I came across a recent publication by Touraine et al in BMC Medical Research Methodolgy
Touraine C, Cuer B, Conroy T, Juzyna B, Gourgou S, Mollevi C. When a joint model should be preferred over a linear mixed model for analysis of longitudinal health-related quality of life data in cancer clinical trials. BMC Med Res Methodol 2023;23:36. When a joint model should be preferred over a linear mixed model for analysis of longitudinal health-related quality of life data in cancer clinical trials | BMC Medical Research Methodology | Full Text.
The paper is quite interesting as it describes the analysis of quality of life data in RCTs where the response is affected by dropout due to death. They demonstrate that that a Joint modelling approach is less likely to produce biased estimates. In the paper they have simulated a joint model where patients who die due to cancer are handled.
This is a very new topic to me but based on the simulation results studied as well as the analysis of the trial results they have done it seems this is a better method to use. I would like the opinion of the statisticians here who have experience with longitudinal patient reported outcomes (quality of life for example) using such modelling techniques. The particular questions I have are:
- How does this handle interim missing data where missing data may be due patient declining to complete assessments at particular time points but filling the assessment later on (or say missing a followup visit).
- In the simultation study, the authors have adopted death as the simulation was done based on a trial in palliative care setting. However when we are dealing with curative treatment trials recurrence of disease is an important factor that affects completion of QoL. Is this method equally valid if we use another time to event endpoint like disease free survival (where disease recurrence is also included in addtion to death) ?
- Finally any guidance to good reading material to understand the modelling assumptions for such models and how to correctly interpret the model with limitations.
I’m still a little concerned that the longitudinal part cannot be easily interpreted if mortality is higher. But I’d like to be educated more about that. Multi-state transition models are the models that most elegantly handle absorbing states that prevent further follow-up assessments, so I’m interested in getting a comparison of that approach with the joint model approach.
The joint model approach may not generalize to more-than-binary events.
Is there any guidance on the use of multi-state transition models w.r.t patient reported outcomes ? I am curious as most of the patient reported outcomes have multiple scales with scores for each scale (e.g EORTC QLQ C30 provides scores for 5 functional scales, 7 symptom scales and 1 global health / quality of life scale). The scores are standardized such that they range between 0 - 100 where 100 means the worst symptom or the best functional outcome. I would like to thank you for the discussion we had on a related topic here but will a multi-state transition model be a better choice for trial design and analysis ?
There are other topics that have been started for this, e.g. https://discourse.datamethods.org/t/presenting-models-of-multiple-potentially-recurring-outcomes. Please move this new question there.
Great questions. I wonder if investigators may be over-anticipating poor adherence rates for QoL PROs? Unlike the response we have to surveys pushed on us all, patients facing serious disease often welcome completing PROs as a way to communicate with the study doctors on a high level. An added incentive is that timely reporting (especially e-PROS) can help guide their supportive care - such as the need for dose adjustments.
Related: Strategies to improve patient-reported outcome completion rates in longitudinal studies Strategies to improve patient-reported outcome completion rates in longitudinal studies - PMC
Regarding the impact of mortality on QoL assessments, if the difference in survival is significant would that not end the study? My thinking is that QoL differences matter most to patients when deciding which roughly equivalent (it’s a coin flip) therapeutic to choose as a management approach (palliative care) where time to relapse is the expectation … and a survival difference is not expected to be demonstrated.