A Report on the NSF ICORS 2002 Computational Workshop

1
A Report on the NSF ICORS 2002 Computational
Workshop

• Arnold Stromberg
• Department of Statistics
• With support from the
• National Science Foundation

ICORS 2003 Antwerp, Belgium July 18, 2003
2
Background

• The NSF sponsered International workshop on
  Computational Methods for Robust Statistics was
  held immediately following ICORS 2002 in
  Vancouver, BC, Canada from May 18 to May 20, 2002

3
Invited Presentations

• Doug Martin of Insighful, Inc., the makers of
  Splus, discussed the need for emphasis on
  computation of robust statistical techniques
• Colin Chen of SAS, Inc. gave a demonstration of
  Proc Robustreg which he wrote to do robust
  regression. It will be available with SAS,
  Version 9 in Fall, 2003

4
Comments Software Availability

• SAS users, primarily applied statisticians, can
  now do robust regression, thus applying what they
  learned in graduate school.
• Nonstatisticians, who will never use SAS, are
  starting to use R, and by extension, Splus.
• We must write code in R. Good methods will be
  incorporated by Splus, and eventually SAS.

5
Research Directions

• Literally hundreds of statistical procedures are
  in need of robustification. By providing
  computational tools, especially in R,
  statisticians can compare classical and robust
  methods.
• The key is to make R code user friendly so
  nonstatisticians can use them.

6
Why make your code user friendly?

• Disadvantages
• It takes extra time.
• It doesnt help my career.
• It isnt fundable.
• Writing code is no fun.
• I cant publish it.
• No one will use it.

7
Why make your code user friendly?

• Advantages
• It takes time, but is publishable in Journal of
  Statistical Software.
• www.jstatsoft.org
• Abstracts published in Journal of Computational
  and Graphical Statistics (JCGS)

8
Journal of Statistical Software

• Types of Papers JSS will publish
• Manuals, user's guides, and other forms of
  description of statistical software.
• The code for new statistical software.
• Data sets that are of use to statisticians.
• Reviews and comparisons of statistical software.

9
Publishing in JSS

• The typical JSS paper will have a section
  explaining the statistical technique, a section
  explaining the code, a section with the actual
  code, and a section with examples. All sections
  will be made browsable as well as downloadable.
  The papers and code should be accessible to a
  broad community of practitioners, teachers, and
  researchers in the field of statistics.

10
Why make your code user friendly?

• More advantages
• It does help your career because
• It increases publications.
• It is fundable! NSF wants innovative and useful
  projects. Useful means, others can and do use
  it. That means user friendly code. NSF rarely
  funds straight theory anymore.
• Nonstatistician who are evaluating you appreciate
  it.
• Many statisticians appreciate it.

11
Why make your code user friendly?

• More Advantages
• Writing code may not be fun, but seeing
  researchers use your methods is lots of fun!
• If you method is useful, researchers will use it
  if they know about it and have the tools.
  Nonstatisticians dont read statistics journals
  so we must publish in their journals!

12
Collaborators in other fields

• Why you need them
• They have real problems.
• They at least double your funding options.
• They at least double your publication options.
• They will support you and your department

13
The Case of Robust Regression

• Everyone agrees its useful.
• Nearly 40 years after M-estimates, they are in
  SAS!
• What happened to the 10 year rule?

14
Why Did it take 40 Years?

• Only theory mattered.
• Computationally difficult at first.
• No two statisticians could agree on the best
  robust method.
• Computationally harder with high breakdown.
• SAS mentality.

15
Funding Statistical Research

• No one funds straight theory.
• NSF wants useful and innovative.
• Useful Collaborators
• NIH want medical applications.
• Not the next robust regression estimator,
  although JASA might accept it.
• Everyone funds conferences and workshops!

16
COMMENTS
17
Collaborations resulting from the workshop

• Ma, Y., Genton, M. G. (2002) "A semiparametric
  class of generalized skew-elliptical
  distributions," Institute of Statistics Mimeo
  Series 2541,under review.
• Nora Muler is working together with Victor Yohai
  on robust estimators for GARCH models. I proposed
  at ICORS 2002 a robust estimator for the general
  GARCH(p,q) model that has pq1 parameters to
  estimate. The algorithm in our paper is
  implemented only for the GARCH(1,1) case. I was
  discussing how to generalize the algorithm to the
  general GARCH(p,q) case.

