|
Activity Number:
|
268
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Tuesday, August 4, 2009 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
ENAR
|
| Abstract - #305637 |
|
Title:
|
A Bootstrap Testing Procedure for Building Piecewise Linear Logistic Regression Models with Free-Knot Splines
|
|
Author(s):
|
Scott W. Keith*+ and David B. Allison
|
|
Companies:
|
Thomas Jefferson University and The University of Alabama at Birmingham
|
|
Address:
|
Division of Biostatistics, Philadelphia, PA, 19107,
|
|
Keywords:
|
free-knots ; splines ; logistic ; nonlinear ; bootstrap
|
|
Abstract:
|
Regression splines are widely used as bases in nonlinear modeling. Estimating both the number and locations of knots as free parameters (i.e., join points in a "free-knot" regression spline) can provide highly flexible and parsimonious spline models if the number of knots can be well specified for the given data. To compute parameter estimates for a given number of knots, we applied nonlinear computation algorithms and grid search methods for locating global optima which maximize likelihood equations constructed with piecewise linear free-knot B-splines. We designed and evaluated a bootstrap procedure for testing the significance of adding knot parameters to these models in the context of nonlinearly regressing the logit of a binary outcome on a continuous predictor in the presence of covariates. Our presentation will include summaries of methodological details and simulation results.
|
