"CANCELLED/POSTPONED AS OF 4-24-97 DUE TO SNOW"
"CANCELLED/POSTPONED AS OF 4-24-97 DUE TO SNOW"
"CANCELLED/POSTPONED AS OF 4-24-97 DUE TO SNOW"
1997 Spring Meeting Colorado/Wyoming Chapter of ASA
Friday, April 25
Guest Speaker: Alan Agresti
****FINAL SCHEDULE ************* Updated 4-17-97************
8:30 ** REGISTRATION **
9:00 . Estimating the Number of Unseen Species and the Prediction Function
in Sampling with Replacement (Shahar Boneh PhD, Metro State)
9:20 . How Many? (Brad Warner PhD, USAFA)
9:40 . Lorgarithmic Pooled Bayesian Inference when Incoherent Priors are
Linked by a Deterministic Simulation Model (Paul Roback, CSU)
10:00 ** BREAK **
10:10 . Causal Inference and the Problem of Estimating the Variance of
Treatment Effects (Gary Gadbury, CSU)
10:30 . Study of Different Upper Bounds for Survivor Distribution of
Nonnegative Random Variable (Philippe Naveau, CSU)
10:50 . A Monte Carlo Evaluation of the Use of the Nelder Mead Modified
Simplex Method to Find the Region of Optimal Dose Therapy Combinations
With and Without Toxicity Constraints in a Phase I/II Clinical Trial
(Lois Larsen, UNC)
11:10 ** BREAK **
11:20 . New Upper Bound for the Distribution Function of a Central Normal
Positive Definite Quadratic Form (Said Al-Karni, CSU)
11:40 . Interrater Reliability: Rating the Raters (Rick Gumina, CSU)
12:00 ** LUNCH **
1:30 . KEYNOTE: Exact vs Approximate Inference for Binomial Proportions
(Alan Agresti PhD, U of Florida)
2:20 **BREAK **
2:30 . A Multivariate Application of the Q Chart (Jay Schaffer, UNC)
2:50 . A Model for Spatially Correlated Binary Responses
(Molly Van Caster, CSU)
3:10 . ML Estimates for Multivariate Double Bounded Survey Data
(Mary Riddel, CSU)
3:30 ** BREAK **
3:40 . Scompile An S-plus Compiler (Matt Calder, CSU)
4:00 . Exact Confidence Regions for Variance Components in Mixed Models
(Chang-Yue UENG, CSU)
4:20 ** CLOSING REMARKS **
Exact vs. Approximate Inference for Binomial Proportions
Alan Agresti
University of Florida
ABSTRACT
This talk presents two short excursions about inference for binomial
proportions. In one case, an exact small-sample method is preferable to
large-sample methods, but in the other case a large-sample method actually
makes for better statistical practice than the exact approach. First we present
a test formulated in response to a consulting question in a clinical trail about
whether a binomial probability is a monotonically increasing function of the
dosage level of a drug. For the sparse multi-center data set, a large-sample
solution is inappropriate, but simulating the exact distributions of likelihood-
ratio test statistics provides reliable results.
Second we discuss confidence intervals for binomial probabilities. Although
the crude approach based on inverting a large-sample test with estimated
standard error has very poor performance, E.B. Wilson’s suggestion in 1927
of inverting the related test with null standard error performs very well even
for small samples. We argue that it is preferable even to the Clopper-Pearson
exact approach. For 95% confidence intervals, Wilson’s approximate method
is essentially the same as the crude approach with estimated standard error
(i.e., sample proportion plus and minus two estimated standard errors) if one
first adds two successes and two failures to the data. This simple formula
works surprisingly well even for very small sample sizes.
|