glm_weightit
,
multinom_weightit
and ordinal_weightit
from
package WeightIt.Added p_significance()
methods for frequentist
models.
Methods for degrees_of_freedom()
have been removed.
degrees_of_freedom()
now calls
insight::get_df()
.
model_parameters()
for data frames and
draws
objects from package posterior also gets an
exponentiate
argument.
equivalence_test()
, which should now be more accurate
related to the proportion of the interval that falls inside the ROPE.
Formerly, the confidence interval was simply treated as uniformly
distributed when calculating the SGPV, now the interval is assumed to be
normally distributed.svy2lme
models from package
svylme.standardize_parameters()
now also prettifies labels of
factors.Fixed issue with equivalence_test()
when ROPE range
was not symmetrically centered around zero (e.g.,
range = c(-99, 0.1)
).
model_parameters()
for anova()
from
mixed models now also includes the denominator degrees of freedom in the
output (df_error
).
print(..., pretty_names = "labels")
for tobit-models
from package AER now include value labels, if
available.
Patch release, to ensure that performance runs with older version of datawizard on Mac OS X with R (old-release).
Deprecated arguments in model_parameters()
for
htest
, aov
and BFBayesFactor
objects were removed.
Argument effectsize_type
is deprecated. Please use
es_type
now. This change was necessary to avoid conflicts
with partial matching of argument names (here:
effects
).
Support for objects from stats::Box.test()
.
Support for glmgee
models from package
glmtoolbox.
Fixed edge case in predict()
for
factor_analysis()
.
Fixed wrong ORCID in DESCRIPTION
.
Fixes issue in compare_parameters()
for models from
package blme.
Fixed conflict in model_parameters()
when both
include_reference = TRUE
and
pretty_names = "labels"
were used. Now, pretty labels are
correctly updated and preserved.
serp
(serp).include_reference
can now directly be set to
TRUE
in model_parameters()
and doesn’t require
a call to print()
anymore.
compare_parameters()
gains a
include_reference
argument, to add the reference category
of categorical predictors to the parameters table.
print_md()
for compare_parameters()
now
by default uses the tinytable package to create markdown
tables. This allows better control for column heading spanning over
multiple columns.
Fixed issue with parameter names for
model_parameters()
and objects from package
epiR.
Fixed issue with exponentiate = TRUE
for
model_parameters()
with models of class clmm
(package ordinal), when model had no component
column (e.g., no scale or location parameters were returned).
include_reference
now also works when factor were
created “on-the-fly” inside the model formula
(i.e. y ~ as.factor(x)
).
exponentiate
argument of
model_parameters()
for
marginaleffects::predictions()
now defaults to
FALSE
, in line with all the other
model_parameters()
methods.model_parameters()
for models of package
survey now gives informative messages when
bootstrap = TRUE
(which is currently not
supported).
n_factors()
now also returns the explained variance
for the number of factors as attributes.
model_parameters()
for objects of package
metafor now warns when unsupported arguments (like
vcov
) are used.
Improved documentation for
pool_parameters()
.
print(include_reference = TRUE)
for
model_parameters()
did not work when run inside a
pipe-chain.
Fixed issues with format()
for objects returned by
compare_parameters()
that included mixed models.
principal_components()
and
factor_analysis()
now also work when argument
n = 1
.
print_md()
for compare_parameters()
now
gains more arguments, similar to the print()
method.
bootstrap_parameters()
and
model_parameters()
now accept bootstrapped samples returned
by bootstrap_model()
.
The print()
method for
model_parameters()
now also yields a warning for models
with logit-links when possible issues with (quasi) complete separation
occur.
Fixed issue in print_html()
for objects from package
ggeffects.
Fixed issues for nnet::multinom()
with wide-format
response variables (using cbind()
).
Minor fixes for print_html()
method for
model_parameters()
.
Robust standard errors (argument vcov
) now works for
plm
models.
Minor improvements to factor analysis functions.
The ci_digits
argument of the print()
method for model_parameters()
now defaults to the same
value of digits
.
model_parameters()
for objects from package
marginaleffects now also accepts the exponentiate
argument.
The print()
, print_html()
,
print_md()
and format()
methods for
model_parameters()
get an include_reference
argument, to add the reference category of categorical predictors to the
parameters table.
Fixed issue with wrong calculation of test-statistic and p-values
in model_parameters()
for fixest
models.
