'***' denotes important announcements
'**' denotes big changes visible to the user
'*' denotes minor changes visible to the user
'.' denotes minor bug fixes
Deprecated and liable to be removed in a future version:
* conv() has been renamed to converged() (as of 2020-02-05)
* 'Old-style' passing of buildmer control parameters in the build...(...)
function call is deprecated (as of 2020-06-05); please use the new-style
'buildmerControl' parameter. That is: change
buildmer(y ~ x,crit='AIC',direction='backward')
to
buildmer(y ~ x,buildmerControl=list(crit='AIC',direction='backward'))
1.7
*** Buildmer control parameters (e.g. crit, direction, cl, ...) are now passed
via a new argument buildmerControl=list(...). The 'old-style' passing of
these arguments as part of the buildX() call itself is still supported,
but is now deprecated. If a control parameter is passed both 'old-style'
and in buildmerControl, the former takes precedence. All parameters are
now documented in a single place: ?buildmerControl.
* Implement gradient and Hessian tolerances in converged(). These can now also
be passed as arguments to buildmer functions.
* New elimination criterion for gam/bam models only: 'F'. This uses the change
in R-squared, and has the advantage of being a valid formal test even for
models that were fitted using PQL.
* If singular.ok=TRUE, converged() will now also test singular fits for
convergence, rather than discard them outright.
* The optional 'quiet' argument has been added back.
* The REML+no-random-effects path now uses mgcv::gam rather than nlme::gls,
because the latter expects a non-compatible form for the 'weights' argument.
. Fixes to environment handling support, in particular concerning
tabulate.formula(), which now preserves the formula's environment.
. glmmTMB models are now also checked for gradient convergence, not just for
positive-definiteness of the Hessian.
. A workaround has been implemented to enable fitting glmmTMB models with
offset terms.
. Offset terms are now understood properly by tabulate.formula().
. Changed the wording of the give-up message in order().
1.6 (2020-05-27)
*** Function conv() has been renamed to converged(). Of course, an alias
remains, but it is now deprecated and will be removed in a future release.
** clm and clmm support
* Argument 'cl' now accepts an integer argument, in which case buildmer will
create and manage a cluster for you.
* REML can now be forced on by passing REML=TRUE. As an optimization, REML is
forced on if the deviance-explained criterion is used. This can be disabled
by passing REML=NA, which will fall back to the default behavior of
differentiating between ML and REML. The ability to force REML off by
passing REML=FALSE remains. This 'hidden' argument is now properly
documented, as is the other hidden argument, 'dep'. For buildcustom(), where
this option was already used explicitly with a different meaning, the old
behavior remains as the new behavior doesn't make much sense if the user is
specifying their own fitting function anyway.
* Remove deprecated 'reduce.fixed' and 'reduce.random' arguments (deprecated
since 2019-11-27).
* Print the failure details when convergence failures cause order() or
backward() to exit early.
. Fix a bug combining backward() with the 'include' argument (GH issue #3)
. Fix a bug in calculating Wald p-values for ANOVA F scores affecting ndf >1
. Allow ddf='KR' as an abbreviation for ddf='Kenward-Roger'
1.5 (2020-03-01)
*** API change to the 'crit' function, which now has a new function signature
(p,cur,alt)
*** Deprecate buildjulia due to the maintenance cost of duplicating all the
important likelihood functions. You can code these yourself using
buildcustom().
** Likelihood-ratio tests for random effects are now modeled as a mixture of
chi squares (Stram & Lee 1994), rather than by dividing the p-value by 2.
This makes a difference only for model terms with >1 change in df.
** PQL is okay with Gaussian errors (Breslow & Clayton 1993); this is now
known to buildbam(), buildgam(), and buildgamm(). These now default to LRT,
and give an error message with advice if errors are not Gaussian.
