see

• plot() for datawizard::means_by_group() gets a flip argument. Use flip = FALSE to show the groups on the horizontal axis (#455).
• geom_violinhalf() and geom_violindot() are now drawn with the *ggdist* package (ggdist::stat_slab() and ggdist::geom_dots()), which they now need (version 3.3.3 or later). They work with facets and free scales (#81, #385), and the plots look a little different. Dodged groups each take half of their dodge cell, from its centre, as before. The dots of geom_violindot() were not dodged before. By default, they now go in the other half of the cell (they overlap a flipped half violin). The new interval argument adds the median with 66% and 95% intervals (#68). The binwidth default of geom_violindot() is now NA (it was 0.05), so ggdist sizes the dots to the data (#44). Other arguments in ... of geom_violindot() now go to ggdist::geom_dots() (instead of geom_dotplot()) as well as to the half violins, except the arguments of the density, the layout, and the interval, which go to the half violins only (see ?geom_violindot). A width in ... is the width of a whole violin, as in geom_violin(), so each half violin is half as wide. bounds must contain all values; otherwise the half violins are not drawn. The stat argument is now ignored, with a warning for any value other than "ydensity" or ggplot2::StatYdensity. geom_violindot() no longer works with coord_polar().
• plot() for modelbased::estimate_contrasts() now draws joint contrasts of two or more variables, such as contrast = c("am", "cyl"). The x axis shows the combinations of levels, such as "auto, 4".
• plot() for modelbased::estimate_contrasts() gains a show_data argument. With show_data = TRUE, the plot draws the raw data of the model as jittered points behind the lighthouses (easystats/modelbased#114).
• plot() for modelbased::estimate_contrasts() now draws more kinds of estimates. It draws means in a Probability, Proportion, or Rate column, such as the probabilities of a logistic model. It draws ratio contrasts from comparison = ratio ~ pairwise as ratios of the means. With the default backend, it draws contrasts of a log response estimated with transform = TRUE as ratios of the back-transformed means. The plot stops with an error for contrasts back-transformed from some other transformations, such as sqrt(y). When insight::find_transformation() reports a transformation other than "identity", it also stops for ratio contrasts estimated with transform = TRUE. It also stops for contrasts and means on different scales, as their predict or transform arguments show. The help page lists the exceptions. With show_data = TRUE, the plot draws a logical or two-level factor response as 0 and 1. It also draws the raw data of y for a response log(y), log2(y), or log10(y). This needs contrasts and means estimated with transform = TRUE.
• plot() for modelbased::estimate_contrasts() now draws ordinal and multinomial models, such as MASS::polr() and mclogit::mblogit(), with one panel per response category. Each panel shows the probabilities of that category and the contrasts within it. Contrasts between two categories are not drawn. show_data = TRUE stops with an error for these models. The plot also draws Bayesian ratio contrasts from comparison = ratio ~ pairwise as ratios. When insight::find_transformation() reports "identity", it draws ratio contrasts estimated with transform = TRUE as ratios. It accepts contrasts and means on the link scale from the marginaleffects and emmeans backends together. For some models, such as lm(y ~ x), the link scale counts as the response scale. These models have an identity link, and insight::find_transformation() reports their response as "identity". The link scale does not count for a model with a zero-inflation or hurdle part.
• plot() for modelbased::estimate_contrasts() now shows only the panels of the by levels that the contrasts have. It leaves out the means and raw data of other levels. A numeric contrast variable at unevenly spaced values, such as contrast = "wt=c(2,2.5,4)", gets a continuous x axis that keeps their spacing. Joint contrasts of three or more variables show one level per line in the axis labels. A by variable named PANEL, ROW, COL, SCALE_X, or SCALE_Y no longer stops the plot.
• If a term with a VIF below 2 has an upper confidence limit that is infinite or above the top of the plot, the collinearity plot now has a caption. The caption says that the interval formula (from *performance*) fails near VIF = 1, and that this is not high collinearity. This applies to plot() for check_collinearity() and to the collinearity panel of check_model(). The y axis now extends below VIF = 1, so the axis no longer cuts off points at 1 (#399).

Bug fixes

• plot() for structural equation models from parameters::model_parameters() drew latent factors as observed variables and indicators as latent variables. Latent factors are now drawn as circles and indicators as squares.
• plot() for structural equation models now gives an informative error when the selected component has no paths to plot, for example component = "regression" for a confirmatory factor analysis (#243).
• geom_binomdensity() now works with facets. It no longer drops the columns other than x and y, so each panel shows only its own data (#192).
• When check_model() failed to create plots, the informative message about troubleshooting and workarounds was no longer printed. This should now work again.
• plot() for modelbased::estimate_contrasts() now draws each contrast estimated with by from the means of its own by level. Previously, it mixed the means of all by levels into each contrast. If the means listed the by variable first, it drew no contrasts. The plot now shows one panel per level, or combination of levels, of the by variables. The x axis now shows the contrasted variable in the order of its levels (easystats/modelbased#267).
• When some observations had a missing Cook's distance, plot() for check_outliers() failed with "arguments imply differing number of rows". An example is rows with missing values in a model fitted with na.action = na.exclude. The scree and count plots now leave these observations out, and the labels still name the right rows.