I see no harm whatsoever in taking a transform of the response in a mixed model to get closer to the model assumptions. However, I would go one step beyond what you propose and calculate what is called a rankit.

rankit = INVNORMAL ((Rank of X - 0.5)/ n)

This will then have a mean of zero and a sd of 1 and will be normally distributed. The q -q or normal probability plots you mention graphs the rankits on the horizontal axis and the data points on the vertical axis.

One word of warning however, the normality assumption concerns the response conditional on the predictors, that is it is about the residuals. It may be that including your predictors may induce non-normality so you should use the Q-Q of the residuals to check.

I echo what Kelvyn Jones said about the normality assumption applying to residuals. I also remind you of what George Box said about normal distributions (and straight lines, for that matter): They don't exist in nature, but may serve as useful models.

Finally, you may find this article useful:

http://www.joophox.net/publist/csda04.pdf

Joop Hox, as you may know, is the author of a well-known book on multilevel modeling (http://www.joophox.net/mlbook2/MLbook.htm).

HTH.

Thanks Kelvyn for "rankit" and Bruce for the article. Much appreciated!

I indeed use residuals for q-q plots.