Grammatical Constraints on Language Switching: Language Control is not Just Executive Control

Tamar H Gollan; Matthew Goldrick

doi:10.1016/j.jml.2016.04.002

. Author manuscript; available in PMC: 2017 Oct 1.

Published in final edited form as: J Mem Lang. 2016 May 18;90:177–199. doi: 10.1016/j.jml.2016.04.002

Grammatical Constraints on Language Switching: Language Control is not Just Executive Control

Tamar H Gollan ¹, Matthew Goldrick ²

PMCID: PMC5033271 NIHMSID: NIHMS781424 PMID: 27667899

Abstract

The current study investigated the roles of grammaticality and executive control on bilingual language selection by examining production speed and failures of language control, or intrusion errors (e.g., saying el instead of the), in young and aging bilinguals. Production of mixed-language connected speech was elicited by asking Spanish-English bilinguals to read aloud paragraphs that had mostly grammatical (conforming to naturally occurring constraints) or mostly ungrammatical (haphazard mixing) language switches, and low or high switching rate. Mixed-language speech was slower and less accurate when switch-rate was high, but especially (for speed) or only (for intrusion errors) if switches were also ungrammatical. Executive function ability (measured with a variety of tasks in young bilinguals in Experiment 1, and aging bilinguals in Experiment 2), slowed production and increased intrusion rate in a generalized fashion, but with little or no interaction with grammaticality. Aging effects appeared to reflect reduced monitoring ability (evidenced by a lower rate of self-corrected intrusions). These results demonstrate robust effects of grammatical encoding on language selection, and imply that executive control influences bilingual language production only after sentence planning and lexical selection.

Keywords: bilingualism, aging, speech errors, intrusion errors, read aloud, switching, inhibition, reversed dominance

Some of the most impressive demonstrations of agility in the domain of spoken language production include the prosodic and expressive speech often exhibited by newscasters or storytellers, unusually rapid production in a single language (e.g., when sportscasters describe team sports action live), and fluid alternation back and forth between languages by bilingual or multilingual speakers. Indeed, language switching has become one of the most broadly studied topics in the field of research on bilingualism—without question, it has played a major role in shaping theories of control over bilingual language selection. With relatively few exceptions the bulk of research on this topic has focused somewhat narrowly on obligatory and intended switches in out of context speech, relatively little on switches in connected speech, and even less on unintended language switches, though these can provide a unique and powerful form of evidence on the cognitive mechanisms underlying bilingual language selection.

Bilingual language switching can be considered as a specific example of the more general problem of regulation and control of production processes, which may be driven by two general types of mechanisms. Language-specific knowledge clearly plays a role in controlling speech production. For example, word exchange errors (I’m writing a mother to my letter) overwhelmingly respect word class – or part of speech (i.e., nouns exchange with nouns, verbs with verbs, etc.; Garrett, 1975), suggesting that knowledge of syntactic properties helps control which lexical items are selected for production (because lexical items in different parts of speech serve distinct syntactic functions; see Dell, Oppenheim, & Kittredge, 2008, for review and discussion). In addition to such domain-specific knowledge, mechanisms that support goal directed behavior in non-linguistic cognitive domains might also support control of speech production. For example, the need to select from among competing response alternatives arises across many situations, and common mechanisms may subserve this function across these cases (e.g., Thompson-Schill, D’Esposito, Aguirre, & Farah, 1997). Few studies have considered how these two types of mechanisms might function jointly to control bilingual language selection, and fewer still have considered how language-specific and domain-general factors might modulate the control of both intended and unintended language switches in production of connected speech.

Factors Modulating Control of Language Switching in Out-of-Context Speech

A great deal of the work in bilingual language control has examined single word production tasks. These have suggested a number of task properties that seem to facilitate switching between languages. Onset of speech in such studies is typically slower when speakers are cued to switch languages relative to when they are cued to continue speaking the same language they used on the previous trial. This difference is assumed to reflect switch costs, i.e., the processing cost associated with switching languages. Switch costs are smaller, but not eliminated entirely, when preceded by longer preparation times (e.g., Costa & Santesteban, 2004; Fink & Goldrick, 2015; Philipp, Gade, & Koch, 2007; Verhoef, Roelofs, & Chwilla, 2009), when they are predictable (Declerck, Koch, & Philipp, 2015), when bilinguals know exactly which words they will produce ahead of time (Declerck, Philipp, & Koch, 2013), and when they are voluntary (Gollan & Ferreira, 2009) rather than forced by an experimentally provided cue (Gollan, Kleinman, & Wierenga, 2015). Switch costs are found even when no overt switch is produced (e.g., when alternating between reading a word in one language and producing a word in the other; Peeters, Runnqvist, Bertrand, & Grainger, 2014). Cost free switches have been reported in just a few cases. In one study, switch costs were found in a task performed in both languages, but not if each task was performed in just one language (e.g., digit naming in one language, and picture naming in the other; Finkbeiner, Almeida, Janssen, & Caramazza, 2006). In another, cost free switches were found when bilinguals switched languages voluntarily, but with an experimental requirement to use each language about equally often which led switching to become the default behavior (Gollan & Ferreira, 2009, Experiment 2). Finally, cost free switches were also found in voluntary switching when a small set of pictures was presented repeatedly and bilinguals chose (Gollan et al., 2015) or were instructed (Kleinman & Gollan, in press) to use just one language to name each picture and then consistently used only that same language on all subsequent appearances of each picture. Together these studies reveal both the persistence of switch costs in a variety of experimental settings, thereby resembling the literature on non-linguistic task-switching, but also the possibility that language switches are sometimes as efficient as (if not more efficient than) using just one language. Such cost-free switches might be easier to observe when presented with contextual support, a property that might be easier to study in the domain of language.

Grammatical structure

Sentence contexts, and connected speech as it is produced when multiple sentences are strung together, include grammatical as well as semantic structure that may facilitate control of language selection. Studies of spontaneous code switching corpora have observed that grammatical properties constrain the distribution of naturally occurring language switches in connected speech (e.g., Poplack, 1980; Muyksen, 2000; Myers-Scotton, 1997; 2005; 2006). However, experimental studies of language switches in sentence contexts have yielded inconsistent results. In one study, cued switches between sentences in unscripted connected speech (descriptions of actions shown in pictures) were found to be costly (Tar owski, Wodniecka, & Marzecová, 2013). Another study revealed speech to be slower when preceding naturally occurring code-switches relative to single-language speech (while controlling part of speech and utterance length; Fricke, Kroll, & Dussias, 2015). However, naturally occurring code-switches might sometimes be initiated to recover from access difficulties in one language, thereby masking the possibility that some switches are fully intentional and cost-free. In a different study, no switch costs were found but connected speech was not measured; sentence context was read silently, and only a single (highlighted) target word within each sentence was produced long after the language switch actually occurred (Gullifer, Kroll, & Dussias, 2013). Finally, another study had bilinguals (professional translators and highly proficient matched controls) read sentences one word at a time with self-paced button presses and exhibited switch costs only if they later had to repeat the sentence aloud (Ibáñez, Macizo, & Bajo, 2010).

