All Times EDT
Keywords: COVID-19, competing-risk analysis, censoring by deaths, estimand
Recent clinical trials for COVID-19 have focused on patients’ time to clinical improvement (TTCI, say) within a specified treatment duration (e.g. 28 days). Improvement is measured on an ordinal scale such as in the recent Remdesivir trials which used an ordinal 6-point clinical status scale of severity (e.g., taking into account degrees of supplemental oxygen support). In the Remdesivir trial clinical recovery was defined as a 2-point improvement from admission.
By comparing treatments on the basis of TTCI as a survival endpoint through 28-days, the analysis can capture important temporal differences that can be present (e.g., early recovery effects), as opposed to a binary endpoint such as the proportion of patients who are discharged alive by day 28 (a single-point in time analysis). An important aspect of TTCI, and a vital part of the effect of treatment, is how to evaluate deaths. A natural approach is to employ a competing-risk analysis (CRA) where death is a competing risk. In the Remdesivir trial deaths were censored at day 28 which effectively considers deaths as “non-responders” throughout day 28. However, this may not be the most efficient analysis in view of the more rigorous approaches that have been developed in the CRA setting (e.g., Fine-Gray Cox model). CRA approaches can also raise important, seemingly subtle, dynamics with regard to the underlying question we are trying to answer -- What is our target estimand?
In this roundtable we discuss the study design implementation (GS design aspects) and interpretation (viz-a-viz estimand) of CRA methods for TTCI in COVID-19 trials. Two key questions are: (1) What should be the target estimand (e.g., TTCI in the presence of death, TTCI as a latent “potential” outcome); and (2) Could other endpoints be more meaningful (e.g., recent proposals are the average of patients’ clinical status scores across follow-up, or the maximum level across follow-up)?