RFC: boxplot_enhanced & paired_stats

I just found some code (http://www.onerussian.com/tmp/plots.py and
pasted below for review/feedback) laying around which I wrote around
matplotlib for plotting primarily pair-wise stats results. Here is a
demonstration:
http://nbviewer.ipython.org/url/www.onerussian.com/tmp/run_plots.ipynb

I wonder if there is a need/place for it in matplotlib and what changes would
you advise. Sorry for the lack of documentation -- I guess I have not finished
it at that point (scipy dependency can easily be dropped, used only for
standard error function iirc):

#!/usr/bin/python
#emacs: -*- mode: python-mode; py-indent-offset: 4; tab-width: 4; indent-tabs-mode: nil -*-
#ex: set sts=4 ts=4 sw=4 noet:
#------------------------- =+- Python script -+= -------------------------
"""
@file paired-plots.py
@date Fri Jan 13 11:48:00 2012
@brief

  Yaroslav Halchenko Dartmouth
  web: http://www.onerussian.com College
  e-mail: yoh@...825... ICQ#: 60653192

DESCRIPTION (NOTES):

COPYRIGHT: Yaroslav Halchenko 2012

LICENSE: MIT

  Permission is hereby granted, free of charge, to any person obtaining a copy
  of this software and associated documentation files (the "Software"), to deal
  in the Software without restriction, including without limitation the rights
  to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  copies of the Software, and to permit persons to whom the Software is
  furnished to do so, subject to the following conditions:

  The above copyright notice and this permission notice shall be included in
  all copies or substantial portions of the Software.

  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  THE SOFTWARE.
"""
#-----------------\____________________________________/------------------

__author__ = 'Yaroslav Halchenko'
__revision__ = '$Revision: $'
__date__ = '$Date: $'
__copyright__ = 'Copyright (c) 2012 Yaroslav Halchenko'
__license__ = 'MIT'

import numpy as np
import pylab as pl
import scipy.stats as ss

def plot_boxplot_enhanced(
    v,
    contrast_labels=None,
    condition_labels=None,
    ccolors=['y', 'b'],
    rand_offsets=None,
    grid=True,
    xticks_rotation=0,
    **bp_kwargs):

    width = bp_kwargs.get('width', 0.5)
    pl.boxplot(v, **bp_kwargs)

    if v.ndim < 2: v = v[:, None]
    ncol = v.shape[1]

    eff = np.mean(v, axis=0) # effect sizes
    sem = ss.sem(v, axis=0)

    if rand_offsets is None:
        rand_offsets = np.random.randn(len(v)) * 0.02

    pl.plot((np.arange(ncol) + 1)[:, None] + rand_offsets,
            v.T, '.', color='k', markerfacecolor='k')
    for i in range(ncol):
        lw = 2
        pl.plot([1 - width/2. + i, 1+i],
                [0, 0],
                '--', color=ccolors[0], linewidth=lw) # first condition
        pl.plot([1+i, 1 + width/2. +i],
                [eff[i]]*2,
                '--', color=ccolors[1], linewidth=lw)

        # place ste
        pl.errorbar(i+1 + 1.1*width/2.,
                    eff[i],
                    sem[i],
                    elinewidth=2, linewidth=0,
                    color='r', ecolor='r')

        if contrast_labels and not i: # only for the first one
            pl.text(1 - 1.1*width/2 + i, 0.1, contrast_labels[0],
                    verticalalignment='bottom',
                    horizontalalignment='right')
            pl.text(1 + 1.2*width/2 + i, eff[i], contrast_labels[1],
                    verticalalignment='bottom', horizontalalignment='left')
    ax = pl.gca()
    if condition_labels:
        ax.set_xticklabels(condition_labels, rotation=xticks_rotation)
    else:
        # hide it
        ax.axes.xaxis.set_visible(False)