Beyond simply asking participants to switch within sentences, some work has explicitly examined whether switch costs are modulated by the extent to which they match habitual, or henceforth grammatical, patterns of switching in natural language use. Some results suggest that the size of switch costs is influenced both by the exact location of the switch within the sentence, and the type of verb participating in the switch (Dussias, 2003; for similar evidence using fMRI see Rossi, Ting, Diaz, Newman, van Hell, & Dussias, submitted). Specifically, faster reading times (measured in gaze duration times) were observed for switches that occurred at a syntactic boundary (i.e., before the auxiliary; terroristas have injured) than between the auxiliary and the main verb (terroristas han injured). However, this was found much more for closely bound syntactic elements (haber + participle), than for less closely bound elements (i.e., estar + participle). The former is more bound to the participle in that it cannot occur on its own but estar can, and there appear to be stronger restrictions on the occurrence of switches that split haber than estar from the participle. These findings imply grammatical constraints on the magnitude of switch costs in bilingual sentence processing that mirror the distributional properties of naturally occurring switches.

However, a recent investigation of language switches in full sentence production provides more mixed support for the notion that switch costs are influenced by grammaticality. In this study, German-English bilinguals memorized and then repeatedly produced four different sentences with five words in them while switching languages on every other word, and with a forced 1500 ms pacing interval between words (Declerck & Philipp, 2015a). Sentences with shared word order in the two languages exhibited negligible switch costs (just 6 ms), whereas sentences with different word order in each language, or scrambled sentences, exhibited larger switch costs (36 and 40 ms respectively). On this basis the authors suggested that language switch costs can be abolished by sentence structure, but only in this very restricted special case (fully overlapping syntax; consistent with the equivalence constraint proposed by Poplack, 1980). Other syntactic constraints within this same study seemed to have no effect on switch costs. Specifically, switches that occurred on functional elements (e.g., you, he, this) were not more costly than switches that occurred on content words (e.g., boy, flowers, shopping), even though naturally occurring switches tend not to involve insertion of single functional elements (Muysken, 2000; Myers-Scotton, 1993).

Part of speech, specifically the contrast between functional elements (which signal grammatical relationships between words) and content words (which communicate more meaningful elements), appears to exert powerful but diametrically opposed constraints on intended versus unintended language switches (e.g., saying pero instead of but), or henceforth, cross-language intrusion errors. Unlike intentional switches (which, as just noted, rarely involve single functional elements), most spontaneously produced intrusion errors involve single function word targets (and interjections; Poulisse, 1999). Although bilinguals rarely lose control over language selection (Gollan, Sandoval, & Salmon, 2011), intrusions can be induced in large numbers by asking bilinguals to read aloud paragraphs that switch languages haphazardly (Kolers, 1966). As discussed in more detail below, based on data from eye-tracking and errors in this task, Gollan, Schotter, Gomez, Murillo, and Rayner (2014) argued that the majority of such intrusions arise within speech production processes. Gollan et al.’s (2014) results confirmed the vulnerability of function word targets to intrusion errors in production in Spanish-English bilinguals. Function words outnumbered content words as targets of intrusion errors, particularly for English targets in paragraphs with Spanish word order. Weaker effects were observed for English targets in English word order (and across paragraphs, Spanish targets exhibited relatively less prominent part of speech effects). The sensitivity of part of speech effects to word order implies that they are not simply an artifact of other non-syntactic differences between function and content words (e.g., function words might draw less attention because they tend to be shorter than content words; Poulisse & Bongaerts, 1994).

Executive functions

Some have suggested that bilingual language control is achieved, at least in part, by utilizing domain-general executive control processes (Abutalebi & Green, 2007; Hernandez, 2009; Kroll, Bobb, Misra, & Guo, 2008), including the same processes that support goal directed behavior outside the domain of linguistic functioning. This proposal has been supported by studies showing bilinguals are sometimes advantaged relative to monolinguals on tests of executive function (Bialystok, Craik, Green, & Gollan, 2009), with such advantages sometimes linked explicitly to language switching ability and frequency (Hartanto & Yang, in press; Prior & Gollan, 2011; Soveri, Rodriguez-Fornells, & Laine, 2011; Verreyt, Woumans, Vandelanotte, Szmalec, & Duyck, 2016), ability to avoid unwanted switches (Festman & Münte, 2012), or proficiency (Tao, Taft, & Gollan, 2015). However, the literature specifically comparing language switching and task switching often suggests that there is relatively little overlap in the underlying processing mechanisms (Calabria, Hernández, Branzi, & Costa, 2012; Gollan, Kleinman, & Wierenga, 2014; Prior & Gollan, 2013), and a number of investigators have been unable to replicate bilingual advantages (for review see Hilchey, Saint-Aubin, & Klein, 2014; Paap, Johnson, & Sawi, 2015; Valian, 2015).

Some of the most often cited support for the role of executive control in bilingual language selection has come from language dominance effects in studies of cued language switching. Most bilinguals have one language that is dominant i.e., relatively more proficient in many or most respects, than the other, non-dominant, language. Switch costs are often larger in the dominant language than in the non-dominant language, a result that could suggest inhibition of the dominant language when the non-dominant language is used, that must then be overcome to return to the dominant language (Meuter & Allport, 1999; for review see Declerck & Philipp, 2015b). These effects are most often attributed to inhibitory control mechanisms, specifically those that might also support non-linguistic task-switching (Green, 1998; Meuter & Allport, 1999; see also Philipp, et al., 2007; Philipp & Koch, 2009; but see Bobb & Wodniecka, 2013). Consistent with strong inhibition of the dominant language in mixed language contexts, under a limited set of conditions fully reversed language dominance effects – in which bilinguals perform less well in their dominant than in the otherwise non-dominant language — have been observed. Gollan et al. (2014) reported that intrusion errors exhibited such reversed dominance effects; words in the dominant language were replaced by words in the non-dominant language more often than vice versa. Reversed dominance effects have also been found in a handful of studies that measured timed picture naming, with both cued (Christoffels, Firk, & Schiller, 2007; Costa & Santesteban, 2004; Verhoef et al., 2009), and voluntary, language switches (Gollan & Ferreira, 2009).

