Discussions on the growth and future of biostatistics

Discussions on the Growth and Future of Biostatistics
and Statistics
Yi Li
Harvard University
SUMMARY
A panel discussion jointly sponsored by the Department of Biostatistics, Harvard School
of Public Health (HSPH) and the Department of Biostatistics and Computational Biology,
Dana-Farber Cancer Institute (DFCI), was held in Boston, Massachusetts on April 3, 2008.
The panel included Colin Begg from Memorial Sloan Kettering/Cornell Medical School,
Bradley Efron from Stanford University and David Harrington from Dana-Farber Cancer
Institute/Harvard School of Public Health, and the meeting was organized and moderated
by Yi Li from Dana-Farber Cancer Institute/Harvard School of Public Health. Discussions
focused on the growth and future of Biostatistics and Statistics and their connections with
the other _elds, e.g. Probability, Biology, Bioinformatics and Computer Science.
Key Words: Biostatistics; Statistics; Growth and Future.
1 Topics and Questions for Discussion
Prior to the meeting, Yi Li polled HSPH and DFCI faculty and researchers. The following
topics were deemed important for the discussion on the growth and future of the profession.
Q.1 Should Biostatistics continue to be a separate discipline from Statistics? Should Departments
of Biostatistics and Statistics merge? In other words, are we seeing a convergence
of biostatistics and statistics? Biostatisticians develop statistical methodology, statisticians
1
are getting involved in biological/clinical data. For example, at Harvard, we are debating
over moving closer to Cambridge - some say that the move might lead to the eventual merge
of the Biostatistics and Statistics departments. What are your thoughts on the division between
the disciplines of statistics and biostatistics in general, whether it is widening or
closing, and how it may a_ect our careers and career choices, especially for starting faculty,
postdoctoral fellows and students?
Q.2 How should statistics/biostatistics as a _eld respond to the increasing development of statistical
and related methods by non-statisticians, in particular computer scientists? It is
striking to some extent that statisticians get involved in applied problems in a rather arbitrary
fashion based on personal connections and whether the statistician's personal methodological

research _ts with the applied problem. Are statisticians su_ciently involved in the
most important scienti_c problems in the world today (at least of those that could make
use of statistical methods) and if not, is there some mechanism that could be developed by
which we as a profession can make our expertise available to the scientists tackling those
problems?
Q.3 How do we close the gap between the sophistication of methods developed in our _eld and
the simplicity of methods actually used by many (if not most) practitioners? Some scienti_c
communities use sophisticated statistical tools and are up to date with the newest developments.
Examples are clinical trials, brain imaging, genomics, etc. Other communities
routinely use the simplest statistical tools, such as single two-sample tests. Examples are experimental
biology and chemistry, cancer imaging, and many other _elds outside statistics.
How do we explain this gap and what can we do to close it?
Q.4 What makes a statistical methodology successful? Some modern statistical methods have
gotten to be very well known in the scienti_c world, even though they are not usually
part of any basic statistics course for non-statisticians. The best examples might be the
bootstrap, regression trees, wavelet thresholding. Even Kaplan-Meier and Cox model are
not in elementary statistics books! But most statistical methods, even when they are good
enough to be published in a statistical journal, might get referenced a few times within the
statistics literature and then forgotten, never making it outside the statistics community.
What makes a statistical methodology gain widespread popularity?
Q.5 Where should computational biology and bioinformatics sit in relation to biostatistics?
Should these subjects be taught as part of cross-departmental programs of which biostatistics
is a part or should they be housed within an expanded biostatistics department?
Q.6 Terry Speed (2007) stated ... the most prominent features of the data were systematic.
Roughly speaking, we saw in the data very pronounced e_ects associated with almost every
feature of the microarray assay, clear spatial, temporal, reagent and other e_ects, all dwarfing
the random variation. These were what demanded our immediate attention, not the
choice of two-sample test statistic or the multiple testing problem. What did we do? We
simply did things that seemed sensible, to remove, or at least greatly reduce, the impact
2
of these systematic e_ects, and we devised methods for telling whether or not our actions
helped, none of it based on probabilities. In the technology-driven world that I now inhabit,
I have seen this pattern on many more occasions since then. Briey stated, the exploratory
phase here is far more important than the con_rmatory... How do we develop the judgement,
devise the procedures, and pass on this experience? I don't see answers in my books
on probability-based statistics." Are or aren't we better o_ without probability?

