In their book Multivariate Data Analysis, Hair et al. have suggested to convert a a metric moderator variable in to a categorical one as the approach of using interaction term is complicated by many factors if the said variable is measured by multiple items.
First, a caveat! Converting a metric or continuous variable to a categorical variable will ALWAYS result in a loss of information. However, sometimes it expedient to do so in order to simplify the analyses. A second reason may be the reliability of the original metric was not adequate enough to justify that level of precision.
If you decide it is appropriate or necessary to do so, the simplest method to convert a metric to a categorical variable is to (1) determine the number of categories you wish to have, (2) divide the number of observations by the number of categories, (3) do a frequency count with the metric, and (4) choose break points so there are approximately the same number of observations in each category.
An alternate way of doing it is to divide into logical groups depending on the meaning of the categorical variable. This may result in vastly unequal sample sizes in each group which could impede the analyses.
The simple answer is don't do it. It results in loss of information and can produce spurious statistical significance.
The more subtle answer is that it may be OK in some contexts, but always use as many categories as you can. Two categories loses about a third of the info on a normally distributed variable. I would strongly advise against this unless you are reasonably expert or unless it is for presentation purposes such as plotting.
Just to follow up some of the existing answers as to why it is generally not a good idea to dichotomise (or split in other ways) continuous variables. I suggest you have a look at (there are many more):
On the practice of dichotomization of quantitative variables.
MacCallum, Robert C.; Zhang, Shaobo; Preacher, Kristopher J.; Rucker, Derek D.
Psychological Methods, Vol 7(1), Mar 2002, 19-40. http://dx.doi.org/10.1037/1082-989X.7.1.19
Bivariate median splits and spurious statistical significance.
Maxwell, Scott E.; Delaney, Harold D.
Psychological Bulletin, Vol 113(1), Jan 1993, 181-190. http://dx.doi.org/10.1037/0033-2909.113.1.181
Ideally for any regression analysis, it is important to convert the moderating variable into a categorical dummy coded variable as it helps to explain the phenomenon better. Categorical variables are considered to be a good moderator and the effect of categorical variables as a moderator can be measured reliably (Hair et al., 2010, pages 85 through 86). Categorization of continuous variables also makes the analysis and interpretation of results simple.
Interaction Between Two Quantitative Predictors:
When both predictors are quantitative, it is necessary to center the scores on each predictor before forming the product term that represents the interaction. The purpose of centering is to reduce the correlation between the product term and the X1 , X2 scores, so that the effects of the X1 and X2 predictors are distinguishable from the interaction. Scores are centered by subtracting the sample mean from the scores on each predictor.
page 632 in the chapter below
https://us.sagepub.com/sites/default/files/upm-binaries/47569_ch_15.pdf
You can find more answers here:
https://public.kenan-flagler.unc.edu/faculty/edwardsj/EdwardsLambert2007.pdf
or
here
http://public.kenan-flagler.unc.edu/faculty/edwardsj/interact.not.pdf
This article answers your question practically: Read page 5 under the heading Testing for two-way interactions
Dawson, J. F. (2014). Moderation in management research: What, why, when, and how. Journal of business and psychology, 29(1), 1-19.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS