NEWS

glmmFEL 1.0.6 (2026-09-18)

Corrected probit tail derivatives through fourth order, safeguarded posterior mode and fixed-effect updates, and kept the E-step distribution fixed during the corrected M step. Added independent derivative, tilted-integral, quadrature, and Gaussian GLS regression checks.
Failed inner solves and iteration limits now produce warnings and explicit convergence diagnostics. Returned posterior moments and PL working quantities are refreshed at the final estimates. Invalid responses, designs, and controls receive earlier errors. A strictly separating fixed-effect direction in binary data is rejected, rather than reporting a small score at an effectively infinite estimate.
Added exact scalar computation paths for disjoint random-effect designs and removed the numerical-Jacobian dependency from estimation. Removed unused suggested packages; local benchmark dependencies are documented separately.
Clarified the modal-EM meaning of Laplace, the diagonal scope of FE_full variance corrections, and the limitations of Hessian standard errors. PL fits now return NA from logLik(); the working Gaussian objective remains available as working_logLik. Cross-method AIC comparisons are inappropriate.
Added a static simulation appraisal in help("glmmFEL-benchmarks"), covering both binary links and Poisson responses, four cluster settings, and comparisons with lme4, glmmML, and MASS::glmmPQL. Benchmarks do not run during CRAN checks.

Established a development repository from the supplied 1.0.5 source archive, with maintenance checks and a local literature index. No algorithm or public API changes were made during repository setup.