Processing Differences between Monolingual and Bilingual Young Adults on an Emotion n-back Task

Abstract

Bilingualism is associated with enhancement of executive control (EC) across the lifespan. Working memory and non-verbal emotion regulation both draw upon EC mechanisms so may also be affected by bilingualism, but these relationships are not fully understood. These relationships were explored using an n-back task with distracting emotional stimuli administered to young adults while continuous EEG was recorded. Monolinguals were faster but less accurate on the 2-back than bilinguals, and monolingual accuracy was more impeded by the presence of emotional stimuli than was that of bilinguals. The P300 event-related potential, a neural signature of working memory processing in the n-back, had smaller amplitudes in both groups on the 2-back than the 1-back, but attenuation in response to distracting emotional stimuli was greater for bilinguals than monolinguals. P300 latencies were also differentially affected by emotional stimuli in each group: Bilingual latencies were constant across emotions but monolingual latencies increased from neutral to angry conditions. In general, bilingual performance was less impacted by the emotional distraction than was that of the monolinguals. Additionally, bilinguals adjusted to the changing demands of the 1-back and 2-back conditions by recruiting neural networks to support different behavioral outcomes than monolinguals.

Bilinguals typically outperform monolinguals on executive control tasks and recruit different functional networks than those used by monolinguals when performing these tasks, speculatively reflecting changes in allocation of attentional resources (Bialystok, 2017). These differences are hypothesized to occur as a result of the need for bilinguals to manage the constant competition arising from semantic representations across the two languages. Both languages are jointly activated in the bilingual mind, so individuals must constantly select the relevant language and avoid interference from the language not in use (Kroll, Bobb, & Hoshino, 2014). This selection is carried out at least in part by domain-general executive control networks (Luk, Green, Abutalebi, & Grady, 2012). Therefore, the need for executive control to negotiate this competition throughout life is thought to reorganize and potentially strengthen executive control in ways that may then generalize beyond language selection. Language group differences in executive control have been demonstrated in young adults using such tasks as the Stroop (e.g., Bialystok, Poarch, Luo, & Craik, 2014; Hernández, Costa, Fuentes, Vivas, & Sebastián-Gallés, 2010), flanker (Costa, Hernandez, & Sebastian-Galles, 2008; Yang & Yang, 2016), and Simon (e.g., Bialystok, Craik, & Luk, 2008; Verreyt, Woumans, Vandelanotte, Szmalec, & Duyck, 2016) tasks, although differences are not always found (e.g., Bialystok, Martin, & Viswanathan, 2005; Paap & Greenberg, 2013; von Bastian, Souza, & Gade, 2016). Nonetheless, differences between language groups on these tasks are more reliably found for children and older adults (review in Antoniou, 2019).

Like bilingual language processing, emotion regulation also requires executive control to manage the expression, duration, and intensity of positive and negative emotions (Calkins & Marcovitch, 2010). Frequently, an automatic emotional response to a stimulus is not optimal and an altered response may be more appropriate. For instance, down-regulating the experience of anxiety that is triggered by the thought of an upcoming job interview could be beneficial to the interviewee. Because emotional stimuli automatically engage attentional resources, the attentional control and top-down processes that are associated with executive control are particularly relevant for emotion regulation (Oschner & Gross, 2005).

Both bilingual language processing and emotion regulation require the ongoing recruitment of executive control processes. Since extensive use of specific networks leads to neuroplastic changes in their efficiency (Pascual-Leone, Amedi, Fregni, & Merabet, 2005), it may be that bilingual experience also impacts the efficiency or automaticity of emotion regulation. Research investigating this possibility of an intersection between bilingualism and emotion processing has largely focused on bilinguals performing verbal tasks in their two languages. The present study examines the hypothesis that bilingualism impacts emotion processing more broadly through modifications in executive control.

Although little evidence exists for that hypothesis, studies addressing a related question show that bilinguals exhibit differences between the way emotional verbal stimuli are processed in their first (L1) and second languages (L2; see Pavlenko, 2012, for review). For example, they rate emotional words and phrases in the L2 as less intense than their L1 equivalents (Caldwell-Harris & Ayçiçeği -Dinn, 2009; Caldwell-Harris, Tong, Lung, & Poo, 2010; for a similar finding with multilinguals see Dewaele, 2008). They also use more emotional words (Marian & Kaushanskaya, 2008) and converse about taboo topics longer (Bond & Lai, 1986) in their L2 than in their L1, indicating greater emotional detachment in that language.

The emotional Stroop task provides a method of comparing the processing of emotional stimuli across languages on a task requiring executive control. Bilingual participants see words printed in colored ink and respond to the color, as typical for a Stroop task. Emotion words are interspersed with emotionally neutral words, and results show that performance is impeded on emotion trials. Some studies have reported a smaller RT cost for emotion trials when bilinguals perform the task in their L2 than in their L1 (Winskel, 2013) but other studies have found similar performance in both languages (Eilola & Havelka, 2011; Eilola, Havelka, & Sharma, 2007; Sutton, Altarriba, Gianico, & Basnight-Brown, 2007).

Physiological measures have also been used to address how emotion processing differs across known languages, with skin conductance being the most common. Consistent with behavioral measures, emotion words and phrases in L2 typically elicit a reduced skin conductance response, indicating lower degree of emotional arousal (e.g., Caldwell-Harris & Ayçiçeği -Dinn, 2009; Harris, 2004; Harris, Ayçiçeği, & Gleason, 2003). Other studies have sought to identify the neural correlates of emotional differences between L1 and L2. Event-related potentials (ERPs) generated by emotional words in L2 show delayed early posterior negativity, suggesting longer processing times for emotion in L2 than for the faster, more automatic processing in L1 (Chen, Lin, Chen, Lu, & Guo, 2015; Conrad, Recio, & Jacobs, 2011 Opitz & Degner, 2012). Similarly, fMRI studies have identified neural differences between languages. Chen and colleagues (2015) found that emotional words in a lexical decision task produced activations in the left cerebellum for both languages, but activation was greater in the L2. The authors attribute this result to the cerebellum’s role in semantic processing and argue that the increased activation indicates more effortful semantic processing for emotional words in L2 than for L1 where processing is relatively automatic processing and does not require additional resources for stimulus recognition. The amygdala, a region relevant for emotion processing, has also been shown to be sensitive to the language in which the emotional content is presented. Here, activation has been found to be greater while reading emotional text in L1 than in L2, suggesting more emotion processing in L1 than L2 (Hsu, Jacobs, & Conrad, 2015).

These studies suggest that bilinguals recruit executive control differently when engaged in emotion regulation in their two languages. What has not been investigated is whether emotion regulation itself is modified by bilingualism. If so, the expectation is that monolinguals and bilinguals will perform differently on emotion regulation tasks. Such evidence would both contribute to our understanding of how emotion regulation is carried out in terms of the involvement of executive control processes and extend out knowledge of the set of processes that are modified by bilingualism.

Working memory (WM) is a natural aspect of executive control in which to investigate effects of language experience and lends itself to incorporating emotion stimuli. Language processing draws upon WM resources, and there is evidence that bilinguals rely on WM to a greater extent than monolinguals during lexical retrieval (Kaushanskaya, Blumenfeld, & Marian, 2011). It follows, then, that such WM practice, in conjunction with domain-general enhancement of executive control found in bilinguals, would lead to better performance by bilinguals than monolinguals on WM tasks. However, studies investigating WM in bilinguals and monolinguals have produced mixed results. Using non-verbal WM tasks, bilinguals have been shown to have similar (e.g., Bonifacci, Giombini, Bellocchi, & Contento, 2011; Bialystok, Craik, & Luk, 2008; Raitu & Azuma, 2015), better (e.g., Bialystok, Poarch, Luo, & Craik, 2014; Luo, Craik, Moreno, Bialystok, 2013; Morales, Calvo, & Bialystok, 2013), and poorer performance than monolinguals on verbal WM tasks (e.g., Luo, Craik, Moreno, Bialystok, 2013; Raitu & Azuma, 2015).

These inconsistent findings were examined in a meta-analysis that showed a significant positive effect of bilingualism on WM performance (Grundy & Timmer, 2017). Another metaanalysis that included second-language proficiency as a continuous measure of degree of bilingualism reported a significant positive relationship between level of proficiency and WM (Linck, Osthus, Koeth, & Bunting, 2014). Moderator analyses in that study showed that effect size was greatest for the association between proficiency and complex span tasks. This result is consistent with the suggestion that the effect of bilingualism on WM is most observable when the task demands are high (Sullivan, Prescott, Goldberg, & Bialystok, 2016).

For these reasons, investigations of the effect of bilingualism on WM performance are best assessed using complex nonverbal tasks. A standard task for evaluating WM is the n-back. Participants decide whether a stimulus matches a previously-presented target from a designated interval, that is, one trial previously, two trials previously, and so on. The task rapidly builds up proactive interference and so executive control is required to respond only to the targets and ignore the other stimuli even though they are familiar. The task is also easily adapted to address specific questions. Therefore, including emotional elements in the display and using nonverbal stimuli allows one to compare performance by monolingual and bilingual participants and separately evaluate the response to the WM demands and the emotion regulation demands.

One study has used this approach with children. Monolingual and bilingual children who were 9-years old performed an n-back task in which they decided whether a letter presented in the center of the display had appeared one (1-back) or two (2-back) trials previously (Janus & Bialystok, 2018). Each letter was flanked by two identical faces conveying happy, angry, or neutral expressions. There were significant differences in reaction time (RT) across the emotion conditions, but these did not interact with language group. For the WM demands, bilingual children were more accurate throughout the task and significantly slower than monolingual children on the 2-back condition, regardless of the emotion. For children, therefore, bilingualism was associated with the use of a different strategy to perform the WM n-back task but not in how the emotional content of the trials was handled.

The present study extends these results by investigating the impact of bilingualism on the intersection of executive control and emotion regulation in young adults performing an adapted n-back task that includes distracting emotional stimuli. Because studies with young adults frequently show no behavioral difference between monolingual and bilingual groups (Antoniou, 2019), the task was performed while EEG was recorded to provide more detailed information about processing. The relevant ERP components are the P300 for working memory and N170 for face processing.

The P300 component is a positive-going voltage deflection over centroparietal regions that occurs approximately 300–500 ms post-stimulus. P300 amplitude reflects stimulus categorization and attention to task stimuli, with larger amplitudes indicating greater attention to target stimuli, whereas P300 latency reveals the speed of processing, with shorter P300 latency indicating more rapid evaluation of stimuli (see Polich, 2007). Using the n-back task, P300 amplitude has been shown to reduce as WM load increases (e.g., Ozen, Itier, Preston, & Fernandes, 2013; Scharinger, Soutschek, Schubert, & Gerjets, 2015), reflecting reallocation of resources from simple stimulus matching to more executive processing of rising memory demands (Mecklinger, Kramer, & Strayer, 1992; Watter, Geffen, & Geffen, 2001). Thus, a reduction in P300 amplitude in centroparietal regions signals an efficient response to increased task demands.

The N170 is an early negative-going waveform that is most prominent in posterior temporal sites. This component is understood to be relevant for face processing (Bentin, Allison, Puce, Perez, & McCarthy, 1996; Eimer, 2011), and there is accumulating evidence that the emotional content of facial expressions amplifies the N170 (Blau, Maurer, Tottenham, & McCandliss, 2007; Hinojosa, Mercado, & Carretié, 2015; Schindler, Zell, Botsch, & Kissler, 2017; Wronka & Walentowska, 2011; although see Rellecke, Sommer, & Schact, 2013, for a contrary interpretation). Therefore, measurement of the N170 provides an index of emotion processing. Consequently, group comparison of N170 amplitude can identify language group differences in processing the emotional content of such stimuli.

The main hypothesis was that the bilingual group would more easily regulate automatic emotion processing in a working memory task than the monolingual group. Although behavioral differences were found in the study with children, no such differences were predicted given language groups typically do not differ behaviorally in the young adult population. Differences were expected, however, in the electrophysiology of the responses, with bilinguals showing smaller P300 amplitude than monolinguals in the more difficult conditions.

Although some previous research has shown larger P300 amplitude for bilinguals than monolinguals performing executive function tasks (Barac, Moreno, & Bialystok, 2016; Moreno, Wodniecka, Tays, Alain, & Bialystok, 2014; Sullivan, Janus, Moreno, Astheimer, & Bialystok, 2014), only one study to date has examined these effects in an n-back task. Morrison, Kamal, and Taler (2018) reported that bilinguals produced larger P300s than monolinguals, but the authors note that participants were performing at ceiling and they did not find typical n-back task effects, undermining any interpretation of the results. Given the sensitivity of the P300 to task difficulty, the prediction for the present study was that P300 will be reduced in the 2-back condition for both groups but that there will be greater modulation across conditions for bilinguals, reflecting more efficient WM performance in light of emotional distraction and given equivalent behavior. For N170, the prediction was that bilinguals will show a smaller amplitude than monolinguals, indicating an attenuated physiological response to the emotional stimuli. Finally, we predicted that bilinguals will generate shorter P300 latencies than monolinguals, reflecting group differences attentional processing speed.

Method

Participants

Sixty-three young adults (31 bilinguals, 32 monolinguals; 18–26 years) participated in this study. Bilingual participants spoke a variety of non-English languages¹ and had been first exposed to a bilingual environment near birth (M = 3.72, SD = 14.76 months). All participants provided written informed consent and received course credit for their participation. Participants were all right-handed and had normal or corrected-to-normal vision, and none indicated a history of neurological disorder or the use of psychoactive medications. Four participants were excluded from analyses because of a technical issue resulting in non-usable event codes, low accuracy stemming from an absence of responses on roughly half of the trials, history of concussion, or EEG data that were greater than four standard deviations from the mean. The final sample, therefore, consisted of 29 bilinguals and 30 monolinguals. A subsample of these participants including 23 bilinguals and 26 monolinguals was also administered a control version of the task to establish a baseline for each group².

Background Tasks

Participants were administered a battery of written measures to assess relevant background variables. The Language and Social Background Questionnaire was used to assign language group membership (LSBQ; Anderson, Mak, Keyvani Chahi, & Bialystok, 2018). This questionnaire quantifies second-language experience in a variety of settings to gain a detailed profile of understanding and speaking ability in known languages, the extent of exposure to various languages, and information about how each language is used.

A measure of verbal and non-verbal intelligence was collected using the Shipley-2 (Shipley, Gruber, Martin, & Klein, 2009). Only the Vocabulary and Block Patterns subtests were administered. This provided standardized scores of crystalized and fluid intelligence for each participant.

The Emotion Regulation Questionnaire (ERQ; Gross & John, 2003) was administered to obtain scores indicating emotion regulation strategies. This 10-question measure yields sub-scores reflecting the use of two strategies: cognitive reappraisal and expressive suppression. The former reflects an internal strategy focused on reconstruing one’s perception of a situation that may alter the related emotional response. The second concerns the modulation of an explicit behavioral response to an emotion-provoking context.

Individual differences in motivational systems were measured with the use of the BIS/BAS scales (Carver & White, 1994). This measure provides four sub-scores that are associated with participants’ expression of the behavioral inhibition and activation systems (BIS/BAS), two constructs that are related to individual emotional differences.

Working Memory n-back Task

WM performance was assessed using an n-back task derived from Ladouceur et al. (2009), incorporating non-verbal emotional stimuli and programmed in Neurobehavioral Systems Inc. Presentation software (version 14.6). Emotional stimuli consisted of facial expressions, classified as neutral, happy, or angry, obtained from the NimStim Face Stimulus Set (Tottenham et al., 2009). A subset of 15 actors from the overall set was used, and all three expressions were included for all actors. Images were elliptically cropped to retain only the face, and resulting images were grey-scaled. A control condition, administered to a subset of the sample, presented shapes instead of emotional faces. Shapes were rounded or angular and were also presented in grey-scale. All stimuli were centered on a black background and up-scaled to 800 × 800 pixels to remain viewable beneath target stimuli. Target stimuli were 60-point font uppercase letters (B, F, H, K, M, Q, R, and X) that were centered and superimposed over the stimulus. Stimuli were presented for 500 ms, or until the participant made a response, followed by a centered fixation cross appearing for 500 ms. The timing between trials was jittered such that the inter-trial interval ranged from 1800 ms, incrementing by 25 ms, to 2000 ms.

Participants completed 1-back and 2-back versions of this task. The 1-back task required participants to decide whether each letter presentation was the same as the letter that immediately preceded it, and the 2-back task required that this judgment be made on the letter presented two presentations prior. Participants indicated their judgments by pressing the innermost button on one of two mouse input devices.

Both the 1-back and 2-back conditions began with a practice block of 10 trials using the control shape stimuli. Each n-back condition contained six emotional blocks of 45 trials each, for a total of 12 blocks over the two conditions. The 45 trials in each block included 15 trials of each facial expression (neutral, happy, and angry) presented in random order. One-third of the trials in each block were randomly assigned as targets, the remainder being non-targets. To assess performance in the absence of emotional stimuli, a subset of 49 participants received non-emotional control trials in both 1-back and 2-back conditions, presented as two blocks of 45 trials sandwiching the six emotional blocks. Participants were given a break between blocks. The order of presentation was 1-back followed by 2-back.

Procedure

Upon arrival, participants completed an informed consent form and were given a description of the electroencephalogram (EEG) system. The experimenter administered the LSBQ in an interview style and then the participant completed the ERQ, BIS/BAS, and Shipley-2 Vocabulary. The Shipley-2 Block Patterns was given during a short between the n-back tasks.

The n-back task was administered in a dimly lit room containing the EEG system. Participants were seated 50 cm from the computer (Dell OptiPlex 760), and the monitor and chair heights were adjusted for the participant’s comfort. The experimenter remained in the room as the participant completed a practice block of the n-back task, and practice was repeated up to three times until participants achieved an accuracy greater than chance. After practice, the experimenter exited to the adjacent room where they could observe the recording for any issues while still being available to the participant via a connected microphone. After the experiment, participants were debriefed on the purpose of the study.

Electrophysiology

Continuous EEG data were collected while participants completed the n-back task using a BioSemi amplifier system (Amsterdam, BioSemi Active 2) with a 64 Ag-AgCl electrode array applied in standard positions (International 10/20 system sites) with a sampling rate of 512 Hz. All electrodes were referenced to the CMS (Common Mode Sense) electrode and grounded using the DRL (Driven Right Leg) electrode. An impedance criterion of 20 kΩ was used to determine which electrodes to adjust after application. If this criterion could not be achieved, impedances were noted and reduced as much as possible prior to recording. Two electrodes were used to measure eye movement. One electrode was placed 1 cm from the canthus while another was located below the eye upon the cheekbone. A third electrode was placed over the mastoid. This placement was performed for both the left and right eyes/mastoids.

EEG preprocessing

Data were preprocessed in Matlab using the EEGLAB and ERPLAB toolboxes (Delorne & Makeig, 2004; Lopez-Calderon & Luck, 2014). A second-order Butterworth filter (0.1–30 Hz bandpass) was applied on the continuous data to retain signal within normal brain activity frequencies while reducing non-brain signal. Interpolation of continuous channels was performed on the filtered data. Electrodes that had been removed at the time of testing were interpolated, otherwise a kurtosis criterion (z-score > 5) was used to automatically identify channels eligible for interpolation. Data were then re-referenced to a common reference and segmented into 1200 ms event-related potentials that comprised a 200 ms baseline (used for baseline correction) and 1000 ms event response window locked to stimuli presentations. Independent component analysis (ICA) was performed on the segmented data and ocular, muscular, line noise, and channel noise components were selected for removal. Candidate components were selected based on stereotypical time courses, spectral, and topological patterns for each source of distortion. Visual inspection verified that artifactual distortions were successfully removed (average components removed = 15.61, SD = 6.64, out of 70 components). Following ICA denoising, data were distorted for one participant. In this case, epochs found to contain substantial drift were removed (12 trials) and ICA re-run. Lastly, ERPs on correct trials were averaged for each stimulus type (i.e., control, neutral, happy, and angry) and target status (target vs. non-target events). Final preprocessed averages were exported as text files for use in partial least squares analysis.

Event-related potential analysis

To investigate group differences in WM processing, mean amplitudes of the P300 event-related component were calculated. The P300 was measured using a time window occupying 300–450 ms from the stimulus presentation. Average potentials from the window were then used in analyses. Peak latencies were also extracted given findings of bilingual differences on this measure. Latency was determined using the same time window to find a peak that was greater than the adjacent 5 data points (9.8 ms). Latency was the timestamp associated with this peak. Based on the typical centroparietal definition of the P300, as well as visual inspection of where the component’s response was most prominent in the data, electrodes Cp1, Cpz, Cp2, P1, Pz, and P2 were averaged for analyses.

To quantify the negative deflection of the N170 component, peak amplitude was used because the waveform was too narrow for a meaningful analysis of mean amplitude. Local peaks were defined as negative-going values with a greater absolute potential than the five adjacent data points (9.8 ms) within a time window of 150–200 ms following stimulus presentation.

Partial least squares analysis

To investigate whole-brain patterns of task-related activity, relative to behavioral outcomes, a behavior partial least squares analysis (PLS) was performed using the Matlab add-on, PLSgui (Shen, 2009; McIntosh, Bookstein, Haxby, & Grady, 1996; McIntosh, Chau, & Protzner, 2004; see Krishnan, Williams, McIntosh, & Abdi, 2011, for a review of the methodology). PLS is a data-driven, multivariate technique, similar to factor analysis, that extracts underlying latent variables (LVs) from the data set that maximally explain the covariance. Given its multivariate nature, this method is particularly well-suited to whole-brain analysis of neuroimaging data (e.g., EEG; see Bailey, Mlynarczyk, & West, 2016; West, Bowry, & Krompinger, 2006; West & Bowry, 2005) that contain a high dimensionality and would otherwise suffer from stringent correction for multiple univariate comparisons.

Behavioral PLS performs singular value decomposition (SVD) on a correlation matrix derived from two separate matrix inputs. A brain matrix is formed by stacking condition-wise submatrices with columns containing measured brain activity and rows representing participants. A complementary behavior matrix contains condition-stacked behavioural measures. These two matrices are then combined to form a single matrix that stores condition-wise correlations between brain activity and behavior in each row vector. The SVD of this matrix generates a set of orthogonal LVs organized as brain and design salience vectors in addition to a diagonal matrix of singular values. Brain saliences indicate brain activity that is most strongly associated with the previous matrix of correlations. Projecting the original brain activity onto these brain saliences and summing within condition produces scalp scores. Scalp scores are a summary score that reflects how closely an individual’s brain data, for a given condition, approximates the patterns evident in the LV. Correlating these scores with behavioral measures may be used to assess condition or group differences in brain-behavior relationships.

Bootstrapping is performed to discern the reliability of the derived salience patterns. Using the bootstrapped sample, a standard error is calculated for each salience value. Dividing saliences by their standard error results in a standard score that is roughly analogous to a z-score (Efron & Tibshirani, 1986). Therefore, these bootstrap ratios (BSRs) may be considered reliable at the 95% confidence level when they exceed a value of two. Additionally, bootstrapping is used to construct bootstrapped confidence intervals around correlations between scalp-scores and behavior. This may be used to assess group differences as well as the reliability of a group’s brain-behavior correlation.

Permutation testing is also carried out on the singular values. Here, each iteration of the permuted singular value is compared to the original value calculated. If the permuted values are greater than the original for less than 5% of the permutations, then the LV corresponding to that singular value is deemed significant. For analysis, bootstrapping and permutation iterations were both set to number 1000.

