Recurrent survival models

Modeling
1

Hi,

I was trying to implement survival model in stan but found that same patients (10%) have one or two recurrences. There are 250 patients with 5 time points per patient. The setting is orally transmitted infectious disease. All patients are in some area and communicate at some extent. 50 patients are treated and they do survive. Also, treated patients help to kill infection in untreated patients (mechanism unknown). The goal is to model survival of 200 untreated patients.Typical timeline for untreated patient is healthy-sick-death but there are occurrences of sick-healthy-sick or healthy-sick-healthy or even healthy-sick-healthy-sick. Can somebody suggest what model would be appropriate?

Thanks for any suggestions.

2

Hey @linas, I suspect the answer may end up being a relatively custom model. To better engage respondents, it may help to write down the model with all the factors that you can take into account and then rephrase the question in terms of adding additional factors and complications to the more basic model. I.e. here’s my model … it takes into account A,B,C but not D,E. How do I include D and E?

1 Like
3

Thanks. I agree that it may end up to be custom model. I really have no model yet so trying to find/build one.

Sorry, I have made a mistake. It appears only one patient gets sick and then becomes healthy again. So I took survival model from here. One problem I have encountered is with censoring. I defined time to event as time elapsed to get sick first time. For those patients that didn’t get sick during the study (and since we had only 5 time points per patient) censored time to event was 5. There were bunch of patients that didn’t get sick. Model didn’t converge at first but then I removed the line for censored patients

target += weibull_lccdf(ycen | alpha, exp(-(mu + Xcen_bg * beta_bg)/alpha));

and model converged. So the first question is how to handle censored patients because I am not comfortable disregarding data. Intuitively it means that for those patients time to event is more than 5.

Another question is how to include frequency of treatment of treated subjects. They were treated regularly (since they were medical doctors). One possibility is to include two covariates: medical doctor and frequency of treatment.

Thanks for any advice