distributions3 vs Google Analytics

An R package giving probability distributions a consistent object interface - d/p/q/r functions, moments, and prodist() methods that pull a fitted distribution out of a regression object. Version 0.3.0 is the first substantive release under Achim Zeileis's maintenance, and it widens what a distribution is allowed to be: Empirical() represents a distribution by a random sample rather than by parameters, and numerical fallbacks now fill in cdf(), pdf(), quantile(), random() and the moments for any object that implements only some of them. New score() and hessian() generics compute first and second derivatives of the log-likelihood with respect to the parameters, analytically for a few distributions and numerically for the rest.

Growth used to arrive as new distribution families contributed from outside - the extreme-value set, Erlang, later the Poisson binomial. This release changes the axis: alongside two new distributions it adds an inference layer (score, hessian) and a forecast-evaluation one (crps() methods against scoringRules), which are capabilities about distributions rather than more of them. Dependency weight is being cut at the same time, with ggplot2 demoted to Suggests and glue replaced by base R sprintf().

With numeric fallbacks and the derivative generics in place, expect analytic score() and hessian() methods to be filled in across more of the distribution catalogue. The constructor-default change is the likeliest source of follow-up fixes, since calls like Poisson() now return a length-zero distribution where they previously errored.