The Role of Determiners in the Processing of Gender Agreement Morphology by Heritage Speakers of Spanish

Abstract

This eye-tracking study examined how heritage speakers of Spanish process gender agreement morphology at a distance, focusing on the activation of the gender feature during sentence processing. Previous work is conceptually replicated and further extended by assessing (1) whether reduced sensitivity to gender agreement mismatches when another word intervenes between the head noun and its modifying adjective stems from weakened gender feature activation, (2) whether a gender-marked determiner enhances this activation, and (3) whether Age of Onset of Bilingualism (AOB) plays a role in this activation. Fifty-three English-dominant heritage speakers of Spanish and a comparison group of 32 Spanish-dominant monolingually raised speakers read sentences with and without gender agreement mismatches while their eye movements were monitored. Sentences contained mismatches in adjectives modified by the intensifier “muy” under two conditions: a No Cue condition (e.g., árboles muy altos/*altas) and a Cue condition with a gender-marked determiner (e.g., unos árboles muy altos/*altas). Statistical modeling of the eye-tracking data suggests similar effects for both groups in the No Cue condition, but AOB and proficiency modulated sensitivity for heritage speakers with a later AOB (4–6). Gender cues on the determiner (Cue condition) impacted the time course of agreement processing for all groups, the total time spent reading mismatches for all heritage speakers as a function of proficiency, and the rereading time for heritage speakers with a later AOB (4–9). We consider the role of Age of Onset of Bilingualism (AOB) and proficiency in morphosyntactic processing, feature retrieval, and cue facilitation in heritage language processing.

1. Introduction

During sentence processing, readers construct mental representations that encode morphosyntactic dependencies between words, including gender and number agreement. Consider the Spanish noun phrases in (1):

(1)	a.	Árboles	altos		(gender match)
		trees	tall-M-pl
		Tall trees
	b.	Árboles	*altas		(gender mismatch)
		trees	tall-F-pl
		Tall trees
	c.	Árboles	muy	*altas	(gender mismatch with intervening word)
		trees	very	tall-F-pl
		Very tall trees

In (1a), the adjective “altos” agrees morphologically in gender and number with the masculine plural noun “árboles.” In (1b), the use of the feminine form “altas” violates gender agreement. While gender agreement mismatches (1b) do not typically create meaning ambiguity, prior research has shown that both L1 and L2 readers are sensitive to gender agreement mismatches during online sentence processing (e.g., Keating, 2009). Moreover, when readers encounter gender agreement mismatches in distant contexts like 1c (e.g., when a word intervenes between a head noun and its adjective), online sensitivity decreases (Alemán Bañón et al., 2012, 2014; Keating, 2009, 2022, 2025).

Interestingly, for heritage speakers (HSs), the age at which majority language input occurs (Age of Onset of Bilingualism; AOB) influences the timing at which sensitivity to mismatches emerges during sentence processing. Keating (2022) reported that sensitivity to gender agreement mismatches emerged in reading measures associated with early processing (e.g., first fixation, gaze duration) for HSs of Spanish with a later AOB, while effects were observed in measures associated with later processing (e.g., go-past time) for HSs with an earlier AOB. Interestingly, when an intervening word increased the distance between the noun and the adjective (like in 1c), HSs and the monolingually raised comparison group demonstrated only later sensitivity, and AOB effects were no longer observed. Although findings were consistent with prior research that shows that agreement processing is weaker in contexts of distance (see Keating, 2009 for eye-tracking and Alemán Bañón et al., 2012, 2014 for ERP), it is unclear why agreement processing at a distance appears to be weaker.

Morphosyntactic cues—such as gender-marked determiners—may offset the effects of distance during agreement processing. Although previous research has shown that gender-marked determiners support morphosyntactic prediction during auditory predictive processing (Fuchs, 2022a, 2022b; Grüter et al., 2012; Hopp, 2013; Lew-Williams & Fernald, 2010), less is known about the role of determiners in the processing of agreement during reading. The present study addresses this gap by testing whether reduced sensitivity to gender agreement at a distance stem from weakened activation of the noun’s gender feature, and whether this activation can be enhanced through earlier morphosyntactic cues, such as determiners. We designed a sentence-level reading task with eye tracking that conceptually replicated Keating (2022, 2025). English-dominant HSs and Spanish-dominant monolingually raised speakers read sentences with and without gender agreement mismatches in adjectives modified by the intensifier “muy” under two conditions: a No Cue condition (e.g., árboles muy altos/*altas) and a Cue condition with a gender-marked determiner preceding the head noun (e.g., unos árboles muy altos/*altas). Statistical modeling of the eye tracking data revealed AOB effects in the time course of agreement processing in both conditions. Additionally, AOB often interacted with Spanish proficiency in some measures. Our results contribute to cue-based models of morphosyntactic processing in reading (Van Dyke & Lewis, 2003) and shed light on the interaction between distance, morphosyntactic cues, and bilingual experience in heritage language processing.

