`NEWS.md`

- Add more imports to NAMESPACE per a
**rstan**developer’s recommendation.

- New vignette about mathematical description of models.
- Prediction and function posterior computation at
`P`

points where`P`

is larger than number of data points should be now much faster and take less memory, as`P`

x`P`

matrices are not computed.

- Adds
`prior_pred()`

for prior predictive sampling and`sample_param_prior()`

for sampling from the parameter prior. - Adds
`read_proteomics_data()`

function. - Relax data type checking, to require that they only inherit from factor or numeric. Allow also
`tibble`

s and`data.table`

s to be passed as data. - Adds more methods for
`lgpfit`

and`lgpmodel`

objects, see their documentation. - Lot of improvements internally. Kernel computations in functions like
`pred()`

should take a lot less memory now. Two separete main Stan models now. One for latent GP (signal where f is sampled) and other for GP with marginalized f. - Improved documentation.
- Improve verbose messages to user.

- Add more informative error message if trying to specify a model like
`y ~ age + id | age`

, which should be`y ~ age + age | id`

, i.e. the continuous covariate on the left of`|`

and categorical on the right. - New startup message that prints also
`rstan`

version - Update citation information

- Add the
`c_hat_pred`

argument to`pred()`

, to be used when`f`

has been sampled and`c_hat`

is not constant. Previously,`c_hat = 0`

was used in all prediction points, which did not make sense in all cases.

- Allow setting
`group_by = NA`

in`plot_pred()`

,`plot_components()`

and`new_x()`

to avoid grouping in plots. - Allow setting
`color_by`

as the same factor as`group_by`

. - Fix bug which caused an error when trying to define a separate prior for parameters of the same type.

- Add option
`do_yrng`

which controls whether to do draws from the predictive distribution. This was previously always done if`sample_f`

was`TRUE`

. That is now considered a bug because it is unnecessary work if the`y_rng`

draws are not needed. So the default is now`do_yrng = FALSE`

, since`do_yrng = TRUE`

can cause errors with the negative binomial model due to numerical problems (see here). These problems should be addressed in a future release to allow more stable prior and posterior predictive sampling.

- Fix bug in
`get_pred()`

, which was caused by not adding the GP mean to the sampled signal. This was causing postprocessing functions like`relevances()`

and`plot_pred()`

to give erroneous results if the GP mean was not a vector of zeros and`sample_f = TRUE`

. - Small edits in documentation and verbose information messages.

- Make
`plot_pred()`

work with any response variable name (fixes issue #12). - Avoid adding
`ggplot2::color_scale_manual()`

if number of colors > 5 (fixes issue #11).

- Formula syntax where
`|`

indicates interaction terms. - Alternative advanced formula syntax with
`gp()`

,`gp_warp()`

,`zerosum()`

etc. - Beta binomial observation model.
- Categorical covariates must now be specified as factors in data, and don’t have to be numeric.
- Component relevance assessment is now separated from model fitting into the
`relevances()`

function and selection into`select()`

. - Easier prior specification with
`normal()`

,`log_normal()`

,`student_t()`

etc. - Better prediction and plotting functionality with
`get_pred()`

,`pred()`

,`plot_pred()`

, and`plot_f()`

. - Extensive argument checking (see
`check_positive_all()`

etc.) to give users informative error messages

- Thorough unit tests using test_that.
- C++ versions of the Stan model functions are now exposed to package namespace and also tested.