Package for Bayesian Variable Selection and Model Averaging
in linear models and generalized linear models using stochastic or
deterministic sampling without replacement from posterior
distributions. Prior distributions on coefficients are
from Zellner's g-prior or mixtures of g-priors
corresponding to the Zellner-Siow Cauchy Priors or the
mixture of g-priors from Liang et al (2008)
Fixed errors identified on cran checks https://cran.r-project.org/web/checks/check_results_BAS.html
initialize R2_m = 0.0 in lm_mcmcbas.c (lead to NA's with clang on debian and fedora )
switch to default of pivot = TRUE
in bas.lm
, adding tol
as an argument to control tolerance in cholregpovot
for improved stability across platforms with singular or nearly singular designs.
valgrind messages: Conditional jump or move depends on uninitialised value(s). Initialize vectors allocated via R_alloc in lm_deterministic.c and glm_deterministic.c.
Included an option pivot=TRUE
in bas.lm
to fit the models using a pivoted Cholesky decomposition to allow models that are rank-deficient. Enhancment #24 and Bug #21. Currently coefficients that are not-estimable are set to zero so that predict
and other methods will work as before. The vector rank
is added to the output (see documenation for bas.lm
) and the degrees of freedom methods that assume a uniform prior for obtaining estimates (AIC and BIC) are adjusted to use rank
rather than size
.
Added option force.heredity=TRUE
to force lower order terms to be included if higher order terms are present (hierarchical constraint) for method='MCMC'
and method='BAS'
with bas.lm
and bas.glm
. Updated Vignette to illustrate. enhancement #19. Checks to see if parents are included using include.always
pass issue #26.
Added option drop.always.included
to image.bas
so that variables that are always included may be excluded from the image. By default all are shown enhancement #23
Added option drop.always.included
and subset
to plot.bas
so that variables that are always included may be excluded from the plot showing the marginal posterior inclusion probabilities (which=4
). By default all are shown enhancement #23
update fitted.bas
to use predict so that code covers both GLM and LM cases with type='link'
or type='response'
Added Code Coverage support and more extensive tests using test_that
.
fixed issue #36 Errors in prior = "ZS-null" when R2 is not finite or out of range due to model being not full rank. Change in gexpectations
function in file bayesreg.c
fixed issue #35 for method="MCMC+BAS"
in bas.glm
in glm_mcmcbas.c
when no values are provided for MCMC.iterations
or n.models
and defaults are used. Added unit test in test-bas-glm.R
fixed issue #34 for bas.glm
where variables in include.always
had marginal inclusion probabilities that were incorrect. Added unit test in test-bas-glm.R
fixed issue #33 for Jeffreys prior where marginal inclusion probabilities were not renomalized after dropping intercept model
fixed issue #32
to allow vectorization for phi1
function in R/cch.R
and added unit test to "tests/testthat/test-special-functions.R"
fixed issue #31 to coerce g
to be a REAL for g.prior
prior and IC.prior
in bas.glm
; added unit-test "tests/testthat/test-bas-glm.R"
fixed issue #30 added n as hyperparameter if NULL and coerced to be a REAL for intrinsic
prior in bas.glm
; added unit-test
fixed issue #29 added n as hyperparameter if NULL and coerced to be a REAL for beta.prime
prior in bas.glm
; added unit-test
fixed issue #28 fixed length of MCMC estimates of marginal inclusion probabilities; added unit-test
fixed issue #27 where expected shrinkage with the JZS prior was greater than 1. Added unit test.
fixed output include.always
to include the intercept issue #26 always so that drop.always.included = TRUE
drops the intercept and any other variables that are forced in. include.always
and force.heredity=TRUE
can now be used together with method="BAS"
.
added warning if marginal likelihoods/posterior probabilities are NA with default model fitting method with suggestion that models be rerun with pivot = TRUE
. This uses a modified Cholesky decomposition with pivoting so that if the model is rank deficient or nearly singular the dimensionality is reduced. Bug #21.
corrected count for first model with method='MCMC'
which lead to potential model with 0 probabiliy and errors in image
.
coerced predicted values to be a vector under BMA (was a matrix)
fixed size
with using method=deterministic
in bas.glm
(was not updated)
fixed problem in confint
with horizontal=TRUE
when intervals are point mass at zero.
suppress warning
when sampling probabilities are 1 or 0 and the number of models is decremented
Issue #25
changed force.heredity.bas
to renormalize the prior probabilities rather than to use a new prior probability based on heredity constraints. For future, add new priors for models based on heredity. See comment on issue #26.