    if grid:
        ax.grid()
    return ax

def plot_paired_stats(
    v0, v1, contrast_labels,
    condition_labels=None,
    style=['barplot_effect',
           'boxplot_raw',
           'boxplot_effect'],
    ccolors=['y', 'g'],
    xticks_rotation=0,
    grid=False,
    fig=None,
    bottom_adjust=None,
    bp_kwargs={}):

    if isinstance(style, str):
        style = [style]

    nplots = len(style) # how many subplots will be needed

    # assure having 2nd dimension
    if v0.ndim < 2: v0 = v0[:, None]
    if v1.ndim < 2: v1 = v1[:, None]
    assert(v0.shape == v1.shape)

    ncol = v0.shape[1]
    v10 = (v1 - v0) # differences
    mv0 = np.mean(v0, axis=0) # means
    mv1 = np.mean(v1, axis=0)

    eff = np.mean(v10, axis=0) # effect sizes
    sem = ss.sem(v10, axis=0)

    # so that data points have are distinguishable
    rand_offsets = np.random.randn(len(v10)) * 0.02

    # interleaved combination for some plots
    v_ = np.hstack((v0, v1))
    v = np.zeros(v_.shape, dtype=v_.dtype)
    v[:, np.hstack((np.arange(0, ncol*2, 2),
                    np.arange(1, ncol*2, 2)))] = v_

    #print v.shape
    #print np.mean(v0, axis=0), np.mean(v1, axis=0)
    #print np.min(v10, axis=0), np.max(v10, axis=0), \
    # np.mean(v10, axis=0), ss.sem(v10, axis=0)
    #pl.boxplot(v10 + np.mean(v1), notch=1, widths=0.05)

    #print v0.shape, v1.shape, np.hstack([v0, v1]).shape

    if fig is None:
        fig = pl.figure()

    bwidth = 0.5
    plot = 1

    if condition_labels:
        xlabels = [ '%s:%s' % (cond, contr)
                    for cond in condition_labels
                    for contr in contrast_labels ]
    else:
        xlabels = contrast_labels

    bp_kwargs_ = {
        #'bootstrap': 0,
        'notch' : 1
        }
    bp_kwargs_.update(bp_kwargs)

    def plot_grid(ax):
        if grid:
            ax.grid()

    if 'barplot_effect' in style:
        if len(style) > 1:
            pl.subplot(1, nplots, plot)
        plot += 1
        # The simplest one
        pl.bar(np.arange(1, ncol*2+1) - bwidth/2,
               np.mean(v, axis=0),
               color=ccolors*ncol,
               edgecolor=ccolors*ncol,
               alpha=0.8,
               width=bwidth)
        #pl.minorticks_off()
        pl.tick_params('x', direction='out', length=6, width=1,
                       top=False)
        ax = pl.gca()
        pl.xlim(0.5, ncol*2+0.5)
        ax.set_xticks(np.arange(1, ncol*2+1))
        ax.set_xticklabels(xlabels, rotation=xticks_rotation)
        # place ste for effect size into the 2nd column
        pl.errorbar(np.arange(ncol)*2+2,
                    mv1,
                    sem, elinewidth=2, linewidth=0,
                    color='g', ecolor='r')

        plot_grid(ax)

    if 'boxplot_raw' in style:
        if len(style) > 1:
            pl.subplot(1, nplots, plot)
        plot += 1

        # Figure 1 -- "raw" data
        # plot "connections" between boxplots
        for i in range(ncol):
            pargs = (np.array([i*2+1, i*2+2])[:, None] + rand_offsets,
                     np.array([v0[:,i], v1[:,i]]))
            pl.plot(*(pargs+('-',)), color='k', alpha=0.5, linewidth=0.25)
            pl.plot(*(pargs+('.',)), color='k', alpha=0.9)
        # boxplot of "raw" data
        bp1 = pl.boxplot(v, widths=bwidth, **bp_kwargs_)
        for i in range(ncol):
            for c in xrange(2):
                b = bp1['boxes'][2*i+c]
                b.set_color(ccolors[c])
                b.set_linewidth(2)

