Seaborn: How to use a weighted mean estimator in seaborn factor plot (incl bootstrapping)?

[timalthoff]

[Note: I am reviving a stackoverflow question that I was unable to figure out with some new insights on how it might work. See: http://stackoverflow.com/questions/32771520/how-to-use-a-weighted-mean-estimator-in-seaborn-factor-plot-incl-bootstrapping]

I have a dataframe where each of the rows has a certain weight which needs to be accounted for in the mean computations. I love seaborn factorplots and their bootstrapped 95% confidence intervals but haven't been able to get seaborn to accept a new weighted mean estimator.

Here is an example of what I would like to do.

tips_all = sns.load_dataset("tips")
tips_all["weight"] = 10 * np.random.rand(len(tips_all))
sns.factorplot("size", "total_bill", 
               data=tips_all, kind="point")
# here I would like to have a mean estimator that computes a weighted mean
# the bootstrapped confidence intervals should also use this weighted mean estimator
# something like (tips_all["weight"] * tips_all["total_bill"]).sum() / tips_all["weight"].sum()
# but on bootstrapped samples (for the confidence interval)

The problem I have is that the estimator function only gets to see the "main variable" (y axis) instead of the full dataframe that would allow the estimator to access more than just "y".
See here: https://github.com/mwaskom/seaborn/blob/f4e777743a93b01ad0499c6f35c2a112161de6ad/seaborn/categorical.py#L1186

Is there any simple way to do this?

If not, what is the easiest way to extend the categorical plotting to allow for weighted estimators?

Thanks a lot,
Tim

PS: couldn't figure out labels. my guess is question and wishlist.

[mwaskom]

It's not really supported, but I think it is possible to hack together a solution. This seems to work?

tips = sns.load_dataset("tips")
tips["weight"] = 10 * np.random.rand(len(tips))

tips["tip_and_weight"] = zip(tips.tip, tips.weight)

def weighted_mean(x, **kws):
    val, weight = map(np.asarray, zip(*x))
    return (val * weight).sum() / weight.sum()

g = sns.factorplot("size", "tip_and_weight", data=tips,
                   estimator=weighted_mean, orient="v")
g.set_axis_labels("size", "tip")

[timalthoff]

Thanks Michael! It looks like this is working. I really appreciate the hacky but clever workaround.

[pkmn99]

mwaskom's solution works when "orient="v" is specified.
Otherwise, there will be error.
"TypeError: zip argument #1 must support iteration"
This is strange...

[adam-katz-rf]

Solution worked for me only when I added: tips["tip_and_weight"] = list(zip(tips.tip, tips.weight))

[MaozGelbart]

The above solution (which is very useful) works with v0.10.1, but not with v0.11.0 which now raises the following error:

 File "/Users/maoz/seaborn/seaborn/categorical.py", line 3714, in factorplot
    return catplot(*args, **kwargs)
  File "/Users/maoz/seaborn/seaborn/_decorators.py", line 46, in inner_f
    return f(**kwargs)
  File "/Users/maoz/seaborn/seaborn/categorical.py", line 3779, in catplot
    p.establish_variables(x_, y_, hue, data, orient, order, hue_order)
  File "/Users/maoz/seaborn/seaborn/categorical.py", line 156, in establish_variables
    orient = infer_orient(
  File "/Users/maoz/seaborn/seaborn/_core.py", line 1310, in infer_orient
    raise TypeError(nonnumeric_dv_error.format("Vertical", "y"))
TypeError: Vertical orientation requires numeric `y` variable.

MWE:

import seaborn as sns, numpy as np
tips = sns.load_dataset("tips")
tips["weight"] = 10 * np.random.rand(len(tips))
tips["tip_and_weight"] = list(zip(tips.tip, tips.weight))
def weighted_mean(x, **kws):
    val, weight = map(np.asarray, zip(*x))
    return (val * weight).sum() / weight.sum()

sns.pointplot(x="size", y="tip_and_weight", data=tips, estimator=weighted_mean, orient='v')