DEPR: Some dropna behaviors in DataFrame.pivot_table

[rhshadrach]

Currently dropna is used in four places within DataFrame.pivot_table:

1. It takes the cartesian product of all index/column levels when there are multiple levels; this was the original use
2. It is passed through to groupby
3. After the groupby aggregation, any rows that are all null are dropped
4. When computing the margins, rows in the original data where the keys and values are all null are dropped

1, 2, and 4 were all implemented for crosstab, which is essentially a call to pivot_table.

The API docs for crosstab document the dropna argument as:

Do not include columns whose entries are all NaN.

The only other documentation in the API and User Guide mentions using dropna=False to include rows/columns for categorical data with missing categorical values.

I think this is too much for a single Boolean argument to handle. I propose the following:

a. Add cartesian_product=[True|False] to pivot_table and crosstab
b. Add observed=[True|False] to crosstab for use with categoricals
c. Deprecate behavior (1) (with dropna), (3), and (4) above. The user may do each of these by dropping null values from the input data if they so desire.

We can implement (c) without affecting the behavior of crosstab by changing the data there to be a mixture of null/non-null values depending on the input and using the aggregation count instead of len.

[rhshadrach]

Instead of adding cartesian_product, we could make observed=False take the cartesian product of both the index and the column, regardless of whether they are categorical. In my opinion, this would only be okay to do if #55261 gets implemented.

Doing this would mean more than just passing observed=False through to groupby - that would only handle the index. We would still need to take the cartesian product of the columns in the case they are a MultiIndex as is done with dropna=False today.

If we're going this route, then I think we should also adhere to groupby semantics with unobserved groupings for various ops. For example:

df = pd.DataFrame(
    {
        'idx1': 1,
        'idx2': [2, 3],
        'col1': 4,
        'col2': [5, 6],
        'val1': [7, 8],
        'val2': [9, 10],
    }
)
df.pivot_table(index=['idx1', 'idx2'], columns=['col1', 'col2'], values=['val1', 'val2'], aggfunc='sum', dropna=False)

currently results in NaN values in various locations; instead with observed=False it should be 0 because the specified aggfunc is sum.

[rhshadrach]

Looking at this again - I think we should also deprecate (2). This can be done by the user as it's just a matter of dropping NA values from the input data.

[tehunter]

+1

As a user, I've learned I have to avoid .pivot_table, especially with margins. I've had issues in the past where the margins didn't aggregate up properly so there was disagreement with the "base" cells. I never had time to thoroughly debug and find where the issue was coming from, but after reading this, I suspect it might be related to the treatment of NAs.