From candy Thu May 2 15:34:07 1996
Subject: Statistics Department Seminar Wednesday, May 8, 1996 (final seminar for the semester)
Status: R
STATISTICS DEPARTMENT SEMINAR
TITLE: PRACTICAL AND THEORETICAL MODEL SELECTION ISSUES
IN SMOOTHING SPLINE ANOVA MODELS FOR ESTIMATION OF
RISK WITH BINARY DATA
SPEAKER: Grace Wahba
DATE: Wednesday, May 8, 1996
TIME: 4:00 P.M.
PLACE: 1221 CSSC
ABSTRACT:
We are interested in the estimation of the probability
of `success' given a sample of observations on subject
predictor variables and outcomes as might be obtained
in a large demographic medical study. Smoothing spline ANOVA
models have been found useful in modeling outcomes
in a variety of data structures and mixes of predictor
variable types with several predictors. These models
reduce to parametric GLIM models when the data so
indicate, otherwise they provide flexible alternatives
to overly rigid parametric models. Smoothing
spline ANOVA models may have indicator variables,
parametric terms and `smooth' main effects, two factor
interactions, and so forth, for the flexible terms.
Interesting questions of both a practical and theoretical
nature arise in deciding which of the terms
should be included in the model. One can look at these
questions with at least three criteria in mind:
Hypothesis testing, prediction, and usefulness and
plausibility to the client, not always leading to
the same result. After a brief review of smoothing spline
ANOVA models with binary data, we will touch on recent
results and approaches from each of these points of view.