2. Literature Review

2.1. Gender in Spanish

Spanish classifies nouns by gender: feminine or masculine. Morphophonological patterns in noun endings offer probabilistic cues to gender: 99.9% of nouns ending in “o” (e.g., libro ‘book’, cuento ‘story’) are masculine, while 96.3% of nouns ending in “a” (e.g., cama ‘bed’, hoja ‘leaf’) are feminine (Teschner & Russell, 1984). These regularities aid in gender acquisition for child L1 (e.g., Hernández Pina, 1984; López-Ornat, 1997; Mariscal, 2009) and adult L2 (e.g., I. V. Alarcón, 2011; Montrul et al., 2008). Despite these transparent patterns, the Spanish gender system includes nontransparent forms—nouns ending in other vowels or consonants—or those with noncanonical endings, like feminine nouns ending in “o” (e.g., mano) and masculine nouns ending in “a” (e.g., problema). Determiner phrases (DPs) in Spanish are headed by determiners, but the inherent noun gender determines agreement: The definite determiners “el” and “los” must precede masculine nouns, while “la” and “las” must precede feminine nouns. Additionally, nouns may be followed by adjectives that must agree in gender with the noun.

The complexity of the Spanish grammatical gender system poses challenges in acquisition across populations, including L1, HSs and L2 learners (E. Bates et al., 1995; Taraban & Kempe, 1999; Montrul, 2011; I. Alarcón, 2020, 2021). Variability disproportionately affects nontransparent nouns (e.g., Alemán Bañón et al., 2014; Montrul et al., 2013) and feminine nouns more broadly, supporting theories of a default masculine gender in Spanish (Domínguez et al., 1999; Harris, 1991; McCarthy, 2008) and the view that gender assignment in Spanish is largely arbitrary (Harris, 1991; Kramer, 2015). Children typically acquire grammatical gender early—around age 4 (Fernández Fuertes et al., 2016; Montrul, 2004). Determiner–noun agreement emerges first (López-Ornat, 1997), followed by noun–adjective agreement (Pérez-Pereira, 1991). However, for speakers whose exposure to the language is reduced or inconsistent—HSs and L2 learners—grammatical gender presents a particular challenge, as its successful acquisition and application require sustained exposure and consistent use (Montrul, 2008; Tsimpli, 2014). This is because the successful acquisition and use of gender depend not only on memorizing arbitrary gender assignments, but also on reliably retrieving these features during real-time language processing (Hopp, 2018). Without sustained and rich input, the lexical representations of gender and the morphosyntactic rules that govern agreement may remain inconsistently applied. Because of the unique experience that HSs bring to the task of language processing, understanding what modulates variability in this population is essential for theories of bilingual processing. What follows is a synthesis of one of the variables (age; Montrul, 2008; Tsimpli, 2014) that explains some variability in HSs of Spanish.

2.2. Age Effects HSs

Heritage speakers (HSs) are bilinguals raised in homes in which a minority language is spoken within a society where a majority language permeates and reduces the input of the minority language (Valdés, 2000). Previous research has consistently reported that the morphosyntax of adult HSs tends to showcase more variability when compared to a monolingually raised baseline (I. V. Alarcón, 2011; Kupisch et al., 2013; Montrul et al., 2013). One of the variables that explains such variability is the age in which majority language input starts (Age of Onset of Bilingualism; AOB; Keating, 2022, 2025), which may coincide with minority language onset for simultaneous bilinguals (at birth), or include a wider range of ages for sequential bilinguals (Montrul, 2008; Tsimpli, 2014).

