geoRglm Developer Page
This page is mainly intended for the two developers of geoRglm
geoR functions used by geoRglm :
* .bilinearformXAY (used by: prepare.likfit.glsm, .maxim.aux1),
* .check.locations (used by: binom.krige, binom.krige.bayes,
glsm.krige, .krige.bayes.extnd, .krige.conv.extnd, pois.krige, pois.krige.bayes, .pred.aux),
* .cond.sim (used by: .krige.bayes.extnd, .krige.conv.extnd),
* .cor.number (used by: .mcmc.bayes.binom.logit, .mcmc.bayes.conj.binom.logit, .krige.bayes.extnd, .krige.conv.extnd, .mcmc.bayes.pois.log, .mcmc.bayes.pois.boxcox, .mcmc.bayes.conj.pois.log, .mcmc.bayes.conj.pois.boxcox),
* .diagquadraticformXAX (used by: .krige.bayes.extnd,
.krige.conv.extnd, prepare.likfit.glsm, .maxim.aux1),
* .geoR_inout (used by: .krige.bayes.extnd,
.krige.conv.extnd, binom.krige.bayes, pois.krige.bayes)
* .ldots.set (used by: image.glm.krige.bayes, persp.glm.krige.bayes),
* .prepare.graph.krige.bayes
(used by: image.glm.krige.bayes, persp.glm.krige.bayes),
* .solve.geoR (used by: binom.krige, binom.krige.bayes, .krige.bayes.extnd, pois.krige, pois.krige.bayes),
* BCtransform (used by: BC.inv),
* coords.aniso (used by: binom.krige, binom.krige.bayes,
glsm.mcmc, .krige.bayes.extnd, .krige.conv.extnd, likfit.glsm, pois.krige, pois.krige.bayes, prepare.likfit.glsm, proflik.glsm), .mcmc.bayes.pois.log, .mcmc.bayes.pois.boxcox, .mcmc.bayes.conj.pois.log, .mcmc.bayes.conj.pois.boxcox),
* cov.spatial (used by: .krige.bayes.extnd, .krige.conv.extnd, lines.covariomodel),
* grf (used by: .pois.log.grf),
* hist.krige.bayes (used by: hist.glm.krige.bayes),
* image.kriging (used in help file for : binom.krige.bayes),
* krige.bayes (used by: binom.krige.bayes, pois.krige.bayes),
* krige.control (used by: .krige.conv.extnd),
* legend.krige (used by: image.glm.krige.bayes),
* locations.inside (used by: .krige.conv.extnd),
* loccoords (used by: binom.krige.bayes, .krige.bayes.extnd, .krige.conv.extnd, pois.krige.bayes),
* model.control (used by: .krige.bayes.extnd),
* output.control (used by: .krige.bayes.extnd, .krige.conv.extnd),
* pars.limits (used by: likfit.glsm),
* plot.1d (used by: image.glm.krige.bayes, persp.glm.krige.bayes),
* prior.control (used by: .krige.bayes.extnd)
* trend.spatial (used by: binom.krige, binom.krige.bayes,
glsm.mcmc, .krige.bayes.extnd, .krige.conv.extnd, likfit.glsm, .pois.log.grf, pois.krige, pois.krige.bayes, prepare.lik.sim, proflik.glsm),
* varcov.spatial (used by: binom.krige, binom.krige.bayes,
glsm.mcmc, .krige.bayes.extnd, .krige.conv.extnd, prepare.likfit.glsm,
.lik.sim, .lik.sim.boxcox, likfit.glsm, pois.krige, pois.krige.bayes,
prepare.likfit.glsm).
In addition, the C-functions corrfctvalue and
geoRmatern are included in both the geoR and the geoRglm
distribution.
geoR classes used by geoRglm :
* kriging (used by: binom.krige, glsm.krige, pois.krige),
* proflik (used by: proflik.glsm).
Future plans/things to be done:
I am currently not working on new features in the package.
The plan for the future is to keep the package running by responding
to bug reports and to changes in R.
Below is list which was compiled some years ago about plans for future
development.
High priority:
- * For fixed parameters the required memory can reduced by
constructing the matrices in the C-part of the code instead of the
R-part of the code.
- * The reparameterisation stuff is not cleverly done from a
numerical point of view. Needs fixing, but is depending on how the
point above (reducing memory by constructing matrices in C) is solved.
- * Default starting value S.start needs some improvement.
- * Consider having a default value of the scaling parameter S.scale
Lower priority (if ever):
- * Approximative methods for large data sets. Presented in a series of
papers by Rue and coauthors. This avoids using MCMC, and is therefore
using a different paradigm than geoRglm. Possibly the better solution is
to start from scratch here.
- * Markov chain Monte Carlo EM algorithm as in Zhang (2002). The
function glsm.mcmc can be used for the MCMC part.
- * One-dimensional profile likelihood in proflik.glsm.
- * Extend likfit.glsm to optimisation w.r.t. anisotropy and
kappa parameters.
- * Using the control() argument in optim in a
automatic way in the optimisation procedures.
- * Extend likfit.glsm to optimisation for fixed phi.
- * Implement the robust MCMC-algorithms (from ideas in the paper with
Gareth and Martin) for Bayesian inference [fixed parameter case is
already implemented].
- * Calculation of predictive mean and variance in binom.krige.bayes (approximative or by simulation).
- * Calculation of quantiles improved in binom.krige.bayes and pois.krige.bayes.
- * Binomial model with probit link.
- * Prediction of the data (incoorporating the measurement error).
- * Composite likelihood as in Varin, Host and Skare
(2005). Christiano Varin may be working on a package for this.
Some day (probably never):
- * Repeated observations at each location.
- * Automatising of sequential procedures (posterior of parameters to be used as priors in a new analysis).
- * Circulant embedding method for correlation matrix in MCMC algorithm (regular grid).
- * Some revision of the functions for the empirical covariogram (other than log-link, and probably some further thinking+simulation studies).
- * Considerations about prior for sigmasq and phi (default priors ?).
History (short) :
- * December 2000 : geoRglm was initiated (originally as an Splus library called glmmS).
- * March 2001 : homepage constructed; version 0.0-0 available on the web; conversion to R.
- * June 2001 : Ole visted Paulo in Lancaster; help pages were made.
- * August 2001 : geoRglm version 0.1-4 used for summerschool in Aalborg.
- * 13 February 2002: geoRglm version 0.4-1 submitted to CRAN.
- * 17 June 2002 : Article in R News about geoRglm.
Last modified: 5 March 2007