augment_longer(), partial_residuals(),
and binned_residuals() now support a label
argument to give human-readable labels to predictors, rather than using
the original variable names.
The documentation for empirical_link() is
improved.
ols_with_error() gives a clearer error message when
the provided error function fails to generate errors, such as if the
wrong arguments are provided.
partial_residuals() and
binned_residuals() now correctly handle models with an
offset() term in their formula. Previously this would be
treated as another predictor and residuals computed for it, which
doesn’t make sense. Instead, treat this the same as a value provided in
the offset = argument to the model fit.
augment_longer() and
augment_quantile_longer() now give one row per
predictor per observation, rather than one row per
regressor. Their documentation is also corrected to be
consistent. This change facilitates plotting residuals against each
predictor, which is more interpretable than having plots of residuals
against multiple regressors derived from the same predictor (e.g.,
against a variable and its square).
The documentation for partial_residuals() now gives
the correct definition of partial residuals for GLMs. (The code gave the
correct output already, but the stated definition was
incorrect.)
parametric_boot_distribution(), and by extension
model_lineup(), now supports models fit with
transformations. Previously, using transformations like
I(x^2) in model formulas (or using ns() or
other functions that transform predictors into matrix columns) would
cause the model refitting to fail.
partial_residuals() now works correctly for models
fit with offsets, by fixing the offset to its mean value.
bin_by_interval() and bin_by_quantile()
now base their default number of bins on the number of
non-NA observations of the variable to be binned, rather
than the total number of rows.
New augment_quantile() and
augment_quantile_longer() functions augment model training
data with randomized quantile residuals.
If a predictor() is multivariate (it generates a
matrix of multiple columns when sampled from), sample_x()
will now use the column names to set the predictor names appropriately,
rather than simply numbering them.
It is now recommended to provide a distribution function to
predictor(), rather than a function name as a string. (For
example, use predictor(rnorm, mean = 1) rather than
predictor("rnorm", mean = 1).) Using a string introduces
problems when the distribution function is defined in an environment not
on the search path when sample_x() needs to find it; by
passing a function directly, sample_x() will always be able
to find it.
partial_residuals() and
binned_residuals() now reject tidyselect syntax that tries
to rename predictors, or that results in no predictors being selected.
This syntax already caused the functions to fail with strange error
messages, so it is now more explicitly rejected.
This version fixes several bugs that arose during classroom use.
Simulation functions (model_lineup(),
parametric_boot_distribution(), and
sampling_distribution()) now check to determine if the
model being simulated from was fit using the data =
argument, and issue an error if it was not. The simulations work by
calling update(fit, data = ...) with newly simulated data,
and update() uses this to call the model fit function again
with the specified data = argument. But if the model was
fit without one, the argument is unused, and the simulations just reuse
the original data.
For example, if you fit this model:
bad_fit <- lm(cars$dist ~ cars$speed)the simulation functions cannot work correctly because even with a
different data = argument, the model fit will still refer
to cars. The model should be fit like this:
good_fit <- lm(dist ~ speed, data = cars)To prevent simulation problems, a suitable error is issued, so the user can refit the model correctly.
response() now correctly detects when the
error_scale argument was missing and issues the appropriate
error.
augment_longer() now supports models with factor
predictors. If there are some factors and some continuous predictors,
the factors are omitted from the result; if the predictors are all
factors, they are kept.
parametric_boot_distribution() now supports
simulations when alternative_fit uses predictors that were
not used in fit. Previously, these would fail because the
simulated data frame only contained the predictors used in
fit. Supply the new data argument to specify
the data frame used in simulations.
First version released to CRAN.