Why does DataFrameGroupBy.boxplot method throw error when given argument "subplots=True/False"? - pandas

Question

I can use DataFrameGroupBy.boxplot(...) to create a boxplot in the following way:
In [15]: df = pd.DataFrame({"gene_length":[100,100,100,200,200,200,300,300,300],
...: "gene_id":[1,1,1,2,2,2,3,3,3],
...: "density":[0.4,1.1,1.2,1.9,2.0,2.5,2.2,3.0,3.3],
...: "cohort":["USA","EUR","FIJ","USA","EUR","FIJ","USA","EUR","FIJ"]})
In [17]: df.groupby("cohort").boxplot(column="density",by="gene_id")
In [18]: plt.show()
This produces the following image:
This is exactly what I want, except instead of making three subplots, I want all the plots to be in one plot (with different colors for USA, EUR, and FIJ). I've tried
In [17]: df.groupby("cohort").boxplot(column="density",subplots=False,by="gene_id")
but it produces the error
KeyError: 'gene_id'
I think the problem has something to do with the fact that by="gene_id" is a keyword sent to the matplotlib boxplot method. If someone has a better way of producing the plot I am after, perhaps by using DataFrame.boxplot(?) instead, please respond here. Thanks so much!

Answer 1

To use the pure pandas functions, I think you should not GroupBy before calling boxplot, but instead, request to group by certain columns in the call to boxplot on the DataFrame itself:
df.boxplot(column='density',by=['gene_id','cohort'])
To get a better-looking result, you might want to consider using the Seaborn library. It is designed to help precisely with this sort of tasks:
sns.boxplot(data=df,x='gene_id',y='density',hue='cohort')
EDIT to take into account comment below
If you want to have each of your cohort boxplots stacked/superimposed for each gene_id, it's a bit more complicated (plus you might end up with quite an ugly output). You cannot do this using Seaborn, AFAIK, but you could with pandas directly, by using the position= parameter to boxplot (see doc). The catch it to generate the correct sequence of positions to place the boxplots where you want them, but you'll have to fix the tick labels and the legend yourself.
pos = [i for i in range(len(df.gene_id.unique())) for _ in range(len(df.cohort.unique()))]
df.boxplot(column='density',by=['gene_id','cohort'],positions=pos)
An alternative would be to use seaborn.swarmplot instead of using boxplot. A swarmplot plots every point instead of the synthetic representation of boxplots, but you can use the parameter split=False to get the points colored by cohort but stacked on top of each other for each gene_id.
sns.swarmplot(data=df,x='gene_id',y='density',hue='cohort', split=False)
Without knowing the actual content of your dataframe (number of points per gene and per cohort, and how separate they are in each cohort), it's hard to say which solution would be the most appropriate.