Competing events and in-hospital follow-up: issues to be aware of

Session Type
Core outcomes
Sandercock J1, Geneen LJ1, Kimber C1, Lamikanra AA2
1Systematic Review Initiative, NHS Blood and Transplant, Oxford, UK
2Clinical Directorate, NHS Blood and Transplant, Oxford, UK

Background: In our review of convalescent plasma or hyperimmune immunoglobulin for severe respiratory viral infections, with 30 RCTs, several outcomes were complicated by competing events and in-hospital follow-up. Survival: Only one trial (RECOVERY) of seven reporting a hazard ratio (HR) explicitly stated that individuals discharged before the end of 28-day follow-up were censored at the end of follow-up. A naïve survival analysis would censor at discharge (in the absence of post-discharge follow-up), inappropriately removing survivors from the denominator at later timepoints. This outcome was meta-analysable as relative risk, which is adequate for short follow-up with very little loss to follow-up. Time to mechanical ventilation (MV): Death is a competing event for MV, usually accounted for by defining the event as “MV or death”. Only one trial (RECOVERY) of the two which reported an HR described their approach to censoring those discharged before the end of follow-up. Some trials reported duration of MV without describing their methods, making results uninterpretable because patients are removed from MV due to either improvement or death. Some trials included individuals already on MV, further complicating this outcome. REMAP-CAP accounted for this using a Bayesian cumulative logistic model to analyse multiple states. This is a useful way to compare the two groups but the resulting odds ratio (and medians) are hard to interpret. Length of hospital stay (LoHS): Here, the competing event (death) is diametrically opposed to the event of interest and cannot be accounted for by combining events. RECOVERY was the only trial to state that people who died in hospital were censored at end of follow-up rather than on the date of death. Most trials reported median LoHS without describing any methods, making their results uninterpretable.
Conclusions: Reviewers need to be aware of the problem of competing events and in-hospital follow-up, and avoid combining results obtained using different (and frequently unspecified) methods of analysis. There will often be insufficient data reported to reanalyse for meta-analysis. Core outcome sets give little or no guidance on analytical methods for trialists; parallel initiatives on core analytical methods would be useful.
Patient, public and/or healthcare consumer involvement: none