2 Meeting Summary
Yi Li convened the session at 2:00pm on the 3rd of April, 2008, by welcoming the audience
and introducing the panelists.
2.1 Opening Statements
Yi Li: Many in our profession feel swamped with an ever increasing demand for statistical
and biostatistical support, and are often stressed by the need to engage in problems
spanning diverse areas. It is helpful to step back and get some perspective on what is happening
with our _eld and where it is likely to go in, say, the next 5 years. It's a privilege
today to hear the personal views from the leaders like the panelists and many others in
the audience on the growth and future of our profession and its connections with the other
_elds, including Probability, Biology, Bioinformatics and Computer Science. Let us start
with Professor Harrington, followed by Professors Efron and Begg.
David Harrington: On the question of whether departments of Statistics and Biostatistics
should merge, it is tempting to say something provocative, that perhaps it is time
for departments of Biostatistics and Statistics to merge. Much of the methods development
in Biostatistics involves di_cult mathematics, and the faculty in Statistics Departments increasingly
recognize the value of collaboration with scientists from other disciplines. But a
moment's thought shows that answer to be glib and super_cial. The rapid and remarkable
growth of of Statistics and Biostatistics departments in the last 20 years shows how much
work there is to do in each of these _elds, and that growth is (thankfully) not reversible.
Since I posed the question to Yi Li to begin with, I am free to say that the question as
posed may be more than a little silly!
Why did I raise the question? The increasing gap between academic departments of
Biostatistics and Statistics raises some important issues that need attention. At most universities,
academic biostatisticians are too distant, both geographically and intellectually,
from undergraduate populations. Undergraduates who have had an introduction to Statistics
lack the opportunity to see the vitality of Biostatistics as a scienti_c discipline that
3
is deeply embedded in cancer, HIV or cardiovascular research. As a discipline, we lose
the opportunity to recruit the best of these students into Biostatistics graduate programs.
Statistics education for undergraduates in Statistics Departments tends to concentrate on
the mathematics of statistics, and that is a narrow slice of the discipline in practice. I
believe that is part of the reason we are increasingly unsuccessful at attracting talented

U.S. citizens to pursue statistics in general, and biostatistics in particular
I think both types of departments have some strong tendencies that need to be mitigated.
On the one hand, biostatisticians can sometimes latch onto solutions too quickly
without carefully thinking about whether the solution is optimal. In some situations, such
quick solution is not only suboptimal but also wrong! On the other hand, some statistics
department faculty members tend to make the perfect be the enemy of the good. The
tendency to focus too much on _ne tuning a solution can sometimes slow down the research.
What steps can be taken to build stronger ties between Statistics and Biostatistics
Departments? Some steps are obvious { jointly sponsored seminars and social activities.
Biostatistics faculty should be come more involved in undergraduate education, including
supervising advanced undergraduates in summer research or internship opportunities. University
administrations should be convinced to fund intra-university sabbaticals that would
allow Statistics faculty to spend signi_cant amounts of time a Biostatistics environment,
perhaps with a focus on collaborative research, and vice versa. Graduate students should
be encouraged to cross-register for courses and to seek members of dissertation committees
from both departments.
I view Q2 as a non-starter! It is terri_c when other _elds dabble in statistics. That
`dabbling' often leads to signi_cant contribution to statistical methods. Imitation is a sure
sign of attery! It appears that data are driving research in almost all _elds now, ranging
from data intensive genomics to traditional Arts and Sciences disciplines, such as, history,
archaeology and architecture. One of my undergraduate students in an introductory course
in Statistics at Harvard College used the distribution of archeological _ndings of clay pot
shards and spatial statistics to determine how villages were organized and distributed in
ancient China. The problem was a very di_cult one, both because of the uneven nature of
the data and the di_culties of spatial inference, but the student had a wonderful intuition
for empirical research.
With the merging of statistics with the other _elds, we should take the opportunity to
learn from other disciplines. Statisticians often overlook the chance to make substantial
contributions to _elds such as biology, arti_cial intelligence, or the social sciences. Fisher
was known as a geneticist to the biologists; Fred Mosteller made important contributions
to educational research. I hope to see a time when people from other _elds ask why
statisticians dabble so often in their domain! As scientists, statisticians should wander
more into other _elds where careful inference is essential to progress.

