Improve how model.cond works

[grst]

Description of feature

• The implementation for finding the baseline level for each variable is currently hacky. Need to reach out to the formulatic devs for input.
• The function should print out the actual contrast used for diagnostics

[const-ae]

Hi Gregor,

I was looking into getting the default values again and was wondering if the following works or if there was a reason why we dismissed it?

import formulaic
import pandas as pd
import numpy as np

data = pd.DataFrame({"celltype": ['A', 'A', 'B', 'C', 'C', 'C'],
        "condition": ['a', 'b', 'a', 'b', 'a', 'b'],
        "cont": [1, 2, 3, 4, 5, 6]})
formulaic.model_matrix('~ condition', data).model_spec.encoder_state['condition'][1]['categories']
data["condition"] = pd.Categorical(data["condition"], categories = ["b", "a"], dtype = "category")
formulaic.model_matrix('~ condition', data).model_spec.encoder_state['condition'][1]['categories']

[grst]

The reason was that if you use the syntax described here, it breaks:

>>> formulaic.model_matrix('~ C(condition, contr.treatment(base="b"))', data).model_spec.encoder_state['C(condition, contr.treatment(base="b"))'][1]['categories']
['a', 'b']

Of course, also something we could decide not to support initially, but it would be nice to have a robust way to achieve this.

[const-ae]

Ah yes. But I think it is fine to nonetheless use 'a' as a reference level. I think the reference level should be based on whatever the data contains and not which additional modifications the formula applies.

I think a helpful analogy is to consider a continuous covariate and formulaic.model_matrix('~ I(cont + 10)', data). The formula shifts all values up by 10, but that doesn't mean that we would move the continuous intercept to -10. Applying the same logic to the categorical covariates means that I think it's fine to just look at the categories.

[grst]

But it does change the design matrix accordingly?

[grst]

Also adding here that I think this function should print out the actual contrast vector used (maybe omit zeros).
It's already a dataframe, so something like

contrasts.T.loc[lambda x: x[0] != 0]

could work nicely.

[grst]

Closed by #36
Follow up in scverse/pertpy#612