Timeline for Cox model vs. Fine-Gray: hazard ratios & predicted cumulative incidence under competing risks
Current License: CC BY-SA 4.0
12 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Aug 2, 2023 at 11:14 | comment | added | user167591 | I think I might have found the answer. In here: discourse.datamethods.org/t/…, @Frank Harrell stated: ."..state occupancy probabilities. For the case of a terminating event (absorbing state) such as death, this is equivalent to a cumulative incidence curve". So is it correct that if each competing event can only occur once, after which the subject is removed from the risk set, then: p(t) = state occupancy probability = fraction in state = absolute risk = cumulative incidence? (sorry for all the = signs I wanted to ensure I had this right!) | |
| Aug 2, 2023 at 10:55 | comment | added | user167591 | In sect. 3.1 of the vignette you cited ("Multi-state models and competing risks"), they did a multistate Cox model and then used survfit.coxph() to compute p(t) (state occupany probability?) from this joint fit. Its a bit unclear as the plot on p17 just labels the y axis as "PCM". Assuming this is probability of being in state PCM (= fraction in PCM as labelled in their other plots?), can we also call this "Predicted cumulative incidence" and then state in the caption that the estimates are conditional, that is, at specific values of the predictors in the multistate Cox model? | |
| Dec 1, 2022 at 17:54 | comment | added | user167591 | Thank you @EdM, I have posted my question here: stats.stackexchange.com/questions/597629/… | |
| Dec 1, 2022 at 13:57 | comment | added | EdM | @user167591 sorry, I'm not expert enough with either case/control studies or conditional logistic regression to provide a reliable answer. Consider posing a new question. | |
| Dec 1, 2022 at 13:32 | comment | added | user167591 | Hi @EdM, in your answer you stated: "censoring at other event types provides correct hazard ratios for Cox models". May I please ask if this is true if we do this kind of censoring (for other competing events) in a sampled nested case control dataset and subsequently fit a conditional logistic regression model? Im guessing that the resulting HR's will be valid estimates under a competing risk setting? | |
| Nov 26, 2022 at 16:01 | comment | added | user167591 | I see thanks @EdM! | |
| Nov 26, 2022 at 15:36 | comment | added | EdM |
@user167591 What's described there is stratification by institution, among which baseline hazards due to un-modeled predictors could vary. In that situation stratification might make sense for FG. But what if sex violated PH and was associated with outcome? If sex wasn't of primary interest you could stratify by it in a Cox model to get hazard ratios for other predictors. But would it make sense (even if technically possible) to stratify by sex for FG analysis in the MGUS-PCM example in the vignette? How would you interpret the results in practice?
|
|
| Nov 26, 2022 at 15:20 | comment | added | user167591 | why would stratifying make less sense in a FG model? A quick google search suggests its possible: pubmed.ncbi.nlm.nih.gov/21155744 | |
| Nov 26, 2022 at 14:19 | comment | added | EdM | @user167591 it's possible to try to deal with violations of assumptions of Cox models in the ways you note, although you can end up with so many parameters to estimate that you end up overfitting the data. The last sentence of your comment summarizes one important difficulty with FG modeling. A standard way of dealing with violation of PH in a Cox model, stratifying by an outcome-associated covariate that violates PH so you can get estimates of coefficients for covariates of interest, wouldn't seem to make much sense in a FG model. | |
| Nov 26, 2022 at 12:02 | comment | added | user167591 | Much appreciated as always @EdM! Unless I'm mistaken, I think we can account for violations of the assumptions you mention (additivity across covariates, linearity, and proportional hazards) using e.g. interactions between covariates, splines and time interaction effects respecively? I guess that you are referring to the Cox model in its basic form? Are you saying that its basically harder to get the model structure "correct" in the FG model due to the combined process you mentioned? | |
| Nov 25, 2022 at 12:57 | history | bounty ended | user167591 | ||
| Nov 23, 2022 at 19:14 | history | answered | EdM | CC BY-SA 4.0 |