Dear Florian

I didn't got your question clearly but are you asked about the method of predicting regression equation such as forward selection, backward elimination.....etc. if yes the backward elimination might be a good method here, it will produce several models which you could judge depends on R2 and the significant regression coefficients.

another approach might be good is the syructural equation models. hope that I got the point.

good luck 

Sounds like you want some of Jeff Edwards' ideas on polynomial regression:

http://public.kenan-flagler.unc.edu/faculty/edwardsj/dif.carma.pdf

http://public.kenan-flagler.unc.edu/faculty/edwardsj/dif.carma.pdf

Hey,

Thank you very much for your suggestions!

To become more precise, I want to model the bias of the Bravais-Pearson in the case of two categorical variables with 3 categories. I assume that the categorical variables are the product of a categorized bivariate standard normal  distribution. So my predictors are the true correlation and two threshold parameters for each categorical variable. Now I try to find a model which predicts the correlation bias.

I hope this helps to explain my issue.

Best regards,

Florian