Release notes

CensoredDistributions.jl follows a hybrid changelog approach:

• GitHub Releases: All releases (automatically generated by TagBot)

• NEWS.md: Major releases requiring detailed context (shown below)

See GitHub Releases for minor/patch release information.

Unreleased

Bug fixes

• double_interval_censored and primary_censored now select the analytic-vs-numeric solver by dispatching on the argument types rather than branching on a Bool value, so the return type stays concrete when the delay parameters are runtime values (e.g. inside a Turing model). Previously the value-level force_numeric branch inferred to a Union of the AnalyticalSolver and NumericSolver specialisations at non-constant-folded call sites, which propagated an abstract element type into downstream pdf loops and poisoned reverse-mode AD. Closes #367.

• Skip the CDF-saturation early-return in the numeric primarycensored_cdf path when the lower bound is at the distribution boundary (lower == minimum(dist)). Evaluating cdf(dist, lower) there is unnecessary (cdf is 0 by construction) and trips degenerate 0·(-Inf) reverse-mode rules in Distributions.jl (e.g. LogNormal via log(0)), contaminating the ReverseDiff tape with NaN. Restores ReverseDiff and Mooncake gradient correctness on PrimaryCensored LogNormal+Uniform numerical. Closes #249.

Added

• primary_censored and double_interval_censored accept a method keyword taking a solver object, AnalyticalSolver() (the default) or NumericSolver(), each optionally given a quadrature solver (e.g. NumericSolver(QuadGKJL())). AnalyticalSolver and NumericSolver are now exported. Passing a concrete method object is the type-stable way to choose the CDF backend, and is preferred over the deprecated force_numeric flag.

Deprecated

• The force_numeric keyword of primary_censored and double_interval_censored is deprecated. Pass method = NumericSolver() (or method = AnalyticalSolver()) instead. force_numeric still works but emits a deprecation warning and, being a runtime Bool, does not guarantee a concrete return type.

Breaking

• primary_censored(...; solver) defaults to GaussLegendre(; n = 64) (was QuadGKJL()). The fixed-node solver traces cleanly through every AD backend, where adaptive quadrature does not. Pass solver = QuadGKJL() explicitly to keep adaptive accuracy.

AD gradient infrastructure

• New test/ad/ sub-environment with DifferentiationInterfaceTest-driven gradient correctness tests for the logpdf of every censored distribution type, plus unit-level tests for the gamma CDF rrule (_grad_p_a_series baseline, Mooncake.TestUtils.test_rule, and defensive guards). Wired into task test-ad, task test, and a dedicated CI job.

• New test/ADFixtures path package is the single source of truth for the scenario set and the working/broken backend matrix. The test suite, benchmark suite (benchmark/src/ad_gradients.jl), and docs tutorial all add it as a path dep so the three AD surfaces stay in lock-step.

• Gradient references are computed with ForwardDiff. Adaptive finite-difference baselines (e.g. central_fdm(5, 1)) disagree with every AD backend by ~10% on Weibull analytical scenarios because PrimaryCensored.logpdf internally finite-differences the CDF with a hardcoded h = 1e-8; ForwardDiff's Dual propagation through that same FD gives the exact derivative the package computes and matches the other backends to ~1e-6.

• New "Automatic differentiation backends" tutorial at docs/src/getting-started/tutorials/ad-backends.jl renders the full backend × scenario matrix via DIT.benchmark_differentiation with failures left visible.

• README gains a per-backend support badge row reflecting the measured state (ForwardDiff full; ReverseDiff tape, Mooncake reverse, and Mooncake forward partial; Enzyme forward/reverse broken).

• Known follow-ups tracked in #217 (gamma_inc Dual dispatch gap on the Distributions.cdf(Gamma) path used by IntervalCensored Gamma arbitrary), #225 (Enzyme + DIT-Mooncake interaction).

v0.1.0 - Initial Release

CensoredDistributions.jl extends Distributions.jl to support primary event censoring and interval censoring, enabling modelling of scenarios where observation times are subject to various forms of censoring and truncation.

Core functionality

Distribution Constructors

• primary_censored(dist, primary_event): Creates primary event censored distributions where an initiating event occurs within a time window, then experiences a delay

• interval_censored(dist, interval): Creates interval censored distributions for continuous values observed only within discrete intervals (regular or arbitrary intervals)

• double_interval_censored(dist; kwargs...): Combines primary event censoring, optional truncation, and optional secondary interval censoring in an appropriate mathematical sequence

Distribution Types

• PrimaryCensored{D, P}: Wraps delay distribution D with primary event distribution P

• IntervalCensored{D, T}: Wraps continuous distribution D with interval boundaries of type T

• DoubleIntervalCensored: Composition of multiple censoring mechanisms

Extensible CDF Method

• primarycensored_cdf(): User-extensible method for computing CDFs of primary censored distributions with analytical solutions for common distribution pairs

Utility Functions

• weight(dist, weights): Creates weighted distributions for efficient likelihood computation

Distributions.jl interface

Partial implementation of Distributions.jl interface including:

• pdf, logpdf: Probability density functions

• cdf, logcdf: Cumulative distribution functions

• quantile: Inverse CDF (where analytically tractable)

• rand: Random number generation

• minimum, maximum, insupport: Support queries

• mean, var, std: Moments (where analytically tractable)

Migration from primarycensored R package

CensoredDistributions.jl provides a Julia-native implementation with enhanced functionality compared to the primarycensored R package:

Function Mapping

• R's dprimarycensored() → Julia's pdf(primary_censored(...))

• R's pprimarycensored() → Julia's cdf(primary_censored(...))

• R's qprimarycensored() → Julia's quantile(primary_censored(...))

• R's rprimarycensored() → Julia's rand(primary_censored(...))

Parameter Differences

• Primary event window: Use primary_event = Uniform(0, pwindow) instead of R's pwindow parameter

• Solver selection: Configure via solver keyword instead of R's integration method parameters

• Truncation: Apply via truncated() from Distributions.jl or use double_interval_censored() convenience function

Contributors

• Sam Abbott (@seabbs)

• Damon Bayer (@damonbayer)

• Sam Brand (@sambrand)

• Michael DeWitt (@dewittpe)

• Joseph Lemaitre