        ax = pl.gca()
        ax.set_xticklabels(xlabels, rotation=xticks_rotation)
        plot_grid(ax)

    if 'boxplot_effect' in style:
        if len(style) > 1:
            pl.subplot(1, nplots, plot)
        plot += 1
        plot_boxplot_enhanced(v10,
                              contrast_labels=contrast_labels,
                              condition_labels=condition_labels,
                              widths=bwidth,
                              rand_offsets=rand_offsets, # reuse them
                              grid=grid,
                              **bp_kwargs_)

    if bottom_adjust:
        fig.subplots_adjust(bottom=bottom_adjust)
    pl.draw_if_interactive()
    return fig

if __name__ == '__main__':

    if True:
        v = np.random.normal(size=(50,8)) * 20 + 120
        if False:
            v[:, 1] += 40
            v[:, 3] -= 30
            v[:, 5] += 60
            v[:, 6] -= 60
        else:
            v -= np.arange(v.shape[1])*10
        v /= 10

    v0 = v[:, ::2]
    v1 = v[:, 1::2]
    d = v1 - v0
    print np.mean(d, axis=0)
    styles = ['barplot_effect',
              'boxplot_raw',
              'boxplot_effect'
              ]
    styles = styles + [styles]
    pl.close('all')

    if False:
        f = plot_boxplot_enhanced((v1-v0)[:,0],
                                  grid=True, xticks_rotation=30, notch=1)

    for s in styles:
        fig = pl.figure(figsize=(12,6))
        f = plot_paired_stats(v0, v1, ['cont1', 'cont2'],
                              style=s, fig=fig,
                              condition_labels=['exp1', 'exp2', 'exp3', 'exp4'],
                              grid=True, xticks_rotation=30)
    pl.show()

···

--
Yaroslav O. Halchenko
Postdoctoral Fellow, Department of Psychological and Brain Sciences
Dartmouth College, 419 Moore Hall, Hinman Box 6207, Hanover, NH 03755
Phone: +1 (603) 646-9834 Fax: +1 (603) 646-1419
WWW: http://www.linkedin.com/in/yarik

I just found some code (http://www.onerussian.com/tmp/plots.py and
pasted below for review/feedback) laying around which I wrote around
matplotlib for plotting primarily pair-wise stats results. Here is a
demonstration:
http://nbviewer.ipython.org/url/www.onerussian.com/tmp/run_plots.ipynb

I wonder if there is a need/place for it in matplotlib and what changes would
you advise. Sorry for the lack of documentation -- I guess I have not finished
it at that point (scipy dependency can easily be dropped, used only for
standard error function iirc):

Looks nice. We'd certainly be interesting in including it in
statsmodels/graphics if there isn't sufficient interest here and/or
you'd like to keep the scipy dependency. :wink:

Skipper

···

On Fri, Nov 16, 2012 at 10:19 AM, Yaroslav Halchenko <sf@...825...> wrote:

#!/usr/bin/python
#emacs: -*- mode: python-mode; py-indent-offset: 4; tab-width: 4; indent-tabs-mode: nil -*-
#ex: set sts=4 ts=4 sw=4 noet:
#------------------------- =+- Python script -+= -------------------------
"""
@file paired-plots.py
@date Fri Jan 13 11:48:00 2012
@brief

  Yaroslav Halchenko Dartmouth
  web: http://www.onerussian.com College
  e-mail: yoh@...825... ICQ#: 60653192

DESCRIPTION (NOTES):

COPYRIGHT: Yaroslav Halchenko 2012

LICENSE: MIT

  Permission is hereby granted, free of charge, to any person obtaining a copy
  of this software and associated documentation files (the "Software"), to deal
  in the Software without restriction, including without limitation the rights
  to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  copies of the Software, and to permit persons to whom the Software is
  furnished to do so, subject to the following conditions:

  The above copyright notice and this permission notice shall be included in
  all copies or substantial portions of the Software.

