We first ran separate logistic mixed effects models for each age-group, predicting whether a participant gave a correct response (i.e., picked the correct word as different out of a choice of three, coded as 1/0) on each trial. The fixed factors were test-session (pre-test, post-test), voice-novelty (trained voice, untrained voice), word-novelty (trained word, untrained non-word), condition (high-variability, low-variability) and all the interactions between these experimental factors. The factor talker and all of the interactions between
talker and the experimental factors were included as control factors. We also included the continuous variable mean-replay-usage as an additional control factor. We ran a further model over the combined data from children and adults which included the same factors as before as well as the fixed effects of age-group (adult, child) and all of the interactions with this factor. This model was specifically inspected to look for the effects of age-group. (Note that, in addition to the means supplied within the text and the difference scores in Figure 5, a full break down of means at pre and post-test by condition can be seen in the R script at http://rpubs.com/ewonnacott/247911).
37
Figure 5. (a) Adult and (b) child discrimination data. Mean increase in percent correct responses
from pre- to post-test (error bars show standard error). Note: for all of the plots within this paper, means are corrected to control for imbalance in counterbalancing of talkers.
Adults: There was no reliable main effect of word-novelty (beta = -0.15, SE = 0.12, z = -1.23, p = .22), suggesting no difference in discrimination for real English words compared to non-words. There was a reliable effect of test-session, indicating an effect of training (beta = -0.52, SE = 0.15, z = -3.44, p = .001). This was qualified by a reliable interaction with talker (beta = -0.59, SE = 0.23, z = -2.53, p = .01), which broke down to show that though there was numerical improvement from pre- to post-test for both talkers, this was only
reliable for the more intelligible talker (female 2: beta = -0.82, SE = 0.23, z = -3.61, p < .001; 87% 93% ; female 1, beta = -0.23, SE = 0.15, z = -1.51, p = .13: 66% 70%) (note that there were no significant higher level interactions involving talker – see S3).
Of critical interest is how participants’ improvement from pre- to post-training was affected by input condition and the novelty manipulations. This is depicted in terms of
38 difference scores in Figure 5. There were no reliable interactions of test-session by word- novelty (beta = 0.07, SE = 0.25, z = 0.29 p = .77), test-session by voice-novelty (beta = 0.19, SE = 0.25, z = 0.76, p = .45), or test-session by word-novelty by voice-novelty (beta = -0.82, SE = 0.50, z = -1.67, p = .10). Contrary to predictions, adults did not show reliably greater improvement in the high-variability than low-variability conditions (test-session by
condition, beta = 0.07, SE = 0.29, z = 0.22, p = .82). There was also no word-novelty by test- session by condition interaction (beta = 0.23, SE = 0.46, z = 0.50, p = .62). However there was a reliable three-way voice-novelty by test-session by condition interaction(beta = -0.91, SE = 0.46, z = -1.99, p = .046) which was qualified by a four-way voice-novelty by word- novelty by test-session by condition (beta = 1.94, SE = 0.92, z = 2.11, p = .035). Breaking down the four-way interaction, voice-novelty by test-session by condition interaction was not reliable for trained words (beta = 0.06, SE = 0.62, z = 0.10, p = 0.93) but was for untrained words (beta = -1.88, SE = 0.67, z = -2.79, p = 0.005). Breaking down the three-way
interaction for untrained words, there was a marginal condition by session interaction for the untrained voice which went in the predicted direction (i.e. more benefit of variability for the untrained voice; beta = 0.89, SE = 0.50, z = 1.76, p = .078) but also a marginal interaction in the opposite direction for the trained voice (beta = -0.99, SE = 0.52, z = -1.91, p = .056). In other words, the interaction rests both on a trend towards a greater benefit of high-variability
input compared with low-variability input for untrained words-untrained voice items (which is predicted since novelty should aid generalization) and a trend towards a greater benefit of
low-variability for untrained words-trained voice items (which is not predicted).
Children: There was no reliable main effect of word-novelty (beta = -0.03, SE = 0.07, z = -0.38, p = .71), suggesting no difference in discrimination for real English words as opposed to non-words. There was a reliable effect of test-session, indicating an effect of training (beta = -0.67, SE = 0.10, z = -6.77, p < .001, pre-test = 62%, post-test = 74%). Again
39 improvement from pre- to post-test is of critical interest and the relevant difference scores are shown in Figure 5. There was a marginal interaction between word-novelty and test-session
(beta = -0.25, SE = 0.15, z = -1.73, p = .084), reflecting a slightly larger improvement for the trained words (60% 75%) than the untrained non-words (63% 74%). In contrast to adults, there was a reliable interaction between test-session and condition (beta = -0.49, SE = 0.20, z = -2.48, p = .013), with children showing a reversed effect to that predicted – i.e., reliably greater improvement in the low-variability condition (18%) than the high-variability
condition (8%). However participants in the low-variability condition were (by chance) lower at pre-test (beta = -0.49, SE = 0.18, z = -2.68, p = .007; low-variability = 56%, high-
variability = 66%) and in fact have not overtaken by post-test (beta = 0.00, SE = 0.19, z = - 0.02, p = .99; low-variability = 74%, high-variability = 74%). The test-session by condition
interaction was not qualified by an interaction with word-novelty (beta = 0.07, SE = 0.29, z = 0.25, p = .80) voice-novelty (beta = -0.02, SE = 0.31, z =- 0.06, p = .95) or word-novelty by
voice-novelty (beta = 0.09, SE = 0.29, z = 0.29, p = 0.77).
Age-group comparison: There was a main effect of age-group, reflecting overall higher performance in adults than children across pre- and post-test (beta = -0.81, SE = 0.13, z = -6.43, p < .001, adults = 79%, children = 68%). Critically, although numerically children improved more from pre- to post-test (see Figure 5) there was no reliable interaction between
age-group and test-session (beta = -0.17, SE = 0.17, z = -1.02, p = .31). The age-group by
test-session interaction was not involved in any reliable higher level interactions with any combination of condition, word-novelty or voice-novelty (p’s > .1) although there was a near
reliable five-way interaction of condition by word-novelty by voice-novelty by age-group by
test-session (beta = -2.00, SE = 1.04, z = -1.92, p = 0.055), reflecting the different effects of these factors in the adult and child models reported above.
40 Summary of discrimination data: Adult participants improved in their
discrimination performance from pre- to post-test, suggesting an effect of training. Although, numerically, adults showed greater improvement in the high- compared to the low-variability condition, the difference was not reliable (or near reliable). There was some tentative
evidence of an interaction between variability and novelty, with the greatest effect of variability evident for maximally novel items. However the key interaction rested both on a predicted benefit of high-variability training for items with untrained words and voices and
an unpredicted benefit of low-variability training for items with untrained words and the trained voice, making it difficult to interpret. Child participants also showed a benefit of training in improvement from pre- to post-test. In contrast to adults, children did show an overall effect of variability, although this was in the opposite direction to that predicted, with greater improvement in the low- compared to high-variability condition, with no evidence that this was affected by the novelty of either word or voice used in the test items or by the
talker used as the new or old voice. However this interaction was driven by (chance)
differences between conditions at pre-test, rather than differences at post-test, and should thus be treated with some caution. In contrast to the previous study by Giannakopoulou et al. (2013a), children did not show reliably greater improvement from pre- to post-tests than adults, i.e. we did not replicate the “plasticity” effect seen in that study.