The Current Study

In the current study we investigated the role of these factors in controlling intended and unintended language switches using the read-aloud task. Both previous studies that used the read-aloud task contained paragraphs with haphazard language switches (Gollan et al., 2014; Kolers, 1966). Here, we asked if bilinguals would still produce intrusion errors when switches followed more typically occurring patterns of bilingual language use. Spanish-English bilinguals read paragraphs that manipulated the degree to which switches conformed to naturally occurring patterns of language switching. We hypothesized that grammatical switches might allow bilinguals to rely on automatic control mechanisms that function in conversational speech between bilingual interlocutors, and therefore might elicit more rapid and also more accurate production of mixed language speech. Because previous evidence concerning the role of grammaticality on switching produced mixed results, we also manipulated the rate of switches, crossing grammaticality with switch rate so that paragraphs either had mostly grammatical or mostly ungrammatical switches, and either a low or a high switch rate (with twice as many switches in high as in low). To provide a further index of grammatical influences on switching, we examined part of speech effects. Based on previous work (Gollan et al., 2014; Kolers, 1966; Poulisse, 1999), we anticipated we would observe greater control difficulties on function vs. content words – but of particular interest here – was whether grammatical encoding would affect function more than content word switches, or perhaps function but not content word switches.

Importantly, though the read-aloud task elicits production via reading, and therefore aspects of the task will at least partially reflect mechanisms underlying reading comprehension, intrusion errors produced in this task primarily (if not exclusively) reflect the mechanisms of speech production (Gollan et al., 2014). Indeed Kolers (1966) found that language mixed text had no effect on comprehension measures, but robust effects on production times and errors produced in the read-aloud task. Several aspects of the errors produced in the read-aloud task supported this conclusion. First, most intrusions were produced with targets that do not share sounds or letters with their translation equivalent targets (e.g., pero—but). The production of intrusion errors even in the presence of substantial formal differences between translations, suggests that bilinguals are not simply mis-reading the paragraphs when they produce these errors. Most importantly, eye-tracking measures in Gollan et al. (2014; see Table 4–5, Figure 3) revealed that although bilinguals skipped function more than content words when reading mixed-language paragraphs, in the majority of cases bilinguals were fixating the target word when they produced the intrusion error, and part of speech effects on intrusion errors remained even when limiting the analysis to targets that were never skipped, again suggesting independence between mechanisms underlying reading processes and production of intrusion errors in the read-aloud task. Finally, if the read-aloud task elicited intrusion errors via different processing mechanisms than those that underlie intrusions in more naturalistic speech, then it might be expected that the read-aloud task would elicit a greater rate of intrusions than found in spontaneous speech. However, intrusions were produced in the read-aloud task approximately or just under 1% of the time, which is the same rate as reported for spontaneous speech (Poulisse, 1999), and in other experimental tasks that measured production more purely (e.g., the verbal fluency task; Gollan et al., 2011).

By-participant mean intrusion error rate (at switch sites) and mean within-language error rate (all positions), English dominant participants. Error bars depict bootstrapped 95% confidence intervals for means.

The read-aloud task appears to elicit speech production errors via the same, or similar processes that give rise to errors in spontaneous speech production. Specifically, we assume that multiple representations are activated in parallel during formulation of the planned utterance. Errors arise because random noise during processing allows non-target representations to become more active than the target (e.g., Dell, 1986). In this task, the code mixed utterance presented for reading is the most strongly activated target representation, along with co-activated representational elements from alternative code mixed utterances (e.g., including translation equivalents of the target words). When the target is weakly activated—e.g., ungrammatical, haphazard mixing sentences that do not fit with the bilingual speaker’s experience—random noise is more likely to give rise to the production of alternative utterances (leading to intrusion errors). Similar phenomena have been observed in monolingual language production, in which phonological structures that violate constraints on sound structure are more likely to result in speech errors than those that respect such constraints (see Goldrick, 2011, for a review).

While such mechanisms provide one means of generating intrusion errors within production processes, other aspects of eye-tracking measures reported in Gollan et al. (2014; Figure 3) did suggest that function words are more vulnerable to failures of language control at least in part because they draw less attention than content words (e.g., function, but not content words, were more likely to involve an intrusion error if bilinguals were looking at words in the nontarget language when they produced the error). To address the possible contribution of reading processes to the results in the present study, we included part of speech as a factor in our analyses; since readers skip function words more than content words – part of speech effects should be differentially sensitive to grammaticality manipulations if these are driven by reading rather than by production mechanisms.

In parallel to analysis of these language-specific factors—which we assumed to be reflecting relatively automatic control processes specialized for language processing — we examined the extent to which language selection might also be driven by more attention demanding control processes subject to individual differences in domain general control abilities. To this end, we measured individual differences in three commonly administered measures of executive control. We hypothesized that bilinguals with stronger general cognitive control ability might also exhibit more efficient ability to mix languages in the read-aloud task. Of particular interest, however, was if individual difference measures would interact with our manipulation of grammaticality, as well as part of speech and language dominance effects. The presence of such interactions would suggest that bilingual language control mechanisms are fully integrated with, and broadly reliant on, domain general control mechanisms, whereas their absence would imply the existence of some language specific control mechanisms that function relatively independently. In Experiment 1 we tested young Spanish-English bilinguals, and in Experiment 2 we tested a group of older Spanish-English bilinguals on the same tasks that were administered in Experiment 1, to provide a second test of the individual differences question. We hypothesized that interactions between executive control measures and grammaticality might be most likely to emerge in older bilinguals who exhibit a broader range of individual differences in executive control ability, including some individuals with relatively impaired executive control ability.

Experiment 1 – Young Bilinguals

Method

Participants

Ninety-seven Spanish-English bilinguals participated in the study in exchange for course credit through undergraduate classes in the Psychology department at the University of California, San Diego (UCSD). Table 1 shows self-reported participant characteristics and ability to name pictures in each language on the Multilingual Naming Test or MINT (see below; Gollan, Weissberger, Runnqvist, Montoya, & Cera, 2012).

Table 1.

Participant characteristics.

	Younger bilinguals (n = 97^a)		Older bilinguals (n = 20)		Matched Younger bilinguals (n = 20)		Comparison of Older and Matched Younger
	M	SD	M	SD	M	SD	t	p
Age	20.3	2.0	75.7	8.8	20.2	1.2	106.19	<.01
Dementia Rating Scale score	--	--	137.8	3.4	--	--	--	--
Mini Mental Status Exam score	--	--	28.6	1.8	--	--	--	--
Education	13.8	1.2	14.3	2.7	14.1	1.0	<1	.82
Age of English Acquisition	3.1	2.5	4.8	3.7	3.9	2.5	<1	.35
English proficiency self-rated^b	6.5	0.8	6.5	0.6	6.4	0.7	<1	.68
Spanish proficiency self-rated^b	5.8	0.9	5.5	1.3	5.9	0.7	<1	.28
Current percent English use	80.9	15.6	77.3	22.9	86.9	9.5	1.40	.09
Percent English use during childhood	56.7	17.9	52.4	31.3	62.0	11.9	<1	.21
English MINT score^c	60.4	3.8	63.1	4.0	63.0	2.1	1.23	.88
Spanish MINT score^c	45.8	10.2	52.4	7.8	52.9	5.9	<1	.82
Trail Making Test^d
Trails A	26.1	7.0	42.0	12.7	26.1	5.6	5.76	<.01
Trails B	54.5	15.2	122.2	53.3	55.6	14.2	6.20	<.01
Trails B errors	0.4	0.7	1.1	1.4	0.4	0.8	1.42	.09
B minus A	28.4	13.5	80.3	49.9	29.5	11.9	4.75	<.01
B/A	2.2	0.6	3.0	1.2	2.2	0.5	2.69	.01
Flanker Task^e
Congruent	527	165	801	202	515	191	4.99	<.01
Neutral	510	160	730	170	493	173	4.68	<.01
Incongruent	571	185	897	276	540	186	5.28	<.01
Flanker Effect^f	61	64	167	169	48	24	3.06	<.01
Flanker Measure^f	12.1	10.2	22	21	10	5	2.52	.02
Stroop^g
Congruent	732	105	884	160	751	108	3.00	<.01
Neutral	756	104	953	147	779	112	4.45	<.01
Incongruent	852	141	1119	180	884	175	4.41	<.01
Stroop Effect^f	97	68	166	93	105	86	1.89	.04
Stroop Measure^f	12.7	8.4	18	11	13	10	1.48	.15

