Do all roads lead to Rome? Modeling register variation with factor analysis and discriminant analysis

Previous theoretical and empirical research on register variation has argued that linguistic co-occurrence patterns have a highly systematic relationship to register differences, because they both share the same functional underpinnings. The goal of this study is to test this claim through a comparison of two statistical techniques that have been used to describe register variation: factor analysis (as used in Multi-Dimensional analysis, MDA) and canonical discriminant analysis (CDA). MDA and CDA have different statistical bases and thus give priority to different analytical considerations: linguistic co-occurrence in the case of MDA and the prediction of register differences in the case of CDA. Thus, there is no statistical reason to expect that the two techniques, if applied to the same corpus, will produce similar results. We hypothesize that although MDA and CDA approach register variation from opposite sides, they will produce similar results because both types of statistical patterns are motivated by underlying discourse functions. The present paper tests this claim through a case-study analysis of variation among web registers, applying MDA and CDA to analyze register variation in the same corpus of texts.