bin_by_quantile() and bin_by_interval() now give a more helpful error
message when an invalid column name is specified.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.