54 Jen Hay and Margaret Maclagan
We wondered whether this effect resulted from speakers of different classes
producing slightly different sets of preceding vowels. To double-check this,
we tried ¿ tting the model with just a single base word, sofa (the base which
has a reasonable frequency of intrusive /r/ production regardless of social
class). In a model ¿ t on just the sofa data, both the F3 of /r/ during Sarah
(p<.001), and social class (p<.005) are signi¿ cant predictors of the F3 during
intrusive /r/. This con¿ rms that social class still remains signi¿ cant when all
variation attributable to the different base words is removed.
In order to double-check that our strategy for pseudo-normalizing the
values was having the desired effect, we also attempted to ¿ t a model which
excluded the normalizing value. In the resulting model the most signi¿ cant
predictor is gender (p<.0001). The base type is still signi¿ cant (p<.0001).
Social class does not quite reach signi¿ cance (p<.07). Gender is not signi¿ -
cant in the model which includes the speaker F3 for a “real” /r/, but it is highly
signi¿ cant in a model which excludes it. We interpret this difference as indi-
cating that the “real” /r/ F3 measurement is doing a decent job of removing
vocal tract length effects from the model.
We were interested in the degree to which there may be a direct link
between the rate of /r/ insertion and the degree to which the /r/ is constricted.
Figure 2.8 Effect of social class in predicting the F3 of intrusive /r/. Lower index
scores are associated with higher social classes.