4
Q.5 is about biostatistics versus bioinformatics and their relationship. I think that
biostatisticians should de_nitely play an extremely large role in the _eld. Technology development
in the last 10 years has encouraged the rapid integration of molecular biology
and biostatistical techniques. We simply cannot a_ord to walk away from the biological
problems arising in this merger, since they are so important to science and contain so many
unsolved statistical questions. We as a profession tend to be reluctant to touch problems
until they are well-formulated mathematically, so that, for instance, we know how to calculate
sampling distributions, based on _nite-sample or asymptotic properties. Sometimes
the epistemic variability (variability other than sampling variability) as mentioned in Speed
(2007) scares us away from the di_cult problems and slows down our engagement in this
_eld.
Brad Efron: First consider Q.1. The training in the two departments is not very
di_erent, but the departments tend to attract di_erent students with more mathematical
students gravitating to statistics departments and biostatistics departments attracting more
scienti_c students. It is very good for the _eld and for universities to have two strong
departments that increase the visibility of statistics. We are likely to see even more growth
in statistical departments including geostatistics and astrostatistics departments. I don't
feel bad at all and actually think this is a healthy intellectual advancement.
What do we do about the computer scientists referred to in Q2? My experience is that
scientists are good at developing probabilistic methods but are less good at analyzing them.
Since other _elds have the advantage of being closer to interesting datasets and problems,
which will be brought to our attention. So this growth should be positive and will not put
us out of business. As a matter of fact, I think this will only increase our business.
Q3 concerns sophistication versus simplicity, which is pretty much the opposite of Q2.
I wouldn't feel bad if complicated methods such as neural networks took longer to be put
to use, but the reverse seems to be the case. The more complicated a method is the harder
it is to critique, but the more it may need criticism.
What makes methods successful (Q.4)? I wish I knew. For example, bootstrap standard
errors caught on immediately, but bootstrap con_dence intervals, which took a lot of my
research time, were not interesting to most people. Another example is the Wilcoxon test.
As one of the most cited statistical papers, it is only marginally better than the standard ttest.
In contrast, James-Stein's Estimator, which o_ers dramatic increases in performance