Results

Background Measures

Mean scores for the background measures are presented in Table 1. One-way ANOVAs were conducted to compare groups on each measure. There were significant differences for L2 speaking ability, F(1, 57) = 410.16, p < .001, and understanding, F(1, 57) = 868.22, p < .001, English speaking ability F(1, 57) = 11.90, p = .001, and English understanding, F(1, 57) = 7.08, p = .01, although all participants had high English proficiency and were studying in English. There was an unexpected group difference in the expressive suppression sub-score of the ERQ, F(1, 57) = 5.00, p = .03, in which bilinguals demonstrated higher scores than monolinguals, suggesting that they were more likely to supress outward, behavioral expressions of emotion such as masking emotional facial expressions (Gross & Levenson, 1993)³. No group differences were found for Shipley-2 Vocabulary, F < 2.0, or non-verbal intelligence, F < 0.10. Similar analyses conducted on the background measures for the subset of 49 participants who completed the control condition showed the same results.

Table 1.

Descriptive Statistics for Background Measures and Significance Level of Group Differences.

	Monolingual	Bilingual
Measure	M (SD)	M (SD)	p
Age	19.87 (2.65)	20.17 (1.95)	.62
LSBQ
English Speaking	9.97 (0.18)	9.05 (1.44)	< .01 *
English Understanding	9.97 (0.18)	9.36 (1.23)	.01 *
L2 Speaking	0.83 (1.14)	8.59 (1.72)	< .01 *
L2 Understanding	1.36 (1.19)	9.47 (0.91)	< .01 *
L2 Speaking Frequency	0.14 (0.35)	5.02 (2.05)	< .01 *
L2 Listening Frequency	0.24 (0.69)	5.6 (2.57)	< .01 *
Shipley-2
Vocabulary Standard	102.9 (12.87)	98.39 (12.16)	.18
Blocks Standard	103.07 (10.76)	103.31 (12.17)	.94
ERQ
Cognitive Reappraisal	5.31 (0.96)	4.86 (1.01)	.08
Expressive Supression	3.67 (1.28)	4.34 (0.99)	.03 *
BIS-BAS
Drive	12.14 (2.14)	11.52 (2.26)	.29
Fun Seeking	12.7 (2.23)	12.14 (2.55)	.37
Reward Responsiveness	17.6 (1.75)	17.93 (1.91)	.49
BIS	20.86 (3.99)	22.24 (3.73)	.18

Note. Scores for the LSBQ measures are out of 10. Shipley scores are standardized around a mean of 100. ERQ is out of 7. BAS Drive and Fun Seeking are each out of 16, Reward Responsiveness is out of 20, and BIS is out of 28.

Working Memory Behavioral Performance

Behavioral results are presented in Figures 1 and 2 for accuracy and reaction time, respectively. Accuracy and reaction times from the shape control task completed by the subset of participants (n = 49) were analyzed using a 3-way ANOVA for group, WM load (1-back, 2-back), and trial type (target, non-target). For accuracy, 1-back scores were higher than 2-back scores, F(1, 47) = 81.95, p < .001, η_p² = .64, and non-target trials were more accurate than target trials, F(1, 47) = 52.11, p < .001, η_p² = .53, with no effect of language group, F(1, 47) = 1.91, p = .17, and no interactions, Fs < 2.0. For RTs (Figure 2A), there was a main effect of memory load, F(1, 47) = 11.88, p = .001, η_p² = .20, whereby participants were faster on the 1-back than 2-back, and a main effect of trial type, F(1, 47) = 11.90, p = .001, η_p² = .20, with faster responding on target trials than non-target trials. There was a marginal interaction between language group and WM load, F(1, 47) = 3.65, p = .06, suggesting a trend for monolinguals to be faster on the 2- back than bilinguals, but no effect of language group, F(1, 47) = 1.56, p = .22, or other interactions, Fs < 1.60.

Figure 1.

Accuracy (proportion correct) by group, n-back level, and trial type, for (A) control and (B) emotion trials. Error bars represent standard error.

Figure 2.

Reaction times by group, n-back level, and trial type, for (A) control and (B) emotion trials. Error bars represent standard error.

Emotion trials were analyzed with a 4-way ANOVA for group, expression (neutral, happy, angry), WM load, and trial type. For accuracy (Figure 1B), there were main effects of memory load, F(1, 57) = 259.93, , p < .001, η_p² = .82, with higher accuracy on the 1-back task than the 2-back task, and trial type, F(1, 57) = 72.40, , p < .001, η_p² = .56, indicating that accuracy was higher for non-target trials than target trials. There was a marginally significant effect of language group, F(1, 57) = 3.82, p = .06, no main effect of emotion expression, F(2, 114) = 0.14, p = .87, but a significant language group by emotion expression interaction, F(2, 114) = 4.35, p = .02, η_p² = .07, whereby bilinguals were more accurate than monolinguals on the angry trials with a sub-threshold tendency to also be more accurate on happy expressions (uncorrected p = .04) but the groups were equivalent on the neutral expressions. No other interactions were significant, Fs < 3.5.

The analysis of RTs (Figure 2B) revealed a main effect of memory load, F(1, 57) = 13.40, p < .001, η_p² = .19, indicating faster performance on the 1-back than 2-back task, and a main effect of trial type, F(1, 57) = 7.63, p = .008, η_p² = .12, with faster RTs on target trials. There was no effect of language group, F(1, 57) = 1.95, p = .17, but a significant interaction between language group and memory load, F(1, 57) = 4.70, p = .03, η_p² = .08. This interaction revealed that bilinguals had a greater difference in RTs between the 1-back and 2-back than monolinguals, F(1, 57) = 5.68, p = .02, η² = .10. No other effects or interactions were significant, Fs < 2.70.

