The more I've thought about it, I've decided that I'm largely okay with the idea of composite endpoints, but like any methodology choices, it has to be an
a priori decision in the study design, not something thrown together to salvage a study. A large part of this conclusion comes from my pediatrics background in which no study is ever large enough to actually garner any real confidence in it's apparent truth (N of 1 studies of parachutes notwithstanding). 500 subjects in a peds study gets categorized as
"HUGE!" and the number of promising, important studies that have shut down due to inability to recruit patients is high, even moreso in the PCCM literature. With a baseline PICU mortality of about 3%, no one ever uses mortality as primary outcome, and yet when I've spoken to adult colleagues I've seen what I thought were good studies with important secondary outcomes get smashed to bits because the mortality didn't change. So I'm primed due to my training and my paradigm to look for alternative means of guiding clinical decision making.
But in accepting a composite endpoint, it's clear that you have to really parse out the effects and your frame of reference, and to understand that it's not a scientific "truth".
@jdh71 's use of the word "signal" is quite appropriate in this context. Highly recommend Nate Silver's book
The Signal and the Noise if anyone wants some easily digestible statistics discussions.
To be clear, endpoints such as all cause death and ventilator free days as primary endpoints could readily be argued are composite endpoints, so let's not pretend this is the first time data points have been lumped together.