Two influential hypotheses make similar predictions for HSs as a function of AOB. Montrul’s (2008) age of onset hypothesis proposes that, if attrition occurs in early (pre-pubertal) bilingualism, it is likely to be more pronounced in simultaneous bilinguals—those exposed to both languages since birth—than in sequential bilinguals, who acquire the L1 first and the L2 later (Hypothesis 2, p. 60). Tsimpli’s (2014) timing hypothesis further distinguishes between macroparametric (broad, structural properties of the grammar) and microparametric (fine-grained, language-specific variations) features as a function of age/timing of acquisition, arguing that macroparameters are acquired early and robustly, while microparameters require more prolonged input and are, therefore, more vulnerable. For grammatical gender, this implies that children may grasp the existence of gender categories (macroparameter) early on (e.g., around age 4 for Spanish; Fernández Fuertes et al., 2016; Montrul, 2004), but the successful application of gender agreement rules (microparameter) relies on rich lexical and morphological input. This distinction aligns with acquisition studies showing earlier mastery of determiner–noun gender agreement compared to noun–adjective gender agreement (Pérez-Pereira, 1991; López-Ornat, 1997) and further predicts variability in the gender agreement system of bilinguals whose exposure to the majority language occurs early in life.

Some studies provide evidence for the role of AOB in the gender agreement system of child HSs. Anderson (1999) observed that a sequential bilingual child produced fewer gender errors than her simultaneous bilingual sibling during a longitudinal analysis of spontaneous oral production data, suggesting that the age at which majority language input occurred had an effect in the younger sibling. Montrul and Potowski (2007) tested child HSs (ages 6–11) by using a picture description task aimed at eliciting determiner–noun–adjective sequences orally. AOB effects emerged: Sequential bilinguals made fewer gender agreement errors in adjectives than simultaneous bilinguals, suggesting a stronger agreement system for children with longer minority language exposure. In contrast, Cuza and Pérez-Tattam (2016) tested Spanish-heritage children (ages 5–10) attending majority-language schooling. They found higher variability for noun-adjective pairs than for determiner-adjective pairs, but no effect of AOB. These findings suggest variability in the gender agreement system of child HSs as a function of AOB during oral production, but that the agreement system (microparameters) may be more vulnerable than assignment (macroparameters).

For adult HSs, the role of AOB has also been observed. Montrul et al. (2008) showed that HSs exposed to English before age 5 produced more gender errors than those exposed to English after age 5. However, in Montrul et al. (2013), task modality played a role: By using a series of tasks on gender agreement among adult HSs and adult L2 learners of Spanish, offline data suggested that HSs were more like L2 learners in performance on gender agreement, while online data suggested that they were more like the monolingually raised baseline, consistently showing greater sensitivity to agreement mismatches. These studies suggest that AOB is a variable that can explain variability in adult HSs, but that the use of more implicit online methods can provide fine-grained evidence for the role of AOB in HSs performance.

Keating (2022) tested the online sensitivity to gender agreement mismatches in HSs of Spanish who differed in AOB (0–3, 4–6, and 7–9) in a sentence-level task with eye tracking. HSs read sentences with and without gender agreement mismatches in adjectives that appeared immediately after the noun it modified (adjacent condition in 2a) and when a word intervened between the head noun and its adjective (nonadjacent condition in 2b).

(2)	a.	Antes	había	árboles	altos/*altas	en la montaña.
		before	there were	trees	tall-MASC/*FEM	in the mountain.
		Before there were tall trees in the mountain.
	b.	Antes	había	árboles	muy altos/*altas	en la montaña.
		before	there were	trees’	very tallMASC/*FEM	in the mountain.
		Before there were very tall trees in the mountain.

Eye-movement data showed an AOB effect for the adjacent condition (2a): Early sequential bilinguals were sensitive to gender agreement mismatches during early reading measures (e.g., first fixation duration), while sensitivity was present in later measures (e.g., go-past time) for simultaneous bilinguals, suggesting a subtle yet interesting AOB effect in the time course of gender agreement processing. Interestingly, when processing DPs with an intervening intensifier word in in the nonadjacent condition (2b), AOB effects disappeared; only HSs with a later (AOB 7–9) and with high proficiency in their heritage language showed late sensitivity. These results support the view that AOB shapes early integration of morphosyntactic information, but it also shows that distance weakens agreement processing for all readers.

To explain why the nonadjacent condition blurred AOB effects, Keating proposed that the distance manipulation made possible by adding intensifiers between nouns and adjectives (2b) may have reduced sensitivity to the establishment of agreement, as has been shown in previous research (Alemán Bañón et al., 2012, 2014). Next, we review findings on the effects of distance during agreement processing.

