I have data that are responses to several multi-item scales. My plan is to use Multiple Imputation (via Proc MI in SAS) to deal with missing values, and then examine the relationships among the scale scores.

The problem is that Multiple Imputation assumes multivariate normality, and almost all the items are skewed. So, I want to do a Box-Cox transformation on the data (all values are positive) to have the distributions more closely resemble normality. I can do this with a TRANSFORM statement in PROC MI, but I need to specify the lambda in the TRANSFORM statement. It is unclear to me how to figure out the optimal lambda.

It seems like it might be possible to do it with PROC TRANSREG but as best I can tell from the examples in PROC TRANSREG you have to specify a regression equation that you are going to test. That doesn't really apply in my case - none of the individual items I am interested in transforming can be thought of as outcome or predictor variables.

I hope someoone can point me in the right direction.

(Another question which may be more appropriate for Cross-Validated is to see if it makes sense to use different values of lambda for different items in the study.)