More or less, I think I understand the subtleties of BBR. I was wondering if it would be a good idea to use the rms capabilities, both the contrast function and the graphs, to analyze the study with a frequentist approach, as described in the notes, and then do another Bayesian analysis as a complimentary analysis, in the same article (to report posterior probabilities under the prior).

I wanted to analyze the effect of a new type of chemotherapy on gastric cancer. There is a previous clinical trial, with an HR 0.77, 95% CI, 0.65-0.99. I could give more details if someone is interested, for now better leave it simple.

I have an observational registry with n > 1500 and I would like to assess if this main effects differs according to some factors (mainly histology). I suspect that might happen and nobody has checked before.

The formula to use for main effects would be something like this: cph (surv ~ new_treat + histology + rcs(age) + rcs(year)

Then for interactions one by one, new_treat * histology, etc. would be introduced. And I use the contrast function.

The notes start with the phrase “Anything resembling subgroup analysis should be avoided”, then I no longer dare to present a subgroup effects table, but I will show the graphs after applying the contrast function, as in the examples. In addition to histology, other interactions could be with tumor extent, age and year of treatment (four in total).

For the Bayesian part I had thought to use this previous knowledge in the literature (HR 0.77) as a prior of the main therapeutic effect. After that I wanted to analyze the effect according to histology, since I have good reasons to think that not all histological types respond equally to treatment. To model the interaction in the bayesian model I could use a relatively skeptical prior (for the interaction term), and report the posterior probability that the new treatment is superior to each group according to histology. But this could be done in parallel to the frequentist analysis, to use the full power of the rms package.

There are two things about BBR notes that I don’t understand.

- Need to use penalized estimation (e.g., interaction effects as random effects) to get sufficient precision of differential treatment effects, if # interaction d.f. > 4

I don’t know how to apply this.
- I also don’t understand the chunk test with all potential interaction R effects combined. I don’t know how this would be done.