compared to the least-squares estimator, is largely ignored. I can never _gure out what
catches people's interest!
Let me deal with Q.5 and Q.6 altogether. I read Terry's article when it came out in
the IMS bulletin, and in fact read all of "Terry's Stu_" in the IMS bulletin because they
are so interesting. Terry must have had a lot of better microarrays than I have, because
in his, everything is clear after a while, but mine are never so! The question is whether
5
we can have statistics without probability. My own preference is for Fisher's attitude
toward statistical inference, which takes a midway point between theoretical models and
informal numerical summaries. I understand, and sympathize with Terry's worry that fancy
modelling calculations are taking the place of common-sense statistical inference, but Tukey
(and Terry) go too far when they want to banish probability theory from our tool chest.
There is a tradition of data driven statistics from a French school (correspondence analysis)
and based on the "green book" (Mosteller and Tukey, 1977). All of that can be fun, but it
does not get very far because real data analyses are dynamic, going back and forth between
inference and descriptive statistics. If there were no probability theory, it would be the end
of serious inference. So I am _rmly against giving up on probability.
Colin Begg: Let me start with Q.1. I address my response from the perspective of a
recruiter of graduate students, having been head of a biostatistics group for 20 years at a
medical center. I am interested in students who show promise as independent investigators,
have a genuine interest in collaborative work, demonstrate an aptitude to analyze and
interpret data and know commonly used biostatistical methods (such as categorical data
and survival analysis, longitudinal analysis and clinical trials). Biostatistics students tend
to have taken more useful methods courses, but statistics students tend to have more of
a focus on independent research. In general, I have not noticed a big di_erence in the
credentials and quality of the graduates from each type of department.
For Q.2, as statisticians, we have all been exposed in our education to what might be
called a statistical tradition" that emphasizes scienti_c norms that characterize the best
features of our work. These include the tradition of validating proposed new methods, a
respect for the importance of data in scienti_c work, and a recognition of the importance of
study design in the relevance/quality of the data. The work of methodologists not trained
in statistics is much less likely to be characterized by these principles.
A short answer to Q.3 is to use simpler methods. Simplicity is a huge advantage in a

method, for very good reasons. It is important to emphasize that simplicity is not equal to
na•ivete. Simpler methods enjoy the bene_ts of reducing the risk of gross interpretive errors,
allowing for transparency of analysis and results and increasing interdisciplinary use. The
additional complexity of a method needs to be balanced by the increased bene_t due to
the complexity.
With respect to Q.4, successful methods often address subject matters that people care
about, are on a topic that is easy to explain and are often accompanied by conceptually
simple statistics (not necessarily technically simple), plots, images and pivotal data summary
that are easy to convey results. It would be interesting to conduct a study of the
characteristics of methods that have been historically successful.
Q.5 is very interesting. Academic _elds do evolve, so it is unclear what will happen
in the future. Computational biology is distinctive from statistics and biostatistics
6
as it emphasizes, for example, sequence alignment algorithms, protein folding and knowledge
of genomic databases, requires knowledge of biomathematics, and depends heavily on
molecular biology knowledge. Given these di_erences, the two _elds should have separate
departments. I, however, do recognize the strong overlap in the interests of statisticians
and computational biologists.
Finally turn to Terry's issue (Q.6), which is the most di_cult one. Prior to most
probabilistic analyses, any dataset requires cleaning, normalization and transformation,
which is often done in an informal non-probabilistic manner. This is especially true for
genomics, as Terry is trying to emphasize, where the data preparation process is even more
intensive. However, does this invalidate the probabilistic principles? Analyses of genomic
data should still adhere to the traditional statistical principles that I mentioned earlier
and so probability theory is still relevant. In particular, validation of the predictions of
data analyses is necessary and this is a key issue in genomics. In summary, the statistical
tradition still has a lot to o_er in this _eld and probability theory is essential in carrying
out hypothesis testing and prediction validation.
2.2 More Discussion
Carl Morris: These aren't easy questions! At Harvard the statistics and biostatistics
departments are physically far apart. We have been working to close that gap, and it will
be good to _nd more ways to move toward a closer relationship. I have the following two
comments.

