Akaike information criterion for Cox proportional hazard models

[Fijoy Vadakkumpadan]

Dear all,

I am conducting an analysis of survival data using Cox proportional hazard (CPH) models, to figure out what is the best model to use. The models I am comparing are non-nested. My plan is to compute the Akaike information criterion (AIC) for each of the models, and identify which models are the contenders for the best model, based on deltaAIC < 2. Here, I am using the definition of AIC corrected for small samples, i.e., AIC = -2*logl + 2k + 2*k*(k+1)/(n-k-1), where logl is the log-likelihood, n the sample size, k the #predictors.

While I have seen multiple published studies that follow the above approach to identify the best CPH model, some publications question the use of AIC to evaluate CPH models, since these models are not fully parameterized.

Can anyone please confirm/clarify whether it is acceptable to use AIC in evaluating CPH models?

Thank you,
Fijoy

P.S. A paper that uses AIC in CPH model selection is available here: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1774402/
A paper that questions the use of AIC for CPH models is available here: http://feb.kuleuven.be/public/ndbaf45/papers/CoxpaperHjortClaeskens.pdf

[David Duffy]

Post by Fijoy Vadakkumpadan
Dear all,
I am conducting an analysis of survival data using Cox proportional
hazard (CPH) models, [...]
My plan is to compute the Akaike >information criterion (AIC)
While I have seen multiple published studies that follow the above
approach to identify the best CPH model, some publications question
the use of AIC to evaluate CPH models, since these models are not
fully parameterized.
Can anyone please confirm/clarify whether it is acceptable to use AIC
in evaluating CPH models?

It's the particular reviewer your paper ends up with that you should worry
about ;). Perhaps find a parametric survival model that does as good a job
(easy in my experience) and show the AIC orders the models the same way?

Just 2c, David Duffy.