Current study

To summarize, there is evidence from separate studies on bilingualism and ageing for comparable disadvantages in language processing, but research on the combined effects of these two factors found no interaction between them. Further investigation of the combined impact of bilingualism and age in language processing will advance our understanding of how normal ageing affects bilingual language processing (e.g., Reifegerste, Reference Reifegerste2021). We therefore tested young and older monolingual and bilingual speakers in both comprehension (Experiment 1) and production (Experiment 2) of L1 sentences where we manipulated linguistic aspects that are thought to underlie disadvantages in language processing in bilinguals and older adults. Furthermore, we tested participants' L1, in order to assess bilingualism effects that are not confounded with language dominance. If bilingualism and ageing have comparable effects, we might expect them to interact, and to observe similar interactions between these two factors and the linguistic variables manipulated. Failing to find interactions between bilingualism and age would, on the other hand, suggest that these factors impact language processing in different ways, which should be further illustrated by different interactions between each of the factors and the linguistic variables.

Experiment 1: Listening comprehension

Experiment 1 investigated bilingualism- and age-related differences in the use of syntactic and semantic information during sentence comprehension. We employed a speech monitoring task where participants listened to a spoken sentence and made a speeded button press when they heard a pre-specified target word (spatula below). Spoken sentences could be lists of words in random order (RWO) (example a), low-constraint (example b), or high-constraint sentences (example c).

1. (a) Tried I find to quickly the spatula without pancake it to flip the breaking.

2. (b) I tried to quickly find the spatula to flip the pancake without breaking it.

3. (c) I flipped the pancake with the spatula without breaking it.

Shorter word monitoring response times (RTs) reflect easier lexical access due to expectations created based on the different types of linguistic representations that comprehenders develop, on a word-by-word basis. In (b), but not in (a), comprehenders can build coherent syntactic representations. And in (c), the semantic information constraints interpretation more than it does in (b). Therefore, differences between RWO and low- constraint conditions index the use of syntactic information, whereas differences between low- and high- constraint sentences index the use of semantic information.

Method

Participants

The data for this study were collected as part of a larger project that was publicly registered on OSF (see https://osf.io/d7aw2/). The current study reports analyses on samples of each group: 40 young English-speaking monolinguals, 40 young Norwegian–English bilinguals, 40 older English-speaking monolinguals, and 40 older Norwegian–English bilinguals (see Table 1 for the demographics). Young monolinguals (aged 18-35) and bilinguals (aged 19-30) did not differ significantly in age (Welch Two Sample t-test, t=−0.13, df = 58.02, p = 0.89), nor did older monolinguals (aged 65-81) and bilinguals (aged 66-80) (t = 0.72, df = 77.94, p = 0.48).

Table 1.

Demographic characteristics of participants in Experiments 1 and 2.

^a Proficiency level based on self-ratings using a scale of 0-10 with 0 being “none” and 10 being “perfect”.

The monolinguals reported themselves to be native speakers of British English, this being the only language they spoke at home. It was also an eligibility criterion that monolinguals should not be able to hold a simple conversation in any other language. Bilinguals completed an adaptation of the Language Experience and Proficiency Questionnaire (LEAP-Q; Marian et al., Reference Marian, Blumenfeld and Kaushanskaya2007), and participant selection was based on the following criteria: (i) Norwegian was the first acquired language; (ii) Norwegian was the dominant language; (iii) participants self-rated both their speaking and reading proficiency in English with at least 3 on a 0-10 scale with 0 being “none” and 10 being “perfect” (see Table 1 and Appendix 1).

Monolinguals were tested at the University of Birmingham, UK, and bilinguals at the University of Agder, Norway. The younger groups were students from these universities, and the older groups were recruited from the respective communities. Older participants were screened for mild cognitive impairment through the Montreal Cognitive Assessment questionnaire (MoCA, Nasreddine et al., Reference Nasreddine, Phillips, Bédirian, Charbonneau, Whitehead, Collin, Cummings and Chertkow2005), and all achieved a score >= 23 (Table 1), the cut-off point recommended by Carson et al. (Reference Carson, Leach and Murphy2018; Völter et al., Reference Völter, Fricke, Faour, Lueg, Nasreddine, Götze and Dawes2023; see Engedal et al., Reference Engedal, Gjøra, Benth, Wagle, Rønqvist and Selbæk2022, for normative data for Norwegian older adults). All participants had normal or corrected-to-normal vision and gave written informed consent. Our study was approved by the University of Birmingham's Institutional Ethics Board (ERN_20-1107), the Norwegian Center for research data (Ref. 239577), and the Committee for Medical and Healthcare Research Ethics in Norway (REK sør-østC, ref. 163931).

Language proficiency

To further control for the potential confound of language proficiency, we administered a reading comprehension task to all participants (mono- and bilinguals) in their L1, where they (self-paced) read sentences of different syntactic complexity followed by a comprehension question (See Appendix 2 for details). We restricted to trials with correct answers to the comprehension question (87.7%) and computed, for each participant, the average reading time across all sentence types, as a measure of proficiency capturing sentence parsing processes, which was entered as a co-variate in the analysis.

