I was reading Frank’s post about CPMs and became curious about how to employ CPMs in competing risks contexts, and multi-state modeling more generally. What would the data setup be, and what would the model setup options be for:
CPM for the purpose of estimating conditional cause-specific cumulative incidence functions
… semi-competing risks context for marginal mean function estimation a la Ghosh-Lin
… general multi-state modeling setup
Thank you.
Please describe the ultimate goals of the analyses first, taking into account that cause-specific cumulative incidence is not an easy-to-interpret estimand. For example if you are analyzing time until cancer returns, with a competing risk from cardiovascular death, the cancer-specific cumulative incidence is the probability that cancer occurs before cardiovascular death. I’m not sure how to really digest that.
Likely not a helpful comment but for me as a clinician, the probability that a cancer recurs before cardiovascular death is an extremely relevant statistic that I would have no problem using in decision making. Most of my patients are 75 or older and most die because of CV disease long before their cancer recurs and for a large segment of the probability spectrum many would opt out of treatment. In fact many already do so on gut feeling already
The question is whether the following would be as or more relevant as that:
How would you treat cancer and cvd-death-related in CPM?
I mean, cvd and cancer are not different levels of the same outcomes. It’s not exactly what I have in mind when I think of death-illness model…
The underlying thought behind this is gaining some Quality-Adjusted Life Years that might be worth treatment-harm?
I hope that someone can think through this is more detail but my thought is that if a non-cancer death is sure to be a random event that is unrelated to the outcomes of interest, we can use an independent censoring assumption to terminate records for a patient who died from an unrelated cause at the last known time alive.
That’s a big “If” 
I think that both independent-censoring and enforcing different outcomes on the same scale are far from perfect heuristics - specially when facing comorbidities.
For prediction models the best heuristic I can think of is trying several approaches as a sensitivity analysis: Competing, Independent-censoring, Ordinal outcome or just plain composite.
I don’t catch the “different outcomes” reference. Censoring on a non-related outcome does not treat it as an outcome. Please elaborate. Thanks.