Thank you very very much for this impressive package;
Within the CoxPHFitter I have observed that when returning the partial hazards, the documentation establishes that the predictions returned have their mean subtracted, this is at odds with the implementation of the CoxPH function itself, where that is not done and normalisation is left to the user.
Moreover, this is done in the case of one-hot encoded columns of categorical variables as well, where the mean subtraction has no meaning, or in other normalisation schemes that may have been selected by the user. The implementation and documentation should be changed.
While it is not a numerically-incorrect result due to the risk scores (at least in the categorical-only case) being shifted, so KMs and concordances are similar, it just looks jarring having negative risk scores with models of only categorical variables with positive associations for example.
Thank you very very much for this impressive package;
Within the CoxPHFitter I have observed that when returning the partial hazards, the documentation establishes that the predictions returned have their mean subtracted, this is at odds with the implementation of the CoxPH function itself, where that is not done and normalisation is left to the user.
Moreover, this is done in the case of one-hot encoded columns of categorical variables as well, where the mean subtraction has no meaning, or in other normalisation schemes that may have been selected by the user. The implementation and documentation should be changed.
While it is not a numerically-incorrect result due to the risk scores (at least in the categorical-only case) being shifted, so KMs and concordances are similar, it just looks jarring having negative risk scores with models of only categorical variables with positive associations for example.