2.3. Distance Effects

Morphosyntactic processing is affected not only by grammatical knowledge but also by structural distance (when agreement relations span across phrase boundaries) and linear distance (when words are intervened by other words between agreeing elements, not necessarily required by the syntax). Alemán Bañón et al. (2014) examined how advanced L1 English learners of Spanish process number and gender agreement using event-related potentials (ERPs), with a particular focus on the role of structural and linear distance during processing. Crucially, the study compared agreement within a local domain (within a determiner phrase, linear distance) and across a more distant structure (across a verb phrase, structural distance). The study found that the establishment of agreement within phrases elicited stronger ERP effects (larger P600, an ERP component typically observed as a positive-going waveform that peaks around 600 milliseconds after the onset of a critical word) than those across phrases, even for grammatical sentences. This suggests that locality facilitates agreement processing for both L1 and L2 speakers. Conversely, distance may weaken agreement processing in both populations.

Similarly, Keating (2009) found that advanced L2 readers only detected gender agreement mismatches when adjectives appeared within the determiner phrase (DP) in a reading task with eye tracking. Sensitivity dropped when dependencies spanned a verb phrase (VP) or clause (CP), unlike L1 readers who showed robust sensitivity across all distances. For HSs, recall that Keating (2022, 2025) found that linear distance similarly reduced agreement sensitivity. In nonadjacent contexts like (2b), readers may fail to retrieve or maintain the gender feature of the noun (árboles), weakening agreement computation at the adjective (altas). Alternatively, the reader may retrieve the gender feature but may not hold it in memory for later agreement processing. Processing the adjective would then depend on the retrieval and maintenance of the previously activated noun gender.

If sentences had included an extra cue for gender, could sensitivity be enhanced? Spanish determiners seem to be a very reliable cue for the acquisition and processing of Spanish grammatical gender. Evidence from auditory predictive processing suggests that gender-marked determiners can support morphosyntactic processing (Lew-Williams & Fernald, 2010; Grüter et al., 2012; Hopp, 2013). Additionally, the Spanish L1 acquisition literature reports that Spanish-acquiring children produce a prenominal vowel in the one-word stage, such as e pie “the foot”, a queca “a doll”, u fo “a flower” (López-Ornat, 1997). This vowel has been argued to represent a ‘protodeterminer’ in the child grammar (Lleó, 1998), given that it coincides with the vowels of gender-marked determiners (el, la, un, una). This raises the question of whether earlier gender cues—such as determiners—could support feature activation and maintenance during sentence reading in heritage speakers. The next section discusses previous research on cue-based processing.

2.4. Cue-Based Processing

Several studies report that cues involving phonology, verb semantics, category structures, and morphosyntax can be used during processing (e.g., Altmann & Kamide, 1999; DeLong et al., 2014; Luke & Christianson, 2016). For instance, Tanner and Bulkes (2015) investigated the effects of a quantification cue in the sensitivity to subject-verb agreement in English. Agreement was manipulated via unquantified (the cookies taste/*tastes) and quantified (most cookies taste/*tastes) cues. By using time-locked ERP components, agreement mismatches generated a P600 for singular marked verbs; however, the P600 was significantly larger for the agreement mismatches when a quantified cue was provided (e.g., most cookies taste/*tastes), suggesting that the quantified cue increased sensitivity to agreement mismatches.

Van Dyke and Lewis (2003) proposed that syntactic processing relies on a cue-based retrieval mechanism in memory, where agreement dependencies are established via retrieval cues that must successfully identify relevant constituents based on features. Therefore, distance and interference weaken retrieval success, delaying the detection of agreement mismatches unless robust cues are present at retrieval. This account has direct implications for the current study: Gender-marked determiners may serve as retrieval-enhancing cues, allowing readers to more efficiently resolve gender agreement across intervening material.

To date, no study has examined whether HSs use gender-marked determiner cues for gender agreement processing during reading. However, emerging evidence from auditory modalities suggests that HSs can indeed use determiners as morphosyntactic cues to support comprehension. In an auditory comprehension study, Vergara and Socarrás (2020) tested the attachment preference of Spanish heritage speakers and L2 learners involving relative clauses with gender-mismatched nouns. Participants were asked to determine which noun was modified by an adjective in a relative clause, relying on gender agreement for disambiguation. While grammatical gender and noun canonicity were not significant predictors in their final model, participants were descriptively more accurate when processing non-canonical nouns that required reliance on gender-marked determiners. This trend suggests that HSs can use determiners as morphosyntactic cues to support auditory sentence comprehension.

