BUG: read_excel with multi-indexed column ignores index_col=None

[jorisvandenbossche]

From SO: http://stackoverflow.com/questions/34020061/excel-to-pandas-dataframe-using-first-column-as-index

@chris-b1 another one on the multi-index excel issues .. :-)

Small test case: content of excel file:

A	A	B	B
key	val	key	val
1	2	3	4
1	2	3	4

gives:

In [2]: pd.read_excel("test_excel_index_col.xlsx", header=[0,1], index_col=None)
Out[2]:
A     A    B
key val key  val
1     2    3   4
1     2    3   4

It's not super clear in the formatting of the dataframe, but the [1, 1] is the index and [A, key] are seen as the level names of the multi-indexed columns.

[chris-b1]

We don't support writing to this format (multi-index columns w/ no row index) because it's ambiguous on the way back in. But it seems reasonable to support reading it, I'll take a look.

[pheman]

I got the same problem

[chris-b1]

From #15180 (comment) - consider changing default of index_col to some kind of sentinel (e.g. 'infer') so that passing None means something.

[stephenrauch]

@chris-b1 @jreback I have a proposed fix here: stephenrauch@1204b31 Before I do all the docs and stuff I wanna make sure I am headed in the right direction. Thanks.

[stephenrauch]

Per review comment from @jreback here is proposed api for read_excel. It is the same as read_csv.

index_col : int or sequence or False, default None

Column (0-indexed) to use as the row labels of the DataFrame. If a
sequence is given, those columns will be combined into a MultiIndex.
If None (default), pandas will use the first column as the
index. If False, force pandas to not use the first column as the index
(row names).

Updates are here:

stephenrauch@9b37ff9

So if this proposed API looks ok, I will do the PR.

[chris-b1]

Thanks @stephenrauch, that api looks good to me, please open it as a PR. One subtlety for the docs, if not using a MultiIndex header, index_col=None won't always use the first column as an index, it takes an inference path based on the shape of the data (same as for read_csv IIRC). e.g.

In [31]: df = pd.DataFrame({'a': [1, 2, 3], 'b': [4, 5, 5]})

In [32]: df.to_excel('temp.xlsx', index=False)

In [33]: pd.read_excel('temp.xlsx')

   a  b
0  1  4
1  2  5
2  3  5

[stephenrauch]

it takes an inference path based on the shape of the data

@chris-b1 @jreback Are there any test cases (or maybe some other description) that show how index_col=None works for something other than the index in the first column? Either for the csv or excel cases? I am having trouble grokking what is meant by the shape of the data in this case.

[chris-b1]

Not sure csv actually works this way. But for current Excel behavior:

Sheet 1

parses a row index

In [297]: pd.read_excel('temp.xlsx', sheetname='Sheet1')
Out[297]: 
      a  b  c
row1  1  2  4
row2  1  2  4
row3  1  2  4

Sheet2

No row index

In [298]: pd.read_excel('temp.xlsx', sheetname='Sheet2')
Out[298]: 
   a  b  c
0  1  2  4
1  1  2  4
2  1  2  4

[chenweisomebody126]

Has this problem been solved? I met the same question.

[araespahan]

As of now I still see the same issue. when using multi headers with read_excel, pandas always assigns the first column as index.

[sdicker8]

Another vote for a fix. I'm running 0.22 and it seems so hackish to have to write the code to save the 'index' to a named column and then reset the index.

Pandas is a 'great' module - thank you -

[the-rccg]

Vote for index_col=False to fix this

[rileymcdowell]

Just encountered this issue and I am looking for a fix. @stephenrauch has the PR been made for this?

[WillAyd]

@rileymcdowell no PRs have been made but would welcome any if you are interested

[rileymcdowell]

I'll put together a PR from @stephenrauch's work in the next couple of days.

[rileymcdowell]

I've dug into this and ran into a decision point. Consider the following spreadsheet (Taken from the test suite).

Right now, this is interpreted by the read_excel function as a MultiIndex where sheet cells A1 and A2 (values c1 and c2) represent the level names of the MultiIndex.

MultiIndex(levels=[['bar', 'foo'], ['a', 'b']],
           labels=[[1, 1, 0, 0], [0, 1, 0, 1]],
           names=['c1', 'c2'])

The situation that brought me to find this github issue is that in this example, I expect cells A1 and A2 (values c1 and c2) to be interpreted as another set of levels in the MultiIndex. For example

MultiIndex(levels=[['bar', 'c1', 'foo'], ['a', 'b', 'c2']],
           labels=[[1, 2, 2, 0, 0], [2, 0, 1, 0, 1]],
           names=['', ''])

The test suite explicitly covers the existing functionality of the former. This behavior differs from that of the read_csv function, which behaves most similarly to the latter.

A workaround is to allow a sentinel value of index_col=False as suggested in this issue, but a change to support this would also touch some of the core logic in pandas.io.parsers. For example this check which limits non-null index_col settings to have numeric values. That seems like it could have far-reaching consequences.

Any thoughts about how best to tackle this?

[WillAyd]

The situation that brought me to find this github issue is that in this example, I expect cells A1 and A2 (values c1 and c2) to be interpreted as another set of levels in the MultiIndex. For example

Hmm well I disagree since this representation matches what you'd see with a normal data frame representation, but regardless of opinions I think it just speaks to what @chris-b1 mentioned earlier that this is really ambiguous so there's not necessarily a right answer

This behavior differs from that of the read_csv function, which behaves most similarly to the latter.

Can you clarify this with an example? CSV doesn't have the concept of a merged cell like you have with the foo and bar cells so I don't think this is apples-to-apples

A workaround is to allow a sentinel value of index_col=False as suggested in this issue, but a change to support this would also touch some of the core logic in pandas.io.parsers. For example this check which limits non-null index_col settings to have numeric values. That seems like it could have far-reaching consequences.

Isn't False a valid value for read_csv? If so then there's got to be something with that implementation that isn't limited by what you've found

Thanks for the investigation!

[ian-contiamo]

@rileymcdowell I would agree with @WillAyd that the former of the two behaviors you describe is the most intuitive way to interpret an Excel file. If you can I would encourage you to submit a PR with this behavior.