  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  THE SOFTWARE.
"""
#-----------------\____________________________________/------------------

__author__ = 'Yaroslav Halchenko'
__revision__ = '$Revision: $'
__date__ = '$Date: $'
__copyright__ = 'Copyright (c) 2012 Yaroslav Halchenko'
__license__ = 'MIT'

import numpy as np
import pylab as pl
import scipy.stats as ss

def plot_boxplot_enhanced(
    v,
    contrast_labels=None,
    condition_labels=None,
    ccolors=['y', 'b'],
    rand_offsets=None,
    grid=True,
    xticks_rotation=0,
    **bp_kwargs):

    width = bp_kwargs.get('width', 0.5)
    pl.boxplot(v, **bp_kwargs)

    if v.ndim < 2: v = v[:, None]
    ncol = v.shape[1]

    eff = np.mean(v, axis=0) # effect sizes
    sem = ss.sem(v, axis=0)

    if rand_offsets is None:
        rand_offsets = np.random.randn(len(v)) * 0.02

    pl.plot((np.arange(ncol) + 1)[:, None] + rand_offsets,
            v.T, '.', color='k', markerfacecolor='k')
    for i in range(ncol):
        lw = 2
        pl.plot([1 - width/2. + i, 1+i],
                [0, 0],
                '--', color=ccolors[0], linewidth=lw) # first condition
        pl.plot([1+i, 1 + width/2. +i],
                [eff[i]]*2,
                '--', color=ccolors[1], linewidth=lw)

        # place ste
        pl.errorbar(i+1 + 1.1*width/2.,
                    eff[i],
                    sem[i],
                    elinewidth=2, linewidth=0,
                    color='r', ecolor='r')

        if contrast_labels and not i: # only for the first one
            pl.text(1 - 1.1*width/2 + i, 0.1, contrast_labels[0],
                    verticalalignment='bottom',
                    horizontalalignment='right')
            pl.text(1 + 1.2*width/2 + i, eff[i], contrast_labels[1],
                    verticalalignment='bottom', horizontalalignment='left')
    ax = pl.gca()
    if condition_labels:
        ax.set_xticklabels(condition_labels, rotation=xticks_rotation)
    else:
        # hide it
        ax.axes.xaxis.set_visible(False)

    if grid:
        ax.grid()
    return ax

def plot_paired_stats(
    v0, v1, contrast_labels,
    condition_labels=None,
    style=['barplot_effect',
           'boxplot_raw',
           'boxplot_effect'],
    ccolors=['y', 'g'],
    xticks_rotation=0,
    grid=False,
    fig=None,
    bottom_adjust=None,
    bp_kwargs={}):

    if isinstance(style, str):
        style = [style]

    nplots = len(style) # how many subplots will be needed

    # assure having 2nd dimension
    if v0.ndim < 2: v0 = v0[:, None]
    if v1.ndim < 2: v1 = v1[:, None]
    assert(v0.shape == v1.shape)

    ncol = v0.shape[1]
    v10 = (v1 - v0) # differences
    mv0 = np.mean(v0, axis=0) # means
    mv1 = np.mean(v1, axis=0)

    eff = np.mean(v10, axis=0) # effect sizes
    sem = ss.sem(v10, axis=0)

    # so that data points have are distinguishable
    rand_offsets = np.random.randn(len(v10)) * 0.02

    # interleaved combination for some plots
    v_ = np.hstack((v0, v1))
    v = np.zeros(v_.shape, dtype=v_.dtype)
    v[:, np.hstack((np.arange(0, ncol*2, 2),
                    np.arange(1, ncol*2, 2)))] = v_

    #print v.shape
    #print np.mean(v0, axis=0), np.mean(v1, axis=0)
    #print np.min(v10, axis=0), np.max(v10, axis=0), \
    # np.mean(v10, axis=0), ss.sem(v10, axis=0)
    #pl.boxplot(v10 + np.mean(v1), notch=1, widths=0.05)

