[FR] Errorbars

[KronosTheLate]

It would be super nice if I could supply a mapping(:x_cord, :y_cord, error=:stddev) mapping. It would perhaps be nice to have a version error_x to allow errorbars in that dimension also.

[KronosTheLate]

Awesome! Lovely to see progress here, AoG holds a lot of promise.

Will the linked PR enable everything described in the seaborn plotting package documentation? That is a great target IMO, as that page is one of the best resources I have seen when it comes to parametric/nonparametric estimates of uncertainty/spread. The ease by which one can chose any of the four combinations with seaborn was of great help for my latest project, and I am dreaming of AoG having the same capabilities.

I am asking because AFAICT, the linked PR does not really close this issue, but facilitates making a PR that closes it. And I mention seaborn because IMO, it has everyone one could wish for when it comes to plotting statistical estimates.

P.S: I know I am just asking for festures without rolling up my sleeves. I am very aware that I am simply making request on other people's time, and I am very grateful for everything you and others have done for making plotting in Julia awesome 😍

[jkrumbiegel]

the linked PR does not really close this issue, but facilitates making a PR that closes it

Huh I guess I must've read the issue incorrectly. I thought this was about errorbars not working correctly because it did not adhere to the x, y, z schema. But you'd actually like to have a generic error keyword added to all plotting functions that take coordinates? I'm not quite sure how that would look. Maybe an analysis would do that one can multiply with some mapping, so that you don't have to determine std or so yourself?

[KronosTheLate]

I am also not at all sure how it would work. I believe it is sort of automagical in other frameworks. Let me sketch out an example:

Given the following DataFrame:

x	y
1	0.9
1	0.1
1	1.1
2	3.8
2	3.9
2	4

If I plot an x-y scatter, I would expect that the package detects that the x-coordinate is the same for three points, and represents them as a single point with an errorbar. That is sort of icky though - what if the x-values are floats and not integers, and are only approximately the same? Of what if you want seperate markers on the same x-coordinate?

If one wants to be more explicit about the fact that we want to aggregate points with the same x-coordinate, one could add a keyword/function/struct est, akin to seaborn's Est object (also related is seaborn's Agg object). This is quite flexible, capable, and intuitive, without the automagic part. It also provides an entry-point for specifying how to estimate the point-estimate (median? mean? Geothmetic Meandean?) and errorbars (+- 1 std? +- 1 standard error? A certain percentile/confidence interval (see 1st figure here?)

[KronosTheLate]

I do suspect that a full treatment of this issue (developting, testing, writing docs) is a rather large task, and more like AoG growing a new small limb rather than a tiny addition. But it would be really really nice, and IMO such statistical estimates is one of the larger value-adds of putting your data in a dataframe and using frameworks like AoG.

[asinghvi17]

This would probably be best as either an analysis or a new recipe - but beeswarm from SwarmMakie does a fundamentally very similar thing (grouping together all rows with the same x-coordinate), so might be some inspiration.

[asinghvi17]

But you could probably do this already, with e.g. violin? So maybe this should just be a convert_arguments for errorbars on the Makie end.

[KronosTheLate]

Yhea I guess violin does a similar thing. But in my greed, I would also love the result to be general enough that if the X-values in my example above take on each value from 1 to 10 000, then doing lines instead of scatter would seemlessly created a shaded uncertainty/spread area around the line.

So while a violin plot definitively shares some functionality, there is also stuff that needs to be added to that functionality.

[asinghvi17]

That's definitely a different recipe I'd say, then you have the choice of how many standard deviations your band takes etc