2.5 Experiments
2.5.2 Shared data removal and analysis procedure
2.5.2.1 Subgrouping criteria
For the reliability of the main response time (RT; measured from the onset of the sound file) analysis, the children in both participant groups were subgrouped into two performance
groups: Group 1 and Group 2. In their lexical decision tasks, Harper-Hill et al. (2014a,b) excluded both the TD and children with ASD with fewer than 60% of raw data points remaining after the removal of inaccurate responses and responses with extreme RTs. In this dissertation, this metric is used to separate the children into two groups. Group 1 comprises the children who have above 60% of the raw data points, including fillers, remaining after the removal of inaccurate and extreme responses. This group of children was included in both the accuracy analysis and the main RT analysis. In contrast, only the accuracy data were explored for the children in Group 2. Since Experiment 2 was embedded as fillers of Experiment 1.2 and vice versa, their two subgroups consist of the same participants. Details of the participants in each group are summarized in Table 2.1.
Table 2.1: Subgroups of participants in Experiment 1.1, 1.2, 2, and 3. Experiment 1.1 Experiment 1.2 and 2 Experiment 3 Group 1 Group 2 Group 1 Group 2 Group 1 Group 2 ASD 16 (4) 13 (1) 13 (2) 19 (1) 24 (0) 8 (0) N (Female N) TD 63 (13) 4(0) 41 (6) 18 (4) 58 (0) 1 (0) ASD 9;11 8;9 10;5 9;1 9;6 10;1 M Age TD 8;10 6;9 8;5 6;11 8 5;7 ASD 109.67 83.21 105.88 88.4 102.74 73.8 M NVIQ TD 113.97 96.98 120.36 110.07 117.44 104.58
The adult participants are native speakers of Thai, demographically mixed, and they were recruited through personal contacts and word-of-mouth. All reported normal hearing and normal or corrected-to-normal vision. None of the adult participants scored lower than 70% in all of the experiments, therefore, no exclusion has been made. Table 2.2 shows the details for each of the experiments.
Table 2.2: Adult participants in Experiment 1.1, 1.2, 2, and 3. Experiment 1.1 Experiment 1.2 and 2 Experiment 3
N (Female N) 30(20) 42(27) 31(20)
2.5.2.2 Pre-analysis outlier treatment
After the overall accuracy percentages were obtained from the data, practice trials and fillers were removed. The remaining experimental trials were used for statistical analyses and models for accuracy. I adopted the outlier treatment methods by Baayen and Milin (2010), which allows for minimal initial data trimming before analyzing the RT data.
After the participants have been subgrouped, the child participants in Group 1 and the adult participants with over 70% accuracy (all of the adult participants) underwent the next steps of data trimming. Prime-target pairs of which the responses to either their prime or their target was incorrect were removed. Pairs with extreme RTs to either the prime or the target were then excluded. The cut-off points for extreme RTs were set differently for children and for adults. Figure 2.1 show the child versus adult individual raw RT distributions in the experimental trials of Experiment 3. For children, the extreme RTs were set to <200 and >5500 ms, while they were set to <250 and >3500 ms for adults.
(a) Children. (b) Adults.
Figure 2.1: A comparison between the child versus the adult individual distribution of raw RTs in the experimental trials of Experiment 3.
The remaining data were visualized and tested with the Shapiro-Wilk Test to determine the best transformation. The natural log transformation was found to be the best trans- formation for data sets of all of the RT experiments in this chapter. Afterwards, outlier trimmings were performed on each participant and on each item for which the distribution was not normal according to their Shapiro-Wilk results. In Experiments 1.1 and 1.2, outlier trimmings were performed only on the log-transformed target RTs. On the other hand,
owing to the fact that Experiment 2 and 3 are repetition priming experiments where the difference between prime and target RTs are crucial to the calculation of priming effects, both the log-transformed target RTs and prime RTs were trimmed. Table 2.3 and 2.4 sum- marize the data trimming procedures and percentages of data removal before the final linear mixed regression models were fitted for each experiment.
