NEWS

flexhaz 0.5.3 (2026-10-28)

Likelihood, score, Hessian, and fitting methods now reject missing or invalid censoring codes and non-numeric, negative, or non-finite observation times. Previously, unrecognized or missing censoring codes silently dropped rows from the log-likelihood. A missing censoring column still denotes exact observations; logical event indicators and zero observation times are supported.

Prototype constructors (dfr_exponential(), dfr_weibull(), dfr_gompertz(), dfr_loglogistic()) now prepend a family-specific S3 subclass (e.g. "dfr_exponential") to the class chain. This enables downstream packages to dispatch on component type via inherits() rather than maintaining parallel lookup tables. Existing inherits(x, "dfr_dist") checks are unaffected (additive change).

Added helper distribution constructors: dfr_exponential(), dfr_weibull(), dfr_gompertz(), dfr_loglogistic() with analytical hazard, cumulative hazard, score, and Hessian functions where available.
Added diagnostic methods: residuals() (Cox-Snell and Martingale), plot() (survival, hazard, cumulative hazard), qqplot_residuals().
Added density() method (alias for pdf).
Added assumptions() method for listing model assumptions.
Added kaplan_meier() internal utility for empirical survival estimation.
Added support for left-censored observations (delta = -1).
Improved numerical stability in log-likelihood computation.
Removed femtograd dependency — users supply their own derivative functions via score_fn and hess_fn, or the package falls back to numDeriv.

Initial release.

dfr_dist() constructor for creating distributions from hazard functions
Complete distribution interface: hazard(), cum_haz(), surv(), cdf(), pdf(), inv_cdf(), sampler()
Likelihood model interface: loglik(), score(), hess_loglik(), fit()
Support for right-censored survival data
Numerical integration for cumulative hazard
Inverse CDF sampling

Supports modeling complex hazard patterns: