108 Zsuzsanna Fagyal
having a substantial effect on its ¿ t of the observed data. The predictor vari-
able removed ¿ rst was the one that had the least impact on the model, while
the predictor removed next was, each time, the one without substantial contri-
bution. The elimination criterion at each step was statistical signi¿ cance, with
a cut-off point of p<0.05. The analyses ran until the best possible ¿ t (no more
improvement) was obtained.
Among all predictors, ¿ ve were signi¿ cant at the initial step. These were:
number of segments in the phrase, articulatory rate, grade-level in school,
moyenne générale, and speaker. The initial ¿ t of the observed data, however,
was barely above chance: 54.3%. After three consecutive runs with several
iterations each, this picture improved. Analyses performed with ¿ ve rhythm
type predictor variables resulted in an overall ¿ t of 68.6% of the model, with
only articulatory rate contributing signi¿ cantly (cut-off p<0.05) to the overall
improvement (r^2 = 0.521). The addition of grade in school as the ¿ rst and only
external variable produced a 67.1 % overall prediction accuracy, while the
addition of moyenne générale alone produced a ratio of 73.1 %. The combi-
nation of the last two variables resulted in 75.9 % accuracy. The addition of
speaker as an independent variable alone led to 80.1 % accuracy in predicting
what speaker belonged to the EF or the AF group.
Table 4.3 shows correlation coef¿ cients for all predictor variables.^34 Out
of ¿ ve rhythmic predictors, only ar ticulator y rate and the number of segments
in the phrase correlated signi¿ cantly and positively with the outcome variable
ethnicity.^35 These two variables also showed co-linearity: their positive rela-
tionship means that the higher was the articulatory rate, the more intervals
(vocalic and consonantal) speakers tended to include in a phrase.^36 Thus, AF
speakers tended to speak faster and inserted overall more segments in their
phrases than EF speakers.
Although none of the rhythm type indices (%V, ǻC or ǻV) predicted sig-
ni¿ cantly which speakers can be grouped in which of the two categories of
ethnic origin, the inverse correlations of %V and ǻC values with ethnicity
indicate that as the outcome variable increases speakers’ %V and ǻC values
would decrease. AF speakers, although slightly, tended to have less vocalic
intervals (low %V) and less variation in the duration of consonantal intervals
(low ǻC values) in their readings. They also showed slightly greater ǻV val-
ues, which is indicative of greater vocalic interval duration variations.
Speakers’ grades in school (approximate ages) and moyennes générales
were statistically signi¿ cant and showed inverse correlation with ethnicity.
Thus EF speakers who were slightly younger, that is enrolled in at a younger
grade level, also tended to be better students than AF speakers. The indi-
vidual speaker as a predictor variable brought the single most important