Table 2.3: Percentages of data removal by experiment (Children in Group 1). Children in Group 1
Experiment 1.1 Experiment 1.2 Experiment 2 Experiment 3 Removal of Experimental Items Initial data points No. removed % removed Initial data points No. removed % removed Initial data points No. removed % removed Initial data points No. removed % removed 1 Inaccurate responses to primes/target 3792 609 16.06% 2430 523 21.52% 2376 483 20.33% 4920 669 13.60% A-priori data trimming
2
Extreme responses (<200 and >5500 ms) both primes and targets
3183 135 4.24% 1907 96 5.03% 1893 63 3.33% 4251 182 4.28% 3 Trimming by subject (target) 3048 77 2.53% 1811 49 2.71% 1830 54 2.95% 4069 63 1.55% 4 Trimming by item (target) 2971 47 1.58% 1762 19 1.08% 1776 26 1.46% 4006 46 1.15% 5 Trimming by subject (prime; If applicable) NA NA NA NA NA NA 1750 44 2.51% 3960 43 1.09% 6 Trimming by item (prime; If applicable) NA NA NA NA NA NA 1706 34 1.99% 3917 25 0.64% Total removal by a-priori data trimming 259 8.14% 164 8.60% 221 11.67% 359 8.45% 7 Model criticism 2924 58 1.98% 1743 38 2.18% 1672 44 2.63% 3892 91 2.34%
Table 2.4: Percentages of data removal by experiment (Adults). Adults
Experiment 1.1 Experiment 1.2 Experiment 2 Experiment 3 Removal of Experimental Items Initial data points No. removed % removed Initial data points No. removed % removed Initial data points No. removed % removed Initial data points No. removed % removed 1 Inaccurate responses to primes/target 1440 60 4.17% 1680 44 2.62% 1610 21 1.30% 1860 37 1.99% A-priori data trimming
2
Extreme responses (<250 and >3000 ms) both primes and targets
1380 28 2.03% 1636 5 0.31% 1554 11 0.71% 1823 9 0.49% 3 Trimming by subject (target) 1352 29 2.14% 1631 20 1.23% 1543 31 2.01% 1814 11 0.61% 4 Trimming by item (target) 1323 35 2.65% 1611 4 0.25% 1512 5 0.33% 1803 3 0.17% 5 Trimming by subject (prime; If applicable) NA NA NA NA NA NA 1507 22 1.46% 1800 25 1.39% 6 Trimming by item (prime; If applicable) NA NA NA NA NA NA 1485 10 0.67% 1775 8 0.45% Total removal by a-priori data trimming 92 6.67% 29 1.77% 79 5.08% 56 3.07% 7 Model criticism 1288 26 2.02% 1607 41 2.55% 1475 28 1.90% 1767 36 2.04%
2.5.2.3 Modeling
Mixed effects logistic regression models were run separately on the accuracy data of children in Group 1, children in Group 2, and adults in each experiment, using the lme4 package (Version 1.1.12; Bates et al. 2015) with the extension lmerTest package (Version 2.0.32;
for obtaining p-values; Kuznetsova et al. 2016) in the R software (Version 3.3.1; R Core Team 2016) and MuMIn package (Version 1.43.6; for obtaining marginal and conditional r- squared; Barton 2019). For logistic regression models, the bobyqa optimizer from the minqa package (Version 1.2.4; Bates et al. 2014) was applied. The default treatment coding scheme was used to compare between each categorical group. Baseline conditions and participant groups were re-leveled to obtain comprehensive pairwise comparisons of factors. Confidence intervals were reported at 95%. As for the RT data, a linear mixed-effects model was only applied to the children in Group 1 and the adults, using the same software and packages. For the RT data, post-fitting model criticism was performed to remove the outliers situating at least 2.5 standard deviations away from the mean in both tails. The results from the refitted models are presented in the next section.