Original Paper

Computational Statistics

, Volume 21, Issue 3, pp 603-620

First online:

Semiparametric estimation of mean and variance functions for non-Gaussian data

  • David NottAffiliated withDepartment of Statistics, University of New South Wales Email author 

Rent the article at a discount

Rent now

* Final gross prices may vary according to local VAT.

Get Access


Flexible modelling of the response variance in regression is interesting for understanding the causes of variability in the responses, and is crucial for efficient estimation and correct inference for mean parameters. In this paper we describe methods for mean and variance estimation where the responses are modelled using the double exponential family of distributions and mean and dispersion parameters are described as an additive function of predictors. The additive terms in the model are represented by penalized splines. A simple and unified computational methodology is presented for carrying out the calculations required for Bayesian inference in this class of models based on an adaptive Metropolis algorithm. Application of the adaptive Metropolis algorithm is fully automatic and does not require any kind of pretuning runs. The methodology presented provides flexible methods for modelling heterogeneous Gaussian data, as well as overdispersed and underdispersed count data. Performance is considered in a variety of examples involving real and simulated data sets.


Overdispersion modelling Double exponential models Generalized linear models Variance estimation