Joint models for longitudinal and time-to-event data

data analysis models
1

In joint models for longitudinal and time-to-event data, all the models that I have seen assume that there is at least one observation for each subject. I am also aware of the principal strata effects when the outcome is truncated by death. However, I could not find any longitudinal model when there are subjects with no observation because of death. I really appreciate it if you can help me find relevant papers.

1 Like
2

I hope that someone who is versed with joint models will answer. I personally would try to use a state transition model for this problem so that it is easy to (1) model death and (2) interpret the results. A discrete-time ordinal longitudinal model based on a Markov process is a good candidate for that. Multiple examples are here. State transition models are unsurpassed for handling absorbing states IMHO. No principal strata needed.

2 Likes
3

Thank you so much for the examples.

4

are you asking out of curiosity or you have data of this kind? if the latter i wonder if they were scheduled visits? and some small proportion did not make it to the first visit? or the data aren’t from a clinical trial?

1 Like
5

This is an observational longitudinal cohort study where almost 20 percent of subjects died before the first measurement.

6

in this open access paper see table 1 in the pdf file under ‘web material’:

they have 265 and 294 people in the two groups at baseline and a1c is first measured at 2 years where there are 181 and 214 people. I don’t know what proportion of this is mortality, but surely some … I’m too rushed to check the detail at the moment

edit: actually they are censored by time to first cvd event (the outcome), not mortality, but it’s perfectly analogous to your problem it seems

1 Like
7

Thank you so much. It seems that they have baseline A1c measurements for the cases (Figure 2). Frankly, the more I read the more I believe that joint models cannot handle this situation.

1 Like
8

With an incidence of death of 0.2 it is hard not to make death a formal part of the outcome. Longitudinal ordinal models can do just that.

1 Like
9

Yes, I am working on it. Thank you again for your great suggestions.