The question of dividing language proficiency into components was explored by determining which of three models best fit the experimental data: a model postulating numerous specific sources of variance (the extreme divisible model), one postulating a single, large source of variance (the unitary model), or one postulating a large general factor and several smaller specific factors. Four measures with clear contrasts in content and mode were constructed and used to identify linguistic and method variance in a correlation matrix of language proficiency variables. The four measures were administered, along with eight other language measures, to a group of beginning-level learners of English as a second language. Factor analysis and comparison with demographic data were performed on the data. The findings suggest that the language proficiency data were best modeled by a large general factor and two specific, content-related factors--grammar and vocabulary. The possibility of specific factors related to mode was not ruled out. (Author/MSE)