* Relax convergence tolerances for negative eigenvalues to -0.002 (mgcv and
glmmTMB models only)
* Relax convergence tolerance for gradients to 0.04 (mgcv models only)
* Convergence messages are now more informative due to an added human-readable
'reason' attribute
. re2mgcv no longer uses 'by', to possibly work around segmentation faults
. Progress messages are now word-wrapped properly
. Fix incorrect filtering of '...' for (g)lmtree models
1.4 (2019-12-03)
*** The 'reduce.fixed' and 'reduce.random' arguments have been deprecated; use
the 'include' argument instead
** Add new criterion 'devexp' (alias: 'deviance') based on the explained
deviance. Modify buildbam() so that, in the generalized case, it only
accepts this criterion. bam() and gam() with non-outer optimization use PQL,
which was incorrectly permitted by the previous version of the PQL guard
** S3 and S4 method passthrough support
** gamm support
** buildgam() has gained an experimental 'quickstart' argument that prefits
each gam model using bam() to obtain starting values. Specifically for the
'scat' family, the optimized theta values indicating the degrees of freedom
and the scale parameter will also be passed on to the gam() call, but only
if the mgcv version is at least 1.8-32
** Added new function 're2mgcv()' that makes it possible to use buildgam()
with lme4-style random effects (with correlations removed)
* Smooth terms are now no longer forced to be evaluated after parametric terms
* For general families that are fitted with REML only, buildgam() now makes
the appropriate modifications to the gam() call and include list
* buildcustom() has gained the ability to use buildmer's ML/REML
differentiation facilities; turn this on by passing REML=TRUE to the call to
buildcustom()
* Split up buildmer() and buildgamm4()
. Filter ... argument more precisely in the various buildmer() fitting paths
for non-mixed models (lm/glm/gls)
. Fix bugs in remove.terms() and in order() affecting certain models containing
smooths
. buildgam() is now properly able to fit intercept-only models
. Convergence for glmmTMB models is now checked correctly if there are/seem to
be (as with REML=TRUE) no fixed effects
. Work around glmmTMB issue with REML for poisson and binomial models
. build.formula() now also handles formulas where fixed and random effects are
not strictly in that order (which fixes 'include' for random effects)
. Implement workaround in buildgls() to work with rank-deficiency
. Wrap examples in requireNamespace()
. Explicitly fit (fixed/random)-intercept models as well, which was skipped in
earlier versions. Add dependency on nlme, as gls() is used when
transitioning from fixed to random effects in lme4 models.
. Fix diag() and add.terms() for one-sided formulas
1.3 (2019-09-28)
** GLMMadaptive support
* lme models (package nlme) now have full support for the random part as well
* glmertree models now have full support for stepwise elimination
* calc.anova is now FALSE by default. Most users will not use it anyway, and
for some lmerTest-based models it can cause an error.
* It is now (experimentally) possible to pass REML=FALSE to all buildmer
functions to disable REML detection and always use ML. This may be useful if
you are fitting GLMMs using glmmTMB and do not want to use their REML
approximation.
. Nonconvergence of the fixed-effects part of the term-ordering step is now
handled properly
. Fix bugs in order() and forward() breaking ML-only GLMM fits (GitHub issue #2)
. The 'include' argument was not being processed correctly when a cluster was
provided. This is now fixed.
1.2.1 (2019-09-03)
. Avoid the use of reformulate(), as its env= argument is only available on R
>= 3.6 (bug reported by Willemijn Heeren)
. Relax the convergence checks for negative eigenvalues (affects mgcv and
glmmTMB models)
. Improve detection of random effects
. Remove terms by block when a reference model fails to converge and detect
empty models
1.2 (2019-08-27)
** Proper environment support
** glmertree support
* More formal support for diagonal covariance structures via a new 'groups'
argument to tabulate.formula
* Retire the 'quiet' argument
* Overhaul documentation
* Significantly improve family handling
* Allow 'include' to take a tabulated formula
* conv() now accepts a singular.ok argument, which is always FALSE inside
buildmer
. Improve REML detection
. Increase robustness with user-provided formula tab
. Make convergence checks more consistent between model types
. Rework add.terms and related random-effects detection to fix obscure bug
when tabulating non-r.e. terms that fool lme4::findbars (e.g. 'ar1(x|g)')
1.1 (2019-05-18)
** Large rewrite of internal code organization. You can now specify arbitrary
fitting functions, criteria, and elimination functions.
* Add 'include' argument to force terms to always be in the formula (fixes
github bug #1)
* New function 'buildcustom'
* Added raw log-likelihood as a criterion: 'LL'
. Fix bug in calculating anova table with lmerTest ddf options
. Make calcWald respect numerator df in anova table
. Remove incorrect gradient check for mgcv models in conv(), instead check for
'full convergence' string
1.0 (2019-03-31)
*** First public release