Effect of noise on speech and gesture

Before proceeding with statistical analyses, we tested all dependent variables (kinematic and acoustic features) for multicollinearity by calculating the variance inflation factor as described by Zuur and colleagues (Zuur, Leno, and Elphick, 2010). Predictors with a variance inflation factor greater than three were excluded from all subsequent analyses (see Supplementary Table ²).

We used linear mixed-effects models to calculate the influence of noise on each of our dependent variables. Mixed-effects models were implemented using the lme4 package⁴⁵. We created nine generalized linear mixed-effects models, each with one of the features of interest (submovements, hold-time, peak velocity, maximum distance, maximum mouth opening, lip movement, lip velocity, speech intensity, speech F0) as the dependent variable, with noise level as a fixed-effect, and a random intercept for each participant and item. To test the significance of these models, we used Chi-square difference tests to compare the models of interest with a null model. As different words may lead to differences in kinematic or acoustic features, we first tested whether a null model containing both participant and word as random intercepts was a better fit to the data than a model with only participant as random intercept. The better fitting model was used as the null model against which the kinematic and acoustic models were tested. We additionally included random slopes for each random term in the model, in keeping with the guidelines of Barr and colleagues⁴⁶. In the case of a singular model fit, the slope for the random effect with the lower explained variance was dropped. If both slopes led to singular fit, we used a model without random slopes. After finding the maximal random effects structure, we again used Chi-square model comparison to determine if this more complex model was a better fit to the data than the null model.

When appropriate, generalized linear mixed models were used together with either Poisson or gamma distributions. We performed this step for variables that are better described as a count of events (i.e., Poisson distribution for submovements) or time accumulation (i.e., gamma distribution for hold-time). We determined that these generalized mixed models were a better fit by visually inspecting the model residuals with a standard Gaussian distribution compared to the alternative Poisson or gamma distribution as well as using Chi-square model comparison between the two. This approach has been advocated to robustly deal with skewed or variable data while maintaining interpretability of results (Lo & Andrews, 2015). This led to submovements being modeled with a Poisson distribution and holdtime being modeling with a gamma distribution. This approach allowed us to keep the full variability of our data, rather than removing outliers⁴⁷. However, to ensure that outliers were not exerting undue influence on our results, we additionally ran our analyses with values beyond the mean + 3 standard deviations removed. These analyses revealed the same pattern of effects as the models that included all data points (see Supplementary Table ³ for an overview of all final models).

Due to the ordinal nature of vertical amplitude, we used a cumulative link linear mixed effects model, implemented with the R package “Ordinal”⁴⁸, rather than the linear model for this aspect of the gesture kinematic analyses. In keeping with the linear mixed models, participant and word were included as random terms in the model. For cumulative link mixed models, significance is determined by directly assessing model parameters, rather than using Chi-square tests model comparisons.