Reversed asymmetrical switch costs in voluntary language switching: Evidence from typed responses

Current study

To further investigate the possibility of (reversed) asymmetrical switch costs during voluntary language switching, we set out to investigate a relatively large group of German-English bilinguals. This would help us determine whether voluntary language switching is a boundary condition for asymmetrical switch costs.

Unlike the studies discussed above, we relied on typed responses instead of spoken responses to further increase the generalizability of (a)symmetrical switch costs during voluntary language switching. A recent study has indicated that switch costs can be observed when bilinguals type their responses in a cued language-switching experiment (Roembke et al., 2023; for a similar effect with handwriting, see Wong & Maurer, 2021). While many similarities can be observed between spoken and typed responses, prior research has also indicated distinct differences (e.g., Pinet & Nozari, 2018; Scaltritti et al., 2016, 2018). For instance, based on a finger-twisting task, Pinet and Nozari (2018) provided evidence for feedback between postlexical and lexical layers during typing, similar to what has been suggested during spoken production (e.g., Dell, 1984). Yet, this study also showed differences between typing and speaking. The authors observed a substantial number of phonotactic violations during typing, a phenomenon that is hardly ever observed in studies focused on spoken production (e.g., Warker & Dell, 2006). This finding indicates differences in postlexical representations between typing and writing. The latter finding is important, as prior studies of language switching have observed evidence for an important role of postlexical representations (e.g., Declerck & Philipp, 2015b). Moreover, the fact that other intrinsic differences have been observed across studies that focused on spoken and typed production that have not been investigated but might also influence language control (e.g., Scaltritti et al., 2018) indicates the necessity to generalize production-based language control to other modalities than spoken production, such as typing.

Participants

Eighty-eight German natives who spoke English as a second language took part in this study. Six of these participants were not taken into the final analyses because they had not finished the experiment and another two participants were not included due to an excess of errors (>40%). Finally, one participant was not included in the final analysis due to a very low number of language switches (<1%). The remaining 79 German-English bilingual participants consisted out of 38 women and 41 men who were on average 23.4 years old (SD = 3.2). Prior to the main experiment, the participants also filled in a questionnaire about their English experience and proficiency (see Table 1). Participants were rewarded with 8€/h (as the experiment was combined with two other, completely independent experiments [cf. section Procedure], the total payment for a participant was between 24 and 30 €).

OPEN IN VIEWER

Table 1.

Overview of the demographic information (SD in brackets).

Demographic information	English
Formal education (years)	8.6 (1.9)
Age-of-acquisition (years)	10.5 (1.2)
Speaking a	5.0 (1.2)
Understanding a	5.7 (1.1)
Writing a	4.8 (1.3)
Reading a	5.8 (1.1)

Note. The information consists of the average years of formal English education, English age-of-acquisition, and average self-rated scores for speaking, understanding, writing, and reading English.

a

1 being very bad and 7 being very good.

Materials and task

Ten pictures (black line drawings, Snodgrass & Vanderwart, 1980) were used in this study, each name of which consisted of five letters in German (word frequency per million: 41.6; Brysbaert et al., 2011) and in English (word frequency per million: 84.7; Brysbaert & New, 2009; see Appendix 1 for an overview of the picture names in German and English). To make sure we had enough stimuli that had an equal number of letters in both languages, half of the names were cognates. Additional analyses with cognate status added as a factor to each of the analyses showed that cognate status did not significantly influence the switch costs or their asymmetry (ts < 1.551; zs < .925), similar to several other spoken production studies (e.g., Santesteban & Costa, 2016; Verhoef et al., 2009; see also Li & Gollan, 2018). Hence, we did not include this factor in the final analysis.

Procedure

The study reported here was conducted in combination with two other studies. The other two tasks involved a persistence task (symbol coding) and a decision-making task and, thus, are completely independent of the language-switching experiment. Testing sessions were conducted with four participants sitting in a 24 m² room, with desks oriented to the walls to minimize distraction and/or interaction. Including breaks of 15 minutes between each substudy, the whole testing session lasted for about 3 hours on average. The three parts were balanced using a Latin square approach. Prior to starting the sessions, participants gave written informed consent, and the study was conducted in accordance with the Declaration of Helsinki.

The language-switching experiment started off with the instructions. The instructions for the voluntary language switching, which was run based on E-Prime, entailed the following:

When naming the pictures, you should always switch back and forth between German and English and decide for yourself when and how often you change the language. However, make sure that you use both languages about equally often and avoid a fixed rhythm, such as changing the language after every (second) picture.

Participants were also informed that each word contained five letters and that corrections were not possible. Furthermore, we asked the participants to type all letters in lowercase (even the first letters for the German words) on a German QWERTZ keyboard. Participants did not receive visual feedback when typing.

Following the instructions, a practice block was presented in which each picture was presented twice for a total of 20 trials. Finally, the 14 experimental blocks, with 40 trials each (total number of trials per participant = 560), were presented. In each of the experimental blocks, every picture was presented four times.

Each trial started with a fixation cross in the middle of the 22″ screen (white background color, viewing distance approximately 60 cm). The fixation cross was replaced by a picture after 500 ms. The picture was visible while participants typed their response and disappeared after the participant’s last (i.e., fifth) keystroke, which in turn was followed by the next trial after 1,000 ms.

Data analyses

The first trial in each block (2.5% of all data) was excluded from the reaction time (RT) and error rate analyses because these were neither switch nor repetition trials. In the RT analyses, error trials and trials immediately following an error were excluded (17.6% of all data), as were trials with an RT faster than 20 ms or 2.5 SDs below and above the mean of each keystroke across participants (4.2% of the data for onset latencies and 1.7% of the data for interkeystroke intervals).