Fixed issue with wrong column header for glm
models
with family = binomial("identiy")
.
Minor fixes for dominance_analysis()
.
nestedLogit
(nestedLogit).model_parameters()
now also prints correct “pretty
names” when predictors where converted to ordered factors inside
formulas, e.g. y ~ as.ordered(x)
.
model_parameters()
now prints a message when the
vcov
argument is provided and ci_method
is
explicitly set to "profile"
. Else, when vcov
is not NULL
and ci_method
is
NULL
, it defaults to "wald"
, to return
confidence intervals based on robust standard errors.
lme
.mipo
for models
with ordinal or categorical outcome.Added support for models of class hglm
(hglm), mblogit
(mclogit),
fixest_multi
(fixest), and phylolm
/
phyloglm
(phylolm).
as.data.frame
methods for extracting posterior draws
via bootstrap_model()
have been retired. Instead, directly
using bootstrap_model()
is recommended.
equivalence_test()
gets a method for
ggeffects
objects from package ggeffects.
equivalence_test()
now prints the SGPV
column instead of % in ROPE
. This is because the former
% in ROPE
actually was equivalent to the second generation
p-value (SGPV) and refers to the proportion of the range of the
confidence interval that is covered by the ROPE. However,
% in ROPE
did not refer to the probability mass of the
underlying distribution of a confidence interval that was covered by the
ROPE, hence the old column name was a bit misleading.
Fixed issue in model_parameters.ggeffects()
to
address forthcoming changes in the ggeffects package.
When an invalid or not supported value for the
p_adjust
argument in model_parameters()
is
provided, the valid options were not shown in correct capital letters,
where appropriate.
Fixed bug in cluster_analysis()
for
include_factors = TRUE
.
Fixed warning in model_parameters()
and
ci()
for models from package glmmTMB when
ci_method
was either "profile"
or
"uniroot"
.
Reduce unnecessary warnings.
The deprecated argument df_method
in
model_parameters()
was removed.
Output from model_parameters()
for objects returned
by manova()
and car::Manova()
is now more
consistent.
Fixed issues in tests for mmrm
models.
Fixed issue in bootstrap_model()
for models of class
glmmTMB
with dispersion parameters.
Fixed failing examples.
flic
and
flac
(logistf), mmrm
(mmrm).model_parameters()
now includes a Group
column for stanreg
or brmsfit
models with
random effects.
The print()
method for
model_parameters()
now uses the same pattern to print
random effect variances for Bayesian models as for frequentist
models.
Fixed issue with the print()
method for
compare_parameters()
, which duplicated random effects
parameters rows in some edge cases.
Fixed issue with the print()
method for
compare_parameters()
, which didn’t work properly when
ci=NULL
.
The deprecated argument df_method
in
model_parameters()
is now defunct and throws an error when
used.
The deprecated functions ci_robust()
,
p_robust()
and standard_error_robust
have been
removed. These were superseded by the vcov
argument in
ci()
, p_value()
, and
standard_error()
, respectively.
The style
argument in
compare_parameters()
was renamed into
select
.
p_function()
, to print and plot p-values and
compatibility (confidence) intervals for statistical models, at
different levels. This allows to see which estimates are most compatible
with the model at various compatibility levels.
p_calibrate()
, to compute calibrated
p-values.
model_parameters()
and
compare_parameters()
now use the unicode character for the
multiplication-sign as interaction mark (i.e. \u00d7
). Use
options(parameters_interaction = <value>)
or the
argument interaction_mark
to use a different character as
interaction mark.
The select
argument in
compare_parameters()
, which is used to control the table
column elements, now supports an experimental glue-like syntax. See this
vignette Printing Model Parameters. Furthermore, the
select
argument can also be used in the
print()
method for
model_parameters()
.
print_html()
gets a font_size
and
line_padding
argument to tweak the appearance of HTML
tables. Furthermore, arguments select
and
column_labels
are new, to customize the column layout of
tables. See examples in ?display
.
Consolidation of vignettes on standardization of model parameters.
Minor speed improvements.
model_parameters().BFBayesFactor
no longer drops the
BF
column if the Bayes factor is NA
.
The print()
and display()
methods for
model_parameters()
from Bayesian models now pass the
...
to insight::format_table()
, allowing extra
arguments to be recognized.
