PS you should consider carefully if Likert-score data are actually continuous... Likert scores are often analysed as continuous but, more rigorously, you should treat them as ordinal data (unless quantitative properties have been previously demonstrated)

Hi Louise,

I do agree with most of what you've just said. When it comes to dichotomizing continuous variables, I would rather be careful when applying this simple statistical manipulation. As a PhD student who has gone through similar problems, I would recommend Your students to read the following papers before making their final decisions concerning analysis:

1. Fitzsimons, G. J. (2008). Death to dichotomizing. Journal of Consumer Research, 35(1), 5-8.

2. Spiller, S. A., Fitzsimons, G. J., Lynch Jr, J. G., & McClelland, G. H. (2013). Spotlights, floodlights, and the magic number zero: Simple effects tests in moderated regression. Journal of Marketing Research, 50(2), 277-288.

I hope this hint will be at least of little help to Your students. Wish them all the best in their studies ;-)

Best,

Lukasz, PhD candidate

For sure, it might be "powerful" to use continuous data rather than discretizing them (and, as you pinpointed, this would avoid cut-off issues), but under certain conditions: for example, your needs in term of sample size may become higher. And, in a few cases, it may actually be better to use discretized data, for instance if you have to stratify an analysis...

It's not totally "up to you" to choose if you'd better use continuous or discretized data, it depends at least on what you want to do, on the study design and on the characteristics of your data...

I think your intuition is correct here; as far as I know, among stats people it is nearly universally agreed on that continuous variables generally shouldn't be dichotomized. Splitting them up might be ok for the purpose of interpretation (for example, if A and B are both continuous variables and they have a significant interaction, often the easiest way to illustrate that is to plot the effect of B when A is high and separately plot the effect of B when A is low, or whatever), but in the stats themselves they should be treated as continuous. In addition to the references Lukasz pointed out, a Google scholar search also turns up tons more making similar arguments. (In fact, I was about to come here and drop a link to whichever paper I first saw making this argument, but now I actually can't even remember what it was, because I've seen so many more papers claiming it.) Bottom line, if you told your students to not to dichotomize when doing their analysis, I think this would be uncontroversial.

I also agree to what you said. I can not imagine any sensible reason to dichotomize (or discretize) quantitative data. No person with a sound mind would ever consider re-painting a color photograph in black-and-white before starting to think what the picture shows. It sounds funny in my ears that especially scientists just do this quite often!

This behaviour also indicates inconsitency in the experimental design and that the scientist has no idea about his research question. If it would be required to have dicotomous data to be analysed w.r.t. the research question - why on earth was the data not collected that way (by design, so to say)? If such a suf´dy desing made no sense, it might be that the dichotomization makes no sense either. If this would just be a "bad study design", it would equally be a "bad analysis style" do analyze the dichotomized data. I don't see any rescue...

PS you should consider carefully if Likert-score data are actually continuous... Likert scores are often analysed as continuous but, more rigorously, you should treat them as ordinal data (unless quantitative properties have been previously demonstrated)