    #print v0.shape, v1.shape, np.hstack([v0, v1]).shape

    if fig is None:
        fig = pl.figure()

    bwidth = 0.5
    plot = 1

    if condition_labels:
        xlabels = [ '%s:%s' % (cond, contr)
                    for cond in condition_labels
                    for contr in contrast_labels ]
    else:
        xlabels = contrast_labels

    bp_kwargs_ = {
        #'bootstrap': 0,
        'notch' : 1
        }
    bp_kwargs_.update(bp_kwargs)

    def plot_grid(ax):
        if grid:
            ax.grid()

    if 'barplot_effect' in style:
        if len(style) > 1:
            pl.subplot(1, nplots, plot)
        plot += 1
        # The simplest one
        pl.bar(np.arange(1, ncol*2+1) - bwidth/2,
               np.mean(v, axis=0),
               color=ccolors*ncol,
               edgecolor=ccolors*ncol,
               alpha=0.8,
               width=bwidth)
        #pl.minorticks_off()
        pl.tick_params('x', direction='out', length=6, width=1,
                       top=False)
        ax = pl.gca()
        pl.xlim(0.5, ncol*2+0.5)
        ax.set_xticks(np.arange(1, ncol*2+1))
        ax.set_xticklabels(xlabels, rotation=xticks_rotation)
        # place ste for effect size into the 2nd column
        pl.errorbar(np.arange(ncol)*2+2,
                    mv1,
                    sem, elinewidth=2, linewidth=0,
                    color='g', ecolor='r')

        plot_grid(ax)

    if 'boxplot_raw' in style:
        if len(style) > 1:
            pl.subplot(1, nplots, plot)
        plot += 1

        # Figure 1 -- "raw" data
        # plot "connections" between boxplots
        for i in range(ncol):
            pargs = (np.array([i*2+1, i*2+2])[:, None] + rand_offsets,
                     np.array([v0[:,i], v1[:,i]]))
            pl.plot(*(pargs+('-',)), color='k', alpha=0.5, linewidth=0.25)
            pl.plot(*(pargs+('.',)), color='k', alpha=0.9)
        # boxplot of "raw" data
        bp1 = pl.boxplot(v, widths=bwidth, **bp_kwargs_)
        for i in range(ncol):
            for c in xrange(2):
                b = bp1['boxes'][2*i+c]
                b.set_color(ccolors[c])
                b.set_linewidth(2)

        ax = pl.gca()
        ax.set_xticklabels(xlabels, rotation=xticks_rotation)
        plot_grid(ax)

    if 'boxplot_effect' in style:
        if len(style) > 1:
            pl.subplot(1, nplots, plot)
        plot += 1
        plot_boxplot_enhanced(v10,
                              contrast_labels=contrast_labels,
                              condition_labels=condition_labels,
                              widths=bwidth,
                              rand_offsets=rand_offsets, # reuse them
                              grid=grid,
                              **bp_kwargs_)

    if bottom_adjust:
        fig.subplots_adjust(bottom=bottom_adjust)
    pl.draw_if_interactive()
    return fig

if __name__ == '__main__':

    if True:
        v = np.random.normal(size=(50,8)) * 20 + 120
        if False:
            v[:, 1] += 40
            v[:, 3] -= 30
            v[:, 5] += 60
            v[:, 6] -= 60
        else:
            v -= np.arange(v.shape[1])*10
        v /= 10

    v0 = v[:, ::2]
    v1 = v[:, 1::2]
    d = v1 - v0
    print np.mean(d, axis=0)
    styles = ['barplot_effect',
              'boxplot_raw',
              'boxplot_effect'
              ]
    styles = styles + [styles]
    pl.close('all')

    if False:
        f = plot_boxplot_enhanced((v1-v0)[:,0],
                                  grid=True, xticks_rotation=30, notch=1)

    for s in styles:
        fig = pl.figure(figsize=(12,6))
        f = plot_paired_stats(v0, v1, ['cont1', 'cont2'],
                              style=s, fig=fig,
                              condition_labels=['exp1', 'exp2', 'exp3', 'exp4'],
                              grid=True, xticks_rotation=30)
    pl.show()

