BayesPIM 2.0
Breaking changes
- Version 2.0 is a substantial rewrite: the user-facing API, the
sampler, and the post-estimation tooling have all changed, and code
written for 1.0.1 will not run unchanged.
- All exported functions and arguments were renamed to snake_case.
bayes.2S() is now bayespim(),
gen.dat() is gen_data(),
get.IC_2S() is get_ic(),
trim.mcmc() is trim_mcmc(), and
search.prop.sd() is search_prop_sd().
Arguments follow the same convention (Vobs ->
v_obs, Z.X -> x_t,
Z.W -> x_g, dist.X ->
dist, tau.w -> tau_g).
bayes.2S_seq() and search.prop.sd_seq()
are removed; the parallel and sequential code paths are unified.get.ppd.2S() is replaced by ppCIF(), which
computes the mixture and the non-prevalent cumulative incidence function
in a single call and has a plot() method.- The
thining, conv.crit,
parallel, vanilla, and
ndraws.naive arguments are gone. Draw storage is controlled
by save_every and the warm-up cutoff by
warmup.
- In
gen_data(), the covariate-correlation argument is
renamed from r to rho, so that r
unambiguously denotes the baseline-test indicator, as it does in
bayespim() and in the returned $r.
- Two defaults changed:
kappa no longer defaults to
0.5 and must be given explicitly when
update_kappa = FALSE, and the effective-sample-size target
min_effss rose from chains * 10 to
chains * 100.
- The generalized gamma is reparameterized to the Prentice form
(location, scale, signed shape
Q) using
flexsurv, replacing the ggamma
parameterisation of 1.0.1. Fitted shape values are not comparable across
versions, and the model is now available only with the collapsed
sampler.
New features
- Two slice samplers.
sampler = "slice_collapsed", the new default, augments only
the latent screening interval and updates the incidence parameters from
the interval-censored likelihood; sampler = "slice"
augments exact event times. Both show lower autocorrelation and faster
convergence than the Metropolis-Hastings sampler of 1.0.1, which remains
available as sampler = "mh".
- Gamma incidence times
(
dist = "gamma"), parameterized through the conditional
mean and coefficient of variation.
summary() and plot()
methods for fitted models, reporting posterior quantiles and
convergence diagnostics for each parameter block.- Revised convergence assessment using
rank-normalized split R-hat and effective sample size from the
posterior package, replacing
coda::gelman.diag().
update_till_converge = TRUE extends sampling automatically
until max_rhat and min_effss are met.
- Reproducibility.
seed_chains sets one
seed per chain and the end-of-chain RNG state is stored, so a run
continued through prev_run is identical to an uninterrupted
run of the same length.
- Covariate standardization
(
standardize_covariates) and internal time rescaling
(rescale_times), both enabled by default, with returned
coefficients on the original scale.
- User-supplied priors through
log_prior_fun; the default is exported as
log_aft_prior().
- Input validation. Malformed inputs produce a
single, specific error before sampling begins.
- Further additions:
save_every to limit memory,
silent to suppress progress output, and fix_q
to hold the generalized-gamma shape fixed.
- Example fit.
data(mod) provides a
converged model so post-estimation examples run without refitting.
- Rewritten vignette
(
vignette("BayesPIM_intro")): a user guide covering
estimation, convergence, model comparison, and posterior CIFs.
- Test suite. The package now ships contract tests
covering the samplers, distributions, draw storage, covariate scaling,
and post-estimation functions.
Other
- Internal code was consolidated and reduced: duplicate
parameterisations, unreachable branches, and redundant state were
removed, and naming is consistently snake_case throughout.
- Dependencies: posterior, survival,
and flexsurv added; ggamma and
mvtnorm removed; coda moved to
Depends.
- Smaller bug fixes and behind-the-scenes efficiency updates and
leaner code.