Can anyone give me an argument for or against using the complementary log-log transformation as opposed to the log-log transformation on survival estimates after cox regression in multiple imputation setting?

This Andrea Marshall paper recommends combining estimates on the complementary log-log scale: https://bmcmedresmethodol.biomedcentral.com/articles/10.1186/1471-2288-9-57

However this seems counter intuitive to me. Suppose

S(t) is the survivorship function, H(t) is baseline cumulative hazard, and X.Beta is the linear predictor from the cox regression.

Then S(t) = exp(-H(t)exp(X.Beta)), from standard survival analysis formulas.

So log(S(t)) = -H(t)exp(X.Beta)

log(-log(S(t)) = log(H(t)) + X.Beta

Which seems like it should have some nice properties. However if you take the complementary log-log, you are working with

log(-log((1-S(t))) = log(-log(1 - exp(-H(t)exp(X.Beta)))

Which doesn't have such a nice form.

Finally, the reference 24, which comes after the statement that complementary log-log transformation should be used, takes you to Hosmer DW, Lemeshow S: Applied survival analysis – Regression modeling of time to event data. 1999, New York: John Wiley & Sons. I have taken this book out of the library and on page 152 there is a section on "confidence interval estimation of the covariate adjusted survivorship function. It has a formula for the variance of the survival estimate after a log-log transformation. This contradicts the statement that it is referenced by (see pic).

enter image description here

enter image description here

Andrea Marshall is a far more accomplished statistician than I am, so someone please point out where I am mistaken! Not only do I need to combine the survival estimates, but also calculate associated standard errors, and so having the right transformation is essential.

Many thanks.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.