Statistics is fundamentally interdisciplinary. Its core concerns the development of theory
and methods for interdisciplinary use. It also fosters generalizing methods developed for
one purpose and making them useable by researchers in other _elds, often to be applied to
quite di_erent topics. Least-squares methods, Kaplan-Meier, and many other well-known
procedures are examples of how a procedure developed for one purpose can have many
other unsuspected applications.
Model building is an essential part of statistics. Doing that is fundamental to our future
development, because good models correspond to advancing science and to understanding
how the given data were generated. Methods arising from models built to _t a given data
set stand in contrast to o_-the-shelf" methods. Even so, it is understandable that there
is not always time for model-building with every data set, and that _elds like law and
medicine often require the tradition of o_-the-shelf" methods.
David Schoenfeld: I also like to add onto the issue of merger. Indeed, location does
matter! If the Harvard School of Public Health were to move to Allston, which would
be one and half mile away from the statistics department, it would be better for the
statistics and biostatistics departments to combine because it would create a superpower
7
of a department. For example, Stanford medical school has a division of biostatistics
which doesn't o_er degrees, whose faculty have joint appointments and whose biostatistics
students get degrees in statistics. The medical school is a short walk from the Statistics
Department. The Department of Statistics at Stanford is a top school in Biostatistics
with many recognized leaders in Biostatistics including Brad Efron who is here today. We
compete with Stanford to attract students. They are strengthened by having a combined
department.
L.J. Wei: I must confess that I have closely followed Brad Efron's work and I even
know that when interviewed by Carl Morris (see Holmes et al., 2003), Brad Efron once
said `inference only exists in statistics departments!' I raise this issue because after having
been working with bioinformatics people for 5 years as of now, I still _nd it rather di_-
cult to convince bioinformaticians that inference is important. Here are my questions to
Efron. Have we not sold probability and inference well or have we explained uncertainty
satisfactorily? Or is the issue something else?
Brad Efron: Our profession has somewhat been bifurcated into two levels: methods
development, which is the fun part, and inference. Inference concerns with what we actually

learn, which can be less fun because it is sometimes discouraging and does not work well.
But it is the crucial part that we provide to the other _elds. If we do not, science su_ers!
David Harrington: This pattern repeats itself whenever a discipline without statistical
training starts relying on empirical methods. For example, statisticians had worked
out most of the details of cancer clinical trials, but at the beginning of the HIV epidemic,
statisticians had to start from scratch and re-explain clinical trial methodology to infectious
disease researchers not used to large clinical trials with less than perfect adherence.
As reproducibility becomes more important in bioinformatics, uncertainty and measuring
uncertainty will become increasingly relevant and the issue will work itself out eventually.
Yi Li: I would like to elaborate on the lack of reproducibility of the statistical _ndings
for modern microbiological experiments. This was actually an issue Marvin Zelen discussed
with me prior to this meeting. For example, given a disease and millions of genes and
through some magic of statistics, one will almost surely _nd several highly signi_cant
genes or _nd some highly signi_cant pathways that are related to this disease. However, if
these experiments were to be repeated, the signi_cance would likely vaporize and some new
signi_cant" genes and pathways would pop up! It seems that we just do not have adequate
validation tools in this _eld. I worry that this will not be bona _de for our profession as
we appear to lose credibility in the eyes of subject-matter scientists. Should we take one
step back and try to conduct more focussed" learning? Perhaps, it is more viable to link
a particular gene to a variety of diseases than blindly _nding a bunch of non-repeatable
genes for a given disease.
Brad Efron: I've been worried about our credibility too. Part of the problem is power:
8
with 10,000 genes and 100 subjects, most of the "interesting" genes will be missed in any
one analysis, just as you say. This is one of the concerns expressed in Efron (2008). But I've
also been concerned about overly optimistic statistical methods that don't take into account
all the things that can go wrong in massive data studies. Both of my talks at Harvard, titled
by Are a Set of Microarrays Independent of Each Other?" and Simultaneous Inference:
When Should Hypothesis Testing Problems Be Combined?", involved things that can go
wrong.
James Ware: I have some comments on Colin Begg's response to Q.3 and Q.4. I
would take the de_nition of a method being successful as its being well-known. Methods
that are well-known, such as Kaplan-Meier and proportional hazards models, generally

