Effects of bilingualism on autobiographical memory: Variation in idea density and retrieval speed

Abstract

Prior work on bilingual memory has largely focused on working memory and less on autobiographical memory. In the present study, we tested the effect of bilingualism on autobiographical memory and examined whether an effect would be moderated or mediated by working memory. Spanish-English bilingual and English-only monolingual adults completed an autobiographical cued-recall task, as well as a working memory measure. Memories were coded for retrieval speed and propositional idea density. Bilingual status was associated with faster memory retrieval but did not affect propositional idea density. Better working memory was associated with slower memory retrieval but did not affect propositional idea density, nor did working memory moderate or mediate the effect of bilingualism. Together, these results indicate an effect of bilingualism on the speed of autobiographical memory retrieval that does not extend to autobiographical memory content and suggest that the effect of bilingualism is independent of the effect of working memory.

Introduction

Effects of bilingualism on memory have garnered much attention, across both the working memory and long-term memory systems. Working memory—the part of the memory system that stores and manipulates information over short periods of time and which includes a central executive component (Baddeley, 1992; McCabe et al., 2010)—is thought to be strengthened by the practice of switching between multiple languages (Bialystok, 2015), though evidence of a bilingual advantage for working memory is mixed (see Grundy & Timmer, 2017 for review). Episodic memory– a system that falls within the long-term memory domain and entails encoding and recall of events that are personally experienced (Tulving, 2002)—is thought to be affected by bilingualism via the presence of language context within the memory trace (Schroeder & Marian, 2014; Bartolotti & Marian, 2012). Episodic memory is further divided into the autobiographical and nonautobiographical, with autobiographical episodic memory referring to memory for events for which the self has a central role (Baddeley, 1992); here, we focus on autobiographical memory.

Existing studies of bilingual autobiographical memory have emphasized the role of language context in accounting for variation in bilinguals’ memory retrieval, often testing bilinguals’ recall in each of their two languages rather than comparing bilingual and monolingual performance. In an early example of this work, Javier and colleagues (1993) asked each of their five Spanish-English bilingual participants to relay the same personal memory twice, beginning in the language in which the event was experienced. Their responses were transcribed and then analysed for richness of detail and organization. Javier and colleagues found that participants described their memories with greater richness and detail when the language at recall matched the language at encoding. This study provided early evidence of the impact of language context on autobiographical memory, prompting further investigation into bilingual autobiographical memory.

Subsequent work has suggested that language context is part of the memory trace and impacts later recall (see Schrauf, 2000 for review). Studies of sequential bilinguals seem to indicate that language context makes particular time periods (e.g., Matsumoto & Stanny, 2006) or cultural contexts (e.g., Marian & Kaushanskaya, 2007) more accessible. Such findings have been reported in studies of Polish-Danish (Larsen et al., 2002), English-Mandarin (Marian & Kaushanskaya, 2007), English-Japanese (Matsumoto & Stanny, 2006), and English-Spanish (Marsh et al., 2015) bilinguals, with congruency of language context facilitating memory retrieval, but such studies rarely consider monolinguals. One exception is Matsumoto and Stanny’s (2006) study of Japanese-English bilinguals and English monolinguals, which indicated that bilinguals and monolinguals retrieve similar quantities of autobiographical memories, but that bilinguals retrieve more numerous and earlier memories when responding to Japanese cue words rather than English cue words. Moreover, bilinguals and monolinguals did not differ in the emotionality of their reported memories, but bilinguals retrieved memories more slowly than monolinguals. These results indicate important differences and similarities in bilingual and monolingual memory retrieval, but they do not touch on potential differences in the quality or richness of retrieved memories.

There is limited work directly comparing the quality of memories retrieved by bilinguals and monolinguals in autobiographical memory tasks. Marsh and colleagues (2015) have come closest to this by assessing the quality of bilingual and monolingual participants’ autobiographical memory based on the number of items recalled, but they tested bilingual participants under relatively more restrictive conditions (i.e., bilinguals were randomly assigned to receive materials in Spanish or English, without consideration for personal preference or dominance, whereas the monolingual control group could not be restricted in a similar manner). Yet, it is possible that availability of additional cues to bilinguals, in the form of linguistic context, might enable bilinguals to be more successful in recalling autobiographical events than monolinguals. Some support for the possibility of broad bilingual advantages in autobiographical recall comes from studies of non-autobiographical episodic memory.

