Deprecated ggplot2 arguments

[BjarkeHautop]

Currently, there are 3 deprecated ggplot2 arguments that appears when running devtools::test():

• The draw_quantiles argument of geom_violin() is deprecated as of ggplot2 4.0.0.
  i Please use the quantiles.linetype argument instead. (Note this deprecation is supposed to say quantile.linetype, as seen in the documentation https://ggplot2.tidyverse.org/reference/geom_violin.html)

This is only used internally, since the public API for ppc_violin_grouped() uses probs. However, when changing all instances of draw_quantiles to quantile.linetype, a test in test-ppc-distributions.R at line 334/335 fails:

> ppc_violin_grouped(vdiff_y, vdiff_yrep, vdiff_group)
Error:
! Problem while converting geom to grob.
ℹ Error occurred in the 1st layer.
Caused by error in `draw_panel()`:
! `geom_path()` can't have varying colour, linewidth, and/or alpha along the line when linetype isn't solid.
Run `rlang::last_trace()` to see where the error occurred.

• ppc_intervals() / ppd_intervals() use deprecated ggplot2 arguments (fatten, size).

The fatten argument of geom_pointrange() is deprecated as of ggplot2 4.0.0.
i Please use the size aesthetic instead.

Using size aesthetic for lines was deprecated in ggplot2 3.4.0.
i Please use linewidth instead.

Changing the public API would be breaking. Furthermore, replacing fatten with size would mean the new size would semantically correspond to the old fatten, not the existing size.

[jgabry]

Thanks @BjarkeHautop! @VisruthSK and I had noticed some of these recently with the release of ggplot2 v4.0. But you mention some additional details I hadn't considered (thank you). A few comments and questions:

This is only used internally, since the public API for ppc_violin_grouped() uses probs. However, when changing all instances of draw_quantiles to quantile.linetype, a test in test-ppc-distributions.R at line 334/335 fails:

Interesting, when I try it I don't see that test failure. There's also mcmc_violin, but changing it there seems to work too. I just put my version of the violin fixes in #410. Would you be willing to take a look and see if there's a difference between my version and your version that might explain the error? I might have made a mistake somewhere. (We'll also see if the error happens with my PR, it didn't locally, but perhaps it will on Github Actions)

Using size aesthetic for lines was deprecated in ggplot2 3.4.0.

I fixed a lot of these uses of size (in fact I just fixed a couple in the intervals plots), but there are probably some that I missed. Since bayesplot requires at least v3.4.0 of ggplot2 we can safely change size to linewidth internally, we just need to find the remaining cases. Like you said, changing the API would be breaking, so for now we'll keep size in the user-facing functions and just change it internally. E.g., mcmc_trace() takes argument size but internally that's passed as the linewidth argument to ggplot2.

The fatten argument of geom_pointrange() is deprecated as of ggplot2 4.0.0.
Furthermore, replacing fatten with size would mean the new size would semantically correspond to the old fatten, not the existing size.

Yeah, you're right and this issue with replacing fatten hadn't occurred to me yet. I'm not sure how we should handle it. We might be forced to make a breaking change in bayesplot and remove the fatten argument and change the behavior of size like you mentioned. Or do you (or anyone else) have any other ideas?

Thanks again for opening the issue!

[VisruthSK]

I'm in favor of deprecating with a warning to swap then doing a breaking change, I think cli or something has a deprecation badge which handles the messaging.

Edit: https://lifecycle.r-lib.org/reference/deprecated.html is the function/package I was thinking about.

[BjarkeHautop]

Interesting, when I try it I don't see that test failure. There's also mcmc_violin, but changing it there seems to work too. I just put my version of the violin fixes in #410.

Pr #410 works for me too, looks good.

Yeah, you're right and this issue with replacing fatten hadn't occurred to me yet. I'm not sure how we should handle it. We might be forced to make a breaking change in bayesplot and remove the fatten argument and change the behavior of size like you mentioned. Or do you (or anyone else) have any other ideas?

Not sure what the best approach is. We have a few options:

• Breaking change, changing the meaning of the arguments so that they match ggplot2 syntax.

• Deprecate them and change public args to a new name(s), so it doesn't clash with ggplot2 syntax.

• Something else?

I'm in favor of deprecating with a warning to swap then doing a breaking change, I think cli or something has a deprecation badge which handles the messaging.

This isn't possible @VisruthSK. The issue is that the meaning of the arguments would change but keep the same name (as explained above), so you can't just deprecate them.

[BjarkeHautop]

Whoops clicked close by accident.

[jgabry]

I'm in favor of deprecating with a warning to swap then doing a breaking change, I think cli or something has a deprecation badge which handles the messaging.

This isn't possible @VisruthSK. The issue is that the meaning of the arguments would change but keep the same name (as explained above), so you can't just deprecate them.