communicate results in a way that is apprehensible to colleagues in other disciplines. It
is important that the results are less opaque so that our colleagues can understand them.
This is also why some mathematically sophisticated methods do not fare very well.
Now let me raise the following issue to Efron, who is an institutional leader at Stanford.
The DFCI has had more success at integrating computational biology and biostatistics
because it is a research institute and so can be more aggressive. On the other hand, the
HSPH, as an educational enterprise, has struggled because computational biology needs
to be placed within a department but departments have professional boundaries, and it is
unclear how far outside of a department's scope we can reach. So my question to Efron is
this. What is the scope of computational biology that comfortably _ts in the statistics?
Brad Efron: Stanford has a bioinformatics group in the medical school, which focuses
on descriptive statistics, database management, and some sophisticated software development.
The statistics department does not have man-power to do. The bioinformatics
group is often a client for the statistics department, but their relationships, as Ware raises,
are institution dependent. It often takes a leader or some key persons to bring groups
together. For example, Lincoln Moses almost single-handedly put Stanford biostatistics together.
However, when we once talked about the merger of the Bioinformatics department
and the Health Research and Policy Department (our version of School of Public Health
which contains Biostatistics) at Stanford, it was decided that the merger was of neither
side's interest.
I'd like to address the issue of useful methods that Ware just raised. Simple and useful
methods are good, but we are often the ones who tell scientists what is simple. If we stick
to our guns we can convince them of the utility of a method. For example, ANOVA is
in fact very complicated, but we have managed to sell it well to people who could _nd it
useful.
Winston Hide: As a bioinformatician, I would like to make two comments. The _rst
is that the statistical contributions to the algorithms and methodological developments,
for example, the gene sequence searching algorithm, in bioinformatics have proved to be
9
very valuable. Secondly, Bioinformaticians recognize and appreciate the need for strong
statistical knowledge for robust analyses and that they need to learn what they do not
know about data from statisticians. The two groups should be brought close together. We
did have some successful experience in my institute in South Africa.

David Harrington: It appears the solution to L.J. Wei's issue is on the way! Now
I would like to ask my fellow panel members and Louise Ryan, the Chair of Harvard
Biostatistics, the following questions. It is a question on whether our statistical methods
developed are sometimes too opaque or at least too opaque to other _elds. Do we train
our students, the future statisticians, to be scienti_cally literate so that when they have
a method they understand it will make a contribution to another _eld? Colin Begg, from
the perspective of a recruiter who has done much hiring, do you _nd your recruits are
scienti_cally literate so that they are good at picking out problems that are actually useful
and important?
Colin Begg: There is wide variety in the breadth of scienti_c knowledge of statistics
graduates. Some graduate are very knowledgeable about the scienti_c background of data,
but it really depends on the candidate.
Brad Efron: These issues are time-dependent. When Carl Morris and I were graduate
students, Stanford statistics was dominated by mathematics. Now statistics has moved
closer to science. This movement is certainly positive since it broadens the base of the
_eld.
Jun Liu: My comments are related to LJ Wei's question and Efron's answer. Under
the current funding mechanisms there is a great emphasis on novel methods, for example,
from non-statistical study sections. These methods appear attractive but are often inferior
to existing methods. I am wondering how we can inuence the funding mechanisms for
promoting truly useful work.
We statisticians can of course do some clean-up work. For example, my recent two
papers were pretty bad" in the sense of criticizing two papers in Cell and Science for poor
work in prediction and robustness of analysis. I do not know, however, whether such e_ort
has any e_ect. If it does, we make enemies; otherwise we are wasting our breath. I am not
sure what to do because based on my review experience simpler methods get shot down
quickly. Simple methods appear successful or vice versa because we are looking at them
retrospectively. Simpler and smarter methods are the hardest ones to _nd.
Colin Begg: As a _eld, which is not only limited to biostatistics/statistics, we are
not self-critical enough of our methods, and would greatly bene_t from more attention
being paid in our journals to critical evaluation of complex new methods, as opposed to
the overwhelming focus on novelty that seems to dominate our collective research output.
That is, a more pro-active debate in the literature of the strengths and weaknesses of new

