odds.ratioargument was given to
summ, users were correctly warned that it is a deprecated argument but the exponentiated coefficients were not returned as they should have been.
effect_plotwhen offsets are specified in a formula or a variable is included more than once in a formula.
To reduce the complexity of this package and help people understand what they are getting, I have removed all functions that directly analyze interaction/moderation effects and put them into a new package,
interactions. There are still some functions in
jtools that support
interactions, but some users may find that everything they ever used
jtools for has now moved to
interactions. The following functions have moved to
Hopefully moving these items to a separate package called
interactions will help more people discover those functions and reduce confusion about what both packages are for.
jtools 1.0.0 release, I introduced
make_predictions as a lower-level way to emulate the functionality of
cat_plot. This would return a list object with predicted data, the original data, and a bunch of attributes containing information about how to plot it. One could then take this object, with class
predictions, and use it as the main argument to
plot_predictions, which was another new function that creates the plots you would see in
effect_plot et al.
I have simplified
make_predictions to be less specific to those plotting functions and eliminated
plot_predictions, which was ultimately too complex to maintain and caused problems for separating the interaction tools into a separate package.
make_predictions by default simply creates a new data frame of predicted values along a
pred variable. It no longer accepts
mod2 arguments. Instead, it accepts an argument called
at where a user can specify any number of variables and values to generate predictions at. This syntax is designed to be similar to the
margins packages. See the documentation for more info on this revised syntax.
make_new_data is a new function that supports
make_predictions by creating the data frame of hypothetical values to which the predictions will be added.
I have added a new function,
partialize, that creates partial residuals for the purposes of plotting (e.g., with
effect_plot). One negative when visualizing predictions alongside original data with
effect_plot or similar tools is that the observed data may be too spread out to pick up on any patterns. However, sometimes your model is controlling for the causes of this scattering, especially with multilevel models that have random intercepts. Partial residuals include the effects of all the controlled-for variables and let you see how well your model performs with all of those things accounted for.
You can plot partial residuals instead of the observed data in
effect_plot via the argument
partial.residuals = TRUE or get the data yourself using the
partialize function. It is also integrated into
In keeping with the “tools” focus of this package, I am making available some of the programming tools that previously had only been used internally inside the
Many are familiar with how handy the
%in% operator is, but sometimes we want everything except the values in some object. In other words, we might want
!(x %in% y) instead of
x %in% y. This is where
%nin% (“not in”) acts as a useful shortcut. Now, instead of
!(x %in% y), you can just use
x %nin% y. Note that the actual implementation of
%nin% is slightly different to produce the same results but more quickly for large data. You may run into some other packages that also have a
%nin% function and they are, to my knowledge, functionally the same.
One of my most common uses of both %in% and %nin% is when I want to subset an object. For instance, assume
x is 1 through 5, y is 3 through 7, and I want only the instances of
x that are not in
%nin%, I would write
x[x %nin% y], which leaves you with 1 and 2. I really don’t like having to write the object’s name twice in a row like that, so I created something to simplify further:
%not%. You can now subset
x to only the parts that are not in
y like this:
x %not% y. Conversely, you can do the equivalent of
x[x %in% y] using the
x %just% y.
As special cases for
%just%, if the left-hand side is a matrix or data frame, it is assumed that the right hand side are column indices (if numeric) or column names (if character). For example, if I do
mtcars %just% c("mpg", "qsec"), I get a data frame that is just the “mpg” and “qsec” columns of
mtcars. It is an S3 method so support can be added for additional object types by other developers.
An irritation when writing messages/warnings/errors to users is breaking up the long strings without unwanted line breaks in the output. One problem is not knowing how wide the user’s console is.
wrap_str takes any string and inserts line breaks at whatever the “width” option is set to, which automatically changes according to the actual width in RStudio and in some other setups. This means you can write the error message in a single string across multiple, perhaps indented, lines without those line breaks and indentations being part of the console output.
wrap_str wrappers (pun not intended) around
summno longer prints coefficient tables as data frames because this caused RStudio notebook users issues with the output not being printed to the console and having the notebook format them in less-than-ideal ways. The tables now have a markdown format that might remind you of Stata’s coefficient tables.
md_tableand can be used by others if they want. It is based on
summno longer prints significance stars by default. This can be enabled with the
stars = TRUEargument or by setting the
TRUE(also available via
summhas been removed.
get_colorsis now available to users. It retrieves the color palettes used in
jtoolsnow have a new theme, which you can use yourself, called
theme_nice. The previous default,
theme_apa, is still available but I don’t like it as a default since I don’t think the APA has defined the nicest-looking design guidelines for general use.
