Cox models and rare events

[tombisho]

A recent paper by Schumie et al (doi:10.1177/09622802211060518) combines results from time to event data but with the likelihood function approximation at each site not necessarily being normal. According to the authors this is suitable for rare events where traditional cox might not give any result in a specific site if no events in one of the groups. The method is implemented in the R package EvidenceSynthesis. Each site needs to use the package, return the aggregated parameters to the analyst who then can use the same package for the meta analysis. In other words, the traditional way of doing meta analysis.

The beauty of the dsSurvival package is that you can perform the Cox regressions in a federated way using DataSHIELD. Although in case of rare events you might have trouble estimating Cox, not only because of disclosure options set in Opal but also because Cox will not give an answer. It might be something to consider using as an option in the dsSurvival package - merging the federated approach you use with a likelihood approximation Schumie et al use.

[tombisho]

Not urgent but an idea from a collaborator