This is the extremely common situation in which we convert a continuous measurement into a binary: “It’s hot today” rather than “It’s 40 degrees Celsius today”.
Details #
Well “clinical” versus “non-clinical” is commoner in our fields. There are understandable reasons why we might do this but there are essentially no circumstances in which it is not throwing away a lot of information so both for individuals’ data, and for analyses of aggregate data, we should be very wary of doing this. When analysing aggregate data the loss of power and precision as opposed to analysing the continuous scores can be quite severe.
The same issues apply when continuous scores are simplified to three, four or even five levels: trichotomisation, quadratomisation (?) and what I call quinchotomisation though the loss of power/precision gets less as the number of levels goes up.
Try also #
Aggregation
Clinically significant change (CSC)
Cutting points
Estimate/estimation
Precision
Reliable and clinically significant change (RCSC)
Reliable change index (RCI)
Statistical power
Chapters #
Not covered in the OMbook.
Online resources #
See my Rblog post (dichotomisation) for a simulation illustrating the issues.
Dates #
First created 25.xi.24, updated with link to Rblog post 4.xii.24.
