Objective Bayesian Distribution Fitting

Fits common univariate distributions using registered objective Bayesian priors, including Jeffreys, reference, and maximal data information priors, and supports user-defined distributions and priors through an extensible model specification. Model-specific posterior propriety and moment conditions are checked before computation when registered or supplied. Exact simulation, marginalization, slice sampling, adaptive Metropolis, and user-supplied posterior samplers share a common interface for summaries, diagnostics, prediction, and pointwise log-likelihood evaluation. A separate interface fits independently right-censored observations using the registered complete-data priors, observed-data likelihood sampling or data augmentation, with sufficient posterior-propriety checks. Optional post-processing provides WAIC, PSIS-LOO, and DIC for observed-data likelihoods. The reference-prior framework follows Bernardo (1979) .

fitdistrBayes

fitdistrBayes 0.6.0 fits common univariate distributions under registered objective Bayesian priors through a deliberately small interface:

fitdistrBayes(x, distr, prior)

For a built-in route, the package checks the sampling support and the known posterior-propriety condition before computation. It also records whether the posterior mean and variance of each parameter are mathematically certified to exist, so a finite chain is not used to justify a divergent moment.

Installation

Install the source archive supplied with this research release and load it:

install.packages("fitdistrBayes_0.6.0.tar.gz", repos = NULL,
                 type = "source")
library(fitdistrBayes)

Dhillon (new in 0.6.0)

The complete-data and censored interfaces accept "dhillon", with survival 1/(1 + theta*t^beta) and positive parameters beta, theta. The aliases "log-logistic" and "loglogistic" use exactly this parameterization. Jeffreys and "reference-beta" coincide; "reference-theta" is a distinct ordered prior. "reference" means "reference-beta". MDI is rejected.

set.seed(19)
x <- exp((rlogis(100) - log(2)) / 3)
fit <- fitdistrBayes(x, "dhillon", "reference-beta", seed = 20)
observed <- pmin(x, 1)
status <- as.integer(x <= 1)
fit_cens <- fitcensBayes(observed, status, "dhillon", "reference-theta", seed = 21)
fit$moment_status
fit$predictive_moments
help("Dhillon")

Both observed-likelihood and augmentation methods are available. Exact observed-data propriety and parameter-moment checks run before sampling. Predictive lifetime means are infinite: use quantiles and survival probabilities. The console tutorial is inst/examples/tutorial_dhillon.R; the censored-data derivation is in inst/doc/dhillon-theory.txt (installed as doc/).

Optional model comparison

Both fitting functions default to criteria = FALSE: no additional likelihood matrix or information criterion is computed. Existing R-hat/ESS diagnostics are independent of this option. Use criteria = TRUE to compute all criteria after sampling, or choose criteria = c("waic", "dic") without extra dependencies. PSIS-LOO requires the optional loo package. The fitting chains are unchanged.

set.seed(20)
x <- rexp(60)
fit <- fitdistrBayes(x, "exponential", "reference", seed = 21,
                    criteria = c("waic", "dic"))
fit$criteria
WAIC(fit)                       # May also be computed later
# install.packages("loo")       # Once, if PSIS-LOO is desired
# LOOIC(fit)

The same functions accept fitcensBayes fits and use log survival for censored observations, never imputed-data densities. DIC is unavailable if necessary posterior means are not certified finite. LOO is withheld if a training posterior is not certified proper. Inspect diagnostic warnings before model comparison. compare_models() requires the same observations, censoring status, and scale. See help("criteria") for definitions, output components, and limitations.

Independent right censoring

The separate function fitcensBayes(x, status, distr, prior) uses status = 1 for exact observations and status = 0 for the strict event T > x. For discrete distributions this distinction is important: a record meaning T >= 3 must be supplied as x = 2, status = 0.

set.seed(20)
lifetime <- rexp(80, rate = 0.7)
censoring <- rexp(80, rate = 0.3)
x <- pmin(lifetime, censoring)
status <- as.integer(lifetime <= censoring)
fit <- fitcensBayes(x, status, "exponential", "reference", seed = 21)
summary(fit)
confint(fit)
predict(fit, type = "survival", times = c(1, 2, 3), draws = 5)
log_lik(fit, draws = 5)

The catalogue fitcensBayes_models() lists 59 registered model-prior routes for 21 distributions. The priors are inherited from the complete-data model; they are not claimed to be Jeffreys or reference priors rederived for a particular censoring design. Posterior propriety is certified by sufficient conditions on the exact-event subset, with an additional analytic Exponential case and an exact observed-data condition for Dhillon. An uncertified case is rejected before sampling. Imputed observations are never used to satisfy these conditions.

method = "auto" uses the observed likelihood when censoring is present; method = "augmentation" alternates truncated lifetime simulation with parameter updates. Without censoring, method = "auto" uses the original complete-data sampler. Posterior medians, credible intervals, rank-based diagnostics and justified moment summaries are returned. Heavy censoring may require substantially longer chains. A diagnostic pass is not a convergence proof. Densities or priors defined by the user remain available through the original fitdistrBayes() interface, but not the new censored interface.

