rbmi vs silx

rbmi implements reference-based multiple imputation for longitudinal clinical trial data with missing values — the estimand machinery regulators expect for handling intercurrent events and dropout. The consequential recent change was 1.3.0 moving rstan from a hard dependency to Suggests, which takes Bayesian imputation out of the default install. Since then the work has been documentation and nomenclature discipline: 1.6.1 standardized on MNAR over a mixed NMAR/MNAR vocabulary and deprecated the nmar.rm argument accordingly.

The package is optimizing for adoption friction over feature breadth. Dropping a compiled Stan dependency from the default install, deprecating a bespoke seed argument in favor of base set.seed(), and aligning lsmeans() behavior and weight naming with emmeans all point the same direction — behave like a conventional R package rather than a specialized one. Documentation work in 1.6.1 covering @return on every exported function and executable examples reads as preparation for validation scrutiny rather than user demand.

Given the FAQ vignette's validation statement and the recent documentation completeness pass, the next work is more likely qualification and estimand documentation than new imputation methods.