saveoptsumandrestoreoptsum!now supportGeneralizedLinearMixedModels #791unfit!(::GeneralizedLinearMixedModel)(called internally byrefit!) now does a better job of fully resetting the model state #791
- lower and upper edges of profile confidence intervals for REML-fitted models are no longer flipped #785
issingularnow accepts comparison tolerances through the keyword argumentsatolandrtol. #783
- Added additional precompilation for rePCA. #749
- Fix a bug in the handling of rank deficiency in the
simulate[!]code. This has important correctness implications for bootstrapping models with rank-deficient fixed effects (as can happen in the case of partial crossing of the fixed effects / missing cells). #778
- Use
publickeyword so that users don't see unnecessary docstring warnings on 1.11+. #776 - Fix accidental export of
datasetanddatasetsand make thempublicinstead. #776
- Use more sophisticated checks on property names in
restoreoptsumto allow for optsums saved by pre-v4.25 versions to be used with this version and later. #775
- Add type notations in
pwrss(::LinearMixedModel)andlogdet(::LinearMixedModel)to enhance type inference. #773 - Take advantage of type parameter for
StatsAPI.weights(::LinearMixedModel{T}). #772 - Fix use of kwargs in
fit!((::LinearMixedModel): #772- user-specified
σis actually used, defaulting to existing value REMLdefaults to model's already specified REML value.
- user-specified
- Clean up code of keyword convenience constructor for
OptSummary. #772 - Refactor thresholding parameters for forcing near-zero parameter values into
OptSummary. #772
- Re-export accidentally dropped export
lrtest. #769
- Properties for
GeneralizedLinearMixedModelnow default to delegation to the internal weightedLinearMixedModelwhen that property is not explicitly handled byGeneralizedLinearMixedModel. Previously, properties were delegated on an explicit basis, which meant that they had to be added manually as use cases were discovered. The downside to the new approach is that it is now possible to access properties whose definition in the LMM case doesn't match the GLMM definition when the GLMM definition hasn't been explicitly been implemented. #767
- Fix for
simulate!when only the estimable coefficients for a rank-deficient model are provided. #756 - Improve handling of rank deficiency in GLMM. #756
- Fix display of GLMM bootstrap without a dispersion parameter. #756
- Support for rank deficiency in the parametric bootstrap. #755
- Use
muladdwhere possible to enable fused multiply-add (FMA) on architectures with hardware support. FMA will generally improve computational speed and gives more accurate rounding. #740 - Replace broadcasted lambda with explicit loop and use
one. This may result in a small performance improvement. #738
- Switch to explicit imports from all included packages (i.e. replace
using Foobyusing Foo: Foo, bar, baz) #748 - Reset parameter values before a
deepcopyin a test (doesn't change test result) #744
- Comment out calls to
@debug#733 - Update package versions in compat and change
Aqua.test_allargument name #733
- Support for equal-tail confidence intervals for
MixedModelBootstrap. #715 - Basic
showmethods forMixedModelBootstrapandMixedModelProfile. #715 - The
hide_progresskeyword argument toparametricbootstrapis now deprecated. Users should instead useprogress(which is consistent with e.g.fit). #717
- Auto apply
Grouping()to grouping variables that don't already have an explicit contrast. #652
- The
.tblproperty of aMixedModelBootstrapnow includes the correlation parameters for lower triangular elements of theλfield. #702
- New method
StatsAPI.coefnames(::ReMat)returns the coefficient names associated with each grouping factor. #709
- More user-friendly error messages when a formula contains variables not in the data. #707
- EXPERIMENTAL New kwarg
amalgamatecan be used to disable amalgation of random effects terms sharing a single grouping variable. Generally,amalgamate=falsewill result in a slower fit, but may improve convergence in some pathological cases. Note that this feature is experimental and changes to it are not considered breakings. #673 - More informative error messages when passing a
DistributionorLinktype instead of the desired instance. #698 - More informative error message on the intentional decision not to define methods for the coefficient of determination. #698
- EXPERIMENTAL Return
finitialwhen PIRLS drifts into a portion of the parameter space that yields a (numerically) invalid covariance matrix. This recovery strategy may be removed in a future release. #616
- Support for check tolerances in deserialization. #703
- Support for different optimization criteria during the bootstrap. #694
- Support for combining bootstrap results with
vcat. #694 - Support for saving and restoring bootstrap replicates with
savereplicatesandrestorereplicates. #694
- New function