Additional support comes from Fuchs (2022a), who tested predictive gender processing in Spanish HSs using a visual-world with eye-tracking paradigm. Participants were shown two images while they heard sentences auditorily. Stimuli were presented in two conditions: gender-match trials, where the two objects shared the same gender, and gender mismatch trials, where objects differed in gender. When participants heard “Dónde está el…” or “Dónde está la…” (“where is the…”), their eye movements were recorded with the premise that, at the onset of the gender-marked determiner (el or la), participants would integrate this information to predictively fixate at the target object. HSs used gender cues in determiners to predictively fixate on the target. However, interpretation of these findings is limited by sample composition: 12 of the 21 HS participants had lived in a Spanish-speaking country before moving to the U.S., while only 9 were U.S.-born. Because AOB was not explicitly analyzed, it remains unclear whether early majority language exposure modulated cue-based predictive processing in this task. Moreover, it is still unknown whether HSs use determiner cues to support gender agreement processing at a distance—i.e., when the agreeing adjective is non-adjacent, during sentence reading.

The present study addresses these gaps by investigating whether Spanish HSs use gender-marked determiners to facilitate gender agreement processing during reading, particularly in distant contexts, as predictive processing studies suggest (Fuchs, 2022a). Using eye-tracking, we designed a sentence-level reading task in which participants read sentences in two conditions: a No Cue condition that replicates Keating’s (2022, 2025) nonadjacent condition (e.g., árboles muy altos/*altas), and a Cue condition in which a gender-marked determiner provides an earlier cue (e.g., unos árboles muy altos/*altas). We also examined the effect of AOB, allowing us to test whether AOB plays a role in cue integration during reading. In doing so, this study contributes new insights into the interplay between morphosyntactic cues, processing demands due to linear distance, and individual bilingual experience in heritage language processing.

In sum, the Spanish grammatical gender is a complex and variable system that typically requires sustained and consistent input to master. As such, AOB is a critical factor shaping long-term language outcomes among HSs, influencing both morphosyntactic competence and processing strategies in adulthood. Because prior research shows that linear distance between agreeing elements (e.g., noun and adjectives) can delay sensitivity to agreement mismatches, we test whether earlier cues—such as gender-marked determiners—can mitigate this challenge by boosting agreement saliency and reducing processing demands during reading, as determiners have been shown to serve as cues during predictive processing (Fuchs, 2022a).

2.5. This Study

Given this background, the research questions that guided the present study were as follows:

• RQ1: To what extent is sensitivity to gender agreement mismatches delayed when processing adjectives that appear distant from the head noun they modify by heritage and nonheritage speakers of Spanish? Does AOB play a role in HSs?
• RQ2: To what extent can the presence of an earlier gender cue (e.g., a determiner) enhance the sensitivity to gender agreement mismatches when processing adjectives that appear distant from the head noun they modify by heritage and nonheritage speakers of Spanish? Does AOB play a role in HSs?

RQ1 replicates the nonadjacent condition in Keating’s (2022, 2025) experiment, in which gender agreement mismatches in adjectives occurred when a word intervened between a noun and its modifying adjective (e.g., árboles muy altos/*altas). We examined the time course of gender agreement processing at the adjective by analyzing reading measures associated with early (First Fixation, Gaze Duration) and late (Go-Past, Rereading, Total Time) stages of morphosyntactic processing. Consistent with Keating (2022, 2025), we expect delayed sensitivity (e.g., in later measures) to distant gender agreement mismatches by both HSs and Spanish-dominant monolingually raised participants. AOB is not expected to modulate sensitivity in this condition, but proficiency, as measured by a score obtained from a modified version of the DELE test (Montrul, 2012), is expected to have an effect in later measures.

RQ2 extends Keating’s design by adding a condition where a gender-marked determiner cue precedes the noun, the intervening word, and the adjective (e.g., unos árboles muy altos/*altas). We examined the time course of gender agreement processing by analyzing reading measures associated with early and late stages of agreement processing at the adjective. If a gender-marked determiner cue boosts the activation and maintenance of the gender feature in memory, we expect the effect of grammaticality to emerge in early reading measures (First Fixation and/or Gaze Duration), with persistence in later measures (Go-Past, Rereading, Total Time). This outcome would suggest that determiners—acquired early in L1 acquisition (López-Ornat, 1997; Lleó, 1998)—serve as strong cue for gender agreement processing as found in predictive processing studies in HSs (Fuchs, 2022a). However, AOB may modulate this integration. If so, HSs with a later AOB may show more immediate effects, while HSs with an earlier AOB may show no effects or only later effects. Alternatively, if the gender-marked determiner cue does not enhance agreement processing at a distance, the results may resemble those in the No Cue condition or show no significant effects across reading measures. This would suggest that distance itself, rather than feature activation or cue presence, determines the memory demands in the time course of gender agreement processing in distant contexts.

