A user-knowledge-based variable selection method for limited information maximum likelihood using principal components