NEWS | R Documentation |
Faster fits by sparse-precision methods when (1) several random effects are present (2) random-coefficient terms are present. (Very) small speed gains in many other fits.
Fixed an inconsequential clang compiler warning (... indeed, no real bug fix in this version).
New 'Suggests': 'memoise', memoisation being optionally used on a few spaMM functions.
New type of random effect handled: "composite" random effects such as corrMatrix(<variable>|ID) or corrMatrix(mv(1,2)|ID), combining features of a random-coefficient model and of a corrMatrix (see help("composite-ranef")). Such random effects are considered in quantitative genetic analyses of multivariate response, for example.
residuals.HLfit() can now also return standardized deviance residuals used in fitting methods and in diagnostic plots.
Much more memory-efficient (and faster) fits for 'nested-geostatistical' random effects (e.g., Matern(1|long+lat %in% year) ).
The 'control.HLfit' argument of fitting functions now handles an 'algebra' parameter for easier control of matrix-algebraic methods. See help("algebra") for possible usage.
Procedures which rest on the stats::gaussian(<"log"|"inverse">) family object to fit models with such response family to data with zero(es) in the response may fail to find valid starting values, which is not totally defensible. spaMM now circumvents this obstruction. Whether asking for fits of such data by such families is a good idea is another matter; the principle adopted here is to allow the fit and let its results contribute to answer this question.
pdep_effects() now handles 'length.out=0L' with the effect that predictions are made for actual values of the predictor variables, then allowing for adding informative plots of fit-residuals on the partial-dependence plots.
For random-coefficient terms fitted by sparse-precision methods, there was a discrepancy between some elements stored in the fitted object. Neither the fits nor predictions were affected, but trying to use together the results of extractors get_ZALMatrix() and ranef.HLfit() would have been confusing. This discrepancy has been removed.
plot.HLfit() should now return a diagnostic plot for standardized deviance residuals in all cases.
More consistent default value of the residual variance of the residual variance model (sic): it is now set to a fixed value in mixed-effect residual-variance models specified by resid.model=<formula>, as was already effectively so for other specifications of the residual-dispersion model.
predict(., <with newdata>) results were incorrect for models with multivariate random-coefficient terms (i.e., using the mv() syntax) fitted by sparse-precision methods.
Prediction variance was incorrect for models with random-coefficient terms.
Prediction variance computation could fail when all random-effect variances were fixed.
residVar(., which="var") had incorrect results for gaussian-response models. This did not affect response-variance computations nor get_residVar(), which did not call this extractor. residVar() could also fail on objects with an offset-only formula for phi.
Improved condition for selection of sparse-precision methods for models with 'corrMatrix' terms, with quite substantial gains in speed for affected fits.
Numerical approximations for COMPoisson-family models have been substantially revised, with disappointing gains in speed. Additional gains are possible if the 'memoise' package has been installed on your R system.
The 'control.HLfit' argument of fitting functions now handles a 'sparse_precision' parameter (previously controllable only globally through spaMM.options()). See help("sparse_precision") for possible usage.
The criterion for switching to Levenberg-Marquardt-like algorithm has been slightly modified and hopefully improved.
New 'Gryphon' data set.
Long-deprecated argument 'uniqueGeo' of HLCor() is now formally removed.
Prediction variance computations could fail for models with random-coefficient terms and not fitted by sparse-precision methods, due to a bug introduced in v3.7.48 (i.e., CRAN versions 3.8.0 and 3.8.9)
Better control of arguments to avoid bugs [eval() looking for the IMRF 'model' in the global environment] when fitting is called in a programming context.
Bug (stop) when calling predict.HLfit() with both 'intervals' calculation and binding of point predictions to data.
Substantially revised documentation for "sparse_precision".
Modified selection between ”outer” optimization versus ”inner” iterations to acheive faster default fitme() fits of some non-spatial GLMMs.
The default cluster type is now "PSOCK" in all cases.
New interface dofuture() for parallelisation. See help("wrap_parallel") for differences from dopar().
The 'control.HLfit' argument of fitting functions now handles a 'sparse_precision' parameter (previously controllable only globally through spaMM.options()). See help("sparse_precision") for possible usage.
Improved condition for selection of sparse-precision methods for models with 'corrMatrix' terms.
Typo in a recently introduced check of geographical coordinates led to a warning when abs(longitude) values >90 occurred.
Tests using the 'adjlg' data set no longer worked. It was the code generating the data set, rather than any spaMM procedure, that required fixing.
Using fork clusters provided statistically correct results but could not be controlled as documented (the random number generator and progress bar could not be properly controlled, and incorrect warnings could be displayed).
New 'Suggests': 'future' and 'future.apply'.
Parallel computations can now use forking via parallel::mclapply.
The conditional AIC can now be evaluated by a bootstrap procedure.
Checks that coordinates are consistent with longitude and latitude for Earth and EarthChord distances.
Previous version had become slower for some spatial GLMMs without residual dispersion parameter. This is fixed.
Fits with "adjacency" random effects failed when data were missing for some locations.
predict(., binding= < char string >, blockSize = < any value > ) no longer worked when predicting more than 'blockSize' values.
For Gamma-response model, get_residVar() (and same computation through call to predict()) returned the dispersion parameter phi rather than the residual variances.
Use of the dot in a formula, to refer "to the remaining variables contained in data" (cf ?terms.formula), is now handled, with a warning about its limitations. It could previously lead to errors (stop), e.g. if all such variables were then removed ( ~ . - ... - ...) to leave only the Intercept.
Substantially revised documentation for prediction variances, with new ?predVar page, and for AIC computations.
Non-canonical links implemented for the COMPoisson family.
More precise computation of prediction variance for some cases with vanishing residual variance (and possibly a few other cases).
New 'which' value in residVar(.,which="fam_parm") to extract the family dispersion parameter of COMPoisson and [T]negbin families.
A "table" attribute in the format of the stats::confint() result is added to the confint.HLfit() result for profile confidence intervals.
Possibly large speed gains for a small class of REML fits on large data sets/random effects with many levels, due to these being now more likely to be fitted by “outer” optimization.
Fits of models with a resid.model of the form ~ 0 + offset(.) should also be faster. This affects the 'mean_fit' performed by IsoriX::isofit(), for example.
how.HLfit()'s message about fitting method is less cryptic. Its return value is unchanged.
Warning added about potential back-compatibility issue in some multinomial fits due to changes in default optimization strategy.
Issue with missing data in multivariate-response fits, that could show up when summary()zing the fit.
Prediction variance calculations with a re.form of the form '[no LHS] ~ <single random-effect>' ignored that random-effect (this did not occur with a dummy LHS nor with several random effects).
VarCorr() would report NA instead of the variance for random-slope terms without correlation parameter [cf terms of the form ( 0 + X | <grouping factor> )].
summary.HLfit() could fail in some programming contexts.
Expressions such as I(y>0) or even y>0 are now accepted in the left-hand side of a model formula (up to now update_resp() and bootstrap functions failed on them, and I(.) was explicitly disallowed).
5-10% speed gains in many fits (at least non-spatial ones), and even more, notably (1) for Levenberg-Marquard-like algorithm (hence for fits affected by convergence issues, which tend to be the longest ones), and (2) for fits by sparse-precision methods for (2a) Gamma(log) GLMMs (>fifthfold in a large problem!) or (2b) other GLMMs when fitted by PQL rather than ML.
In addition, the criterion for using “inner” iterative methods rather than “outer” optimization ones for fitting dispersion parameters has been modified, resulting in substantially faster fits in some non-spatial models.
New diagnostic message pointing a potential issue when fitting models with group-specific geostatistical effects without matching Intercept terms.
Back-compatibility fixes for summaries and AIC computations on old fit objects (tested against v2.5.0 objects).
(major) New function fitmv() to fit multivariate-response models, extending most of the features of fitme().
LRT() (aka anova.HLfit()) can now perform a bootstrap test for models differing (also) by their random effects.
fitme(), and the new fitmv(), accept partially-fixed random coefficients parameters, as in 'fixed=list(ranCoefs=list("1"=c(NA,0.5,NA)))'
get_inits_from_fit() now gets also the estimates of the shape parameters of the negbin and COMPoisson families.
New function map_ranef() to plot a map of a spatial random effect (using only graphic functions from base R, but other functions can use the returned grid of values).
Improved control of names of user-provided fixed lambda values (new error message may result).
poly() in random-coefficient terms is now handled.
New extractor residVar() for residual variances, with features different from get_residVar().
Most GLMM fits using sparse-precision algorithm should be substantially faster (near 50% for long fits).
Better control of progress bars as the global 'barstyle' option is no longer used in some cases where a local argument should be used.
The 'uniqueGeo' argument of HLCor() has long been obsolete and is now deprecated.
VarCorr.HLfit() gets a new argument 'add_residVars' to make information about residual variance optional (also fixing potential bugs).
residuals() did not always return a vector.
confint() on PQL[/L] fits had several issues, in particular easily producing spurious messages suggesting that the fit was inaccurate.
One could not use a Matern(RHS|.) (for non-unit RHS) together with a distMatrix for this Matern term.
Improve a backward incompatibility issue in the (internal) post-fit function determining which fitting function was used to produce a fit.
The recently introduced extractor get_RLRsim_args() did not correctly handle marginal LR tests (as opposed to REML ones).
the global 'barstyle' option was always directly used, so that the 'showpbar' argument did not control the progress bar in dopar().
plot.HLfit() reported incorrect standardized deviance residuals for binomial, non-binary response.
Several post-fit functions such as predict() could report incorrect results for 'IMRF' (and perhaps 'adjacency') models since version 3.4.1.
The (hardly ever tested) 'sqrt' family link was not fully implemented for families that allows it.
summary() printed an incorrect number of groups for random-coefficients terms (long known and not bothered about until now).
Documentation for the 'distMatrix' argument of HLCor() and 'higher' fitting functions was incomplete.
simulate(., re.form=NULL) was broken.
random-coefficient parameters fitted by "outer optimization" where not counted in the degrees of freedom for AIC.
Fitting some LMMs with fixed 'lambda' could return incorrect results.
Typo in printed summary for rand.family=Gamma (wrong 'scale' value).
It was given a new scrutiny, leading to a display bug fix, a revised fitted() output, improved summary() information, a new how.HLfitlist() extractor, and revised documentation. But fitmv() might ultimately be more useful to fit multinomial data.
'Nested' Matern() or Cauchy() random effects are now fitted, using a syntax of the form Matern(1|x+y %in% grouping_factor).
'barstyle' option controls more progress bars than previously.
New extractor get_RLRsim_args() extends previousy existing get_RLRTSim_args().
get_ranPars() handles new 'which' values for extraction of variance parameters.
More judicious and limited production of warnings in some prediction variance computations.
An 'invalid type/length (S4/0) in vector allocation' bug occurring for recent spaMM versions on 3.6-ish versions of R.
Glitch in VarCorr() for conditional autoregressive fits (a line is now filled with NA's rather than with inappropriate values).
IMRF()
terms now also handle objects created by INLA::inla.spde2.matern() with non-integer alpha values, and therefore a broader range of smoothness values (PS: doc corrected wrt this only in version 3.6.0).
objects created by INLA::inla.spde2.pcmatern() are also handled.
'sparse_precision' fitting method can be notably faster than before for large data sets. By default, this will affect fits of models with adjacency or IMRF terms.
Fitting family=Gamma(log) mixed models will be faster than before, twofold or possibly more for large data sets.
confint.HLfit() now returns a list of confidence intervals when parm
defines a vector of parameters. In the bootstrap case, the same bootstrap simulations can thus be used for several parameters.
get_predCov_var_fix() can now handle 'multIMRF' random effects correctly (for the happy few who would combine these two relatively cryptic functionalities).
formula_env
specification to control a formula environment: see help("formula_env").
The structure of the confint.HLfit() result is modified as this function no longer hacks the boot::boot.ci() results for printing purposes.
The 'distMatrix' argument was no longer effective.
Prediction variance computations could fail for 'IMRF' random effects with 'newdata'.
confint() could fail when the object had been fitted by a "non-vanilla" call of the fitting function (e.g., 'spaMM::fitme' instead of 'fitme').
confint() by sparse-precision method could return incorrect results.
is_separated.formula() was not returning results properly for binomial non-binary response.
A case where LevenbergM for PQL could underfit was found.
confint(.,parm=<fixed-effect coefficient>) failed (stop) on sparse-precision fits with a single fixed-effect coefficient (typically "(Intercept)").
spaMM is now formally declared to enhance 'INLA'.
confint.HLfit() can now compute bootstrap confidence intervals by acting as a wrapper for boot::boot.ci(), and is then not limited to fixed-effect coefficients.
New hatvalues() extractor.
Faster and more memory-efficient step in Levenberg-Marquardt-like algorithm for "big" random effects fitted by sparse-correlation algorithms.
Faster and more memory-efficient step in prediction code for "big" random effects.
New option 'barstyle' to control globally the style of progress bars. New argument 'showpbar' for function dopar() and spaMM_boot().
Informative warning when some variables are not found in parallel context. New argument 'pretest_fn' for function dopar().
Better criteria for selection of sparse-precision algorithm, though far from perfect.
VarCorr() extractor should now handle all spaMM fit results.
Added 'seed' argument to spaMM_boot() and fixedLRT().
'ROI.plugin.glpk' moved from Imports to Suggests, removing a "hard" dependency to the external library glpk.
'boot' in Imports.
Added formal argument 'intervals' for get_intervals(). This does not affect any functionality, but explicitly shows that different intervals can be computed by changing this argument.
All mixed-effect fitting functions now have a formal argument 'method' equivalent to 'HLmethod'. 'HLmethod' can still be used in a backward-compatible way.
simulate(<Gamma-response fit>) could return incorrectly formatted results when nsim>1.
get_intervals(, which="respVar") (and equivalent syntaxes) returned an incorrect result when the family link was not 'identity'.
Failure of fitting by sparse-precision method a model specified by a 'corrMatrix' argument with more levels than present in the data.
ranef() extractor failed on sparse-precision fits with correlated random effects.
prediction variance could be wrong for models with random-coefficient terms and fitted by sparse-precision methods.
get_cPredVar() and and confint() no longer stop on fixed-effects models.
plot.HLfit() could stop in some cases.
eval_replicate() could stop in some cases.
Non-default optimizer optim() failed for models with random-coefficient terms. Selecting optim() is not recommended anyway.
Re-emergent glitches in how() when the object had been fitted by a "non-vanilla" call of the fitting function.
Substantial sections of the documentation have been rewritten, including new documentation items 'method' and 'random-effects', for easier navigation, updating information, and removing obscurities tracing back to early versions of spaMM.
New documentation 'external-libraries'.
Modified criterion for default selection of sparse precision methods.
is_separated.formula() to test separation with the same arguments as for a fitting function call.
Newly exported and documented function dopar() for parallel execution of a function on columns of a matrix. Not logically restricted to mixed-effect applications, hence it can be used more widely.
Finer control of verbosity in simulate.HLfit() and predict.HLfit().
New 'format' argument for how.HLfit().
'ROI' and 'ROI.plugin.glpk' replace 'lpSolveAPI'.
Several errors in simulate() called on COMPoisson-reponse models.
Anecdotal glitches in how() when the object was fitted by a "non-vanilla" call of the fitting function.
non-default 'REMLformula', as illustrated in example("update.HLfit"), could not be fitted by sparse-precision algorithms.
Sparse-precision fits of CAR models could crash for some format of the adjacency matrix.
simulate(<Gamma-response fit>) simulated according to an incorrect residual variance.
Faster code for some autoregressive models (mostly for autoregressive LMMs fitted by fitme(), as in tests described at the end of the ”gentle Introduction” which are ca. 50% faster).
Faster and more memory-efficient code computation of prediction variance for large designs. 'variances$as_tcrossfac_list' control for more memory-efficient (albeit less human-readable) reporting of large prediction covariance matrices.
The obsolete functions designL.from.Corr() and sym_eigen() have been removed.
'RSpectra' suggested for a step in fitting autoregressive models.
Better numerical precision of some computations for binomial models (including a more accurate alternative to binomial(.)$variance(mu)) may avoid numerical imprecisions (or worse) for extreme 'mu'.
More rigorous variant of the Levenberg-Marquardt-like algorithm will give better (albeit slower) fits in difficult cases such as binary models with strong spatial correlations, and Gamma GLMMs with response values approaching 1.
Default optimizer is not always nloptr(); nloptr() or bobyqa() are now selected depending on input (though nloptr() remains the most common choice).
Stricter checking of arguments for adjacency models (the adjacency matrix and any fixed 'rho' parameter).
Prediction variances more accurate when correlation matrices of random effecs are nearly singular, for random-coefficient and Matern/Cauchy types of random effects.
Algorithms for fitting COMPoisson-response models made more numerically stable for extreme values of the linear predictor and for low COMP_nu parameter.
New function eval_replicate() is the new default 'simuland' argument of the likelihood-ratio testing functions, which themselves better handle imperfections of the bootstrap results.
Multiple changes to spaMM_boot(), to make it consistently usable in programming, to control parallel::makeCluster(), etc.
The '...' argument of LRT() is now passed to spaMM_boot().
New extractor get_inits_from_fit().
By default, spaMM.options() now checks whether in arguments are preexisting options or not. This helps to avoid errors due to typos. The argument 'warn' controls this check.
New argument 'verbose' for how().
plot_effects() now produces more suitable output than before for binomial models.
simulate.HLfit() can perform simulation from linear predictors randomized in various ways (see Details section of its documentation).
New argument in get_predVar(., which) provides an alternative extraction method for various components of prediction variance.
spaMM.options(separation_max=.) can now be used to control also the checks of separation using the lpSolveAPI package.
'help("post-fit")' explains how to use multiple-comparison procedures implemented in multcomp::glht on spaMM's fit objects.
fitme() will select iterative methods, rather than "outer" optimization, in some cases where it did not but where it is faster.
Computations using the Levenberg-Marquardt-like algorithm tend to be longer than before (there may be exceptions due to other changes).
Changes in formal arguments of LRT() (a.k.a. anova.HLfit()) and fixedLRT(). These should be backward compatible, as important removed arguments are now handled through the '...'.
SpatialPolygonsDataFrame objects 'worldcountries' and 'oceanmask' removed from the package. See help("seaMask") for availability.
'crayon'; one may not yet see any effect unless verbose=c(TRACE=TRUE) is used.
'minqa' and 'gmp' moved from Suggests to Imports.
Updated dependency to lpSolveAPI >= 5.5.2.0.17.6 as it seems faster than previous versions on "big problems".
Interval computation by predict() was not correct for zero-truncated response families.
Fitting a truncated negbin model could fail due to numerical precision issues.
HLfit() stopped on non-gaussian response models with several random effects including at least one random-coefficient one (although fitme() allowed to fit such models). Gaussian-response HLfit()'s with the same random-effect structure were imprecise, with typically third-decimal imperfections on the logLik.
Substantially faster fits of random-coefficient models.
Faster procedure for some simple LMMs.
New VarCorr() extractor.
confint() failed (stop) on models fitted by fast procedures suitable for a class of LMMs.
The names appearing in certain outputs of lambda values were incorrect.
Fits could fail (stop) in the uncommon case where some fixed-effect coefficients were fixed by the user and non-standard REML, as defined by a non-default REML formula, was requested.
Prediction variance failed (stop) given more than two random effects.
'help(fixed)' completed ('corrPars' information missing), corrected and clarified.
Trendy stuff: geostatistical analyses based on a Markov Random Field on a lattice are now implemented through the new random-effect syntax 'IMRF' (Interpolated Markov Random Field).
Improved implementation of sparse-precision algorithms lead to increases in speed that can exceed 50 % for fits that use such methods.
New syntax introduced to fit some heteroscedastic non-Gaussian random effects (see help("spaMM")).
Better choice of default optimization strategy by fitme() may lead to much faster fits of Gamma-response models.
Restriction of [filled.]mapMM() to plots with a single spatial random effect removed, as long as the spatial effects remain together in 2D space.
Restriction against predict.HLfit(.,newdata=...), on models with corrMatrix or adjacency terms, lifted when the 'newdata' do not involve levels absent from the correlation structure for such terms.
New functions pdep_effects() and plot_effects() to quantify and plot the effects of a variable.
Better control of damping factor of Levenberg-Marquardt algorithm.
Better assessment of convergence of Levenberg-Marquardt algorithm for PQL fits.
New features of LRT() and spaMM_boot() to facilitate debugging of bootstrap simulations.
New package option 'Gamma_min_y' to control numerical problems with low response values in Gamma-response models.
update_resp() now handles additional arguments as update() does.
get_boot_response() is deprecated (and removed), and the 'boot_fn' argument of fixedLRT() is deprecated, and its formerly possible value, the function .eval_boot_replicates(), is deprecated (and removed).
Predictions function could fail (stop) for some combinations of fixed vs. estimated dispersion parameters.
Prediction variance computation could fail (stop) when 'etaFix' argument was used (cryptic case indeed).
LRT() could fail (stop) when prior weights were used.
anova.HLfit() now properly detect non-nestedness of fixed effects.
Information about random number generator was missing from bootstrap output, contrary to the documentation.
Incorrect dfs for AIC computation in the case of inner-optimized rho parameter of a CAR model.
Revised 'help(spaMM_boot)', in particular with respect to the 'simuland' argument.
Informative message when plot.HLfit does not produce a 'mean' plot.
Fix problem in corMatern example (only on CRAN tests for debian+gcc, which started with a system upgrade to glibc 2.28, and which may reveal weaknesses of nlme and/or nlminb rather than spaMM).
New syntax for autocorrelated random-coefficient effects, allowing e.g. the formula terms Matern(male|.) + Matern(female|.) to fit distinct Matern effects for males and females. See help(spaMM) for details.
New function update_resp() to refit on a updated response vector.
New simulate() argument 're.form' for finer control of conditioning.
New predict() argument 'type' to allow returning the linear predictor.
New summary() argument 'verbose' may be convenient when using summary() as an extractor.
New function response() to extract the response from the fit.
The double-vertical syntax for random effects, ( rhs || lhs ), is now handled. See help(spaMM) for details.
HLfit can now fit random-coefficient models by sparse-precision methods.
Faster implementation of some computations in the sparse-precision code.
The warnings that occurred when 'e1071' availability was tested, but the package was not available, has been suppressed.
Modified approach to controlling numerical precision of prediction covariance matrices, should impact these matrices when they are nearly singular.
as_precision() return type documented, and type of the matrix within is itself modified.
Bug, reintroduced in the previous version, in predict()'ing from models with poly() terms.
Bug in simulate(.,type="marginal", newdata=<non-NULL>)
confint() had weaknesses, more likely to impact intervals for simple LMMs since the introduction of new fitting algorithms in version 2.5.0.
Bug in predict()'ing with an unusual combination of random effects and newdata (documented in test/testthat/test-more-predict.R).
Marginal AIC computation did not exclude fixed correlation parameters from number of degrees of freedom.
Sparse-precision code could fail when all fixed-effect coefficents were set by the user using the 'etaFix' argument.
Failure to catch some incorrect or inefficient calls, notably of HLfit to fit random-coefficient models by sparse-precision methods (which is now a correct call!).
Function get_matrix() to extract various matrices from the fit.
Function remove_from_parlist() for manipulating a structured list of parameters.
spaMM_boot() now rbind's into a data frame the results of the 'simuland' calls when each call returns a 1-row data frame (the previous behaviour was undocumented and effectively inconsistent accross different parallel backends).
More explicit error message when variables for prior weights are not in the 'data'.
logLik(., which="logL_Lap") to compute Laplace approximation using observed Hessian (currently only for family=Gamma(log), for demonstration purposes).
Low response variance could result in poor fits. A warning is now issued and several steps have been taken for more robust results.
New family() extractor.
predict() did not generally handle correctly offset terms.
A hack perfomed by spaMM on the value of 'options(error=<.>)' could conflict in some cases with the user setting 'options(error=<expression>)'.
Prediction variance computation with a 'fix_X_ZAC.object' failed when the model included a non-autocorrelated random effet.
Prior-weights information was not shown in model summaries in some cases where it should.
Documentation bug in help("fixed") (with respect to 'etaFix' vs. 'fixed' elements).
predict() with 'newdata' did not work from a fit with non-null 'etaFix$beta'.
A faster-fitting procedure is implemented (through fitme()) for fitting a class of LMMs (without heteroscedasticity, without fixed random-effect variances...).
The restriction, to models with a single random effect, of the computation of the component of prediction variance for uncertainty in dispersion parameters, has been removed.
simulate.HLfit() has new arguments for controlling prior weights and simulation from fits with structured dispersion.
EarthChord distance: see help("EarthChord").
Models with 'geostatistical' random-effect terms (Matern, Cauchy) can now be fitted by sparse-precision methods (although this is not necessarily very efficient).
Possible improvements in numerical accuracy, notably by sparse-precision methods when (1) fitting models with AR1 terms, that are numerically unstable for extreme AR1 correlation values; (2) in REML fits with low residual variance.
Improvement in numerical accuracy of computation of prediction variance when residual variance is low or spatial correlation matrices are nearly-singular.
New function MSFDR() implements the multiple-stage false discovery rate method of Benjamini and Gavrilov 2009.
Improved .LRT() function (called through fixedLRT()) which makes it more suitable for programming, and which better handles generic-optimization issues.
fixedLRT(), LRT() and anova() now call by default spaMM_boot() for the bootstrap. They handle new arguments to control the spaMM_boot() call, allowing finer control of poorly- or non-informative bootstrap samples.
is_separated() convenience function to check for (quasi-)separation in binomial-response models (this was previously implemented but not exported nor documented).
Possibly improved initial value for lambda may affect speed and results of some fits.
New extractors formula(), terms(), nobs(), and extract.AIC(), being methods for 'HLfit' objects of generics defined by 'stats'.
The syntax '.' can now be used in formulas, e.g. HLfit(y ~ (.)^2, ...)
Automatic conversion of correlation matrices to precision factor matrices when sparse-precision algorithm is requested; new as_precision() function for such conversion.
More transparent control of the residual variance parameter of a residual dispersion model. However, this affects the default results if the residual dispersion model is a mixed-effect one. Use 'fixed=list(phi=NA)' in the 'resid.model' (forcing estimation of its phi) to recover past results.
More efficient computation of conditional AIC. Users can also control whether it should be computed along with other information criteria.
New extractor how() to cleanly extract information about how a model fit was obtained.
An hidden memory-dependent constraint on the size of data that could be fit by fixed-effect models by ML has been alleviated.
Numerous improvements for residual dispersion models with random effects makes fitting of the joint residual and main-response models more likely to converge in reasonable time when both models contain spatial terms.
The Predictor() function is deprecated. See help("Predictor") for details.
The designL.from.Corr() function is superseded by mat_sqrt().
The structure of the return value of sym_eigen() is modified.
'resid.formula' argument of fitting functions is formally deprecated (it has been declared obsolete since v1.8.0).
spaMM now enforces variables for prior weights to be in the 'data'.
The computation of prediction variance (specifically its component for uncertainty in dispersion parameters) could be incorrect when several random effects were considered. Fixed as noted in 'New Features'.
The 'random-slope' (Intercept+slope) code was OK, but the more general 'random-coefficients' code (>2 covarying random effects) was not.
Bug in initialization of adjacency correlation parameter, which prevented at least some bootstrap computations.
Fixed a (previously documented) issue in spaMM's update() method, stemming from the distinct interpretation of formulas without explicit fixed-effect terms by spaMM vs. stats. Formula updates should still be carefully checked, as getting them fluent is not on the priority list.
Fits failed when spaMM.options(sparse_precision=TRUE) but a corrMatrix was not specified as a precision matrix.
Prediction failed from models involving poly(., raw=FALSE) terms.
Prediction should use contrast information contained in the fitted object rather than in the session's options().
confint() for fixed-effect models has various problems.
An horrendous bug was found in the implementation of Gamma family for inverse link. Fortunately, no reasonable people would use Gamma(inverse)...
Stricter control of numerical accuracy of COMPoisson-family fits, which does not affect high-nu fits (which were already accurate), but which improve accuracy particularly around nu=1.
Use of the 'lpSolveAPI' package is reintroduced to check (quasi-)separation.
New extractor get_ranPars() (with limited functionality so far).
Parallel computation of some bootstrap computations by foreach+doSNOW has been reinstated.
plot.HLfit has a new argument 'ask', and no longer has a side effect on subsequent plots.
Bug in COMPoisson-family fits, which impacted most clearly the results when some response values were zero.
Crashing bug could occurs in zero-truncated models (for various cases of the Laplace approximation).
Fixed some bootstrap code which could fail (because the original try() statement had disappeared through mis-editing :-( ).
simulate(., type = "(ranef|response)") failed, or returned incorrectly formatted results.
Non-trivial re.form argument now works for models including a 'corrMatrix' term.
Precision matrices with more levels than there are in the data are now handled.
COMPoisson-family fits could crash in some cases (bug introduced in version 2.4.0 in code for initial values).
Calls not requesting any form of least-squares fit, such as fitme(y ~ 0, data = <.>), failed.
Prediction interval computation failed for zero-truncated distributions.
Prediction variance computation could fail for large newdata ('slicing' algorithm).
Cauchy spatial correlation family has been implemented.
For binomial family the response can now be specified as a factor.
Better default selection of methods for autoregressive fits with many random-effect levels may lead to much faster fits.
fitme() now selects estimation methods for phi and lambda by default according to nrow(data), which may change details of the results. To overcome the new defaults, fitme() now allows better control of estimation method and of initial values for phi and lambda.
get_predCov_var_fix() may now handle a non-trivial 're.form'.
New 'max.print' argument for summary.HLfit().
New function spaMM_boot() for parametric bootstrap.
New 'resp_testfn' argument for simulate.HLfit(), to simulate samples that satisfy a given condition.
Revised summary of random-effect parameters.
revised format of ranef() screen output for random-coefficient terms, using new method 'print.ranef'.
Imports package 'pbapply' to draw progress bars, including in parallel computations; no longer Suggests the package 'foreach' (and optionally 'doSNOW') for the same effect.
An error was produced when model formulas included a nested random effect (.|./.) followed by a correlated random effect such as Matern(.).
Incorrect prediction variance with non-trivial 're.form' with random-coefficient term(s) or permutations of the random effects.
Errors when attempting to fit models with correlated random effects and 'negbin' response family.
A wrong algorithm could sometimes be selected for 'inner' estimation of the correlation parameter of CAR models.
Wrong documentation for ranef().
anova() failed when a fit had non-estimable coefficients because of collinearity.
Revised 'help(ranFix)' to clarify the usage of argument for fixing parameters in the different fitting functions.
Revised default initial values of lambda may slightly change results.
Revised default 'fittingFunction' for fixedLRT() may slightly change results, or may call for a user-provided explicit value of 'fittingFunction'.
Corrected the poorly written documentation for COMPoisson.
Zero-truncated variants of Poisson and negative-binomial families are now implemented.
Models with first-order autoregressive (AR1) random-effects terms are now fitted.
fitme() now fits models with random-coefficient terms, using by default a different strategy than the other fitting functions.
Prediction variance computations extended to most models handled by spaMM, including random-slope models.
The 'ranFix' argument of HLfit(), and similar 'ranPars' or 'fixed' arguments of other functions, allow a finer control of variance parameters of random effect than before.
predict.HLfit() should be substantially faster in most practical cases, particularly when the 'newdata' have many rows.
Better approximations are implemented for the infinite sums required for fitting COMPoisson models, in particular for low 'nu'.
New extractor get_rankinfo().
Some gains in speed, particularly for fit by sparse-precision algorithms, which also benefit from better memory management for large model matrices.
Prediction variance computation now take into account uncertainty in the random effect variance for adjacency random effects.
Control of optimizers has been revised, with distinct 'optimizer1D' and 'optimizer' in the spaMM.options(); lower case is now used for 'optimizer' keyword in all user-level code; and minqa::bobyqa() can now be called. See ?spaMM.options for further information.
Argument 'ZALMatrix' of predictor() is deprecated.
Argument 'as_matrix' of get_ZALMatrix() is deprecated.
predict.HLfit() gets a new argument 'blockSize'.
predict.HLfit( . , binding=NA) newly defined to return a vector instead of a one-column matrix. Attributes will differ too.
Removed dependency on package 'ff'.
Since version 2.0.0, the ranef() extractor did not return "correlated" effects except for random-coefficient models.
Prediction variance computation failed for models involving an adjacency random effect and other random effects.
Revision of 'help(autoregressive)' following introduction of AR1 models.
Revision of 'help(fitme)' to account for its new features.
New 'help(fixed)' gathering information on how to fix parameter values.
Simplification and updating of documentation pages for fitting functions.
New or updated documentation for 'negbin', 'Tnegbin', 'Poisson', 'Tpoisson' following implementation of the truncated variants.
Reimplementation of corMatern (forced by effectively backward-incompatible changes in Rcpp) might very slightly change results.
Added '...' argument to LRT(). The function does not yet make anything of it.
Widespread improvement in speed for large data sets, thanks to some of the following changes:
Sparse precision matrix algorithm extended to all HGLMs whose linear predictor involves only a restricted set of random effect terms (excluding mainly Matern and random-coefficient terms). It is used by default for conditional autoregressive models, or when the new argument 'covStruct' is used to provide a sparse precision matrix: see help(covStruct). See also help(pedigree) for a type of applications where the latter syntax may be of particular interest, and help(sparse_precision) for more information.
Significantly faster algorithm for fitting models with large dense correlation matrices (as in Matern correlation models).
new extractor function get_boot_replicate() to get the response vector for a given bootstrap replicate
‘spaMM_glm()’ and ‘spaMM_glm.fit()’ get the same ‘singular.ok=TRUE’ argument that ‘glm()’ and ‘glm.fit()’ got in R-devel r73380 (2017-09-28).
More reliable version of Levenberg-Marquardt algorithm for mixed models (see 'help("LevenbergM")'). Revised rules for using it, so that users typically do not need to think about it.
fixedLRT() results might be modified in particular when the likelihood has local maxima.
Changes in the way the generic optimizers are used may lead to changes in the fitme() and corrHLfit() results.
Usage of 'control.corrHLfit' argument has been modified.
Fitted values always get their names from the row names of the data.
Fixed numerical issues for binomial(cloglog) family with large values of linear predictor.
Random-coefficient models could not be properly fitted by Levenberg-Marquardt algorithm.
simulate.HLfit() can now simulate samples from a fitted GLMM under the distribution of random effects conditional on the data and given the fixed-effect estimates (using a Gaussian approximation in the GLMM case).
predict.HLfit() could no longer return covariance matrices as a result of a bug introduced in version 2.1.0.
'respVar' computation by predict.HLfit() did not take into account non-identity links, although the documentation suggested it did.
fitme() can now use a fast sparse matrix algorithm for CAR (adjacency) (G)LMMs. See revised help(adjacency) for further information.
confint.HLfit() now computes confidence intervals for GLMs.
spaMM now handles cases where nloptr::nloptr() can be trapped in an infinite loop.
residuals.HLfit() has been redefined to bring it closer to stats::residuals.glm(). Up to now it only returned response residuals, now it returns (unscaled) deviance residuals.
'method' argument of spaMM_glm() redefined to allow more experiments (default behaviour is unchanged).
simulate.HLfit() now also simulates GLMs.
predict(<formula ith random-coefficient term>, <newdata>) is now handled.
binomial(cauchit) is now handled.
simulate() member function added to COMPoisson().
simulate() member function modified in negbin() and in spaMM's version of Gamma() to return matrices when nsim>1, as described in help(family).
simulate.HLfit() now simulates COMPoisson and negbin (mixed) models.
HLCor() no longer fitted adjacency models in Version 2.0.0.
spaMM.options(<new values>) did not return old values.
For GLMs, confint.HLfit() (computed no interval but) did return an empty interval.
COMPoisson$linkfun sometimes returned theta=0, causing failure of fit.
Several problems corrected in degrees of freedom computation for information criteria retruned by get_any_IC().
Several problems corrected in prediction variance calculation with non-null 're.form'.
Core functions have been extensively rewritten to manage expensive matrix computations more efficiently. As a result, version 2.0.0 is 2-3 times faster when correlation matrices of random effects are dense (as typically the case for Matern correlation matrices), and can be much faster when correlation matrices are sparse.
Simple interface for using orthodromic distances: see help(Earth).
bootstrap computations can now use parallelisation; see new argument 'nb_cores' for fixedLRT() and related functions.
The term for uncertainty in dispersion parameters is now included in computation of prediction variance (1) by default from objects returned by fitme() (special default 'control' argument no longer needed); and (2) for models with multiple random effects.
New extractors get_predCov_var_fix() and preprocess_fix_corr() for more efficient computation of specific blocks of prediction covariance matrix.
get_ZALMatrix() gets a new argument 'as_matrix' to control the return value.
confint() now works for HLCor fits.
nloptr() call replaces optim() call in random-slope code, with notable improvements in some toy examples. Random-slope models with >1 predictor variables, such as (x+y|z), are now fitted.
New function stripHLfit() to remove some large objects from a fit object before saving it to file.
Time to fit a model is now recorded in the fit object (as element 'fit_time').
Structure of bootstrap output in return value of fixedLRT() modified.
Structure of ranef() result modified.
Several efforts to reduce the size of fitted objects.
Likelihood-ratio chi-square statistic renamed 'chi2_LR' (was 'LR2') in return value of anova() and some bootstrap outputs.
Raster.Rd removed to reduced package dependencies. A link to a web page is instead provided.
spaMM now provides a slightly less informative diagnostic for (quasi-)separation in binary response models, but no longer 'imports' package lpSolveAPI, instead 'suggesting' e1071.
Tracing stuff, such as 'trace' argument of some functions, has been removed.
Possible values of 'control.corrHLfit$Optimizer' are (as before) 'L-BFGS-B' and (new) 'NLOPT_LN_BOBYQA'. nlminb can no longer be selected.
get_ZALMatrix() is now a standard function rather than part of the object resulting from a fit.
fitme() now fits CAR model parameters using its own optimization procedure rather than through HLfit's procedures.
Message added to warn users about the interpretation of formulas without fixed effects.
More meaningful display of results for random-coefficients models; new argument 'details' for summary.HLfit().
ML fit summaries no longer display a 'p_bv' value, to prevent confusion about what this value means.
Version 1.10.0 failed on R < 3.2.0 because cBind() and rBind() calls from earlier spaMM versions have been replaced by cbind() and rbind() calls (see help(cBind)). The Depends field is now corrected to R (>= 3.2.0).
A few stopping bugs in predict() have been corrected.
LRT() did not work on HLCor fit objects. This is corrected.
spaMM now fits models with random effects in the residual dispersion model, and thus implements a class of “double hierarchical GLMs”.
fitme() now fits models with offset in the dispersion model.
Modified convergence criteria for the iterative algorithm in HLfit().
Examples run generally faster by about 30%.
New extractor function get_intervals() and get_ZALMatrix().
Intervals returned by predict.HLfit() (and hence by get_intervals()) use Student's t distribution when the fitted model is a simple linear model.
Prediction variance can now be computed from fits of models with a corrMatern term.
An HLfit object's $APHLs$p_bv for a *fixed-effect* model now returns the restricted likelihood (for estimation of residual dispersion parameters), which it has never done so far (out of disregard for this case). This allows REML fits of fixed-effect models using fitme() (but has no bearing on similar fits using HLfit()).
More informative summary() when non-default spatial distance function, or non-default prior weights, are used.
New 'refit' control in fitme().
New extractor function get_RLRTSim_args() for use with the RLRsim package to perform small-sample LRT of the presence of a random effect.
Summary() outlook revised, in particular to more accurately display the meaning of phi for Gamma response.
The objects returned by the fitting functions should be of much smaller size than before.
Better reporting of terms that have to be ignored in the computation of prediction variance.
New 'inverse.Gamma', 'spaMM-conventions' and 'member-extractor' documentation pages. HLfit() documentation revised.
Some numerical issues that could affect GLM fitting with HLfit() have been fixed.
More accurate detection of unidentifiable random effect components, when residual variance formulas involve a pure offset term.
confint() on a LMM fitted by fitme() was returning incorrect results. This is corrected.
Computation of prediction variance has been reimplemented so that it requires storing smaller matrices than before. Further, some terms required for this computation are computed only when prediction variance is requested, rather than being always returned by HLfit(). This allows some gains in computation time for large data sets.
simulate.HLfit() gets new arguments and prints some information, in anticipation of future developments.
spaMM.colors() gets new argument 'adjustcolor_args' for finer control.
Filled plot functions gets new argument 'plot.title'.
spaMM_glm(), which has been designed to 'rescue' stats::glm() when the latter fails, did not rescue it for some Gamma(log) GLMs. This is corrected.
Computation of AIC (and other information criteria) failed in many cases. This is corrected.
confint() did not correctly handle models with several random effects. This is corrected.
fitme() does not fit random-slope models, nor models with offset in the dispersion model, but this was not documented. This is corrected.
C++ code has been fixed for compatibility with recent standards.
New fitme() function, intended to become the main fitting function in spaMM, can be substantially faster than HLfit() or corrHLfit(), particularly for large data sets, but can occasionally be slower.
Some other moderate but noticeable improvements in speed.
Users no longer need to explicitly request AIC computation when calling the fit of a model. The new extractor function get_any_IC() (or its more conventionnally named alias, AIC()) can be used to compute them from a fitted model object.
Negative binomial response models can be fitted by 'family=negbin()'.
The syntax 'family=COMPoisson()', with unknown 'nu' parameter, works. This means that COMPoisson() models can now be fitted with unknown values of this parameter, by corrHLfit() or by the new fitme() function. The same is true for the parameter of the 'negbin' family.
Initial or fixed values of the variance of random effects can be controlled independently when there are several random effects.
spaMM_glm() gains a new argument 'strict' for finer control. spaMM_glm.fit() has increased default 'control$maxit'.
filled.mapMM() now returns a predictor of the response surface. It also gains a new argument 'smoothObject' that takes such a predictor, allowing for faster redrawing.
fixedLRT() gets a new argument 'fittingFunction' to select between corrHLfit() and fitme().
Revised usage of 'corrMatrix' argument of HLCor().
'spaMM', as well as 'RcppEigen', had become broken for Windows 32bit. See https://github.com/RcppCore/RcppEigen/issues/34 for details. A suitable compilation flag is now used to overcome this issue.
simulate.HLfit() could fail when a design matrix was stored as a _M_atrix. This is corrected.
predict.HLfit() can return prediction intervals.
New spaMM_glm() and spaMM_glm.fit() functions provide a robust alternative (using the Levenberg-Marquardt algorithm) to glm.fit() when the latter diverges. They also circumvent the initial-value issues of glm().
Slight general increase in speed.
The Conway-Maxwell-Poisson family is implemented.
New extractor functions get_fixefVar(), get_predVar(), get_residVar() and get_respVar().
Subsetting the result of a predict() call also subsets its variance attributes, thanks to a new internal 'Extract' method '[.predictions'.
spaMM is protected from disturbing numerical problems of the Gamma()$dev.resids() function, thanks to an internal redefinition of the Gamma family.
New 'control.glm' argument of HLfit(), to control calls to spaMM_glm() within HLfit().
Once-again revamped 'variances' argument and documentation of predict.HLfit(). The 'variances' argument of filled.mapMM() is replaced by 'variance'.
New argument 'resid.model' replaces 'resid.formula' and 'resid.family' specifications in the main fitting function.
The default behaviour of mapMM() for argument 'map.asp=NULL' has been modified.
The obscure control parameters 'RHOMAX' and 'NUMAX' previously accessible through spaMM.options() are no longer needed and have been removed.
Better control of some plot margins.
Fitting random-slope models may be substantially less slow (but still slow).
Printing bug for non-standard REML formula in summary.HLfit() corrected.
Conditions where predict() return an object of class 'predictions' corrected to avoid a printing bug.
The 'xrange' and 'yrange' arguments of spaMMplot2D() were not effective. This is corrected.
Printing bug in summary.fixedLRT() corrected.
spaMM.options(maxLambda=. . .) did not fully control extreme lambda estimates. The revised code is more effective, and also better bypasses some numerical problems that can occur for such values.
Fitting phi using the corrHLfit() outer optimization mechanism was inefficiently implemented and could return some confusing output. This is corrected.
Less inefficient code for fixed-effect models, faster and handling larger data sets.
Changed default value of 'upper' argument of optim call() within corrHLfit(), and better reporting of issues with this call.
Argument 'which' replaces arguments 'REML' in logLik().
New control 'break_conv_logL' in argument control.HLfit of function HLfit(). Mainly for development purposes.
Improved reporting of convergence issues in HLfit().
More comfortable print method for vcov() value.
This new version includes several improvements of predict(), the most notable being a correction of prediction variances for new values of predictor variables, which were incorrect in CRAN version 1.6.2 (but not in CRAN version 1.5.1). Many other more trivial bugs or features of predict() have been corrected.
summary.fixedLRT() failed to display correctly the bootstrap results. This is corrected.
A bug in C++ code, introduced in version 1.7.1, has been corrected in version 1.7.2; and code applying some Eigen functions on a matrix with a dimension equal to zero (generating an UBSAN runtime error but a correct retrun value in other conditions) has been fixed.
Prediction variance can now account for uncertainty is dispersion parameters (though not for all models; see Details of 'disp' in the 'predict' documentation).
Now allows prediction for a large number of spatial positions by storing some large correlation matrices on disk rather than in memory (using the ff package, and as controlled by the spaMM option 'ff_threshold'). (NB: this was removed in a later version)
Convenience extractor functions 'dev_resids' and 'deviance' have been added.
New data objects 'worldcountries' and 'oceanmask' added for better plots of spatial predictions (see the new documentation '?raster' for plot of spatial predictions using rasters and geographical projections). (PS: removed since, see version 3.1.2)
More efficient implementation of a conditional autoregressive (CAR) model.
're.form' argument of 'predict' function has been extended to handle values of class 'formula'.
etaFix$beta usage has been extensively revised to provide a convenient interface for fixing (some of) the fixed effects (beta coefficients) values.
If anyone on Earth ever tried inverse.Gamma("-1/mu") random effects, s-he might have been confronted with negative estimates of u_h. This problem is inherent to the algorithms described in the HGLM literature, but spaMM now brings a solution, which allows one to fit more securely models with v ~ Gamma( lambda/(lambda+1), lambda+1). A new link identity is also allowed for Gamma random effects [i.e., v=u ~ Gamma(1/lambda,lambda) ], albeit with the restriction that lambda is < 1.
New 'symSVD' argument for designL.from.Corr function.
New keyword 'predict' for 'which' argument of 'plot.HLfit', for easier plotting of prediction diagnostic plot.
'make.scaled.dist' has been renamed 'make_scaled_dist', and gets a new 'return_matrix' argument for programming purposes.
Substantial gains in speed for moderately large data sets (hundreds of lines), in particular by minimizing the use of the QR module from Eigen.
Users no longer have to think about pre-computing prediction coefficients. 'predictionCoeffs' function has been removed. 'coeffs' argument of predict.HLfit() has been removed.
More compact printing of predict() results.
The output format of information criterias computed by HLfit, and related documentation, have been revised.
confint() did not correctly handle PQL fits. This is corrected. In addition, it better handles some non-convergence issues.
Gamma response models with non-unit 'prior.weights' produced an inaccurate correction from EQL to 'HL(1,1)'. This is corrected.
anova() and LRT() declared some nested models as non-nested, and sometimes stopped. This is corrected.
HLfit could stop on some fits involving moderately large sparse design matrices (Z) for random effects. This is corrected.
A printing bug for ranef(), which prevented the display of its result when lambda was fixed, has been corrected
New example_maxtime package option for controlling execution of examples.
Tests are now controlled by a local environment variable _LOCAL_TESTS_ so that they are not run on CRAN.
Faster code for (some) sparse models with a single random effect on large data sets. It is also possible (by using spaMM.options(QRmethod=Matrix::qr)
) to use sparse matrix algorithms in other models, but this will be slow in many applications.
Various other changes were made to speed up computation in dense models (by about 10% in the tests).
The non-Euclidean distance functions implemented in the proxy
package (e.g. Geodesic) can now be used (although only Geodesic was tested, and the Matern correlation function with arbitrary rho and nu is not mathematically feasible for all distance functions). See control.dist
argument of HLCor
and corrHLfit
.
Reorganized graphic functions, with new spaMMplot2D
function and new envir
argument for mapMM.HLfit
.
predict.HLfit
modified for better consistency with other widely used predict methods: added newdata
argument and modified default behaviour of binding
argument.
New getDistMat
extractor for extracting the distance matrix used in the Matérn model
This NEWS.Rd file is created
Correction of a bug that caused incorrect prediction on new data for 1D (only) spatial model.