custom error bars for factorplot/bar plot? #question #enhancement

[szeitlin]

Missing from the tutorial/documentation, not sure if this is doable or how to do it:
dataframe -> factor plot #works great

How do I get it to use one of my other columns containing the errors (calculated elsewhere)?

Tried following matplotlib instructions for 'tweaking', doesn't work. Can't find good examples.

[mwaskom]

This isn't functionality that is currently possible to do in any easy way, and it would be enough of a departure from the way things currently work that it isn't likely to be added any time soon. Of course, there is plt.errorbar and plt.bar has a yerr argument which should do what you want.

[szeitlin]

Ah, that's what I suspected.

I have tried plt.errorbar and plt.bar... the truth is, I hate
matplotlib. It's just UGLY.

Anyway thanks for the fast reply! I really like seaborn.

Sam

Michael Waskom mailto:notifications@github.com
October 15, 2014 at 5:43 PM

This isn't functionality that is currently possible to do in any easy
way, and it would be enough of a departure from the way things
currently work that it isn't likely to be added any time soon. Of
course, there is |plt.errorbar| and |plt.bar| has a |yerr| argument
which should do what you want.

—
Reply to this email directly or view it on GitHub
#331 (comment).

[mwaskom]

Just out of curiosity is there some major difference in the way you're calculating your error bars that can't be covered by what's in factorplot?

[szeitlin]

No, I don't think so. I just happened to have a situation where I needed
to collapse a few categories into one (some genius made this table so it
had '0', 'None' and 'Unknown' as categories in the same column (!).

And I already had calculated the descriptive statistics using the
revised list of categories, so what I ended up with was a small df with
something like ('category names', 'means', 'errors') and I just want to
plot that.

Anyway, I think I can make the plot I want, I just have to re-factor the
categories in the original dataframe, which is not a big deal.

Sam

Michael Waskom mailto:notifications@github.com
October 16, 2014 at 11:17 AM

Just out of curiosity is there some major difference in the way you're
calculating your error bars?

—
Reply to this email directly or view it on GitHub
#331 (comment).

[linbug]

Hi Michael,

Related to this question, as I understand it the confidence intervals for factorplot and barplot are calculated using a bootstrapping procedure, and are therefore suitable for use with small or nonparametric datasets. Is this correct? (apologies if this is stated somewhere, but I can't find the method for their generation explicitly written in the docs)

Thanks,
Lin

[mwaskom]

Correct, all the confidence intervals are computed using the bootstrap.

[alkalait]

I see two ways to using precomputed errors on a seaborn.barplot by plotting the errorbars directly onto the existing axis:

import matplotlib.pyplot as plt
plt.errorbar(x=x, y=y, fmt='none', xerror=error, ecolor='k', elinewidth=2)

or

seaborn.barplot(..., **{'xerr':error})

In the former, the fmt='none' option draws just the errorbars. The latter passes arguments to the underlying plt.bar caller, but is less flexible, e.g. does not accept an elinewidth parameter.

The only aesthetic difference I can see to a pure seaborn solution, is the round barline capstyle. errorbar returns a Linecollection object, but it doesn't seem to store/control capstyle information.

[mwaskom]

It's redundant to use the ** operator and a dictionary, you can just do xerr=error. This won't work for nested bars, by the way.

[nileracecrew]

One use case for custom errors is large datasets where bootstrap confidence intervals take a long time to compute, and sensible CIs can be computed by other means.

[buddrball]

This thread was very helpful. Thank you, everyone. Out of curiosity, in my field we frequently have very few replicates (n = 3 - 5) because of difficulty/expense per replicate. How well does bootstrapping work for representing error of very small datasets? I typically use standard deviation rather than standard error to avoid of over-stating the significance of my data. Does anyone have a recommendation on how to best use bootstrapping in this case? Thanks for your time. And Seaborn rocks.

[mwaskom]

If you have an N = 3-5, why not just show all the datapoints?

[domenico-somma]

Hi. Thanks for seaborn. I love it.

I have an issue with the custom error bars
I have a dataframe like:
df

n	gene_short_name	Sample	Average	CI
0	IL6	Sample 1	0.000000	0.000000
1	IL17RC	Sample 1	2.409620	0.756890
2	CCL15	Sample 1	0.000000	0.000000
3	IL6	Sample 2	1.000000	0.900000
4	IL17RC	Sample 2	2.556600	0.781060
5	CCL15	Sample 2	0.000000	0.000000
6	IL6	Sample 3	0.029063	0.029063
7	IL17RC	Sample 3	2.093810	0.682930
8	CCL15	Sample 3	0.000000	0.000000
9	IL6	Sample 4	0.016800	0.016800
10	IL17RC	Sample 4	1.727750	0.594910
11	CCL15	Sample 4	1.000000	0.800000

And I can plot easily with
g = sns.factorplot(x="Sample", y="Average", hue="gene_short_name", data=df)

If I understood correctly, there is not an easy way to plot the error bars.

Here I found that a possible workaround could be:

• Find the x,y coordinates for each point

x_coords = []
   y_coords = []
   for point_pair in ax.collections:
       for x, y in point_pair.get_offsets():
           x_coords.append(x)
           y_coords.append(y)

• Calculate the type of error to plot as the error bars, Make sure the order is the same as the points were looped over

errors = tips.groupby(['smoker', 'sex']).std()['tip']
colors = ['steelblue']*2 + ['coral']*2
ax.errorbar(x_coords, y_coords, yerr=errors,
    ecolor=colors, fmt=' ', zorder=-1)

But I am not sure about the order. Is it the Reverse Alphabetical Order for hue?

I tried plt.errorbar as explained here, but of course didn't work

g = sns.FacetGrid(df, col="gene_short_name")
g.map(plt.errorbar, "Sample", "Average", yerr="CI")

ValueError: could not convert string to float: 'Sample 4'

Am I missing something? Any suggestion?
Thanks

[m-beau]

@mwaskom

It's redundant to use the ** operator and a dictionary, you can just do xerr=error. This won't work for nested bars, by the way.

Indeed, I cannot feed in something like yerr=[1,2,1,1] to a barplot displaying 4 bars for instance. Has this issue been addressed? It is very sad to be forced to use matplotlib to plot a barplot with pre-calculated errorbars.

[mwaskom]

That doesn’t sound sad at all. That sounds like using the correct tool for its intended purpose.

[m-beau]

Sorry but I do not understand, why wouldn't seaborn (or actually pyplot if I get it right) barplot function argument 'yerr' be the correct tool to plot custom nested error bars?

I simply meant that seaborn is much much nicer than matplotlib, hence it is sad to have to use matplotlib to get custom error bars, I was criticizing matplotlib rather than your great module :)

[loic001]

@mwaskom That doesn’t sound sad at all. That sounds like using the correct tool for its intended purpose.

My opinion is that seaborn is a (general purpose) visualization tool, it is not up to seaborn to calculate the confidence intervals. Yeah, it's practical to be able to do so and it is a plus, but it should above all be able to display the given data. No ?

[romanwerpachowski]

If the confidence interval is symmetric, the data can be massaged pretty easily to work around the lack of this feature:

data_up = data.copy()
data_down = data.copy()
data_up["value"] = data_up["value"] + data_up["error"]
data_down["value"] = data_down["value"] - data_down["error"]
data = pd.concat([data_up, data_down])
sns.barplot(x="category", y="value", hue="group", data=data)

And Bob's your uncle.

[mwaskom]

I would really encourage everyone who wants to add errorbars from prespecified error values to use matplotlib, a tool that handles this use case well, rather than trying to hack functionality from seaborn that is intended to do something different.

[mwaskom]

#2407