2015-07-02 Isaac Gravestock
* Fixed comilation bug due to reusing memory and constant vectors.
* Introduced hack to stop clashing of lapack definitions
2013-11-07 Daniel Sabanés Bové
* start adaptations for R version 3.0.2
2013-08-26 Daniel Sabanés Bové
* R/getMarginalZ.R: introduce verbose option
2013-08-22 Daniel Sabanés Bové
* Fixed a bug in sampleGlm which occurred when local empirical
Bayes estimation of g was used, on 64bit machines.
2013-07-03 Daniel Sabanés Bové
* Add offsets.
2013-04-17 Daniel Sabanés Bové
* R/sampleGlm.R: Fix computation of bfp curves and fixed form
covariate coefficients. Before there was a wrong extraction of the
coefficients samples from the grand samples matrix. However, the
fitted values were correct.
2013-04-16 Daniel Sabanés Bové
* src/sampleGlm.cpp: Fix sampling from Cox model. Before the
covariance of the coefficients was wrong.
2013-02-06 Daniel Sabanés Bové
* R/glmBayesMfp.R: Remove default value for "family" to avoid
erroneous fitting of Gaussian models when instead a logistic
regression is to be performed.
2013-01-24 Daniel Sabanés Bové
* Include option "fixedg" to do a model search with fixed g.
2013-01-23 Daniel Sabanés Bové
* R/plotCurveEstimate.R: Mention colors of pointwise (blue) and
simultaneous (green) credible intervals in the help page.
2012-12-10 Daniel Sabanés Bové
* R/glmBayesMfp.R: fix bug: order according to survival times when
Cox model is requested, otherwise the computed deviances are
wrong!!
2012-12-07 Daniel Sabanés Bové
* R/GlmBayesMfp-methods.R: change default for freq in
as.data.frame.GlmBayesMfp to FALSE, to not generate error messages
on objects resulting from exhaustive model search.
2012-12-04 Daniel Sabanés Bové
* Finished inclusion of Cox models.
2012-11-29 Daniel Sabanés Bové
* Start developing functions for Cox models.
2012-11-26 Daniel Sabanés Bové
* R/plotCurveEstimate.R: add option to include zeros resulting
from models which do not include a covariate in the computations
for the curve estimates.
Also correct colors.
2012-11-21 Daniel Sabanés Bové
* Start adding likelihood ratio statistic based Bayes factors (TBF
methodology).
2012-10-24 Daniel Sabanés Bové
* R/GlmBayesMfp-methods.R: correct extraction of Monte Carlo model
probability estimates.
* Add vignette with Pima Indian data set example.
2012-07-26 Daniel Sabanés Bové
* Small changes in documentation.
2011-10-07 Daniel Sabanés Bové
* inst/CITATION: Include updated citation info.
2011-07-29 Daniel Sabanés Bové
* various: now the option "higherOrderCorrection" allows the
higher order Laplace correction also for other canonical GLMs
besides the logistic regression. (And it should speed up the
computation, too!)
2011-07-08 Daniel Sabanés Bové
* various: add the "dependent" model prior
2011-06-10 Daniel Sabanés Bové
* R/glmBayesMfp.R: Correctly abort when asked and not "y" or "n"
is entered (thanks Alex!)
2011-06-09 Daniel Sabanés Bové
* R/sampleGlm.R: Fix bug in construction of "newDesign" for
"predictions" (thanks Alex!)
2011-02-18 Daniel Sabanés Bové
* inst/CITATION: Add citation info.
2010-12-09 Daniel Sabanés Bové
* src/various: remove commented code fragments, e.g. for
eighth-order Laplace approximation, which were not used. See the
archives for this code.
* various: Switch from FLENS to RcppArmadillo, because then we can
install the package on other machines easily.
2010-08-16 Daniel Sabanés Bové
* src/zdensity.cpp: Higher-order correction is now O(n^-2) after
including two additional terms.
2010-08-03 Daniel Sabanés Bové
* src/sampleGlm.cpp: verbose was falsely initialized with the
estimateMargLik value in the Options class; this was corrected.
* R/sampleBma.R: Always save z samples (even for null / EB)
* R/plotCurveEstimate.R: Now rug=TRUE works for plotCurveEstimate.
2010-08-02 Daniel Sabanés Bové
* src/zdensity.cpp: If g is numerically zero, we do not longer
throw an R error, but a C++ exception which is turned into an NaN
value by the catch-block.
2010-07-29 Daniel Sabanés Bové
* Included a higher-order correction of the Laplace approximation,
which is currently SPECIFIC ONLY FOR BINARY LOGISTIC REGRESSION.
It is available through a new option for glmBayesMfp.
2010-07-26 Daniel Sabanés Bové
* R/sampleBma.R: And also here.
* R/sampleGlm.R: Do not drop dimensions when saving UC
coefficients samples.
2010-07-23 Daniel Sabanés Bové
* R/GPrior-classes.R: only warn when the integral of the g-prior
density is not equal to unity instead of producing an error.
2010-07-08 Daniel Sabanés Bové
* Further shorten the interval where the mode of z is searched to
(-100, 200) because the posterior will usually be very flat and
low below -100.
* Fix a bug in dataStructure.h/Model: Now two models are compared
by their parameters if the info is the same. Before, in the
exhaustive search functions two model could have the same info and
only the first model would be included in the result list.
* Start including an empirical Bayes option, where g is estimated
by maximizing the (approximated) conditional marginal likelihood.
2010-07-05 Daniel Sabanés Bové
* Fixed a bug in sampleGlm.cpp where the unnormalized posterior of
the high density point was rounded to integer. This lead to almost
uniform errors of ILA wrt MCMC marginal likelihoods.
2010-07-02 Daniel Sabanés Bové
* R/sampleGlm.R: extend the functionality to the case with fixed
z. This can be useful later on in the empirical Bayes setting!
For now it's important for checking the marginal likelihood in the
normal distribution case.
2010-06-30 Daniel Sabanés Bové
* R/getNullModelInfo.R: include the 2 * pi constant
2010-06-28 Daniel Sabanés Bové
* R/getNullModelInfo.R: switch back to the Laplace approximation
for the computation of the marginal log likelihood of the null
model.
2010-05-25 Daniel Sabanés Bové
* Remove unnecessary testing code from the package (e.g. the
optimization routines, openMP testing etc.). You can find it in
the old packages in the archive directory if you need it.
* Introduce formal class "GlmBayesMfpSamples". Objects of this
class are produced by the single and model average sampling
functions. The "plotCurveEstimate" function expects this as input.
* Add model averaging function "sampleBma".
2010-05-20 Daniel Sabanés Bové
* Convergence of IWLS is now measured on the coefficients scale,
rather than the linear predictor scale. This is faster (because
there are fewer coefficients). Non-convergence is treated by
issuing a warning message (no R warning, though) in debug mode.
* IWLS errors etc. are no longer breaking the model sampling, but
throw exceptions which are processed. If the error is only
occurring at single z values, it is no problem for the
optimization algorithm -- it just turns away from these values. If
the resulting log marginal likelihood estimate is nevertheless
NaN, the model is not included (but the algorithm proceeds) in the
model cache.
2010-05-17 Daniel Sabanés Bové
* "linear" is now the default approximation method for the
marginal z density, because it always works, does not need long
initialization times like the Runuran generators and gives equally
good acceptance probabilities (provided the density was explored
well enough by the optimization routine).
* "optimize" is now the default optimization routine, because it
is more robust than bfgs (although it is slower, but that is the
price).
2010-05-12 Daniel Sabanés Bové
* Include the dispersions phi_i correctly in the g-prior matrix
(they were not included at all before...) and also the factor c
which is determined by the distribution family and the link
function. So we want the "generalized g-prior" instead of the
usual "linear model g-prior".
2010-04-14 Daniel Sabanés Bové
* Fixed some bugs.
* Added expert options to glmBayesMfp.
2010-03-18 Daniel Sabanés Bové
* Add code for the model sampling, because we now want to explore
larger model spaces where exhaustive computation of all models is
no longer possible.
* Also use OpenMP to speedup some loops. Adapt makefile to new
BLAS/LAPACK situation (we do not need to include ATLAS things any
longer, because we now use GotoBLAS2).
2010-03-10 Daniel Sabanés Bové
* Use FLENS instead of newmat for better performance.
2010-02-17 Daniel Sabanés Bové
* Finalize sampling (better spline approximations for the marginal
z density, correct treatment of the null model, ...)
2010-01-05 Daniel Sabanés Bové
* Use the new unuran.details function in sampleGlm to avoid
another integration of the unnormalized pdf.
2009-12-10 Daniel Sabanés Bové
* Start implementing the MCMC sampler for a specific GLM from the
list returned by "glmBayesMfp".
2009-11-18 Daniel Sabanés Bové
* First only focus on the new glm methodology in this package, and
gradually build it up. So move the hyper-g C code to separate
directory (because I already changed some things there) and remove
the hyper-g R code. (It is still present in the bfp package of
course!) Start with testing of marginal likelihood approximation
in small model spaces where we can do exhaustive searches.
2009-10-26 Daniel Sabanés Bové
* branch off this package ("glmBfp"), which shall include
(one-parameter) generalized linear models. This will *not* be on
R-forge in the first development time, as opposed to the old "bfp"
package.
2009-09-30 Daniel Sabanés Bové
* Add some documentation to R functions.
* Clearer structure for src directory, so that all newmat stuff is
in a separate directory.
2009-09-29 Daniel Sabanés Bové
* Upload to R-forge SVN as package "bfp".
2009-09-21 Daniel Sabanés Bové
* Only allow identical FP degrees for the covariates, to ensure
that the SWITCH always works.
* Start implementing the new SWITCH move type, which shall guard
us against bad mixing in data sets with near-collinear covariates.
* Note that the R CMD check warning for the marked UTF8 character
is the degree symbol in the labels of the ozone covariates.
2009-09-14 Daniel Sabanés Bové
* start cleaning before the next major revisions
2009-02-12 Daniel Sabanes Bove
* faster implementation for findModel.
2008-12-26 Daniel Sabanes Bove
* test the corrected computation of marginal likelihood (see
hyperg.cpp).
2008-11-29 Daniel Sabanes Bove
* Added flat model prior option as the default prior specification.
2008-10-24 Daniel Sabanes Bove
* First version which is tested for the ozone data.
2008-08-22 Daniel Sabanes Bove
* copied everything from original BayesMfp package. This shall be
a seperate implementation for the new methodology with
hyper-g-prior & Co .