Adds option to plot average in boxplot, besides the median - #2520

Is there anything else I need to do?

This looks quite good. It needs a regression test.

@pelson, @efiring: You guys look like the most recent major editors of the boxplot code. Any comments?

Looks good, though I'd like to see an existing test updated to include this functionality.

I tried writing another boxplot test at matplotlib/lib/matplotlib/tests/test_axes.py but many tests are failing on my machine. The simple test_boxplot() test fails, and looking at the images the difference seems to be the font of the labels. Does that have something to do with Mac vs Linux?

I could just add another test function like this:

@image_comparison(baseline_images=['boxplot_with_averages'],
                  extensions=['pdf', 'png'])
def test_boxplot_with_averages():
    x = np.linspace(-7, 7, 140)
    x = np.hstack([-25, x, 25])
    fig = plt.figure()
    ax = fig.add_subplot(111)

    # show 1 boxplot with mpl medians/conf. intervals, 1 with manual values
    ax.boxplot([x, x], bootstrap=10000, usermedians=[None, 1.0],
               conf_intervals=[None, (-1.0, 3.5)], averages=True)
    ax.set_ylim((-30, 30))

But first I'd need help figuring out how to run this test and get the previous test to pass.

@miggaiowski What OS are you using?

Mac OS X.

Miguel Gaiowski

On Oct 26, 2013, at 8:13 PM, Thomas A Caswell notifications@github.com wrote:

@miggaiowski What OS are you using?

—
Reply to this email directly or view it on GitHub.

It's been a while, can we get this in without the tests?

I've been using it just fine, along with a colleague.

It no longer merges cleanly so it will need a re-base.

The unit-tests are not because we don't believe it doesn't work now, but so that in the future a seemingly unrelated changes does not break it.

@miggaiowski

I think this is a good addition. If you and core team can hold off, over the US Thanksgiving holiday, I plan to submit a PR that will split the boxplot into three functions: _compute_boxplot, _draw_boxplot, and then an existing API-compliant boxplot that relies on the other two methods.

The idea here is that _compute_boxplot will return a dict of possible values to draw (e.g., confidence intervals, means), and _draw_boxplot will accept that dictionary blindly and make it happen. Then end result being that users can easily define their own dictionary and go straight to _draw_boxplot for advanced functionality.

The main advantage here is that there are a ton of different ways to render a boxplot[1] and the API is getting pretty bloated.

[1] e.g., I have 1 project manager who insists on putting the caps at the 5th and 95th percentiles, and another at the max and min

First pass at list of keys that will be in that dictionary:

1. Medians
2. Confidence intervals around the median
3. Quartiles 1 and 3
4. Whisker top and bottom (caps)
5. outliers (fliers)
6. arithmetic means (averages)
7. anything else?

Another advantage to this approach is that the _draw_boxplot method should easily accommodate a style kwarg. That way people can select different ways to render the same values. (see http://nbviewer.ipython.org/5432378 for Tufte's take on boxplots)

@phobson

Sounds good!
We can hold if off and get it done when you finish your refactoring.

Thanks!

Should I close this or leave it open? =)

@miggaiowski Is there anything is this PR that is not include in #2643?

I think it has everything.

On Sun, Jan 5, 2014 at 3:07 PM, Thomas A Caswell
notifications@github.comwrote:

@miggaiowski https://github.com/miggaiowski Is there anything is this
PR that is not include in #2643#2643
?

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/2520#issuecomment-31618735
.

Miguel Gaiowski

ok, then I am going to close this PR.