ERP Waveforms

P300 mean amplitude and latency for the control condition are presented in Figure 3 and were analyzed using 3-way ANOVAs for group, WM load, and trial type. For mean amplitude, there was a main effect of memory load, F(1, 47) = 31.084, p < .001, η_p² = .40, with larger amplitudes for the 1-back than 2-back, a main effect of trial type, F(1,47) = 85.87, p < .001, η_p² = .65, with target trials eliciting a greater P300 response, and an interaction between them, F(1, 47) = 9.71, p = .003, η_p² = .17. Computing difference scores between 1-back and 2-back amplitude revealed that differences between WM conditions were greater on target trials than on non-target trials, F(1, 48) = 9.86, p = .003, η_p² = .17. There were no other significant effects, Fs < 2.0. Peak latency showed a significant effect of trial type, F(1, 47) = 6.53, p = .01, η_p² = .12, with longer latencies on target trials than non-target trials, and this was qualified by an interaction between trial type and memory load, F(1, 47) = 8.10, p = .006, η_p² = .15, whereby latencies were longer for target than non-target trials on the 1-back, but not for the 2-back. No other effects were significant, all Fs < 1.0.

Figure 3.

The P300 component evoked by control trials. Panel A depicts ERP waveforms, with the analysis window (i.e., 300 – 450 ms) highlighted. (B) Mean amplitudes and (C) peak latencies from the window are summarized. Error bars represent standard error.

P300 amplitude for emotion trials are shown in Figure 4 and were analyzed with a 4-way ANOVA for group, expression, WM load, and trial type (Figure 4). There was a main effect of memory load, F(1, 57) = 38.56, p < .001, η_p² = .40, indicating that the 1-back produced higher P300 potentials than the 2-back, a main effect of trial type, F(1, 57) = 57.89, p < .001, η_p² = .050, whereby target trials elicited a larger P300 than non-target trials, and a main effect of emotion, F(2, 114) = 3.25, p = .04, η_p² = 056, in which angry trials produced larger P300 amplitudes than neutral condition, F(1,58) = 5.08, p =.03, η_p² = .08, with no differences between any other conditions. There was no main effect of language group, F(1, 57) = 0.36, p = .66, but there was a significant language group by WM load interaction, F(1, 57) = 6.60, p = .01, η_p² = .10, and interaction between language group, WM load, and trial type, , F(1, 57) = 7.93, p = .007, η_p² = .12. The interaction showed that the difference between the P300 amplitude in the 1-back and 2-back conditions was larger for bilinguals than monolinguals, F(1, 57) = 9.29, p = .003, η² = .16, and found on target trials but not on non-target trials, F(1, 57) = 1.41, p = .24. These effects are illustrated in the comparison of difference scores between conditions shown in Figure 3C. No other interactions were significant, Fs < 1.80.

Figure 4.

The P300 component evoked by emotion trials. (A) ERPs are plotted by group and n-back level for each emotion condition. (B) Mean amplitudes, (C) n-back target amplitude difference (1-back – 2-back), and (D) peak latency costs of expression (expression - neutral) from the highlighted analysis window (i.e., 300 – 450 ms), collapsed over trial type, are depicted. Error bars represent standard error.

For P300 latency on emotion trials (Figure 4D), there was a significant effect of trial type, F(2, 57) = 12.84, p < .001, η_p² = .18, with longer latencies for target trials than non-target trials, and a language group by emotion interaction, F(2, 114) = 5.87, p = .004, η_p² = .09. The interaction indicated that latencies differed across emotion conditions for the monolingual group, F(2, 58) = 5.26, p = .008, but not the bilingual group, F < 2.0. Post-hoc contrasts demonstrated that monolingual latencies were longer in the angry condition than the neutral condition, F(1, 29) = 9.23, p = .005, η_p² = .24, and latencies on the happy condition were marginally longer than the neutral condition, F(1, 29) = 3.49, p = .07, η_p² = .11. No other effects or interactions were significant, Fs < 1.60.

Results of the analyses for the N170 are shown in Figure 5 for the control condition and Figure 6 for the emotion condition. The analysis of peak amplitude in the control condition revealed no significant effects or interactions, all Fs < 3.10. For emotion trials, there was a significant effect of emotion, F(2, 114) = 5.26, p = .007, η_p² = 09, in which the happy and angry trials produced larger amplitudes than neutral trials, with no difference between happy and angry trials. No other effects or interactions were significant, Fs < 3.50.

Figure 5.

The N170 component evoked by control trials. (A) ERPs are plotted by group and n-back level with (B) peak amplitudes obtained from the highlighted window (i.e., 150 – 200 ms). Error bars represent standard error.

Figure 6.

The N170 component evoked by emotion trials. (A) ERPs are plotted by group and n-back level across emotion conditions. (B) The peak amplitudes obtained from the analysis window (i.e., 150 – 200 ms). Panel C emphasizes the main effect of emotion with waveforms averaged over group, n-back level, and trial type. Error bars represent standard error.

Behavioral Partial Least Squares

A behavioral PLS analysis was performed on waveforms from the 1-back and 2-back, collapsed across emotion, to examine how behavioral measures covaried spatially and temporally with electrophysiology for each group. Three significant LVs were revealed that depicted how patterns of neural activity supported behavior for each language group. The first LV (Figure 7; p < .001, 39.10% crossblock covariance explained) produced a scalp salience map (Figure 7A), with positive saliences occupying frontal electrodes from Fz to Fpz and were most robust for a negative deflection during a mid to late time window in the ERP. Negative saliences were associated with a wide region covering centroparietal sites. This area corresponded with the P300 component, including electrodes Cz to POz along the posterior-anterior axis and CP5 to CP6 along the lateral axis, with the most reliable saliences representing a positive amplitude ranging from roughly 200 to 500 ms. Average scalp score correlations by group (Figure 7B) indicated that the negative, centroparietal pattern of negative saliences was associated with faster response times in both groups across all conditions. The positive pattern, however, uniquely supported bilingual accuracy on non-target trials. No other correlations were significant.

Figure 7.

LV1 of the behavioral PLS. (A) Whole-brain ERPs averaged over emotion conditions and with reliable saliences annotated to highlight patterns of activity associated with (B) behavioral correlations. Error bars represent 95% bootstrapped confidence intervals, and differences may be inferred where bars do not overlap.