18
Collaborations resulting from the workshop

• Robust Methods for Microarray Data Analysis by
  Hanga Galfalvy, Steven Grambow,Johanna Hardin and
  Arnold Stromberg was started and extensive
  progress has since be made.
• Rocke, D.M., and D.L. Woodruff, "Multivariate
  Outlier Detection and Cluster Identification",
  Working Paper

19
Collaborations resulting from the workshop

• Chen, et. al. extensively discussed smoothing
  algorithms for quantile regression which will
  become part of SAS soon.
• Discussions lead to Salibian-Barreras
  Estimating the p-values of robust tests for the
  linear model. Now under revision for JSPI

20
Workshop Benefits

• Salibian-Barrera (2003). Estimating the p-values
  of robust tests for the linear model. Now under
  revision for JSPI
• Attending the workshop assisted Matias
  Salibian-Barrera and two colleagues application
  for a large grant for a computer lab here (over
  980K)... The agencies were CFI
  (www.innovation.ca) and OIT (www.oit.on.ca). They
  were awarded the grant in October 2002.

21
Computational Issues

• Discussed possibilities for new algorithmic
  strategies for efficient computation of robust
  estimators such as LTS
• Discussed current research on theoretical
  properties of algorithmic estimators... in
  particular, the recent work of Hawkins and Olive.
• Discussed the impact of new robust procedures in
  SAS and Splus on data analysis as well as the
  impact on the ability to develop and run
  simulations for research in robust methods.
• Discussed potential applications of traditional
  robust estimators to be emerging field of
  microarray/gene expression analysis.

22
More Computational Issues

• Many robust techniques are only computable for
  small data sets, but the larger the data set, the
  more likely robust techniques should be used.
• Success stories
• Fast LTS, Fast MCD.
• Others?????????

23
More Computational Issues

• The need for computation of robust singular value
  decomposition for large matrices (with
  application to microarray data).
• The need for computation of robust variogram
  estimator in spatial statistics
• The need for investigation of similarities
  between support vector machines and robust
  regression
• The need to detect outliers in asymmetric
  distributions

24
More Computational Issues

• differences between SAS's and Splus' s
  implementations of robust regression
• The appropriateness of using the robust Wald or
  robust F tests when using ANOVA to compare two
  nested robust regression models

25
More Computational Issues

• how to deal with categorical variables for high
  breakdown methods
• how to handle multiple root problems for
  re-descending M-estimators
• Issues in robust SVD. One thing proposed is to
  use norm L1 instead of the usual norm L2, the
  problem is the orthogonality property is lost.

26
More Computational Issues

• In the Skew-Symmetric type distributions, the
  coefficients in the skewing function is very
  sensitive to even small amount of outliers. In
  fact, a small amount of outlier will demand
  increasing the order of the polynomial in the
  skewing function, yet the extra coefficients are
  very hard to estimate.

27
More Computational Issues

• Subsampling strategies
• Empirical analysis of estimator performance
• Robust metrics
• Compute high breakdown value estimates with both
  continuous and categorical variable
• Compute multivariate robust estimates
•   Smoothing algorithm for regression quantile

28
More Computational Issues

• Fast and robust bootstrap methods for robust
  regression estimates
• Fast and robust estimates for p-values for robust
  regression
• Fast computation of MM-regression estimates for
  high-dimensional data...Maybe related to
  Hoaglin-Mosteller-Tukey's sweeping method?

29
Would you do it again?

• I would definitely be interested in
  participating in another workshop. Attendance at
  the 2002 workshop proved to be invaluable. I
  learned about new research going on in the robust
  statistics field, initiated several new research
  collaborations with other conference attendees,
  and had the opportunity to meet several key
  researchers in the robust field. All in all, it
  was an excellent experience.

30
Would you do it again?

• Yes, especially after SAS and Splus release their
  robust procedures formally.

31
Workshop Participants

• Chen, Colin (Lin) . SAS INSTITUTE, INC.
• Galfalvy Hanga C. New York State Psychiatric
  Institute
• Garcia Ben Marta Universidad de Buenos Aires
• Genton Marc North Carolina State University
• Grambow Steve Duke University Medical Centre

32
Workshop Participants

• Hardin,Johanna, PostDoc, Fred Hutchinson Cancer
  Research Center, Seattle Washington
• He Xuming, Professor, Department of Statistics,
  University of Illinois
• Kafadar, Karen, Professor, Department of
  Mathematics, University of Colorado at Denver
• Lin Nan
• Ma,Yuanyuan North Carolina State Univerisity

33
Workshop Participants

• Muler,Nora, Junior Faculty,Universidad Torcuato
  di Tella, Argentina
• Stromberg Arnold, Professor, Department of
  Statistics, University of Kentucky
• Tyler, David, Professor, Department of
  Statistics, Rutgers University"
• Vanden Branden Karlien, Graduate Student,
  Katholieke Universiteit Leuven, Belgium.
• Werner, Mark, Graduate Student, University of
  Colorado at Denver

34
Workshop Participants

• Woodruff ,David, Professor, Department of
  Mathematics, University of California at Davis
• Zamar, Ruben, Professor, Department of
  Statistics, University of British Columbia
• Ekblom Hakan, Professor, Lulea University of
  Technology, Sweden
• Sinha, Sanjoy Assistant Professor, University of
  Winnipeg, Canada