A separate RT analysis was conducted for onset latencies, which consisted of the interval between picture onset and the first keystroke, and inter keystroke intervals, which consisted of the intervals between consecutive keystrokes while producing a word (e.g., keystroke 1 → keystroke 2 and keystroke 2 → keystroke 3; cf. Scaltritti et al., 2016), with factors Trial type (switch vs. repetition trial) and Language (German vs. English). The latter is a novel dependent variable that has not been used in language-switching research yet but can provide additional insight into whether language control also affects the actual execution of a word. Finally, one overall error analysis was conducted with the factors Trial type (switch vs. repetition trial) and Language (German vs. English) since assigning errors to a keystroke (initial vs. keystroke 2–5) is not always straightforward because mistakes were sometimes made at several keystrokes or a keystroke was missing. Furthermore, most of these more specific analyses would be underpowered due to a lack of errors. In total, we had 43.134 data points across all participants to analyze error data, 33.490 data points across all participants to analyze the onset latencies data, and 138.478 data points across all participants to analyze the interkeystroke intervals data.

The RTs were analyzed using linear mixed-effects regression modeling, whereas the error rates were analyzed using logistic mixed-effects regression modeling. Both participants and items were considered random factors with all fixed effects and their interactions varying by all random factors. Yet, convergence and singularity issues were taken into account. ² We used effect coding for all two-level factors (i.e., −0.5 and 0.5). Finally, t- and z-values larger or equal to 1.96 were deemed significant (Baayen, 2008).

(Reversed) Asymmetrical switch costs

Therefore, in line with Liu et al. (2021) and Sánchez et al. (2022), we found that switch costs can be modulated by language dominance in a voluntary language-switching setting, resulting in larger switch costs for L2 compared to L1. This is, however, not in line with the voluntary language-switching studies that came before those studies, which all observed symmetrical switch costs (de Bruin et al., 2018, 2020; Gollan et al., 2014; Gollan & Ferreira, 2009; Gross & Kaushanskaya, 2015; Grunden et al., 2020; Jevtović et al., 2020). One possible explanation for this discrepancy across voluntary language-switching studies is that studies that observed symmetrical switch costs typically relied on highly proficient bilinguals, whereas the three studies that observed (reversed) asymmetrical switch costs relied on participants that would more likely be characterized as second-language learners (Declerck & Koch, 2023). Prior research has indicated that highly proficient bilinguals tend to show symmetrical switch costs and second-language learners asymmetrical switch costs (e.g., Costa et al., 2006; Costa & Santesteban, 2004). This explanation could clarify the discrepancy across studies and, thus, also indicate that voluntary language switching is not a boundary condition for asymmetrical switch costs.

That still leaves the question of why the asymmetrical switch cost pattern was reversed (L2 switch costs > L1 switch costs) in the current study from what is typically observed (L1 switch costs > L2 switch costs) in nonvoluntary language-switching studies (e.g., when using cued language switching or alternating language sequences). Yet, larger L2 than L1 switch costs is not unique to voluntary language switching. There have been several language-switching studies that relied on other variants of language switching that also observed a reversed asymmetrical switch cost pattern (e.g., Declerck, Stephan, et al., 2015; Timmer et al., 2019; Zheng et al., 2020). These studies typically explained this pattern through conditions that make L2 the relatively more activated language, for instance, through L1 proactive language control or through a high L2 proficiency/exposure. While it seems unlikely that the reversed asymmetrical switch cost pattern in the current study is due to a very high L2 proficiency/exposure, since these were second-language learners who lived in Germany and were tested by a German experimenter at a German-speaking university, it could very well be that the participants implemented L1 proactive language control throughout the experiment to make both languages equally accessible. An argument against this explanation is that one would expect to observe overall faster L2 than L1 responses if this were the case. However, it could very well be that the typical language dominance pattern (i.e., faster L1 than L2 responses) would only be slightly affected by L1 proactive language control, which in turn would not necessarily result in overall faster L2 than L1 responses (cf. Declerck et al., 2020).

Another possibility is that the bilinguals strategically avoided producing the more “difficult” words (e.g., words with a low word frequency) in their L2 and thus chose to mainly produce these in their L1 (Gollan & Ferreira, 2009; Sánchez et al., 2022). The choice to produce these “difficult” words in their L1 would stem from the fact that these words should be easier to produce in their L1 than their L2 since their L1 is the more dominant and more used language. Together, this should result in a relatively higher L2 activation, at least for the words used in the task. According to the inhibitory account of asymmetrical switch costs (Green, 1998; Meuter & Allport, 1999), this should lead to more inhibition on L2 and thus result in more persisting inhibition to overcome for L2 than L1 switch trials. In turn, larger L2 than L1 switch costs should be observed, which was the case in the current study.

Language control can also affect execution

It is typically assumed that language control is implemented during word selection and thus prior to execution. Our results show that language control can also have an influence on the execution process of language production. More specifically, switch costs and their reversed asymmetry were also observed in the interkeystrokes (i.e., the intervals between the five keystrokes). Since the word should already be selected at this point, this indicates that execution is also affected by language control.

There are several ways of interpreting this result. On the one hand, it could be that there is a spillover of language control from the word selection process into the execution process. On the other hand, it could be that language control has a direct influence on language execution. While the current study does not provide evidence for either interpretation, there is some evidence along the lines of the latter: Using a go/no-go delay (100 or 1,500 ms) setup in a language-switching task, Philipp and Koch (2016) showed that when a response was selected, but not produced, no switch costs were observed in the next trial. In contrast, when the response was produced, substantial switch costs were observed. While this indicates that language control can occur at the execution level, it does not preclude the possibility that there is spillover from word selection to execution.