39 Neural Oscillations as Predictors of Variability in Second Language Proficiency Victoria Ogunniyi1*, David Abugaber2, Irene Finestrat2, Alicia Luque4, Kara Morgan-Short2,3 1Department of Biological Sciences, University of Illinois at Chicago, Chicago, IL ²Department of Hispanic Linguistics, University of Illinois at Chicago, Chicago, IL ³Department of Psychology, University of Illinois at Chicago, Chicago, IL 4Department of Language and Culture, UiT The Arctic University of Norway, Tromsø (Norway) ABSTRACT: Understanding what traits facilitate second language (L2) learning has been the focus of many psycholinguistic studies for the last thirty years. One source of insight comes from quantitative electroencephalography (qEEG), i.e., electrical brain activity recorded from the scalp. Using qEEG, [1] found that functional brain connectivity is predictive of language learning ability. This study extends Prat et al. in investigating the association of qEEG measures for two mea- sures of L2 proficiency, namely: 1. a grammaticality judgment task, wherein participants read and identified Spanish sentences as either correct or incorrect based on possible grammar violations, and 2. a standardized Spanish proficiency test (DELE). Participants were low-intermediate L2 learners recruited from third- and fourth-semester university Spanish classes. Spectral power and coherence within and across six different regions were analyzed for correlations with either scores on the grammaticality judgment task or on the DELE. Follow-up linear regression models based on significant qEEG correlates explained up to 11% of variance in DELE scores but none of the variance in grammaticality judgment task performance. Negative correlations were found between theta frequency coherence and the DELE. Because theta activity has been associated with epi- sodic and working memory performance, these findings suggest that less proficient learners might utilize memory-based strategies more often to compensate for their lack of familiarity with the L2. KEYWORDS: Second language acquisition, language proficiency, quantitative EEG, psycholinguistics, resting-state studies INTRODUCTION Understanding what characteristics underlie successful learning is not only pedagogical- ly important but also intriguing from a cogni- tive standpoint. In the study of linguistics, this question has frequently been examined in re- gard to second language (L2) acquisition (for a review, see [2]). Researchers have used a variety of approaches to investigate which cog- nitive abilities are correlated with higher L2 pro- ficiency. Going beyond behavioral measures for psycholinguistic and cognitive constructs © 2021 Ogunniyi, Abugaber, Finestrat, Luque, Morgan-Short. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits the user to copy, distribute, and transmit the work provided that the original authors and source are credited. Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 40 [2], neuroimaging techniques such as electro- encephalography (EEG) allow us to directly quantify and connect brain activity to the pro- cesses involved in language processing [3]. This study addresses the question of whether EEG can explain variance on out- comes of L2 proficiency. We first review the literature on qEEG and its use as a neurocog- nitive measure in L2 studies, focusing on the studies that examine resting-state qEEG as a potential factor in language learning ability and language proficiency. Then, in a conceptual replication and extension of [1], we describe the results of our study in which EEG measures of resting-state brain rhythms were explored for relationships with behavioral measures of L2 proficiency. The potential neurocognitive and pedagogical implications of these findings are expounded on in the discussion section. Through our conceptual replication of [1], which focused on measures of L2 learning, we aimed to examine whether prior research on resting-state qEEG and L2 learning rate would extend to the construct of L2 proficiency, that is, whether a learner’s intrinsic pattern of brain rhythms is associated with their L2 abilities. Background on qEEG EEG is an electrophysiological technique that utilizes electrodes placed on the scalp to mea- sure changes in voltage between electrodes [4, 5]. These transient shifts in electric po- tential are caused by the electrical activity of neurons. Due to their proximity to the scalp, pyramidal neurons, which project information to neurons in local regions, produce most of the postsynaptic potentials recorded by EEG [5]. Although EEG data cannot attribute the electrical activity to specific brain regions, its temporal resolution allows researchers to track changes in brain activity to the millisec- ond. Thus, since the 1960s, EEG has been a widely used tool in cognitive studies [6]. There are various methods that can be used to analyze EEG data. The analysis of raw EEG, or qEEG, data yields useful information via neural oscillations, that is, rhythmic or re- petitive patterns of neural activity. In contrast to more common methods that analyze voltage amplitudes within given time windows tied to a stimulus, such as event related potentials, qEEG has the advantage of providing infor- mation about neural activity occurring before and after the onset of a stimulus, or even in the absence of any particular stimulus. In pri- or literature, qEEG has been used in dispa- rate domains, such as serving as evidence of mental dysfunction in criminal cases [7], pro- viding neurofeedback for therapy patients with ADHD [8], and predicting learners’ aptitude for learning computer programming languages [9]. Neural oscillations can be quantified through three measures: synchrony, amplitude, and coherence. The first measure, synchro- ny, describes whether neural oscillations are increasing or diminishing during a cognitive process [6]. By measuring the phase synchro- nization and desynchronization of the neural oscillations in this way, researchers can map larger interactions among the brain’s networks and demonstrate patterns of activation. This may provide insight into a possible solution to the binding problem, which asks how the brain integrates separate streams of information into one cohesive mental representation [5]. The second measure, amplitude, describes local changes in synchrony. Amplitude within a cer- tain frequency band is also often referred to as power. Though an increase in power does not always reflect the presence of oscillations, sustained power increases within a narrow frequency range is usually a good indicator that oscillations are likely present at that fre- quency [6]. The third measure, coherence, describes the similarity in waveform proper- ties and the stability of phase differences be- tween two oscillations across brain regions [6]. Though these three measures do not en- compass all possible properties of oscillations, they can describe how oscillations