Research comparing bilingual and monolingual non-autobiographical episodic memory has found evidence for bilingual advantages in non-autobiographical memory across the lifespan. In one study, Schroeder (2019) tested college-aged Spanish-English bilinguals and English monolinguals on a visual recall task, where participants were presented with images of shapes and asked to draw what they remembered at recall. To measure the effect of language context on shape recall, learning trials were accompanied by the phrase “this drawing looks like this” in either English or Spanish. Participants were told that they only needed to remember the images they saw, and that they could ignore the person talking. Monolinguals and bilinguals completed the exact same task and were exposed to both English and Spanish blocks. Schroeder found that changes in language between blocks improved bilingual participants’, but not monolingual participants’, visual memory recall. These results are particularly interesting when considering the characteristics of Schroeder’s bilingual group – many of his bilingual participants experienced only a few years of exposure to their second language. These findings suggest that variation in language context may aid in the organization and retrieval of remembered information, even for bilinguals who are not highly proficient in both of their languages. Evidence for bilingual advantages in non-autobiographical memory has also been found in studies with children (Kormi-Nouri et al., 2003; Kormi-Nouri et al., 2008) and older adults (Schroeder & Marian, 2012), although the mechanisms underlying such advantages remain unknown. In the present study, we consider the possibility that the effect of bilingualism on autobiographical memory may be moderated or mediated by working memory, the part of the memory system that stores and manipulates information over short periods of time, and which includes a central executive component (McCabe et al., 2010).

The relationship between working memory and episodic memory is well documented empirically (Unsworth et al., 2011; Unsworth & Spillers, 2010) and debates regarding the theorical unity of the memory system continue (Norris, 2017; Cowan, 2019; Norris, 2019). Successful episodic remembering is dependent on successful encoding and retrieval, attentiondemanding processes that tap into working memory (McCabe et al., 2010). However, to our knowledge, the relationship between working memory and autobiographical episodic memory has not yet been examined. Evidence does exist for a relationship between working memory and long-term memory, more broadly construed. In a study of autobiographical semantic memory (i.e., memory for information about the self that is not tied to a specific event; Baddeley, 1992), Unsworth and colleagues (2012) found that participants categorized as having high- or low-working memory capacity performed differently on an autobiographical semantic memory retrieval task. Compared to participants with low-working memory capacity, participants with high-working memory capacity recalled a greater number of friends’ names and generated more retrieval cues. Here, we asked whether verbal working memory may moderate or mediate the relationship between bilingualism and autobiographical recall.

In investigating a potential moderating effect of working memory, we follow logic similar to Kaushanskaya and colleagues (2011), who found a relationship between working memory and receptive vocabulary in low-proficiency bilinguals but not in high-proficiency bilinguals. Their proposed explanation for this finding was that low-proficiency bilinguals would be more reliant on working memory for vocabulary retrieval whereas vocabulary retrieval would be more automatized and less effortful for high-proficiency bilinguals. In the context of our study, we hypothesized a similar moderating effect of working memory on the relationship between bilingualism and autobiographical memory. We predicted an interactive effect of bilingualism and working memory on autobiographical memory retrieval such that bilinguals would perform similarly across levels of working memory performance, whereas monolinguals with higher working memory scores would outperform monolinguals with lower working memory scores.

We also asked whether working memory could be the mechanism underlying any bilingualism-autobiographical memory relationship. The relationship between working memory and bilingualism is complex, and some studies have reported null effects of bilingualism on working memory performance in young adults (e.g., Bialystok et al., 2008; Bialystok & Craik, 2010; Namazi and Thordardottir, 2010; Bonifacci et al., 2011; Engel de Abreu, 2011; Ratiu & Azuma, 2015; McVeigh et al., 2017) whereas other studies have observed bilingual advantages in working memory for this age group (e.g., Bialystok et al., 2004, Morales et al., 2013; Blom et al., 2014; Jiao et al., 2019; Luo et al., 2013; Antón et al., 2019). A recent meta-analysis by Grundy and Timmer (2017) including 88 effect sizes supported the existence of a bilingual advantage for working memory, with the caveat that larger effects were seen in children relative to other age groups. Alternatively, Calvo and colleagues (2016) argued that bilinguals more frequently demonstrate a working memory advantage on nonverbal (visuospatial) tasks than verbal tasks, consistent with the common finding that bilinguals tend to do less well than monolinguals on tasks involving lexical retrieval (Sullivan et al., 2018). At the same time, verbal (rather than visuospatial) working memory capacity may be particularly important for autobiographical verbal recall. In the present study, we assessed whether our results would be consistent with a mediation model whereby bilingualism affects autobiographical memory through verbal working memory.

