ols_plot_resid_regressor(),
ols_plot_resid_hist(), ols_plot_diagnostics()
and ols_plot_comp_plus_resid() take the residuals and
fitted values of the rows the fit used, so models fitted with
na.action = na.exclude plot the used observations instead
of failing with a differing-row-count error or pairing padded residuals
with used rows (M1, L1).
ols_step_all_possible() and
ols_step_best_subset() report the number of estimable terms
rather than the number of candidate columns, so an exactly collinear
predictor is not counted in n; the fit statistics are
unchanged (L4).
ols_vif_tol() warns that VIFs may not be sensible
for a model without an intercept, as car::vif() does (also
through ols_coll_diag(), which calls it); the tolerance and
VIF values are unchanged (L3).
ols_test_breusch_pagan() now computes Holm adjusted
p values with stats::p.adjust(p, "holm"). Earlier versions
sorted the raw p values, multiplied by a reversed rank and mapped them
back, which dropped the cumulative maximum: the smallest raw p was
returned unchanged, so the first predictor could be rejected at .05 when
its Holm p was above the cutoff. The simultaneous test is unchanged;
Bonferroni and Sidak were already correct (B1).
ols_plot_comp_plus_resid() keeps every non-intercept
column of the model matrix when it is built, so single-predictor and
no-intercept models return one plot per predictor with one point per
observation. Earlier versions dropped the only column
([, -1]) for a single predictor and the intercept handling
silently reduced no-intercept models to one column, giving a one-row
plot and losing predictors (B2).
ols_plot_resid_hist() draws the histogram with the
breaks used for the normal curve, so the curve and the bars share one
bin width. Earlier versions scaled the curve to the hist()
bin width while geom_histogram(bins = 6) drew
different-width bars, so the fitted normal density was drawn at the
wrong height (0.54x or 1.03x) (B3).
ols_plot_resid_regressor() and
rvsr_plot_shiny() align the regressor with the rows the fit
used, so a model fitted with an omitted NA row plots
instead of failing with a differing-row-count error (B4).
ols_plot_diagnostics(print_plot = FALSE) no longer
draws the ten sub-plots before returning them; each sub-plot is built
with print_plot = FALSE (B5).
ols_prep_rvsrplot_data() reports every non-intercept
model-matrix column, so no-intercept and weighted models lose no
predictor names and a single predictor returns a one-column data frame
instead of a bare vector (B6).
ols_test_outlier() caps the Bonferroni p values at
1, so they remain probabilities (earlier versions could report values
above 1 for large samples) (B7).
The grob-returning plots (ols_plot_added_variable(),
ols_plot_comp_plus_resid(), ols_plot_dfbetas()
and ols_plot_diagnostics()) draw the grob and return it
invisibly when print_plot = TRUE, so they print in scripts,
functions and loops instead of only at top level (#161)
(B8).
ols_correlations() reports that models without an
intercept are not supported with “more than one predictor coefficient”,
matching the guard that counts coefficients rather than term labels (#232)
(B9).
Forward selection on a saturated model (n <= p,
residual degrees of freedom 0) now starts from the null model and
records the predictors selected before the next candidate saturates,
instead of returning an empty result and losing the selection (#199)
(B10).
Stepwise selection plots no longer cut off the highest value when the plotted metric (e.g. AIC) has both negative and positive values (F13).
Stepwise offset identity checks compare exact numeric values, allowing labelled or attributed responses without weakening the rejection of a different fit (P02).
Stepwise selectors choose one term per step. Selection rankings and criterion improvement tests use 10 significant digits: differences below that precision are treated as ties, equal criterion keys mean no improvement, and any remaining tie goes to the first candidate in formula order, after F ties are broken by p value (smaller for entry, larger for removal). P-value entry/removal cutoffs are compared without rounding. Histories identify the term actually removed with forced predictors and report SBIC against the original full model (P03, M1-M4, L1-L2).
P-value selection ignores candidates with missing F tests and ends cleanly when none remain; an initial search with no eligible candidate gives the existing entry-criteria error (P04).
