lme4

• simulate() with re.form=NULL now works when a predictor is a one-column matrix (e.g. from scale()) used inside poly(); poly() with pre-specified coefficients (stored in predvars) rejected matrix input even though it accepted it during initial model fitting (GH #948)
• allFit() could fail or emit spurious warnings when refitting a model whose original control argument specified a non-default optimizer or optCtrl (GH #985)
• fixed a heap-use-after-free bug in merPredD construction. This caused intermittent (GC-timing-dependent) "Downdated VtV is not positive definite" errors and other spurious failures in glmer() fits, most visible under memory pressure (e.g. during R CMD check) or under AddressSanitizer.
• fixed a second, unrelated heap-use-after-free with the same symptoms, triggered by an interaction between R's ALTREP-wrapped vectors and Rcpp's data-pointer accessors (exposed by the Rcpp 1.1.2 upgrade; invaluable help from Ivan Krylov).

NEW FEATURES

• a new vignette (vignette("glmer", package = "lme4")) describes the details of glmer implementation
• parallelization via the future framework is available by setting argument parallel to "future".

USER-VISIBLE CHANGES

• getData() now gives more informative error messages when the original data cannot be retrieved: if the object is not found, the message names the object and suggests re-loading it; if an object of the wrong type is found (e.g. the name data resolving to utils::data), the message reports the type and location of what was found and notes the likely masking (GH #808)
• the Gauss-Hermite quadrature (GH) nodes and weights used when nAGQ > 1 are now symmetrized (nodes averaged with their negatives in reverse order; weights averaged in forward and reverse order), matching the convention in MixedModels.jl (GH #968). This will produce small changes to numerical results for any glmer() fit with nAGQ > 1.
• exported covariance methods documented in help("Covariance-class") (getPar, getTheta, etc.)
• the dependency on rlang has been removed; the once-per-session deprecation warnings for formula-processing functions that moved to reformulas are now implemented internally.
• performing *de novo* simulation (i.e., calling simulate() with a one-sided formula rather than a fitted model or a two-sided formula) without specifying family or newparams now gives a more informative error. This only affects the one-sided-formula dispatch path that is internal to lme4; the two-sided simulate.formula/simulate.formula_lhs dispatch mechanism shared with ergm (see GH #566) is unchanged, so ergm's behavior should not be affected.
• *de novo* simulation (see previous item) for a family with an estimated scale parameter (e.g. Gaussian, Gamma) without specifying sigma as part of newparams now gives an error (it previously gave a bogus result based on the residual dispersion from a merMod object with default beta and theta values)