Lifelines

New features
- Some performance improvements to regression models.
- lifelines will avoid penalizing the intercept (aka bias) variables in regression models.
- new `utils.restricted_mean_survival_time` that approximates the RMST using numerical integration against survival functions.

API changes
- `KaplanMeierFitter.survival_function_`'s' index is no longer given the name "timeline".

Bug fixes
- Fixed issue where `concordance_index` would never exit if NaNs in dataset.

New features
- model's now expose a `log_likelihood_` property.
- new `conditional_after` argument on `predict_*` methods that make prediction on censored subjects easier.
- new `lifelines.utils.safe_exp` to make `exp` overflows easier to handle.
- smarter initial conditions for parametric regression models.
- New regression model: `GeneralizedGammaRegressionFitter`

API changes
- removed `lifelines.utils.gamma` - use `autograd_gamma` library instead.
- removed bottleneck as a dependency. It offered slight performance gains only in Cox models, and only a small fraction of the API was being used.

Bug fixes
- AFT log-likelihood ratio test was not using weights correctly.
- corrected (by bumping) scipy and autograd dependencies
- convergence is improved for most models, and many `exp` overflow warnings have been eliminated.
- Fixed an error in the `predict_percentile` of `LogLogisticAFTFitter`. New tests have been added around this.

New features
- lifelines is now compatible with scipy>=1.3.0

Bug fixes
- fixed printing error when using robust=True in regression models
- `GeneralizedGammaFitter` is more stable, maybe.
- lifelines was allowing old version of numpy (1.6), but this caused errors when using the library. The correctly numpy has been pinned (to 1.14.0+)

New features
- New univariate model, `GeneralizedGammaFitter`. This model contains many sub-models, so it is a good model to check fits.
- added a warning when a time-varying dataset had instantaneous deaths.
- added a `initial_point` option in univariate parametric fitters.
- `initial_point` kwarg is present in parametric univariate fitters `.fit`
- `event_table` is now an attribute on all univariate fitters (if right censoring)
- improvements to `lifelines.utils.gamma`

API changes
- In AFT models, the column names in `confidence_intervals_` has changed to include the alpha value.
- In AFT models, some column names in `.summary` and `.print_summary` has changed to include the alpha value.
- In AFT models, some column names in `.summary` and `.print_summary` includes confidence intervals for the exponential of the value.

Bug fixes
- when using `censors_show` in plotting functions, the censor ticks are now reactive to the estimate being shown.
- fixed an overflow bug in `KaplanMeierFitter` confidence intervals
- improvements in data validation for `CoxTimeVaryingFitter`

New features
- Ability to create custom parametric regression models by specifying the cumulative hazard. This enables new and extensions of AFT models.
- `percentile(p)` method added to univariate models that solves the equation `p = S(t)` for `t`
- for parametric univariate models, the `conditional_time_to_event_` is now exact instead of an approximation.

API changes
- In Cox models, the attribute `hazards_` has been renamed to `params_`. This aligns better with the other regression models, and is more clear (what is a hazard anyways?)
- In Cox models, a new `hazard_ratios_` attribute is available which is the exponentiation of `params_`.
- In Cox models, the column names in `confidence_intervals_` has changed to include the alpha value.
- In Cox models, some column names in `.summary` and `.print_summary` has changed to include the alpha value.
- In Cox models, some column names in `.summary` and `.print_summary` includes confidence intervals for the exponential of the value.
- Significant changes to internal AFT code.
- A change to how `fit_intercept` works in AFT models. Previously one could set `fit_intercept` to False and not have to set `ancillary_df` - now one must specify a DataFrame.

Bug fixes
- for parametric univariate models, the `conditional_time_to_event_` is now exact instead of an approximation.
- fixed a name error bug in `CoxTimeVaryingFitter.plot`

I'm skipping 0.21.4 version because of deployment issues.

New features
- `scoring_method` now a kwarg on `sklearn_adapter`

Bug fixes
- fixed an implicit import of scikit-learn. scikit-learn is an optional package.
- fixed visual bug that misaligned x-axis ticks and at-risk counts. Thanks christopherahern!