Exponential AFT Fitter

I used the WeibullAFTFitter to produce an ExpAFTFitter.

The code I’ve changed:

    def __init__(self, alpha=0.05, penalizer=0.0, l1_ratio=0.0, fit_intercept=True, model_ancillary=False):
        self._ancillary_parameter_name = "rho_"
        self._primary_parameter_name = "lambda_"
        super(ExpAFTFitter, self).__init__(alpha, penalizer, l1_ratio, fit_intercept, model_ancillary)

    def _cumulative_hazard(self, params, T, Xs):
        lambda_params = params["lambda_"]
        log_lambda_ = Xs["lambda_"] @ lambda_params

        return safe_exp(np.log(np.clip(T, 1e-25, np.inf)) + log_lambda_)

    def _log_hazard(self, params, T, Xs):
        lambda_params = params["lambda_"]
        log_lambda_ = Xs["lambda_"] @ lambda_params

        return log_lambda_

This produces numerically correct estimates, with wrong signs though.

However, it also produces some warnings and when printing the summary it spits out an error too. Obviously because of the rho. I tried to exclude rho, but then it started to whine about some _ancillary_parameter_name.

So, my suggestion is to include an ExpAFTFitter(). And why wouldn’t there be a single AFT function which would take the distribution as an argument?

For instance:

from lifelines import AFTFitter
aft = AFTFitter()
aft.fit(rossi_dataset, duration_col='week', event_col='arrest', distribution='Weibull')

[CamDavidsonPilon]

To do the Exponential AFT model, try something like: https://lifelines.readthedocs.io/en/latest/jupyter_notebooks/Custom Regression Models.html

That is the preferred way to extend AFT models.

The latter model, AFTFitter might be included in a future release (it’s something I’ve thought about), but it bucks the trend of models being equal to classes.

[sursu]

Yes, coefficients are identical in all 3.