represent activation and suppression of different neu- ral networks, how wave amplitude is related Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 41 to increased general activation, and how os- cillations communicate over long distances. Though their specific ranges vary be- tween studies, there are five primary frequency bands that neural oscillations can be divided into [5, 6, 10]: delta (1-2 Hz), theta (3-7 Hz), alpha (8-12 Hz), beta (13-30 Hz), and gamma (30-200 Hz). These bands have been implicated in a va- riety of cognitive processes. For instance, prior research has suggested a relationship between oscillation frequency and the range of neural network interactions: the lower frequencies, alpha and below, represent local interactions, whereas higher frequencies represent interac- tions between more distant brain regions [10]. In regard to the more general applica- tions of qEEG, all of the frequency bands have been shown to play a role in memory. For in- stance, the delta band, which is the predomi- nant frequency found in slow wave sleep, has been associated with memory consolidation [5]. In particular, delta oscillations facilitate the for- mation of declarative memory, the memory of one's experiences and explicit knowledge. Sim- ilarly, theta oscillations have been implicated in both working memory and long-term memory retrieval [5, 11]. Theta oscillations have primar- ily been observed in the cortex, further echoing these memory functions [10, 11]. Alpha oscilla- tions, which are the most prominent in the adult brain, are related to attention paid to external stimuli [5, 6]. More importantly, alpha plays a role in blocking irrelevant information in work- ing memory. Additionally, alpha desynchro- nization and reductions in alpha power result in more successful information encoding [10]. Recent studies have shown that the beta band, which is mostly generated in the fronto-central region of the brain, also plays a role in regulat- ing information stored in working memory [12]. The gamma band, which has been observed in the cortex [5], has been associated with short- and long-term memory maintenance [6]. Additionally, increases in gamma activity is anti-correlated with beta activity levels [12]. Additionally, the bands have been shown to play significant roles in stimulus-based lan- guage tasks. For instance, increased power in the theta band, which occurs during grammat- ical violations and sentence contexts that are difficult to interpret, reflects its involvement in lexical-semantic memory retrieval. Similarly, the alpha band helps organize information stored in short-term memory during sentence compre- hension [10]. Regarding the higher frequency bands, gamma and beta have been associated with unifying related word meanings and similar grammatical forms, respectively. More specifi- cally, the gamma band has been attributed to semantic unification [6, 10], which is supported by the observed decreases in gamma power in response to phrases with unclear meanings and idiomatic expressions [10]. Conversely, the beta band is involved in syntactic unification [5, 6, 10]. Beta oscillations sentence to lower processing regions [10]. By reflecting both do- main-general and stimulus-specific cognitive functions, qEEG has proven to be a useful neu- rocognitive measure in psycholinguistic studies. The Use of qEEG in L2 Studies Several L2 studies have tested the relationship between qEEG measures and L2 constructs. These studies have generally addressed two issues: L2 proficiency [13-16], which describes a learner’s language abilities at a given point in time, and L2 grammatical learning [17-19], which describes how learners better under- stand the rules of a language with increasing proficiency. Regarding L2 proficiency studies, results have shown that highly proficient L2 learners differ in qEEG measures from less proficient L2 learners [14-16]. When compar- ing differently proficient participant groups, significant differences are found in the timing and location of oscillatory activity, especially regarding lateralization of the location of the oscillations between the right and left hemi- sphere. Though the delta band is not frequently analyzed in these studies, one study found no significant group differences [13]. For gram- matical learning studies, in both natural and ar- Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 42 tificial grammar learning tasks, higher relative power and coherence in the higher frequency bands ( >8 Hz) were associated with higher proficiency and increased in prevalence over time, whereas higher relative power and co- herence in the lower frequency bands were as- sociated with lower proficiency and decreased over time [16, 17, 19]. Gauging learning by assessing oscillatory responses to grammat- ical violations, other research has also found that the higher frequency bands were elicited in both semantic and syntactic conditions [17]. A relatively new approach in L2 stud- ies is to examine whether resting-state qEEG measures (i.e., that are taken when the brain is “idling” in the absence of any explicit task, as opposed to stimulus-related qEEG measures) are associated with the rate at which an individ- ual acquires second language abilities, or L2 learning, and proficiency. As of now, very few studies have tested the correlation between neural activity occurring in the absence of a stimulus, or resting-state qEEG, and L2 learn- ing rates [1, 20]. In these studies, native-En- glish speakers learned a second language over the span of a few months. Prior to this learning period, resting-state EEG was performed, and various behavioral measures were collected as additional outcome measures. In the first of these studies, [20] found that, when entered as predictors into a regression model, rest- ing-state qEEG measures explained up to 60% of the variance observed in L2 learning rates, meaning that learning rate accounted for more than half of the variability in the data. Though the most predictive frequency range was found to be the low-beta range (13-14.5 Hz), power in the beta and gamma frequency bands record- ed primarily over right hemisphere electrode regions were found to be the strongest predic- tors of L2 learning ability in general. The au- thors also found that alpha power over frontal and temporal electrodes and low-beta power over temporal regions were indicative of better language learning ability. As with other studies, greater activation in left hemisphere electrode sites was associated with lower L2 proficiency. In a later study conducted by [1] with a higher sample size, the authors tested whether resting-state qEEG measures were significant predictors of different L2 learning measures. Similar to their earlier study, the results impli- cated the qEEG activity in the right hemisphere with greater L2 learning ability. Simultaneous regression analyses were run on three out- come variables: L2 learning