Inadequacy of the chi-squared test to examine vocabulary differences between corpora

Pearson's chi-squared test is probably the most popular statistical test used in corpus linguistics, particularly for studying linguistic variations between corpora. Oakes and Farrow (Literary and Linguistic Computing, 2007, 22, 85-99) proposed various adaptations of this test in order to allow for the simultaneous comparison of more than two corpora, while also yielding an almost correct Type I error rate (i.e. claiming that a word is most frequently found in a variety of English, when in actuality this is not the case). By means of resampling procedures, the present study shows that when used in this context, the chi-squared test produces far too many significant results, even in its modified version. Several potential approaches to circumventing this problem are discussed in the conclusion.