Open in a new tab

For executive function measures, means exclude the two individuals that were excluded from analysis due to failure to complete executive function tasks.

Self-rated proficiency level was averaged across ratings for speaking, comprehension of spoken speech, reading, and writing on a scale from 1 (“little to no knowledge”) to 7 (“like a native speaker”).

Maximum possible score is 68.

Error rates were extremely low for both young and older bilinguals and there were no significant aging effects in any condition

Error rates were extremely low (<1%) in all conditions except for older adults in the incongruent condition (2%).

Flanker and Stroop effects were calculated by subtracting neutral from incongruent times. Flanker and Stroop measures were calculated as follows (incongruent minus neutral)/neutral

Stroop performance is reported for the language with the fastest incongruent trial reaction times. Error rates were very low in all conditions except for the incongruent condition in which both young and older bilinguals produced 5% errors. Aging effects on error rates were not significant.

Materials and procedure

Stimuli were presented using PsyScope X software (Build 57; Cohen, MacWhinney, Flatt, & Provost, 1993; http://psy.ck.sissa.it) on an iMac 7 computer with a 20-inch color monitor. On each trial an entire paragraph appeared centered on the screen and participants were instructed to read the paragraph aloud as accurately as possible at a comfortable pace. Each bilingual read 16 paragraphs, four in each of four conditions: (a) grammatical, low-switch; (b) grammatical, high-switch; (c) ungrammatical, low-switch; (d) ungrammatical, high-switch. On average, paragraphs had 109 (SD=10) words (range 88–134 words), and on average 72% (SD=3%) were in English (and the rest in Spanish).

Paragraphs were modified from Gollan et al., (2014) so that all were presented in English word order, and varied with respect to switch rate and grammaticality of switches. An example of each paragraph is presented in the Appendix. Low switch paragraphs had 13 language switches in them, and high switch paragraphs had twice as many (i.e., 26 switches). A native Spanish-English bilingual adapted the paragraphs (which were originally composed to contain “haphazard” language switches) to manipulate switch rate and grammaticality without following any specific rules about what types of switches are or are not natural. A second native Spanish-English bilingual read through the paragraphs and confirmed the grammaticality manipulation (any disagreements were discussed and settled and paragraphs modified accordingly). Note that classifying switches as grammatical or not is a topic of some debate in the field (see e.g., Bullock & Toribio, 2009, for discussion). Our manipulation here was deliberately not restricted to one type of constraint on switching, and was intended as a first step towards examining the role of naturalness in facilitating control over language selection in production of connected speech.

However, to consider the extent to which the grammaticality manipulation might generally match theoretical constraints on switching we compared grammatical and ungrammatical paragraphs in frequency of occurrence of three constraints. First, we asked if ungrammatical paragraphs violate the Equivalence Constraint (Poplack, 1980), a structural constraint on language switching. Switches are grammatical, satisfying this constraint, if the word order matches across languages (e.g., a switch on a noun in a phrase like the man would be more acceptable than a switch in a the old man because the Spanish translation equivalents would appear in the same order el hombre for the former but not in the latter el hombre viejo which word for word translates to the man old). To consider this possibility we identified all phrases in the paragraphs with a Determiner Modifier Noun structure (e.g., an old car, an old carro), and then counted the number of switches on the noun. We excluded those cases with pronominal adjectives in Spanish (e.g., misma) along with their translation equivalents. We also excluded two examples where Spanish words were in the incorrect word order (e.g, su fresca sombra). These counts revealed that violations of the equivalence constraint occurred just 3 times in grammatical paragraphs (the total number of opportunities for switches in such constructions was n=47, for a violation rate of 6.4%), but 9 times in ungrammatical paragraphs (with n=45 and a violation rate of 20%); thus, ungrammatical paragraphs violated this particular equivalence constraint three times as often as grammatical paragraphs.

Though these numbers are relatively small, each subsequent comparison revealed a similar pattern. Another commonly accepted constraint on switching was mentioned above; this is that switches on a single function word are rare (Muysken, 2000; though, paradoxically, these are the most common targets of naturally occurring intrusion errors, Poulisse, 1999). To consider if such switches occurred more often in ungrammatical than in grammatical paragraphs, we counted the number of single function word switches (e.g., Spanish word, English function word, Spanish word). About 1.9% of all function word switches in grammatical paragraphs fit this profile (9/483 switches on function words), and more than twice as often, or 4.2% in (13/309) in ungrammatical paragraphs. Finally, it is well known that switches on single isolated words most commonly involve nouns (Myers-Scotton, 2002; Muysken, 2000), and much less often involves verbs (which may be switched within a sentence less often because they are less easily integrated into syntactic structure across different spoken languages; Myers-Scotton & Jake, 1995; but see Emmorey, Borinstein, Thompson, & Gollan, 2008, for contrasting patterns in speech-sign bilinguals). In ungrammatical paragraphs singleton switches on verbs (n=16/111 total verb switches, 14.4%) outnumbered singleton noun switches (n =10/173, 5.8%), whereas in grammatical paragraphs singleton switches on nouns (n=27/65, 41.5%) outnumbered singleton switches on verbs (n =15/63, 23.8%), as would be expected in naturally occurring code-switches. Thus, though grammatical paragraphs had some switches that violated constraints on switching proposed in the literature, for each and every comparison considered, there was a greater number of violations in ungrammatical than in grammatical paragraphs. Thus, the intuitions of the bilinguals who designed the paragraphs for the present study, generally match what has been reported in the literature for characterizing commonly occurring switches.