rate, total speech attempts, and performance on a declarative memory posttest. Mean right posterior beta power was found to be a significant predictor of L2 learning rate and total speech attempts. Frontotemporal to posterior coherence in the right hemisphere was found to be a signifi- cant predictor of performance on the declar- ative memory posttest, whereas mean within left posterior coherence across all frequencies was a significant predictor of total speech at- tempts. Altogether, these results indicate that mean beta power over posterior electrode regions plays a significant role in L2 ability. In all, due to their potential advantag- es over stimulus-locked measures, qEEG measures have been increasingly used in L2 studies. Though some frequency bands have been more frequently studied than others, all five of the classic bands have been implicat- ed in various language functions in some way. Across various research designs, qEEG has been used to illustrate large-scale patterns of brain activation during language processing. However, relatively little research has explored the potential of resting-state paradigms for qEEG in L2 psycholinguistics. For example, al- though [1, 20] have examined whether qEEG can predict individual differences in L2 learning rate, speech attempts during learning, and L2 declarative knowledge, no study has yet exam- ined whether or how resting-state qEEG mea- sures may be predictive of L2 proficiency, which is the desired final outcome of L2 learning. Purpose of Research The goal of this study is to investigate the Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 43 potential association between resting-state qEEG and L2 proficiency. The current study ex- pands on the research design of previous rest- ing-state L2 studies. At present, resting-state qEEG studies have involved extensive training sessions with participants in the initial stag- es of L2 learning. However, the participant population utilized in this study comprised L2 learners at a low-intermediate stage who were recruited from third- and fourth-semester uni- versity-level Spanish courses. In connecting the qEEG measures with observed variabili- ty in language proficiency, we reasoned that second-year Spanish learners with the qEEG profiles most conducive to effective processing (e.g., through memory functions and other re- lated cognitive processes) would have gained the most Spanish proficiency from their class- es. Additionally, this study has the advantage of analyzing a greater number of electrodes than prior research [1, 20], providing higher spatial resolution for measuring electrical ac- tivity on the scalp. As with prior resting-state studies, a variety of qEEG measures were ana- lyzed, including spectral power and coherence. Considering the issues above, this study spe- cifically investigated two research questions: Research Question 1: Is mean spectral power calculated from resting-state qEEG data asso- ciated with L2 proficiency, as assessed by two Spanish proficiency tasks? Research Question 2: Is mean within- and between-network coherence calculated from resting-state qEEG data associated with L2 proficiency, as assessed by two Spanish profi- ciency tasks? Following the results of [1], we predicted that the qEEG measures that were most likely to show a relationship with the two L2 proficiency tasks were mean beta power and frontotempo- ral-to-posterior coherence. In spite of these ten- tative predictions, we sought to replicate [1]’s methods as closely as possible in reproducing their two-step exploratory analysis strategy (i.e., pairwise correlations followed by multiple regression using significant correlates as pre- dictors) on all frequency bands. As we intended our analysis itself to be strictly exploratory rath- er than confirmatory, we included all frequency bands in the analysis. In order to assess L2 pro- ficiency, we decided to administer two assess- ments, one that reflects specific grammatical knowledge acquired through the learners’ class (the grammaticality judgment task), and one that is a more general and widely measure in the L2 literature (Diplomas de Español como Len- gua Extranjera), see Methods. Given that prior research [21] has found a distinction between automatic and controlled language processing, using both a timed and untimed L2 proficiency measure would allow us to examine the appli- cation of L2 knowledge in two distinct ways. METHODS Participants Forty-nine participants (29 female; 20 male; mean age = 21.54; age SD = 4.14; range = 18- 38) were initially recruited for participation in a two-part EEG study in which they were tested in English and in Spanish in separate testing sessions. Participants were recruited at a large, public urban university in Chicago from third- and fourth-semester Spanish language cours- es, which centered on developing communica- tive abilities and aimed to help students obtain low-intermediate proficiency by the end of the fourth-semester course. Thirteen of these par- ticipants were recruited from the third-semes- ter Spanish course, while twenty-two of these participants were recruited from the fourth-se- mester Spanish course; thirteen participants did not report their course level. The recruit- ment process involved advertising to these language courses, after which participants self-selected whether to participate. Regarding the racial distribution of the participant popula- tion, nineteen participants identified as White/ Caucasian, fourteen participants identified as Asian, six participants identified as Black/Af- Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 44 rican-American, four participants identified as multiracial, and six participants’ races were un- reported. All the participants reported having English as a native language (even if they were simultaneous bilinguals with early exposure to other languages) and having no Spanish ex- posure growing up. Additionally, per our EEG criteria, all participants were right-handed as assessed by the abridged version of the Edin- burgh Handedness Inventory [22], with normal or normal-to-corrected vision. This was done to ensure that our results were not confounded by uncorrected vision or differing brain activity in those who are left-handed. None of the partici- pants reported having psychiatric, neurological, or learning disorders. All participants gave in- formed consent according to the standards of the University of Illinois at Chicago institutional review board and were financially compensat- ed for their participation based on the number of hours spent participating in the study, re- ceiving $5 for every hour spent on completing the proficiency measures, $15 for every hour spent during the EEG recording process, and a $45 bonus for completing both sessions. Participants were only included in the analysis if they completed the resting state EEG recording along with at least one of our Spanish proficiency measures. The