The new interaction-plot scales and contrast-based inference require interactions 1.3.0 or later. Earlier interactions releases retain their old plotting behavior and do not provide the corrected panelr simple-slopes or Johnson-Neyman methods. Update both packages to use these features.
Added a contrast-design hook for the updated interactions S3 methods for
sim_slopes() and johnson_neyman(). Slopes, confidence intervals, and
Johnson-Neyman regions use the fitted coefficients and full covariance
matrix without refitting. They match interaction plots' reference settings
and predictor units, including standardized fits. probe_interaction() is
supported. Inference is conditional, fixed-effect, and link-scale with
explicit degrees-of-freedom assumptions and unsupported cases rejected.
Reworked prediction to replay fitted variable roles, factor coding, scaling,
and basis parameters. Predictions preserve input row order and do not require
the response. Panel predictions require entity identifiers even when random
effects are excluded; prepared raw = TRUE rows are no longer transformed.
Fixed zero fixed-effect predictions and duplicated random effects by using lme4's point-prediction implementation directly. Prediction and simulation now share panel preprocessing, and offset values are retained.
Added a distinct hypothetical-grid path for interactions::interact_plot()
and cat_plot(). With the corresponding interactions update, plots default
to within-person deviations in original units; numeric moderator defaults use
within-person variation. Constant moderators retain their original scale.
plot.scale = "original" restores original-level displays. Labels and captions
identify the scale and reference means, and out-of-range deviations warn about
extrapolation. Added numerical checks of plotted values and confidence bands.
Fixed namespaced basis functions and replay of polynomial coefficients. Fixed interaction-column lookup during common-slope detrending.
Unsupported uncertainty/plotting options now fail explicitly. In particular,
use.re.var = TRUE no longer adds variances to standard errors, and automatic
grids for detrending, lags/leads, and basis expansions require prepared data.
Simple slopes instead use direct contrasts because refitting after
recentering is not generally appropriate for within-person interactions.
Fixed compatibility with lme4 2.1-0 by explicitly binding weights and offsets
in the model formula's environment. This also fixes failures in wbm() when
neither argument is supplied.
Requires lmerTest 3.2-1 or later for compatible Satterthwaite calculations.
Preserved lme4 covariance metadata when converting model classes, fixing model display and likelihood calculations with recent lme4 versions.
Fixed make_wb_data() to pass the specified wave argument through to data
preparation instead of using the id argument.
Added a warning when double-demeaned interactions involving an even number of time-varying predictors are used with at most two observations per entity (#31). Cross-level interactions with one time-varying predictor are exempt.
Added clearer error messages when variables in a model formula are missing from the input data.
Fixed the summary.panel_data() fallback when the optional skimr package
is not installed.
Added elongate_panel() as an exported alias for long_panel().
Added regression tests for these changes and for treating wave factors as
unordered factors in wbm().
Fixed bind_rows() dropping the panel class and metadata (#11), and fixed
source labels being assigned to the wrong rows when using its .id argument.
Direct vctrs operations now preserve row order, and row binding refreshes
wave metadata.
Fixed conversion from plm::pdata.frame() when index columns have been
dropped (#9), restoring those columns from the saved index.
Documented reference categories and indicator coding for factor predictors in within-between models (#19).
This is a CRAN resubmission following the March 18, 2026 archive.
mutate_() and filter_(), which are now defunct in current dplyr
releases and caused CRAN checks to fail.This is a point release, albeit without breaking changes. There are numerous long-delayed improvements to performance as well as some new features.
vctrs integration: Added vctrs support for panel_data objects. This
improves attribute preservation during tidyverse operations and provides
better type-safe coercion. New methods include vec_restore.panel_data,
vec_proxy.panel_data, vec_ptype2, and vec_cast methods.
Panel balancing functions: Three new functions help work with unbalanced panel data:
has_gaps(): Check if panel data has implicit missing rowsscan_gaps(): Identify which entity-wave combinations are missingbalance_panel(): Make implicit gaps explicit by adding NA rows for
missing entity-wave combinations. Optionally fill with custom values.wbm() now supports matrix-returning basis expansion terms in the time-varying
part of the formula, including splines::ns(), splines::bs(), and
stats::poly(). These terms are expanded into within- and between-person
components during data preparation. (#36)
New build_panel_data() internal helper provides fast reconstruction of
panel_data objects without full re-validation, improving performance of
operations that preserve panel structure (~8x faster than full reconstruction).
New is_panel_sorted() internal function provides O(n) validation that data
is properly sorted by id then wave, avoiding unnecessary O(n log n) sorting.
If data is out of order, it is automatically re-sorted.
Updated reconstruct.panel_data() to use the lightweight build_panel_data()
helper for faster attribute restoration.
Removed deprecated dplyr method overrides (arrange_, mutate_, summarise_,
summarize_, slice_). These were deprecated in dplyr 0.7.0 (2017) and are
no longer needed.
Refactored interaction effects processing in wbm() and related functions.
The scattered boolean flags (demean.ints, old.ints, detrend) are now
encapsulated in an InteractionConfig object for cleaner conditional logic.
Added WBFormula S3 class for structured representation of parsed formulas
(groundwork for future improvements).
sim_slopes() and johnson_neyman()
from the interactions package (requires jtools >= 2.3.1).
(#57)long_panel() has been substantially sped up. (#51)tibble. (#55)skimr, dplyr, and purrr. Users should not notice any change in behavior from these.long_panel() to check whether values are varying,
it no longer performs a series of operations that are rendered unnecessary. This
will slightly speed up performance when check.varying = FALSE. (#44)long_panel() and
are_varying() functions.Compatibility update for upcoming changes to the clubSandwich. Thanks to
James Pustejovksy for submitting the necessary fixes.
Re-release: There are no changes, but panelr was removed from CRAN because
one of the packages it depended on had also been removed. That package is
now back on CRAN, so panelr will return as well.
Re-release: There are no changes, but panelr was removed from CRAN because
one of the packages it depended on had also been removed. That package is
now back on CRAN, so panelr will return as well.
Bugfixes:
dplyr updates. (#28, #29)wave variable.broom package. (#30)Bugfix:
long_panel() now handles numeric waves correctly when the input data are
unbalanced.brms package's interface for autocorrelated
errors.tidyr package.Bugfixes:
wbm() (#14; thanks
@strengejacke).pdata.frame to panel_data has been fixed.interaction.style argument to make_wb_data().predict.wbm() and predict.wbgee() has been improved. Notably,
the DV does not need to be included in newdata and the ID variable is only
required when necessary.Lots of new stuff! CRAN coming soon as well.
wbgee() works just like wbm(), except uses GEE (via the geepack package)
for estimation. This can give you more trustworthy results under some
circumstances and is much less likely to have convergence problems.fdm() estimates first differences models via GLS (from the nlme package).asym() estimates the linear asymmetric effects model described by
Allison (2019)
via first differences.asym_gee() estimates a similar asymmetric effects model to the one using
cumulative differences described in Allison (2019), but using GEE rather than
conditional logit.heise() produces stability and reliability estimates via the popular
method described in Heise (1969).nlsy and teen_poverty).New stuff:
are_varying() can now also assess individual-level variation, so using
the type = "individual" argument you can instead assess variables like age
that vary over time but change equally for every case.wbm() can now handle transformed dependent variables (e.g. log(y)).
Transformations on the right-hand side of the equation were always supported.panel_data objects are now quite a bit more difficult to break by
accidentally subsetting the ID and wave columns out of existence. Now,
subsetting via data[], select() and implicitly via transmute() will
never remove the ID and wave columns. You will also be warned if you arrange()
a panel_data object since it will generally break lag() functions.panel_data objects now store information about what the periods are for
the data, which you can access with the get_periods() function. For example,
if the waves in your data are the numbers 1 through 7, that's what you'll get.
This is more useful when the periods are irregular, such as if the waves are
the years of a biennial survey.Bugfixes:
wave variable to wbm() in the formula without
running into cryptic errors.lag(x)) could not be
included as a user-specified random effect. Pre-0.5.0, these could be included
if they were surrounded by backticks, but now that hack is unnecessary and
does not work.make_wb_data() is now updated to work with other internal updates
introduced in 0.5.0.long_panel() was never really working right when the source data's labels
were located at the beginning (i.e., label_location = "beginning"). It is
now much more robust.wbm()'s wave.factor argument had become non-functional for some time but
is now fixed.Starting to polish things up for CRAN.
Key changes:
panel_data frames now always place the id and wave columns first (in
that order).wbm() can now handle time-varying factors appropriately. Do note that it
only uses treatment contrasts, however. (#8)line_plot(), to help you explore trends in data.
It's a little rough around the edges for now.wbm objects are now a bit more streamlined and nice-looking.tidy() and glance() methods (from the broom package) for
wbm objects. (#4)as_panel_data() is an alias for panel_data() when supplying a data frame
and an S3 method otherwise. It can be used to convert pdata.frame objects
from the plm package to panel_data.wbm() are now converted to Formula objects to make
working with their multiple parts easier (see the Formula package for more
info).summary method for panel_data frames, which works best
if you have skimr installed. You can use dplyr::select() style syntax
to select which variables you want to describe and choose to get descriptives
by wave and/or entity.This version has switched the default degrees of freedom calculation for
linear wbm models to Satterthwaite, which are more computationally efficient
and less prone to breaking R. They are also calculated on a per-variable basis.
Kenward-Roger standard errors and degrees of freedom can be requested with
the t.df = "Kenward-Roger" argument.
This version includes some major under-the-hood changes, converting from an
S3 object representation to S4. This allows the wbm objects to formally be
extensions of merMod objects, meaning any method that could apply to wbm but
isn't formally implemented will fall back to the merMod implementation.
The panel_data class no longer hardcodes the id and wave variables as "id"
and "wave". Instead, they remain whatever they are named and the panelr
functions will simply know which variables are these special ones.
A new function, make_wb_data, allows users to do the data prepping that
wbm does internally without having to use all the modeling choices made by
wbm.
A series of helper functions have been added to make wbm objects behave
more like regular model objects. Now update, formula, terms,
model.frame, coef, predict, and several more are defined for wbm.
The summary function for wbm has been refined and had some minor bugs
squished.
More tweaks to widen_panel, giving users the option to opt out of the
feature introduced in 0.3.2 that stores data about varying and constant
variables from long_panel. Since poor data labeling in the original wide
data can cause those stored attributes to be wrong, users can use
ignore.attributes = TRUE with widen_panel to force checking for varying
variables with are_varying. Users can now also supply a vector of varying
variables, similar to reshape in base R.
This small update adds an enhancement to long_panel and widen_panel. If
you start with wide data, convert it to long format, and then want to convert
back to wide, the panel_data object in long format will cache information
about the variables to drastically speed up widen_panel when you run it again.
Additionally, are_varying was sped up by about 50%, though it slows
widen_panel down for data with many variables.
Tiny bugfixes:
long_panel would error when supplied a tibble rather than a base
data.frame.magrittr operators used
internally.New functions:
widen_panel converts your panel_data object to wide format, with one row
per entity. This can be useful for SEM analysis and some other things.long_panel does a much more difficult thing, which is convert wide-formatted
data to the more conventional long panel data format. It contains several
means for parsing the variable names of the wide formatted data to produce
a sensible long data frame with all the time-variant variables accounted for
properly. Unlike reshape, it can deal with unbalanced data.are_varying is a function that can let you check whether variables in
long-formatted panel data vary over time or not.New feature:
detrend and balance_correction arguments were added to wbm to implement
the procedures described in Curran and Bauer (2011). These, respectively,
account for over-time trends in the predictors and correcting between-subject
effects when panels are unbalanced.NEWS.md file to track changes to the package.