profilefor computing likelihood profiles forLinearMixedModel. The resultantMixedModelProfilecan be then be used for computing confidence intervals withconfint. Note that this API is still somewhat experimental and as such the internal storage details ofMixedModelProfilemay change in a future release without being considered breaking. #639 - A
confint(::LinearMixedModel)method has been defined that returns Wald confidence intervals based on the z-statistic, i.e. treating the denominator degrees of freedom as infinite. #639
raneftablesreturns aNamedTuplewhere the names are the grouping factor names and the values are someTables.jl-compatible type. This type has been changed to aTablefromTypedTables.jl. #682
- Precompilation is now handled with
PrecompileToolsinstead ofSnoopPrecompile. #681 - An unnecessary explicit
Varargin an internal method has been removed. This removal eliminates a compiler warning about the deprecatedVarargpattern. #680
- The pirated method
Base.:/(a::AbstractTerm, b::AbstractTerm)is no longer defined. This does not impact the use of/as a nesting term in@formulawithin MixedModels, only the programmatic runtime construction of formula, e.g.term(:a) / term(:b). If you requireBase.:/, thenRegressionFormulae.jlprovides this method. (Avoiding method redefinition when usingRegressionFormulae.jlis the motivating reason for this change.) #677
raneftablesreturns aNamedTuplewhere the names are the grouping factor names and the values are someTables.jl-compatible type. Currently this type is aDictTablefromTypedTables.jl. #634
- Rank deficiency in prediction is now supported, both when the original model was fit to rank-deficient data and when the new data are rank deficient. The behavior is consistent but may be surprising when both old and new data are rank deficient. See the
predictdocstring for an example. #676 - Multithreading in
parametricbootstrapwithuse_threadsis now deprecated and a noop. With improvements in BLAS threading, multithreading at the Julia level did not help performance and sometimes hurt it. #674
- Support
StatsModels0.7, drop support forStatsModels0.6. #664 - Revise code in benchmarks to work with recent Julia and PkgBenchmark.jl #667
- Julia minimum compat version raised to 1.8 because of BSplineKit #665
- Use
SnoopPrecompilefor better precompilation performance. This can dramatically increase TTFX, especially on Julia 1.9. #663
- Don't fit a GLM internally during construction of GLMM when the fixed effects are empty (better compatibility with
dropcollinearkwarg in newer GLM.jl) #657
- Allow predicting from a single observation, as long as
Grouping()is used for the grouping variables. The simplified implementation ofGrouping()also removes several now unnecessaryStatsModelsmethods that should not have been called directly by the user. #653
- More informative error message for formulae lacking random effects #651
- Replace separate calls to
copyto!andscaleinflate!inupdateL!withcopyscaleinflate!#648
- Avoid repeating initial objective evaluation in
fit!method forLinearMixedModel - Ensure that the number of function evaluations from NLopt corresponds to
length(m.optsum.fitlog) whenisone(thin)`. #637
- Relax type restriction for filename in
saveoptsumandrestoreoptsum!. Users can now pass any type with an appropriateopenmethod, e.g.<:AbstractPath. #628
- Attempt recovery when the initial parameter values lead to an invalid covariance matrix by rescaling #615
- Return
finitialwhen the optimizer drifts into a portion of the parameter space that yields a (numerically) invalid covariance matrix #615
- Support transformed responses in
predict#614 - Simplify printing of BLAS configuration in tests. #597
- Add precompile statements to speed up first
LinearMixedModeland BernoulliGeneralizedLinearModelfit #608
- Efficiency improvements in
predict, both in memory and computation #604 - Changed the explanation of
predict's keyword argumentnew_re_levelsin a way that is clearer about the behavior when there are multiple grouping variables. #603 - Fix the default behavior of
new_re_levels=:missingto match the docstring. Previously, the default was:population, in disagreement with the docstring. #603
- Loosen type restriction on
shortestcovint(::MixedModelBootstrap)toshortestcovint(::MixedModelFitCollection). #598
- Experimental support for initializing
GeneralizedLinearMixedModelfits from a linear mixed model instead of a marginal (non-mixed) generalized linear model. #588
- Allow constructing a
GeneralizedLinearMixedModelwith constant response, but don't update theLmatrix nor initialize its deviance. This allows for the model to still be used for simulation where the response will be changed before fitting. #578 - Catch
PosDefExceptionduring the first optimization step and throw a more informativeArgumentErrorif the response is constant. #578
- Fix type parameterization in MixedModelsBootstrap to support models with a mixture of correlation structures (i.e.
