getX() threw an error when trying to construct model
design matrices using factors with less than two levels (where contrasts
can’t be applied). Such an error might occur where e.g. the dataset
called by a model contained at least one character/factor variable with
only one unique value, which was incorrectly coerced to a factor by
getX() (typically in order to gain access to individual
dummy variables not in the original design matrix).
bootEff() failed to name unnamed model lists where
the names attribute was a character vector of empty strings (rather than
NULL). Such names are apparently generated by
piecewiseSEM::psem().
12/10/2021
semEff() and bootCI(), accessed via new
print() and summary() methods.predEff()."psem" objects
(piecewiseSEM::psem()) in
bootEff()/semEff()."semEff" object (for reference; does
not include bootstrapped effects). Extract using
getAllInd().bootEff()/semEff() (names are generated
automatically).R2() (control of negative values, new
improved default method for adjusted R-squared – Olkin-Pratt exact
estimator).getX(), for more flexible construction of
model design matrices (mostly for internal use).… argument of stdEff() – arguments
to R2() are now passed as named list to
R2.arg.dirEff()
-> getDirEff()) and added some new ones.xNam() was not evaluating factor/character terms correctly.
The function now explicitly treats all non-numeric predictor variables
as factors and coerces where necessary. It also has improved handling of
factor contrasts when evaluating names.pSapply() did not work with
parallel = "multicore", due to relying completely on
parallel::parSapply() for parallel processing (which is
"snow" only). The function now wraps
parallel::mcmapply() for "multicore" (not
available on Windows systems).09/04/2021
R2() can now calculate R-squared based on Spearman’s
Rho.RVIF(), to calculate ‘root variance
inflation factors’ (square root of VIFs).unique.x argument of stdEff() to
unique.eff (old name temporarily allowed).R argument of
bootEff(). The number of bootstrap resamples must now be
explicitly specified, which is probably better practice (10,000 is often
recommended for confidence intervals).type argument to bootEff(), to
specify the type of bootstrapping to perform (for mixed models). This
replaces ran.eff = "crossed", previously used to indicate
parametric bootstrapping (although it’s temporarily allowed).bootEff() will now treat a list containing both mixed
and non-mixed models as all mixed (with a warning). Previously all such
models were treated as non-mixed (unintentionally). This is presumably a
relatively rare scenario.R2() did not calculate adjusted R-squared correctly for
beta regression models (i.e. did not incorporate the ‘phi’ parameter in
degrees of freedom calculations).xNam() produced an error when attempting to evaluate
factor contrasts in data, expecting that character vectors were factors
(related to the change to stringsAsFactors = FALSE as
default in R 4.0.0, but would have occurred in some cases
regardless).01/10/2020
incl.raw argument to
stdEff() (stdCoeff()), to append raw effects
(unstandardised coefficients) to the output. This facilitates
simultaneous bootstrapping of both sets of effects, allowing raw effects
to be used alternatively for calculating
(semEff(..., use.raw = TRUE)) or predicting
(predEff(..., use.raw = TRUE)) effects/CIs.stdCoeff() to stdEff(),
to better reflect the concept of standardised model coefficients as
‘effects’ (calling stdCoeff() will still work – with a
warning – until the next version at least).offset argument to getY() and
R2(), to explicitly retain/remove an offset (where present)
in/from the response variable or fitted values. Offsets are removed by
default, which ensures, for example, that standardised effects are
scaled appropriately.env argument to multiple functions, for
explicitly specifying the location of data used to fit models (not
necessary in most circumstances). This replaces the ...
argument in many instances, which was previously used to pass an
environment to eval() (via getData()).
env (and data) can also now be passed
(...) to bootEff() and
predEff().bootCI()/semEff() output (i.e. confidence
level, type).R2() no longer calculates predictive R-squared for
GLMMs, as the interpretation of the hat matrix used in calculations is
not reliable (see https://rdrr.io/cran/lme4/man/hatvalues.merMod.html).getY() to
glt(), allowing more controlled output of
getY(..., link = TRUE).bootEff() specified with correlated errors failed for
mixed models of class "lmerModLmerTest" (issue with
re-fitting models using update()).predEff() failed to evaluate some complex model terms
(e.g. polynomials).stdEff() (stdCoeff()) did not re-fit model
properly to calculate correct VIFs for a fully ‘centred’ model (i.e. did
not account sufficiently for complex terms such as polynomials or
transformations, where mean-centring should occur as the final
step).xNam() generated incorrect term names for categorical
predictors under certain circumstances (different contrast types,
interactive effects with no ‘main’ effects).stdEff() (stdCoeff()) incorrectly
calculated ‘centred’ intercept for models with an offset specified.predEff() failed when a nested list of models and list
of numeric weights were supplied (i.e. a model averaging scenario).stdEff() (stdCoeff()) did not return the
‘phi’ parameter(s) for beta regression models.25/03/2020
"lmerModLmerTest".glt(), for calculating ‘generalised’ link
transformations for non-gaussian variables.getY() to
glt().bootEff() and
getY().stdCoeff() to use variables not
present in the model design matrix (e.g. a ‘missing’ main effect for an
interaction).bootEff()) to
the effects argument of predEff().refit.x argument to stdCoeff(),
allowing control over whether to refit the model with centred predictors
(for correct VIFs).xNam() did not generate correct term names for
categorical variables with contrast types other than
contr.treatment().stdCoeff() did not correctly adjust for
multicollinearity for a model containing categorical variables when
centring was specified (cen.x = TRUE).getY() failed to generate an estimated working response
when a variable with missing values (NA) was supplied (this
functionality now in glt()).predEff() failed for models with categorical variables
(did not access dummy variables in model matrix).15/01/2020
semEff() did not output effects properly.08/01/2020
"gls")."betareg").xNam() did not generate correct term names for
interactions involving multi-coefficient terms (e.g. factors).xNam() did not generate correct term names for
factors when the model intercept is suppressed.R2() with argument pred = TRUE
threw an error for models where any weights = 0.04/11/2019
New package semEff, allowing the automatic calculation
of effects for ‘piecewise’ structural equation models.