1. Introduction
The speakers of today's globalizing world are increasingly bi- or multilingual. A psycholinguistic perspective on multilingualism asks how multiple languages are represented and processed in a speaker's mind. A central problem is how a speaker manages to pick out words only from the intended language and to avoid intrusion from the non-intended language, a process known as lexical selection or, more generally, language control. Different theories diverge in how they account for language control, but most postulate a central role to domain-general executive functions (de Groot, Reference de Groot2011). For example, Meuter and Allport (Reference Meuter and Allport1999) suggest that language switching is not different from other types of task shifting, and that it engages the same general executive processes. The presupposition that general executive functions are central in multilingual language use also underlies the hypothesis that multilingualism enhances executive control (see e.g., Bialystok, Reference Bialystok2011).
A common experimental setup to examine processing costs related to language switching in production is picture naming, where the subject names pictures in either L1 or L2 according to a cue. A language switch typically produces a switching cost, defined as the difference between switch trials (where the language shifts from the previous trial) and repetition trials (where the language is the same as on the previous trial). There is evidence that the switching cost in language production in unbalanced bilinguals is asymmetric, with a larger switching cost when switching from the weaker L2 into the dominant L1 than vice versa (e.g., Meuter & Allport, Reference Meuter and Allport1999; Jackson, Swainson, Cunnington & Jackson, Reference Jackson, Swainson, Cunnington and Jackson2001; Schwieter & Sunderman, Reference Schwieter and Sunderman2008; Linck, Schwieter & Sunderman, Reference Linck, Schwieter and Sunderman2012; but see also Hernandez & Kohnert, Reference Hernandez and Kohnert1999; Declerck, Koch & Philipp, Reference Declerck, Koch and Philipp2012). Intuitively, one might expect just the opposite, namely that switching would be easier into the dominant language. The prominent explanation for the switching cost asymmetry is the inhibitory control (IC) model (Green, Reference Green1998), which holds that the dominant L1 is strongly inhibited by the general executive system during L2 trials to prevent its intrusion into the weaker L2, leading to a larger switching cost when moving back into L1 by resolving this inhibition. In contrast, during L1 trials the weaker L2 need not be strongly inhibited, making a switch from L1 into L2 less demanding. The IC model also predicts that the switch cost asymmetry should be proportional to the proficiency imbalance between the two languages: the weaker L2 in relation to L1, the more L1 has to be inhibited in order to produce L2.
In their review of language switching costs and asymmetry, Bobb and Wodniecka (Reference Bobb and Wodniecka2013) note that the switching asymmetry is not consistently found in language switching studies. However, in experiments where the asymmetry is discovered, it is almost always larger for the dominant L1. If the effect were due to chance, we would expect larger switch costs for L2 in approximately half of the experiments. This suggests that even if this asymmetry can be dependent on specific experimental features (e.g., cue-to-stimulus interval; see Verhoef, Roelofs & Chwilla, Reference Verhoef, Roelofs and Chwilla2009), when being present it is consistently larger for L1. The question is, then, whether this asymmetry is due to stronger L1 inhibition, as suggested by the IC model. Our main focus in this article is on the mechanisms underlying switch costs and their asymmetry.
Moreover, although the asymmetry is often considered a litmus test of inhibition, it can be argued to stem from other sources than inhibition (Bobb & Wodniecka, Reference Bobb and Wodniecka2013; see also Declerck & Philipp, Reference Declerck and Philipp2015). For instance, Philipp, Gade and Koch (Reference Philipp, Gade and Koch2007; Experiment 2) examined n-2 repetition costs, which can be considered as more direct evidence of inhibition than simple switch costs. With A, B and C representing a multilingual's different languages, the IC model predicts n-2 repetition slowing in an ABA sequence compared to a CBA sequence. In the ABA sequence, language A has been active on the n-2 trial, and according to the IC model it needs to be strongly inhibited on the n-1 trial. In the CBA sequence, A has not been recently active, and has to be inhibited less on the n-1 trial. The inhibition of A on the n-1 trial, in turn, affects the reaction time on the n trial, where inhibition of A has to be resolved. Thus, the IC model predicts an n-2 repetition cost; i.e., longer reaction time in the last trial in the ABA sequence compared to the CBA sequence. Philipp et al. (Reference Philipp, Gade and Koch2007) did find n-2 slowing, but the n-2 repetition costs did not vary depending on the speaker's proficiency in the three languages, as predicted by the IC model. Thus, the authors conclude that their results do not support the IC model. They propose an alternative framework where target language activation works in parallel with inhibitory processes. The activation model also predicts a switch cost, as the activation of a language B in a BA sequence has to be resolved before A can be produced.Footnote 1 (There are also other studies that report findings that are to some extent inconsistent with the IC model; see e.g., Costa & Santesteban, Reference Costa and Santesteban2004; Gollan & Ferreira, Reference Gollan and Ferreira2009; Christoffels, Firk & Schniller, Reference Christoffels, Firk and Schiller2007).
Efforts to more directly examine the hypothesized connection between switch cost and inhibition have been made, most of them being ERP studies. Jackson et al. (Reference Jackson, Swainson, Cunnington and Jackson2001) examined costs in predictable language switching and their electrophysiological correlates. The switch cost was larger for L1. Moreover, L2 switch trials were accompanied with a small increase in frontal negativity (the N2 component), hypothetically indicating that L1 inhibition is needed to produce a word in the weaker L2. On the other hand, Christoffels et al. (Reference Christoffels, Firk and Schiller2007) found a stronger N2 component in L1 repetition trials. This is against the IC model, which implies no inhibition during L1 repetition trials. Misra, Guo, Bobb & Kroll (Reference Misra, Guo, Bobb and Kroll2012) used a setup where identical pictures were named in single language blocks, once in English and once in Chinese. They found repetition priming effects for L2 but not for L1, which they took to indicate L1 suppression. The ERPs were overall more positive when naming in L2 after L1, and more negative when naming in L1 after L2. The authors take the results to indicate sustained inhibition of L1. All in all, the evidence from the ERP studies concerning the role of inhibition in language switching is somewhat inconsistent.
In addition to the ERP studies, some behavioral studies have assessed the connection between subjects’ inhibitory capacity and switch costs. Linck et al. (Reference Linck, Schwieter and Sunderman2012) examined the relationship between subjects’ (N = 56) performance in the Simon task, assumed to reflect inhibitory capacity, and their language switching costs. They discovered that performance in the Simon task correlated with the switching cost in a picture naming task. In particular, the Simon effectFootnote 2 correlated positively with costs when switching into the dominant L1 (English) but the correlation was lower when switching into the weaker L2 (French) or L3 (Spanish). Better inhibitory capacity (a smaller Simon effect) predicted a smaller switching cost. The authors interpret the finding in line with the IC model. However, the prediction of the IC model could also be interpreted to be the opposite. According to the IC model, switch cost asymmetry is due to a stronger inhibition of L1 during L2 (or L3) trials than vice versa, and it is plausible that subjects with a higher inhibitory capacity could inhibit L1 more strongly than those with a lower inhibitory capacity. This would produce a larger switching cost asymmetry in subjects with a higher inhibitory capacity, contrary to the Linck et al. (Reference Linck, Schwieter and Sunderman2012) findings. They note this as well, and argue that better inhibitory control need not imply stronger inhibition of the non-target lexical items, but instead more rapid inhibition. Hence, according to them, better inhibitory control may imply more brief inhibition of the non-target items. Moreover, one could argue that good inhibitors are also good at resolving inhibition so that it would not impede performance on the following trial.
