Bimodal distribution - moderator

[BeD]

In one of my constructs I find a kind of bimodal distribution. There are many people who answered with 1 (strongly disagree), only a few with 2 and three (so there is a gap), most of the people answered with 4 (neither agree, nor disagree) and every step from 5 to 7 a little less people have chosen these numbers to answer (like the half of a normal distribution). Therfore 1 looks like a mode and 4 like another mode, so there is no normal distribution.

I want to use the variable as a moderator. Do I have to transform the data or can I split the data in different models (1 and the rest) or what could be other opportunities what I am allowed statistically?
I just tried to split the data in two groups and got better results (higher significance and effect size).

[jmbecker]

Usually, the literature agrees that it is better to model a continuous variable as continuous moderator an not dichotomize. However, there are some papers discussing exceptions. I don't know if they cover your case, but it would sound plausible to me that your situation is better treated with a median split or similar putting the two modes in separate groups.
The following reference points to a special issue that covers some of the discussions.
http://www.sciencedirect.com/science/ar ... 0815000807

[BeD]

Many thanks for your quick answer!