|
Activity Number:
|
116
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, August 3, 2009 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistics in Epidemiology
|
| Abstract - #304814 |
|
Title:
|
Use of Variance Component Estimators to Assess Predictive Model Stability
|
|
Author(s):
|
Michael Jones*+
|
|
Companies:
|
Macquarie University
|
|
Address:
|
Psychology Department, North Ryde, International, 2080, Australia
|
|
Keywords:
|
Predictive model stability ; bootstrapping
|
|
Abstract:
|
Background: Assessment of predictive model stability through independent samples is not always feasible in small samples. Aim: Develop a measure of model stability from training sample data that concords well with estimates from independent samples. Method: Variance components obtained from cross-validated, predicted probabilities across bootstrapped logistic models are used to derive an index of model stability. In a simulation study sw/sb is compared with differences between independent training and validation samples (?). Results: The variance ratio exhibits good stability across models under varying conditions but responds when expected. Importantly, the index correlates well with the difference in model parameters across independent samples. Conclusion: Intensive investigation of a single sample via simulation can yield insights into model stability.
|
