Closed mattwarkentin closed 2 years ago
In your implementation of the "joint" model, the covariate representing the transition only affects the Weibull scale parameter and not the shape parameter. So it will make the accelerated failure time assumption, and assume the shape is the same for every transition. Your "separate" models will have different shape and scale parameters per transition. So to mimic that in a single model fit, just do Surv(time, status) ~ trans + shape(trans)
or add anc = list(shape = ~trans)
.
Thanks for the clarification, @chjackson. Much appreciated.
Hi @chjackson,
I have a pretty simple multi-state model where participants start in a healthy state and can only transition into one of two states (death from cause-of-interest or death from any other cause). Pretty simple competing risks setup. I have read through your multi-state vignette (and some other papers), and I understand that you can either fit a single model call for the joint likelihood, if you restructure/stack your data, or you can fit separate cause-specific models and then combined them to form the full multi-state model.
I am wondering if I have done this correctly in the toy example below. I am trying to obtain the "same" prediction for each approach (i.e. five-year risk of death due to cause-of-interest), but I am not sure if this is correct. Hoping to get your insights. For the joint likelihood approach, I am not confident I have stacked the data into the necessary format, and I am not sure if I am making predictions the correct way using this joint model.