Design and materials

All participants completed the tasks in their L1. We created four different sets of stimuliFootnote ¹, two sets in British English and two sets in Norwegian BokmålFootnote ². Each set comprised 60 target word items embedded in sentences with low-constraining context, high-constraining context, or in the RWO condition (created by randomizing the order of words in the low-constraint sentence). To determine the extent of constraint of each sentence context, we conducted a pretest with a different set of participants completing a cloze task on 132 sentences, for each language, based on which we selected 120 items in each language, which were matched for length and mean cloze scores (see Appendix 4 for details).

For each set, each of the 60 items appeared in each of the three context conditions across three different lists (Latin Square Design). In addition to the experimental items, 12 fillers were created using different target words/ sentences. Therefore, each list contained 72 trials. These were divided into four blocks of 18 items (containing 15 experimental and 3 filler items). The order of trials within a block was pseudo-randomized so that no more than three trials of the same condition occurred after each other. Two versions of each list were created with a different order of presentation of the blocks. The final experiment therefore consisted of two unique sets with 6 lists each, per language. Participants were pseudo-randomly assigned to lists. The same four practice sentences were presented in the beginning of the experiment, that preceded the presentation of each individual list.

The sentences were recorded by female native speakers of Standard British English and Norwegian Bokmål in a soundproof booth using a high-quality USB microphone (Røde NTG) at a sampling rate of 44.1 kHz. The sound files were equalized for intensity, cut to the exact length of the sentences at zero crossings using Praat (Boersma & Weenink, Reference Boersma and Weenink2011), and were not further manipulated.

The design crossed two between-subjects variables (age group: old vs. young, and bilingualism group: bilingual vs. monolingual), and one within-subjects variable (context: RWO, low-constraint, high-constraint).

Procedure

The stimuli were presented using Presentation® (Version 20.1, Neurobehavioral Systems, Inc.). Spoken sentences were presented through speakers connected to the computer. Each trial began with a central fixation cross (‘+’) presented for 500ms, followed by a 1000ms blank screen and then the presentation of the target word, in the centre of the screen, for 1000ms (font ‘Consolas’, size 30). 500ms after target word offset, the spoken sentence started playing. Participants were instructed to pay attention to the written target word and to press the space bar on the keyboard as soon as they heard it in the spoken stimuli. Each trial ended 2000ms after the end of the audio file. Monitoring RTs were measured from the onset of each target word in the spoken stimulus. The task took around 15 minutes to complete.

Analysis

We first discarded RTs below 150ms or above 1500ms (3.18% of data; refer to OSF for an explanation of the trimming procedure and its impact on the analyses). To normalise the time responses and avoid the confound of potential differences in cognitive speed that do not reflect specific group differences in fundamental mental operations (such as, e.g., general slowing in older adults, but also potential cognitive speed differences between mono and bilinguals), we z-scored RTs, for each participant (Faust et al., Reference Faust, Balota, Spieler and Ferraro1999; each observation is subtracted from the mean and divided by the standard deviation of RT). We then excluded z-scores <-2.5 or >2.5 – that is, observations distancing more than 2.5SD from the mean – for each participant (2% of data).

For the analysis of (z-scored)RT, we ran linear mixed-effects models (LMMs; Baayen et al., Reference Baayen, Davidson and Bates2008), which allow for simultaneous estimation of between-participants and between-items variance. We used LMMs as implemented by blmer in R. blmer closely follows the function lmer, the primary distinction being that blmer allows the user to do Bayesian inference or penalized maximum likelihood, with priors imposed on the different model componentsFootnote ³, avoiding ‘singular fits’ in the models when using lmer (Chung et al., Reference Chung, Rabe-Hesketh, Dorie, Gelman and Liu2013). We fitted full models (all main effects and interactions) with a maximal-random structure when justified by the design (Barr et al., Reference Barr, Levy, Scheepers and Tily2013) and if convergence occurred. In particular, the random structure included intercepts by participant and item, but random slopes for each fixed-effect only by-participant (excluding Age Group, Bilingualism Group and Proficiency scores, as these are not justified) and not by item, as all of these resulted in non-convergence, or a worse fit, as assessed through Likelihood ratio testsFootnote ⁴ (refer to the summaries of models in the tables for its syntax). We report the predictors’ coefficients (β values), SE, t (or z) values, and the derived p significance values (by treating the t-statistic using the standard normal distribution as a reference; e.g., Baayen et al., Reference Baayen, Davidson and Bates2008, footnote 1).

We had participants (160) and items (240) as random effects in the analysis. For the fixed effects, we coded the context condition using forward difference coding, where each contrast compares adjacent levels (each level minus the next level): the first contrast is ‘RWO’ minus ‘low-constraint’ and the second one is ‘low-constraint’ minus ‘high-constraint’. Other fixed effects were Bilingualism group (bilingual vs. monolingual) and Age group (old vs. young), both contrast coded by centering (refer to summary of the model for coefficients). In addition, reading scores were added as a predictor to account for individual differences in Proficiency (Prof). Finally, Education level (see Table 1) was added as a confounding variable (with four categorical levels: ‘CompulsoryEducation’, ‘PostgraduateDegree’, ‘UndergraduateDegree’, and ‘UpperSecondary’; the first level was the reference level).