Paragraphs appeared in one of four different fixed orders with conditions counterbalanced across subjects and items in a Latin square design. Within each fixed order, all four conditions were presented on each group of four consecutive trials but in different condition orders across subjects so that the average list position of paragraphs was the same for every condition, and participants never saw paragraphs in the same condition on consecutive trials. Each trial began with a fixation point (+) that appeared at the location on the screen where the first word appear and remained on the screen until the participant pressed the space bar, after which the fixation was replaced by the paragraph which remained on the screen until the participant pressed the space bar again. A 1,500 ms ITI was presented and then replaced by the fixation point for the following trial. Before beginning bilinguals completed 4 shorter practice paragraphs (approximately 30 words in each) including one of each type of paragraph (counterbalanced in the same way as the critical paragraphs).

To provide an objective measure of English and Spanish proficiency, all bilinguals completed the Multilingual Naming Test (MINT; Gollan et al., 2012). This consists of 68 black-and-white line drawings, administered in order of progressing difficulty (e.g., item #1 is hand, and item # 68 is axle). This test was designed to assess picture-naming ability in four languages (English, Spanish, Mandarin, Hebrew), and also both dominant and non-dominant language proficiency.

A number of tests of executive function were also administered to provide an objective measure of domain-general executive control ability. Choice of tests was motivated by previous reports of an association between bilingualism and performance on these measures (for review see Bialystok et al., 2009). Means (and standard deviations) by condition for these tasks are shown in Table 1, which also shows that all three tests exhibited significant interference effects typically reported for these measures.

Stroop (Stroop, 1935)

Target stimuli consisted of four color names (black, blue, green, and red in English; negro, azul, verde, and rojo in Spanish) and four neutral words (big, prime, deep, and legal in English; boda, dulce, mesa, and usar in Spanish). Each bilingual completed the task in both languages in counterbalanced order between participants as described below. The use of neutral words as a baseline and unblocked presentation of trial types (congruent, neutral, incongruent; see below) was based on the recommendations of Spieler, et al. (1996). There were 36 congruent trials, 36 incongruent trials, and 32 neutral trials. In the congruent trials, each of the four color names appeared nine times in its corresponding color. In the incongruent trials, each color name appeared three times in each of the three non-matching colors. In the neutral trials, each of the four neutral words appeared twice in each of the four colors. Sixteen randomized lists were created for each language, with the restriction that a word or color was not repeated more than twice on consecutive trials. Participants were instructed to produce the name of the color of ink in which the words were presented. On each trial, the target stimulus was presented until a response was recorded. The interval between trials was 500 ms. Before the critical trials, participants completed 16 practice trials in the same language as the critical trials. Participant responses were coded for accuracy online by an experimenter, and were also audio-taped for subsequent verification of accuracy.

Trail Making Test A and B

(TMT A and TMT B; from the Halstead Reitan Neuropsychological Test Battery; see Reitan, 1958; cf. Mickes et al., 2007): In Part A (TMT A), participants draw a line to connect the numbers 1–25 in consecutive order as quickly as possible within a 150 second time-limit. In Part B (TMT B), participants draw a line to connect 25 numbers and letters in alternating, consecutive order as quickly as possible within a 300 second time-limit. Time-to-complete each task is scored.

Flanker task

The flanker task was adapted from Fan et al. (2002). Stimuli consisted in a central arrow with two flankers on each side, and appeared in three conditions: congruent (central arrow pointing in the same direction as the flanker), incongruent (central arrow pointing in the opposite direction of the flanker), and neutral (central arrow flanked by lines without arrowheads). Direction of arrows was counterbalanced in each condition. There were 16 trials in each condition. Trials began with a central fixation point for 900 ms, then the target stimulus replaced the fixation point until a response was made, and there was an ITI of 1000 ms. Participants pressed a button on the left or right of a button box to indicate the direction of the central arrow. Practice blocks began with six neutral trials, followed by six congruent trials, then six incongruent trials, and ended with six trials with equal numbers of the different trial types presented in a random order. Feedback “correct” or “incorrect” was included for practice trials only.

The paragraph reading task was completed first after which bilinguals completed the remaining tasks in a counterbalanced order. Some bilinguals completed English MINT and English Stroop next or Spanish MINT and Spanish Stroop next (counterbalanced between participants). The two different language versions of MINT and Stroop were always separated in time by the Flanker Task or by the Trail Making Test (see below). An equal number of bilinguals completed Flanker or Trails first (counterbalanced between participants), followed by English or Spanish Stroop and MINT separated by whichever executive task they had not yet completed.

Results

Following the methods of Gollan et al., (2014) a native Spanish-English bilingual research assistant transcribed the errors which were classified as intrusions (n = 899) e.g., saying el instead of he), partial intrusions (n = 208; starting to produce an intrusion but self-correcting before producing the error), accent errors (e.g., n = 844; saying the correct word with the accent of the non-target language), or within-language errors (n = 2,271; e.g., saying such instead of much). All but one bilingual produced at least one intrusion and up to as many as 24 (M=9; SD=5). Accent errors are necessarily subjective (see also Kolers, 1966), and other error types are difficult to interpret given indeterminacy with respect to error content (i.e., omissions, n = 291 and insertions, n = 222); thus we focused our analyses primarily on intrusions and within-language errors. A small number of accent errors (n = 5) and intrusion errors (n = 6) were classified as more than one type of error (e.g., accent and within, or intrusion and accent). These cases were classified by default as intrusions rather than as within-language errors, with only one type of language control failure coded if both intrusion and accent error were produced. For example, if the intended target was monastery and the speaker said monasterio and then self-corrected to monastery but produced the English word with a Spanish accent, these would be counted as a single intrusion error.

Partial intrusions, i.e., intrusion errors self-corrected in mid-utterance, exhibited a noticeably different pattern from other intrusion errors. As discussed below, intrusions in this paradigm (as in naturally occurring errors) are more likely to target function than content words. Partial intrusion errors exhibit the opposite pattern. At switch sites (the primary location of intrusion errors; see below), partial intrusions occurred on 1.2% of content words (95% CI 0.9%, 1.4%), significantly higher than the rate of 0.1% of function words (95% CI 0.008%, 0.2%). This likely reflects the relatively difficulty of monitoring for errors on function vs. content words (Gollan et al., 2014; Poulisse & Bongaerts, 1994). Given this clear difference in sensitivity to word class across the two error types, we excluded partial intrusions from our analysis of intrusion errors.

Analysis methods

Behavioral data were analyzed using mixed effects regressions (linear for reading times and logistic for error rates; Jaeger, 2008). All categorical predictors (e.g., switching type, grammatical vs. ungrammatical) were contrast coded, and continuous predictors (e.g., Stroop interference) were centered. Random intercepts were included for participants; more complex random effects structures failed to converge for some analyses, so this simplified random effects structure was used in all cases. Significance was assessed via model comparisons (Barr, Levy, Scheepers, & Tily, 2013).