Evidence for the role of inhibition in language switching also comes from a study by Liu, Rossi, Zhou, and Chen (Reference Liu, Rossi, Zhou and Chen2014) who examined the switching cost asymmetry in a picture naming task in unbalanced Chinese–English bilinguals (N = 47). They discovered symmetrical switching costs for subjects with a high inhibitory control (small Simon effect), but asymmetrical switching costs for those with a low inhibitory control (large Simon effect). These findings are in line with Linck et al.’s (Reference Linck, Schwieter and Sunderman2012) results, where lower inhibitory control also correlated with larger L1 switching costs.
Liu, Fan, Rossi, Yao and Chen (Reference Liu, Fan, Rossi, Yao and Chen2016), in turn, investigated the role of cognitive flexibility in language switching in a picture naming task. The participants (N = 52), unbalanced Chinese–English bilinguals, were divided into two groups depending on their cognitive flexibility (CF) as measured with the Wisconsin Card Sorting Task. Subjects with high CF showed symmetrical switching costs, whereas those with low CF showed asymmetrical switching costs. In other words, L1 switching costs were reduced in the high CF group, indicating that good task switchers may also be good language switchers. The Liu et al. (Reference Liu, Fan, Rossi, Yao and Chen2015) study did not directly address the role of inhibition in language switching, but instead the relationship between a subject's general set shifting capacity and language switching. Inhibition and set shifting may, however, be related, as in set shifting the non-target task schema has to be inhibited.
In addition to the switching cost, a mixing cost can be examined in picture naming tasks. Assessing the mixing cost requires using single language blocks where the language does not shift, in addition to a mixed language block. The mixing cost is defined as the difference between the repetition trials in the mixed block and the single language block trials. The mixed block repetition trials are typically processed more slowly and produce more errors than the single block trials. One way to account for the mixing cost is in terms of a central executive monitoring process. In the mixed language block, there is an ongoing process of monitoring possible task switches, whereas in the single language block these kinds of sustained attentional demands are smaller as the language does not shift.
Most studies have reported larger mixing costs for L1 than for L2 (Hernandez & Kohnert, Reference Hernandez and Kohnert1999; Christoffels et al., Reference Christoffels, Firk and Schiller2007; Prior & Gollan, Reference Prior and Gollan2011), while some have reported a mixing cost for L1 but a mixing benefit for L2 (Gollan & Ferreira, Reference Gollan and Ferreira2009; Philipp et al., Reference Philipp, Gade and Koch2007). As monitoring demands should be fairly similar for each language in a given block, these findings arguably cannot be explained solely in terms of monitoring. Instead, they could be due to either carryover effects of reactive inhibition or sustained inhibitory processes. As to the carryover effect of reactive inhibition, it is possible that L1 inhibition present during processing of the L2 items in the mixed block is not fully resolved after a language switch but carries over to the L1 repetition trials in that block as well. This would cause a larger processing cost for L1 than for L2 in the mixed block when compared to the respective single block items, i.e., an asymmetric mixing cost. Thus, the asymmetric mixing cost could be due to the same general executive mechanisms as the asymmetric switching cost, along the lines of the IC model.
Another possible source of the asymmetric mixing cost is sustained or global L1 inhibition (cf. Christoffels et al., Reference Christoffels, Firk and Schiller2007). On this suggestion, L1 is globally inhibited in the mixed block to facilitate L2 production. In line with this suggestion, some studies have reported a global L1 slowing effect, where L1 performance is slower than L2 in the mixed block, but faster in the single block; a finding suggested to indicate sustained L1 inhibition (e.g., Christoffels, Ganushchak & La Heij, Reference Christoffels, Ganushchak, La Heij and Schwieter2016).
In our study, we investigated the switching and mixing costs and a possible L1-L2 asymmetry in a picture naming task, and their relationships with performances on the Simon and Flanker tasks (commonly used to assess reactive inhibition). If reactive inhibition exerted by the general executive system is central in language switching costs as the IC model suggests, we would expect to see a relationship between the switching costs and the Simon and Flanker task performance. Supposing that the Simon and Flanker tasks tap specifically into reactive inhibition, not sustained inhibition, we expected the two tasks to correlate specifically with switch costs and their asymmetry. We hypothesized that the Simon and Flanker tasks could also correlate with the mixing cost, due to carryover effects of reactive inhibition. We also assessed the subjects’ proficiency in L2. Following Meuter and Allport (Reference Meuter and Allport1999), we hypothesized that subjects with a lower L2 proficiency need to inhibit L1 during L2 production to a greater extent than subjects with a higher L2 proficiency, thus showing a greater switching cost asymmetry. Additionally, we were interested in the more general question whether language switching and mixing costs correlate with a subject's general set shifting capacity, measured with the number-letter task. If language switching engages general set shifting processes (cf. Meuter & Allport, Reference Meuter and Allport1999), we would expect to see a relationship between the number-letter switching and mixing costs, and the language switching and mixing costs.
Compared to earlier related studies, the current study has the advantage of using multiple executive tests. Moreover, our setup also enables assessing the language mixing costs and their possible relationships with executive performance, a question not addressed before.
2. Method
2.1. Participants
The participants were 51 Finnish speaking neurologically healthy Finns (33 females) recruited via e-mail lists at the Abo Akademi University and the University of Turku in Finland. They had acquired L2 (English) mainly in elementary school as their first foreign language, starting from the 3rd grade (age 9 or 10). Home language of all participants was Finnish, and they reported having learned L1 (Finnish) from birth. The participants self-estimated their reading, writing, speaking, and listening skills on a scale from one to seven, seven representing the skills of a native speaker. The participants reported significantly better Finnish than English skills on all the measures (Z’s > 5.5). The participants’ L2 proficiency was assessed with a test where they were presented with English words and non-words, and the task was to decide whether the word is a true English word or not. On a 0–100 scale, the L2 proficiency of the subjects was on average 58 (SD = 11.8, range 31 – 80), indicating that they knew on average 58% of English words. The test is developed by the Ghent University Center for Reading Research and is available online at http://vocabulary.ugent.be. Key participant characteristics are reported in Table 1.
Participant characteristics and self-ratings.
