According to Essential Science
Reuters, the paper currently ranked
at #3 in the field of Mathematics is "Bayesian measures
of model complexity and fit," (Spiegelhalter DJ,
et al., J. Roy. Stat. Soc. Ser. B-Stat. Met. 64:
583-616, Part 4, 2002). For the period ending October
31, 2008, this paper garnered 651 cites.
Lead author Dr. David Spiegelhalter is a Senior Scientist in the
MRC Biostatistics Unit at the Institute of Public Health in Cambridge, UK.
Since October 2007, he also serves as the Winton Professor of the Public
Understanding of Risk in the Statistical Laboratory of the Centre for
Mathematical Sciences at the University of Cambridge. His record in
Essential Science Indicators includes 32 papers cited a total of 1,235
times. He is also a
Highly Cited Researcher in the field of
talks with Dr. Spiegelhalter about this paper and its
impact on the research
What was your inspiration for your
I should first give a bit of background. Bayesian methods for data analysis
allow great flexibility for handling complex problems and including
multiple sources of evidence, but had been hampered for years by
computational problems. Around 1990 it was realized that simulation-based
techniques could be adapted, and this led to an explosion of activity, and
we developed our BUGS software, which has become the most popular program
for carrying out Bayesian analysis.
"… I feel confident that the
basic idea is sound, but still hope that a
better version can be
But we quickly realized this new flexibility meant that people needed a way
of comparing all the different models they could now fit, and no existing
method would do. So it was a real practical demand that drove the research,
and the paper summarizes what our multi-institution team (me from
Cambridge, Nicky Best from Imperial College London, Brad Carlin from the
University of Minnesota, and Angelika van der Linde from the University of
Bremen) came up with after a lot of effort.
How did you develop this new method?
People had come up with all sorts of theoretical ideas, but we needed
something that was reasonably straightforward, could be put into our
software, gave equivalent answers to existing methods in simple
circumstances, and gave sensible answers in new areas. We tried all sorts
of options from 1995 onwards, but then in 1997 we came up with an idea that
seemed to work well and had a reasonable theoretical justification. This
was a very exciting time—we knew we had got something
powerful—although it was another five years before the paper finally
Would you sum up your findings for our readers?
If you fit a complex model to some data, possibly making use of additional
expert opinion, the method provides a way of assessing the essential
complexity of the model that you have fitted and the adequacy with which it
explains the past data; adding these two things up gives you an overall
measure of how well you would expect the fitted model to predict new data.
This final measure is called the DIC (Deviance Information Criterion), and
can be used to compare any models you want. It can be applied generally and
does not require specific software.
How was this paper received by the community?
"Bayesian methods for data analysis
allow great flexibility for handling complex
problems and including multiple sources of
evidence, but had been hampered for years by
Shall we say that there was a mixed reception? Methodological purists have
not particularly liked it because, although the theoretical justification
is reasonable, the specific estimation techniques are quite basic in order
to make them widely applicable. But practitioners have loved it as it did
something they just could not do before.
Where have you taken your research since the publication
of this paper?
We and others have tried to widen the theory but with limited success. We
have identified potential problems with the specific implementation, and
alternative versions have been suggested—I feel confident that the
basic idea is sound, but still hope that a better version can be developed.
An added spin-off is increased interest in the general problem of
estimating the essential complexity of a model and the residual "effective
degrees of freedom."
What should the "take-away lesson" be for this paper and
for your work?
If there is a real practical problem, and you worry away at it long enough
and follow your intuition, then a solution can be found. An additional help
is to provide the means so that others can try your ideas on problems that
really concern them: our software is free and is used by thousands
worldwide, and the citations we have received have largely been from
practical papers that were trying to answer scientific questions. So
another lesson is that a good "product" that fills a hole in the market
will be successful. I just wish that everything I did had the same sort of
MRC Biostatistics Unit
Institute of Public Health
Winton Professor of the Public Understanding of Risk
Statistical Laboratory Centre for Mathematical Sciences
University of Cambridge