In a large Ngbaka dictionary (Henrix et al. 2015), segmental identity between vowels is systematically overrepresented: cases of identity like [kɔkɔ] 'to cover oneself' and [ɓɛnɛ] 'to consult an oracle' occur significantly more often than predicted by the frequencies of the individual vowels alone. I show that the best-fit statistical model of the lexical data includes a predictor for identity: non-identical vowels are dispreferred. The implication of this finding is that the phonological grammar must be able to reference identity directly (following MacEachern 1999, Coetzee & Pater 2008, Gallagher & Coon 2009, Gallagher 2013, 2014, a.o.).