3dLME can handle missing data at some factor levels, but it can't deal with missing covariate values. You have two options: either don't use those subjects with missing covariate values, or don't include the covariate. Regardless, it seems that you should use 3dMVM, not 3dLME.
Gang