final num- ber of participants included in the current anal- ysis is 47, with 44 of these participants having completed both Spanish proficiency measures (see Table I below). A similar EEG recording procedure was performed for an L1 English reading task (not reported here). The partici- pant data included in this study was collected over the span of two years, from 2018 to 2020. Grammaticality Judgment Task Participants read various Spanish sentences and were asked to determine whether they fol- lowed Spanish grammatical rules. The gram- maticality judgment task consisted of three experimental conditions: phrase structure, sub- ject-verb agreement, and noun-phrase viola- tions (see Table II for examples). The phrase structure condition, wherein word order viola- tions were introduced into a sentence by pre- senting a noun instead of a verb or vice-versa, consisted of 60 correct and 60 violation sen- Table I. Participant Language Characteristics M (SD) [Range] Number of native lan- guages 1.29 (0.45) [1-2] Number of L2s 1.44 (0.53) [1-3] Age of acquisition Spanish (years) 14.53 (4.88) [0-29] Self-rated Spanish listening proficiencya 4.88 (2.01) [1-9] Self-rated Spanish reading proficiency 5.42 (1.72) [1-8] Note: aSelf-rated proficiency on 0 (‘none) to 10 (‘perfect) scale. Table II. Examples of stimulus sentences Item Type Example Phrase Structure Ella tiene mucho dinero/*gastar que gastar/dinero en ropa. [She has a lot of money/*spend to spend/money on clothes.] Subject-Verb La mujer dibuja/*- dibujan en su ha- bitación. [The lady draws/*draw in her bedroom.] Noun Phrase El hombre prepara estas papas/ *papa para su esposa. [The man prepares these pota-toes/*po- tato for his wife.] Note: * = violation word. Italics indicate the critical correct/violation word in each sen- tence. Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 45 tence frames, totaling 120 sentences over- all. The subject-verb agreement condition, for which the verb ending did not agree with the plurality of the subject, consisted of 60 correct sentences with a singular subject, 60 correct sentences with a plural subject, and 120 vio- lation sentences, totaling 240 sentences. The noun-phrase condition, for which either the singular/plural status or the grammatical gen- der of an article (e.g., los, “the [MASC. PLU- RAL]”; esta, “this [FEM. SINGULAR]”) did not match the noun, consisted of 124 number vi- olation frames, 124 gender violation frames, and 248 correct sentence frames, for a total of 496 sentences. In total, 856 sentences were used across all three conditions. The sen- tences ranged from 5 to 12 words in length. None of the sentences contained violations in initial or final sentence positions, so as to avoid sentence “start-up” and “wrap-up” ef- fects in the EEG [23-24], and none of the crit- ical words were repeated between frames. To ensure that the participants would be familiar with the vocabulary contained in the sentence frames, all the words for this task were taken from a Spanish textbook used at the universi- ty at the time that data collection began [25]. Sentences were presented one word at a time on the computer screen (see Figure 1) using E-Prime 2.0 software. Instructions for the task were read orally to the participants by the experimenter. Preceding each sentence, there was a screen that read, “Rest your eyes.” After three seconds, the sentence was then visually presented one word at a time. Each word was displayed in the center of the screen for 350 ms, with a 150 ms interval of blank screen before the onset of the subsequent word. Once the entire sentence was presented, a screen followed that said, “Good/Bad?” In response to this prompt, participants pressed a keyboard button to cat- egorize the sentence as either grammatical or ungrammatical. Participants first completed a short practice block containing 8 sentences. The stimuli sentences were then presented over 4 experimental blocks. There were three 3-minute breaks during the experiment, one at the end of each block. Another EEG recording was performed over the duration of this task but was not analyzed in the current study. Par- ticipants’ responses were used to calculate a d-prime (dʹ) score, which is a metric for signal detection that accounts for response bias by comparing how often a participant correctly identifies a signal to their false-alarm rate [26]. Figure 1: Diagram of a typical trial in the gram- maticality judgment task. Diplomas de Español como Lengua Ex- tranjera (Diploma of Spanish as a Foreign Language): A modified version of the Diplomas de Español como Lengua Extranjera (DELE, [27]) was used to assess Spanish proficiency. The three-part test was completed on a computer in the lab- oratory through a Qualtrics survey form. Par- ticipants were asked to read through the DELE questions and answer them at their own pace. In the first section, participants were required to read through a passage in Spanish and answer 20 fill-in-the-blank questions. Each question had 3 possible answer choices. In the second sec- tion, participants were given 10 sentences and asked to choose the answer choice that best defines the bolded word in the sentence. Each sentence also had 3 possible answer choices. The third section consisted of 19 grammatical questions . Participants were asked to select the answer choice that fit best in the context of each of the sentences. Eight of these questions had 2 possible answer choices, and the remain- ing questions had 4 possible answer choices. In total, participants answered 49 questions. Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 46 Procedures Prior to testing, all participants completed pre-testing questionnaires that verified their el- igibility and provided more detailed information about their language history. This included a language background questionnaire, a test-ses- sion questionnaire, and a handedness question- naire. The language background questionnaire assessed each participant’s demographic back- ground and language history and experience (LEAP-Q, [28]). The test-session questionnaire assessed how much sleep a participant had and whether they had taken any psychoactive substances that may affect their ability to per- form the task. The handedness questionnaire was used to gauge left-/right-handedness by assessing hand preference during various ac- tivities, following the standard Edinburgh Hand- edness Inventory [22]. The items were read to participants, who provided their answers verbally to the experimenter. Answers were recorded in computer-based survey forms. In replicating [1]’s procedure, we collect- ed five minutes of eyes-closed resting-state EEG following completion of the pre-testing surveys. Participants sat in a chair inside of a sound-at- tenuating booth. After fitting the participants for an EEG cap and placing eye electrodes, an electrolyte solution was applied to the scalp electrodes to minimize electrical