Quartile/percentile reducers?

[mbostock]

Maybe I’m overoptimizing for brevity here, but I feel like it’d be nice to have named quartile reducers for boxplots. You can always use data => d3.quantile(data, 0.25) instead, but "quartile1" is shorter.

[tophtucker]

fwiw when i was recreating visnu's vl.markErrorband({ extent: 'iqr', color: 'gray', interpolate: 'monotone' }) i was indeed lookin for quartile reducers

(sorta interesting that vl has markErrorband. at first it feels too specific to me! why not just an area? but maybe it's sorta analogous to a boxplot compound mark, and it's useful to have ergonomic names that combine a form with a reducer that has a prominent standard cultural function. it's easy to say 'iqr' or 'stderr'…)

[Fil]

Nothing specific against this, but I'd tend to avoid having many statistics tool—because statisticians will never be happy with them, and will want to create a rich and complete plugin with all the stats.

[mbostock]

Maybe it’s fine as-is, I dunno. I’m uncertain how much we want to build-in to Plot and how much we want to rely on third-party functions (e.g. D3 and Arquero). On the one hand I like not introducing a new API when one already exists: we can focus efforts on learning those existing tools and keep a separation of concerns. On the other hand built-in functionality (e.g. named color schemes and curves) usually feels nicer and “safer” (more constrained) than calling out to a function.

I agree that quartile is quite specific, but some shorthand for a p-quantile would be highly desirable, e.g. p = 0.25 for first quartile and p = 0.75 for third. I guess you’d have to say something like x: {quantile: 0.25} though, or introduce a string language like x: "quantile(0.25)"?

[Fil]

Need to see how to plug-in simple-statistics in a facet-aware/group-aware fashion?

[yurivish]

A common shorthand for tail latencies in the computer performance world is “p99” for 99th percentile and p999 for the 99.9th (and so on, eg. p50 for median).

Slightly related, I’ve been thinking how cool it would be if one could easily facet a continuous variable into some number of quantile-based groups, such as quartiles or deciles.