Ratio-of-Uniforms Sampling for Bayesian Extreme Value Analysis

Provides functions for the Bayesian analysis of extreme value models. The 'rust' package < https://cran.r-project.org/package=rust> is used to simulate a random sample from the required posterior distribution. The functionality of 'revdbayes' is similar to the 'evdbayes' package < https://cran.r-project.org/package=evdbayes>, which uses Markov Chain Monte Carlo ('MCMC') methods for posterior simulation. Also provided are functions for making inferences about the extremal index, using the K-gaps model of Suveges and Davison (2010) . Also provided are d,p,q,r functions for the Generalised Extreme Value ('GEV') and Generalised Pareto ('GP') distributions that deal appropriately with cases where the shape parameter is very close to zero.


Travis-CI Build Status AppVeyor Build Status Coverage Status CRAN_Status_Badge

What does revdbayes do?

The revdbayes package uses the ratio-of-uniforms method to produce random samples from the posterior distributions that occur in some relatively simple Bayesian extreme value analyses. The functionality of revdbayes is similar to the evdbayes package, which uses Markov Chain Monte Carlo (MCMC) methods for posterior simulation. Advantages of the ratio-of-uniforms method over MCMC in this context are that the user is not required to set tuning parameters nor to monitor convergence and a random posterior sample is produced. Use of the Rcpp package enables revdbayes to be faster than evdbayes. Also provided are functions for making inferences about the extremal index, using the K-gaps model of Suveges and Davison (2010).

A simple example

The two main functions in revdbayes are set_prior and rpost. set_prior sets a prior for extreme value parameters. rpost samples from the posterior produced by updating this prior using the likelihood of observed data under an extreme value model. The following code sets a prior for Generalised Extreme Value (GEV) parameters based on a multivariate normal distribution and then simulates a random sample of size 1000 from the posterior distribution based on a dataset of annual maximum sea levels.

mat <- diag(c(10000, 10000, 100))
pn <- set_prior(prior = "norm", model = "gev", mean = c(0,0,0), cov = mat)
gevp  <- rpost(n = 1000, model = "gev", prior = pn, data = portpirie)

From version 1.2.0 onwards the faster function rpost_rcpp can be used. See the vignette "Faster simulation using revdbayes and Rcpp" for details. The functions rpost and post_rcpp have the same syntax. For example:

gevp_rcpp  <- rpost_rcpp(n = 1000, model = "gev", prior = pn, data = portpirie)


To get the current released version from CRAN:



See vignette("revdbayes-vignette", package = "revdbayes") for an overview of the package and vignette("revdbayes-using-rcpp-vignette", package = "revdbayes") for an illustration of the improvements in efficiency produced using the Rcpp package. See vignette("revdbayes-predictive-vignette", package = "revdbayes") for an outline of how to use revdbayes to perform posterior predictive extreme value inference. Inference for the extremal index using the K-gaps model is described in vignette("revdbayes-kgaps-vignette", package = "revdbayes")


revdbayes 1.3.3

Bug fixes and minor improvements

  • LF line endings used in inst/include/revdbayes.h and inst/include/revdbayes_RcppExports.h to avoid CRAN NOTE.

  • The format of the data supplied to rpost() and rpost_rcpp() is checked and an error is thrown if it is not appropriate.

  • In rpost() and rpost_rcpp() an error is thrown if the input threshold thresh is lower than the smallest observation in data. This is only relevant when model = "gp", model = "bingp" or model = "pp".

  • The summary method for class "evpost" is now set up according to Section 8.1 of the R FAQ at (https://cran.r-project.org/doc/FAQ/R-FAQ.html).

  • A bug in grimshaw_gp_mle has been fixed, so that now solutions with K greater than 1 are discarded. (Many thanks to Leo Belzile.)

  • In grimshaw_gp_mle using the starting value equal to the upper bound can result in early termination of the Newton-Raphson search. A starting value away from the upper bound is now used (lines 282 and 519 of frequentist.R). (Many thanks to Jeremy Rohmer for sending me a dataset that triggered this problem.)

  • In set_prior() if prior = "norm" or prior = "loglognorm" then an explicit error is thrown if cov is not supplied. (Many thanks to Leo Belzile.)

  • The mathematics in the reference manual has been tidied.

revdbayes 1.3.2

Bug fixes and minor improvements

  • The arguments to d/p/q/rgev and d/p/q/rgp now obey the usual conventions for R's dpqr probability distribution functions.

  • In pp_check.evpost the argument subtype is now documented properly.

  • The conf argument to kgaps_mle didn't work properly: conf = 95 was always used. This has been corrected.

revdbayes 1.3.1

New features

  • Bayesian and maximum likelihood inference for the K-gaps model for inferring the extremal index using threshold inter-exceedances times. [Suveges, M. and Davison, A. C. (2010), Model misspecification in peaks over threshold analysis, The Annals of Applied Statistics, 4(1), 203-221. doi:10.1214/09-AOAS292.]

  • New vignette: "Inference for the extremal index using the K-gaps model".

Bug fixes and minor improvements

  • Added the attribute attr(gom, "npy") (with value 3) to the gom dataset. This is for compatability with the threshr package.

  • Give an explicit error message if plot.evpost is called with the logically incompatible arguments add_pu = TRUE and pu_only = TRUE.

  • The documentation for set_bin_prior has been corrected: only in-built priors are available, i.e. it is not possible for the user to supply their own prior.

revdbayes 1.2.1

Bug fixes and minor improvements

  • In some extreme cases (datasets with very small numbers of threshold excesses) calling predict.evpost with type = "q" and x close to 1 returns an imprecise value for the requested predictive quantiles. This has been corrected by using stats::uniroot rather than stats::nlminb.