effect_plotnow can plot categorical predictors, picking up a functionality previously provided by
effect_plotnow uses tidy evaluation for the
predargument (#37). This means you can pass a variable that contains the name of
pred, which is most useful if you are creating a function, for loop, etc. If using a variable, put a
rlangpackage before it (e.g.,
pred = !! variable). For most users, these changes will not affect their usage.
make_predictions(and by extension
effect_plotand plotting functions in the
interactionspackage) now understands dependent variable transformations better. For instance, there shouldn’t be issues if your response variable is
y. When returning the original data frame, these functions will append a transformed (e.g.,
log(y)) column as needed.
lme4has a bug when generating predictions in models with offsets — it ignores them when the offset is specified via the
offset =argument. I have created a workaround for this.
This is a minor release.
plot_predictionshad an incorrect default value for
interval, causing an error if you used the default arguments with
make_predictions. The default is now
effect_plotwould have errors when the models included covariates (not involved in the interaction, if any) that were non-numeric. That has been corrected. (#41)
FALSE) were not handled by the plotting functions appropriately, causing them to be treated as numeric. They are now preserved as logical. (#40).
sim_slopesgave inaccurate results when factor moderators did not have treatment coding (
"contr.treatment") but are now recoded to treatment coding.
summoutput in RMarkdown documents is now powered by
kableExtra, which (in my opinion) offers more attractive HTML output and seems to have better luck with float placement in PDF documents. Your mileage may vary.
rmdformatsrather than the base
huxtable) will now have conditional namespace registration for users of R 3.6. This shouldn’t have much effect on end users.
This release was initially intended to be a bugfix release, but enough other things came up to make it a minor release.
broomupdate when using
plot_coefsarising from the latest update to
interact_plotno longer errors if there are missing observations in the original data and quantiles are requested.
summ.merMod, the default p-value calculation is now via the Satterthwaite method if you have
lmerTestinstalled. The old default, Kenward-Roger, is used by request or when
pbkrtestis installed but not
lmerTest. It now calculates a different degrees of freedom for each predictor and also calculates a variance-covariance matrix for the model, meaning the standard errors are adjusted as well. It is not the default largely because the computation takes too long for too many models.
johnson_neymannow allows you to specify your own critical t value if you are using some alternate method to calculate it.
johnson_neymannow allows you to specify the range of moderator values you want to plot as well as setting a title.
sim_slopesin a way similar to
modx.values = "plus-minus"). [#31]
plot_summsnow supports facetting the coefficients based on user-specified groupings. See
summvariants now have pretty output in RMarkdown documents if you have the
huxtablepackage installed. This can be disabled with the chunk option
render = 'normal_print'.
pred.valuesin place of
predvals. Don’t go running to change your code, though; those old argument names will still work, but these new ones are clearer and preferred in new code.
sim_slopesobjects. Just save your
sim_slopescall to an object and call the
plotfunction on that object to see what happens. Basically, it’s
huxtableinstalled, you can now call
sim_slopesobject to get a publication-style table. The interface is comparable to
This release has several big changes embedded within, side projects that needed a lot of work to implement and required some user-facing changes. Overall these are improvements, but in some edge cases they could break old code. The following sections are divided by the affected functions. Some of the functions are discussed in more than one section.
These functions no longer re-fit the inputted model to center covariates, impose labels on factors, and so on. This generally has several key positives, including
lmmodels, 60% for
svyglm, and 80% for
merModin my testing). The speed gains increase as the models become more complicated and the source data become larger.
log) in the formula, the function would previously would have a lot of trouble and usually have errors. Now this is supported, provided you input the data used to fit the model via the
dataargument. You’ll receive a warning if the function thinks this is needed to work right.
As noted, there is a new
data argument for these functions. You do not normally need to use this if your model is fit with a
y ~ x + z type of formula. But if you start doing things like
y ~ factor(x) + z, then you need to provide the source data frame. Another benefit is that this allows for fitting polynomials with
effect_plot or even interactions with polynomials with
interact_plot. For instance, if my model was fit using this kind of formula —
y ~ poly(x, 2) + z — I could then plot the predicted curve with
effect_plot(fit, pred = x, data = data) substituting
fit with whatever my model is called and
data with whatever data frame I used is called.
There are some possible drawbacks for these changes. One is that no longer are factor predictors supported in
effect_plot, even two-level ones. This worked before by coercing them to 0/1 continuous variables and re-fitting the model. Since the model is no longer re-fit, this can’t be done. To work around it, either transform the predictor to numeric before fitting the model or use
cat_plot. Relatedly, two-level factor covariates are no longer centered and are simply set to their reference value.