Fixed footer message regarding the approximation method for CU and p-values for mixed models.
Fixed issues in the print()
method for
compare_parameters()
with mixed models, when some models
contained within-between components (see wb_component
) and
others did not.
Arguments that calculate effectsize in
model_parameters()
for htest
, Anova objects
and objects of class BFBayesFactor
were revised. Instead of
single arguments for the different effectsizes, there is now one
argument, effectsize_type
. The reason behind this change is
that meanwhile many new type of effectsizes have been added to the
effectsize package, and the generic argument allows to make use
of those effect sizes.
The attribute name in PCA / EFA has been changed from
data_set
to dataset
.
The minimum needed R version has been bumped to
3.6
.
Removed deprecated argument parameters
from
model_parameters()
.
standard_error_robust()
, ci_robust()
and p_value_robust()
are now deprecated and superseded by
the vcov
and vcov_args
arguments in the
related methods standard_error()
, ci()
and
p_value()
, respectively.
Following functions were moved from package parameters
to performance: check_sphericity_bartlett()
,
check_kmo()
, check_factorstructure()
and
check_clusterstructure()
.
Added sparse
option to
principal_components()
for sparse PCA.
The pretty_names
argument from the
print()
method can now also be "labels"
, which
will then use variable and value labels (if data is labelled) as pretty
names. If no labels were found, default pretty names are used.
bootstrap_model()
for models of class
glmmTMB
and merMod
gains a
cluster
argument to specify optional clusters when the
parallel
option is set to "snow"
.
P-value adjustment (argument p_adjust
in
model_parameters()
) is now performed after potential
parameters were removed (using keep
or drop
),
so adjusted p-values is only applied to the parameters of
interest.
Robust standard errors are now supported for fixest
models with the vcov
argument.
print()
for model_parameters()
gains a
footer
argument, which can be used to suppress the footer
in the output. Further more, if footer = ""
or
footer = FALSE
in print_md()
, no footer is
printed.
simulate_model()
and
simulate_parameters()
now pass ...
to
insight::get_varcov()
, to allow simulated draws to be based
on heteroscedasticity consistent variance covariance matrices.
The print()
method for
compare_parameters()
was improved for models with multiple
components (e.g., mixed models with fixed and random effects, or models
with count- and zero-inflation parts). For these models,
compare_parameters(effects = "all", component = "all")
prints more nicely.
dominance_analysis()
, to compute dominance
analysis statistics and designations.ci_random
in model_parameters()
defaults to NULL
. It uses a heuristic to determine if
random effects confidence intervals are likely to take a long time to
compute, and automatically includes or excludes those confidence
intervals. Set ci_random
to TRUE
or
FALSE
to explicitly calculate or omit confidence intervals
for random effects.Fix issues in pool_parameters()
for certain models
with special components (like MASS::polr()
), that failed
when argument component
was set to
"conditional"
(the default).
Fix issues in model_parameters()
for multiple
imputation models from package Hmisc.
It is now possible to hide messages about CI method below tables
by specifying options("parameters_cimethod" = FALSE)
(#722). By default, these messages are displayed.
model_parameters()
now supports objects from package
marginaleffects and objects returned by
car::linearHypothesis()
.
Added predict()
method to cluster_meta
objects.
Reorganization of docs for
model_parameters()
.
model_parameters()
now also includes standard errors
and confidence intervals for slope-slope-correlations of random effects
variances.
model_parameters()
for mixed models gains a
ci_random
argument, to toggle whether confidence intervals
for random effects parameters should also be computed. Set to
FALSE
if calculation of confidence intervals for random
effects parameters takes too long.
ci()
for glmmTMB models with
method = "profile"
is now more robust.
Fixed issue with glmmTMB models when calculating confidence intervals for random effects failed due to singular fits.
display()
now correctly includes custom text and
additional information in the footer (#722).
Fixed issue with argument column_names
in
compare_parameters()
when strings contained characters that
needed to be escaped for regular expressions.
