combined the full data, the advantage for etc and went away. So we went to present this in court before a judge. And he was I might add a judge in his 90s. So you did wonder what his ability was to, to follow arguments. So we did when the case for Tylenol, and it said it was written up in the Wall Street Journal, for obvious reasons. So the Wall Street Journal was interested in this finding. And they said, well, the defendants used an extremely sophisticated statistical argument to win the day. That was all it was. But there was a lot of feedback from a study I did with Rebecca bear simonian. On the LSAT scores, there's a lot of feedback in the press on that one.
John Bailer And that was more of the meta analytic one?
Nan Laird
Yes, that meta analysis was something that I was always interested in, I think, even as a grad student, because Fred Mosteller was quite interested in meta analysis. And it was just beginning to be popular back in the early 70s, when I was a student. And it was used mostly in sociology, where you'd have 1000s of studies of psychotherapy versus psychopharmacology, and who got better, faster. And so there was, there was beginning to be some interest in that in Madison, although it wasn't much used at the time. But somehow, Fred got ahold of a data set that had been collected by two professors at Harvard Medical School, and that's another story unto itself, which I won't go into. But they had collected data from about 20 studies that have been done on whether or not there was an effect of coaching on the LSAT. And their particular interest in this was basically the LSAT is its Scholastic Aptitude Test. So aptitude doesn't mean, what have you learned? What does it mean? What are your abilities to learn? So their reasoning was, if you have a test that's designed to measure aptitude, that if you then coach people, it should not have much of an influence on your aptitude. And I thought this was really ripe for looking at because if you look at the studies, it was kind of clever because for many studies, they didn't have a control group, but the essay key publishes all these statistics on how people do how the random person does. And if they take it at one at the end of their junior year, and then three months later to the beginning of their senior year, how much do they gain. So they have all these statistics that you can use, estimate what you would expect the ordinary kid to gain and compare that to what people actually gained. And so there are a whole variety of different studies. There were very few studies that were randomized, where you randomize kids into a coach and a coached group, but there were quite a few studies. And so when I looked at that I said, you know, it's just clear that you've got to build a harrowed genius mix of studies. And you know, that's something that people worry about a lot. At the time when doing a meta analysis, they all said, there was a school of thought that said, No, you can't combine heterogeneous studies in there's another school of thought, yes, you got to combine everything you can find. And to me, it seemed quite natural to use the concept of variance components to indicate to what extent there was variation in the results. So I had a graduate student at the time, who also worked with Fred and she was familiar with the data. So we did an analysis that showed how you could use variance components to characterize the degree of variation in these studies. And also, we showed that if you stratified the subject according to the study, according to the design, it can make a big difference in the results. And the conclusions were that basically, if you looked at the more tightly controlled studies, there wasn't much indication that coaching could be helpful. And I remember Fred Mosteller said to me, Well, of course coaching is helpful, we wouldn't be here if coaching wasn't helpful, of course. He said the question is, what's the evidence for coaching? Well, there wasn't any evidence very strong for coaching at the time, but the press, then they made hay out of it. MIT, it was reported in many newspapers. And, you know, the reports are small. They all said, there's not much evidence that coaching has a big effect. And of course, the Kaplan testing Institute didn't really like it, I'm sure. Anyway, it was somewhat controversial, but I don't recall that the press was negative about the study. I don't recall what they were.
Rosemary Pennington
Well, that's all the time we have for this episode of stats and stories. Man, thank you so much for being here today. And again, congratulations on the honor.
John Bailer Thanks, Nan. This was wonderful. What a delight to speak with you.
Rosemary Pennington
stats and stories is a partnership between Miami University's Department of Statistics and media, journalism and film, and to the American Statistical Association. You can follow us on Twitter, Apple podcast, or other places where you find podcasts. If you'd like to share your thoughts on the program. Send your email to stats and stories at Miami. Oh h.edu or check us out at stats and stories dotnet and be sure to listen for future editions of stats and stories where we discuss the statistics behind the stories and the stories behind the statistics.