(apologies as usual for cross-posting)
I will be giving a one-day short course on Bayesian hierarchical modeling
on Saturday 8 April 2000 from 9am to 5pm at the Interface 2000 meeting (the
32nd symposium on the interface between computing science and statistics)
in New Orleans (the meeting runs from 5-8 April and has as its theme this
year the slightly ungrammatic phrase "modeling the earth's systems:
physical to infrastructure").
The course will be based on a book of the same name which will be published
later this year. Chapters from the book may be downloaded for free, along
with data sets and accompanying software, from my home page
http://www.bath.ac.uk/~masdd
If you may be interested in attending this short course, please see
http://www.neptuneandco.com/interface/
for more details on registration, the Interface meeting itself, etc.
New Orleans is one of the most interesting American cities, a place with a
strong French Creole influence that feels quite European, and it is lovely
to visit in April.
INTERFACE 2000 SHORT COURSE
Saturday 8 April 2000, 9am to 5pm
New Orleans, Louisiana, USA
Sponsored by LearnSTAT
Bayesian Hierarchical Modeling
David Draper
Department of Mathematical Sciences
University of Bath
England
Overview Data Examples
The course "Bayesian I will meet these
Hierarchical Modeling" objectives by exploring
provides an introduction three case studies -- from
to the formulation, education, health policy,
fitting and checking of and engineering risk
hierarchical or multilevel assessment -- with emphasis
models from the Bayesian on the practical
point of view. interaction between
Hierarchical models (HMs) scientific,
arise frequently in three decision-making, and
main kinds of statistical
applications: considerations. Several
other real examples will
1. HMs are common in also be used to illustrate
fields such as health particular concepts.
and education, in Software details required
which data -- both for carrying out the
outcomes and analyses will be provided
predictors -- are in the course materials.
often gathered in a
nested or Target Audience
hierarchical fashion,
e.g., patients within The principal target
hospitals, or audience includes applied
students within statisticians (1) who work
classrooms within with clustered data on a
schools. HMs are thus regular basis, or are
also ideally suited about to begin doing so;
to the wide range of (2) who wish to gain
applications in experience in the modern
government and fitting of random-effects
business in which and mixed models, in
single- or meta-analysis and other
multi-stage cluster settings; and (3) who wish
samples are routinely to learn about current
drawn, and offer a methods for coping with
unified approach to problems of model
the analysis of selection and model
random-effects uncertainty (with all
(variance-components) kinds of data, not just
and mixed models. cluster samples).
2. A different kind of Application areas in which
nested data comes up hierarchical modeling
in meta-analysis in, occurs frequently include
e.g., medicine and policy analysis and other
the social sciences. governmental activities,
In this setting, the agriculture, medicine and
goal is combining health, education, and
information from a biology. Others who may be
number of studies of interested in this course
essentially the same include applied and
phenomenon, to methodological workers who
produce more accurate wish to learn more about
inferences and (4) comparisons in
predictions than complexity and performance
those available from between Bayesian and
any single study. frequentist methods, and
Here the data (5) Markov Chain Monte
structure is subjects Carlo techniques and how
within studies, and to ensure that they work
as in the clustered well in practice. There
case above there will are no formal mathematical
generally be prerequisites, but a
predictors available working knowledge of
at both the subject probability at the
and study levels; and master's level (from such
3. Hierarchical books as Hogg and Craig,
modeling also Bickel and Doksum or
provides a natural Casella and
way to treat issues Berger) -- particularly the
of model selection ability to conceptualize
and model uncertainty and manipulate conditional
with all types of probabilities -- will be
data, not just helpful.
cluster samples. For
example, in Learning Outcomes
regression if the
data appear to Participants will develop
exhibit residual and/or extend facility in:
variation that Formulating appropriate
changes with the hierarchical
predictors, you can (random-effect and/or
expand the model that mixed) models for
assumes constant clustered outcomes in
variation, by meta-analyses and other
embedding it in a studies, both qualitative
family of models that and quantitative, and in
span a variety of situations with predictor
assumptions about information available at
residual variation. some or all levels of the
In this way, instead hierarchy; using Bayesian
of having to choose reasoning and Markov Chain
one of these models Monte Carlo methods to
and risk making the compute posterior
wrong choice, you can distributions for
work with several parameters of greatest
models at once, interest in a given
weighting them in hierarchical model;
proportion to their diagnosing problems with a
plausibility given given hierarchical model
the data. by looking for
discrepancies between
The Bayesian approach is predictive distributions
particularly effective in for observables and the
fitting hierarchical actual values the
models, because other observables take on; and
model-based hierarchically expanding
methods -- based an existing model (for all
principally on maximum kinds of data, not just
likelihood -- sometimes do cluster samples) which
not capture all relevant does not pass all
sources of uncertainty, diagnostic checks, by
leading to over-confident embedding it in a richer
decisions and scientific model class of which it is
conclusions. a special case.
In this course the The Instructor
principles of Bayesian
hierarchical modeling are David Draper is a
described with emphasis on Professor of Statistics in
practical rather than the Department of Mathematical
theoretical issues, and Sciences at the University
illustrated with real data of Bath in England. David
drawn from case studies. did his Ph.D. work at the
The course is intended for University of California,
applied statisticians with Berkeley, finishing in 1981,
an interest in learning and has since taught and
more about hierarchical done consulting and public
models in general, and the policy research at the
Bayesian analysis of such University of Chicago
models in particular. An (1981-84), the RAND
understanding of Corporation (1984-91),
probability at the level UCLA (1991-93), and the
typically required for a University of Bath
master's degree in (1993-Present), with a
statistics provides sabbatical visit to the
sufficient mathematical University of Washington
background. No previous in 1986. He is a fellow of
experience with Bayesian the Royal Statistical
methods is needed -- all Society and a member of
relevant ideas are covered both the IMS and ASA.
in the course in a David has served as
self-contained fashion. Associate Editor for the
Journal of the American
Objectives Statistical Association
and the Journal of the
Participants will learn Royal Statistical Society,
how to translate and is the author or
scientific and coauthor of four books and
decision-making problems 49 articles and other
involving nested or substantial contributions
clustered data into to refereed journals.
appropriate hierarchical
models (including David has been nominated
random-effects and mixed for teaching awards at
models), and will also every university in which
learn how -- with any kind he has taught and was the
of data, not just a recipient of the Quantrell
cluster sample -- to embed Award of Excellence in
a given model Undergraduate Teaching at
hierarchically in a richer the University of Chicago.
model class, as a way to He was also the recipient
realistically approach of an Excellence in
issues of model selection Continuing Education award
and model uncertainty. for his course offerings
Participants will learn on Bayesian hierarchical
methods for computing modeling at the Anaheim and
posterior and predictive Dallas Joint Statistical
distributions for Meetings through the ASA
quantities of interest Continuing Education Program.
arising in the
hierarchical models and to
examine the results of the
model-fitting for
weaknesses and for
sensitivity to modeling
assumptions.
Registration Price:
$375 Interface Participants and ASA Members
$475 Non-Interface Participants and Non-ASA Members
$200 Full-time Students
Registration Code:
3530-2004-01
Please go to the conference website at
http://www.neptuneandco.com/interface/
for further registration information.