Fixed issues with unknown arguments in
model_parameters()
for lavaan models when
standardize = TRUE
.
model_parameters()
now no longer treats data frame
inputs as posterior samples. Rather, for data frames, now
NULL
is returned. If you want to treat a data frame as
posterior samples, set the new argument
as_draws = TRUE
.sort_parameters()
to sort model parameters by
coefficient values.
standardize_parameters()
,
standardize_info()
and
standardise_posteriors()
to standardize model
parameters.
model_parameters()
model_parameters()
for mixed models from package
lme4 now also reports confidence intervals for random effect
variances by default. Formerly, CIs were only included when
ci_method
was "profile"
or
"boot"
. The merDeriv package is required for this
feature.
model_parameters()
for htest
objects
now also supports models from var.test()
.
Improved support for anova.rms
models in
model_parameters()
.
model_parameters()
now supports draws
objects from package posterior and deltaMethods
objects from package car.
model_parameters()
now checks arguments and informs
the user if specific given arguments are not supported for that model
class (e.g., "vcov"
is currently not supported for models
of class glmmTMB).
The vcov
argument, used for computing robust
standard errors, did not calculate the correct p-values and confidence
intervals for models of class lme
.
pool_parameters()
did not save all relevant model
information as attributes.
model_parameters()
for models from package
glmmTMB did not work when exponentiate = TRUE
and
model contained a dispersion parameter that was different than sigma.
Furthermore, exponentiating falsely exponentiated the dispersion
parameter.
Added options to set defaults for different arguments. Currently supported:
options("parameters_summary" = TRUE/FALSE)
, which sets
the default value for the summary
argument in
model_parameters()
for non-mixed models.options("parameters_mixed_summary" = TRUE/FALSE)
, which
sets the default value for the summary
argument in
model_parameters()
for mixed models.Minor improvements for print()
methods.
Robust uncertainty estimates:
vcov_estimation
, vcov_type
, and
robust
arguments are deprecated in these functions:
model_parameters()
, parameters()
,
standard_error()
, p_value()
, and
ci()
. They are replaced by the vcov
and
vcov_args
arguments.standard_error_robust()
and
p_value_robust()
functions are superseded by the
vcov
and vcov_args
arguments of the
standard_error()
and p_value()
functions.Fixed minor issues and edge cases in n_clusters()
and related cluster functions.
Fixed issue in p_value()
that returned wrong
p-values for fixest::feols()
.
Improved speed performance for model_parameters()
,
in particular for glm’s and mixed models where random effect variances
were calculated.
Added more options for printing model_parameters()
.
See also revised vignette:
https://easystats.github.io/parameters/articles/model_parameters_print.html
model_parameters()
model_parameters()
for mixed models gains an
include_sigma
argument. If TRUE
, adds the
residual variance, computed from the random effects variances, as an
attribute to the returned data frame. Including sigma was the default
behaviour, but now defaults to FALSE
and is only included
when include_sigma = TRUE
, because the calculation was very
time consuming.
model_parameters()
for merMod
models
now also computes CIs for the random SD parameters when
ci_method="boot"
(previously, this was only possible when
ci_method
was "profile"
).
model_parameters()
for glmmTMB
models
now computes CIs for the random SD parameters. Note that these are based
on a Wald-z-distribution.
Similar to model_parameters.htest()
, the
model_parameters.BFBayesFactor()
method gains
cohens_d
and cramers_v
arguments to control if
you need to add frequentist effect size estimates to the returned
summary data frame. Previously, this was done by default.
Column name for coefficients from emmeans objects are now more specific.
model_prameters()
for MixMod
objects
(package GLMMadaptive) gains a robust
argument, to
compute robust standard errors.
Fixed bug with ci()
for class merMod
when method="boot"
.
Fixed issue with correct association of components for ordinal
models of classes clm
and clm2
.
Fixed issues in random_parameters()
and
model_parameters()
for mixed models without random
intercept.
Confidence intervals for random parameters in
model_parameters()
failed for (some?) glmer
models.
Fix issue with default ci_type
in
compare_parameters()
for Bayesian models.
Following functions were moved to the new datawizard package and are now re-exported from parameters package:
center()
convert_data_to_numeric()
data_partition()
demean()
(and its aliases degroup()
and
detrend()
)
kurtosis()
rescale_weights()
skewness()
smoothness()
Note that these functions will be removed in the next release of parameters package and they are currently being re-exported only as a convenience for the package developers. This release should provide them with time to make the necessary changes before this breaking change is implemented.
Following functions were moved to the performance package:
check_heterogeneity()
check_multimodal()
The handling to approximate the degrees of freedom in
model_parameters()
, ci()
and
p_value()
was revised and should now be more consistent.
Some bugs related to the previous computation of confidence intervals
and p-values have been fixed. Now it is possible to change the method to
approximate degrees of freedom for CIs and p-values using the
ci_method
, resp. method
argument. This change
has been documented in detail in ?model_parameters
, and
online here:
https://easystats.github.io/parameters/reference/model_parameters.html
Minor changes to print()
for glmmTMB with
dispersion parameter.
Added vignette on printing options for model parameters.
model_parameters()
The df_method
argument in
model_parameters()
is deprecated. Please use
ci_method
now.
model_parameters()
with
standardize = "refit"
now returns random effects from the
standardized model.
model_parameters()
and ci()
for
lmerMod
models gain a "residuals"
option for
the ci_method
(resp. method
) argument, to
explicitly calculate confidence intervals based on the residual degrees
of freedom, when present.
model_parameters()
supports following new objects:
trimcibt
, wmcpAKP
, dep.effect
(in
WRS2 package), systemfit
model_parameters()
gains a new argument
table_wide
for ANOVA tables. This can be helpful for users
who may wish to report ANOVA table in wide format (i.e., with numerator
and denominator degrees of freedom on the same row).
model_parameters()
gains two new arguments,
keep
and drop
. keep
is the new
names for the former parameters
argument and can be used to
filter parameters. While keep
selects those parameters
whose names match the regular expression pattern defined in
keep
, drop
is the counterpart and excludes
matching parameter names.
When model_parameters()
is called with
verbose = TRUE
, and ci_method
is not the
default value, the printed output includes a message indicating which
approximation-method for degrees of freedom was used.
model_parameters()
for mixed models with
ci_method = "profile
computes (profiled) confidence
intervals for both fixed and random effects. Thus,
ci_method = "profile
allows to add confidence intervals to
the random effect variances.
model_parameters()
should longer fail for supported
model classes when robust standard errors are not available.
n_factors()
the methods based on fit indices have
been fixed and can be included separately
(package = "fit"
). Also added a n_max
argument
to crop the output.
compare_parameters()
now also accepts a list of
model objects.
describe_distribution()
gets verbose
argument to toggle warnings and messages.
format_parameters()
removes dots and underscores
from parameter names, to make these more “human readable”.
The experimental calculation of p-values in
equivalence_test()
was replaced by a proper calculation
p-values. The argument p_value
was removed and p-values are
now always included.
Minor improvements to print()
,
print_html()
and print_md()
.
The random effects returned by model_parameters()
mistakenly displayed the residuals standard deviation as square-root of
the residual SD.
Fixed issue with model_parameters()
for
brmsfit objects that model standard errors (i.e. for
meta-analysis).
Fixed issue in model_parameters
for
lmerMod
models that, by default, returned residual degrees
of freedom in the statistic column, but confidence intervals were based
on Inf
degrees of freedom instead.
Fixed issue in ci_satterthwaite()
, which used
Inf
degrees of freedom instead of the Satterthwaite
approximation.
Fixed issue in model_parameters.mlm()
when model
contained interaction terms.
Fixed issue in model_parameters.rma()
when model
contained interaction terms.
Fixed sign error for model_parameters.htest()
for
objects created with t.test.formula()
(issue #552)
Fixed issue when computing random effect variances in
model_parameters()
for mixed models with categorical random
slopes.
check_sphericity()
has been renamed into
check_sphericity_bartlett()
.
Removed deprecated arguments.
model_parameters()
for bootstrapped samples used in
emmeans now treats the bootstrap samples as samples from
posterior distributions (Bayesian models).
SemiParBIV
(GJRM), selection
(sampleSelection), htest
from the survey
package, pgmm
(plm).summary()
method for
model_parameters()
, which is a convenient shortcut for
print(..., select = "minimal")
.model_parameters()
model_parameters()
gains a parameters
argument, which takes a regular expression as string, to select specific
parameters from the returned data frame.
print()
for model_parameters()
and
compare_parameters()
gains a groups
argument,
to group parameters in the output. Furthermore, groups
can
be used directly as argument in model_parameters()
and
compare_parameters()
and will be passed to the
print()
method.
model_parameters()
for ANOVAs now saves the type as
attribute and prints this information as footer in the output as
well.
model_parameters()
for htest-objects now
saves the alternative hypothesis as attribute and prints this
information as footer in the output as well.
model_parameters()
passes arguments
type
, parallel
and n_cpus
down to
bootstrap_model()
when
bootstrap = TRUE
.
bootstrap_models()
for merMod and
glmmTMB objects gains further arguments to set the type of
bootstrapping and to allow parallel computing.
bootstrap_parameters()
gains the
ci_method
type "bci"
, to compute
bias-corrected and accelerated bootstrapped intervals.
ci()
for svyglm
gains a
method
argument.