--
Yaroslav O. Halchenko
Postdoctoral Fellow, Department of Psychological and Brain Sciences
Dartmouth College, 419 Moore Hall, Hinman Box 6207, Hanover, NH 03755
Phone: +1 (603) 646-9834 Fax: +1 (603) 646-1419
WWW: http://www.linkedin.com/in/yarik

------------------------------------------------------------------------------
Monitor your physical, virtual and cloud infrastructure from a single
web console. Get in-depth insight into apps, servers, databases, vmware,
SAP, cloud infrastructure, etc. Download 30-day Free Trial.
Pricing starts from $795 for 25 servers or applications!
http://p.sf.net/sfu/zoho_dev2dev_nov
_______________________________________________
Matplotlib-devel mailing list
Matplotlib-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/matplotlib-devel

I was going to suggest either the same thing or adding it to pandas. I
think statsmodels if the better fit, though. I also noticed scipy is only
used for scipy.stats.sem -- so it might be easy enough to loose the scipy
dependency. Just a thought.
-paul

···

On Fri, Nov 16, 2012 at 7:58 AM, Skipper Seabold <jsseabold@...149...>wrote:

On Fri, Nov 16, 2012 at 10:19 AM, Yaroslav Halchenko <sf@...825...> > wrote:
> I just found some code (http://www.onerussian.com/tmp/plots.py and
> pasted below for review/feedback) laying around which I wrote around
> matplotlib for plotting primarily pair-wise stats results. Here is a
> demonstration:
> http://nbviewer.ipython.org/url/www.onerussian.com/tmp/run_plots.ipynb
>
> I wonder if there is a need/place for it in matplotlib and what changes
would
> you advise. Sorry for the lack of documentation -- I guess I have not
finished
> it at that point (scipy dependency can easily be dropped, used only for
> standard error function iirc):
>

Looks nice. We'd certainly be interesting in including it in
statsmodels/graphics if there isn't sufficient interest here and/or
you'd like to keep the scipy dependency. :wink:

Skipper

Thanks everyone for the feedback.

yeah -- scipy dependency is just a joke, as I said, since only sem is in
use, so would be trivial to 'fix'.

As for where to contribute unless to matplotlib -- I would have just
sticked it in our own PyMVPA :wink: but I think those would be generally
useful, so I will just try to cook up a proper PR against matplotlib
some time next week and then go from there.

Cheers!

···

On Fri, 16 Nov 2012, Paul Hobson wrote:

     [4]http://nbviewer.ipython.org/url/www.onerussian.com/tmp/run_plots.ipynb

     > I wonder if there is a need/place for it in matplotlib and what
     changes would
     > you advise. Sorry for the lack of documentation -- I guess I have not
     finished
     > it at that point (scipy dependency can easily be dropped, used only
     for
     > standard error function iirc):
     Looks nice. We'd certainly be interesting in including it in
     statsmodels/graphics if there isn't sufficient interest here and/or
     you'd like to keep the scipy dependency. :wink:
   I was going to suggest either the same thing or adding it to pandas. I
   think statsmodels if the better fit, though. I also noticed scipy is only
   used for scipy.stats.sem -- so it might be easy enough to loose the scipy
   dependency. Just a thought.

--
Yaroslav O. Halchenko
Postdoctoral Fellow, Department of Psychological and Brain Sciences
Dartmouth College, 419 Moore Hall, Hinman Box 6207, Hanover, NH 03755
Phone: +1 (603) 646-9834 Fax: +1 (603) 646-1419
WWW: http://www.linkedin.com/in/yarik