Changed License to GPL 3.0
variable.names
to extract variable names in the highest probability model, median probability
model, and best probability model for objects created by predict
.predict.basglm
which had that type = "link"
was the default for prediction issue #18add na.action for handling NA's for predict methods issue #10
added include.always
as new argument to bas.lm
. This allows a formula to specify which terms should always be included in all models. By default the intercept is always included.
added a section to the vignette to illustrate weighted regression and the
force.heredity.bas
function to group levels of a factor so that they enter
or leave the model together.
fixed problem if there is only one model for image
function;
github issue #11
fixed error in bas.lm
with non-equal weights where R2 was incorrect.
issue #17
predict
argument in predict.bas
, predict.basglm
and internal functions as it is not utilizedconfint.coef.bas
when parm is a character stringBayes.outlier
if prior
probability of no outliers is providedfixed issue with scoping in eval of data in predict.bas
if dataname is defined in local env.
fixed issue 10 in github (predict for estimator='BPM' failed if there were NA's in the X data. Delete NA's before finding the closest model.
fixed bug in 'JZS' prior - merged pull request #12 from vandenman/master
fixed bug in bas.glm when default betaprior (CCH) is used and inputs were INTEGER instead of REAL
removed warning with use of 'ZS-null' for backwards compatibility
updated print.bas to reflect changes in print.lm
Added Bayes.outlier function to calculate posterior probabilities of outliers using the method from Chaloner & Brant for linear models.
Added new method for bas.lm
to obtain marginal likelihoods with the Zellner-Siow Priors for "prior= 'JZS' using QUADMATH routines for numerical integration. The optional hyperparameter alpha may now be used to adjust the scaling of the ZS prior where g ~ G(1/2, alpha*n/2) as in the BayesFactor
package of Morey, with a default of alpha=1 corresponding to the ZS prior used in Liang et al (2008). This also uses more stable evaluations of log(1 + x) to prevent underflow/overflow.
Priors ZS-full
for bas.lm is planned to be deprecated.
replaced math functions to use portable C code from Rmath and consolidated header files
Fixed unprotected ANS in C code in glm_sampleworep.c and sampleworep.c after call to PutRNGstate and possible stack imbalance in glm_mcmc.
Fixed problem with predict for estimator=BPM when newdata was one row
Fixed non-conformable error with predict
when new data was from a dataframe with one row.
Fixed problem with missing weights for prediction using the median probability model with no new data.
Extract coefficent summaries, credible intervals and plots for the HPM
and MPM
in addition to the default BMA
by adding a new estimator
argument to the coef
function. The new n.models
argument to coef
provides summaries based on the top n.models
highest probability models to reduce computation time. 'n.models = 1' is equivalent to the highest probability model.