Results and discussion

Figure 1 presents the mean word monitoring RT (z-scored, refer to Appendix 6 for raw RTs), and Table 2 presents the fitted LMM model. Participants were faster to detect words in sentences with a low-constraint compared to random word lists (RWO). This effect was qualified by an interaction with bilingualism, indicating that the difference was larger in monolinguals than bilinguals. To explore this interaction, we conducted a follow-up LMM on the data from these context conditions where we created a four-level variable combining the two context levels (RWO vs. low-constraint) with the two bilingualism levels (bilingual vs. monolingual). We found that monolinguals were faster than bilinguals in low-constraint sentences (Bilingualxlow_Monolingualxlow; β = -0.13, SD = 0.06, p = 0.02), suggesting they were more able to use basic syntactic information to predict upcoming words. This result shows that the bilingual disadvantages in L2 syntactic processing (e.g., Pozzan & Trueswell, Reference Pozzan and Trueswell2016) are also observed for L1.

Figure 1.

Mean word monitoring RT (z-scored, ms) for the Younger and Older Monolingual and Bilingual groups, in the three experimental conditions. Error bars represent standard errors on means.

Table 2.

Summary of the linear mixed effects model fitted to the z-scored time to word monitoring (RT).

Note: The syntax of the model is: blmer(depM ~ 1 + RWO.lowConst + lowConst.highConst + Bilingualim + Prof + AgeGroup + HighestEdu + RWO.lowConst:AgeGroup + lowConst.highConst:AgeGroup + Bilingualim:AgeGroup #+ compProf:AgeGroup + RWO.lowConst:Bilingualim + lowConst.highConst:Bilingualim + Bilingualim:AgeGroup:Prof + RWO.lowConst:Bilingualim:AgeGroup + lowConst.highConst:Bilingualim:AgeGroup + RWO.lowConst:Prof:AgeGroup + lowConst.highConst:Prof:AgeGroup + RWO.lowConst:Bilingualim:Prof + lowConst.highConst:Bilingualim:Prof + Prof:Bilingualim:AgeGroup + RWO.lowConst:Bilingualim:Prof:AgeGroup + lowConst.highConst:Bilingualim:Prof:AgeGroup + + (1 | subj) + (1 | item) + (0 + RWO.lowConst | subj) + (0 + lowConst.highConst | subj), data = dataset, control = lmerControl(optimizer =“Nelder_Mead”))

Interestingly, bilinguals were faster than monolinguals in RWO (BilingualxRWO_MonolingualxRWO; β = 0.11, SD = 0.06, p = 0.05). This result adds to previous research that only contrasted low- and high-constraint sentences (e.g., Gollan et al., Reference Gollan, Slattery, Goldenberg, Van Assche, Duyck and Rayner2011Footnote ⁵). Our data suggest that bilinguals outperform monolinguals in detecting words that are embedded in linguistic material that does not allow listeners to construct a coherent syntactic or semantic analysis. In this case, comprehenders should benefit from ignoring the irrelevant linguistic stimuli of the random word lists, focussing attention only on the (pre-specified) target word. Although our study was not set to test this hypothesis, we suggest that bilinguals have an advantage in such circumstance, due to their improved language control (e.g., Bialystok & Craik, Reference Bialystok and Craik2022).

Comprehenders were also faster in high- compared to low-constraint sentences, and this effect was qualified by two-way interactions with bilingualism and with ageing. To follow up the first interaction, we had a four-level variable combining the context (high- and low-constraint) with the two bilingualism levels. Monolinguals were faster than bilinguals in low-constraint (as reported already above), but there were no group differences when high-constraint context was provided (Bilingualxhigh_Monolingualxhigh; p = 0.49). Thus, bilinguals are not at a disadvantage when they have contextual information to guide comprehension. This result parallels findings for L2 comprehension, where differences between monolinguals and bilinguals were apparent in low-constraint contexts, but not when semantic information could be used for prediction (e.g., Schwartz & Kroll, Reference Schwartz and Kroll2006; Shook et al., Reference Shook, Goldrick, Engstler and Marian2015). This is also consistent with the findings for L1 processing in Gollan et al.'s (Reference Gollan, Slattery, Goldenberg, Van Assche, Duyck and Rayner2011) Experiment 2 (but see their discussion on the possible trade-off between reading times and skipping rates).

Finally, we ran a similar four-level variable model contrasting age groups. No age differences were found in the low-constraint condition (youngerxlow_olderxlow; p = 0.16), supporting evidence for preserved syntactic processing with age (e.g., Tyler et al., Reference Tyler, Shafto, Randall, Wright, Marslen-Wilson and Stamatakis2010). Moreover, older adults tended to be faster than young to detect words in high-constraint contexts (youngerxhigh_olderxhigh; β = -0.05, SD = 0.03, p = 0.09). Although this effect did not reach significance, it would at least indicate the absence of a disadvantage in semantic processing in older adults, consistent with behavioural evidence for preserved semantic processing in old age. Tyler et al. (Reference Tyler, Shafto, Randall, Wright, Marslen-Wilson and Stamatakis2010) found no age-differences in the processing of sentences like ‘The church was broken into last night. Some thieves stole most of the LEAD off the roof’ (their Normal prose condition). Note that this type of sentence is similar to the ones we used in our low-constraint condition, where we too found no age-differences. In contrast, we had an additional condition where coherent semantic information was not only present, but also strongly constrained the interpretation (high-constraint condition). Our results suggest that, in these circumstances, there is a trend for older adults to outperform young.

