Dynamic Logistic Regression and Dynamic Model Averaging for Binary Classification

advertisement
David Madigan (Columbia)
Dynamic Logistic Regression and Dynamic Model Averaging for Binary Classification
We present a model for dynamic model averaging for binary classification. The model accounts for
model uncertainty when parameters are expected to change over time by applying a Markov chain
model to the "correct" model and a state space model for the parameters within each model.
Parameterization is done in terms of forgetting and approximation of the marginal likelihood is
computed via Laplace Approximation. We evaluate the method using simulated data and present an
application to laparoscopic surgery in children. (with Tyler McCormick, Adrian Raftery, and Randall
Burd)
Download