Yeah, I agree this isn't possible in this case unfortunately. Normally we would deprecate and then eventually make the breaking change like @VisruthSK suggested, but because the meaning of the arguments is changing (and the fact that they interact with each other) I don't see how we could do that here.

Not sure what the best approach is. We have a few options:

• Breaking change, changing the meaning of the arguments so that they match ggplot2 syntax.

• Deprecate them and change public args to a new name(s), so it doesn't clash with ggplot2 syntax.

• Something else?

I guess the second is closest to what @VisruthSK was recommending, but I'm not sure which new names would make sense.

I think the best option might be just to make the breaking change, especially since it would match ggplot2.

For example, for ppc_intervals() we have defaults size=1, fatten=2.5. In older versions of ggplot2 the point size in geom_pointrange() would be size * fatten and in the new version I think it's just size, right? So what we could do is change the default to size=2.5 and throw a deprecation warning if the user specifies fatten (e.g. using fatten = deprecated() like at the link @VisruthSK shared above). Then we can eventually remove fatten, but we wouldn't remove it for a while so we don't cause any errors.

What do you both think of that idea?

[BjarkeHautop]

I guess the second is closest to what @VisruthSK was recommending, but I'm not sure which new names would make sense.

I think the best option might be just to make the breaking change, especially since it would match ggplot2.

I agree.

For example, for ppc_intervals() we have defaults size=1, fatten=2.5. In older versions of ggplot2 the point size in geom_pointrange() would be size * fatten and in the new version I think it's just size, right?

Yes, point size is only controlled through size in ggplot2 4.0.0+.

So what we could do is change the default to size=2.5 and throw a deprecation warning if the user specifies fatten (e.g. using fatten = deprecated() like at the link @VisruthSK shared above). Then we can eventually remove fatten, but we wouldn't remove it for a while so we don't cause any errors.

What do you both think of that idea?

Sounds good to me.

[jgabry]

Progress update:

• @BjarkeHautop fixed the rest of the deprecated internal uses of size (changed to linewidth) in Fix all size warnings calls to ggplot2 (deprecated at 3.4.0).  #414
• I updated the violin plots to use the correct quantile arguments depending on whether the user has ggplot2 4.0 or an earlier version (in Fix violin plot deprecation warnings from ggplot2 4.0 #410)

I think the remaining tasks are:

• Deprecate user-facing size arguments that control line width and add a linewidth argument
• Make breaking change with fatten and size for geom_pointrange as described in Deprecated ggplot2 arguments #408 (comment)

[utkarshpawade]

hi @jgabry i did an audit of the codebase for all remaining ggplot2 v4 deprecated patterns beyond whats been resolved in #410 and #414. Here's what I found:

Pattern	Production	Test	Affected Files
fatten argument	~12	~3	ppc-intervals.R, ppd-intervals.R, ppc-discrete.R, ppc-loo.R
coord_flip()	2	0	ppc-errors.R
theme_get()/theme_set()	4	4	bayesplot-ggplot-themes.R, test-aesthetics.R
$ access on ggplot objects	~4	~20	mcmc-intervals.R, ppc-discrete.R, multiple test files
fortify()	1	1	ppc-censoring.R

i also checked for aes_string(), after_stat(), last_plot(), geom_errorbarh(), coord_trans(), and borders(). none of these are used in bayesplot.
regarding the open questions from pr 430:
Default size when replacing fatten: size = 2.5 overshoots because geom_pointrange() still applies its internal fatten multiplier (default 4) on top. size = 1 is worth trying first. we can compare against current output with vdiffr snapshots to calibrate.

geom_linerange() + geom_point() vs geom_pointrange(): this would sidestep the fatten multiplier entirely and give direct control over point size. more code, but size becomes unambiguous. worth exploring.

ppc_rootogram(): here size controls line width, not point size, so it should follow the size → linewidth path instead of the fatten/size path.

[utkarshpawade]

hi @jgabry about the new default size, i ran an empirical sweep on ppc_intervals so we have actual renders to compare against the master baseline.

setup: a copy of ppc_intervals that uses the exact production code path (same intervals_inner_aes, scale_color_ppc, etc.) and only varies the geom_pointrange size / fatten arguments. sanity-checked by rendering it with master defaults next to the real ppc_intervals(y, yrep) pixel-identical. used a sparse 30-point subset of example_y_data so individual dot sizes are clearly visible.

what i see in the grid:

variant	result
size = 0.625 (no fatten)	visually indistinguishable from baseline (0.625 * 4 = 2.5)
size = 1 (no fatten)	moderately bigger than baseline
size = 2.5 (no fatten)	catastrophically big, matches the regression @jgabry flagged
geom_linerange + geom_point(size = 1.5)	clean, decoupled, size means what users expect

so under the current geom_pointrange approach size = 0.625 is the math-faithful drop-in, but it's an ugly default. the split-geom path looks like the better long-term fix from the renders, size gets its expected ggplot2 semantics and the result is closer to baseline at any reasonable value.