An R package for spatial extreme modelling

Home Learn More Gallery Contact

Max-stable processes

Conditional max-stable simulations

Latent Variable

Copula approaches

Miscellaneous

References

Reference Manual

What are Bayesian Hierarchical models?

Hierarchical models have different layers of variations which must be modelled. When trying to model spatial extremes we can think of (at least) two layers: a layer that determines the marginal behaviour of extremes and another layer that controls the spatial dependence of extremes. Unfortunately because the likelihood of max-stable processes is not available it prevents the use of a spatial dependence layer. Consequently this type of models will never be able to model the dependence between say two weather stations. However its flexibility for modelling the marginal behaviour and will be of great interest if one is interested in modelling only the pointwise distribution of extreme events.
If we dispose of block, e.g., annual, maxima observed at locations , univariate extreme value arguments suggests that these block maxima might be modelled by a GEV distribution. The key idea of the latent process approach is to assume that the GEV parameters vary smoothly over space according to a stochastic process . The SpatialExtremes package use Gaussian processes for this and assume that the Gaussian processes related to each GEV parameter are mutually independent. For instance we take

where is a trend surface depending on regression parameters and is a zero mean stationary Gaussian process with covariance function depending on parameters . Similar expressions might be used for the two remaining GEV parameters.
Then conditional on the values of the three Gaussian processes at the weather stations, the block maxima are assumed to follow a GEV distribution

independently for each location and where and .

Function "latent": Draw a Markov chain from the posterior distribution

The full conditional distributions are easily shown to be

where are prior distributions. The full conditional distributions related to the two remaining GEV parameters give similar expressions.