Language proficiency, as measured by reading scores, did not influence word monitoring responses.

Experiment 2: Phrase production

Experiment 2 tested bilingualism- and age-related differences in a production task assessing planning scope of utterances. Participants described two pictures using phrase types that are known to have different planning scopes in young monolinguals, coordinate NPs (CNP; e.g., ‘The cone and the grape’) and NPs modified by prepositional phrases (PP; ‘The cone above the grape’), as indexed by longer production onsets on the first than on the latter (Allum & Wheeldon, Reference Allum and Wheeldon2007). We further manipulated complexity by having the second NP modified or not modified by an adjective (simple vs. complex; e.g., ‘The cone and the grape’ vs. ‘The cone and the pink grape’). Following previous studies, we take speech onset as a measure of the amount of preparation needed to start producing utterances, i.e., planning scope (e.g., Allum & Wheeldon, Reference Allum and Wheeldon2007; Konopka & Meyer, Reference Konopka and Meyer2014).

Method

Participants

The participants were the same as in Experiment 1 and completed the phrase production task after finishing the listening comprehension experiment.

Language proficiency

To control for language proficiency effects in the production experiment, we used scores from a vocabulary task. Participants were presented with words and asked to choose between four options: one correct answer which was either a synonym or antonym of the word, and three foils. We computed, for each participant, the percentage of accurate responses (number of correct answers divided by number of trials; see Appendix 3 for further details). Lexical knowledge should modulate picture naming (e.g., Gollan et al., Reference Gollan, Montoya, Fennema-Notestine and Morris2005), and we thus use the vocabulary scores as an additional predictor controlling for proficiency in the production experiment.

Design and materials

We adapted the multi-picture description task introduced by Smith and Wheeldon (Reference Smith and Wheeldon1999) such that participants were presented, on each trial, with a four-picture display. Each experimental item was a 600 x 600-pixel image containing four individual pictures. On each display, two of the pictures (target word-pair) were surrounded by a rectangular red line frame, indicating the phrase type to be produced. Coordinate phrases were cued by a horizontal red frame (that could appear at the top or the bottom; Figure 2 (a) and (b)), and prepositional phrases were cued by a vertical red frame (that could appear at the right or left; Figure 2 (c) and (d)). The second NP on each phrase type could also be simple or modified by an adjective. In the latter case, the target picture corresponding to the second noun appeared twice, once as the target modified picture and once as the original non-modified picture (Figure 2 (b) and (d)). The design crossed phrase type (coordinate; prepositional) and complexity (simple: not modified; complex: adjective modified).

Figure 2.

Example of the stimuli images. (a) to (d) illustrate an experimental item in the four conditions crossing phrase type and complexity, whereas (e) and (f) are examples of filler items. The utterances corresponding to (a) to (d) are, respectively (English version): ‘The cone and the grape’, ‘The cone and the pink grape’, ‘The cone above the grape’, and ‘The cone above the pink grape’.

We created four different sets of stimuliFootnote ⁶, each using 20 different target pictures, from the MultiPic picture database (Duñabeitia et al., Reference Duñabeitia, Crepaldi, Meyer, New, Pliatsikas, Smolka and Brysbaert2018), combined in 20 unique picture pairs (each picture occurred both as the first or second element of a different pair). In each set, each of the 20 word-pair items appeared in the four experimental conditions so that each participant experienced every item in every condition. Thus, each set comprised 80 experimental items distributed across four conditions, with each individual experimental picture appearing 8 times (rotated across the screen positions). The other two of the four pictures on each display were pictures that did not appear in another experimental item but only as part of the filler displays. In addition to the 80 experimental items, we had 48 filler items.

The length of the names of the target pictures did not differ significantly between English and Norwegian, as to the number of syllables and phonemes (see Appendix 5 for further details).

Two practice blocks were administered before the experiment started (containing the 20 experimental pictures re-arranged and 16 fillers). For each set of stimuli, we created two different orders of stimuli presentation. Therefore, we had 8 different lists, and participants were pseudo-randomly assigned to these.

Procedure

Pictures were named in speakers’ L1. Following common practice in the picture naming literature, participants were first given a booklet to familiarise with the pictures and respective names, and they were encouraged to look carefully at each image-name pair, and to confirm that they recognized it. Participants were also shown previously examples of the grids of pictures that would appear on the screen, for which they would be asked to uniquely identify the red framed pictures. They should describe the pictures having a horizontal red frame around them from left to right using ‘and’ (e.g., ‘the cone and the grape’; ‘kjeglen og druen’), and the pictures having a vertical red frame around them from top to bottom using ‘above’ (e.g., ‘the cone above the grape’; ‘kjeglen over druen’). They were also told that one of the objects to be named could have a different colour or size but be of the same type as one object outside the red frame. In this case, they should use the adjectives to distinguish the picture within the red frame. Finally, they were told there were other cases where all the pictures would be identical within the display or where the display would be empty, in which case they should say ‘all the images are the same/ alle bildene er like’ or ‘there are no pictures/ det er ingen bilder’. Participants were asked to start speaking as quickly and accurately as possible, after seeing each display.