2.2. Procedure
The study was approved by the Joint Ethics Review Board of the Departments of Psychology and Logopedics at the Åbo Akademi University. The study was part of a larger project where both a picture naming and a semantic categorization task (not examined here) were used. The participants conducted both the picture naming and semantic categorization task in one session (duration approximately 1.5 hours). At first they filled in an informed consent form and then a background information form: these probed, among other things, language and educational background and possible neurological conditions. Then the subjects performed the language tasks. The picture naming task included three blocks, Finnish and English single task blocks and a mixed block where the language switched pseudo-randomly. The order of the picture naming and semantic categorization tasks and blocks was counterbalanced between participants. However, all the three blocks in a given task were presented consecutively for each participant (i.e., the task did not change between blocks).
After the language switching tasks, the subjects were presented with the Bilingual Switching Questionnaire (Rodriguez-Fornells, Krämer, Lorenzo-Seva, Festman & Münte, Reference Rodriguez-Fornells, Krämer, Lorenzo-Seva, Festman and Münte2012; not analyzed here), the executive functions tests, and finally the L2 proficiency test. The executive function (EF) tasks used were the Simon, Flanker, and Number-Letter task. The presentation order of the executive tasks was counterbalanced.
2.3. The picture naming task
Participants named aloud photographs of common items (both animate and inanimate) in Finnish or English, depending on a color cue. Each object was isolated on a white rectangular background and standardized in size. The cue color formed the background of the whole screen, on which the photo on the white background was centered. The subject was instructed to name the picture in English if the background was red and in Finnish if the background was blue. The cues were pseudorandomized so that there were always 2–4 consecutive same-language trials. A trial began with a blank white screen (1000 ms), followed by a black fixation cross in the center of the white screen (500 ms), and then the stimulus picture on the cue-colored background was shown for 1500 ms (irrespectively of whether a response was given).
To obtain a mixing cost measure, two single language blocks (English and Finnish) were used in addition to a mixed language block. The single language blocks included 90 trials each and the mixed language block 180 trials. We had altogether 90 photos of items, and their names in Finnish and English were matched in terms of length in phonemes and log frequency (p’s > .5). The single language blocks consisted of 90 trials, i.e., all the photos in random order. The same photos were used in both single language blocks. The mixed language block consisted of 180 trials, i.e., the 90 photos named once both in Finnish and in English. Every photo was thus presented twice in the mixed block. There was at least a 10-trial interval before a photo was presented for the second time. The naming language was assigned randomly to each photo, i.e., the stimuli were bivalent. The mixed language blocks included 119 repetition trials and 60 switch trials (in sum 179 switch and repetition trials; the first trial is neither a repetition nor a switch trial), with an equal number of switches into both languages.
We had 4 variants of each block, each separately pseudorandomized. The presentation order of the blocks and variants was counterbalanced between subjects.
2.4. Determination of reaction time
We did not use the built-in voicekey of Presentation to determine reaction time, because it was easily triggered by sounds other than the actual response (e.g., “err”). To attain more reliable estimates of reaction time, we recorded the responses with a microphone and analyzed the recordings in Matlab™ using a script where we could set the threshold for volume (percent of maximum volume in a particular file) that would determine the reaction time.Footnote 3,Footnote 4 Different thresholds were examined in 4 participant recordings and compared to manual timings. The correlation between the manual timings and the script-determined timings was highest (r’s > .9) when 100% of the maximum amplitude was used as a threshold. This method does not yield the onset of vocalization, but we employed this threshold because we were primarily interested in switching and mixing effects, which are based on subtractions that eliminate the absolute differences.
The validity of the automatic timing method was examined by manually timing the vocal onset in a randomly selected group of 19 participants (10 female) from random blocks (3060 samples). Mean automatically determined reaction time was 1042 ms (SD = 259), and mean manually determined reaction time was 923 ms (SD = 229). Mean difference was thus 119 ms. The correlation between the manually and automatically determined reaction times was .87. To attain approximations of vocal onset reaction times, we subtracted 119 ms from the automatically determined RTs.
2.5. The executive tasks
The following descriptions of the Simon, Flanker, and Number-Letter task are adapted from Jylkkä, Soveri, Wahlström, Lehtonen, Rodríguez-Fornells and Laine (Reference Jylkkä, Soveri, Wahlström, Lehtonen, Rodríguez-Fornells and Laine2017), which utilized the same tasks.
The Simon task (Simon & Rudell, Reference Simon and Rudell1967) was used as a measure of inhibition. In the task, a blue or red square appears on either the left or right side of the screen. The participant has to press the left button each time a blue square appears and the right button when a red square appears, irrespective of the location of the square. On congruent trials, the square is on the same side as the correct response key (e.g., a blue box on the left side) and on incongruent trials the square is on the opposite side. On incongruent trials, the participant has to suppress the conflicting spatial information. The Simon effect is calculated by subtracting the average reaction time or error rate on the congruent trials from the average reaction time or error rate on the incongruent trials.
In the present version of the test, we used 100 trials, of which half were congruent and half incongruent, separately randomised for each subject. The trials were divided into four blocks, with 5 s breaks in-between. Before the actual test took place, all participants performed a practice sequence. Each experimental trial began with an 800 ms fixation cross, followed by a 250 ms blank interval. After this, a red or blue box appeared and remained on the screen for 1000 ms, unless a response was given. Finally, the screen was blank for 500 ms.
The Flanker task (adapted from Eriksen & Eriksen, Reference Eriksen and Eriksen1974) is another measure of inhibition. In this task, the subject is presented with an array of five horizontal arrows and has to determine the direction of the middle arrow by a left or right key press. On congruent trials, the arrow in the middle points towards the same direction as the other four arrows (the flankers); on incongruent trials, the middle arrow points towards the opposite direction to the other four. The Flanker effect is calculated by subtracting the mean reaction time or error rate of the congruent trials from the mean reaction time or error rate of the incongruent trials.
Our version of the Flanker test consisted of 100 trials, half congruent and half incongruent, separately randomised for each subject. The trials were divided into four blocks with 5 s breaks in-between. Before the actual test took place, a practice sequence was presented to each participant. Each trial began with an 800 ms fixation cross, which was immediately followed by a row of five arrows remaining on the screen for 800 ms unless a response was given. Finally, the screen went blank for 500 ms.
The Number-Letter task (adapted from Rogers & Monsell, Reference Rogers and Monsell1995) is a measure of shifting and monitoring abilities. In this task, a number-letter pair (e.g., “3A”) appears in one of two boxes appearing in one column, and the subject has to decide either whether the number is even or odd, or whether the letter is a vowel or a consonant, depending on which box the pair appears in. Each time the number-letter pair appears in the upper box, the subject has to determine the number and each time the pair appears in the lower box the subject has to determine the letter. The response was given with two keys, one for vowels or even numbers, and another for consonants or odd numbers. The task consisted of three blocks: two single task blocks (where either the number or the letter only in the number-letter pair was categorized) and one mixed tasks block. In the mixed tasks block, a trial is either a repetition trial, where the task of the subject is the same as on the previous trial, or a switch trial, where the task shifts.
