Problem with predicting survival probabilities from stpm2 model with TVC in external data

[Anita Lynam]

I have built a prognostic model using stpm2 which includes a tvc (all covariates in model are centered, Sb variables are restricted cubic splines). The model is intended for clinical use which will involve generating survival probabilities at 2 and 5 years for individual patients.
Is it possible to calculate/predict survival probabilities in external data for a stpm2 model including a tvc?
I am having problems re-creating the tvc element of the model.

Many thanks in advance.
Kind regards
Anita

stset survtime, failure(iT==1) scale(12) id(ID)
stpm2 Pred_A_cent Sb1 Sb3 Sb2 Sb4 Pred_Hb_cent , tvc(Pred_Hb_cent) knotstvc(Pred_Hb_cent 5) scale(normal) df(3) noorthog

Coefficients from output table -
xb
Pred_A_cent -0.0223022 .
Sb1 -0.0196823
Sb3 0.0813105
Sb2 -0.0655064
Sb4 -0.0227443
Pred_Hb_cent 0 .0171128
_rcs1 0 .247728
_rcs2 0.14239
_rcs3 -0.3269927
_rcs_Pred_Hb_cent1 0. 0021463
_rcs_Pred_Hb_cent2 0.0003087
_cons -2.574488

[Paul Dickman]

Yes, it's possible to predict survival probabilities in external data for a stpm2 model including a tvc. Here's a link to an example doing just that.

http://pauldickman.com/software/stat...out-of-sample/

You haven't shown code illustrating your attempts at prediction, so I can't say what the problem is, but hopefully the linked example will help.

[Anita Lynam]

Thank you so much for the link Paul, very helpful. I can now externally validate my model. Would I need to follow the same process to generate probabilities if I wanted to incorporate the model into an app/website for use by doctors in clinical practice?

[Paul Dickman]

You could just code the formulas in the predict command into your website/app. Have a look at https://interpret.le.ac.uk/

I didn't do any of the coding (see the 'about' section for the names and contact details of those who did) but the predictions are based on the estimates from -stpm2-. Essentially, the -predict- commands are coded in javascript.

[Markos Valsamis]

Hi Paul Dickman , I am trying to make an out of sample prediction using a stpm2 model that has two tvc's, both factor variables (which I have made into dummies so stpm2 can run). However, following the link you gave, I fail to create the rcs for each of the dummies (presumably as they just have a 0 or 1 value). Is there a way around this?

[Markos Valsamis]

Specifically, I realise the issue is that, after using the model originally specified as, for instance,

stpm2 female age, df(2) tvc(female) dftvc(2) scale(h)

, when I have the out of sample dataset, I am trying to :

predict lp, xbnobase

in order to then calculate the c-index model performance metric. However, I get the following error code: "_rcs_female1" not found.