Chapter 5: Quantitative analysis of MNZE rhoticity
5.2 Modelling the MNZE data
5.2.8 More specific models of rhoticity
Some of the results obtained from the models described above seemed worthy of more detailed scrutiny. In particular, there was a lack of clarity in relation to the relevance of gender for the articulation of both the pre-vocalic and the pre-consonantal tokens. It therefore seemed worthwhile to explore the effect of gender, as well as MCI scores, by fitting models to the data for each town separately. A more specific model would also verify whether there is indeed a clear age difference between the town N adults and teenagers when the town C speakers are not included in the model. In addition, I wanted to verify whether the regional differences identified in the models described thus far, would hold for the teenage data once the adults were removed.
The aim of the first more specific model12 (Model Teenage PreV), was to probe the robustness of the main social predictors identified in Model PreV (i.e. linking /r/) when the town N adult data was removed. I therefore fitted a model to the teenage data only and entered the variables: MCI, region and gender. The estimated coefficients for Model
TeenagePreV are presented in table 5.15 with none of the variables dropped from the model.
MCI continues to be a significant predictor of linking /r/ articulation in this model. Region is also identified as a predictive factor. Town N teenagers are estimated to produce less linking /r/ than town C teenagers (p=<0.001). This regional difference is shown in figure 5.15. The model does not identify gender as influential on linking /r/.
Table 5.15: Model estimates for Model TeenagePreV
Estimate Std. Error z-score Pr(>|z|) Intercept / baseline 1.583580 0.231460 6.842 < 0.001 MCI -0.190877 0.038590 -4.946 < 0.001 Region N -0.559603 0.208204 -2.688 0.00719
Gender M 0.008239 0.209181 0.039 0.96858
12 The random effects for word and for speaker continued to be included in all of the specific models described in this section.
Figure 5.15: Log odds of pre-vocalic /r/ for teenagers by region
In the next more specific model I tested the predictive value of MCI, age and gender for linking /r/ within region N only, in order to see if age and other social fixed effects would hold constant when only the town N speakers were considered. I entered only age, MCI and gender into this model (Model RegionNPreV) and display the results in table 5.16 with none of the variables removed.
Table 5.16: Model estimates for Model RegionNPreV
Estimate Std. Error z-score Pr(>|z|) Intercept / baseline 2.21073 0.29633 7.460 < 0.001 Age Y -1.48556 0.30832 -4.946 < 0.001 MCI -0.16565 0.03889 -4.818 < 0.001 Gender M 0.48102 0.27958 1.721 < 0.1
As with previous models, MCI remains a factor which predicts linking /r/ use. Speakers with higher MCI scores are estimated to produce less linking /r/. Age is also identified as
significant in Model RegionNPreV with the teenagers predicted to produce less linking /r/
than adults. Once again, gender is not identified as a significant factor (p=0.08). In order to be
sure of the absence of a gender effect, I repeated the model with the adults removed. Model RegionNTeensPreV includes only MCI and gender as explanatory factors. The outcome is shown in table 5.17. Only MCI is identified as significant. Gender is apparently not a significant predictor of /r/ for the teenagers‟ use of linking /r/ in town N, though the male teenagers are predicted to articulate slightly more /r/ than the females.
Table 5.17: Model estimates for Model RegionNTeensPreV
Estimate Std. Error z-score Pr(>|z|) Intercept / baseline 0.6862 0.2799 2.452 < 0.1 MCI -0.16565 0.0450 -3.626 < 0.001 Gender M 0.5101 0.3162 1.613 0.106645
A similar model, with only MCI and gender included, was fitted to the data for the town C speakers only. The model estimates for this model (Model RegionCTeensPreV) are shown in table 5.18.