The experiment was presented using Presentation® (Version 20.1, Neurobehavioral Systems, Inc.). Each trial began with a central fixation cross (‘+’) presented for 500ms, followed by a 500ms blank screen, after which the multi-picture display was presented and the recording of production started. The program registered speech onset, and an experimenter was present throughout the experiment to record the accuracy of responses (responses in which participants did not use the expected names or phrase type, where they did not mention the adjective, or disfluent responses were all categorized as errors). The picture-display disappeared after the participant finished speaking (or after 3000ms with no production), triggering the presentation of the next trial. The task took approximately 25 minutes.

Analysis

Our analyses of speech onset were restricted to trials with correct utterances (82.5% of data). Due to a technical problem, the voice onsets for one monolingual older adult were not registered and the data from this participant are missing from the analyses.

We first excluded extreme speech onset measures <250ms or >2500ms (2.1%). We then z-scored RTs to avoid spurious interactions between groups and conditions, due to general slowing (see Analysis for Experiment 1), and further discarded observations distancing more than 2.5SD from the mean, for each participant (i.e., zscores <-2.5 or >2.5, corresponding to 1.8% of data).

Speech onsets were analysed with LMMs, as in Experiment 1. Participants (159) and items (80) were entered as random factors. The fixed effects were bilingualism group (bilingual vs. monolingual), age group (old vs. young), phrase type (coordinate vs. prepositional), and complexity (complex vs. simple), all contrast coded by centering (refer to summaries of models for the coefficients). Moreover, vocabulary scores (voc; numeric, scaled) were added as a co-variate, to account for individual differences in language proficiency, and Education level was added as a confounding variable (as in Experiment 1). We fitted full models with a maximal-random structure when justified by the design and if convergence occurred (refer to the summaries of models for its syntax). We report the predictors’ coefficients (β values), SE, t values, and the derived p significance values.

Results and discussion

Figure 3 presents the mean speech onset latencies (z-scored, refer to Appendix 7 for raw RTs), and Table 3 presents the fitted LMM model to (z-scored) RTs.

Figure 3.

Mean onset latencies (z-scored, ms) for the Younger and Older Monolingual and Bilingual groups, in the four experimental conditions. Error bars represent standard errors on means.

Table 3.

Summary of the linear mixed effects model fitted to the z-scored speech onset times (RT).

Note: The syntax of the model is: blmer(depM ~ 1 + PhraseType + BilingualismGroup + Complexity + voc + AgeGroup + HighestEdu + PhraseType:AgeGroup + BilingualismGroup:AgeGroup + Complexity:AgeGroup + voc:AgeGroup + PhraseType:voc + BilingualismGroup:voc + Complexity:voc + PhraseType:BilingualismGroup + PhraseType:Complexity + BilingualismGroup:Complexity +

PhraseType:AgeGroup:BilingualismGroup + PhraseType:AgeGroup:Complexity + PhraseType:AgeGroup:voc + BilingualismGroup:AgeGroup:Complexity + BilingualismGroup:AgeGroup:voc + Complexity:AgeGroup:voc + PhraseType:voc:BilingualismGroup + PhraseType:voc:Complexity +

BilingualismGroup:voc:Complexity + PhraseType:BilingualismGroup:Complexity + PhraseType:BilingualismGroup:Complexity:AgeGroup + (1 | subj) + (1 | item)

+ (0 + Complexity | subj), data = dataset, control = lmerControl(optimizer =“Nelder_Mead”)

Participants were slower to produce coordinate than prepositional utterances, consistent with prior evidence for young monolinguals (Allum & Wheeldon, Reference Allum and Wheeldon2007). This confirms that, in the face of the syntactically simpler coordinates, speakers engage in longer planning, arguably preparing the whole coordinated phrase (CNP). Conversely, to produce more complex prepositional phrases (PP) speakers plan less, leaving preparation of (at least some aspects of) the second NP for later. Our data are the first to show that differences in the planning of coordinate and prepositional phrases are used in similar ways by monolinguals and bilinguals, and by young and older adults. Speakers also took longer to start producing utterances where the second NP was modified by an adjective, compared to utterances with non-modified NPs. This shows that the additional complexity of the adjective elicited longer preparation time, and that, despite a general strategy whereby, for more complex phrases (PP), the second NP is not fully prepared in advance, there is some level of processing of it before speech (see, Allum & Wheeldon, Reference Allum and Wheeldon2007, for discussion).

The effect of phrase type (CNP, PP) interacted with age group. In a follow-up analysis using a four-level variable combining the variables age and phrase type, we found that young and older participants did not differ in the onsets of coordinate utterances (olderxcoord_youngerxcoord; p = 0.47), but older adults were slower than young adults in prepositional utterances (olderxprep_youngerxprep; β = -0.07, SD = 0.02, p < 0.01). A possible interpretation is that older adults planned more in advance than young adults when the phrase to be uttered was syntactically more complex (PP). Yet, another possibility is that older adults engage in longer planning when possible (that is, in PP structures where some planning of the second NP could occur in parallel with planning of the first NP). In fact, the difference in syntactic complexity between coordinated and prepositional phrases is already reflected in the fact that, for prepositional phrases (the more complex structure), speakers prioritise preparation of the first noun phrase before speech. Therefore, we argue that this result indicates an overall larger planning scope in older age, whereby speakers prepared some aspects of the second NP (note that this would not affect CNP as in that case there is already full planning of the utterance). As we mentioned, advanced planning allows for more fluency once speech starts, and this should be particularly relevant for older adults, who have increased speech disfluencies (Cooper, Reference Cooper1990; Kemper, Reference Kemper1992; though see Spieler & Griffin, Reference Spieler and Griffin2006, for contrasting evidence).

Phrase type also interacted with complexity, indicating that structures with adjective modified NPs evoked slower responses in coordinates than prepositional phrases, which confirms extensive pre-speech processing for coordinates but the prioritising of the first |NP in prepositional phrases. The interaction between phrase type and complexity was further qualified by a three-way interaction with Bilingualism. We conducted follow-up LMMs on the data from the complex and simple conditions, separately, and in each case, we created a four-level variable combining the two phrase type levels (coordinate vs. prepositional) with the two bilingualism levels (bilingual vs. monolingual). Bilinguals tended to be slower than monolinguals producing coordinate utterances that had a modified second NP (complex CNP; Bilingualxcoord_Monolingualxcoord; β = -0.07, SD = 0.04, p = 0.069; the effect of bilingualism was not observed in complex PPs, nor in any of the no adjective conditions: all ps > 0.4). This effect only approached significance, and therefore should be interpreted with caution. Yet, it is consistent with the idea that the effect of bilingualism is to slow speech onset, but only for the phrase productions that demand the longest planning – that is, coordinated noun phrases with the second NP modified by an adjective.

Finally, language proficiency, as measured by vocabulary scores, did not affect speech onset latencies.

General discussion

Many studies on bilingual language processing have focused on L2, and have examined comprehension and production separately. In the current study, we tested the same participants in comprehension and production of L1 (while controlling for language proficiency), and demonstrated comparable findings – namely, poorer syntactic processing in bilinguals compared to monolinguals. We thus provide evidence for bilingualism effects per se, that are modality-general. Moreover, we tested young and older adults, and observed age-preserved syntactic processing across modalities. In our comprehension experiment, we also tapped onto semantic processing, and we found no differences between mono and bilinguals, whereas older adults tended to outperform younger adults. Importantly, we found no interactions between ageing and bilingualism, suggesting that the two factors independently impact language processing. Any disadvantages in bilingual syntactic processing were not counteracted by the preservation of that ability in ageing, and thus the bilingual differences remained in old age. On the other hand, what seem to be advantages in semantic processing in old age are not larger for bilinguals than monolinguals, suggesting that a higher reliance on accrued knowledge by the older adults does not accumulate with, or is distinct in nature from, any potential increased reliance on lexical-semantics by bilinguals.

The results of Experiment 1 suggest that bilingualism adversely impacts syntactic processing during comprehension, a finding resembling evidence for L2 processing (e.g., Pozzan & Trueswell, Reference Pozzan and Trueswell2016). On the contrary, syntax seems to be preserved in old age. Whereas some research reported age-related decline in processing of complex syntactic structures (e.g., Caplan et al., Reference Caplan, Dede, Waters, Michaud and Tripodis2011) or morpho-syntactic agreement interpretation (e.g., Poulisse et al., Reference Poulisse, Wheeldon and Segaert2019), we replicated Tyler et al. (Reference Tyler, Shafto, Randall, Wright, Marslen-Wilson and Stamatakis2010) in finding no age-related differences in reaction times to detect words during online comprehension of sentences with ‘normal’ syntactic complexity.

Moreover, we used sentences where the target words were highly predictable given the sentence context to assess semantic processing. Consistent with some evidence (e.g., Gollan et al., Reference Gollan, Slattery, Goldenberg, Van Assche, Duyck and Rayner2011; Tyler et al., Reference Tyler, Shafto, Randall, Wright, Marslen-Wilson and Stamatakis2010), we found no processing differences between monolinguals and bilinguals, and older adults even tended to outperform young adults when sentences provided a high-constraint context. The reliance on semantic information might reflect strategies to compensate for detriments in other processes, such as general cognitive decline, in healthy ageing (Opdebeeck et al., Reference Opdebeeck, Martyr and Clare2016), and mirrors findings for L2 processing (e.g., Clahsen & Felser, Reference Clahsen and Felser2006). We should note, however, that there is also physiological evidence against identical semantic processing mechanisms, both in bilingualism and in older age, from studies using sentences with low- and high-constraints (Federmeier & Kutas, Reference Federmeier and Kutas2005; Federmeier et al., Reference Federmeier, McLennan, De Ochoa and Kutas2002, Reference Federmeier, van Petten, Schwartz and Kutas2003; Martin et al., Reference Martin, Thierry, Kuipers, Boutonnet, Foucart and Costa2013). This highlights that, even when performance between groups is comparable, the underlying brain function mechanisms may not be. For our present findings, it may also be the case that seemingly comparable behaviour across groups is supported by different neural mechanisms.

In Experiment 2, we examined how the grammatical complexity of to-be-produced phrases affected planning scope. We showed, for the first time, that similar planning scope patterns of larger planning for coordinated phrases than prepositional phrases (Allum & Wheeldon, Reference Allum and Wheeldon2007) are used across age and bilingualism groups. However, older adults were slower in producing prepositional phrases than younger adults, suggesting an overall larger planning scope in older age.

We further manipulated syntactic complexity by including phrases with an adjective modifying the second NP. We found that the adjective modification was particularly demanding when speakers planned coordinates, confirming the preparation of the whole utterance in advance. Thus, modified coordinated phrases required most preparation, as indicated by overall longer latencies. Complexity did not differentially affect old and young adults. This is in line with the absence of age-differences in syntactic comprehension (Experiment 1), and further supports our suggestion that older adults were slower in producing prepositional phrases because they engaged in more extensive planning overall, and not because of the additional syntactic complexity of these structures. However, bilinguals showed a disadvantage (longer speech onset latencies) relative to monolinguals in modified coordinated phrases. Therefore, syntactic complexity affected bilinguals more than monolinguals, which is in line with our findings from Experiment 1.

We were interested in the types of linguistic information that could underlie disadvantages in language processing in bilinguals and older adults. For that purpose, we used whole sentences and complex phrases as stimuli, and interpreted our results as reflecting syntactic and semantic processing effects on lexical access.

Lexical access has been at the heart of accounts for bilingual disadvantages. One of the main findings in bilingualism research refers to parallel activation of bilinguals' two languages in both comprehension (e.g., Schwartz & Kroll, Reference Schwartz and Kroll2006; Spivey & Marian, Reference Spivey and Marian1999) and production (e.g., Colomé, Reference Colomé2001; Costa et al., Reference Costa, Caramazza and Sebastian-Galles2000). According to the competition hypothesis, the bilingual disadvantage results from the competition between activated lexical items and the need for conflict resolution during lexical access.

Semantically constraining contexts reduce the extent of dual language lexical activation during comprehension (e.g., Schwartz & Kroll, Reference Schwartz and Kroll2006) and, therefore, reduce language competition and the bilingual disadvantage (e.g., Gollan et al., Reference Gollan, Slattery, Goldenberg, Van Assche, Duyck and Rayner2011). We replicate this finding by showing a reduced bilingual disadvantage in comprehending words embedded in high-, compared to low-constraint sentences. The fact that syntactic and semantic information are used to guide comprehension is furthermore supported by the comparison between the low-constraint sentences and the random word order condition. When no coherent syntactic or semantic information is present, more lexical candidates are active, slowing processing. Yet, although both mono- and bilinguals were slower in this condition compared to normal sentences, bilinguals outperformed monolinguals in monitoring words embedded in random word lists. This effect is compatible with better attentional control in bilinguals, compared to monolinguals (Bialystok & Craik, Reference Bialystok and Craik2022). Future research is required to determine the role of attention and executive function in more complex language processes.

On the other hand, the frequency lag hypothesis (Gollan et al., Reference Gollan, Montoya, Cera and Sandoval2008, Reference Gollan, Slattery, Goldenberg, Van Assche, Duyck and Rayner2011) proposes that the bilingual disadvantage reflects frequency effects emerging from patterns of language use: bilinguals use word forms less frequently than monolinguals, and thus lexical items are less accessible, in both L1 and L2. Studies evaluating the frequency lag hypothesis have shown that the bilingual disadvantage is larger for low- than high-frequency words, and propose that the greater frequency of use in monolinguals leads to a ceiling in activation. Our study was not set to test this hypothesis, and does not speak directly to frequency effects. Also some aspects of our study would minimize the potential differences in frequency of use. First, our stimuli were matched for frequency in English and Norwegian. Second, we found no effects of language proficiency, which should correlate with language experience. Finally, and more importantly, our bilingual population had Norwegian as both their first acquired and dominant language (in contrast to switched-dominance bilinguals; see Hanulová et al., Reference Hanulová, Davidson and Indefrey2011, for discussion), they were early bilinguals (Table 1), and resident in Norway (the L1 speaking country). Although we had no direct measure of language exposure, these are conditions that promote an experience of L1 closer (though necessarily smaller) to that of monolinguals, and therefore our reported bilingual disadvantage likely reflects the (additional) influence of something other than a frequency lag.

We also argued that syntactic complexity affected bilinguals in Experiment 2, where modification of the second NP reflected higher syntactic complexity. In a previous study, Sadat et al. (Reference Sadat, Martin, Alario and Costa2012) asked mono- and bilinguals to name pictures by producing either single words (e.g., ‘airplane’) or adjective modified NPs (e.g., ‘the red airplane’). Bilinguals always took longer to start speaking than monolinguals, but the bilingual disadvantage in speech onset was of similar magnitude for single word and modified-NP utterances. Although production is arguably more complex for modified NPs than bare nouns, it is possible that Sadat et al.'s (2012) stimuli were not complex enough to elicit differences. In contrast, we elicited utterances comprising two NPs. When these were coordinated (CNP; ‘The cone and the grape’), speakers across groups planned more extensively, but the additional complexity of having the second NP adjective-modified (‘The cone and the pink grape’) only affected the bilingualism group, where bilinguals showed a cost in speech onset relative to monolinguals. Of course, in the CNP condition, there are more lexical items to retrieve, which makes our result compatible with a bilingual disadvantage being observed due to lexical competition. However, this finding is also compatible with an explanation based on the increased complexity of the syntactic processes needed to produce these utterances.

