Comparing weights of cues with different numbers of levels.
Vsevolod M. Kapatsinski, Irina A. Shport, Susan Guion‐Anderson · The Journal of the Acoustical Society of America · 2011
Cue weighting is a useful methodological tool in speech perception research: it allows to access within-group and between-group biases in sound categorization. Examination of cue weighting may also have major implications for phonological theory. For instance, significant individual variation in reliance on cues to a phonemic contrast within a speech community challenges the traditional assumption of the language-specific feature system. Morrison argued that logistic regression coefficients in identification tasks provide good estimates of cue weights [Studies in Second Lang. Acquisition 597–606 (2005)]. Unfortunately, coefficient estimates vary with the number of levels a cue has, which makes it impossible to directly compare weights of cues with different numbers of levels, a serious limitation. The present paper shows that, using a null hypothesis of zero cue weight and a fixed sample size, one can employ Monte Carlo techniques to estimate the effect of the number of levels on observed cue weights (regression coefficients). One can then test whether the difference in the number of levels between cues can account for the observed difference in weights. If not, then the cue weights are reliably different. The method is demonstrated on data investigating individual variation in attention to cues to pitch accent.