zerocorrin some but not all RE terms) #577
- Add a constructor for the abstract type
MixedModelthat delegates toLinearMixedModelorGeneralizedLinearMixedModel. #572 - Compat for Arrow.jl 2.0 #573
- Add support for storing bootstrap results with lower precision #566
- Improved support for zerocorr models in the bootstrap #570
- Add support for zerocorr models to the bootstrap #561
- Add a
Base.length(::MixedModelsFitCollection)method #561
- Add support for specifying a fixed value of
σ, the residual standard deviation, inLinearMixedModel.fittakes a keyword-argumentσ.fit!does not exposeσ, butσcan be changed after model construction by settingoptsum.sigma. #551 - Add support for logging the non-linear optimizer's steps via a
thinkeyword-argument forfitandfit!. The default behavior is 'maximal' thinning, such that only the initial and final values are stored.OptSummaryhas a new fieldfitlogthat contains the aforementioned log as a vector of tuples of parameter and objective values.#552 - Faster version of
leverageforLinearMixedModelallowing for experimentation using the sum of the leverage values as an empirical degrees of freedom for the model. #553, see also #535 - Optimized version of
condVarwith an additional method for extracting only the conditional variances associated with a single grouping factor. #545
- Drop dependency on
BlockArraysand use aVectorof matrices to represent the lower triangle in packed, row-major order. The non-exported functionblockcan be used for finding the correspondingVectorindex of a block. #456 simulate!now marks the modified model as being unfitted.- Deprecated and unused
namedargument removed fromranef#469 - Introduce an abstract type for collections of fits
MixedModelFitCollection, and makeMixedModelBootstrapa subtype of it. Accordingly, rename thebstrfield tofits. #465 - The response (dependent variable) is now stored internally as part of the
the renamed
FeMatfield, now calledXymat#464 - Replace internal
statscholeskyandstatsqrfunctions for determining the rank ofXbystatsrank. #479 - Artifacts are now loaded lazily: the test data loaded via
datasetis downloaded on first use #486 ReMatandPCAnow support covariance factors (λ) that areLowerTriangularorDiagonal. This representation is both more memory efficient and enables additional computational optimizations for particular covariance structures.#489GeneralizedLinearMixedModelnow includes the response distribution as one of its type parameters. This will allow dispatching on the model family and may allow additional specialization in the future.#490saveoptsumandrestoreoptsum!provide for saving and restoring theoptsumfield of aLinearMixedModelas a JSON file, allowing for recreating a model fit that may take a long time for parameter optimization. #506- Verbose output now uses
ProgressMeter, which gives useful information about the timing of each iteration and does not flood stdio. Theverboseargument has been renamedprogressand the default changed totrue. #539 - Support for Julia < 1.6 has been dropped. #539
- New
simulate,simulate!andpredictmethods for simulating and predicting responses to new data. #427
- It is now possible to construct
RandomEffectsTerms at run-time fromTerms (methods forBase.|(::AbstractTerm, ::AbstractTerm)added) #470 RandomEffectsTerms can have left- and right-hand side terms that are "non-concrete", andapply_schema(::RandomEffectsTerm, ...)works more like other StatsModels.jlAbstractTerms #470- Methods for
Base./(::AbstractTerm, ::AbstractTerm)are added, allowing nesting syntax to be used withTerms at run-time as well #470
- Add support for NLopt
maxtimeoption toOptSummary#524
- Add support for
condVarfor models with a BlockedSparse structure #523
- Add
condVarandcondVartablesfor computing the conditional variance on the random effects #492 - Bugfix: store the correct lower bound for GLMM bootstrap, when the original model was fit with
fast=false#518
- Add
likelihoodratiotestmethod for comparing non-mixed (generalized) linear models to (generalized) linear mixed models #508.
- Explicitly deprecate vestigial
namedkwarg inranefin favor ofraneftables#507.
- Fix MIME show methods for models with random-effects not corresponding to a fixed effect #501.
