Evaluating Assumptions of the Accelerated Failure Time Model

[pete-trenk]

My understanding is that the primary assumption in the AFT model is the baseline survival distribution follows some specified distribution (e.g.weibull, lognormal, loglogistic).

What methods can we use to inspect this assumption?

I've come across several mentions of plotting some sort of log transformation of the kaplan-meier estimate, derived from a particular covariate, against log(time) to inspect for a linearity (straight line). For instance, this paper from the The Journal of Risk Management and Insurance suggests that “The AFT assumption holds for the Weibull model if a plot of log[−log S(t)] against log(time) yields a straight line, where S(t) is the KM survival estimate.” This databricks notebook also uses a similar approach for inspecting the validity of the loglogistic aft where the plot of log(1 - S(t)) / S(t) against log(time) should be a straight line - again S(t) is the KM estimate of a particular covariate.

The lifelines documentation, Selecting a parametric model using QQ plots, seems to allude to a similar strategy:
“This graphical test can be used to invalidate models. For example, in the above figure, we can see that only the log-normal parametric model is appropriate (we expect deviance in the tails, but not too much). Another use case is choosing the correct parametric AFT model.”
However, the API for these plots only support univariate models (AFT models aren’t supported).

Do the above-mentioned approaches sound reasonable? Or is there a reason these methods aren’t supported in lifelines?