impedances. Participants were then instructed to close their eyes and remain still and awake during the re- cording. While recording, the lights were turned off and the door of the sound booth was closed. The EEG data was recorded using asa™ software with an ANT Neuro wave- guard™ elastic cap with 32 Ag/AgCl electrodes distributed in standard and extended 10-20 system locations. Scalp impedances were low- ered to 10 kΩ or below. Scalp electrodes were referenced to the common average of all the electrodes. To detect artifacts caused by eye movements, electrodes were placed above and below the right eye and on the left and right outer canthi to record a vertical electrooc- ulogram and a horizontal electrooculogram, respectively. Using an ANT Neuro bioamplifier system (AMP-TRF40AB Refa-8 amplifier), the EEG signal was amplified to 22 bits. The signal was also recorded in DC mode, digitized with a 512 Hz sampling rate, and filtered online us- ing a low-pass filter with a cutoff of 138.24 Hz. Following the resting-state EEG ses- sion, the grammaticality judgment task was im- plemented in the sound booth (see above), and after disassembly of EEG equipment and a short break the DELE task was performed on a com- puter outside of the sound booth (see above). Analyses The qEEG data was pre-processed using the EE-GLab toolbox [29] for MATLAB [30]. To ensure that the resting-state recording was exactly five minutes, each recording was lim- ited to 300 seconds. Seven participants had recordings that were slightly less than 300 seconds (minimum = 283 seconds) but were still included in the analysis. Each participant’s recording was divided into epochs of two-sec- ond duration, with 50% overlap across epochs. These epochs were then cleaned for artifacts (e.g., from muscle movements, eyeblinks, faulty electrodes, etc.) using the pop_autorej() function from EEGLab. Participant datasets with less than 75 seconds of epoch-free re- cording were omitted from the final analysis, which resulted in the loss of 6 participants (12% of the data). The mean number of sam- ples per participant was 144.44 (S.D. = 39.73). The pre-processed data were subse- quently analyzed using a modified version of the script used in the Prat et al. study [1] (avail- able at: https://github.com/UWCCDL/QEEG) for the R scripting language [31]. Six electrode networks were defined (Figure 2): medial fron- tal (electrodes FP1, FPz, FP2, and Fz), left hemisphere fronto-temporal (electrodes F7, FC5, T7, C3), right hemisphere frontotemporal (F8, FC6, T8), left hemisphere posterior (CP5, CP1, P7, P3, O1), right hemisphere posterior (Cz, CP6, CP2, C4, P4, Pz), and right hemi- sphere posterior occipital (Oz, O2, POz, P8). Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 47 In defining the electrode clusters, we aimed to replicate Prat et al. (2019), in which qEEG channels were collapsed into networks based on phase synchrony results from an earlier L2 qEEG study. Prat et al. 2019’s network defini- tions are technically slightly different from ours in that they used 14 electrodes rather than a 32-electrode cap, but our network definitions were aligned as closely with theirs as possible based on visual inspection of scalp maps (and in fact having a higher spatial resolution is a point in our favor in a sense). As data-driven results in favor of our network definitions (which we left out due to space limits), we replicated [1] (2019, Table I) in that independent samples t tests found that all within-network qEEG coher- ence values in our networks were significantly greater than all between-network coherence values, with all independent samples t-tests at p < .001. We then extracted the qEEG measures of interest, which were power and within- and between-network coherence for each of the fre- quency bands: theta (4-7.5 Hz), alpha (8-12.5 Hz), beta (13-29.5 Hz), and gamma (30-40 Hz). Figure 2: Diagram of the six network regions analyzed Finally, to address the research questions, we first conducted correlations between mean power and performance on the grammaticality judgment task and the DELE, followed by cor- relations between mean coherence and perfor- mance on the grammaticality judgment task and the DELE. Here we report statistically significant correlations. These exploratory correlations were not corrected for multiple comparisons following the main analyses reported by [1]. For power and coherence measures that showed statistically significant correlations for either pro- ficiency variable, we then entered them as pre- dictors into two, separate regression analyses. RESULTS Individual Differences in Indicators of L2 Proficiency Before examining the qEEG measures, de- scriptive statistics were examined for the two outcome measures of Spanish proficiency (see Figures 3 and 4 and Table III). With respect to the DELE, the group mean was 19 out of 49, which illustrates that the participants were overall low proficiency speakers [27]. The most proficient participant scored twice as much as the least proficient participant. With respect to the grammaticality judgment task, participants were given two scores: mean accuracy and dʹ. The average accuracy on the grammaticality judgment task was 76%. The average dʹ was 0.91. A bootstrapped simulation of chance-lev- el dʹ values on 244 trials with 10,000 iterations performed using the psycho package for R [32] found a 95% confidence interval of -0.26 to 0.26. This suggests that our participants’ dʹ values were above chance at α = 0.05. DELE scores did seem to be above chance, as indi- cated by a mean accuracy of 39.5%. Consid- ering that most of the DELE test items had 2, 3, or 4 answer choices, a minimum accuracy of 25% would at least reflect chance levels on the items with the most answer choices. DELE scores and grammaticality judgment task dʹ scores were not significantly correlated with one another, r(44) = 0.27, p = 0.069. As per [21], our results suggest that the grammati- cality judgment task and DELE might capture somewhat different facets of L2 proficiency. Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 48 Table III. Performance on the two proficiency tasks Performance measure M (SD) [Range] DELE score 19.3 (3.15) [0.13-0.26] Grammaticality judgment task dʹ score 0.91 (0.74) [-0.27-2.88] Grammaticality judgment task accuracy 0.76 (0.10) [0.51-0.98] Figure 3: Participant performance on the grammaticality judgment task. The maximum possible dʹscore on the grammaticality judgment task was effectively 4.9. Figure 4: Participant performance on the Spanish proficiency test (DELE). The maxi- mum possible score on the DELE was 49. Relating Individual Differences in Resting-state qEEG Power to L2 Proficien- cy Variables In order to determine the relationship between resting-state qEEG power and performance on the two proficiency tests, we performed correlation analyses between either the gram- maticality judgment task or DELE scores (in separate analyses) and mean power across six electrode networks. The frequency bands of interest were theta (3-7 Hz), alpha (8-12 Hz), beta (13-30 Hz), and gamma (30-200 Hz). After conducting these analyses, none of the correlations were found to be significant. How- ever, two positive correlations were approach- ing significance: medial frontal alpha power and DELE scores, r(47) = 0.28, p = .052; and left hemisphere frontotemporal alpha pow- er and DELE scores, r(47) = 0.27, p = 0.057. Relating Individual Differences in Resting-state qEEG Coherence to L2 Profi- ciency Variables In order to determine the relationship between resting-state qEEG coherence and performance on the two proficiency tests, we performed cor- relation analyses between the L2 proficiency variables and mean within- and between-co- herence across six electrode networks. The frequency bands of interest were theta, alpha, beta, and gamma. For the between-coherence values, each of the networks were paired to- gether and coherence across the four frequency bands was calculated for every pair. Regarding the DELE, two significant negative correlations were found: theta coherence within the right hemisphere posterior occipital network, r(47) = -0.31, p = 0.028; theta coherence between the me-dial frontal and right hemisphere posterior occipital networks, r(47) = -0.35, p = 0.012. Re- garding the grammaticality judgment task, there were no signif-icant correlations (all p > 0.05). Simultaneous Linear Regression Analyses When the two predictors of performance on the DELE (theta coherence within the right hemi- Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 49 sphere posterior occipital network and theta coherence between medial frontal and right hemisphere posterior occipital) were entered into a simultaneous regression analysis, the overall model was found to be statistically sig- nificant, F(2, 46) = 3.97, p = 0.026, explaining up to 11% of the observed variance. However, neither theta coherence within the right hemi- sphere posterior occipital network (β = -13.91, t = -1.07, p = 0.290) nor theta coherence be- tween medial frontal and right hemisphere posterior occipital (β = -41.78, t = -1.61, p = 0.114) were found to independently predict per- formance on the DELE (see Figures 5 and 6). Figure 5: Regression line between the DELE and theta coherence within right hemisphere posterior occipital networks. Figure 6: Regression line between the DELE and theta coherence between the medial frontal and right hemisphere posterior occipital networks. DISCUSSION Our results suggest that certain resting-state qEEG measures, particularly over the theta fre- quency band, are associated with L2 proficien- cy. Regarding the first research question, none of the correlations run between resting-state mean power and L2 proficiency reached signif- icance. However, the two correlations that did approach significance were related to alpha power: the positive correlation between medi- al alpha power and the DELE, and the positive correlation between left hemisphere frontotem- poral alpha power and the DELE. Regarding the second research question, two significant negative correlations were found between rest- ing-state within- and between-network coher- ence and the DELE, one within right hemisphere posterior networks and another between medi- al frontal and right hemisphere posterior net- works. Both significant correlations were found over the theta frequency band. After performing regressions on the significant qEEG predictors for DELE performance, the model was found to explain up to 11% of the variance. None of the variance in grammaticality judgment task per- formance was explained by qEEG measures. In relating our results with those of previ- ous studies, the correlations found between the theta frequency band and L2 proficiency were anticipated, although the direction of the rela- tionship varied by study. In [1], theta coherence within frontal electrode regions was positively correlated with several outcome measures of L2 learning. In [16], highly proficient speakers experienced increased theta synchronization in right frontal regions during a grammar learn- ing task. However, in the L2 proficiency study conducted by [13], lower theta coherence in frontal and occipital electrodes was observed among highly proficient speakers. In our study, the relationship between theta coherence mea- sures and L2 proficiency was also negative. Why do we see these contradictory patterns among studies? This may be explained by the differences in language learning and lan- guage proficiency. For instance, the participant Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 50 population used for language learning studies consists of speakers who were just beginning to learn an L2, whereas proficiency studies involve participants who have already had experience learning the additional language. We understand activity in the theta fre- quency band to reflect several memory func- tions, including specifically short- and long-term memory maintenance and memory retrieval [5, 10, 33, 34]. Additionally, the theta frequency is believed to originate from the cortex [5], which plays a role in the formation of new memories. Generally, studies have found positive rela- tionships between theta activity and learning rate in earlier stages of learning, negative re- lationships between theta and L2 ability in lat- er stages [1, 16, 19]. The negative relationship found in this study may signify that more pro- ficient participants are good at applying and retrieving grammar rules and no longer need to rely on working memory, which would result in a decreased prevalence of theta oscillations. Altogether, as suggested by [16], the negative correlations found between theta coherence and DELE performance suggest that less pro- ficient L2 speakers have a greater reliance on memory-based strategies to compensate for their lack of familiarity with the language. Even though the alpha frequency band was not significant, it was approaching signif- icance, which reflects the inverse theta-alpha relationship expressed in the literature [5]. Increases in the alpha frequency have been associated with diminished attention paid to a linguistic task [10]. Interpreting the negative correlation between theta and proficiency to be the consequence of decreased working memo- ry load, a positive correlation with alpha would suggest that more proficient participants were able to pay less attention to the task and still be successful. This may signify that the more automatic a language task is to a participant, the more likely they are to be more proficient. The results of this study need to be considered in light of its limitations. One lim- itation was the proficiency measures that were employed. As mentioned in the results sec- tion, our participants did not score statistically above chance on average on the DELE. Thus, it is somewhat surprising that significant results were found for