Our proposed methodology for dealing with multivariate configurations heavily relies on the dichotomisation#3633#> of variables. The dichotomization process is interesting because it reduces the potentially large scale of the original data into a more schematic, compact result. But its major drawback is the crude, radical reduction of data range. Is it a too radical procedure? Is it within acceptable limits?
The method for answering this question is... entropy ! Simply because entropy is a measure of the quantity of information#3634#> contained in a variable. If we compare the entropy of the original scale with the entropy for the recoded variable, we will get how much information#3635#> is lost in the process.
This simple operation even allows to compare different recoding schemes. The investigator may wonder if a dichotomisation#3636#> or a trichotomisation#3637#> would be more efficient, or if dichotomisation#3638#> around the mean or the median would be better. The entropy comparison yields a sound guideline.