As Steven V. Rouse has said, there are often good conceptual and statistical reasons at times to keep the full scale of measurement rather than splitting a sample into groups. However, it would also surely be the case that at times you have a theoretical prediction about groups where it makes sense to group people based on a score of some kind. For example, the Beck Depression Inventory produces a score from 0 to 21, but there are various points that are considered clinical cut-offs, 0-8 is not depressed, 9-16 (I think) is mildly depressed. You may want to compare non-depressed with mildly depressed people, and you don't care whether your non-depressed people scored 0, 8, or somewhere in between those numbers.

If you have a god reason for using a continuous variable to seperate your sample into groups, the most common approach would most likely be to use a median split. From a quick search in Google Scholar with the search words: median split moderator, I got over 25,000 results. On a quick scan of these papers lots of them created groups by using a median split (or similar) of a continuous score in order to use the variable as a moderator: https://scholar.google.com.au/scholar?hl=en&as_sdt=0%2C5&q=median+split+moderator&btnG= This being the case, you're likely to find papers among these where this approach is convincingly justified.

Hi, Muhammad,

I would argue that this is not recommended. When you have a dimensional variable, splitting it into two groups causes you to lose a lot of statistical power. Imagine two people, both close to the cutoff level but one slightly above the cutoff and one slightly below the cutoff. If the construct truly is dimensional, these two people are very similar on the construct-- depending on the level of measurement error, perhaps even slightly reversed from each other. To arbitrarily place one in the lower group and one in the upper group will weaken your analyses because you are analyzing data as though they represent two different populations. If a variable is dimensional in nature, I would argue that it's generally best to choose a data analytic strategy that treats it as dimensional.

Now, with moderation analysis, sometimes it can be helpful to split the sample into two group just for the purpose of creating a figure. By splitting the sample into two groups you can create two different regression lines, which can make the nature of the moderation effect easier to visualize. But that's for the purpose of creating figures, not for the analyses themselves.

Take care,

Steve

As Steven V. Rouse has said, there are often good conceptual and statistical reasons at times to keep the full scale of measurement rather than splitting a sample into groups. However, it would also surely be the case that at times you have a theoretical prediction about groups where it makes sense to group people based on a score of some kind. For example, the Beck Depression Inventory produces a score from 0 to 21, but there are various points that are considered clinical cut-offs, 0-8 is not depressed, 9-16 (I think) is mildly depressed. You may want to compare non-depressed with mildly depressed people, and you don't care whether your non-depressed people scored 0, 8, or somewhere in between those numbers.

If you have a god reason for using a continuous variable to seperate your sample into groups, the most common approach would most likely be to use a median split. From a quick search in Google Scholar with the search words: median split moderator, I got over 25,000 results. On a quick scan of these papers lots of them created groups by using a median split (or similar) of a continuous score in order to use the variable as a moderator: https://scholar.google.com.au/scholar?hl=en&as_sdt=0%2C5&q=median+split+moderator&btnG= This being the case, you're likely to find papers among these where this approach is convincingly justified.

Respected Steven V. Rouse and Guy Curtis thank you very much for the guidance on this matter. Best regards

Most programs split moderator variables into groups to compare relationship b/w predictor and outcome. Usually this split is at interquartile ranges of moderator. Simple slopes are then computed for these groups to test if relationship b/w predictor and outcome differ at different levels of moderator variable. But interaction shall be generated using the continuous measure.