Robust confidence intervals: Plotting robust standard errors for compatible models (tested on
glm). Just use the
robust argument like you would for
Preliminary support for confidence intervals for
merMod models: You may now get confidence intervals when using
merMod objects as input to the plotting functions. Of importance, though, is the uncertainty is only for the fixed effects. For now, a warning is printed. See the next section for another option for
merMod confidence intervals.
Rug plots in the margins: So-called “rug” plots can be included in the margins of the plots for any of these functions. These show tick marks for each of the observed data points, giving a non-obtrusive impression of the distribution of the
pred variable and (optionally) the dependent variable. See the documentation for
effect_plot and the
Facet by the
modx variable: Some prefer to visualize the predicted lines on separate panes, so that is now an option available via the
facet.modx argument. You can also use
plot.points with this, though the division into groups is not straightforward is the moderator isn’t a factor. See the documentation for more on how that is done.
To let users have some more flexibility,
jtools now lets users directly access the (previously internal) functions that make
interact_plot work. This should make it easier to tailor the outputs for specific needs. Some features may be implemented for these functions only to keep the
_plot functions from getting any more complicated than they already are.
The simplest use of the two functions is to use
make_predictions just like you would
cat_plot. The difference is, of course, that
make_predictions only makes the data that would be used for plotting. The resulting
predictions object has both the predicted and original data as well as some attributes describing the arguments used. If you pass this object to
plot_predictions with no further arguments, it should do exactly what the corresponding
_plot function would do. However, you might want to do something entirely different using the predicted data which is part of the reason these functions are separate.
One such feature specific to
make_predictions is bootstrap confidence intervals for
You may no longer use these tools to scale the models. Use
scale_mod, save the resulting object, and use that as your input to the functions if you want scaling.
All these tools have a new default
centered argument. They are now set to
centered = "all", but
"all" no longer means what it used to. Now it refers to all variables not included in the interaction, including the dependent variable. This means that in effect, the default option does the same thing that previous versions did. But instead of having that occur when
centered = NULL, that’s what
centered = "all" means. There is no
NULL option any longer. Note that with
sim_slopes, the focal predictor (
pred) will now be centered — this only affects the conditional intercept.
This function now supports categorical (factor) moderators, though there is no option for Johnson-Neyman intervals in these cases. You can use the significance of the interaction term(s) for inference about whether the slopes differ at each level of the factor when the moderator is a factor.
You may now also pass arguments to
summ, which is used internally to calculate standard errors, p values, etc. This is particularly useful if you are using a
merMod model for which the
pbkrtest-based p value calculation is too time-consuming.
The interface has been changed slightly, with the actual numbers always provided as the
data argument. There is no
x argument and instead a
vars argument to which you can provide variable names. The upshot is that it now fits much better into a piping workflow.
The entire function has gotten an extensive reworking, which in some cases should result in significant speed gains. And if that’s not enough, just know that the code was an absolute monstrosity before and now it’s not.
There are two new functions that are wrappers around
center, which call
gscale but with
n.sd = 1 in the first case and with
center.only = TRUE in the latter case.
Tired of specifying your preferred configuration every time you use
summ? Now, many arguments will by default check your options so you can set your own defaults. See
?set_summ_defaults for more info.
Rather than having separate
center.response arguments, each
summ function now uses
transform.response to collectively cover those bases. Whether the response is centered or scaled depends on the
robust.type argument is deprecated. Now, provide the type of robust estimator directly to
robust. For now, if
robust = TRUE, it defaults to
"HC3" with a warning. Better is to provide the argument directly, e.g.,
robust = "HC3".
robust = FALSE is still fine for using OLS/MLE standard errors.
summ.merMod previously offered an
odds.ratio argument, that has been renamed to
exp (short for exponentiate) to better express the quantity.
vifs now works when there are factor variables in the model.
One of the first bugs
summ ever had occurred when the function was given a rank-deficient model. It is not straightforward to detect, especially since I need to make a space for an almost empty row in the outputted table. At long last, this release can handle such models gracefully.