10
and existing methods would be bene_cial.
Louise Ryan: The conversation has moved so quickly that I really do not know where
to jump in! Returning though to David Harrington's question on producing scienti_cally
literate student and on what methods actually are good or useful, I would like to say that
methods do not need to be simple to be popular. Methods that _ll a void or gap related
to an interesting problem or question are often the best. The problem however lies in that
methods developed for applied questions are sometimes not rewarded su_ciently because,
for example, promotion committees tend to emphasize methodological journals too heavily.
The issue is also tied to the funding mechanisms as Jun Liu mentioned. Sometime we are
jealous of people from other _elds, for example, computer scientists who take the o_-theshelf"
methods and come up with quick answers for compelling problems, because we often
like to dig into the problems and _nd the optimal solutions. However, by the time we have
found the optimal answers, the problem has been solved and the question of _nding the
optimal solution has become irrelevant. So our struggle lies in the fact that we are trained
to establish the mathematical foundation of problems but our e_ort is often unappreciated
by the academic reward systems. It appears that we do need to learn to how balance theory
and application, especially under the tight funding situations that we currently face.
While I am here, I can't resist commenting on Q.6! I very much like what Speed (2007)
says. In my own _eld of application, environmental health research and risk assessment,
the variability and uncertainty arising from scienti_c mechanisms, bias attributable to data
problems and so on, often completely swamp the variability we statisticians tend to focus on.
My concern is that we as a profession take too narrow a view of inference and uncertainty.
Probability is surely very important, but, as Terry argues, we do need to think about it in
a much broader way. We need to look beyond the con_nes of traditional statistical analysis
and think about the real underlying questions of interest. Sometimes there are exciting
opportunities to use our tools of probability and inference to explore bias, to incorporate
multiple sources of data and to think in terms of synthesizing information to aid in e_ective
and appropriate decision-making.
David Harrington: I strongly believe that biostatistics is most successful when it is
embedded in a _eld. For example, biostatisticians can avoid bias and reduce uncertainty in
AIDS and cancer trials by good design. As I mentioned to our statisticians at the DFCI, a
well designed clinical trial is one that can be explained to a high school student! Of course,

outside the experimental world, like in the observational studies, it is much harder to do
so.
In fact the integration of our _eld with the science has put us in the frontlines to handle
the uncertainty arising from applied projects. A merger of biostatistics and statistics would
probably not be bene_cial as it would take biostatistics away from the front-lines and take
biostatistics to a place where it is di_cult to reward for interdisciplinary research.
11
Cheng Li: Going back to L.J. Wei's comments on the di_culty of convincing bioinformaticians
to believe in statistical inference, I'd like to share some observations from
my collaborative experience in genomics. At the initial stage, molecular biologists want
to identify top genes from genomics studies and use these genes for further experimental
validation or hypothesis generation. Complex inference methods tend to give similar conclusions
as simpler and faster methods in these initial analyses. This may partially explain
why inference is not widely appreciated by biologists in genomics analysis.
It is a rare chance that we are sitting in the same room with Brad Efron, Colin Begg
and many other top statisticians! As a junior faculty member, I would like to ask the senior
folks on what problems junior faculty should work on. What kind of work will likely attract
funding or have long-term impact?
David Harrington: Ask biologists you work with for questions, because they know
what is going on. You may need to translate those questions into those which can be posed
as statistical inference problems. We hire junior faculty like those at this meeting because
of their potential to both extend our discipline and to conduct signi_cant joint research
with other scientists as well.
Brad Efron: Work on what is interesting to you personally because good science
cannot come out of rote work. Biological questions have opened up the statistical _eld,
with massive inference being a typical example. Classical 20th Century statistics developed
superb theories for estimation and testing in small problems, but not for the kind of massive
inference situations we see for example in microarray situations. Notably missing are
convincing theories of prediction and model selection, so these are promising open areas for
21st Century statisticians. Personally, I _nd it much easier to work on something when I
have a compelling data example, so hanging around a good source of data is a useful tactic.
Armin Schwartzman: Are statisticians su_ciently involved in the most important
scienti_c problems in the world today? For example, climate change is an important and