Stepwise selectors return offset models fitted from the original call, preserving offset expressions for correct predictions on new data and a compact printed call (P01). If that fit cannot exactly reproduce the frozen-offset selection fit, selection stops with an error.
ols_regress() and selection refits reject custom
named contrast functions; fit with an explicit contrast matrix instead.
Base contrast names are resolved from the stats namespace and refits use
matrices, so changes to search-path functions or contrast options cannot
change the fitted coding.
Standardized-beta and selection refits carry the fit’s frozen total offset with formula offset terms removed, preserving cancelling formula and call offsets exactly instead of recovering a contribution by subtraction.
ols_step_both_p() with every term in
include returns the included model with empty step metrics
instead of failing.
The deleted-studentized-residual, standardized-residual,
studentized-residual, Cook’s D, DFFITS and DFBETAS plot data/plot
functions now compute their influence measure on the positive-weight fit
rows and map the original observation IDs back. Fits that combine zero
prior weights with na.action = na.exclude no longer fail
with a differing-row-count error, and their values and observation IDs
agree with the same model fitted after removing the unused
rows.
ols_prep_rstudlev_data() and
ols_plot_resid_lev() compute the leverage threshold from
the model rank rather than the number of coefficient slots, so adding an
aliased predictor no longer raises the threshold and reclassifies
genuine leverage points as “normal”.
ols_press(), ols_pred_rsq() and the
other leverage-based diagnostics stop with an error (“refit with qr =
TRUE”) for a model fitted with qr = FALSE, instead of
rebuilding the projection from the design matrix. The fitted tolerance
and pivot cannot be recovered from the call, and a rebuilt decomposition
can span different columns than the fit and give a wrong PRESS; only the
fit’s own QR is used.
Standardized betas and every stepwise selection routine now
rebuild the fit from the data found through the model call with
model.frame() and model.matrix() (the call’s
subset, weights, offset, na.action and the fit’s contrasts) and require
the response, model matrix (values and column names, so factor coding),
weights and offset to be identical to the fit’s. If the data have
changed since fitting, a factor was releveled, or the call symbol is
bound to different data, they stop with “the data used to fit this model
has changed or cannot be found; refit the model”. When the call data
cannot be found, the stored model frame is used only for formulas made
of plain variables; transformed terms and offsets require the data.
Unchanged fits with NULL weights/offset symbols, subsets,
na.exclude, poly()/scale() terms
and multiple offsets are accepted.
ols_regress() no longer silently refits a model
fitted with qr = FALSE or model = FALSE:
re-evaluating the call could use another environment, tolerance or data
and report a different model than the one supplied. It stops with “refit
with lm(…, qr = TRUE, model = TRUE)”.
ols_press() and ols_pred_rsq() return
NA with a warning naming the observations when a used row
has leverage 1 (for example a singleton factor level or an isolated
indicator), because its deleted residual cannot be estimated; previously
rounding error gave a finite value. ols_regress() and
ols_model_performance() show NA for predicted
R-squared in that case.
Cook’s D (types 2 and 4) and DFFITS thresholds use the model rank and residual degrees of freedom rather than the number of coefficient slots, so an aliased predictor no longer changes the cutoff or hides influential observations.
Observations with leverage 1 (for example the only member of a
factor level) keep undefined (NaN) standardized/studentized
residuals, Cook’s D and DFFITS in the plot data and plots, as in base R,
and a DFBETAS of 0, instead of arbitrary finite values flagged as
normal. Plot ranges ignore the undefined values.
ols_prep_rstudlev_data() computes leverages and
studentized residuals for the positive-weight fit rows and maps the
original observation IDs back, instead of calling the NA-padded
hatvalues()/rstudent() path. Fits that combine
zero prior weights with na.action = na.exclude, or that
omit rows at the end of the data, no longer fail with a
differing-row-count error; the returned leverages, studentized residuals
and IDs agree with the same model fitted after removing the unused
rows.
Standardised betas keep the original coefficient names when a factor level contains text that looks like one of the internal placeholder names. The placeholder is now chosen so that it does not occur anywhere in the data names, factor levels or model-matrix columns, so translating the refitted names back no longer rewrites the factor level itself. The affected Std. Beta cells are printed against their own coefficient rows again.