The current study

The present study was designed to evaluate the effects of bilingualism on autobiographical episodic memory. Our predictions are based on evidence of bilingual advantages in non-autobiographical memory and theorizing on the structure of bilingual memory. If bilingual advantages in non-autobiographical memory extend to autobiographical memory, then we would find that bilinguals retrieve autobiographical memories faster and with greater richness and detail than monolinguals.

Our predictions regarding a possible moderating and/or mediating effect of working memory capacity are rooted in prior evidence that the same skills that support efficient storage and manipulation of information in working memory are involved in the efficient retrieval of long-term memories (Unsworth, 2019), and that bilingualism may impact working memory or enable participants to demonstrate strong task performance even if their working memory is relatively low (Kaushanskaya et al., 2011). The relationship between working memory and autobiographical memory may vary by language group, such that working memory affects performance for monolinguals—whom we expect to be disadvantaged relative to bilinguals—but is less important for bilingual performance.

Accurate operationalizing of autobiographical memory was critical to our study. Response time is a relatively straightforward operationalization of the efficiency of the long-term memory system, one that Unsworth et al. (2012) and Matsumoto and Stanny (2006) have used in prior work. Measuring the quality of a retrieved memory can be a controversial and timeconsuming process. Quality can mean the fidelity of a retrieved memory—i.e., how well it reflects actual events (Rubin et al., 2003)—or the richness of a memory. Richness, in turn, can be assessed subjectively via self-report measures that query the vividness of the retrieved memory (Luchetti & Sutin, 2016; Rubin et al., 2003; Boyacioglu & Akfirat, 2014) or coded by an experimenter based on idea content (Javier et al., 1993; Snowdon et al., 1996; Marsh et al., 2015). Here, we followed Javier and colleagues (1993) in asking participants to describe memories with no attempt to verify the authenticity of these memories – a difficult premise at best; the transcribed memories were then analysed for idea content under the assumption that memory for an event aligns with verbalization of that event (Kintsch, 1974). Similar to Javier and colleagues, we then evaluated these memories in terms of propositional content. Propositional idea density (commonly referred to as idea density) is a measure of semantic richness that has been particularly useful for researchers who study cognitive decline (Snowdon et al., 1996).

The utility of propositional density in predicting and measuring cognitive decline has led to its adoption by Alzheimer’s researchers and, subsequently, the creation of automated propositional idea density calculators. Propositional density is traditionally calculated by dividing the amount of propositional content included in a language sample by the total number of words in a language sample (Kintsch, 1974), and it can be a tedious and error-prone task. Automated propositional density calculators, such as the Computerized Propositional Idea Density Rater (CPIDR; Brown et al., 2008) and the Dependency-based Propositional Idea Density calculator (DEPID; Sirts et al., 2017) streamline propositional density coding by automating the bulk of the process.

CPIDR (Brown et al., 2008) is perhaps the most widely used of these calculators. CPIDR operates according to a claim shared by Kintsch (1974) and Turner and Greene (1977): certain syntactic categories embed propositional content whereas others are propositionally empty. Using a part-of-speech tagger, CPIDR determines the categories of each word in a language sample and calculates idea density based on the number of proposition-containing words to total number of words. The idea densities calculated by CPIDR are reported to be highly reliable and well-correlated with manual idea density scores (Brown et al., 2008), but the program has been criticized for overestimating propositional content (Chand, 2012; Sirts et al., 2017). CPIDR remains a popular option for calculating idea density and has been integrated into TalkBank’s Computerized Language Analysis (CLAN) program (MacWhinney, 2000), but a newer calculator, DEPID, has been introduced and used in dementia research (Shibata, Ito, Nagai, Okahisa, Kinoshita & Aramaki, 2018). In contrast to CPIDR’s method of associating certain word classes with propositional content, DEPID estimates propositional structure of a language sample based on its dependency structure (Sirts et al., 2017). This method may result in somewhat less inflated estimates of idea density relative to CPIDR (Sirts et al., 2017). Given the advantages and disadvantages of each calculator and the novel application to bilingual memory, we chose to calculate and report idea density via both calculators.

In summary, the current study examined the effects of bilingualism on autobiographical episodic memory through a cued-recall task. We hypothesized effects of bilingualism on autobiographical episodic memory, operationalized as retrieval speed and propositional idea density of retrieved memories. We also tested the possibility that verbal working memory may be differently advantageous to autobiographical recall for bilinguals and monolinguals, and that the hypothesized effect of bilingualism on autobiographical recall would be mediated by working memory.

Materials and methods

Participants

Thirty-four Spanish-English bilingual (24 women, 10 men) and 38 English-only monolingual (27 women, 10 men) adults participated in this study. Of these participants, 3 monolinguals were excluded from analyses because of partial data loss.

All participants were recruited from the University of Wisconsin-Madison and reported no substantial exposure to languages besides English and Spanish or history of language impairment. Participants self-identified as bilingual or monolingual at recruitment. The language backgrounds of bilingual and monolingual participants, collected via the Language Experience and Proficiency Questionnaire (LEAP-Q; Marian et al., 2007), are summarized in Table 1. Notably, the majority of our bilingual participants were English-dominant (n = 32). Eighteen participants in the monolingual group reported some knowledge of Spanish (n = 15) or another second language (n = 3), but their self-reported proficiency in Spanish (mean spoken language proficiency of 3.00, SD = 1.13) or other second language (mean spoken language proficiency of 1.33, SD = .58) was minimal.

Table 1.

Participant Characteristics

	Monolinguals	Bilinguals
	M(SD)	M(SD)	p
N	35	34
Age (years)	20.76 (1.43)	20.69 (1.27)	.83
Years of education	15.24 (1.63)	15.68 (1.25)	.22
Non-verbal IQ a	105.03 (11.99)	104.56 (13.44)	.88
Verbal working memory – English b	101.23 (15.52)	102.62 (15.76)	.71
English age of acquisition (years)	0.14 (0.43)	0.32 (1.09)
English receptive vocabulary c	108.85 (12.0)	105.71 (10.15)	.0.25
Self-rated English speaking proficiency d	9.60 (0.77)	9.68 (0.64)	.66
Verbal working memory – Spanish b	-	91.8 (12.04)	-
Spanish age of acquisition (years)	-	8.50 (4.95)	-
Spanish receptive vocabulary e	-	102.44 (6.32)	-
Self-rated Spanish speaking proficiency d	-	7.12 (1.74)	-

^aStandard scores of matrices subtest of Kaufmann Brief Intelligence Test-II

^bStandard score of numbers reversed subtest of Woodcock-Johnson Test of Cognitive Abilities-III

^cStandard score of Peabody Picture Vocabulary Test-IV

^dSelf-rated 0–10 score from Language Experience and Proficiency Questionnaire; 0 is “none” and 10 is “perfect”

^eRaw score of Test de Vocabulario en Imagenes Peabody

In addition to the self-report language background information collected via the LEAP-Q, we administered standardized vocabulary measures to assess participants’ receptive English and Spanish abilities. English vocabulary knowledge was assessed with the Peabody Picture Vocabulary Test 4 (PPVT-4; Dunn & Dunn, 2007). For bilingual participants, Spanish vocabulary was measured with the Test de Vocabulario en Imagenes Peabody (TVIP; Dunn et al.,1986). Because the TVIP is intended for and normed on children 2:6 through 17:11, we report raw scores for this measure in line with common practices in the literature (Stadthagen-Gonzalez et al., 2013). Nonverbal intelligence was measured with the Visual Matrices subtest of the Kaufman Brief Intelligence Test (KBIT-2; Kaufman & Kaufman, 2004). Verbal working memory was assessed via the Numbers Reversed subtest of the Woodcock-Johnson III Tests of Cognitive Abilities (Woodcock et al., 2005). The backward digit-span task has been extensively used to assess verbal working memory in bilingual populations (Wodniecka, et al., 2010; Yoo & Kaushanskaya, 2012, Blom et al., 2014; Kaushanskaya & Crespo, 2019). The subtest was administered in English to monolingual participants and in both English and Spanish to bilingual participants. Bilinguals and monolinguals did not perform significantly differently on any of these measures (see Table 1).

Materials and procedure

Ten pictorial cues based on the cue words used by Marian and Neisser (2000) were used to prompt participants in the autobiographical memory recall task. The cues were photographs of an airplane, a birthday cake, a bride, a diploma and mortarboard, a doctor, a woman with her mouth open in fright, a baby, a dog and cat, a beach, and a teacher. All images were freely available online. Cues were presented in a random order one at a time via E-Prime 2.0 software (Psychology Software Tools, Inc., 2002); participants manually advanced through the task by pressing a button. At the onset of each image, a beep sounded. All participants received and responded to all ten cues.

For each pictorial cue, participants were instructed to orally describe the first personal memory that came to mind. Participants could take as much time as they needed to think of and begin describing a memory, and their response could be as long or as short as they wished. The instructions were written in English, but participants were told that they could respond in either English or Spanish. The task was self-paced; participants advanced by pressing a button. Responses were audio recorded.

Response times for the autobiographical memory task were extracted using Praat (Version 6.1.16, Boersma & Weenink, 2020). For each trial, coders marked the offset of the beep accompanying the image and the onset of the participant’s response. These two marked points were extracted using a Praat script and then used to calculate response times for each observation. Coders also noted trials on which a participants’ response was preceded by a disfluency (n= 448, or 65% of all trials) and/or discussion between the experimenter and participant (n = 14, or 2% of all trials).

All English responses were transcribed by one coder. Responses were transcribed in full, excepting commentary on the task and questions regarding the instructions, non-lexical (i.e., uh, um) and lexical (i.e., like, you know, I mean) fillers, repetitions, and false starts. Excluding disfluencies from transcripts is typical in prior work evaluating oral language samples for propositional idea density (Fromm et al., 2016; Brown et al., 2008), as disfluencies are considered empty of propositional content (Bosker et al., 2014). For the observations included in our model, monolingual transcriptions had a mean length of around 60 words as calculated by both CPIDR 5 (M = 61.46, SD = 66.16) and DEPID (M = 62.32, SD = 67.19), and bilingual transcriptions were slightly shorter (CPIDR 5: M = 48.21, SD = 63.94; DEPID: M = 48.88, SD = 64.83). Group differences in word count were significant across calculators (CPIDR 5: p = .009; DEPID: p = .009).

When coding response times, we extracted time to first disfluency as well as time to first semantically meaningful utterance; time to first semantically meaningful utterance was used for all response time analyses. Transcripts were segmented into conversational-units comprised of independent clauses and modifiers (Miller et al., 2018) prior to analysis. A second trained coder transcribed 10% of the responses according to these guidelines. To assess interrater reliability, we calculated two-way mixed effects, absolute agreement, single rater intraclass correlation coefficients in R (version 3.2.2; R Core Team, 2015) via the Psych package (version 2.0.12; Revelle, 2020). Interrater reliability was good for CPIDR 5 density (ICC of .835) and excellent for DEPID density (ICC of .976), based on guidelines from Koo and Li (2016).

Responses that were entirely (n =10) or partly in Spanish (n = 2) were transcribed by a bilingual coder fluent in both English and Spanish. The small number of Spanish responses reflects the English-dominant profile of the bilingual participants. Because the idea density calculators used (CPIDR 5 and DEPID) do not accept text in languages besides English, two fluent Spanish-English bilinguals, one with professional training in translation, worked collaboratively to translate these responses to English. Because the calculators use word count and class to compute idea density, translators were instructed to preserve these features as best they could. The final translations were agreed upon by both translators. To further ensure that translated responses did not impact our conclusions, all models involving idea density were run without these 12 observations. As the pattern of results remained the same, we report models with translated responses.

Propositional idea density was calculated for all responses using two idea density calculators, CPIDR 5 (Brown et al., 2008) and DEPID (Sirts, 2017). We used both calculators because this is a novel application of idea density calculators, and we wanted to see if the relationships between bilingualism, working memory capacity, and idea density generalized across different methods for calculating idea density.

Results

Data analysis

Item-level data were used in all analyses, except in some tests for mediation that relied on between-subjects variables. Linear mixed effects models were constructed to analyse data in R using the lme4 package (Bates et al., 2015), and normal linear models were constructed using the stats package (R Core Team, 2020), where appropriate. The correlation between CPIDR 5 and DEPID density estimates was calculated at the transcript-level, using the rmcorr package (Bakdash & Marusich, 2017) to account for non-independence at the subject-level. Although idea density estimates for CPIDR 5 and DEPID were significantly correlated, r_rm (596) = .33, p < .001, the correlation was modest, and separate models were constructed with CPIDR 5 density and DEPID density as dependent variables. The distribution of idea density values for each calculator is visualized in Figure 1. Another model was constructed with log-transformed response time serving as separate dependent variables. Reaction times (RTs) below 150ms and RTs that were more than 2.5 SDs above or below the individual participant’s mean were excluded. Following these data-trimming procedures, 10 trials were removed from analyses. An additional 14 trials were removed because the experimenter either provided further instructions or answered a question prior to the participant responding to the image prompt. Trimmed RTs were log-transformed to reduce skewness. The item-level analyses included a total of 666 observations. Data and analyses are available online (https://osf.io/3kgft/).

Figure 1.

Density plots describing the spread of idea density values for both CPIDR and DEPID. Idea density values must fall between 0–1 for both calculators.

