Predicting the survival function for uncensored test data?

When we predict the survival function for a given data point (this data point is in my test set and is uncensored), do we use the conditional_after argument? To me, it doesn’t make sense to because when I evaluate, I will take the median percentile as the predicted duration and measure that to the ground truth duration. The predicted duration shouldn’t be dependent on the ground truth duration, correct? Or am I thinking about this wrong?

[jlim13]

Thanks for the prompt reply! So, the docs don’t directly say to set conditional_after = 0 when predicting uncensored data. I just quickly looked through the docs just now, so I could have missed it!. I wouldn’t say they are confusing, but the docs don’t address the various questions surrounding the topic of evaluating on censored data. For example, I want to get a sense of how close my model can predict the actual duration in days. To evaluate this, I would have to get the L1-Loss for my predictions on the uncensored data in my test set and their ground truth durations. Maybe a section about doing this type of evaluation of uncensored data would make it clear to not use conditional_after on uncensored subjects? Thoughts?

[konradsemsch]

Agreed! doing right now 😉