Failure with categorical variables run in pyfixest?

[donbowen]

Hi @vincentarelbundock

Tried to get this working, but failed.

from marginaleffects import hypotheses
import pyfixest as pf

data = pf.get_data() 
fit = pf.feols("Y ~ i(f1)", data=data)
pf.etable(fit)

# pick any two coefs to test
hypotheses(fit, f"Q({fit._coefnames[1]}) + Q({fit._coefnames[2]}) = 0")

---------------------------------------------------------------------------
ColumnNotFoundError                       Traceback (most recent call last)
<ipython-input-17-b451ce25fdde> in ?()
----> 1 hypotheses(fit, f"Q({fit._coefnames[1]}) + Q({fit._coefnames[2]}) = 0")

c:\Users\deb219\AppData\Local\anaconda3\Lib\site-packages\marginaleffects\hypotheses.py in ?(model, hypothesis, conf_level, vcov, equivalence, eps_vcov, joint, joint_test)
    106     eps_vcov=None,
    107     joint=False,
    108     joint_test="f",
    109 ):
--> 110     model = sanitize_model(model)
    111 
    112     if joint:
    113         out = joint_hypotheses(

c:\Users\deb219\AppData\Local\anaconda3\Lib\site-packages\marginaleffects\sanitize_model.py in ?(model)
     37 
     38     elif is_pyfixest(model):
     39         from .pyfixest import ModelPyfixest
     40 
---> 41         return ModelPyfixest(model)
     42 
     43     raise ValueError(
     44         "Unknown model type. Supported modelling packages include `statsmodels` and `pyfixst`. In addition, users can call `fit_sklearn()` or `fit_linearmodels()` to fit models using the Scikit-Learn and LinearModels modelling packages."

c:\Users\deb219\AppData\Local\anaconda3\Lib\site-packages\marginaleffects\pyfixest\model.py in ?(self, model, vault)
     13             "modeldata": modeldata,
     14             "formula": formula,
     15         }
     16         vault.update(cache)
---> 17         super().__init__(model, vault)
     18 
     19         # after super init & validation
     20         if hasattr(model, "_fixef"):

c:\Users\deb219\AppData\Local\anaconda3\Lib\site-packages\marginaleffects\model_abstract.py in ?(self, model, vault)
      8     def __init__(self, model, vault):
      9         self.model = model
     10         self.vault = vault
---> 11         self.validation()

c:\Users\deb219\AppData\Local\anaconda3\Lib\site-packages\marginaleffects\validation.py in ?(self)
     11         self.validate_response_name()
     12         self.validate_formula()
     13         self.validate_modeldata()
     14         self.vault.update(
---> 15             variables_type=get_type_dictionary(self.get_formula(), self.get_modeldata())
     16         )

c:\Users\deb219\AppData\Local\anaconda3\Lib\site-packages\marginaleffects\utils.py in ?(formula, modeldata)
    169         ]
    170         t_c = [pl.Utf8, pl.Categorical, pl.Enum]
    171         t_n = [pl.Float32, pl.Float64]
    172         t_b = [pl.Boolean]
--> 173         if modeldata[v].dtype in t_i:
    174             if modeldata[v].is_in([0, 1]).all():
    175                 out[v] = "binary"
    176             else:

c:\Users\deb219\AppData\Local\anaconda3\Lib\site-packages\polars\dataframe\frame.py in ?(self, key)
   1399         │ 2   ┆ 5   ┆ 3   │
   1400         │ 3   ┆ 6   ┆ 2   │
   1401         └─────┴─────┴─────┘
   1402         """
-> 1403         return get_df_item_by_key(self, key)

c:\Users\deb219\AppData\Local\anaconda3\Lib\site-packages\polars\_utils\getitem.py in ?(df, key)
    159     # Single string input, e.g. df["a"]
    160     if isinstance(key, str):
    161         # This case is required because empty strings are otherwise treated
    162         # as an empty Sequence in `_select_rows`
--> 163         return df.get_column(key)
    164 
    165     # Single input - df[1] - or multiple inputs - df["a", "b", "c"]
    166     try:

c:\Users\deb219\AppData\Local\anaconda3\Lib\site-packages\polars\dataframe\frame.py in ?(self, name, default)
   9035         try:
   9036             return wrap_s(self._df.get_column(name))
   9037         except ColumnNotFoundError:
   9038             if default is no_default:
-> 9039                 raise
   9040             return default

ColumnNotFoundError: "i" not found

[leostimpfle]

Hi @donbowen! Responding here because as far as I can tell this is not related to pyfixest but to the way categorical variables are processed by pymarginaleffects. There are two separate issues:

First, the fixest-style factor interaction operator i is currently not natively supported by formulaic which is used to parse formulas in pymarginaleffects (see here). This fails because formulaic doesn't know about i and parses this as a separate variable that needs to be present in the data. I believe that the R version marginaleffects implements custom parsing of the fixest interaction operator i here which avoids this problem. So we could potentially implement something similar in pymarginaleffects. However, I should also mention that there is a chance that formulaic might implement the i-operator natively which would render a custom parsing unnecessary: matthewwardrop/formulaic#265

Second, it looks to me like the categorical encoding fails if the encoded variables are not valid Python expressions. I've implemented a very simplistic fix in #239. As far as I can see, it basically boils down to two things:

• pymarginaleffects expects that no categorical variable is numeric (see here) but pyfixest uses formulaic's categorical encoding which handles the conversion internally. My adjustments here convert numeric values to categorical if needed instead of throwing an error.
• pymarginaleffects uses Python's eval to construct hypothesis tests which may break if variable names are not valid Python expressions (which is the case when using C(f1) in a formula). The adjustments here work around this.

I'm very new to pymarginaleffects so may miss a lot of context. @vincentarelbundock would love to get your feedback, in particular if there's something we should change on pyfixest's side. Thanks!

cc @s3alfisc

[vincentarelbundock]

Thanks a lot @donbowen for the report and @leostimpfle for the fix! I really appreciate that you both took some time to look into this issue.

I believe that version 0.5.0 on PyPI should solve this issue.

Please note that future development will take place on the marginaleffects monorepo here:

https://github.com/vincentarelbundock/marginaleffects