I also think that a GLM (what you call ANCOVA) is the best solution and that a dichotomization is the worst one.

You should check if the chosen functional relationship between the trait variable and the response is reasonable (the relationship between possibley transformed trair var. and possibly transformed response should be linear; the transformations may be chosen to just get such a linear relationship).

Do you have repeated measures? For instance: are different stimulus types applied to the same individual? In this case a mixed model with "subject" as random factor might be better.

What is the response? Did you chose a reasonable error model? For instance, if the response are counts and the variance depends on the mean, a Poisson-, Gamma-, or NegBinom-model would be indicated.

Regarding the interpretability:

You could center the metric covariable. The coefficients of the model will then refer to the mean value of this covariable.

If there is an interaction of this covariable with the other predictors then the story is complicated. You may plot the results/predictions for a low, a medium, and a high value of this variable to get an impression and a guide to interpretation.

I also think that a GLM (what you call ANCOVA) is the best solution and that a dichotomization is the worst one.

You should check if the chosen functional relationship between the trait variable and the response is reasonable (the relationship between possibley transformed trair var. and possibly transformed response should be linear; the transformations may be chosen to just get such a linear relationship).

Do you have repeated measures? For instance: are different stimulus types applied to the same individual? In this case a mixed model with "subject" as random factor might be better.

What is the response? Did you chose a reasonable error model? For instance, if the response are counts and the variance depends on the mean, a Poisson-, Gamma-, or NegBinom-model would be indicated.

Regarding the interpretability:

You could center the metric covariable. The coefficients of the model will then refer to the mean value of this covariable.

If there is an interaction of this covariable with the other predictors then the story is complicated. You may plot the results/predictions for a low, a medium, and a high value of this variable to get an impression and a guide to interpretation.

I am not sure what you mean with "within"-design. Normally, this means that the information is derived from the same person (e.g. by repeated measurement). Another thing I don't understand is what you want to do with the trait variable. Is is more of a bias variable you want to control? Or is it a third independent variable? In any case, you could try multiple regression. You would not need to dichotomize any variables. In case of dependent data like repeated measures, you could use a multilevel regression model to cope with that.

Many thanks for the quick replies and the help. I will try an mixed model design in the next step.

@Jochen: yes there is an interaction with the covariable and this is also of great interest, which, as you said, does not make it easier...

@Marc: long time, no see, hope you are doing fine ;-)

@Roh: the trait variable is a third independent variable. Did you use a mixed model/multilevel model and if yes, which program did you use?

I attended an MPlus-Workshop in 2009 on multi-level regression models. But I don't work with MPlus, so this will not be helpful except for the knowledge how to do it in general. If you want to compute using R, you could ask Timo Gnambs for help.

If you never did something like this before, I also would suggest to read Tabachnick and Fidells Book on Multivariate Methods. It's a practical handguide including examples for SPSS. They discuss and compare the different methods and make suggestings on when to choose which, and show the reader how to perform the tests, plot the slopes, etc.

Yes, I have the Tabachnick & Fidell book (great book) at hand but ignored this chapter too long I suspect. I also have West, Welch, and Galecki - Linear mixed models. Roh, I thought if you already used R, you could give a quick introduction ;-)

I think we both attended to the same MPlus workshop. Maybe I'll ask Timo