use of newdata that is a vector is now depricated for predict.bas; newdata must be a dataframe or missing, in which case fitted values based on the dataframe used in fitting is used
factor levels are handled as in lm
or glm
for prediction when there may be only level of a factor in the newdata
fixed issue for prediction when newdata has just one row
fixed missing id in plot.bas for which=3
bas.lm
to agree with documentationrenormalize
that selects whether the Monte Carlo frequencies are used to estimate posterior model and marginal inclusion probabilities (default renormalize = FALSE
) or that marginal likelihoods time prior probabilities that are renormalized to sum to 1 are used. (the latter is the only option for the other methods); new slots for probne0.MCMC, probne0.RN, postprobs.RN and postprobs.MCMC.coefficients
function.bas.lm
and bas.glm
na.action
for bas.lm
and bas.glm
to omit missing data.confint.pred.bas
or confint.coef.bas
. See the help files for an example or the vignette.se.fit
option in predict.basglm
.testBF
as a betaprior
option for bas.glm
to implement Bayes Fatcors based on the likelihood ratio statistic's distribution for GLMs.A vignette has been added at long last! This illustrates several of the new features in BAS
such as
confint.pred.bas()
confint.coef.bas()
predict.bas()
type
to specify estimator in fitted.bas and replaced with estimator
so that predict()
and fitted()
are compatible with other S3 methods.bas
to avoid NAMESPACE conficts with other librariesfitted.bas
or predict.bas
bas.glm
diagnostic()
function for checking convergence of bas
objects created with method = "MCMC"
"plot.bas
that appears with Sweavecoef.bma
when there is just one predictorbas
rather than
bma
to avoid name conflicts with other packages- added weights for linear models
- switched LINPACK calls in bayesreg to LAPACK finally should be
faster
- fixed bug in intercept calculation for glms
- fixed inclusion probabilities to be a vector in the global EB
methods for linear models
- added intrinsic prior for GLMs
- fixed problems for linear models for p > n and R2 not correct
- added phi1 function from Gordy (1998) confluent hypergeometric
function of two variables also known as one of the Horn
hypergeometric functions or Humbert's phi1
- added Jeffrey's prior on g
- added the general tCCH prior and special cases of the hyper-g/n.
- TODO check shrinkage functions for all
- new improved Laplace approximation for hypergeometric1F1
- added class basglm for predict
- predict function now handles glm output
- added dataframe option for newdata in predict.bas and predict.basglm
- renamed coefficients in output to be 'mle' in bas.lm to be consistent across
lm and glm versions so that predict methods can handle both
cases. (This may lead to errors in other external code that
expects object$ols or object$coefficients)
- fixed bug with initprobs that did not include an intercept for bas.lm
- added thinning option for MCMC method for bas.lm
- returned posterior expected shrinkage for bas.glm
- added option for initprobs = "marg-eplogp" for using marginal
SLR models to create starting probabilities or order variables
especially for p > n case
- added standalone function for hypergeometric1F1 using Cephes
library and a Laplace aproximation
-Added class "BAS" so that predict and fitted functions (S3
methods) are not masked by functions in the BVS package: to do
modify the rest of the S3 methods.
- added bas.glm for model averaging/section using mixture of g-priors for
GLMs. Currently limited to Logistic Regression
- added Poisson family for glm.fit
- cleaned up MCMC method code
- removed internal print statements in bayesglm.c
- Bug fixes in AMCMC algorithm
- fixed glm-fit.R so that hyperparameter for BIC is numeric
- added new AMCMC algorithm
- bug fix in bayes.glm
- added C routines for fitting glms
- fixed problem with duplicate models if n.models was > 2^(p-1) by
restricting n.models
- save original X as part of object so that fitted.bma gives the
correct fitted values (broken in version 0.80)
- Added `hypergeometric2F1` function that is callable by R
- centered X's in bas.lm so that the intercept has the correct
shrinkage
- changed predict.bma
to center newdata using the mean(X)
- Added new Adaptive MCMC option (method = "AMCMC") (this is not
stable at this point)
-Allowed pruning of model tree to eliminate rejected models
- Added MCMC option to create starting values for BAS (`method = "MCMC+BAS"`)
-Cleaned up all .Call routines so that all objects are duplicated or
allocated within code
- fixed ch2inv that prevented building on Windows in bayes glm_fit
- fixed fortran calls to use F77_NAME macro
- changed allocation of objects for .Call to prevent some objects from being overwritten.
- fixed EB.global function to include prior probabilities on models
- fixed update function
- fixed predict.bma to allow newdata to be a matrix or vector with the
column of ones for the intercept optionally included. - fixed help file for predict - added modelprior argument to bas.lm so that users may now use the beta-binomial prior distribution on model size in addition to the default uniform distribution - added functions uniform(), beta-binomial() and Bernoulli() to create model prior objects - added a vector of user specified initial probabilities as an option for argument initprobs in bas.lm and removed the separate argument user.prob