It has been argued that the bilingual disadvantage due to a frequency lag may be restricted to low-frequency lexical items and syntactic structures. Again our study does not speak directly to this issue, as we tested structures that are frequent in the tested languages (see Runnqvist et al., Reference Runnqvist, Gollan, Costa and Ferreira2013, for related evidence on a bilingual disadvantage in production of infrequent structures and the teasing apart of lexical and syntactic retrieval). Crucially, the fact that we did find reliable bilingual disadvantages in processing frequent stimuli, for early and non-switched dominant bilinguals, suggests that the effects of being bilingual on language processing go beyond effects of frequency of experience with a language.

More important for the contrast between the competition and the frequency lag theories is the lack of interactions between bilingualism and age that we found. According to the frequency lag hypothesis, increased language experience should result in reduced bilingual disadvantages in older compared to younger adults, as frequency of use of words and structures accrues with age, potentially reaching the ceiling effects of activation observed for monolinguals (Gollan et al., Reference Gollan, Montoya, Cera and Sandoval2008, Reference Gollan, Slattery, Goldenberg, Van Assche, Duyck and Rayner2011; Runnqvist et al., Reference Runnqvist, Gollan, Costa and Ferreira2013). However, we observed no interactions between ageing and bilingualism, against the predictions of the weaker links account. Consistent with prior evidence, in our additional proficiency tasks, older adults were slower in reading comprehension but had better vocabulary scores than younger adults, across bilingualism groups. In addition, bilinguals scored lower in the vocabulary task, but this difference was only significant in the young group (see Appendices 2 and 3). However, these individual measures of proficiency did not affect comprehension and production, further suggesting that the effects assessed by our tasks were not mediated nor modulated by language competence.

On the other hand, accounts of bilingualism appealing to control processes (Green, Reference Green1998) do not make clear predictions about how bilingualism would interact with age. We might assume also a smaller bilingual disadvantage in older age, if mechanisms like inhibition ameliorate with age (Markiewicz et al., Reference Markiewicz, Rahman, Fernandes, Limachya, Wetterlin, Wheeldon and Segaert2024; Veríssimo et al., Reference Veríssimo, Verhaeghen, Goldman, Weinstein and Ullman2022), but it is also conceivable that other cognitive resources involved in language, which decrease with age (e.g., working memory), would counteract any potential advantages (as suggested by Gollan et al., Reference Gollan, Montoya, Cera and Sandoval2008).

Whereas the frequency lag hypothesis is appealing – namely, in providing a single explanation for both language dominance and bilingualism effects – it seems to be insufficient for a comprehensive account of the bilingual disadvantage, particularly in the face of the evidence for a bilingual advantage in executive control tasks (e.g., Bialystok et al., Reference Bialystok, Craik and Luk2012), which could result from bilinguals negotiating cross-language competition (Kroll et al., Reference Kroll, Bobb, Misra and Guo2008). Future research should examine how age and bilingualism effects in language processing are affected by frequency of use, language competition, and other measures of broader cognitive function.

Our study joins research that failed to find interactions between bilingualism and age in language processing (Gollan et al., Reference Gollan, Montoya, Cera and Sandoval2008), and our findings suggest that bilingualism and ageing are two factors that independently impact on language processing. Nevertheless, given the lack of studies examining combined effects of bilingualism and ageing, further research is important for elucidating the underlying mechanisms involved, and for evaluating the possibilities for ‘connecting models’ of language processing in bilingualism and healthy older age (Rossi & Diaz, Reference Rossi and Diaz2016).

Moreover, future studies could address some limitations of the research conducted so far. In the current study, we made every attempt to make the materials comparable across languages (English and Norwegian). However, this is always difficult to achieve, and ultimately there is always a potential confound of language.

Different studies have used varying tasks and stimuli to tap into language processes, which may be differently affected by both age and bilingualism. For example, whereas our comprehension experiment measured comprehension of target words in spoken sentences, it would also be important to investigate the interaction between age and bilingualism in written comprehension of complex sentences. Likewise, we accounted for effects of proficiency, but neither reading times nor vocabulary are catch-all measures of all domains that affect language proficiency, and so it is possible that the effects would be modulated by aspects of proficiency or language use that we did not measure.

Finally, bilinguals are a diverse population, and researchers have been encouraged to consider more in-depth differences in bilingual profile (e.g., Rothman et al., Reference Rothman, Bayram, DeLuca, Di Pisa, Duñabeitia, Gharibi, Hao, Kolb, Kubota, Kupisch, Laméris, Luque, van Osch, Pereira Soares, Prystauka, Tat, Tomić, Voits and Wulff2022). In the current study, we tested Norwegian–English bilinguals, who can be thought to be a relatively homogeneous population concerning language use and exposure, but it would be important to test other bilingual populations to determine how variability in bilingual profile can affect processing.