Seeking out inter-language differences in the execution of a contrastive analysis is given priority over looking for manifestations of language universals. But unless a contrastive study is based upon an understanding of language universals and contains a set of instructions for how each language realizes them, the common ground for contrast can only be arbitrarily decided for each set of languages, and the contrasts themselves remain largely on the surface. In evaluating observed transfer and interference in language learning, the final step of error prediction is possible only with a general linguistic model that operates within the framework of deep and surface grammar. Taxonomic studies, ignoring the theoretical problem of determining interlinguistic correspondences, established phoneme classes, morphemes and grammatical categories separately for each language; no justification other than the anticipation of probable learning errors was required for contrasting the specific units and arrangements of one language with those of another. The present paper discusses the theoretical orientation of current contrastive studies and proposes to incorporate the notions of deep and surface grammar in contrastive analysis. Illustrations of the operations of deep categories, surface realization rules and contrastive statements are given. (Author/FWB)