Standardised betas standardise a predictor even when it is a raw
variable that also appears inside a transformed response, such as
wt in I(mpg / wt) ~ wt * hp or
mpg in log(mpg) ~ mpg * hp. The response
itself is still carried through unchanged, but its separate predictor
occurrence now uses its own scale instead of being left
unstandardised.
ols_pure_error_anova() derives the regression, error
and lack-of-fit degrees of freedom from the model rank and the number of
distinct predictor values, and obtains the regression sum of squares as
the part of the total not explained by the residual sum of squares.
No-intercept fits now report the same error degrees of freedom, overall
F and nested lack-of-fit F as summary() and
anova(). A model with no estimated direction beyond the
intercept — an intercept-only fit or a constant/fully aliased predictor
— is rejected instead of reporting a significant regression for an
intercept-only fit.
PRESS, predicted R-squared and the internal leverage calculation
use only the first rank columns of the fit QR
decomposition. Rank-deficient models (for example a predictor that is an
exact multiple of another) previously included the non-estimable
orthonormal directions, so ols_press() and the predicted
R-squared printed by ols_regress() disagreed with the
equivalent estimable fit.
ols_press(), ols_pred_rsq() and
ols_regress() take the leverage of each observation from
the fit itself by row position in the model frame instead of matching
row names. Fits that combine zero prior weights with
na.action = na.exclude no longer pair residuals with
another observation’s leverage, and no longer fail when the omitted rows
are at the end of the data.
ols_test_breusch_pagan() and
ols_test_score() stop with an informative error when the
auxiliary regression has no direction beyond the intercept — a constant
or fully aliased regressor, or an intercept-only model — instead of
reporting a statistic with zero degrees of freedom, which gave a false
p = 0 rejection for ols_test_breusch_pagan()
and a degenerate p = 1 for
ols_test_score().
ols_pure_error_anova() removes a fixed offset from
the response before splitting the sums of squares, so the decomposition
adds up for offset models: the error sum of squares equals
deviance(model) and the overall regression F equals
summary(). It previously grouped the raw response and
reported an error sum of squares larger than the model’s.
ols_regress() standardised betas and the selection
routines carry both offset contributions of a model that has
offset() terms in its formula and an
offset argument to lm(), instead of silently
dropping the argument offset and standardising or selecting on a
different model.
Standardised betas read the term structure from the model terms rather than splitting deparsed term text, so a predictor whose name contains a colon is standardised correctly, and a term that cannot be evaluated in the model data raises an error instead of leaving that predictor unscaled.
Standardised betas keep the model’s contrast coding, so the names
of a fit made with contrasts = list(...) match
coef() and its values come from the same parameterisation.
Selection refits keep the contrasts too, so an include-all selection
returns the input fit instead of re-parameterising its factors.
A predictor that is also used as an offset is standardised on its own scale; the offset itself stays a fixed, unstandardised vector.
ols_step_all_possible_betas() stops with a clear
error for weighted models instead of returning the unweighted
coefficients of every subset. It now also documents that it requires a
model with an intercept, and counts predictor terms rather than
coefficients when checking the model size.
ols_step_all_possible() and
ols_step_best_subset() enumerate only the requested subset
orders, so a bounded search with many predictors no longer builds the
combination matrices for every order up to the number of predictors. A
max_order larger than the model is rejected before any
allocation.
Predicted R-squared and PRESS now use the model weights for weighted models. PRESS is the weighted sum of squared deleted residuals, so it scales with the weights instead of moving towards one as the weights grow; predicted R-squared, the ratio of PRESS to the weighted total sum of squares, is invariant to a constant rescaling of the weights.
Observations with zero prior weights are excluded consistently
from PRESS, predicted R-squared, MAE and the information criteria, which
now use the effective sample size and the fitted model’s deviance.
Zero-weight fits now agree with the same model fitted after removing
those rows, and na.exclude and na.omit fits
agree with each other.
ols_aic() and ols_sbc() with
method = "SAS" now use the model’s residual sum of squares
(honouring weights and na.exclude) instead of the
unweighted, NA-padded residuals, so the criteria no longer depend on the
missing-value action.
ols_regress() aligns the coefficient table by name
when the model has aliased coefficients, so standard errors, t
statistics, p values and confidence limits stay with their own terms
instead of shifting onto the next one; standardized betas are computed
only for estimable coefficients.