Whole Paragraph Reading times

We first examined how whole paragraph reading times were affected by switching rate (low vs. high), and paragraph type (grammatical vs. ungrammatical), along with the interactions of these factors. As shown in Figure 1, bilinguals read paragraphs with ungrammatical switches more slowly than those with grammatical language switches (β = 3.75, SE β = 0.26, χ²(1) = 199.67, p < .0001), and paragraphs with a high switch rate more slowly than those with a low switch rate (β = −2.16, SE β = 0.26, χ²(1) = 70.08, p < .0001). Reading times were especially slow in ungrammatical paragraphs with a high switching rate; i.e., switching rate and paragraph type interacted (β = −1.81, SE β = 0.51, χ²(1) = 12.53, p < .0005). Follow-up regressions at each switching rate showed that while grammaticality had a significant effect on readings time in both types of paragraphs, grammaticality effects were stronger at high switching rates (β = 4.64, SE β = 0.34, χ²(1) = 159.44, p < .0001) than at low switch rates (β = 2.83, SE β = 0.35, χ²(1) =60.77, p < .0001). Additional, follow-up regressions within each paragraph type showed that while there were switch costs on readings times in both types of paragraphs (i.e., a significant effect of switch rate), switch costs were stronger in ungrammatical (β = −3.07, SE β = 0.35, χ²(1) = 72.21, p < .0001) than in grammatical paragraphs (β = −1.26, SE β = 0.34, χ²(1) =13.94, p < .0005).

Intrusion Errors

Bilinguals were much more likely to produce intrusion errors at switch sites (mean: 1.4%, 95% CI 1.2%, 1.6%) than at non-switch positions in the paragraphs (mean: 0.3%, 95% CI 0.3%, 0.4%). We therefore focused our analysis at switch points, analyzing the likelihood of intrusion errors with a logistic mixed effects regression. In addition to the experimental factors (grammaticality and switch rate), we examined if intrusions were modulated by part of speech, and allowed part of speech to interact with grammaticality (to examine if grammaticality effects were driven exclusively or primarily by function vs. content words). A random intercept was included for participants.

As shown in Figure 2, bilinguals produced more intrusion errors when reading paragraphs with ungrammatical than with grammatical switches (β = 0.66, SE β = 0.12, χ²(1) = 30.6, p < .0001). While there was no main effect of switch rate (β = 0.14, SE β = 0.11, χ²(1) = 1.62, p < .21), there was a significant interaction between rate and grammaticality (β = −0.74, SE β = 0.21, χ²(1) = 11.5, p < .001). Follow-up regressions at each switching rate revealed grammaticality effects on intrusions only in paragraphs with high switch rates (β = 1.13, SE β = 0.16, χ²(1) = 61.1, p < .0001), whereas in paragraphs with low switch rates grammaticality effects were not significant (β = 0.14, SE β = 0.19, χ²(1) = 0.58, p < .45). Additional follow-up regressions within each paragraph type revealed switch costs in ungrammatical paragraphs, such that bilinguals produced more intrusion errors in high than in low switch paragraphs (β = −0.26, SE β = 0.13, χ²(1) = 3.89, p < .05), whereas in grammatical paragraphs, switch costs were absent, in fact there was a significant effect in the unexpected direction (i.e., more intrusion errors in low than in high switch paragraphs; β = 0.51, SE β = 0.17, χ²(1) = 8.53, p < .005). We do not interpret this effect any further, given the presence of a significant effect in the opposite direction in the reading times (see above), possibly suggesting a speed-accuracy trade-off.

By-participant mean intrusion error rate at switch sites within each experimental condition, Experiment 1. Error bars depict bootstrapped 95% confidence intervals for means.

There was a nonsignificant trend for bilinguals to produce more errors on function words (mean 1.5%, 95% CI 1.2%, 1.7%) compared to content words (mean 1.3%, 95% CI 1.1%, 1.6%; β = 0.20, SE β = 0.12, χ²(1) = 2.82, p < .10). The absence of robust part of speech effects on intrusion errors resembles results reported for English word order paragraphs as used here and in Gollan et al., (2014), and contrasts notably with highly robust part of speech effects on within-language errors (see below). Crucially, the effect of grammaticality was not significantly stronger for function vs. content words (β = 0.20, SE β = 0.12, χ²(1) = 2.82, p < .10¹).

To examine if intrusion errors were modulated by the language of production, we repeated the analysis above, focusing on the 88 English-dominant speakers (those with higher English than Spanish MINT scores). Language of production was included as an additional predictor, interacting with part of speech (to examine if dominance effects were driven by function vs. content words). As shown in Figure 3, there was a significant reversed-dominance effect; bilinguals were more likely to produce intrusions in the dominant language (replacing written English words with spoken Spanish translations), than in the non-dominant language (replacing written Spanish words with spoken English translations; β = 0.41, SE β = 0.11, χ²(1) = 13.7, p < .0005). Interestingly (and contrasting with a result reported below for within-language errors), this did not interact with part of speech (β = 0.22, SE β = 0.22, χ²(1) = 0.98, p <.35).

Though our primary focus was on language switching ability and language intrusion errors, as a point of contrast we repeated our analyses of intrusion errors with errors produced within a single language as the dependent measure. To the extent that our manipulations targeted language control specifically, they should have no effect on within-language errors.

Within-Language Errors

Within-languages errors were equally likely to occur at switch sites (mean: 1.3%, 95% CI 1.1%, 1.6%) and other positions in paragraphs (mean: 1.3%, 95% CI 1.2%, 1.5%). We therefore analyzed errors at all points. Regression analyses were structured following those above.

We first examined the experimental factors (grammaticality and rate) and part of speech. As shown in Figure 4, bilinguals were no more likely to produce within-language errors when reading paragraphs with ungrammatical than with grammatical switches (β = −0.02, SE β = 0.04, χ²(1) = .24, p < .65). There was a non-significant trend for more errors to occur at high switch rates (β = 0.07, SE β = 0.04, χ²(1) = 2.83, p < .10). In contrast with the results reported above for intrusion errors, there was no significant interaction between rate and grammaticality (β = 0.11, SE β = 0.08, χ²(1) = 1.58, p < .21). Also in contrast with results reported for intrusions, bilinguals produced within-language errors much less frequently with function word targets (mean 0.9%, 95% CI 0.7%, 1.0%) compared to content word targets (mean 2.1%, 95% CI 1.8%, 2.4%; β = −0.92, SE β = 0.04, χ²(1) = 458.39, p < .0001). However, as for intrusion errors, there was no interaction of part of speech with grammaticality (β = −0.09, SE β = 0.09, χ²(1) = 1.02, p <.32).

Mean within-language error rate (across participants) within each experimental condition, Experiment 1. Error bars depict bootstrapped 95% confidence intervals for means.