emerging issue that intrinsically involves statistical questions which remain unaddressed.
Where do we stand and where should we go?
Colin Begg: It is important to think big, but unfortunately most statisticians think
small because it helps move our careers along more quickly. This question goes to the
heart of the issue of what value statisticians add to scienti_c debates. We have value to
add, but not everyone thinks statisticians are crucial to this area of research so we need to
prove ourselves. Also, we cannot assume we are the best people to analyze the data simply
because we are trained statistically. Some scientists do have a good sense of making right
decisions for their data even without in-depth statistical training.
Chris Paciorek: As a junior statistician, I am asking whether the current incentive
structure is appropriate in our _eld for faculty evaluation. Can we go beyond counting
12
the number of papers published and pay more attention to the work that generates more
discussion and point-counterpoint debate?
Colin Begg: This is a very live issue in our _eld and beyond. Promotion committees
struggle with the issue since we are increasingly involved in Team Science. In this situation
everyone on a team needs recognition in a promotion system, but it is often hard to pick
out individual contributions for evaluation. This is an issue because throughout science
there is a deep seated tradition to give priority to leaders and to independent research.
David Harrington: Promotion evaluation involves so much uncertainty because we
have to use 6-10 years of research as a surrogate for a lifetime of research. It is an imperfect
surrogate, but we do need to make a decision somehow. In some cases at the Harvard
medical school, the process can take as long as 15 years, but still I am not sure how well it
works. The easy solutions can be extreme: tenure for everyone or tenure for no one. But
when it may take 15 or even more years for work to ourish, I really do not see there is an
easy _x!
Layla Parast (_rst year Harvard PhD student): I would like to go back to Q.1.
Many students think hard about whether to apply to statistics or biostatistics departments
and merging the departments would get rid of the di_erences and bene_ts that drove students'
decisions. Most biostatistics students have math backgrounds so there is a lot they
can learn from biologists. It is good that the departments are separate and allow for more
freedom for biostatistics students to take classes and pursue interests in non-quantitative
_elds. It would be very nice if there was more collaboration between the Harvard biostatistics

and statistics departments, for example, with inter-departmental talks.
David Harrington: I think this is a great idea.
3 Conclusion
Yi Li thanked the panel again and adjourned the meeting at 3:30pm.
Acknowledgements
We thank the panelists, Colin Begg, Brad Efron and David Harrington, for their e_ort and
passion. We are indebted to David Schoenfeld, David Harrington, Tianxi Cai, Cheng Li,
Christopher Paciorek, Armin Schwartzman, Dave Zhao and Layla Parast who contributed,
among many others, the interesting and stimulating discussion questions and topics. We
thank Marvin Zelen, Louise Ryan and James Ware for their constant encouragement. We
are grateful that Andrew Gelman presented his views on the panel discussion in his blog:
13
http://www.stat.columbia.edu/_cook/movabletype/archives/2008/03/some thoughts o 4.html
Reference
Efron, B. (2008) Microarrays, Empirical Bayes, and the Two-Groups Model. Statistical
Science, in press.
Holmes, S., Morris, C. and Tibshirani, R. (2003) Bradley Efron: a conversation with good
friends. Statistical Science, 18, 268-281.
Speed, T. (2007) Statistics without Probability. IMS Bulletin, 36, page 12.
Mosteller, F. and Tukey, J. (1977) Data Analysis and Regression: A Second Course in
Statistics Addison-Wesley Series in Behavioral Science.
14

Discussions on the growth and future of biostatistics

More Related Content

What's hot (20)

Similar to Discussions on the growth and future of biostatistics (20)

More from haramaya university (20)

Discussions on the growth and future of biostatistics