the DELE and not for the gram- maticality judgment task. Perhaps the more dif- ficult DELE test, which had been developed to test up to near-native speaker status, allowed us to detect a role related to which learners are more successful on a more challenging task. Conversely, the grammaticality judgment task was designed to reflect specific grammatical structures taught in intermediate-level Spanish courses, so we would expect participants to perform more successfully on this task overall, which they did, as evidenced by above-chance performance. More generally, because the DELE and grammaticality judgment task both reflect performance accuracy on grammatical tasks, it is important to note that they are not holistic measures. As a multidimensional con- struct, language proficiency encompasses all the skills necessary to engage with the lan- guage in a real-life context [2]. Thus, it would be beneficial for future research to include more time-pressured proficiency measures in future research, such as an oral elicited imita- tion (EIT) task. Unlike the DELE and grammat- icality judgment task, which are both primarily prescriptive grammar tasks, the EIT can assess more implicit language knowledge by asking participants to listen to sentences and repeat them [35, 36]. This testing method has been found to engage long-term memory and require a higher level of language comprehension. It is also worth noting that, while we found significant correlations with theta coher- ence, the majority of the qEEG measures were not associated with L2 proficiency. Though we interpreted theta to reflect the engagement of working memory, one difficulty in interpreting the results of this study, is in interpreting what cognitive processes may be reflected by the different frequency bands. Although previous research does suggest associations between activity in the frequency bands and different Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 51 cognitive processes, future research will need to strengthen the validity of these claims re- garding theta. In regard to the lack of a rela- tionship between L2 proficiency and the other frequency bands, null results can be difficult to explain, but the results could be at least par- tially due to our processing procedures where a certain amount of the data was not included. For future analyses, we plan on implement- ing independent component analyses per- formed to correct for eye and muscle artifacts in EEG data (using ICLabel; [37]), which is expected to lead to lead to cleaner data, high- er sample sizes, and improved model fits for both DELE and grammaticality judgment task. We note three further limitations in our study that should be addressed in future re- search. First, future research might want to analyze data from a particular semester rather than data spanning participants from two Span- ish course levels, or the course level could be included as a covariate in analyses. Second, regarding the analysis, a more precise way to define the frequency ranges of the specific frequency bands is to use individual alpha fre- quency (IAF) peaks. For each person, the IAF peaks at a different number, which affects the ranges of the other frequency bands [38]. Fi- nally, given the highly exploratory nature of this study, we did not correct for multiple correlation analyses, and we entered regression predic- tors based on significance from the correlation- al analyses. Future research should conduct more conservative, confirmatory analyses on a dataset with higher statistical power to mitigate possible Type I errors. Indeed, post hoc analy- ses for our dataset showed that the significant correlations reported above did not survive cor- rection for Type I error inflation using the fam- ily-wise discovery rate, which further suggests that a confirmatory study would need to be con- ducted to validate any of the exploratory find- ings reported in this study. Using this method in future research may further solidify the validity of our results or may lead to different findings. CONCLUSION The purpose of the current study was to inves- tigate whether mean qEEG power and coher- ence are significant predictors of L2 proficiency. Based on our results, within- and between-net- work coherence over the theta frequency band is closely related to Spanish L2 proficiency. Because the theta frequency has been asso- ciated with memory retrieval and load, these results suggest that there is an inverse rela- tionship between L2 proficiency and reliance on memory-based strategies for interpreting linguistic inputs. Additionally, increased the- ta activity may be characteristic of individuals in earlier stages of language learning. More research is needed to further validate the sig- nificance of theta in L2 proficiency, as well as to determine the importance of the other fre- quency bands. Ultimately, this study adds to growing literature of resting-state L2 qEEG studies, echoing the implication of intrinsic pat- terns of neural activity as sources of individual variation in linguistic ability. Over time, qEEG may help to reveal individual neurophysiolog- ical variations among students within a class- room, enabling educators to develop language learning strategies that will be most conducive to successful L2 outcomes for them. In other words, from the conclusions of this body of re- search, we might be able to identify particular cognitive processes that are associated with L2 learning and proficiency. With such infor- mation, further research could then examine how to leverage these processes in instruction. AUTHOR INFORMATION Corresponding Author Victoria Ogunniyi *vogunn2@uic.edu Author Contributions Victoria Ogunniyi assisted with data collec- tion and performed the analyses along with Irene Martinez and David Abugaber. Vic- toria Ogunniyi wrote the manuscript. Kara Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 52 Morgan-Short, David Abugaber, Irene Fin- estrat, and Alicia Luque reviewed the man- uscript and devised the experimental plan. Competing Interests The authors declare no competing financial and non-financial interests. ACKNOWLEDGMENTS Thank you to Kinsey Bice for her help with per- forming the qEEG analyses. Thank you also to all undergraduate research assistants affil- iated with the Cognition of Second Language Acquisition laboratory who contributed towards data collection for this project, including Han- nah Chaddah, Sarah Hassan, Gayatri Chavan, Will Camacho, Ana Hernandez, Anahi Gaytan, Krishna Bhatia, Amber Lewis, Mallory Webber, Evelyn Delgado, Daniel Henreckson, Yasiel La- calle, and Nethaum Mizyed. We would also like to thank Sarah Malone, Christopher Powell, Crys- tal Galvan, and Allen Bryson of the University of Illinois at Chicago’s Summer Research Oppor- tunities Program for Undergraduates for their support in the development of this manuscript. ABBREVIATIONS DELE- Diplomas de Español como Lengua Extranjera EEG – Electroencephalography L2 – Second Language qEEG – Quantitative Electroencephalography REFERENCES [1] C.S. Prat, B.L. Yamasaki, E.R. Peterson, Individual Differences in Resting-state Brain Rhythms Uniquely Predict Second Language Learning Rate and Willingness to Communi- cate in Adults. J Cogn Neurosci 31(1), 78-94 (2019). [2] P. Leclercq, A. Edmonds, H. Hilton, Eds. Measuring L2 proficiency: Perspectives from SLA (Multilingual Matters, 2014). [3] C. J. Doughty, M.H. Long, “The scope of inquiry and goals of SLA” in The handbook of second language acquisition (John Wiley & Sons, 2003), pp. 3-16. [4] S. J. Luck, An introduction to the event-re- lated potential technique (MIT press, 2014). [5] K. Majumdar, A brief survey of quantitative EEG, 2nd ed. 2018 (CRC Press, 2018). [6] M. Bastiaansen, A. Mazaheri, O. Jensen, “Beyond ERPs: Oscillatory neuronal dynam- ics” in The Oxford Handbook of event-related potential components, S. Luck & E. Kappen- man, Eds. (Oxford University Press, 2012), pp. 31-50. [7] G. Gluck, QEEG accepted in death penalty trial in Florida v. Nelson. Biofeedback 39(2), 74-77 (2011). [8] M. Arns, W. Drinkenburg, J. Leon Kene- mans, The effects of QEEG-informed neuro- feedback in ADHD: an open-label pilot study. Appl Psycho-physiol Biofeedback 37(3), 171- 180 (2012). [9] C.S. Prat, T.M. Madhyastha, M.J. Mot- tarella, C.H. Kuo, Relating Natural Language Aptitude to Individual Differences in Learning Programming Languages. Sci Rep 10(1), 1-10 (2020). [10] Y. Prystauka, A.G. Lewis, The power of neural oscillations to inform sentence compre- hension: a linguistic perspective. Lang Linguist Compass 13(9), e12347 (2019). [11] L. Meyer, The neural oscillations of speech processing and language compre- hension: state of the art and emerging mech- anisms. Eur J Neurosci 48(7), 2609-2621 (2018). [12] M. Lundqvist, P. Herman, M.R. Warden, Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 53 S.L. Brincat, E.K. Miller, Gamma and beta bursts during working memory readout sug- gest roles in its volitional control. Nat Commun 9(1), 1-12 (2018). [13] S. Reiterer, C. Hemmelmann, P. Rap- pels-berger, M.L. Berger, Characteristic func- tional networks in high- versus low-proficiency second language speakers detected also during native language processing: an explor- ative EEG coherence study in 6 frequency bands. Brain Res Cogn Brain Res 25(2), 566- 578 (2005). [14] S. Reiterer, E. Pereda, J. Bhattacharya, Measuring second language proficiency with EEG synchronization: how functional cortical networks and hemispheric involvement differ as a function of proficiency level in second language speakers. Second Lang Res 25(1), 77-106 (2009). [15] S. Reiterer, E. Pereda, J. Bhattacharya, On a Possible Relationship between Linguistic Expertise and EEG Gamma Band Phase Syn- chrony. Front Psychol 2, 334 (2011). [16] O. Kepinska, E. Pereda, J. Caspers, N.O. Schiller, Neural oscillatory mechanisms during novel grammar learning underlying language analytical abilities. Brain Lang 175, 99-110 (2017). [17] A. Kielar, J.A. Meltzer, S. Moreno, C. Alain, E. Bialystok, Oscillatory responses to semantic and syntactic violations. J Cogn Neu- rosci 26(12), 2840-2862 (2014). [18] A.G. Lewis, K. Lemhӧfer, J.M. Schoffelen, H. Schriefers, Gender agreement violations modulate beta oscillatory dynamics during sentence comprehension: A comparison of second language learners and native speak- ers. Neuropsychologia 89, 254-272 (2016). [19] Z.R. Cross, A. W. Corcoran, M. Schle- sewsky, M.J. Kohler, I. Bornkessel-Schlesews- ky, Oscillatory and aperiodic neural activity jointly predict grammar learning. BioRxiv (2020). [20] C.S. Prat, B.L. Yamasaki, R.A. Kluender, A. Stocco, Resting-state qEEG predicts rate of second language learning in adults. Brain Lang 157, 44-50 (2016). [21] A. Godfroid, S. Loewen, S. Jung, J. H. Park, S. Gass, R. Ellis, Timed and untimed grammaticality judgments measure distinct types of knowledge: Evidence from eye-move- ment patterns. Studies in Second Language Acquisition 37(2), 269-297 (2015). [22] R.C. Oldfield, The assessment and anal- ysis of handedness: the Edinburgh inventory. Neuropsychologia 9(1), 97-113 (1971). Avail- able at: https://www.psytoolkit.org/survey-li- brary/handedness-ehi.html. [23] V. Kuperman, M. Dambacher, A. Nuth- mann, R. Kliegl, The effect of word position on eye-movements in sentence and paragraph reading. Q J Exp Psychol (Hove) 63(9), 1838- 1857 (2010). [24] L.A. Stowe, E. Kaan, L. Sabourin, R.C. Taylor, The sentence wrap-up dogma. Cogni- tion 176, 232-247 (2018). [25] J.A. Blanco, P.R. Donley, Vistas: Introduc- cion a la lengua espan̆ola, 5th ed. 2016 (Vista Higher Learning, Inc., 2016). [26] N.A. Macmillan, C.D. Creelman, Detec- tion theory: A user's guide (Psychology press, 2004). [27] S. Montrul, Second language acquisition and first language loss in adult early bilinguals: exploring some differences and similarities. Second Lang Res 21(3), 199-249 (2005). Columbia Undergraduate Science Journal Vol. 15, 2021 Ogunniyi et. al. 54 [28] V. Marian, H.K. Blumenfeld, M. Kaushanskaya, The Language Experience and Proficiency Questionnaire (LEAP-Q): assessing language profiles in bilinguals and multilinguals. J Speech Lang Hear Res 50(4), 940-967 (2007). Available at: https://bilingual- ism.northwestern.edu/leapq/. [29] A. Delorme, S. Makeig, EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J Neurosci Methods 134(1), 9-21 (2004). Available at: https://sccn.ucsd.edu/ee- glab/index.php. [30] MathWorks, Inc., MATLA8B 8.0 and Statistics Toolbox 8.1, Natick, Massachusetts, United States (2020). Available at: https:// www.mathworks.com/downloads/. [31] R Core Team, R: A language and environ- ment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria (2020). Available at: https://www.r-project.org/. [32] D. Makowski, The Psycho Package: An Efficient and Publishing-Oriented Workflow for Psychological Science. Journal of Open Source Software 3(22), 470 (2018). Available at: https://github.com/neuropsychology/psy- cho. [33] W. Klimesch, EEG alpha and theta oscil- lations reflect cognitive and memory perfor- mance: a review and analysis. Brain Res Brain Res Rev 29(2-3), 169-195 (1999). [34] N.A. Herweg, E.A. Solomon, M.J. Kahana, Theta Oscillations in Human Memory. Trends Cogn Sci 24(3), 208-227 (2020). [35] R. Erlam, Elicited imitation as a measure of L2 implicit knowledge: an empirical valida- tion study. Applied Linguistics 27(3), 464–491 (2006). [36] N. Spada, J.L.J. Shiu, Y. Tomita, Validat- ing an elicited imitation task as a measure of implicit knowledge: comparisons with other validation studies. Lang Learn 65(3), 723-751 (2015). [37] L. Pion-Tonachini, K. Kreutz-Delgado, S. Makeig, ICLabel: An automated electro- encephalographic independent component classifier, dataset, and website. Neuroimage 198, 181-197 (2019). [38] O.M. Bazanova, Alpha EEG activity depends on the individual dominant rhythm frequency. J Neurother 16(4), 270-284 (2012).