The Number-Letter task yields two executive measures. The switching effect (NLSE) is calculated by subtracting the average reaction time or error rate of the repetition trials in the mixed task block from the average reaction time or error rate of the switch trials in the mixed task block. NLSE can be considered as a measure of a subject's general set-shifting capacity, with higher capacity correlating with smaller switching effect. The mixing effect (NLME) is calculated by subtracting the average reaction time or error rate of the single block trials from the average reaction time or error rate of the repetition trials in the mixed task block. Whereas the NLSE is typically taken to measure simple set shifting abilities, the NLME is generally considered as a monitoring or preparedness cost. On repetition trials of the mixed block, unlike single task trials, the subject has to monitor for possible task switches.
The single task blocks consisted of 32 trials each. The mixed task block consisted of 32 switching trials and 48 repetition trials. Of the repetition trials, 24 were number trials and 24 were letter trials. The number-letter pairs appeared in the two squares randomly. Each block was preceded by a practice sequence.
3. Results
In both the picture naming and EF tasks, an individual's reaction time on a trial was deleted as an outlier if it deviated more than three standard deviations from the individual's mean reaction time. Moreover, if a subject's overall error rate exceeded 15% in any of the EF tasks or the picture naming task, his/her reaction times and error rates in that task were excluded from analysis. In the EF tasks, one subject was thus excluded in the Simon task, and two subjects in the Number-Letter task. No subjects were excluded in the picture naming task.
Mean reaction times and error rates in the picture naming and executive tasks are summarized in Tables 2 and 3, respectively (in the picture naming task, the estimated vocal onset times are reported). All the EF cost effects in log reaction times were significant and in the expected direction (|t|(48-50)’s > 10, p’s < .001). In error rates, the effects were significant and in the expected direction on all measures (|Z|’s > 2.6, p’s < .05) except for the number-letter mixing cost (|Z| = 1.8, p > .05).
Correct reaction times (approximated onset) and error rates in the picture naming task by Language and Condition.
Reaction times (correct trials) and error rates in the EF tasks.
Normality of the item-level log-RTs in the picture naming task was inspected visually using quantile-quantile plots for each subject and stimulus separately. As a result, 15 outlier trials (.08 % of all trials) with log-transformed reaction times lower than 3.0 (equivalent to 380 ms) were deleted.Footnote 5 The resulting quantile-quantile plots showed no serious violations of normality.
3.1. Switching and mixing cost reaction times
Switching and mixing effects were examined using linear mixed models with Subject and Stimulus as random variables. The analyses were conducted in R using the package lme4 (Bates, Maechler, Balker, Walker, Christensen, Singmann, Dai & Grothendieck, Reference Bates, Maechler, Bolker, Walker, Christensen, Singmann, Dai and Grothendieck2015). The advantage of linear mixed models is that each item and subject is given an intercept of their own in the analysis. Thus, the method takes into account individual differences in response speed as well as item-specific differences. This makes the estimates for variables of interest more accurate compared to ANOVA, which can only take into account either subject or item means.
We first examined the basic language switching and mixing effects in a linear mixed effects model that included the log-transformed RT as the dependent variable and Condition (mixed-block repetition, mixed-block switch, or single block repetition) and Language (Finnish or English) as predictors. Subject and Stimulus were included as random effects. Visual inspection of the model's residual plots did not show signs of heteroscedasticity or bias.
The factors Condition and Language utilized simple coding, so that the model compares a given level of the factor to a selected baseline level. In the model, the baseline can be changed to attain estimates for the different factor levels without affecting the overall model fit. For instance, if Condition baseline is set as mixed block repetition, the remaining two levels mixed block switch and single block are compared against this baseline. In this example, the estimate for mixed block switch tells us whether the mixed block switch trials differ from the baseline (i.e., the estimate is for switching cost). With multiple predictors, an estimate is interpreted in contrast to the baseline levels of all the predictors. For instance, when the baseline of Language is set to Finnish and Condition to mixed block repetition, the estimate for mixed block switch gives the switching cost in Finnish.
Estimates from the mixed effect model are summarized in Table 4 and illustrated in Figure 1. Three different intercepts were used. Mixed-block repetition trials in L1 serve as the baseline for estimating the contrast between the baseline and L1 mixed-block switch trials (switching cost in L1), between the baseline and L1 single block trials (mixing cost in L1), and the language × switching cost as well as language × mixing cost interactions (switching and mixing cost asymmetry). The baseline mixed-block L2 repetition trials were used to estimate the switching and mixing costs in L2.
Estimated coefficients (logRT) Footnote 1 of the switching and mixing costs in the picture naming task.Footnote 2
1 Note that the log-transformed reaction times were calculated from the raw reaction times produced by the Matlab script, which were constantly 500ms too long due to a programming error (see section 2.4., fn. 3). This, however, does not affect the estimates in the mixed model, as they indicate only differences between different levels of factors; only the intercepts are affected.
2 † p < .1; * p < .05; ** p < .01; *** p < .001.
Estimated reaction times (ms) and standard errors in the picture naming task by Language and Condition.
In Table 4, under the baseline L1 mixed-block repetition the significant positive estimate for Switch indicates a significant switching cost in L1. The significant negative estimate for Single-block shows that the subjects named a picture faster in the L1 single block trials than in the L1 mixed-block repetition trials, i.e., there was a significant mixing cost in L1.
Under the baseline L2 mixed-block repetition, the significant positive estimate of Switch signals a significant switching cost in L2, while the estimate of Single-block under this baseline indicates no significant mixing cost in L2.
The significant estimates for the interactions L2 × Switch and L2 × Single indicate that the switching and mixing costs differed across the two languages. Both the switching and mixing costs were larger for L1 than for L2 (see Figure 1).
3.2. The mixing and switching cost reaction times and the EF measures
All the executive task reaction time variables were log-transformed prior to calculating the cost effects. In all the executive measures, larger effect indicates poorer executive functioning; for instance, higher Simon effect indicates weaker inhibitory capacity. Thus, a positive correlation between Simon effect and language switching cost would mean that better inhibitors are better language switchers.
Each of the four cost measures was inserted into a mixed model separately. Each model thus included logRT as dependent measure and Condition, Language, and one of the four EF cost measures as predictors, and Subject and Stimulus as random variables. In the Number-Letter task analyses, we focused on the relationship between the Number-Letter switching effect (NLSE) and the language switching cost, and the Number-Letter mixing effect (NLME) and the language mixing cost. All the interactions of the EF tasks with the language switching and mixing effects are summarized in Table 5.
Estimated coefficients of the interactions between the EF tasks and the language switching and mixing costs.
The Simon task
The Simon effect did not predict the switching cost in L1, but a larger Simon effect was associated with a smaller switching cost in L2. From Figure 2 we see that this effect was mainly due to the repetition trials: the correlation of the Simon effect with the repetition trials (E = .23, t = 1.3, p > .05) was stronger than with the switch trials (E = -.0079, t = .042, p > .05). The Simon effect also significantly predicted the positive mixing cost in L1: the L1 mixing cost became smaller as the Simon effect became larger. From Figure 3 we see that this effect was mainly due to the single block trials, where reaction times were slower the higher the Simon effect (E = .37, t = 2.1, p < .05); performance in the repetition trials did not correlate with the Simon effect (E = .045, t = .25, p > .05).
