Skip to main content

Table 2 Summary of treatment effect measures

From: Treatment estimands in clinical trials of patients hospitalised for COVID-19: ensuring trials ask the right questions

Treatment effect measure Explanation
Hazard ratio The hazard ratio provides a weighted average of the hazards across all follow-up time points. In some cases, this interpretation can be difficult to understand; in Fig. 1, the hazard ratio is 0.90, indicating some treatment benefit. However, there is no difference in events between treatment groups at 28 days, and the hazard ratio gives no indication of how much additional time is conferred through the intervention.
Risk difference at a specific time point A difference in percentage points (or risk or odds ratio) at a specific time point provides an overall measure of benefit within that time period. However, it does not take into account the timing of events within that time span, and so, its appropriateness will depend on whether trial objectives relate to the occurrence of an event within a time period, or altering the time until an event.
Difference in means or difference in restricted mean time A difference in means provides a measure of benefit across the entire distribution, while the difference in restricted mean time (commonly referred to as ‘restricted mean survival time’) provides a measure of benefit within a certain time period; for instance, in Fig. 1, the difference in restricted mean time is − 1.0 days, meaning that over a 28-day period, patients in the intervention group had on average 1 additional day before an event. The two measures will typically differ, with earlier restrictions typically leading to greater differences. When feasible, for outcomes such as the number of days in hospital/on a ventilator/on oxygen/in ICU, using the mean or a later restriction is usually more informative.
Difference in medians A difference in medians provides a measure of benefit seen at the midpoint of the distribution. Although this can be informative in some settings, it can also mask what happens in other parts of the distribution.