Using the CDF in Graphical Modeling


Hey everyone,

I’m currently working on Survival Analysis and would particularly like to model the ability to accommodate for “censoring”. If for a while we forget about censoring, the I think this falls in domain of known ways of modeling. We can build a generative process (based on whatever covariates) which emits the Survival Time.

My general question would be - How do I create a generative process for censored observations?

In particular, is there a way to simulate “if” conditions in a Bayesian Network in Pyro?


this hasn’t been merged into the codebase yet but if your censoring is univariate you could do something like this:

would that cover your use case?


Oh interesting!

Can you verify if I’m understanding the process correctly here?

A censored distribution would return log(1-CDF) when beyond the censoring time and log(PMF) otherwise. (I’m assuming log is only for numerical stability reasons).

If this is correct, how does the variational guide look like in this case then?