The second LV (Figure 8; P < .001, 14.51% crossblock covariance explained) revealed a set of regions and components that consistently supported monolingual behavior on the task. The positive saliences represented an early frontal deflection that roughly corresponded to N200 component, and this was associated with faster reaction times for monolinguals across the entire task. For bilinguals, this activity was only correlated reliably with 2-back accuracy on non-target trials. Negative saliences covered a centroparietal region and occurred late in the time window, roughly 600–900 ms and a posterior, negative deflection occurring within the same time-window as the N170 followed by a P200. This was correlated with monolingual accuracy in every condition, but only correlated with bilingual accuracy on 2-back target trials.

Figure 8.

LV2 of the behavioral PLS. (A) Whole-brain ERPs averaged over emotion conditions and with reliable saliences annotated to highlight patterns of activity associated with (B) behavioral correlations. Error bars represent 95% bootstrapped confidence intervals, and differences may be inferred where bars do not overlap.

The final significant LV (Figure 9; P < .001, 9.92% crossblock covariance explained) revealed an interaction whereby a positive set of saliences occupying frontal and right temporal electrodes were associated with faster responding for monolinguals on the 1-back, but were supportive of more accurate responding on target trials across n-back for bilinguals. The negative saliences highlighted a similar LPC-like and N200 component to LV2 and were associated with the opposite trend.

Figure 9.

LV3 of the behavioral PLS. (A) Whole-brain ERPs averaged over emotion conditions and with reliable saliences annotated to highlight patterns of activity associated with (B) behavioral correlations. Error bars represent 95% bootstrapped confidence intervals, and differences may be inferred where bars do not overlap.

Discussion

Bilingual and monolingual young adults were examined for their performance on an EC task using working memory and emotion regulation demands, two processes known to draw upon EC, while EEG was recorded. For reaction time, responses were faster on the 1-back than the 2-back and faster on target than non-target trials in both the control and emotion conditions. However, on the emotion condition, an interaction revealed that bilinguals showed a greater difference in RTs between n-back levels than monolinguals. Regarding accuracy, both groups were more accurate on the 1-back than 2-back and on non-target than target trials, for both the control and emotion conditions. In the emotion condition, a marginal interaction indicated the language groups achieved similar scores on the 1-back, but bilinguals were more accurate than monolinguals on the more effortful 2-back. Although this effect failed to reach the conventional level of significance (p = .06), this may simply reflect a power issue in detecting the effect, and it is possible that with a larger sample this effect will become reliable. Finally, an emotion by language group interaction identified group differences for angry trials and a trend for a difference on happy trials, with no group difference on neutral trials, indicating larger differences for the most salient emotional stimuli.

Similar patterns were found for the EEG results. There were no group differences on the 1-back, but the groups diverged on target trials when performing the more effortful 2-back task. P300 amplitude was reduced for both groups, an expected finding based on previous literature using the n-back task, but there was greater attenuation by the bilinguals during the emotion condition.

Previous studies investigating the effect of bilingualism on P300 in cognitive control tasks have reported greater P300 amplitude for bilinguals than monolinguals. However, Kousaie and Phillips (2012) found that bilinguals produced smaller P300s than monolinguals on a Simon task which measures non-verbal response conflict. This suggests that, at least in some cases, bilinguals do not generate larger P300 amplitude during executive control tasks.

The reduction of P300 amplitude on the 2-back reflects increased WM load and is thought to originate from the reallocation of resources to process these new demands (Watter, Geffen, & Geffen, 2001). Importantly, the greater attenuation of the P300 for the bilingual group was accompanied by higher accuracy and longer RT than monolinguals. Such a brain-behavior relationship appeared in LV1 of the behavioral PLS, whereby larger P300 amplitude supported faster responding across the task but greater loading on to the P300 pattern was associated with lower accuracies, specifically for non-target trials. This pattern indicates that when the task required the greatest EC, bilinguals more readily engaged resources elsewhere in the brain necessary to meet increasing WM demands than monolinguals.

These results suggest that bilinguals approached the task with a greater emphasis on controlled processing than monolinguals as EC demands increased. The emotion condition scaled EC demands through the need to overcome emotional distraction. Bilinguals slowed more than monolinguals on the 2-back, and this may reflect the use of a strategy that emphasized more decision-making prior to initiating a response when WM demands were high, an approach that is consistent with expertise (Incera & McLennan, 2016). On this condition, however, bilinguals were more accurate than monolinguals in response to the challenging 2-back. Therefore, when EC demands were highest, bilinguals adopted a controlled processing approach that was associated with accurate responding whereas monolinguals used a strategy that more closely resembled how they performed the easier conditions. This strategy allowed RT to remain fast, but at the cost of reduced accuracy for monolinguals.

Although a main effect of emotion in P300 amplitudes was not predicted, earlier studies have found that P300 positivity can be amplified by visual emotional stimuli (Johnston, Miller, & Burleson, 1986; Mini, Palomba, Angrilli, & Bravi, 1996). This finding has since been replicated using emotional face stimuli, with angry faces eliciting the largest P300, followed by happy faces, then neutral faces (Tortosa, Lupiáñez, & Ruz, 2013). Similarly, we found that angry faces were associated with larger P300 amplitudes, although there was no difference between neutral and happy faces. This may be due to differences in the stimuli used between both studies. The P300 is well-known to increase in response to deviant stimuli in oddball paradigms (Squires, Squires, & Hillyard, 1975), and the emotion effect may reflect a similar response to stimulus saliency.

The relationship between P300 and emotion raises an additional interpretation for the greater attenuation of P300 amplitude found on the 2-back for bilinguals. It is possible that amplification of the P300 response acted in contradiction to the typical reduction associated with increasing WM load in the monolingual group. This is supported by the lack of a language group interaction in the control condition: only when emotion regulation was present were bilinguals found to attenuate the P300 to a greater extent than monolinguals. Other work has investigated the impact of complex face stimuli on P300 morphology and found that typical cognitive task effects associated with this component are obscured when face stimuli are employed (Zhang, Li, Qian, Zhou, 2012). Therefore, the smaller reduction in amplitude found for monolinguals may reflect greater processing of the distractor stimuli when WM load increased. This interpretation is consistent with bilinguals using controlled processing on the task, and this strategy would be particularly effective given the emotion regulation demands of the task. Namely, controlled processing may have protected bilinguals from distraction and facilitated accurate WM performance.

