TY - JOUR
T1 - Efficient MCMC for binomial logit models
AU - Fussl, Agnes
AU - Frühwirth-Schnatter, Sylvia
AU - Frühwirth, Rudolf
PY - 2013
Y1 - 2013
N2 - This article deals with binomial logit models where the parameters are estimated within a Bayesian framework. Such models arise, for instance, when repeated measurements are available for identical covariate patterns. To perform MCMC sampling, we rewrite the binomial logit model as an augmented model which involves some latent variables called random utilities. It is straightforward, but inefficient, to use the individual random utility model representation based on the binary observations reconstructed from each binomial observation. Alternatively, we present in this article a new method to aggregate the random utilities for each binomial observation. Based on this aggregated representation, we have implemented an independence Metropolis-Hastings sampler, an auxiliary mixture sampler, and a novel hybrid auxiliary mixture sampler. A comparative study on five binomial datasets shows that the new aggregation method leads to a superior sampler in terms of efficiency compared to previously published data augmentation samplers.
AB - This article deals with binomial logit models where the parameters are estimated within a Bayesian framework. Such models arise, for instance, when repeated measurements are available for identical covariate patterns. To perform MCMC sampling, we rewrite the binomial logit model as an augmented model which involves some latent variables called random utilities. It is straightforward, but inefficient, to use the individual random utility model representation based on the binary observations reconstructed from each binomial observation. Alternatively, we present in this article a new method to aggregate the random utilities for each binomial observation. Based on this aggregated representation, we have implemented an independence Metropolis-Hastings sampler, an auxiliary mixture sampler, and a novel hybrid auxiliary mixture sampler. A comparative study on five binomial datasets shows that the new aggregation method leads to a superior sampler in terms of efficiency compared to previously published data augmentation samplers.
UR - https://dl.acm.org/citation.cfm?doid=2414416.2414419
U2 - 10.1145/2414416.2414419
DO - 10.1145/2414416.2414419
M3 - Journal article
SN - 1049-3301
VL - 22
SP - 1
EP - 21
JO - ACM Transactions on Modelling and Computer Simulation
JF - ACM Transactions on Modelling and Computer Simulation
IS - 3
ER -