Chapter 5: Quantitative analysis of MNZE rhoticity
5.2 Modelling the MNZE data
5.2.7 Model 1
5.2.7.5 Model PreC
Model PreC was fitted only to the non-phrase final pre-consonantal /r/s. The model initially included preceding vowel, word frequency, age, region, MCI, gender, interactions between region and MCI, region and gender, MCI and gender and between region, MCI and gender.
The random effects for word form and speaker were also included.
In this model collinearities were identified between MCI and region and between MCI and gender. This is because there are more region N speakers with higher MCI scores than region C speakers and because there are more female speakers with higher MCI scores than male speakers. This issue was solved by re-running the model using centred variables. The best fitting model for the pre-consonantal tokens excluded the interactions, as well as gender and word frequency, but it included MCI even though MCI was not identified as a significant predictor.
The best fitting model retains vowel, age, region and MCI plus the random effects for speaker and for word form. Although the pre-consonantal tokens have a very low likelihood of articulation, the model identifies several effects as significant on articulation of /r/. The coefficients for Model PreC are provided in table 5.12.
Table 5.12: Estimates for the best fitting Model PreC
Estimate Std. Error z-score Pr(>|z|) Intercept / baseline -9.57767 0.95387 -10.041 < 0.0001 Vowel FIRE -12.17702 2709.91489 -0.004 0.99641 Vowel lettER 0.32038 0.77820 0.412 0.68056 Vowel NEAR -0.37012 1.35651 -0.273 0.78497 Vowel NORTH -0.17507 0.82874 -0.211 0.83270 Vowel NURSE 3.43358 0.65838 5.215 <0.0001 Vowel OUR -12.28837 807.94189 -0.015 0.98787 Vowel START -0.33024 1.05708 -0.312 0.75473
Age Y 1.86176 0.64577 2.883 0.00394
Region N 1.66048 0.37695 4.405 <0.0001
MCI 0.09239 0.05971 1.547 0.12178
In contrast to Model PreV, the model for pre-consonantal /r/ identifies an effect for the preceding vowel. A preceding NURSE vowel is a significant predictor for pre-consonantal /r/.
The model identifies a hierarchy of preceding vowel contexts in which pre-consonantal /r/
tokens are most likely to be articulated. The difference in likelihood of articulation for NURSE
compared with other preceding vowel contexts is apparent in figure 5.12.
Figure 5.12: The effect of a preceding NURSE vowel on pre-consonantal /r/ articulation
As shown in table 5.12 and figure 5.12, the lettER tokens have the second highest coefficient value. /r/s with a preceding FIRE and OUR vowel have the lowest.
As with the pre-vocalic /r/ tokens age is identified as a significant factor in the model of pre-consonantal /r/. Although the likelihood of articulation is very low for pre-consonantal tokens, it increases for teenagers in contrast to adults (see figure 5.13). This lends support to a hypothesis of change. Perhaps younger speakers are beginning to exhibit a slight tendency to articulate some pre-consonantal /r/s.
Figure 5.13: Log odds of pre-consonantal /r/ articulation by age
As noted previously, it is important to be cautious about the age results for the MNZE data, given that there are only a few adults available for analysis and especially since in this model, all teenagers, regardless of their town, are being compared with adults from only 1 town.
Model PreC predicts an effect for region such that pre-consonantal /r/ has a higher likelihood of articulation for the speakers in region N when compared to region C (shown in figure 5.14).
Figure 5.14: Log odds of pre-consonantal /r/ articulation by region
The difference between the region N and region C log odds of pre-consonantal /r/ is not huge, but it is important to take into account that pre-consonantal /r/ is seldom articulated across the data set. A higher likelihood of pronunciation in this context for a particular region may be indicative of an innovation in its early stages of adoption, especially in light of the
corresponding age difference.
In contrast to the model of pre-vocalic tokens, MCI does not appear to have a significant effect in Model PreC although the model is a better fit when it is included than when it is not.
It is possible that MCI has some relevance but that it is not identified as statistically
significant due to the low number of articulated tokens. It seems surprising that gender does not appear to be a significant predictor of pre-consonantal /r/ articulation in this model and I explore the relevance of gender further below.
