pygformula does not support categorical predictors for covariates declared as "truncated normal"

[sagarsimha]

Hi,

I discovered that when a covariate is declared as "truncated normal", pygformula does not support a categorical predictor in its model signature. Is this a known issue? I can provide a minimal example if it helps

[sagarsimha]

Hi, could you kindly have a look at the above issue?

[ryan-odea]

Hello, Apologies for the delay, I have been on holiday.

Would you be able to provide a minimal working example?
Feasibly, it should support categorical variables - you would need to encode them as numeric first since the pass to pytruncreg does not include such an encoding step for modeling.

[LilJing]

Hi, thanks for posting this issue. Yes, categorical predictors can be supported, but not in the same way as in the non-truncated normal case (i.e., using the C(·) syntax in the model formula). This is because, when fitting truncated normal covariates, pygformula internally relies on the pytruncreg package, which does not handle categorical predictor in the same way.

In this setting, a categorical predictor can be incorporated in two ways:
(1) encoding it as a single numeric variable, or
(2) representing it using one-hot (dummy) encoding.

Users should carefully consider which encoding strategy is most appropriate. If the categorical variable has a natural ordering, encoding it as a numeric variable is often reasonable. For nominal categories without an inherent order, one-hot encoding is generally more appropriate. However, one-hot encoding introduces additional (time-varying) dummy covariates, which can substantially increase model complexity and computational cost - particularly when the number of categories is large.