
version 0.7.9
- predict now also works for models with bym2 components
- improvements to conjugate gradients and constrained multivariate normal
  sampling methods, which now also work in more cases (e.g. in case of
  inferred AR1 parameter, local scale parameters or bym2 components)
- new argument add.eps.I of sampler_control to add a small multiple of the
  identity matrix to the coefficients' posterior precision matrix to avoid
  singularities. Both this option and the constrained multivariate normal
  (cMVN) sampler allow faster sampling for large structured multilevel models
  with many identifiability or other linear equality constraints.
- linear equality and inequality constraints should now be specified using
  function set_constraints, which returns an environment including functions
  to check whether a numeric vector obeys all constraints
- now imports R package collapse instead of matrixStats
- new argument weights of predict method for mcdraws object, which can be used
  to pass population sizes of poststratification cells for the P-step of MRP
  (Multilevel Regression and Poststratification)
- family-specific arguments should now be passed to thw family argument of
  create_sampler of generate_data by means of the family functions f_gaussian,
  f_binomial, f_negbinomial, f_poisson, f_multinomial, f_gamma
- fixed a bug in prediction with type="response" for multinomial family (thanks
  to Sumonkanti Das for reporting)
- fixed a bug where generating data for a model with only random iid effects at
  the data level failed
- spline factor component renamed to splines to avoid naming conflict with
  stats::spline
- further (small) efficiency improvements
- more tests, documentation updates and further small bug fixes


version 0.7.8
- added bym2 model component
- in spatial random effect factor, argument poly.df has been replaced by graph
  which now also accepts neighbours lists, and argument derive.constraints has
  been deprecated as it is no longer needed
- AR1 random effect factors used to be limited to fixed autoregressive
  parameter values, whereas now these parameters can be inferred by
  assigning them uniform or truncated normal priors
- prediction for models with offsets did not always work correctly, especially
  for Poisson family
- fixed a bug in blocked Gibbs sampler for models including a measurement model
  component
- more control over Metropolis-Hastings proposal distributions for some
  parameters through new control arguments
- small documentation updates
- some code restructuring to facilitate future extensions
- more tests


version 0.7.7
- fixed a bug introduced in mcmcsae 0.7.5 that affected the outcomes of
  the blocked Gibbs sampler for non-gaussian models with random effects
  defined at the data level
- user-defined equality constraints on regression coefficients or random
  effects now work again for blocked Gibbs sampler
- default prior for shape of gamma family has been changed from gamma(1, 1)
  to gamma(0.1, 0.1)
- more sensible default start values for gamma shape parameter
- renamed model component name bart to brt, to avoid name clash with main
  fitting function of package dbarts
- prediction now also works for models with a BART model component by
  specifying keepTrees=TRUE in brt()
- corrected and updated the still somewhat experimental SBC_test function for
  simulation-based calibration; it now also supports parallel computation


version 0.7.6
- blocked Gibbs sampler now also works with nonzero prior means of regression
  effects
- fully-blocked Gibbs sampler is now the default, and argument block has been
  moved to sampler_control
- model components for fixed and random effects now also usable to model
  log-variance of gaussian multilevel models
- added support for random effects for log mean of gamma sampling distribution
- shape parameter now given default gamma(1, 1) prior in gamma multilevel
  models


version 0.7.5
- added gamma regression (family = "gamma")
- several improvements to the "softTMVN" truncated multivariate normal sampler
- added a few more methods to class tabMatrix to prepare for Matrix 1.6.2
  (thanks to Mikael Jagan)
- removed a few obsolete arguments from exported functions
- more consistent prior specification; normal priors can now be specified using
  function pr_normal; arguments b0, Q0 for prior mean and precision in several
  model components have been deprecated
- function pr_fixed for specifying a degenerate prior can now be used in more
  places
- global option setting function set_opts has been replaced by several
  control functions sampler_control and chol_control that can be used
  to pass computational options to various functions


version 0.7.4
- replaced maptools in Suggests by sf for reading shape files; now both
  SpatialPolygonsDataframe (for backward compatibility) and sf spatial data
  frames are supported
- updated documentation of spatial() (in help topic 'correlation') and added
  an example of a CAR spatial random effects model
- fixed a bug so that conjugate gradients algorithm works again
- added control functions to set computational options for create_sampler
  and setup_CG_sampler
- updated a few unit tests to be compatible with upcoming Matrix 1.6.0
- small documentation and code improvements


version 0.7.3
- improved handling of out-of-sample categories by predict method
- further improvements to prepare for upcoming version of Matrix package
  (thanks to Mikael Jagan)
- clean-up of create_TMVN_sampler, in which now the method for truncated
  multivariate normal sampling can be specified by means of a method function
  that allows to pass method-specific options
- added HMC ZigZag TMVN sampler
- fixed a bug in soft-TMVN sampler, which did not work in case of a sparse
  equalities constraint matrix
- option to add a Bayesian Additive Regression Trees model component to the
  linear predictor through package dbarts


version 0.7.2
- prediction for new data now handles out-of-sample random effects (at least
  for iid random effect terms), so that it becomes easier to account for
  cluster effects from cluster samples, say
- several other small improvements to predict method
- small fix in preparation for upcoming Matrix 1.5-4 (thanks to Mikael Jagan)
- model_matrix: allow single-level factor/character variables if no contrasts
  are applied
- bug fix: inequality constraints did not work in combination with blocked
  Gibbs sampler
- some parts of truncated multivariate normal samplers have been converted
  to C++ (using Rcpp and RcppEigen) for better performance
- argument sampler of computeDesignMatrix has been removed
- to_draws_array can now also convert an mcdraws object (or a subset of
  components from it) to a draws_array object for further analysis using
  R package posterior


version 0.7.1
- compute_WAIC can now run using multiple cores
- predict method with option ppcheck=TRUE now also works in parallel
- prepare for coercion deprecations in upcoming version of Matrix package


version 0.7.0
- renamed class 'draws' to 'mcdraws' to avoid name clash with R package
  posterior
- added function to_draws_array to convert a draws component to an object of
  class draws_array, as defined in R package posterior
- support for multinomial family
- support for Poisson family, approximately, in terms of negative binomial
- it is now possible to use weights to specify irregularly spaced AR1 or RW1
  correlation structures
- initial support for conjugate gradient coefficient sampler
- experimental function for simulation-based calibration


version 0.6.0
- measurement in covariates model component mec() added
- new function pr_gig to specify a Generalized Inverse Gaussian prior
- new argument logJacobian for create_sampler to allow comparisons of
  information criteria between model fits based on different transformations
- added function to set labels of draws component object
- data is now second argument of create_sampler and generate_data functions,
  in line with many model fitting functions in R
- generate_data gains argument linpred, which is convenient for generating
  both data and latent quantities of interest for area-level models
- solved a bug in function split_iters
- print.dc_summary now correctly handles max.lines argument
- adapted to new version of Matrix package
- more input checks and small code improvements


version 0.5.0
- initial CRAN release
