Theodore Lystig

"Goodness of fit in hidden Markov models"

abstract:

Hidden Markov models (HMMs) provide an elegant framework for
specifying long range dependencies in longitudinal data.  They can be
thought of as prototypical examples of the increasingly popular class
of models known as latent variable models.  I will present graphical
and numerical techniques for evaluating goodness of fit (GOF) in HMMs.
An explanatory introduction to HMMs will be provided, illustrated with
a dataset of recurrent human papillomavirus (HPV) infections.  The GOF
techniques presented will be used to assess time related model
misspecification, as well as for screening of potentially omitted
covariates.  Both an expansion of the score process and examination of
conditional residuals are demonstrated to be very powerful GOF
techniques in certain situations.  These techniques are applied to the
recurrent HPV infection dataset and (time permitting) to simulated
datasets with known properties.