Like the rest of R, when
summ rounded your output, items rounded exactly to zero would be treated as, well, zero. But this can be misleading if the original value was actually negative. For instance, if
digits = 2 and a coefficient was
-0.003, the value printed to the console was
0.00, suggesting a zero or slightly positive value when in fact it was the opposite. This is a limitation of the
trunc) function. I’ve now changed it so the zero-rounded value retains its sign.
summ.merMod now calculates pseudo-R^2 much, much faster. For only modestly complex models, the speed-up is roughly 50x faster. Because of how much faster it now is and how much less frequently it throws errors or prints cryptic messages, it is now calculated by default. The confidence interval calculation is now “Wald” for these models (see
confint.merMod for details) rather than “profile”, which for many models can take a very long time and sometimes does not work at all. This can be toggled with the
summ.svyglm now will calculate pseudo-R^2 for quasibinomial and quasipoisson families using the value obtained from refitting them as binomial/poisson. For now, I’m not touching AIC/BIC for such models because the underlying theory is a bit different and the implementation more challenging.
summ.lm now uses the t-distribution for finding critical values for confidence intervals. Previously, a normal approximation was used.
summ.default method has been removed. It was becoming an absolute terror to maintain and I doubted anyone found it useful. It’s hard to provide the value added for models of a type that I do not know (robust errors don’t always apply, scaling doesn’t always work, model fit statistics may not make sense, etc.). Bug me if this has really upset things for you.
One new model type has been supported:
rq models from the
quantreg package. Please feel free to provide feedback for the output and support of these models.
To better reflect the capabilities of these functions (not restricted to
lm objects), they have been renamed. The old names will continue to work to preserve old code.
center.response now default to
FALSE to reflect the fact that only OLS models can support transformations of the dependent variable in that way.
There is a new
vars = argument for
scale_mod that allows you to only apply scaling to whichever variables are included in that character vector.
I’ve also implemented a neat technical fix that allows the updated model to itself be updated while not also including the actual raw data in the model call.
A variety of fixes and optimizations have been added to these functions. Now, by default, there are two confidence intervals plotted, a thick line representing (with default settings) the 90% interval and a thinner line for the 95% intervals. You can set
NULL to get rid of the thicker line.
plot_summs, you can also set per-model
summ arguments by providing the argument as a vector (e.g.,
robust = c(TRUE, FALSE)). Length 1 arguments are applied to all models.
plot_summs will now also support models not accepted by
summ by just passing those models to
plot_coefs without using
summ on them.
Another new option is
point.shape, similar to the model plotting functions. This is most useful for when you are planning to distribute your output in grayscale or to colorblind audiences (although the new default color scheme is meant to be colorblind friendly, it is always best to use another visual cue).
The coolest is the new
plot.distributions argument, which if TRUE will plot normal distributions to even better convey the uncertainty. Of course, you should use this judiciously if your modeling or estimation approach doesn’t produce coefficient estimates that are asymptotically normally distributed. Inspiration comes from https://twitter.com/BenJamesEdwards/status/979751070254747650.
broom’s interface for Bayesian methods is inconsistent, so I’ve hacked together a few tweaks to make
stanreg models work with
You’ll also notice vertical gridlines on the plots, which I think/hope will be useful. They are easily removable (see
drop_x_gridlines()) with ggplot2’s built-in theming options.
Changes here are not too major. Like
plot_summs, you can now provide unsupported model types to
export_summs and they are just passed through to
huxreg. You can also provide different arguments to
summ on a per-model basis in the way described under the
plot_summs heading above.
There are some tweaks to the model info (provided by
glance). Most prominent is for
merMod models, for which there is now a separate N for each grouping factor.
New arguments have been added to
remove.y.gridlines, both of which are
TRUE by default. APA hates giving hard and fast rules, but the norm is that gridlines should be omitted unless they are crucial for interpretation.
theme_apa is also now a “complete” theme, which means specifying further options via
theme will not revert
theme_apa’s changes to the base theme.
Behind the scenes the helper functions
drop_gridlines are used, which do what they sound like they do. To avoid using the arguments to those functions, you can also use
drop_y_gridlines which are wrappers around the more general functions.
pf_sv_test — now handle missing data in a more sensible and consistent way.
There is a new default qualitative palette, based on Color Universal Design (designed to be readable by the colorblind) that looks great to all. There are several other new palette choices as well. These are all documented at
This release is limited to dealing with the
huxtable package’s temporary removal from CRAN, which in turn makes this package out of compliance with CRAN policies regarding dependencies on non-CRAN packages.
Look out for
jtools 1.0.0 coming very soon!
sim_slopeswere both encountering errors with
merModinput. Thanks to Seongho Bae for reporting these issues and testing out development versions.
export_summshad an extra space (e.g.,
( 1)) due to changes in
huxtable. The defaults are now just single numbers.
TRUE. It was reporting
alpha * 2in the legend, but now it is accurate again.
johnson_neymannow handles multilevel models from
Jonas Kunst helpfully pointed out some odd behavior of
interact_plot with factor moderators. No longer should there be occasions in which you have two different legends appear. The linetype and colors also should now be consistent whether there is a second moderator or not. For continuous moderators, the darkest line should also be a solid line and it is by default the highest value of the moderator.
export_summs, but that has been fixed.
cat_plotby providing a vector of colors (any format that
ggplot2accepts) for the
summthat formats the output in a way that lines up the decimal points. It looks great.
