I'm in the social sciences (psych) using cross sectional data. I'm completing a multiple hierarchical regression looking at interaction effects.

If one of my predictor (non-moderator) variables that I'm using in the interaction is highly non-normal (skewness = 2.1 and kurtosis 4.7), should I transform it to improve it? When is it ok to leave it alone? In terms of the other predictors I'm using, these pass the normality test and I'm centering them to minimize collinearity problems.


thanks ....
Linear regression has no distributional assumptions for the predictors.

The only reason to change the scale of a predictor is if you have influential outliers or to improve interpretibility of parameter estimates.

response to 'help!'

Thanks, this is what I thought, though read in a few spots to the contrary. Can I trouble you to elaborate a bit on your point about interpreting parameter estimates? I assume a 'parameter' is the model calculated from the regression equation.

Parameters are the regression coefficients for the population (the betas). The model contains parameters. Since you are running the regression on sample data, not the population, you have parameter estimates (the Bs).

It's a really common misconception about independent variables needing to be normal.