I am investigating how one blood biomarker (biomarker_1, measured at hospital discharge) predicts the longitudinal trajectory of another biomarker (biomarker_2) in a cohort of patients after discharge.
Data structure
-
Every patient has baseline measurements of both biomarkers at hospital discharge (time = 0).
-
During follow-up, biomarker_2 is only measured again at the time of revisit/rehospitalization → observation times are irregular and patient-specific. Biomarker_1 is not remeasured.
-
Most patients have multiple revisits/rehospitalizations and therefore several follow-up measurements.
-
Some patients die during follow-up (informative dropout/truncation by death).
The timing of biomarker_2 measurements could be informative (higher/lower biomarker_2 may trigger rehospitalization) and censoring due to death could also be informative.
We decided to use a joint modelling approach with three submodels linked by shared random effects:
R:Package, JMbayes2 jm(….,reccurent=”calendar”)
-
Linear mixed-effects model for the longitudinal trajectory of biomarker_2
-
Cox model for outcome revisit (recurrent)
-
Cox model for outcome death
Specific questions
-
How should the two survival submodels (rehospitalization and death) be adjusted for covariates? Some papers include no covariates in the survival part (only the shared random effects and the predicted value or slope of biomarker_2). Is it acceptable/statistically sound to leave the survival submodels unadjusted for classic risk factors (age, sex, comorbidities, etc.), or should we adjust them the same way we would in a “stand-alone” Cox model for death or rehospitalization?
-
For the longitudinal mixed-effects submodel of biomarker_2: My interest is predictive (does baseline and/or time-varying biomarker_1 predict a different trajectory of biomarker_2?), not explanatory. Therefore, should I include in the mixed model only the covariates that are known from the literature to strongly predict biomarker_2 levels, right?