306 Thomas C. Purnell
the Wisconsin English speaker groups where percent glottal pulsing is greater
than vowel duration. For the controls, there is a fairly equal relation between
the two measures. Second, note that the differential relation favoring percent
glottal pulsing is most extreme in older groups. Third, that all other things
being equal, the youngest speakers mirror the group immediately preceding
it, i.e., the 1920–1939 group, and not the controls.
To this point, the analysis has focused on the voiced-voiceless compari-
son. Tables 13.6 and 13.7 compare within-group variation. That is, the principal
component analysis explains how latent factors may shape the within-category
data, and provides insight into the variability of the weighting of variables such
as percent glottal pulsing and vowel duration, which dominate all ¿ ve groups’
data. The cumulative proportion for the second principal component by voic-
ing category on Table 13.7 reveals how much within-category variation is
accounted for by the measures used by the model. The voiceless tokens in each
group ranged from 63% (group 4) to 85% (group 2). The voiced tokens in each
group ranged from 65% (group 3) to 87% (group 2). Group 3 perception data
had similar results with the voiced and voiceless tokens (63% and 89%, respec-
tively). All of the eigenvalues on Table 13.7 exceed a value of 1, although the
eigenvalues for group 4, group 5, and the perceptual group 3 data in the multi-
variate canonical discriminant analysis (Table 13.5) do not reach a value of 1. In
order to visualize the separation of tokens by voicing, the strongest eigenvectors
for the relevant principal component for voiced and voiceless tokens were used
as weights of each measure. The addition of the two strongest measures per
principal component were plotted against each other and shown in Figure 13.3
through Figure 13.8. Selection of the component eigenvector was established by
¿ nding the largest absolute value and then using the same component assign-
ment for voiced and voiceless tokens. Coherence should prevail over these
resultant groupings. Consider, for example, the ¿ rst group. The voiceless ¿ rst
component eigenvector for percent glottal pulsing (0.541) and the voiced ¿ rst
component eigenvector for vowel duration (0.610) suggest that these two mea-
sures are grouped together. Such a grouping suggests a latent factor of TEMPO-
RAL CHANGE, while the other pair of measures, change in F0 and change in
F1, suggests a competing latent factor of SPECTRAL CHANGE.^13 The group-
ing of percent glottal pulsing and pulsing duration in group 2 suggests a latent
factor of PULSING in contrast to CONSONANT DURATION (consonant gap
duration and V:C ratio). The latent factors in group 3 mirror group 1. The ¿ rst
latent factor for group 4 is the TEMPORAL CHANGE factor (percent glottal
pulsing, vowel duration). However, the coherence of an interpretation of the
grouping of consonant gap duration and change in F0 is less clear (independent
from the other two measures). The latent factors in group 5 represent RELA-
TIVE DURATION (percent glottal pulsing, V:C ratio) as the ¿ rst component