We then repeated the analysis above for the 88 English-dominant speakers, including language of production as an additional predictor, interacting with part of speech. As shown in Figure 3, there was a significant dominance effect; bilinguals were less likely to produce within-language errors in the dominant language (English), than in the non-dominant language (Spanish; β = −0.92, SE β = 0.05, χ²(1) = 338.62, p < .0001). Additionally, unlike results reported above for intrusions, language dominance effects on within-language errors were stronger for content than for function words, i.e., dominance interacted with part of speech (β = 0.28, SE β = 0.09, χ²(1) = 9.54, p < .005). Follow-up regressions showed that while there was a significant dominance effect for both parts of speech, the effect was stronger for content (β = −1.03, SE β = 0.06, χ²(1) = 315.28, p < .0001) as compared to function words (β = −0.74, SE β = 0.07, χ²(1) = 97.32, p < .0001).

To further investigate these effects, we re-ran these regressions, incorporating a continuous measure of each participant’s language dominance (the ratio of the English to Spanish MINT scores), allowing this to interact with the factors above. The continuous dominance measure not only modulated the overall effect of language of production (such that the tendency for within-language errors to occur on Spanish vs. English targets was stronger for more English-dominant participants; β = −0.73, SE β = 0.11, χ²(1) = 49.46, p < .0001); it also modulated the interaction between language of production and part of speech (β= 0.56, SE β = 0.21, χ²(1) = 7.36, p < .01). Follow-up regressions revealed that the general effect of the dominance measures—the tendency for more English-dominant participants to produce more within-language errors on Spanish vs. English targets—was much stronger for content (β = −1.03, SE β = 0.14, χ²(1) =63.39, p < .0001) vs. function words (β= −0.42, SE β = 0.15, χ²(1) = 7.59, p < .01).

Executive function effects

We next examined the extent to which individuals with relatively poor performance on executive function tasks had greater difficulties with language control. In particular, we were interested if relatively poor executive function abilities enhanced the effects observed above; such interactions would support the integration of language-specific and domain general control mechanisms.

To address these issues, we constructed three separate new regression models, with each one including one of three executive function measures:

Stroop: The language with the faster mean reaction times (RTs) on neutral trials was used to calculate each participant’s degree of Stroop interference, relative to neutral trial RT: (mean RT on inconsistent trials – mean RT on neutral trials)/mean RT on neutral trials)
Flanker: Similar to Stroop, this measure was interference relative to the neutral trial baseline: (mean RT on inconsistent trials – mean RT on neutral trials)/mean RT on neutral trials)
Trails: Increase in completion time for Part B relative to Part A: Trail Marking Test Part B completion time/Trail Marking Test Part A completion time

Each (centered) executive function measure was included as a main effect, allowing us to examine its contributions to language control. To examine whether language control mechanisms are fully integrated with domain-general mechanisms, we also allowed each measure to interact with the significant effects observed above: grammaticality and its interaction with rate. Two participants were excluded from these analyses for failure to complete one of the executive function tasks. The results of the regressions for the 95 remaining bilinguals are summarized in Table 2.

Table 2.

Summary of regressions examining effects of individual differences in executive function in younger bilinguals. Significant effects are shown in bold; the second column summarizes the predicted effect in words (if observed) and the cells to the right provide coefficient values (standard errors in parentheses; *: p < 0.05, **: p < 0.005, ***: p< 0.0005). Non-bold cells give coefficients with 0.15 > ps > 0.05. Cells with dashes have coefficients with ps > 0.15.

		Stroop	Flanker	Trails
Reading Times	Main Effect: Longer reading times (RTs)	--	--	2.03 (1.13)
	Interaction with Type: Reading Time Even Longer on Ungrammatical	10.96 (3.1)***	−4.63 (2.54)	1.79 (0.43)***
	Interaction with Type, Rate	--	--	--

Intrusion Errors	Main Effect: Increased intrusions	--	--	0.31 (0.12)*
	Interaction with Type	--	--	--
	Interaction with Type, Rate	--	--	--

Within Errors	Main Effect	--	--	--

Open in a new tab

If language control relies on domain-general mechanisms, individuals with lower executive function performance should show a greater rate of intrusion errors. There is some evidence consistent with this. While there were no main effects of either Stroop or Flanker interference, individuals with relatively worse performance on Trails produced more intrusion errors. To consider if this effect was specific to language control, we examined whether any of these executive function measures were correlated with the rate of within-language errors; as shown in Table 2, there were no significant effects.

If language control mechanisms are fully integrated with domain-general control mechanisms, we expect these differences in executive function to interact with grammaticality. There is some evidence consistent with this. Individuals with greater Stroop and Trails interference effects showed exaggerated effects of grammaticality on whole paragraph reading times. However, there were no such interactions in our analyses of intrusion errors. Furthermore, there was a marginal interaction in the unexpected direction for Flanker performance (such that individuals with greater Flanker interface showed decreased effects of grammaticality).

Discussion

The results of Experiment 1 revealed a number of key findings. First, grammaticality of switches had robust effects on both reading times and production of intrusion errors; this implies that grammatical encoding plays an important role in helping bilinguals maintain control over language selection. We also observed robust switch rate effects, so that switch costs increased with switch rate, and interactions with grammaticality, so that bilinguals found it particularly difficult to produce larger numbers of switches if those switches were also ungrammatical. Importantly, grammatical switches were not cost free; though intrusions did not reveal switch costs in grammatical paragraphs (they revealed an unexpected effect in the opposite direction), bilinguals’ whole paragraph reading times were significantly slower for high- than for low-switch paragraphs. Additionally, intrusion errors exhibited robust part of speech effects, consistent with theories that posit distinct control mechanisms for retrieval of targets that assign thematic roles versus those that do not (Myers-Scotton, 2006), the finding of robust part of speech effects in naturally occurring intrusion errors (Poulisse, 1999), and inconsistent with the absence of part of speech effects in other recently reported experimental paradigms (Declerck & Philipp, 2015a).

In contrast, within language errors exhibited no sensitivity to our language control manipulations, including grammaticality and switch rate. Also replicating previous work (Gollan et al., 2014), intrusions and within-language errors exhibited different part of speech and language dominance effects. Function word targets were equally likely to elicit intrusions as content words, but were far less likely to elicit within language errors relative to content words, and language dominance effects were significantly reversed for intrusions relative to within-language errors. Moreover, of great interest, in the current study, part of speech effects interacted with language dominance in our analysis of within-language errors (dominance effects were driven more by content than by function words), but not in our analysis of intrusion errors. These contrasting effects across different error types imply different underlying processing mechanisms, a point that will be elaborated in the General Discussion section.

Finally, our analyses of individual differences in executive control ability revealed that bilinguals who exhibited greater interference effects on the Trails measure produced a greater number of intrusion errors but no interaction with grammaticality implying independence of language-specific and domain-general control mechanisms. Whole paragraph reading times revealed some evidence for integration of control mechanisms; bilinguals who exhibited greater interference effects in both Trails and Stroop tests were particularly slow to read paragraphs with ungrammatical switches in them. However, we defer further discussion of these contrasting patterns to the General Discussion, after we present the results of Experiment 2.

