Objective: Mortality is frequently used as an outcome in critical care trials, being a patient-orientated variable and robust against information/selection bias. Mortality frequency, however, should be measured at a defined time point of follow-up. Practice of meta-analysis shows that follow-up times of trials in critical care medicine differ substantially. This may have substantial implications on potential pooling of effect estimates. We aimed to describe the current practice of mortality follow-up time definitions in a representative sample of published critical care randomized controlled trials and to analyze the influence of different follow-up times on subsequently pooled effect estimates.
Data sources: Cochrane CENTRAL, EMBASE, MEDLINE, PASCAL Biomed, and PsycINFO.
Study selection: Databases were searched for critical care randomized controlled trials published after 2000. A random sample of 50% was drawn for further review.
Data extraction: Study characteristics were extracted, as well as the number and time points of mortality ascertainment. Additional data were extracted from Kaplan-Meier plots, as available.
Data synthesis: Meta-regression and multilevel mixed-effects linear regression were used to analyze the influence of follow-up time (independent variable) on deviation of pooled risk ratios from study baseline (dependent variable). From 9,246 retrieved references, we included 106 studies representing 63,713 participants. Among these, 45 studies (43%) reported more than one time point, with 24 different time points at all, only three (28, 30, and 90 d) being reported in more than 10% of studies. Limiting meta-analyses to only one predefined time point would reduce the number of eligible studies by at least 60%. No influence of time points on meta-analytic summary effect estimates was found.
Conclusions: In a large sample of critical care randomized controlled trials, numerous different mortality time points are reported. Mortality time points did not influence pooled point estimates of the effects. Consequently, it seems possible to pool effect estimates, which in turn will increase the precision of these effect estimates.