Which competing risks survival analysis to choose

#1
Hi guys and girls,

First of all, thank you for taking the time to read this and hopefully provide me with an answer. Context: I'm an MD, currently in my third year of my PhD in adult oncology.

I'm doing a study regarding treatment-related mortality. In this study, we have roughly 2000 patients included (5-year cohort) , all adults with cancer. In this population we want to look at factors that increase the chance of treatment-related mortality. I've done previous studies looking at overall mortality for which we used Cox regression analyses however in this case I don't think these will do. This is because of the fact that if you die from cancer, you can't die from the treatment thereafter and vice versa. Also for some cancers the portion of treatment-related deaths might be relatively high, but that is because the more intensive therapy leads to an overall better survival.

So from diving into the literature I've learned I can do a competing risks analysis, which take these incompatible events into account. Examples of interesting articles I came across are 3328633 and 4741409 (PMIDs). However there is a question that remains.

In all literature about these analyses, a distinction between cause-specific hazard models and subdistribution hazard models are made. From what I've understood, the latter is for more prognostic questions, the first for more etiological questions.

Am I correct to think that for our questions (e.g. does being underweight at diagnosis lead to an increased chance of dying from complications of treatment) we should use a cause-specific hazard model?

And if so, what model would suit this (type of) questions the best? We will include multiple factors which we think are risk factors (e.g. age, nutritional status, ethnicity, sex - thus some ordinal and some continuous) but also factors we want to correct for (e.g. intensity of treatment rating, expected 5 year survival based on diagnosis).

Any help would be greatly appreciated.

Cheers,

Franz
 

hlsmith

Less is more. Stay pure. Stay poor.
#2
Another issue you may run into, it seems, could be a "Well defined Treatment". It gets difficult to establish causality/association when you have multiple version of treatments (dosing) and incomplete compliance. I am slightly familiar with the types of models you are writing about, but not enough so as to answer your specific questions. :(