  • A bug (missing drop = FALSE in subsetting a matrix) in plot.evpred produced an error message if n_years was scalar in the prior call to predict.evpost. This bug has been corrected.

  • The placing of ... in the function definitions of rpost and rpost_rcpp meant that it was not possible to supply the argument r to be passed to rust::ru or rust::ru_rcpp to change the ratio-of-uniforms tuning parameter r. Furthermore, if model = "os" then trying to do this sets ros in error. This has been corrected.

  • A bug meant that the values returned by predict(evpost_object, type = "d") being incorrect if evpost_object was returned from a call to rpost using model = bingp. The values returned were too small: they differ from the correct values by a factor approximately equal to the proportion of observations that lie above the threshold. This bug has been corrected.

revdbayes 1.2.0

New features

  • Faster computation, owing to the use of packages Rcpp and RcppArmadillo in package rust (https://CRAN.R-project.org/package=rust).

  • New function: rpost_rcpp.

  • New vignette. "Faster simulation using revdbayes".

  • set_prior has been extended so that informative priors for GEV parameters can be specified using the arguments prior = "prob" or prior = "quant". It is no longer necessary to use the functions prior.prob and prior.quant from the evdbayes package to set these priors.

Bug fixes and minor improvements

  • The list returned from set_prior now contains default values for all the required arguments of a given in-built prior, if these haven't been specified by the user. This simplifies the evaluation of prior densities using C++.

  • The GEV functions dgev, pgev, qgev, rgev and the GP functions dgp, pgp, qgp, rgp have been rewritten to conform with the vectorised style of the standard functions for distributions, e.g. those found at ?Normal. This makes these functions more flexible, but also means that the user take care when calling them with vectors arguments or different lengths.

  • The documentation for rpost has been corrected: previously it stated that the default for use_noy is use_noy = FALSE, when in fact it is use_noy = TRUE.

  • Bug fixed in plot.evpost : previously, in the d = 2 case, providing the graphical parameter col produced an error because col = 8 was hard-coded in a call to points. Now the extra argument points_par enables the user to provide a list of arguments to points.

  • All the (R, not C++) prior functions described in the documentation of set_prior are now exported. This means that they can now be used in the function posterior in the evdbayes package.

  • Unnecessary dependence on package devtools via Suggests is removed.

  • Bugs fixed in the (R) prior functions gp_norm, gev_norm and gev_loglognorm. The effect of the bug was negligible unless the prior variances are not chosen to be large.

  • In a call to rpost or rpost_rcpp with model = "os" the user may provide data in the form of a vector of block maxima. In this instance the output is equivalent to a call to these functions with model = "gev" with the same data.

revdbayes 1.1.0

New features

  • A new vignette (Posterior Predictive Extreme Value Inference using the revdbayes Package) provides an overview of most of the new features. Run browseVignettes("revdbayes") to access.

  • S3 predict() method for class 'evpost' performs predictive inference about the largest observation observed in N years, returning an object of class evpred.

  • S3 plot() for the evpred object returned by predict.evpost.

  • S3 pp_check() method for class 'evpost' performs posterior predictive checks using the bayesplot package.

  • Interface to the bayesplot package added in the S3 plot.evpost method.

  • model = bingp can now be supplied to rpost() to add inferences about the probability of threshold exceedance to inferences about threshold excesses based on the Generalised Pareto (GP) model. set_bin_prior() can be used to set a prior for this probability.

  • rprior_quant(): to simulate from the prior distribution for GEV parameters proposed in Coles and Tawn (1996) [A Bayesian analysis of extreme rainfall data. Appl. Statist., 45, 463-478], based on independent gamma priors for differences between quantiles.

  • prior_prob(): to simulate from the prior distribution for GEV parameters based on Crowder (1992), in which independent beta priors are specified for ratios of probabilities (which is equivalent to a Dirichlet prior on differences between these probabilities).

Bug fixes and minor improvements

  • The spurious warning messages relating to checking that the model argument to rpost() is consistent with the prior set using set-prior() have been corrected. These occurred when model = "pp" or model = "os".

  • The hyperparameter in the MDI prior was a in the documentation and a_mdi in the code. Now it is a everywhere.

  • In set_prior with prior = "beta" parameter vector ab has been corrected to pq.

  • In the documentation of rpost() the description of the argument noy has been corrected.

  • Package spatstat removed from the Imports field in description to avoid NOTE in CRAN checks.

Reference manual

It appears you don't have a PDF plugin for this browser. You can click here to download the reference manual.


1.3.9 by Paul J. Northrop, a year ago

https://paulnorthrop.github.io/revdbayes/, https://github.com/paulnorthrop/revdbayes

Report a bug at https://github.com/paulnorthrop/revdbayes/issues

Browse source code at https://github.com/cran/revdbayes

Authors: Paul J. Northrop [aut, cre, cph] , Scott D. Grimshaw [ctb]

Documentation:   PDF Manual  

Task views: Bayesian Inference, Extreme Value Analysis, Probability Distributions

GPL (>= 2) license

Imports bayesplot, graphics, Rcpp, rust, stats, utils

Suggests evdbayes, ggplot2, knitr, microbenchmark, rmarkdown, testthat

Linking to Rcpp, RcppArmadillo

Imported by lax, threshr.

Suggested by exdex, fitteR, mev, rust, smovie.

See at CRAN