Consistent with our prediction, monolingual P300 latencies were longer than those of bilinguals, but this effect interacted with emotion: P300 latency increased from neutral to happy, and marginally to angry emotions for monolinguals but not bilinguals. Additionally, bilinguals were more accurate on happy and angry trials than monolinguals. This indicates that the emotional distractors uniquely affected monolingual response accuracy and processing speed regardless of WM load.

N170 peak amplitude did not differ between groups but increased for happy and angry expressions relative to neutral expressions. Bilinguals were predicted to show less of an N170 response, but this finding indicates that both groups processed the face stimuli with the same amount of neural resources. It has been suggested that bilingual face processing ability differs from that of monolinguals (Kandel et al., 2016), but no support for this hypothesis was found in the present sample. Behavioral performance was largely similar when using control stimuli that presented no face processing demands. The lack of N170 group differences indicates that faces were processed in the brain similarly between groups.

The behavioral PLS analysis yielded three significant patterns of brain-behavior relationships. LV1 depicted a set of components that were universally beneficial to task performance. The P300 component was associated with faster responding for both groups throughout the task. However, bilingual accuracy on non-target trials was associated with modulation of a mid-frontal negativity. This component has been characterized as indexing familiarity in recognition paradigms (Curran & Cleary, 2003; Woodruff, Hayama, & Rugg, 2006) and may reflect a relationship between recollective processes, above familiarity, aiding this group in accurately responding. Since n-back stimuli were constrained to eight letters and repeated frequently, reliance on familiarity in response judgements may increase susceptibility to false alarms in this task for non-target trials. This would suggest that bilinguals were able to leverage more explicit, controlled processes when probed for recognition to positively influence accuracy.

LV2 highlighted regions that consistently covaried with monolingual behavior across the task. A late-positive complex (LPC) and posterior P200 component were associated with accurate responding for the monolingual group. This relationship is understandable as the LPC has been related to working memory processing (Addante, Ranganath, & Yonelinas, 2013; Luu et al., 2014) and the P200 has been noted to reflect encoding (Finnigan, O’Connell, Cummins, Broughton, & Robertson, 2011). Monolingual reaction times, however, were linked to an anterior N200 component which is linked to inhibitory control (for review see, Folstein & van Petten, 2008) and may reflect inhibition processes overriding the prepotent tendency to attend to the salient emotional distractors in the current task. In contrast, only bilingual 2-back performance correlated with the latent variable. Target accuracy for this group was supported by the same LPC/P200 regions, however non-target accuracy was associated with the N200 component.

LV3 accounted for the least covariance and produced the least clear pattern of relationships. However, the LPC and N200 were again identified and correlated with bilingual accuracy on target trials, whereas a frontal and right-temporal pattern of mid-late activity was associated with monolingual reaction times on the 1-back.

Overall, these results align with the only other study that has investigated WM and emotion regulation using an n-back task (Janus & Bialystok, 2018). Studying children, the authors similarly found that bilinguals were slower but more accurate than monolinguals on the 2-back condition. This finding was interpreted as evidence that bilinguals more easily adapted to new demands and switched their task strategy than monolinguals. In other words, bilinguals approached the task with more strategic flexibility than monolinguals.

The present study extends this line of work by revealing how this behavioral pattern was supported in the brain of young adults. P300 results indicate that bilinguals more readily processed increasing WM demands and were less distracted by emotional stimuli. The brain-behavior relationships produced by the PLS analysis revealed that both groups recruited the same neural resources, however behavior was differentially affected. Monolingual accuracy was associated with electrophysiology in only one latent variable, but correlations were detected for bilingual accuracy in every latent variable. In contrast, monolingual RTs were repeatedly found to correlate with neural activity where bilingual RTs did not. This suggests that the behavioral differences found between monolinguals and bilinguals, when posed with concurrent WM and emotion regulation demands, indicate that the neural patterns present during the task were leveraged to yield different behavioral outcomes between groups.

This study investigated the relationship between executive control and non-verbal emotion regulation in bilingualism. The present findings suggest that bilinguals adopt a different neural processing strategy than monolinguals when confronted with increasing EC demands, although the consequence of that strategy were sometimes beneficial (accuracy) and sometimes not (RT). Bilinguals more readily processed novel task demands and produced slower but more accurate responses as WM load increased, mirroring expert behavior. In the context of nonverbal emotion regulation demands, this approach was associated with less impact of emotion stimuli on bilingual neural electrophysiology and behavioral performance.

The results contribute to our understanding of how bilingualism impacts executive control in WM and emotion regulation given inconsistent outcomes of previous research using behavioral measures. As in previous studies, behavioral differences between monolinguals and bilinguals are small and appear most reliably in conditions that have higher processing demands. Thus, behavioral differences in the present study were found in 2-back but not in 1-back conditions. More importantly, however, the inclusion of EEG recording showed that even when behavioral outcomes are similar, monolinguals and bilinguals are recruiting neural resources differently to complete the task successfully. In the present study, bilinguals showed better accommodation to changing task demands than did monolinguals. Although the effect of this accommodation on behavior was small, this greater neural flexibility may be one of the reasons that bilinguals demonstrate greater cognitive reserve in older age, including greater resilience with cognitive decline and dementia (review in Bialystok, 2017). This neural flexibility, therefore, may be the most important outcome of bilingualism.

Highlights.

• Monolingual and bilingual young adults performed an emotional n-back task in 1-back and 2-back conditions

• Bilinguals were slower and more accurate than monolinguals in the difficult 2-back condition

• EEG recordings showed that bilinguals adjusted their processing strategies to meet the demands of the difficult condition but monolingual did not

• Presence of emotional stimuli impacted performance of monolinguals but had no effect on bilinguals