Abstract

This paper describes the construction and analysis of a corpus of harmonic progressions from 12-bar blues forms included in the jazz repertoire collection The Real Book. A novel method of coding and analyzing such corpus data is developed, with a notion of “possible harmonic change” derived from the corpus and logit mixed-effects regression models that describe the difference between actually occurring harmonic events and possible but non-occurring ones in terms of various sets of theoretical constructs. Models using different sets of constructs are compared using the Bayesian Information Criterion, which assesses the accuracy and efficiency of each model. The principal results are that: (1) transitional probabilities are better modeled using root-motion and chord-frequency information than they are using pairs of individual chords; (2) transitional probabilities are better described using a mixture model intermediate in complexity between a bigram and full trigram model; and (3) the difference between occurring and non-occurring chords is more efficiently modeled with a hierarchical, recursive context-free grammar than it is as a Markov chain. The results have implications for theories of harmony, composition, and cognition more generally.