Possibility in additional functionality with effect modification

[alockhar]

Hey Robin!

Thank you very much for such a great package!

I had a question as I was working in AdjustedCurves (also considered this in your contsurvplot but realized it would make sense in AdjustedCurves) for method/plot functionality possibility.

One thing I am/have been interested in doing outside of comparing groups (let's say via G-computation/outcome approach) in a survival context in both complete-case or using multiply imputed data, is to consider plotting survival probabilities as a function of another covariate. In other words, I am also interested in generating my adjusted survival probabilities as a function of let's imagine a binary or even continuous covariate so I can estimate on the x-axis my covariate (say age) and then survival probabilities by group (or even difference/ratios of an effect). The output in the object is obviously the group(s) estimate(s) of interest but if using direct standardization and a model with interaction as input, it would be possible to generate in the output object the hypothetical covariate values per iteration. The issue would be computational time, obviously, but still possible.

[RobinDenz1]

Thanks for the suggestion and for your interest in both of my survival packages!

I will try to break this down a little. If your interest lies in plotting the effect of a continuous variable and a categorical variable at the same time, you can use the contsurvplot package as is already. Multiple functions support the group argument, which takes the name of a categorical variable (just in case you were not aware). Since only g-computation allows the dependency on continuous variables anways, I don't see the point of integrating this into adjustedCurves.

If your interest is to plot adjusted survival curves as a function of two categorical variables, you could use the adjustedCurves package with some small workarounds, depending on what you are interested in. If you want to plot the actual interaction, you can simply create one variable out of the two using interaction() and use that new variable as variable in regular adjustedsurv() calls. If your interest is in effect modification instead, you could stratify your data by the moderator variable and use one adjustedsurv() call for each strata, combining the output later.

Neither of the last two options are great, but it is possible. I agree that it would be nice to have some more options directly implemented in adjustedCurves, but I currently do not have the time to work on this, as I am focusing on the simDAG package and a yet-to-be-released package on time-dependent matching. A possible alternative is the marginaleffects package, which directly supports what you are asking for, albeit only with g-computation. I am currently supporting the author of that package to write a new vignette on this topic, which will explain this in more detail.

[alockhar]

Thanks so much, Robin! I may reconsider the contsurvplot for a binary by continuous possibility but given it does not handle MI objects I may do some thinking regarding your adjustedCurves possibility, though. Best regards.

[RobinDenz1]

No problem. I will keep this issue open as a possible future enhancement.