Standardized betas of interaction models now standardize
transformed predictors after the transformation (log(wt),
I(x^2)), keep the levels and names of categorical terms,
and preserve model offsets, both calling styles giving the same values.
Matrix-valued terms such as poly() inside an interaction
now give a clear error instead of a value from a different
model.
ols_pure_error_anova() gives the same lack-of-fit
and overall regression statistics for a unit-weight model as for the
unweighted model, and works for na.exclude fits.
Stepwise and variable selection routines preserve the intercept
and any fixed offset of the input model, so selecting with
include set to every predictor returns the input fit, and
models without an intercept are not refitted with one.
Selection routines keep the formula environment of the input model for every candidate and final fit, so locally defined transformations and predictors outside the data frame work.
ols_test_breusch_pagan() and
ols_test_score() report the rank of the auxiliary
regression (independent tested directions) as degrees of freedom, so
aliased predictors no longer inflate the degrees of freedom and p
values. Overlapping or duplicated vars selections are
de-duplicated.
ols_test_f(), ols_test_score() and
ols_test_breusch_pagan() work for models fitted with
na.action = na.exclude, giving the same results as the
na.omit fit.
ols_vif_tol() works for single-predictor models and
for models without an intercept, returning tolerance and VIF for every
model-matrix column instead of failing.
ols_test_bartlett() stops with the same informative
errors as stats::bartlett.test() when all observations are
in one group or a group has fewer than two observations, instead of
printing a result with NaN or NA
statistics.
ols_regress() now supports weighted linear models
(lm(..., weights = ...)), matching summary(),
anova(), and confint() for coefficient
estimates, standard errors, test statistics, confidence intervals,
R-squared, residual standard error, and the ANOVA decomposition,
including models with interaction terms, whose standardised betas come
from a weighted refit of the standardised predictors. Passing
weights through the formula interface and passing a
glm object now give clear errors, and stepwise and variable
selection routines stop with a clear error message when passed a
weighted model instead of selecting on unweighted refits (#236).
Diagnostics that mix weighted and unweighted sums of squares or
refit auxiliary regressions without the model weights now stop with a
clear error for weighted models: ols_pure_error_anova(),
ols_plot_added_variable(),
ols_plot_comp_plus_resid(),
ols_plot_resid_regressor(),
ols_test_breusch_pagan(), ols_test_score(),
ols_test_f(), ols_correlations(),
ols_test_correlation() and
ols_prep_avplot_data(). Collinearity diagnostics
(ols_vif_tol(), ols_eigen_cindex(),
ols_coll_diag()) remain available because they are computed
from the design matrix (#236).
plot() on a 0-step stepwise selection result (such
as models with near-perfect fit) returns invisibly with an informative
message instead of failing with a differing row count error (#234).
ols_step_forward_aic() (and related forward
selection routines) handles models where all candidate predictors are
forced in include, returning a valid result object with the
included predictors that can be printed and plotted alongside an
informative message that no candidate predictors were left (#235).
Stepwise and selection routines now preserve the model formula
environment and call so the returned $model can be used
directly with ols_regress() and predict() when
transformed terms (log(), poly(),
factor(), I()) are present (#246).
ols_test_normality()’s Cramér–von Mises test now
compares the data with a normal distribution whose mean and standard
deviation are estimated from the data
(nortest::cvm.test()); it previously compared with the
uniform distribution and rejected normality for almost any
data.
ols_test_normality() no longer fails for samples
larger than 5000 or smaller than 8; tests outside their sample-size
range are skipped with a warning.
