RMS Semiparametric Ordinal Longitudinal Model

I wanted to report a finding that slightly surprised me. I had the chance to reanalyze trial data from COV-BARRIER, which recorded the 8-level NIAID ordinal score for 28 days. What surprised me was that all the differences between the groups comes from mortality, not something we saw in ACTT-2 for example. There’s no difference in ventilation, high-flow oxygen etc. This results in the unexpected situation that P(Benefit) is larger if one just runs a logistic regression for death at 28 days, compared to the longitudinal ordinal model over the whole 28 days using all 8 levels. It’s probably quite unusual / unlucky, but made me more aware that including more levels can reduce power if we include levels which are not affected by treatment.