PairGrid and missing values?

[jseabold]

Should the dropna argument work for this and do pairwise dropping?

import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt

dta = pd.util.testing.makeMissingDataframe()
g = sns.PairGrid(dta)
g.map_diag(plt.hist)

[jseabold]

Workarounds for histogram/scatter.

def na_hist(x, *args, **kwargs):
    x.plot(kind='hist', *args, **kwargs)


def na_scatter(x, y, *args, **kwargs):
    dta = pd.DataFrame([x, y]).T.dropna()
    x, y = dta
    x = dta[x]
    y = dta[y]
    plt.scatter(x, y, *args, **kwargs)

Though it looks like I'm missing some things with these. E.g., xaxis labels and ticks.

[DizietAsahi]

this still does not work. Am I missing how dropna works?

> from matplotlib import pyplot as plt
> import numpy as np
> import pandas as pd
> import seaborn as sns
> sns.__version__
'0.7.dev'
generate an example DataFrame
> a = pd.DataFrame(data={
    'a': np.random.normal(size=(100,)),
    'b': np.random.lognormal(size=(100,)),
    'c': np.random.exponential(size=(100,))})
> sns.pairplot(a) # this works as expected
(...)
> b = a.copy()
> b.iloc[5,2] = np.nan # replace one value in col 'c' by a NaN
> sns.pairplot(b) # this fails with error 
                  # "AttributeError: max must be larger than min in range parameter."
                  # in histogram(a, bins, range, normed, weights, density)"
> sns.pairplot(b, dropna=True) # same error as above

[wehlutyk]

Any news on this or where it's going? 0.7.0 still has the same behaviour.

[Alsanis]

Is any progress on the problem?

[cbrnr]

In fact, the dropna argument seems to be misleading (or at least I haven't really figured out what it is supposed to do). IMO the expected behavior should be equivalent to the following line in the example by @DizietAsahi above:

sns.pairplot(b.dropna())

[DizietAsahi]

The problem with using
sns.pairplot(b.dropna())
is that b.dropna() drops the whole line, even though there are some pairs of values that could be plotted. NaNs should only be dropped at the level of each pair of values IMHO.

[cbrnr]

I agree. Or it might even be useful to specify the dropping strategy with the argument dropna (such as implemented in R, e.g. complete cases, pairwise, ...). In any case, the current implementation of dropna simply does not work at all with the default settings of having histograms on the diagonal (it does work with kdes though).