The package no longer accesses the network or offers to update
itself when attached, and xplorerr is now a suggested
package; ols_launch_app() explains how to install it when
it is missing.
ols_step_all_possible() now counts the models it
will actually fit when warning about a large search or refusing one
without force = TRUE, taking max_order into
account instead of the number of predictors. It warns above
2^15 - 1 models and refuses more than 2^20 - 1
models unless force = TRUE.
ols_test_f(), ols_test_score() and
ols_test_breusch_pagan() accept factor terms in
vars. ols_test_breusch_pagan() prints the
degrees of freedom of the tested term(s) instead of the number of
vars, and rejects factor terms with
multiple = TRUE with an informative error instead of
failing with object 'cyl' not found. vars are
now validated against the model matrix, so model-matrix column names
keep working and components of interaction terms that are not in the
model get the standard model-variable error.
Stepwise selection is hardened against edge cases: the step
counter is always initialized, backward elimination falls back to an
intercept-only model when every predictor is dropped (#186), selection
stops with an informative message when the residual sum of squares is
below 1e-10 (#199), tied criteria are broken
deterministically (#206), and forced include variables are
no longer duplicated or removed by position (#172).
ols_regress() also handles intercept-only models.
ols_regress(), ols_step_forward_p(),
ols_step_backward_p(), ols_step_best_subset(),
ols_step_all_possible() and the AIC, SBC, R2 and adjusted
R2 stepwise routines now refit models that contain transformed terms
(log(), factor(), poly(),
I()) correctly, returning the same estimates as the
original lm() fit, and printing a stepwise result no longer
fails in the final model output for such terms (#244).
The final model from ols_step_backward_p() no longer
repeats variables passed through include.
Refits in ols_regress() and the selection routines
now take the model’s observations by row name, so models fitted on
filtered data use the right rows and report the correct number of
observations instead of shifted rows and a reduced sample size.
Standardised betas of models with interaction terms no longer
standardise the response variable, so interaction models with a
transformed response (log(), sqrt()) return
the correct values instead of a NaN warning and estimates
from a subset of the data.
ols_step_both_p() and the hierarchical selection
routines stop with the near-perfect fit message instead of failing
inside Anova(), and ols_step_both_p() ignores
duplicated include/exclude variables.
ols_regress() and
ols_model_performance() now compute residual Mean Square
Error (MSE) as deviance(model) / model$df.residual,
aligning with the ANOVA residual mean square, and Root Mean Square Error
(RMSE) as sqrt(mse) rather than using the sample variance
denominator mean(residuals^2) (#221).
Information criteria (ols_msep(),
ols_fpe(), ols_apc(), ols_hsp(),
ols_mallows_cp(), ols_sbic(),
ols_aic(), ols_sbc()) now evaluate model
complexity using model$rank, fixing parameter counting for
factor variables and rank-deficient models; ols_msep() now
uses the residual mean square error rather than sum of squares;
ols_aic(corrected = TRUE) now implements the Hurvich-Tsai
correction for R and STATA methods; and ols_sbic() and
ols_aic() include guards returning NA with warnings on
boundary conditions (#224).
ols_test_bartlett() coerces group_var
to a factor, so numeric grouping variables return the correct degrees of
freedom, statistic and p-value (audit F4).
ols_pure_error_anova() reports the overall
regression F using the error mean square, matching
summary.lm(), and returns NA for the lack of
fit statistics when the model has no replicates (audit F5).
ols_prep_avplot_data() drops the intercept only for
models that have one, preserves drop = FALSE, and keeps the
real predictor names (audit F7).
ols_correlations() works for single-predictor
models, returning the zero-order correlation with NA
partial and part correlations, and gives a clear error for
multi-predictor models without an intercept (audit F8).
ols_test_outlier() reports the most extreme
significant outliers when n_max truncates the result (audit
F11).
ols_get_predictors() always returns a
data.frame, including for single-predictor models (audit
F14).
ols_test_breusch_pagan() accepts
vars = NULL and falls back to the fitted values test;
calling with fitted.values = FALSE without
vars or rhs now errors (audit F15).
ols_regress() forwards conf.level,
iterm and title to the formula method when it
is called with a model, so conf_lm uses the requested
confidence level instead of always 95% and the title is printed in the
parameter estimates table (audit F2, #244).