- The Progressbar for
parametricbootstrapandreplicateis not displayed when in a non-interactive (i.e. logging) context. The progressbar can also be manually disabled withhide_progress=true.#495 - Threading in
parametricbootstrapnow uses aSpinLockinstead of aReentrantLock. This improves performance, but care should be taken when nesting spin locks. #493 - Single-threaded use of
paramatricbootstrapnow works when nested within a larger multi-threaded context (e.g.Threads.@threads for). (Multi-threadedparametricbootstrapworked and continues to work within a nested threading context.) #493
- The value of a named
offsetargument toGeneralizedLinearMixedModel, which previously was ignored #453, is now handled properly. #482
shortestcovintmethod forMixedModelsBootstrap#484
- HTML and LaTeX
showmethods forMixedModel,BlockDescription,LikelihoodRatioTest,OptSummaryandVarCorr. Note that the interface for these is not yet completely stable. In particular, rounding behavior may change. #480
- Markdown
showmethods forMixedModel,BlockDescription,LikelihoodRatioTest,OptSummaryandVarCorr. Note that the interface for these is not yet completely stable. In particular, rounding behavior may change. White-space padding within Markdown may also change, although this should not impact rendering of the Markdown into HTML or LaTeX. The Markdown presentation of aMixedModelis much more compact than the REPL summary. If the REPL-style presentation is desired, then this can be assembled from the Markdown output fromVarCorrandcoeftable#474.
- [experimental] Additional convenience constructors for
LinearMixedModel#449
- Compatibility updates
rankUpdate!method forUniformBlockDiagonalbyDense#447
- Compatibility updates
rankUpdate!method forDiagonalbyDense#446- use eager (install-time) downloading of
TestDataartifact to avoid compatibility issues withLazyArtifactsin Julia 1.6 #444
- Compatibility updates
- Better
loglikelihood(::GeneralizedLinearMixedModel)which will work for models with dispersion parameters #419. Note that fitting such models is still problematic.
simulate!and thusparametricbootstrapmethods forGeneralizedLinearMixedModel#418.- Documented inconsistent behavior in
sdestandvarestGeneralizedLinearMixedModel#418.
- Compatibility updates
- Minor updates for formatting in various
showmethod forVarCorr.
- Nested grouping factors can be written using the
/notation, as in@formula(strength ~ 1 + (1|batch/cask))as a model for thepastesdataset. - The
zerocorrfunction converts a vector-valued random effects term from correlated random effects to uncorrelated. (See theconstructorssection of the docs.) - The
fulldummyfunction can be applied to a factor to obtain a redundant encoding of a categorical factor as a complete set of indicators plus an intercept. This is only practical on the left-hand side of a random-effects term. (See theconstructorssection of the docs.) Grouping()can be used as a contrast for a categorical array in thecontrastsdictionary. Doing so bypasses creation of contrast matrix, which, when the number of levels is large, may cause memory overflow. As the name implies, this is used for grouping factors. #339
- Checks for rank deficiency in the model matrix for the fixed-effects parameters have been improved.
- The distinction between
coef, which always returns a full set of coefficients in the original order, andfixef, which returns possibly permuted and non-redundant coefficients, has been made consistent across models.
- The
parametricbootstrapfunction and the struct it produces have been extensively reworked for speed and convenience. See thebootstrapsection of the docs.
- The PCA property for
MixedModeltypes provides principal components from the correlation of the random-effects distribution (as opposed to the covariance) - Factor loadings are included in the
printmethod for thePCAstruct.
- An
AbstractReMattype has now been introduced to support #380 work on constrained random-effects structures and random-effects structures appropriate for applications in GLM-based deconvolution as used in fMRI and EEG (see e.g. unfold.jl.) - Similarly, a constructor for
FeMat{::SparseMatrixCSC,S}has been introduced #309. Currently, this constructor assumes a full-rank matrix, but the work on rank deficiency may be extended to this constructor as well. - Analogous to
AbstractReMat, anAbstractReTerm <: AbstractTermtype has been introduced #395. Terms created withzerocorrare of typeZeroCorr <: AbstractReTerm.
- Several data sets from the literature were previously saved in
.rdaformat in thetestdirectory and read using theRDatapackage. These are now available in anArtifactin theArrowformat #382. - Call
MixedModels.datasets()to get a listing of the names of available datasets - To load, e.g. the
dyestuffdata, useMixedModels.dataset(:dyestuff) - Data sets are loaded using
Arrow.Tablewhich returns a column table. Wrap the call inDataFrameif you prefer aDataFrame. - Data sets are cached and multiple calls to
MixedModels.dataset()for the same data set have very low overhead after the first call.
describeblockshas been dropped in favor of theBlockDescriptiontype- The
namedargument toranefhas been dropped in favor ofraneftables