Engaging proactive control: Influences of diverse language experiences using insights from machine learning

Abstract

We used insights from machine learning to address an important but contentious question: Is bilingual language experience associated with executive control abilities? Specifically, we assess proactive executive control for over 400 young adult bilinguals via reaction time (RT) on an AX continuous performance task (AX-CPT). We measured bilingual experience as a continuous, multidimensional spectrum (i.e., age of acquisition, language entropy, and sheer second language exposure). Linear mixed effects regression analyses indicated significant associations between bilingual language experience and proactive control, consistent with previous work. Information criteria (e.g., AIC) and cross-validation further suggested that these models are robust in predicting data from novel, unmodeled participants. These results were bolstered by cross-validated LASSO regression, a form of penalized regression. However, the results of both cross-validation procedures also indicated that similar predictive performance could be achieved through simpler models that only included information about the AX-CPT (i.e., trial type). Collectively, these results suggest that the effects of bilingual experience on proactive control, to the extent that they exist in younger adults, are likely small. Thus, future studies will require even larger or qualitatively different samples (e.g., older adults or children) in combination with valid, granular quantifications of language experience to reveal predictive effects on novel participants. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

Figure 1.

Illustration of the group-level, aggregate effects for AX-CPT reaction time in milliseconds (A) and percent accuracy (B). Error bars illustrate 1 standard error of the mean (SEM). AY trials were overall slower and less accurate compared to BX trials, suggesting that, overall, participants approached the task with a proactive control strategy.

Figure 2.

Illustration of the linear mixed effects regression model-estimated effects for the best fitting linear mixed effects regression model that included interactions between trial type and language entropy components (i.e., Model 2b). The plot illustrates AX-CPT reaction time as a function of language entropy components: general entropy (A) and work entropy (B). An increase in general entropy was associated with a divergence between the AY and BX trial types, consistent with a greater engagement of proactive control. An increase in work entropy was associated with a general increase in reaction time. Error bands illustrate 1 standard error of the mean (SEM).

Figure 3.

Illustration the LASSO regression results for AX-CPT reaction time using leave-one-out cross-validation. A. Illustrates prediction error (RMSE) as a function of the regularization parameter. The left-most dotted vertical line illustrates the value that minimized RMSE, and the right-most dotted line illustrates the parameter resulting in the simplest model within 1 SEM of the minimizing model (i.e., the most parsimonious, predictive model). Effects that survive the minimizing value of lambda are colored in purple (i.e., a main effect of work entropy). Error bands illustrate 1 SEM. B. Illustrates model estimates (unstandardized B) as a function of the penalizing parameter. As the value of the parameter increases (left-to-right) effects are regularized to zero. Dotted lines again indicate minimizing value of lambda and the value leading to the most parsimonious, predictive model. The model that minimized RMSE contained several effects and interactions related to individual differences in bilingual language experience, but the most parsimonious model contained only effects of trial type.

Figure 4.

A zoomed-in reproduction of Figure 3B. This figure illustrates model estimates (unstandardized B) as a function of the penalizing parameter.

Figure 5.

Illustration of magnitude of the model estimates (unstandardized B) for cross-validated LASSO regression at the minimizing value of lambda.