The influence of individual word forms on the articulation of pre-consonantal /r/s can be probed by considering the individual word item intercepts provided by Model PreC. The intercepts represent the estimated adjustment to the baseline intercept value for each word
form when the general trends / fixed effects identified in Model PreC are taken into account (i.e. word frequency is not influential but a preceding NURSE vowel is).
There are 936 different word items for the pre-consonantal tokens in Model PreC. The majority of these word items have negative intercept values. This is to be expected given that pre-consonantal /r/ is seldom articulated. However, 41 of the word forms have positive intercepts. These 41 word forms are shown in table 5.13a, in order of highest to lowest intercept value (i.e. most likely to least likely to be articulated). Table 5.13b shows the 40 word forms with the lowest intercept values.
Tables 5.13a also includes the numerical position of each word form on an ordered list of most to least frequent across the 936 word items. Both tables also show, where appropriate (i.e. if the following consonant occurred within the same word), the phonological target of the consonant which immediately follows the /r/ (though this is not necessarily how the
following consonant was articulated).
There may be individual word effects beyond preceding vowel context on the words with the highest intercepts. Although Model PreC identified a preceding NURSE vowel as a
significant predictor, tables 5.13a and 5.13b indicate that there may be additional factors affecting the likelihood of articulation for different lexical items. Different word items with a
NURSE vowel adhere to the model‟s trend to different degrees.
Model PreC did not identify word frequency, when calculated in relation to the whole data set, as having a significant predictive effect. Looking at the frequency of items in relation to other pre-consonantal /r/ word forms specifically there is again no identifiable effect on the likelihood of articulation. Words items with positive intercepts are some of the most frequently occurring as well as some of the least frequently occurring pre-consonantal /r/ words, although most of the words are positioned within the top (i.e. most frequent) half of the word items.
It is also apparent from a comparison of tables 5.13a and 5.13b that different forms of the same lemma can have quite different intercept values. For example, work is amongst the 40 highest intercept values, while works is amongst the 40 lowest.
Table 5.13a: 41 word forms with positive intercept values in Model PreC
Word forms Intercepts Vowel Position Following phonological target
for 1.20 NORTH 2
burning 1.07 NURSE 397 alveolar nasal
are 0.911 START 8
burnt 0.780 NURSE 153 alveolar nasal
weren‟t 0.670 NURSE 109 alveolar nasal
work 0.633 NURSE 14 voiceless velar plosive
higher 0.615 lettER 287
permission 0.614 lettER 486 bilabial nasal
performing 0.611 lettER 199 voiceless labiodental fricative perform 0.606 lettER 197 voiceless labiodental fricative
air 0.602 SQUARE 152
turn 0.593 NURSE 103 alveolar nasal
worst 0.584 NURSE 232 voiceless alveolar fricative
other 0.576 lettER 12
nerdy 0.568 NURSE 765 voiced alveolar plosive thirty 0.546 NURSE 58 voiceless alveolar plosive workplace 0.538 NURSE 934 voiceless velar plosive
before 0.536 NORTH 29
birthday 0.525 NURSE 95 voiceless dental fricative person 0.507 NURSE 43 voiceless alveolar fricative
better 0.503 lettER 30
births 0.500 NURSE 387 voiceless dental fricative purple 0.495 NURSE 814 voiceless bilabial plosive shirt 0.495 NURSE 855 voiceless alveolar plosive murders 0.489 NURSE 470 voiced alveolar plosive years 0.459 NEAR 20 voiced alveolar fricative
heard 0.448 NEAR 81 voiced alveolar plosive
prefer 0.429 NURSE 124
persons 0.428 NURSE 356 voiceless alveolar fricative
we‟re 0.419 SQUARE 19
third 0.410 NURSE 87 voiced alveolar plosive
where 0.395 SQUARE 17
workers 0.389 NURSE 181 voiceless velar plosive
her 0.343 NURSE 13