Fixed issue in model_parameters()
for
emmGrid objects with Bayesian models.
Arguments digits
, ci_digits
and
p_digits
were ignored for print()
and only
worked when used in the call to model_parameters()
directly.
print()
method for
model_parameters()
.blrm
(rmsb), AKP
,
med1way
, robtab
(WRS2),
epi.2by2
(epiR), mjoint
(joineRML), mhurdle
(mhurdle),
sarlm
(spatialreg), model_fit
(tidymodels), BGGM
(BGGM),
mvord
(mvord)model_parameters()
model_parameters()
for blavaan
models
is now fully treated as Bayesian model and thus relies on the functions
from bayestestR (i.e. ROPE, Rhat or ESS are reported)
.
The effects
-argument from
model_parameters()
for mixed models was revised and now
shows the random effects variances by default (same functionality as
random_parameters()
, but mimicking the behaviour from
broom.mixed::tidy()
). When the group_level
argument is set to TRUE
, the conditional modes (BLUPs) of
the random effects are shown.
model_parameters()
for mixed models now returns an
Effects
column even when there is just one type of
“effects”, to mimic the behaviour from broom.mixed::tidy()
.
In conjunction with standardize_names()
users can get the
same column names as in tidy()
for
model_parameters()
objects.
model_parameters()
for t-tests now uses the group
values as column names.
print()
for model_parameters()
gains a
zap_small
argument, to avoid scientific notation for very
small numbers. Instead, zap_small
forces to round to the
specified number of digits.
To be internally consistent, the degrees of freedom column for
lqm(m)
and cgam(m)
objects (with
t-statistic) is called df_error
.
model_parameters()
gains a summary
argument to add summary information about the model to printed
outputs.
Minor improvements for models from quantreg.
model_parameters
supports rank-biserial, rank
epsilon-squared, and Kendall’s W as effect size measures for
wilcox.test()
, kruskal.test
, and
friedman.test
, respectively.
describe_distribution()
gets a quartiles
argument to include 25th and 75th quartiles of a variable.Fixed issue with non-initialized argument style
in
display()
for compare_parameters()
.
Make print()
for compare_parameters()
work with objects that have “simple” column names for confidence
intervals with missing CI-level (i.e. when column is named
"CI"
instead of, say, "95% CI"
).
Fixed issue with p_adjust
in
model_parameters()
, which did not work for
adjustment-methods "BY"
and "BH"
.
Fixed issue with show_sigma
in print()
for model_parameters()
.
Fixed issue in model_parameters()
with incorrect
order of degrees of freedom.
Roll-back R dependency to R >= 3.4.
Bootstrapped estimates (from bootstrap_model()
or
bootstrap_parameters()
) can be passed to
emmeans
to obtain bootstrapped estimates, contrasts, simple
slopes (etc) and their CIs.
model_parameters()
and
related functions to obtain standard errors, p-values, etc.model_parameters()
now always returns the confidence
level for as additional CI
column.
The rule
argument in equivalenct_test()
defaults to "classic"
.
crr
(cmprsk), leveneTest()
(car), varest
(vars), ergm
(ergm), btergm
(btergm),
Rchoice
(Rchoice), garch
(tseries)compare_parameters()
(and its alias
compare_models()
) to show / print parameters of multiple
models in one table.Estimation of bootstrapped p-values has been re-written to be more accurate.
model_parameters()
for mixed models gains an
effects
-argument, to return fixed, random or both fixed and
random effects parameters.
Revised printing for model_parameters()
for
metafor models.
model_parameters()
for metafor models now
recognized confidence levels specified in the function call (via
argument level
).
Improved support for effect sizes in
model_parameters()
from anova objects.
Fixed edge case when formatting parameters from polynomial terms with many degrees.
Fixed issue with random sampling and dropped factor levels in
bootstrap_model()
.