Estimated reaction times in the L2 mixed block repetition and switch trials, and the Simon effect (both log-transformed).
Estimated reaction times in the L1 mixed block repetition and single block trials, and the Simon effect (both log-transformed).
The Flanker task
The Flanker effect did not mediate the switching or mixing costs in either L1 or L2.
The Number-Letter switching effect
The Number-Letter switching effect (NLSE) did not predict the L1 switching cost, but it did predict the L2 switching cost. The L2 switching cost was larger for higher values of the NLSE (see Figure 4).
Estimated reaction times in the L2 repetition and switch trials, and the NLSE (both log-transformed).
The Number-Letter mixing effect
The Number-Letter mixing effect (NLME) did not predict the mixing cost in either L1 or L2.
3.3. The mixing and switching cost reaction times and L2 proficiency
To test whether the mixing and switching costs were predicted by L2 proficiency as measured by the participants’ score in the English vocabulary test, a model with logRT as dependent variable and Condition, Language, and L2 proficiency as predictors was used. Subject and Stimulus were used as random variables. L2 proficiency did not predict any of the switching or mixing costs (p’s > .1), except for the L2 mixing cost (E = .00027, t = 2.19, p < .05). The mixing cost was negative for subjects with low L2 proficiency, but positive for subjects with high L2 proficiency (see Figure 5). Figure 5 can be interpreted so that in both repetition and single block trials responses were faster the higher the L2 proficiency, but in the single block trials the response speed correlated more strongly with L2 proficiency than in the repetition trials.
Estimated reaction times in the L2 mixed block repetition and single block trials (in logRT), and L2 proficiency.
3.4. Analyses on error rates
The switching and mixing effects in error rates were analyzed using a binomial generalized linear mixed model with error rate as a dependent variable, Condition and Language as predictors, and Subject and Stimulus as random variables.
In L1, the error rate was higher in switch than in repetition trials (E = .82, z = 3.81, p < .001), and the error rate was lower in the single block trials than in the repetition trials (E = -.88, z = 3.74, p < .001). That is, in L1 there was a switching and mixing cost in error rates. In L2, neither the switching nor mixing costs were significant (p’s > .1). The L2 × switch interaction (switching cost asymmetry) was significant (E = -.79, z = 3.02, p < .01); there was a switching cost for L1 but not for L2. Also the L2 × single-block interaction (mixing cost asymmetry) was significant (E = .95, z = 3.67, p < .001); there was a mixing cost for L1 but not for L2.
The relationship between the executive measures and the language switching and mixing costs in error rates was examined using a binomial generalized linear mixed model with error rate as dependent variable, Condition and one of the executive measures at a time as predictors, and Subject and Stimulus as random variables. Analyses were conducted separately on L1 and L2 due to convergence issues. None of the interactions were significant (p's > .05).
3.5. Post hoc analyses
The finding that the Simon effect correlated with L1 single block performance but not with the L1 mixed block trials was unexpected. Thus, we examined the main effects of the Simon task in each trial type separately. Moreover, we assessed the effects of the congruent and incongruent trials of the Simon task separately to see whether the Simon interactions were specifically due to either condition. We also included in the analyses a new variable, global Simon RT, assumed to assess a subject's monitoring capacity (cf. Hilchey & Klein, Reference Hilchey and Klein2011). Finally, we wanted to examine possible global L1 slowing (cf. Christoffels et al., Reference Christoffels, Ganushchak, La Heij and Schwieter2016) in the mixed block and its possible correlations with the EF tasks.
We first examined the correlation of the Simon effect with performance in each naming trial type separately. The Simon effect correlated positively with naming latencies in the single blocks of both L1 (E = .37, t = 2.1, p < .05) and L2 (E = .35, t = 2.0, p < .05), but not with the mixed block trials in either language (E’s = [.0 - .23], t’s = [0 - 1.3], p’s >.1).
Secondly, we examined whether the correlation between the Simon effect and single block performance was due to either the congruent or incongruent trials in the Simon task. Neither the congruent nor congruent trials alone correlated with the single block performance in either language (E’s = [.025 - .095], t’s = [.40 – 1.43], p’s > .1), indicating that the Simon × single block RT interaction was not due to either condition alone.
Thirdly, we assessed the correlations between the global reaction time in the Simon task (Simon global), hypothesized to measure monitoring capacity (see e.g., Hilchey & Klein, Reference Hilchey and Klein2011), and picture naming performance in the different conditions. Simon global correlated with naming latencies in the mixed block (E’s = [.14 - .16], t’s = [2.1 – 2.4], p’s < .05), except for the L1 repetition trials, where the effect was nearly significant (E = .13, t = 1.97, p = .054). Simon global did not predict naming latencies in the single blocks (E’s = [.05 - .07], t’s = [.84 – 1.0], p’s > .1).
Fourth, global L1 slowing was examined with a linear mixed model with logRT as dependent variable, using Block (mixed or single) and Language (L1 or L2) as predictors, and Subject and Stimulus as random factors. In the mixed block, L2 was in fact faster than L1 (E = -.0066, t = 5.15, p < .001), whereas in the single block L2 was slower than L1 (E = .031, t = 24.23, p < .001). In other words, there was a global L1 slowing in the mixed block. This global L1 slowing did not correlate with any of the inhibitory measures or the Simon global reaction time (p’s > .05).
All in all, the additional analyses indicate that the Simon effect systematically predicted reaction times in the single blocks, and the Simon global systematically predicted reaction times in the mixed block. Global L1 slowing was significant, but did not correlate with any of the inhibitory measures.
4. Discussion
By administering a picture naming task and a set of executive tasks to Finnish–English non-balanced bilinguals, we aimed to examine the switching and mixing costs and their determinants in the dominant L1 (Finnish) and weaker L2 (English). Theoretically, we were mainly focusing on the IC model (Green, Reference Green1998) that accounts for an asymmetric switching cost in bilingual language production by relating this finding to language-specific activation levels and general inhibitory functions. According to the IC model, the dominant L1 has a higher base level activation than the weaker L2, and L1 needs to be inhibited more during L2 production than vice versa. The stronger L1 inhibition needs to be resolved when switching from L2 to L1, leading to a larger switching cost than in the opposite case. A mixing cost, in turn, can stem from two possible sources, a monitoring process and/or inhibition. In the mixed block, a central executive process monitors for possible task switches, and this can lead to a monitoring cost in the mixed block. Moreover, larger inhibition of L1 than L2 in the mixed block can cause higher L1 mixing costs, presuming that L1 is inhibited to some extent throughout the mixed block. This effect could be due to carryover effects of reactive inhibition, sustained/global L1 inhibition, or both.
The Simon and Flanker tasks were used to assess general inhibitory capacity. Both tasks were expected to tap specifically into reactive inhibition. According to the IC model, the lower a subject's inhibitory capacity (i.e., the larger the Simon or Flanker effect), the higher the switching cost should be, particularly in L1 trials which are supposedly inhibited to a greater extent (cf. Linck et al., Reference Linck, Schwieter and Sunderman2012). The mixing costs, in contrast, are often considered to be due to sustained or global inhibition (Christoffels et al., Reference Christoffels, Firk and Schiller2007), not to reactive inhibition which the Simon and Flanker tasks are assumed to measure. Reactive inhibition could, however, have carryover effects in the repetition trials, affecting the mixing costs indirectly. In the mixed block, L1 is inhibited during L2 production, and this inhibition is assumed to be resolved on the following L1 switch trial. The inhibition need not, however, be fully resolved, but instead only to the extent that the activation of L1 exceeds a critical threshold for its production. Supposing that the extent to which L1 inhibition is resolved is related to performance in the Simon and Flanker tasks, the two tasks could be related to mixing costs as well.
The Number-Letter task was used to examine general set shifting capacity. If language switching and mixing engage general set shifting or monitoring processes, we would expect the Number-Letter switching effect to correlate with the language switching cost, and the Number-Letter mixing effect to correlate with the language mixing cost (cf. Meuter & Allport, Reference Meuter and Allport1999). Finally, we assessed the possible effects of L2 proficiency. We expected, in line with the IC model, that the lower the L2 proficiency is, the more L1 would be inhibited, causing a higher L1 switching cost for low proficiency L2 speakers.
4.1. Switching costs, mixing costs, and asymmetry
In the picture naming task, a significant switching cost was observed for both languages. Moreover, the switching cost was significantly larger for L1 than for L2. This is in line with the IC model, and with earlier findings of asymmetry in production tasks when switching from a weaker language to a stronger one (see the review by Bobb & Wodniecka, Reference Bobb and Wodniecka2013).
There was also a significant mixing cost in L1 but not in L2. The L1 mixing cost could be accounted for by a monitoring process, but that should also cause a mixing cost in L2. A more plausible explanation of the asymmetric mixing cost can be made in terms of inhibition. First, it is possible that sustained/global inhibition of L1 is engaged in the mixed block to facilitate L2 processing. Second, the L1 mixing cost could be due to carry-over effects of reactive inhibition. As to the lack of a mixing cost in L2, on the other hand, the IC model implies that L2 need not be strongly suppressed during L1 production, because the baseline activation of L2 is lower than that of L1 to start with. Thus, only minimal L2 inhibition takes place in the mixed block, and performance in the mixed block repetition trials is similar to that in the single block.
The analyses on error rates yielded results largely similar to the reaction time analyses. The L2 switching cost was not significant in the error rate analysis, which is probably due to lack of statistical power in the error rate analysis.
In sum, the asymmetries in the mixing and switching costs that we observed in the present study are in line with previous results and follow the IC model. We expected that these effects would be related to the inhibitory task performances.
4.2. The mediating effects of the executive tasks
The Simon task performance did not mediate the switching cost in L1, in contrast to the prediction of the IC model and the finding of Linck et al. (Reference Linck, Schwieter and Sunderman2012) who reported that a smaller Simon effect predicted a smaller L1 switching cost. The mixing cost, in turn, showed a negative relationship with the Simon effect in L1 (the larger the Simon effect, the smaller the mixing cost). This is in contrast with the IC model, according to which a larger Simon effect should predict larger mixing costs particularly in L1, which needs to be inhibited in the mixed block. The effect was mainly due to a correlation between the Simon effect and single block performance, not repetition trials. This is surprising, as the IC model predicts a central role for inhibition particularly in the mixed block, not in the single block. Our post hoc analysis indicates that in both languages, the Simon effect correlated with single block performance but not with mixed block performance. We will discuss these effects in more detail below.
Our results concerning the mediating effects of the Simon task performance neither support the IC model nor replicate the Linck et al. (Reference Linck, Schwieter and Sunderman2012) findings. The failure to replicate the Linck et al. results could be due to a number of factors. First of all, the mixed model estimates for L1 switch cost and switch cost asymmetry were stronger in Linck et al. than in our study, which makes it more probable to find an association with the Simon task, presuming such an effect exists. The stronger switch cost and asymmetry in the Linck et al. study could stem from the use of three languages instead of two, which arguably loads the cognitive system more. Moreover, the subjects in the Linck et al. study were somewhat less proficient in their L2 (overall L2 proficiency was 7.5 on a 1–10 scale) than our subjects (overall L2 proficiency transformed into a 1–10 scale was 8.4). Lower L2 proficiency could increase L1 switch costs and asymmetry (cf. Meuter & Allport, Reference Meuter and Allport1999). Also the response-to-stimulus interval (RSI) was shorter in the Linck et al. (Reference Linck, Schwieter and Sunderman2012) study than in ours. In Linck et al. the RSI was fixed at 500 ms, whereas in our study the RSI was varying, since the stimulus remained on the screen irrespective of response. In our study, the RSI was 1500 ms plus the time when the stimulus remained on the screen. On average the RSI in our study was 2070 msFootnote 6 , which can be considered to be quite long. This could have contributed to the small switch cost and asymmetry compared to Linck et al., as longer preparation time may diminish cognitive load (cf. Declerck et al., Reference Declerck, Koch and Philipp2012).
A further potentially important difference between ours and the Linck et al. (Reference Linck, Schwieter and Sunderman2012) study is that they used only ten black-and-white line drawings which were repeated in the naming task 60 times each. Presuming that the naming language for each picture was balanced between the three languages, each picture was named 20 times in each language. In contrast, in our experiment we used 90 color photographs that were repeated only twice in the mixed block and shown only once in the single blocks. The setup of Linck et al. resembles to some extent the rapid automatized naming (RAN) task, where each stimulus is repeated several times. As suggested by Bexkens, Wildenberg and Tijms (2014), in a RAN-type setup many stimuli may be maintained in working memory in a highly accessible state. They argue that this causes competition between the previously activated stimuli and the current stimulus, which has to be resolved using inhibition. Accordingly, Bexkens et al. (Reference Bexkens, van den Wildenberg and Tijms2014) found that interference control, measured with the Simon task, predicted RAN performance in dyslexic children. This finding appears to be similar to that of Linck et al. (Reference Linck, Schwieter and Sunderman2012), although the Bexkens et al. study did not involve any language switching. Despite this difference, the picture naming task of Linck et al. can be interpreted as having characteristics that load on interference control. Thus, it is possible that in their naming task the stimulus-language pairings were to some extent maintained in working memory, possibly interfering with the current task. This interference would be resolved using inhibition. If this is the case, then the Linck et al. (Reference Linck, Schwieter and Sunderman2012) results might not reflect lexical control processes underlying language switching per se, but instead specific features of their experimental setup. Our experiment does not have this potential confound.
The Flanker effect did not mediate any of the switching or mixing costs. These results do not support the IC model or the Linck et al. (Reference Linck, Schwieter and Sunderman2012) findings, as the Simon and Flanker tasks should have a similar mediating effect, assuming that both measure similar inhibition abilities. The failure to discover similar results using the Simon and Flanker tasks could be due to differences in which aspects of inhibition they tap. The Flanker effect is based on stimulus-stimulus congruency whereas the Simon effect is based on the spatial correspondence between the stimulus and the response (cf. Costa, Hernández & Sebastián-Gallés, Reference Costa, Hernández and Sebastián-Gallés2008). In our data, the Flanker and Simon effects did not correlate with each other (r = .057, p > .1)Footnote 7 .
The Number-Letter task was taken to provide a measure of a subject's general set shifting and monitoring capacity. Comparing it to subjects’ performances in the language switching task should thus indicate to what extent language switching engages these more general executive processes. More specifically, we focused on the relationship between the Number-Letter switching effect (NLSE) and the language switching cost, and the Number-Letter mixing effect (NLME) and the language mixing cost.
The NLSE did not predict the switching cost in L1 but it did so in L2, where the switching cost was higher the larger the NLSE. This is in line with the assumption that language switching engages general set shifting processes when switching into L2. Finding the mediating effect only in L2 and not in L1 was, however, unexpected. This might indicate that switching into L1 is a more specialized process than switching into L2. In our sample, L2 was learnt relatively late (at 9 years of age on average), which might cause switching into L2 to engage general set shifting processes to a greater extent than switching into L1, a language acquired from birth. A more general reason for the lack of consistent correlations between the Number-Letter and the language switching task could be that the two tasks were not wholly analogous: L1 and L2 were discrepant in terms of dominance, while that was likely not the case for the number and letter tasks.
The NLME did not mediate the language mixing costs in either language, indicating that language mixing does not engage general monitoring processes. This was to be expected in L2, where no mixing cost was observed to begin with, but not in L1, where there was a strong mixing cost. This indicates that the L1 mixing cost is due to some other factors besides monitoring.
The interactions between the Number-Letter task and language switching and mixing costs provide only modest evidence for the hypothesis that language switching and mixing engages non-lexical or more general set shifting mechanisms. It could be that language switching and mixing are specialized processes that do not load heavily on a general set shifting capacity. Alternatively, the null results could be due to some features in our experimental setup. First, the ratio between switch and repetition trials in the Number-Letter task and the picture naming task was not equal. In the Number-Letter task, there were 40% switch trials whereas in the picture naming task there were 34% switch trials. Moreover, the task sequence was randomized in the Number-Letter task whereas it was pseudorandomized in the picture naming task. We hold that the greatest risk of these discrepancies is that they reduce the power to find significant results. In our results, only the L2 switch cost × NLSE interaction was statistically significant, and all the other interactions were far from significance. Thus, we would hold that even if the Number-Letter task and the language switching task were comparable in terms of their features, none of the currently non-significant interactions would have reached significance. One final discrepancy between the picture naming task and the Number-Letter task (as well as the inhibition tasks) is that in the former, the response was vocal, while in the latter it was a key press. As suggested by Gollan, Kleinman and Wierenga (Reference Gollan, Kleinman and Wierenga2014), a difference in response modality could affect performance. Finally, the RSI in all of the executive tasks was constant, (stimulus disappeared on response), whereas in the picture naming task the RSI was varying. All these differences could have contributed to lack of positive results, but are not likely to account for the positive effects.
In the error rate analyses, none of the interactions with the executive tasks were significant. This is probably due to the weakness of the effects and lack of statistical power when analyzing error rates, which have less variance than reaction times.
4.3. The mediating effect of L2 proficiency
L2 proficiency was examined to investigate the implication of the IC model that L1 has to be suppressed more the weaker the second language is. Thus, we expected the L1 switching and mixing costs to be larger when L2 proficiency is lower. Contrary to this prediction, L2 proficiency did not mediate the L1 switching or mixing costs, but it did mediate the L2 mixing cost. From Figure 5 we see that the lines of the mixed and single block trials cross as L2 proficiency increases. Therefore, speaking of a mixing cost in this case is somewhat misleading. Instead, a more plausible way to describe the effect is that in the single block, naming latency correlated more strongly with L2 proficiency than in the mixed block. This might be because in the mixed block the language switches reduce the effect of L2 proficiency on naming latencies.
4.4. The post hoc analyses
The interactions of the switching and mixing costs with the Simon task were unexpected. A particularly surprising result was the negative correlation between the Simon effect and L1 mixing cost, which was mainly due to the Simon effect correlating with the single block performance. Hence, we conducted post hoc analyses to see how the Simon effect correlated with the single and mixed block performance. Moreover, we introduced a new predictor, i.e., global reaction time in the Simon task, as a measure of monitoring capacity in the Simon task. Finally, we examined whether the observed L1 mixed block slowing was significant, and mediated by the inhibitory task performance. The IC model implies that global L1 slowing is due to sustained L1 inhibition in the mixed block, or possibly carryover effects of reactive inhibition.
The Simon effect correlated with single block performance in both languages, but not with mixed block performance in either language. This finding is contrary to the IC model, which postulates a central role to inhibition particularly in the mixed block, not in the single block. One possible explanation for this finding is that the Simon task taps sustained inhibition, which is required in the single language block, but not so much with reactive inhibition, which is more central in the mixed block. Typically, however, the Simon task is taken to measure reactive inhibition. Another possible explanation can be formulated along the lines of the adaptive control (AC) hypothesis, proposed by Green and Abutalebi (Reference Green and Abutalebi2013). The model distinguishes between different interactional contexts and the demands they place on cognitive control processes. Interference control is supposed to be central in single and dual language contexts, where it is important to keep the two languages separate. In a dense code switching context, in contrast, a speaker may use any language s/he is proficient in, a control process the authors call opportunistic planning. Interference control is not required in a dense code switching context. If we suppose that the single language block is a single language context and the mixed language block a dense code switching context, then the AC model would predict a correlation between the Simon effect and naming latencies particularly in the single block, not in the mixed block. The problem with this suggestion is that the mixed block does not appear to be a dense code switching context. Central to a dense code switching context is the possibility of using any language, whereas in the mixed block the subject names pictures in a pre-determined language depending on a cue. Thus, it is implausible that opportunistic planning would be utilized in the mixed block. All in all, it is unclear where the interaction between the Simon effect and single block performance stems from, and why no such interaction occurs in the mixed block. Further research is clearly needed to see whether this effect is replicable.
In contrast to the Simon effect, Simon global reaction time predicted performance particularly in the mixed block but not in the single blocks. The effect is not arguably due to differences in the subjects’ overall response speed, as in this case we would expect Simon global to correlate with naming latencies in both the mixed and single blocks. This finding could be interpreted as indicating a more central role of monitoring in the mixed block in contrast to the single blocks. On a related note, several studies have found a bilingual advantage in global reaction times in the Simon task or similar tasks, but not in the interference effect (see Hilchey & Klein, Reference Hilchey and Klein2011). Our finding could be interpreted in line with these results: for some reason, language mixing engages monitoring capacity but not reactive inhibition.
4.5. Possible alternative accounts of the results
Our main goal in this study was to assess whether language switching and mixing costs and their asymmetries, often considered as hallmarks of inhibition, correlate with subjects’ performance in inhibitory tasks. The observed correlations did not support the IC model and were partly contrary to it. We will next examine as to what could account for the present results.
Our results are somewhat compatible with the suggestion by Philipp et al. (Reference Philipp, Gade and Koch2007), according to which the switching cost asymmetry is due to top-down activation rather than inhibition. The weaker L2 has to be more strongly activated than the stronger L1, and this L2 activation has to be resolved on the following L1 switch trial to avoid cross-language intrusion, causing the asymmetric switching cost. L2 activation could also underlie the mixing cost and its asymmetry: although L2 activation is strongest during the L2 trials and the immediately following L1 switch trial, it could have carryover effects in the following L1 repetition trials as well. Also sustained or global L2 activation could occur in the mixed block to facilitate L2 processing, which would disrupt L1 production and cause the asymmetric mixing cost. According to this activation model, inhibition is not central in lexical selection, which would explain the lack of interactions of the switching and mixing cost measures with the inhibitory tasks. This suggestion does not, however, explain the correlations between the Simon effect and single block performance.Footnote 8
One could raise a potential problem with both the IC and the activation model. Let us look more closely at how the two models account for the L1 switching cost and the L2 switching cost. The IC model explains the L1 switching cost in terms of resolving L1 inhibition from the previous L2 trial. The activation model, on the other hand, would explain the L1 switching cost in terms of resolving the L2 activation from the previous trial. But due to the same mechanisms, both models arguably imply a switch cost for L2 as well. On the IC model, L1 inhibition is engaged on an L2 switch trial, which takes time. On the activation model, L2 activation is engaged on an L2 switch trial, which also takes time. At issue is why activation or inhibition would take less time than resolving that activation or inhibition. This is necessary in order for the L1 switch cost to be larger than the L2 switch cost. As a response to this, one could argue that the engagement of activation/inhibition is a fast, active process, but they fade away passively and slowly. This would explain why switching into L2 is faster than switching into L1: on an L2 switch trial L2 is activated or L1 is inhibited quickly, whereas on an L1 switch trial the L2 activation or L1 inhibition is still lingering.
If we endorse this interpretation of the IC model, the prediction of the IC model with respect to the mediating effect of the Simon task arguably differs from our and Linck et al.’s (Reference Linck, Schwieter and Sunderman2012) interpretation. If the L1 switch cost is mainly due to slowly fading L1 inhibition, this process may not correlate at all with a subject's inhibitory capacity as measured with the Simon or Flanker tasks, where better performance is arguably correlated with faster engagement of inhibition. Better inhibitory capacity would more likely be correlated with L2 switch costs, as on these trials L1 inhibition is engaged. The more rapidly L1 is inhibited, the faster L2 can be produced. Our results did not support this interpretation.
The activation model can be considered to be more in line with the results than the IC model, although the evidence is mainly negative: the activation model is compatible with the lack of interactions with the inhibitory tasks. Moreover, the activation model cannot account for the correlation between the Simon task and single block performance.
The basic switching and mixing cost results could also be accounted for by a model that does not rely on general inhibitory capacity, but instead on within-lexicon excitatory and/or inhibitory connections. One could argue along the lines of the bilingual interactive activation (BIA) model of language reception (Dijkstra, van Heuven & Grainger, Reference Dijkstra, van Heuven and Grainger1998) that lexical representations are connected to a language node that globally activates all of the same language representations and inhibits other language representations. Asymmetry could follow if the excitatory connections between L2 language representations were stronger than those between L1 representations, or if the inhibitory connections from L2 to L1 were stronger than from L1 to L2. On an L1 switch trial, the L2 activation or L1 inhibition from the previous L2 trials has to be resolved. On an L2 switch trial, on the other hand, L1 is not strongly activated or L2 inhibited on the previous trials, making a switch into L2 less demanding than a switch into L1. This could also account for the mixing cost for L1: the activation of L2 or inhibition of L1 in the mixed block hinders L1 production; hence L1 activation is submaximal throughout the mixed block. In contrast, in the L1 single block no L2 activation or L1 inhibition takes place, and L1 activation is maximal. The problem with this approach is that it does not account for the effect of the Simon task in the single blocks.
Yet another alternative explanation of our results could be formulated along the lines of the L1 repeat-benefit hypothesis (Verhoef et al., Reference Verhoef, Roelofs and Chwilla2009), which holds that in language production in unbalanced bilinguals, task schema competition is present in all other conditions except for the L1 repetition trials, which are disproportionately fast. In L1 switch trials the L2 schema is still active, hindering L1 production. In L2, the subjects suffer from interference of the stronger L1 in both repetition and switch trials. Thus, according to the repeat-benefit hypothesis, the asymmetric switching cost is due to the subjects responding disproportionately fast to the L1 repetition trials, not their being exceptionally slow in the L1 switch trials. However, the model is not compatible with our results as in our data the L2 repetition trials were actually responded to faster than the L1 repetition trials, making it unlikely that the switching cost asymmetry is due to an L1 repeat benefit.
The most problematic finding with respect to forming a consistent theoretical account of the results is the correlation between the Simon effect and single block naming performance. The finding contradicts the IC model, which implies a central role for inhibition in the mixed block but not in the single blocks. Above we speculated how the finding could be accounted for by the AC model: on this account, the mixed block would be considered a dense code switching context, where opportunistic planning is used and no interference control is required, whereas the single block would be a single language context, where interference control is central (Green & Abutalebi, Reference Green and Abutalebi2013). This account would be consistent with the pattern of results overall, but the problem is that it is implausible that opportunistic planning would be used in the mixed block. In opportunistic planning, the subject may use any language most readily available, but in the mixed block the correct language to be used is pre-determined.
In sum, our results do not support, and partially conflict with, the IC model. The results are somewhat consistent with the top-down activation model or a BIA-type within-lexicon lateral inhibition/activation model, as these models do not imply a role for general inhibition in language switching and mixing. However, none of the models are compatible with the results overall, particularly the correlation between the Simon effect and single block naming performance and the lack of correlation in the mixed block. The AC model could account for the results if the mixed block could be considered as a dense code switching context. This, however, is implausible. All in all, the results are not compatible with any existing theoretical framework.
5. Conclusion
In language production tasks, the asymmetric switching cost is often considered as a litmus test for inhibition, as suggested by the IC model. We examined the mediating effect of the Simon and Flanker tasks, presumed to measure a subject's general inhibitory capacity, on the switching and mixing costs. In our study, neither the Simon nor the Flanker task performance mediated the language switching and mixing costs in the direction predicted by the IC model. Moreover, the Simon task performance predicted naming latencies in the single language blocks but not in the mixed language block, a finding that directly contradicts the IC model. These findings indicate that the language switching and mixing effects and their asymmetries are not due to inhibition. On the other hand, our results lend some support for the hypothesis that conflict monitoring is central in the mixed language block and that switching into the weaker language may engage general set shifting capacity. All in all, the results were not compatible with any existing theoretical framework. This calls for further investigation to see whether the findings (or lack of them) are robust.