Table 5.18: Model estimates for Model RegionCTeensPreV
Estimate Std. Error z-score Pr(>|z|) Intercept / baseline 2.02491 0.29769 6.802 < 0.001 MCI -0.27279 0.06413 -4.254 < 0.001
Gender M -0.38669 0.24850 -1.556 0.12
The model for the town C teenagers also predicts that MCI, but not gender, is a significant predictor of linking /r/ use. What is interesting with respect to gender is that the direction predicted for males in town C is opposite to that for males in town N. In town N male speakers are predicted to use slightly more linking /r/ and in town C males are predicted to use slightly less. It would seem then, that because the males and females show trends in a different direction in each town with respect to linking /r/, the model is unable to identify a clear trend for gender in the data overall. Figure 5.16 shows the estimated log odds of
articulation for males and females in each region, though I emphasise that the model does not identify these gender differences as having statistical significance.
Figure 5.16: Log odds of pre-vocalic /r/ by for males versus females in each region
In light of this complicated patterns of distribution with respect to gender, the individual speaker data may be particularly informative with respect to linking /r/.
I also fitted more specific models to the non-pre-vocalic /r/ data. In all previous models, only pre-consonantal tokens were considered because the inclusion of phrase final and absolute final tokens caused complications for the models. However, if the models are fitted with a much smaller set of explanatory variables, it is possible to include all of the non-pre-vocalic phonological contexts of /r/ use (i.e. pre-consonantal non-phrase final tokens of /r/
plus all pre-vocalic and pre-consonantal phrase final tokens of /r/ plus all absolute final tokens of /r/).
I decided to test whether the influence of the social factors identified as significant in the models of pre-consonantal /r/ would also hold when all non-pre-vocalic tokens were
included. I fitted the first such model (Model TeenageNONPreV) to the teenage data only and included only the variables region, MCI and gender. The results are shown in table 5.19 with none of the variables removed. The model confirms the regional difference that was
identified in Model PreC. Region N teenagers are predicted to produce more non-pre-vocalic /r/ than region C teenagers.
Table 5.19: Model estimates for Model TeenageNONPreV
Estimate Std. Error z-score Pr(>|z|) Intercept / baseline -10.24241 0.82082 -12.478 < 0.001 MCI 0.05908 0.06911 0.855 0.393 Region NR 1.67078 0.41642 4.012 < 0.001 Gender M -0.38326 0.42477 -0.902 0.367
This model also confirms the Model PreC finding that neither MCI nor gender is influential on /r/ use. This finding holds with all non-pre-vocalic /r/ tokens included. Figure 5.17 shows the model‟s estimated regional difference in the log odds of non-pre-vocalic /r/.
Figure 5.17: Log odds of non-pre-vocalic /r/ for teenagers by region
In order to test whether MCI or gender would become relevant for all non-pre-vocalic /r/
tokens when only the data for town N was considered, I fitted a model to the town N data, including the variables age, MCI and gender only. According to this model (Model
RegionNNONPreV), teenagers are predicted to have a higher log odds of articulation than adults. Again, MCI and gender are not identified as influential. The model estimates are provided in table 5.20 with none of the variables removed. Figure 5.18 shows the age difference in the log odds of non-pre-vocalic /r/ articulation within town N.
Table 5.20: Model estimates for Model RegionNNONPreV
Estimate Std. Error z-score Pr(>|z|) Intercept / baseline -12.19577 1.36903 -8.908 < 0.001 Age Y 2.20504 0.79409 2.777 < 0.01
MCI 0.11777 0.07716 1.526 0.12693
Gender M -0.69893 0.59230 -1.180 0.23799
Figure 5.18: Log odds of non-pre-vocalic /r/ by age for region N speakers only
A corresponding model was fitted to the town C data. Again, this was to check the
relevance of gender and MCI when only town C speakers are considered. Only age, MCI and gender were included in the model. As with Model RegionNNONPreV, neither gender nor MCI appear to have a significant effect on the articulation of /r/ when all non-pre-vocalic tokens are included. The model estimates are shown in table 5.21 with all of the entered variables retained in the model.
Table 5.21: Model estimates for Model RegionCNONPreV
Estimate Std. Error z-score Pr(>|z|)
Intercept / baseline -8.39116 0.88633 -9.467 < 0.001
MCI -0.08548 0.15306 -0.558 0.577
Gender M -0.02183 0.57871 -0.038 0.970