Do we need to assume some parametric model to utilize lifelines?

[ogencoglu]

Do we need to assume some parametric model to utilize lifelines? Also can one have an arbitrary code as the underlying assumed model (if-else, non-linear and possibly non-convex stuff - basically a black box)?

[pzivich]

No, there is support for both semiparametric and nonparametric models. For example, the Cox proportional hazards model is semiparametric, since no model is specified for the baseline hazard. Similarly, the Kaplan-Meier and Nelson-Aalen estimators are nonparametric estimators. Finally, you can also pair lifelines with other estimators. For example, you can pair the Kaplan-Meier with inverse probability weighting for causal effect estimation.

However, I think your question is trying to see if there is an option for machine learning. lifelines does support some penalized models (which you could manually pair with splines to get a GAM). To my knowledge, lifelines does not easily integrate support for arbitrary code for the model but someone might know better

[ogencoglu]

Thanks for the answer. Indeed that was my question: can I use a machine learning model or arbitrary code.

[CamDavidsonPilon]

A lot of survival analysis in lifelines involves a likelihood equation, and this is provided by starting with a parametric model. The likelihood equation is where one can define to the model that there is censoring (of any type), see chapter 2 here.

However, you can have a black box produce a likelihood (just make sure it output a non-negative value), use that in the censoring likelihood equation, and maximize this likelihood equation . That's kinda what I did with my experimental project lifelike.