The optimal cutpoint, at which the EFS difference between low versus high LDH was maximized, was sought. This was clarified in the following table (supplementary to this article)
I wonder if this is the best method to choose a cutoff; I went through the documentation of maxstat R package and I see it relies on “Maximally Selected Rank Statistics”.
Should I use this “Maximally Selected Rank Statistics” for getting the cutpoints for a marker ? Also is it possible to have more than one cutpoint and is there a way to get the number of possible cutpoints ?
In that particular setting (and all other settings I’ve ever seen) a cutpoint does not actually exist. That is, you can’t demonstrate that the relationship between X and Y is flat on either side of the cutpoint. So any attempt to find a cutpoint is futile. It may be better to move this discussion to https://discourse.datamethods.org/t/categorizing-continuous-variables
Cutpoints do not work correctly for either of those two applications. As one simple example, a patient whose value is right at the cutpoint should not be pooled with other patients far from the cutpoint, and should be classified as “we don’t know; get more data”. Equivocal values lead to tentative decisions that are revisited. If you use a cutpoint you’ll forget to revisit shaky decisions. Think about a track coach who is testing runners in preparation for the olympics. Does her stopwatch read “fast/slow” or does it read in seconds?
In principle I agree with this. But I also know, through experience and market research, that in a number of clinical applications users/physicians essentially demand cutpoints. So that’s a dilemma facing companies developing diagnostic products. No straightforward resolution, to my knowledge
We need to insist on not using things that don’t exist. Everyone finds different cutpoints. The lack of replication comes from seeking something that didn’t exist in the first place.
