|
Activity Number:
|
267
|
|
Type:
|
Invited
|
|
Date/Time:
|
Tuesday, July 31, 2007 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #308013 |
|
Title:
|
Reducing the Variability in Least Squares Cross-Validation Bandwidths
|
|
Author(s):
|
Jeffrey Hart and Simon Sheather*+
|
|
Companies:
|
Texas A&M University and Texas A&M University
|
|
Address:
|
Department of Statistics, College Station, TX, 77843,
|
|
Keywords:
|
density estimation ; bandwidth selection ; cross validation
|
|
Abstract:
|
Least squares cross validation (LSCV) is a popular way of selecting bandwidths for density estimates. A major advantage of LSCV over other methods is that it is widely applicable and it performs better than its competitors when the true density has many modes. However, LSCV is highly variable in practice, and it can dramatically undersmooth, and as such lead to spurious bumpiness in the resulting density estimate. In this talk we show that is possible to reduce the variability in LSCV and to overcome the sporadic, but persistent, cases of dramatic under smoothing by LSCV.
|