first 0.330 NURSE 25 voiceless alveolar fricative
terms 0.279 NURSE 225 bilabial nasal
christchurch 0.256 NURSE 167 voiceless (post)alveolar affricate
worked 0.202 NURSE 87 voiceless velar plosive
they‟re 0.0178 SQUARE 6
learning 0.117 NURSE 82 alveolar nasal
or 0.0850 NORTH 4
Table 5.13b: 40 word forms with the lowest intercept values in Model PreC Word forms Intercepts Vowel Following phonetic context
were -0.898 NURSE
university -0.659 NURSE voiceless alveolar fricative learnt -0.628 NURSE alveolar nasal
theres -0.532 SQUARE voiced alveolar fricative alternative -0.384 NURSE alveolar nasal
there -0.355 SQUARE
works -0.355 NURSE voiceless velar plosive thirteen -0.317 NURSE voiceless alveolar plosive
sir -0.303 NURSE
birth -0.272 NURSE voiceless dental fricative
term -0.262 NURSE bilabial nasal
worker -0.238 NURSE voiceless velar plosive
over -0.233 letter
sort -0.215 NORTH voiceless alveolar plosive certain -0.204 NURSE voiceless alveolar plosive nurse -0.204 NURSE voiceless alveolar fricative circle -0.197 NURSE voiceless velar plosive learners -0.194 NURSE alveolar nasal
hurt -0.193 NURSE voiceless alveolar plosive word -0.191 NURSE voiced alveolar plosive
here -0.190 NEAR
turned -0.168 NURSE alveolar nasal
prefers -0.168 NURSE voiced alveolar fricative services -0.163 NURSE voiced labiodental fricative words -0.163 NURSE voiced alveolar plosive perfect -0.157 NURSE voiceless labiodental fricative personal -0.156 NURSE voiceless alveolar fricative Thursday -0.155 NURSE voiced alveolar fricative furtherest -0.147 NURSE voiced dental fricative dirty -0.143 NURSE voiceless alveolar plosive earth -0.127 NURSE voiceless dental fricative percent -0.123 letter voiceless alveolar fricative brothers -0.122 letter voiced dental fricative
their -0.113 SQUARE
learn -0.112 NURSE alveolar nasal
sisters -0.111 letter voiced alveolar fricative nursing -0.111 NURSE voiceless alveolar fricative
you‟re -0.109 NORTH
murdered -0.107 NURSE voiced alveolar plosive worth -0.104 NURSE voiceless dental fricative
It is also worth considering the relevance of the following phoneme. For both the positive and the negative intercept values, a range of different consonants follow the /r/ across the word forms (note that the contexts are phonemes and the precise articulations are likely to differ).
In table 5.13a there are 28 word forms for which information about the following phoneme is available. For table 5.13b there are 33. In all other word forms the /r/ was followed by another word beginning with a consonant and information about the following word was not entered into the model. In table 5.13a, 9 of the 28 (32%) word items have a following fricative consonant. In table 5.13b, there are 16 out of 33 (48%). It is possible then that there is a slight tendency for a following fricative consonant to have a disfavouring effect on the articulation of pre-consonantal /r/11. In chapter 4 I noted that there are comments in the historical literature suggesting that pre-consonantal /r/ declined at a very early date in some words where it appeared before /s/ and /ʃ/. This observation therefore warrants further scrutiny in any future analyses.
It is important to also consider whether there are simply more /r/ tokens with a preceding
NURSE vowel relative to other preceding vowel contexts across the data set. Table 5.14 shows that this is not the case.
Table 5.14: Number of word forms for each preceding vowel context
Vowel context FIRE OUR NEAR START NURSE SQUARE lettER NORTH
Number of items 23 191 682 1033 1483 2103 2299 2503 Pre-consonantal
and pre-vocalic
25 243 1132 1195 1579 2815 2927 3155
It is clear that in addition to a statistically higher likelihood for pre-consonantal /r/s to be articulated in the context of a preceding NURSE vowel, there are word form specific effects (perhaps especially the following context) on /r/ articulation. It is not unusual for changes to commence in a lexically specific way, progressing from 1 lexical item to another.
11 I would like to thank Jen Hay for bringing this to my attention.