Standardised betas of models without an intercept now include
every coefficient and are aligned with the coefficients by name, instead
of dropping the first predictor and shifting the printed
Std. Beta column (audit F3, #244).
ols_plot_added_variable() stops with a clear message
when the model has fewer than two predictors, instead of failing while
preparing the plot data (audit F10, #244).
goftest is no longer a dependency: the Cramér–von Mises
test for normality is now nortest::cvm.test().This is a minor release for bug fixes and other enhancements.
ols_step_best_subset() unable to force more than one
variable in selection process (#210)plot() with ols_step_both_aic() doesn’t
show anything when AIC values are negative (#212)This is a patch release for urgent bug fixes.
ols_step_all_possible()
(#202)ols_step_all_possible() (#211)ols_regress() (#213)geom_segment() warning in
ols_plot_obs_fit() (#217)This is a minor release for bug fixes and other enhancements.
p
values as variable selection metricols_plot_dffits()ols_test_outlier() does not find any outliers, it
returns largest positive residual instead of largest absolute residual
(#177)ols_step_all_possible() with Model created from
dynamic function leads to
"Error in eval(model$call$data) . . . not found" (#176)ols_step_both_p(): Error in if (pvals[minp] <= pent) {: argument is of length zero
(#175)ols_correlations() returns error for models with 2
predictors (#168)ols_step_both_aic() doesn’t return model (#167)ols_regress() returned residual standard error instead
of RMSE (@jens-daniel-mueller, #165)This is a patch release to reduce the number of packages imported and fix other CRAN errors.
The following functions will now require the variable names to be enclosed within quotes
ols_test_bartlett()ols_plot_resid_regressor()This is a minor release to fix bugs from breaking changes in recipes package and other enhancements.
lm (#81)This is a patch release to fix minor bugs and improve error messages.
olsrr now throws better error messages keeping in mind beginner and intermediate R users. It is a work in progress and should get better in future releases.
Variable selection procedures based on p values now handle categorical variables in the same way as the procedures based on AIC values.
This is a minor release for bug fixes and API changes.
We have made some changes to the API to make it more user friendly:
ols_step_*ols_test_*ols_plot_*ols_regress returns error in the presence of interaction terms in the formula (#49)
ols_regress returns error in the presence of interaction terms in the formula (#47)
return current version (#48)
ols_launch_app() to launch a shiny app for building
modelsols_all_subset() (#41)A big thanks goes to (Dr. Kimberly Henry) for identifying bugs and other valuable feedback that helped improve the package.
This is a minor release containing bug fixes.
This is a minor release containing bug fixes and minor improvements.
ols_avplots) returns error when
model formula contains inline functions (#3)ols_all_subset) returns an
error when the model formula contains inline functions or interaction
variables (#4)ols_best_subset) returns an
error when the model formula contains inline functions or interaction
variables (#5)ols_srsd_plot) returns an
error when the model formula contains inline functions (#6)ols_step_backward)
returns an error when the model formula contains inline functions or
interaction variables (#7)ols_step_backward) returns
an error when the model formula contains inline functions (#8)ols_stepaic_backward)
returns an error when the model formula contains inline functions (#9)ols_stepaic_forward)
returns an error when the model formula contains inline functions (#10)ols_stepaic_both) returns an error
when the model formula contains inline functions (#11)ols_cooksd_barplot)
cook’s d bar plot (#12)ols_regress) returns an error when the
model formula contains inline functions (#21)ols_stepaic_backward) is not properly formatted (#22)ols_stepaic_both) is
not properly formatted (#23)ols_cooksd_barplot) returns the
threshold value used to classify the observations as outliers (#13)ols_cooksd_chart) returns the threshold
value used to classify the observations as outliers (#14)ols_dffits_plot) returns the threshold
value used to classify the observations as outliers (#15)ols_dsrvsp_plot) returns the threshold value used to
classify the observations as outliers (#16)ols_rsdlev_plot) returns the threshold value used to
detect outliers/high leverage observations (#17)ols_srsd_chart) returns
the threshold value used to classify the observations as outliers (#18)ols_srsd_plot) returns the
threshold value used to classify the observations as outliers (#19)There were errors in the description of the values returned by some functions. The documentation has been thoroughly revised and improved in this release.
First release.