This may be the single biggest update yet. If you downloaded from CRAN, be sure to check the 0.8.1 update as well.
New features are organized by function.
control.fdroption is added to control the false discovery rate, building on new research. This makes the test more conservative but less likely to be a Type 1 error.
line.thicknessargument has been added after Heidi Jacobs pointed out that it cannot be changed after the fact.
sim_slopesfor 3-way interactions is much-improved.
alpha = .05the critical test statistic was always 1.96. Now, the residual degrees of freedom are used with the t distribution. You can do it the old way by setting
df = "normal"or any arbitrary number.
plot.points(see 0.8.1 for more). You can now plot observed data with 3-way interactions.
mod2valsspecification has been added:
"terciles". This splits the observed data into 3 equally sized groups and chooses as values the mean of each of those groups. This is especially good for skewed data and for second moderators.
linearity.checkoption for two-way interactions. This facets by each level of the moderator and lets you compare the fitted line with a loess smoothed line to ensure that the interaction effect is roughly linear at each level of the (continuous) moderator.
plot.points = TRUE.
jitterargument added for those using
plot.points. If you don’t want the points jittered, you can set
jitter = 0. If you want more or less, you can play with the value until it looks right. This applies to
pbkrtestare slowing things down.
r.squaredis now set to FALSE by default.
plot_summs: A graphic counterpart to
export_summs, which was introduced in the 0.8.0 release. This plots regression coefficients to help in visualizing the uncertainty of each estimate and facilitates the plotting of nested models alongside each other for comparison. This allows you to use
summ features like robust standard errors and scaling with this type of plot that you could otherwise create with some other packages.
plot_coefs: Just like
plot_summs, but no special
summ features. This allows you to use models unsupported by
summ, however, and you can provide
summ objects to plot the same model with different
summ argument alongside each other.
cat_plot: This was a long time coming. It is a complementary function to
interact_plot, but is designed to deal with interactions between categorical variables. You can use bar plots, line plots, dot plots, and box and whisker plots to do so. You can also use the function to plot the effect of a single categorical predictor without an interaction.
Thanks to Kim Henry who reported a bug with
johnson_neyman in the case that there is an interval, but the entire interval is outside of the plotted area: When that happened, the legend wrongly stated the plotted line was non-significant.
Besides that bugfix, some new features:
johnson_neymanfails to find the interval (because it doesn’t exist), it no longer quits with an error. The output will just state the interval was not found and the plot will still be created.
interact_plothas been added. Previously, if the moderator was a factor, you would get very nicely colored plotted points when using
plot.points = TRUE. But if the moderator was continuous, the points were just black and it wasn’t very informative beyond examining the main effect of the focal predictor. With this update, the plotted points for continuous moderators are shaded along a gradient that matches the colors used for the predicted lines and confidence intervals.
Not many user-facing changes since 0.7.4, but major refactoring internally should speed things up and make future development smoother.
This function outputs regression models supported by summ in table formats useful for RMarkdown output as well as specific options for exporting to Microsoft Word files. This is particularly helpful for those wanting an efficient way to export regressions that are standardized and/or use robust standard errors.
The documentation for j_summ has been reorganized such that each supported model type has its own, separate documentation.
?j_summ will now just give you links to each supported model type.
More importantly, j_summ will from now on be referred to as, simply, summ. Your old code is fine; j_summ will now be an alias for summ and will run the same underlying code. Documentation will refer to the summ function, though. That includes the updated vignette.
One new feature for summ.lm:
part.corr = TRUEargument for a linear model, partial and semipartial correlations for each variable are reported.
More tweaks to summ.merMod:
pbkrtestpackage is installed. If it is, p values are calculated based on the Kenward-Roger degrees of freedom calculation and printed. Otherwise, p values are not shown by default with lmer models. P values are shown with glmer models, since that is also the default behavior of
r.squaredoption, which for now is FALSE by default. It adds runtime since it must fit a null model for comparison and sometimes this also causes convergence issues.
Returning to CRAN!
A very strange bug on CRAN’s servers was causing jtools updates to silently fail when I submitted updates; I’d get a confirmation that it passed all tests, but a LaTeX error related to an Indian journal I cited was torpedoing it before it reached CRAN servers.
The only change from 0.7.0 is fixing that problem, but if you’re a CRAN user you will want to flip through the past several releases as well to see what you’ve missed.
Bug fix release:
A lot of changes!
More goodies for users of interact_plot:
Other feature changes: