Should bins implicitly filter data that is out of range?

Vega-lite’s default for binning is to assign out-of-range data to the nearest defined bin. This is not a default I’ve come across in other charting libraries, and is potentially very misleading. For example, here is a zoomed-in normal distribution (vega editor):

{
  "data": {"sequence": {"start": 0, "stop": 1000}},
  "transform": [{"calculate": "sampleNormal()", "as": "x"}],
  "mark": "bar",
  "encoding": {
    "x": {"field": "x", "type": "quantitative", "bin": {"extent": [-1, 1]}},
    "y": {"type": "quantitative", "aggregate": "count"}
  }
}

If you don’t know to look for this (and my guess is most users wouldn’t) it would lead you to make erroneous conclusions about the content of the data.

[jheer]

I like the proposal to output either null or infinite values for out-of-range bins. I will look into that soon!

[zanarmstrong]

+1 to Jake’s point that this is “potentially very misleading” and “it would lead you to make erroneous conclusions about the content of the data.” Can this get a higher priority?

In Jake’s example, the chart is showing that there are >200 points in the dataset that have a value between 0.8 and 1. But, there aren’t >200 points in the dataset with values between 0.8 and 1. There are ~50.

This breaks the fundamental concept of a histogram that the height of the bar corresponds to the number of points >=x0 and <x1 (where x0 is the left edge of a bin and x1 is the right edge).

Some options

(I’m writing from a user’s perspective, not knowing how this fits into Vega structure)

a. Values outside the extent aren’t shown at all: If a bin extent is defined (either explicitly or via a selection) it seems much more reasonable that those values would be excluded from the chart altogether than end up in a bin that doesn’t match their value. This is what Numpy, ggplot2, and d3 all do.

b. n+2 bins exist: what if there is always an additional bin for “values < the left-most bin” and “values > the right-most bin”? With perhaps a flag to include/exclude these? Or, a flag to “show-bucket for left side”, “show bucket for right side”, “show both”, or “show neither”.

c. Single out of range bin: Something like what kanitw suggested above.

d. Do (a) with a warning saying that “x points have been excluded from the chart due to being out of range”. ggplot2 also does this.

e. The x-axis text clearly indicates that the left/right most bins are for values infinity to -0.8 and 0.8 to infinity, and not -1 to -0.8 and 0.8 to 1.

My personal preference is for (a) and (d), with a flag to enable (b). Option © could easily mess up the scale on the y-axis if you are purposefully excluding a lot of data (more than is any single bin).

Option (e) would at least make it clear what’s happening in the current chart, without having to change how the data is being rendered.

What other libraries do:

Numpy: “Values outside the range are ignored”

ggplot2: “Note that, by default, any values outside the limits will be replaced with NA.”]. A warning is also printed with the number of datapoints not shown.

d3: if you set a domain for the bin, it ignores value outside of those ranges.

In the example notebook Fil adds some amusing warning labels.