In each of the interactive models, fixed effects included language group (monolinguals vs. bilinguals), working memory, and the interaction between language group and working memory. All models included a by-subject and by-item random intercept. The addition of a by-item random slope for the effect of working memory, but not language group, significantly improved model fit in only two models (ps < .05). However, only by-item random intercepts were retained in our final analyses due to singularity and convergence issues (Barr et al., 2013; Brauer & Curtin, 2018). The pattern of results remained the same with or without the inclusion of by-item random slopes for the effect of working memory. Years of formal education improved model fit for the model examining CPIDR 5 idea density performance and was therefore included as a covariate. The addition of age, gender, nonverbal IQ scores, and receptive vocabulary scores, did not improve model fit for models examining reaction time data or idea density calculated by CPIDR 5 or DEPID.

To find support for working memory as a mediator in a bilingualism-autobiographical memory relationship (1-1-1 mediation; Figure 2), four conditions must be satisfied (Baron & Kenny, 1986). Bilingualism must have an effect on both autobiographical memory (the total effect; Path C) and working memory (Path A). Moreover, working memory must have an effect on autobiographical memory after controlling for the effect of bilingualism on autobiographical memory (Path B), and the effect of bilingualism on autobiographical memory becomes smaller after controlling for the effect of working memory on autobiographical memory (Path AB). In testing the first condition of mediation, only by-item random intercepts were retained in our final analyses due to singularity issues (Barr et al., 2013; Brauer & Curtin, 2018). The pattern of results remained the same with or without the inclusion of by-item random slopes, and the inclusion of random slopes did not improve model fit for any of the dependent variables. As none of our models satisfied both of the first two conditions, we did not test conditions three or four.

Figure 2.

Hypothesized mediation models, with estimates and significance levels for tested pathways.

Moderation models

Bilingual language experience, working memory capacity, and their interaction did not significantly predict idea density calculated by CPIDR 5 or DEPID (ps > .05). For logRT, A significant main effect of language group was observed (β = - 0.12, SE = 0.06, t = −2.13, p = .04), such that adults with bilingual language experience were quicker to produce memories than monolingual adults. A significant main effect of working memory was also observed (β = 0.08, SE = 0.03, t = 2.90, p = .005). Adults with higher working memory capacity took more time to produce a memory when prompted by an image. The interaction between language group and working memory was not significant (t = 0.04, p = .97). See Table 2 for full model results.

Table 2.

Full moderation model results

	CPIDR 5		DEPID		RTlog
	β(SE)	t	β(SE)	t	β(SE)	t
Intercept	.46 (.04)	12.51***	0.55 (0.006)	85.27***	0.57 (0.04)	16.25***
Language Group	−0.0008 (0.007)	−0.11	−0.02 (0.009)	−1.77	− 0.12 (0.06)	−2.13*
Working Memory	−0.0001 (0.003)	−0.03	−0.002 (0.005)	−0.33	0.08 (0.03)	−2.90**
Years Education	0.005 (0.002)	1.99
Language Group X Working Memory	−0.003 (0.007)	−0.43	−0.01 (0.009)	−1.25	0.002 (0.06)	0.97
Observations	666		666		666
Akaike Inf. Crit.	1637.80		−1313.18		6.05
Bayesian Inf. Crit.	−1601.79		−1281.67		37.56

^*p <0.05;

^**p <0.01;

^***p <0.001

Mediation models

To test our hypothesis that bilingualism would confer an autobiographical memory advantage through working memory, we examined whether working memory mediated any effect of bilingualism on autobiographical memory. The first condition for mediation, a significant total effect of bilingualism, was not met for the CPIDR, DEPID, or logRT models (ps > .05). See Figure 2 for a visualization of these results. These results are not consistent with the hypothesis that working memory may mediate the relationship between bilingualism and autobiographical memory efficiency or quality. See Table 3 for full model results.

Table 3.

Full mediation model results

	CPIDR 5, Path C		DEPID, Path C		RTlog, Path C
	β(SE)	t	β(SE)	t	β(SE)	t
Intercept	.46 (.04)	12.88***	0.56 (0.008)	70.47***	0.62 (0.05)	13.57
Language Group	−0.0008 (.007)	−0.13	−0.02 (0.01)	−1.81	− 0.11 (0.06)	−1.91
Years Formal Education	0.005 (0.002)	2.08*
Observations	666		666		666
Akaike Inf. Crit.	−1659.32		−1331.90		0.85
Bayesian Inf. Crit.	−1632.32		−1309.39		23.36

^*p <0.05;

^**p <0.01;

^***p <0.001

General discussion

We observed a positive effect of bilingualism and a negative effect of working memory capacity on the speed with which autobiographical memories were retrieved. At the same time, we did not see an interactive effect of bilingualism and working memory, nor did we observe any significant predictors for the propositional idea density of the retrieved memories. Moreover, we found no evidence that working memory mediated the effect of bilingualism on autobiographical memory.

The finding of a bilingualism effect on autobiographical memory retrieval has novel and important implications for theories of bilingualism and its effects on cognition—a topic that is mired in controversy. While numerous studies have focused on the hotly contested effects of bilingualism on various attention-related cognitive skills (termed cognitive control, executive function, attention control, mental flexibility, etc.), less consideration has been given to possible effects of bilingualism on processes that may be less attention-dependent, like retrieval of autobiographical memories. Yet, the finding that the bilingualism was associated with faster retrieval but not with richer content of retrieved memories does point to an attention-based mechanism rather than a representational one – that is, bilinguals may be more efficient at retrieving autobiographical memories because their enhanced attentional skills enable them to more quickly select an appropriate memory to relate. However, this possibility seems at odds with the finding that working memory did not mediate the bilingualism effect. In the future, it would be important to probe whether participants had multiple memories that they decided among when selecting the particular memory, and whether that number might influence the speed with which memories were retrieved and interact with bilingualism.

Our finding of a bilingual advantage for retrieval speed is contrary to findings reported by Matsumoto and Stanny (2006), but this is likely due to differences in task demands and participant characteristics between the two studies. Whereas our monolingual and bilingual participants performed the same memory retrieval task, bilingual participants in Matsumoto and Stanny (2006) performed a relatively more challenging task than monolingual controls. Specifically, bilinguals were required to retrieve memories in response to cues across two languages, whereas monolingual controls responded to cues in a single language. The researchers noted that 11 of the 18 bilingual participants requested translations of some English cues, indicating that the bilingual group struggled when receiving cues in their non-dominant language. Our task involved picture cues, presented in the same manner to the two groups, and the equal processing demands were likely at the core of the faster response times observed in the bilingual group. Yet the mechanisms underlying this effect of bilingualism on response times in our task remain unknown. Prior work has indicated that differences in response time may reflect differences in retrieval strategy rather than differences in efficiency for a single strategy (Uzer et al., 2012; Conway & Loveday, 2010), and this possibility should be investigated in future work. In testing working memory as a potential mediator in the relationship between bilingualism and autobiographical memory, we explored the possibility that bilingualism, should it have an effect, may do so via the working memory system. Yet, this possibility was not realized, as bilingualism and working memory had independent, and opposite, effects on the retrieval of autobiographical memories. Moreover, working memory and bilingualism were unrelated in our sample, a finding that is not uncommon but deviates from the relationship indicated by Grundy and Timmer’s (2017) meta-analysis. Replicating this study with older adults or children, populations that show more consistent effects of bilingualism on working memory, would elucidate the role of working memory in the bilingualism-autobiographical memory relationship across the lifespan.

The finding that greater working memory capacity was associated with slower retrieval of autobiographical memories was unexpected and does not align with prior findings (i.e., Unsworth et al., 2012). A notable difference between Unsworth and colleague’s (2012) sample and ours concerns the “extremeness” of participants’ working memory capacities. Whereas we measured participants’ working memory capacities during the same visit at which they completed the memory task, Unsworth’s participants were recruited based on performance on an earlier measure of working memory capacity. Unsworth’s participants fell at the extremes of working memory capacity, whereas our participants’ working memory capacities followed a normal distribution. At the same time, it is not clear why graded vs. categorical approaches to capturing working memory capacity would lead to the dramatically different working memory and long-term memory relationships observed in our study vs. the Unsworth et al. (2012) study. One possible reason for the finding regarding the relationship between working memory and autobiographical memory retrieval relates to cue generation. Prior studies have indicated that high-working memory capacity individuals generate cues with greater efficiency and specificity in response to a retrieval prompt (Unsworth et al., 2012; Unsworth et al., 2013). It may be that participants with higher working memory capacity in our study were slower to relate memories because they were doing more retrieval work upfront. However, if high-working memory capacity participants were relatively slow to relay memories because of this retrieval work, we would expect high-working memory capacity to produce richer, more detailed memories, yet we did not see any association between idea density and working memory capacity. To investigate this possibility, future research could use retrieve-aloud task that requires participants to verbalize their retrieval processes (e.g., Mace et al., 2020; Mace et al., 2017; Uzer et al., 2012) and measure the number and types of cues generated. That said, high-working memory capacity individuals, who tend to generate more effective retrieval cues, have been found to be more impacted by mismatch in encoding and retrieval environments than low-working memory capacity individuals (Unsworth et al., 2011; Cokely et al., 2006; Delaney & Sahakyan, 2007; Aslan et al., 2010). In our task, where participants retrieved memories in a lab setting and were prompted for memories via an image on a computer screen, participants with high workingmemory capacity may have experienced a similar effect, leading to a disadvantage in retrieval speed.

While the speed with which the memories were retrieved was sensitive to effects of both bilingualism and working memory, propositional idea density was not sensitive to either effect. The null results could be related to the quality of our language samples, as many of our language samples fell below the minimum sample length of 60 words recommended by Alzheimer’s disease researchers (Ferguson et al., 2014). Our monolingual samples had a mean length near this guideline as calculated by both CPIDR 5 and DEPID, whereas our bilingual samples had a shorter mean length across calculators. In a more structured recall task, researchers could solicit longer language samples that may be more sensitive to effects of bilingualism and working memory. However, it may be that the idea density calculators are simply not sensitive enough to detect propositional idea density variation in non-clinical samples, even though the ranges of propositional idea density scores yielded by both calculators suggest significant variability within our sample (see Figure 1). While we focused on propositional idea density as the index of memory content in our primary analyses, in exploratory post-hoc analyses we re-operationalized memory quality in terms of memory specificity (Williams & Broadbent, 1986) and internal and external detail counts (Levine et al., 2002). We tested the possibility that these alternative measures of autobiographical memory retrieval would be more sensitive to differences in retrieval quality than propositional idea density. However, like propositional idea density, these measures did not yield effects of bilingualism, working memory, or their interaction (see Supplementary Materials for these analyses). Given the consistency across these measures of retrieval quality, the findings suggest that bilingualism and working memory affect retrieval speed without a trade off in retrieval quality.

In conclusion, the present study lays the foundation for future work focusing on bilingualism and autobiographical memory. While the mechanisms behind the effects of bilingualism we have uncovered remain to be examined, we have identified an existing and easyto-implement toolkit for measuring both the efficiency of autobiographical recall and the richness of the content that is recalled. To further investigate the underlying mechanism of bilingualism, future researchers should account for the vast individual differences within bilingual populations. Although we treated bilingual status as a dichotomous variable, we recognize that bilinguals vary greatly in their language use, proficiency, and dominance (Cox & Zlupko, 2019). In their study of individual differences in autobiographical memory in heritage speakers, Cox and Zlupko (2019) found that variables related to language use (e.g., country of birth, frequency of language birth) predicted switches out of the language used at test whereas proficiency across languages was not a significant predictor. In our study, bilingual participants were generally English-dominant, and while this was not true of all participants (see Table 1), our bilingual sample was too homogeneous to allow for analysing performance according to individual differences across bilinguals. The advantage to testing English-dominant bilinguals as opposed to English L2 immigrant samples (Marian & Neisser, 2000; Larsen et al., 2002; Marian & Kaushanskaya, 2007; Matsumoto & Stanny, 2006) or heritage speakers (Cox & Zlupko, 2019; Javier et al. 1993) is that our bilingual and monolingual participants were evenly matched in their English language proficiency, and characterized by similar acquisition trajectories of English, the target language of the study. Thus, unlike studies comparing bilinguals and monolinguals with distinct L1s, our study is not confounded by issues of target-language proficiency.

Understanding the mechanisms that underlie an apparent bilingual advantage in autobiographical memory retrieval will require tasks that probe memory context. One possible mechanism for the observed difference in retrieval speed is language context, as bilinguals experience variation in language context at memory encoding and retrieval that monolinguals do not experience (Marian & Neisser, 2000). Future research could compare bilingual and monolingual performance whilst manipulating or querying bilingual language context, assessing retrieval speed using a task similar to that employed by Marsh et al. (2015). Comparing bilingual performance across language contexts to baseline monolingual performance would make it possible to ask under what circumstances bilingual participants outperform monolinguals during autobiographical recall and investigate language context as the mechanism underlying the group differences observed here. In the meantime, our work suggests that bilingualism effects on the autobiographical memory system are nuanced and independent of working memory effects, with different operationalizations of autobiographical memory recall (e.g., retrieval speed, propositional idea density, emotionality) yielding different effects of bilingualism.