Enable `filter` aesthetic à la ggforce

[teunbrand]

From a discussion with Thomas, there was an idea to propagate the filter aesthetic that is found in the ggforce package to geoms in ggplot2. In ggforce, the filtering step is applied after the stat is computed in the Geom$setup_data() step, see e.g. here. We can consider making this a separate method so it will propagate to all extensions as well.

[z3tt]

Interesting idea! I’ve always appreciated that kind of filter option in ggforce, but I’m not entirely sure how well it fits into the broader aesthetics philosophy. In my workshops, less experienced participants often find it confusing when filtering suddenly appears as another aesthetic; especially since we typically filter in the data step, and this doesn’t quite “feel” like an aesthetic.

Just my two cents from a user and instructor perspective.
Maybe I’m missing a rationale as to why this should be considered an aesthetic in the strict sense.

[thomasp85]

I agree that this is more of a "meta-aesthetic" and I'd group it together with the group aesthetic in this regard...

The rationale is that you'd quite often want to have a layer only focusing on a subset of the global data in a plot. Rather than needing to first prepare a subset of this, it feels more ergonomic and more self-documenting to have the filtering together with the layer specification.

[teunbrand]

I think the notion to implement it as an aesthetic is a pragmatic one, rather than from theoretical considerations. You'd want this column to participate in default group determination, to survive stat computations as a within-group constant column and all the other steps aesthetics go through. I'm happy to entertain alternatives, but implementation-wise it'd be the most straightforward. I'd agree that this is nice to have as part of the layer specification.

For reference, this is currently a viable strategy to filter data.

library(ggplot2)

ggplot(mpg, aes(displ, hwy)) +
  geom_point() +
  geom_point(
    data = ~ subset(.x, drv == 'f'),
    colour = 'red'
  )

But a few things make this awkward, namely (1) the lambda syntax, (2) the lack of function-factories for filtering and because this happens up-front, rather than after stat computations, (3) this may misplace data through position adjustments (for geom_boxplot() for example).

[z3tt]

Don’t get me wrong, a simplified filter option would be fantastic! And the pragmatic notion is exactly the rationale I was ignoring.

@teunbrand I’d typically usesubset(mpg, drv == "f"), but I assume the underlying issues remain the same. And of course, that approach breaks down when the data is constructed on the fly and piped straight into ggplot(), so the lambda syntax is definitely appreciated as a solution until then.

Looking forward to whatever you come up with!