Experiment 2 – Older Bilinguals

Our analyses in Experiment 1 revealed some evidence for language control mechanisms, related to grammatical encoding, that facilitate language switches in mixed-language connected speech, but some ambiguity as to whether they function independently of general executive control or not. In these analyses we assumed that there is a continuous relationship between domain general control and language switching. However, an alternative possibility is that language control does systematically rely on domain general control ability, but that only a minimal degree of executive control is recruited for this job of controlling language selection. On this view, a constant relationship between domains would be found regardless of individual differences in executive control, and evidence for interactions would be especially obvious when executive control ability dipped below the minimal level needed to serve language switching. Consistent with this possibility, some have suggested that bilingualism enhances executive control ability, but that this bilingual advantage is more apparent in older age, when individuals are no longer at peak levels of functioning in executive control ability (Bialystok et al., 2009; see Valian, 2015 for a comprehensive discussion; see Hilchey et al., 2014; Paap, et al., 2015 for a differing perspective). To test this possibility, in Experiment 2 we tested a group of older bilinguals with the same materials and procedure as in Experiment 1 – aiming to answer two key questions: (a) Do aging bilinguals exhibit significantly larger grammaticality effects, and (b) do aging bilinguals reveal stronger evidence for a relationship between domain-general executive control and language control?

We anticipated that older bilinguals would exhibit a broader range of abilities in executive function measures than in younger bilinguals. Although there is some disagreement as to the underlying cognitive mechanism (Salthouse, 1996, 2010; Verhaegen, 2011), it is widely reported that older individuals have more difficulty with measures of executive control ability, including the measures used herein. Consistent with this view, the prefrontal cortex, which is thought to control planning of complex behaviors, declines earlier, and more rapidly, than other parts of the brain as part of the normal aging process (see reviews in Raz, 2000; West, 1996). Additionally, aging related declines affect cognitive processes related to inhibition; e.g., older adults appeared to lose the ability to suppress activity in the ‘default mode network’ during word generation (Meinzer et al., 2012). Possibly related aging deficits are found in measures of selective attention, including Stroop and Flanker tasks (Spieler, Balota, & Faust, 1996; West & Baylis, 1998). According to the Inhibitory Deficit Hypothesis, activation of possible responses to stimulus cues remains intact, but aging leads to specific impairments in the ability to inhibit irrelevant information from attention and working memory, and difficulty with suppression of inappropriate responses (Hasher & Zacks, 1988; Zacks & Hasher, 1994).

The notion that aging leads to deficits in inhibitory control is not universally accepted, and one problem with trying to answer this question is that the construct of inhibition is defined differently by different investigators and in different subfields (Aron, 2007). Empirically, the hypothesis is difficult to test because there is disagreement as to which tasks constitute valid measures of inhibition (Burke & Osborne, 2007). In the context of language processing, attempts to test specific versions of the inhibitory deficit hypothesis have failed to confirm its predictions (Burke, 1997, 1999; Burke & MacKay, 1997). Experiment 2 aimed to contribute unique evidence to these long-standing questions by investigating aging effects on bilingual language control. This may offer a more ecologically valid measure of inhibitory control than those typically investigated. As discussed in the introduction, dominance reversal effects have been attributed to the use of inhibitory control in bilingual language processing (Gollan & Ferreria, 2009; Kroll, et al., 2008); assuming this analysis is correct, such effects may offer a more naturalistic and powerful way to measure age related decline in inhibitory control.

Thus, in Experiment 2 we hypothesized we might find (a) larger effects of grammaticality for older than for younger bilinguals. Several possible factors could give rise to such effects. For example, if grammaticality is related to the degree of experience with language switching, a longer lifetime would provide greater experience and expectations that switches should follow typical patterns of language use. Greater grammaticality effects might also be expected in aging bilinguals if reduced ability to rely on domain-general executive control to plan, produce, and monitor language switches) leads older bilinguals to rely more heavily on language specific control mechanisms. Alternatively, if typical switch patterns arise because some types of switches are motivated by syntactic structure, and therefore easier than others, older bilinguals have more difficulty with difficult tasks in a generalized way (but see Gollan et al., 2008). For similar reasons, older bilinguals might be expected to show larger part of speech and language dominance effects. Finally, we also predicted that (b) older bilinguals would show greater evidence for interaction between executive function measures and our experimental manipulation of grammaticality, on the assumption that such interactions become apparent only when minimal levels of executive control ability are absent.

Previous studies of aging and language control

Several studies have examined switch costs in out-of-context speech and revealed switching deficits in aging. For example, older bilinguals were more likely to fail to switch when cued to do so than young bilinguals (Hernandez & Kohnert, 1999). A more recent study confirmed the presence of larger cued language switching costs for older than for young bilinguals (Weissberger, Wierenga, Bondi, & Gollan, 2012), but also demonstrated that age-related slowing was far smaller in language switching than it was in a comparable measure of non-linguistic task-switching. Another related study used cued switching tasks configured a bit differently relative those in Weissberger et al. (2012), and found no aging related increase in language switching costs, but greater task-switching costs in older than in younger bilinguals (Calabria, Branzi, Marne, Hernández, & Costa, 2015). Though the precise pattern of results in the two studies was different, in both cases dissociations between aging effects across linguistic and non-linguistic domains were reported, and language control appeared to be relatively more intact than non-linguistic control in aging bilinguals. This in turn implies that at least partially different cognitive mechanisms support non-linguistic task-switching versus language-switching. Finally, very limited aging effects were found in another study, in which bilinguals were instructed to choose for themselves on each trial which language to use to name the pictures. If the “bilingual switch mechanism” were impaired in aging, then older bilinguals should have chosen to switch less often than young bilinguals. Instead, older bilinguals chose to switch voluntarily as often as young bilinguals, and though they responded more slowly overall than young bilinguals, switch costs were similarly sized across age groups (Gollan & Ferreira, 2009). This again implies relatively intact ability to produce intended language switches in aging bilinguals.

Another study examined unintended language switching in young and older bilinguals and produced mixed results (Gollan, Sandoval, & Salmon, 2011). In this study, bilinguals completed a verbal fluency task, on each trial producing as many members of semantic and phonemic categories (e.g., animals) as they could when given a minute for each category. Increased age was associated with a higher rate of intrusion errors, however, the total number of intrusion errors produced was very low even for older bilinguals who exhibited substantial difficulty with a non-linguistic control task. More recently, limited aging effects were again found in a verbal fluency task even though bilinguals had to switch languages on every trial, repeating the same category but in the other language (e.g., first they produced as many animal names as they could in English, and then they completed the same category in Spanish; Ivanova, Montoya, Murillo, & Gollan, in press). These studies again imply some limited aging effects, on ability to prevent unintended language switches. However, the verbal fluency task does not does not elicit production of full sentences, and elicits only a small number of intrusions, and therefore, leaves many open questions about how aging may or may not influence language control when switches are produced in connected speech, and with a greater number of observations of both intended and unintended language switches.