bvarnet 1.0.4
- Removed
skip_lag immediately, including its zero-filled
gap likelihood, which could bias shared parameters and temporal
estimates. Passing this argument now errors for any value. All model
families use complete-transition analysis: responses with any
unavailable required lag are excluded from the likelihood. Observation
counts and estimates can therefore differ from the former default.
- Added
beep_col. With beep_col = NULL,
time_col is the scheduled integer occasion index within a
subject. With both supplied, time_col labels days or
sessions and beep_col indexes scheduled occasions within
them. Temporal lags never cross subject/day boundaries or missing
scheduled occasions. Subject random effects remain shared across
days.
- Switching to a completely centered backend. The Stan models now
center all fixed effects parameter internally. This can lead to improved
posterior geometry in complex models. This does not change parameter
estimates or Bayes factors!
- Internal centered blocks are excluded from stored public
draws/diagnostics.
- Custom initial values are mapped to fit the transformation.
- Bugfix:
save_data argument now works correctly.
bvarnet 1.0.3
- Bugfix: Bayes factors for interaction effects were sometimes
displayed wrong.
bvarnet 1.0.3
- Minor bugfix that let to an error when running a model without
covariates.
extract_draws() now accepts several parameters at
once.
- Minor documentation fixes
- Bugfix: Removing order constrained from adjacent category
model.
bvarnet 1.0.2
New features
- Introduce toolchain free installation via
bvarnet_setup_models().
Bug fixes affecting results
- Fixed automatic scaling of Gaussian default priors and related
misscalulation of BFs
- Fixed that
extract_random_effects(what = "mean_u") and
what = "draws_u" returned scrambled subject estimates and
RE mislabeling in mixed-family fits.
- Change default CI to 95% and introduce argument to vary CI
width.
Other changes
- Defined the
%||% operator internally.
print() now reports the effective scale of default
Gaussian priors, which are widened by the outcome SD before reaching
Stan. Previously this scaling was invisible.
- Minor bug and documentation fixes.
bvar() gains ..., forwarding additional
arguments (e.g. init, refresh,
thin, step_size) to CmdStanR’s
$sample() method.
- Added a safeguard against duplicated
(id, time) rows,
which previously produced an ambiguous, silently-contaminated lag
design; these now error with guidance to deduplicate or aggregate.
time_col must now be integer-valued (one time unit =
one lag step).
bvarnet 1.0.1
- Add installation safeguard on CRAN.
bvarnet 1.0.0