See help("fitcensBayes") and the installed English tutorial:

system.file("examples", "tutorial_fitcensBayes.R", package = "fitdistrBayes")

Quick start

An exact-posterior example:

set.seed(10)
x <- rexp(40, rate = 2)
fit_exp <- fitdistrBayes(x, "exponential", "jeffreys", seed = 11)
fit_exp
confint(fit_exp)

An MCMC example with a parameter-specific reference prior:

set.seed(20)
x <- rgamma(50, shape = 2.5, rate = 1.3)
fit_gamma <- fitdistrBayes(
  x, "gamma", "reference-shape",
  iter = 6000, warmup = 1000, chains = 4, seed = 21
)
summary(fit_gamma)
plot(fit_gamma, type = "trace")
plot(fit_gamma, type = "acf")

fit$summary contains posterior medians, equal-tail intervals, R-hat, bulk and tail ESS, and MCSE when the relevant moments exist. Acceptance rates labelled overall use post-warmup iterations; warmup and all-iteration rates are stored separately.

Posterior prediction and pointwise log likelihood are computed on demand:

yrep <- predict(fit_gamma, draws = 500, size = length(x), seed = 22)
ll <- log_lik(fit_gamma, draws = 500, seed = 23)
dim(yrep)
dim(ll)

Supported routes

The package has 59 enabled model--prior combinations for 21 distributions. For Student-t with unknown degrees of freedom, the independence Jeffreys prior requires at least two pairwise distinct observations; samples with ties are rejected because the posterior is improper. Fixed-df Student-t models use their separate, multiplicity-dependent conditions. The public machine-readable catalogue is the concise way to inspect them:

fitdistrBayes_routes()
fitdistrBayes_routes("gamma")
fitdistrBayes_routes("t")
fitdistrBayes_routes("weighted lindley")

The catalogue reports the estimated parameters, any required fixed parameter, the computational engine, and a concise propriety condition. The fitting function performs the authoritative sample-dependent check. In particular, the negative-binomial routes estimate mu with known positive size:

x <- rnbinom(40, size = 5, mu = 3)
fit_nb <- fitdistrBayes(
  x, "negative binomial", "reference", fixed = list(size = 5), seed = 30
)

Automatic initialization

When start = NULL and numerical initialization is required, the package uses classical estimators: ordinary moments where they exist, quantile matching for Cauchy and heavy-tailed Student-t cases, and closed-form L-moments for Weibull, Frechet, and Lomax. The Exponential-Logarithmic model uses a stable scalar moment equation. Exact independent posterior simulation does not require a starting point. The selected rule and center are stored in fit$initialization.

For weighted Lindley data, the package uses the model's closed-form likelihood estimator (with a numerical maximum-likelihood fallback) and samples in the exactly Fisher-orthogonal mean/shape parameterization. The public output remains on the original lambda, phi scale. The one-group reference prior equals Fisher-information Jeffreys; reference-lambda and reference-phi select the two exact ordered reference priors.

For the implemented Frechet parameterization, F(x) = exp(-scale * x^(-shape)); hence the argument named scale is the positive coefficient in the exponent, while the conventional quantile scale is scale^(1 / shape).

Custom densities

A custom density requires named starting values and a prior function. Log mode is inferred only from an explicit log argument; the presence of ... alone does not imply that log = TRUE is honored. The contract can be set explicitly with control$density_is_log and control$prior_is_log. Use control$prior_style = "scalar" or "vector" when automatic prior calling is ambiguous. Bounds are supplied through control$lower and control$upper; the package constructs the componentwise transform and Jacobian. A predictive RNG and its optional support validator can be supplied through control$rng and control$rng_validator.

Posterior propriety and posterior-moment existence remain the user's responsibility for every custom target.

Scope and limitations

The package handles complete iid univariate samples and independent, noninformative right censoring through fitcensBayes(). Optional WAIC, DIC, and PSIS-LOO calculations use the observed-data likelihood. It refuses routes known to be improper, including the Lomax independent Jeffreys/reference posterior and the Nakagami-m MDI posterior, and it does not substitute an approximation when full-model propriety has not been established. Numerical overflow or underflow stops with an explicit error; values are not silently clipped into the floating-point range.

Teaching example, citation, and support

A console-oriented walkthrough is installed at inst/examples/teaching.R. For the complete mathematical catalogue and computational validation, cite the accompanying manuscript, fitdistrBayes: Objective Bayesian Distribution Fitting in R. Until a public issue tracker is announced, reproducible problem reports can be sent to the maintainer address in DESCRIPTION.

This research version is licensed under GPL-3.