The following seminar will take place on Wednesday 5th February
at 2pm in Room A, Wellcome Trust Building at the Institute of Child Health,
30 Guilford Street, London WC1N 1EH.
==========================================================
Multilevel modelling of ordered and unordered categorical responses
Sophia Rabe-Hesketh
Department of Biostatistics and Computing
Institute of Psychiatry, King's College London
Abstract
Multilevel models for categorical responses will be
specified and estimated for a number of datasets using
the Stata program gllamm. Gllamm uses adaptive quadrature
to estimate the models by maximum likelihood, a superior
approach to approximate methods such as marginal quasi
likelihood (MQL) and penalized quasi likelihood (PQL).
The most commonly used multilevel models for ordinal
responses are the random effects proportional odds model
and the random effects ordinal probit model. For these models
the effects of covariates represent unit (or cluster)-specific
effects conditional on the random effects. Population average effects
tend to be smaller and can be investigated by integrating out
the random effects using gllamm. The variance
components are most easily interpreted by viewing the
observed response as arising from an underlying continuous
response crossing thresholds.
Unordered categorical responses such as treatment
or diagnosis can be modelled using the multinomial logit model.
When the response corresponds to a decision, such as
treatment given to a patient, the models can include
category specific covariates (e.g. cost of treatment).
As in the ordinal case, viewing these models in terms of underlying
continous responses facilitates the specification and
interpretation of the random part of the model.
_______________________
Mario Cortina Borja
Senior Lecturer in Statistics
[log in to unmask] Phone +44(0)20 7905 2113 Fax +44(0)20 7242 2723
Paed. Epid. & Biostats, Institute of Child Health, UCL, London WC1N 1EH, UK
|