3. Materials and Methods

3.1. Participants

All participants were recruited via a flier distributed in a large midwestern university. Fifty-three English-dominant HSs of Spanish (M_age = 20.68, 18–27, 42 females and 11 males) participated in the current study, all of whom reported active enrollment (48 undergraduates, five graduates). The comparison group was comprised of 32 Spanish-dominant monolingually raised Spanish speakers (M_age = 27.16, 19–38, 19 females, 13 males), mostly graduate students (n = 31, except one undergraduate) at the same university. The monolingually raised speakers were all born in a variety of Spanish-speaking countries (Colombia, n = 7; Spain, n = 6; Argentina, n = 3; Mexico n = 2; Venezuela n = 4; Perú, n = 2; Ecuador n = 2; and at least one participant from Chile, Costa Rica, Ecuador, El Salvador, Honduras, and Nicaragua) and grew up in households where only Spanish was spoken. Their average length of residence in the US was 3.64 years (SD = 1.81, 1–8). They all reported being late L2 English acquirers (age > 11).

All participants were screened via a language background questionnaire that served to classify them into their respective groups based on the place and age of onset of Spanish and English acquisition. Crucially, onset of Spanish was at birth for all speakers, while onset of English (AOB) occurred before puberty in a range of ages (0–9) and in the United States for the HSs, and after puberty and in a Spanish-speaking country for the comparison group. Consistent with Keating (2022, 2025), the HSs were divided into three cohorts based on their AOB: 0–3 (n = 18), 4–6 (n = 22), and 7–9 (n = 13). For all participants, proficiency in English was assessed via self-ratings, while proficiency in Spanish was assessed via self-ratings in addition to the modified version of the DELE test, a 50-item written proficiency test (Montrul, 2012) also used in Keating (2022, 2025). A one-way ANOVA confirmed no statistically significant difference in proficiency scores across HSs subgroups, F (2, 50) = 0.39, p = 0.682.

In addition, all participants responded to the Bilingual Language Profile (BLP) as a measure of dominance (Birdsong et al., 2012). In the BLP, dominance scores are obtained by subtracting the score of one language to the score of the other (scores range 1–250 in each language). More negative scores indicate Spanish dominance, while more positive scores indicate English dominance. Figure 1 displays the scores for the heritage and comparison groups, which shows that the HSs were more English dominant than the comparison group (overall more positive scores), except for one HS whose dominance score was −76.84 (AOB 5, DELE 42/50). A one-way ANOVA confirmed no statistically significant differences in dominance across AOB cohorts, F (2, 50) = 0.97, p = 0.388. Descriptive statistics for proficiency (self-reported and DELE) scores, as well as dominance scores for both groups overall, and for the HSs cohorts, are provided in

Table 1. Participants’ scores in proficiency and dominance.

Measure	Heritage Speakers n = 53	Monolingually Raised (Comparison) n = 32
Proficiency in English (Self-Reported, 1 to 5)	All (n = 53) = 4.83, SD 0.38, 4–5 AOB 0–3 (n = 18) = 4.94, SD 0.24, 4–5 AOB 4–6 (n = 22) = 4.91, SD 0.29, 4–5 AOB 7–9 (n = 13) = 4.54, SD 0.52, 4–5	4.38, SD 0.71, 3–5
Proficiency in Spanish (Self-Reported, 1 to 5)	All (n = 53) = 4.26, SD 0.68, 3–5 AOB 0–3 (n = 18) = 4.44, SD 0.7, 3–5 AOB 4–6 (n = 22) = 4.23, SD 0.69, 3–5 AOB 7–9 (n = 13) = 4.08, SD = 0.64, 3–5	5, SD 0, 5–5
Proficiency in Spanish (DELE, 1 to 50)	All (n = 53) = 41.34, SD 4.5, 31–49 AOB 0–3 (n = 18) = 41.56, SD 4.8, 31–48 AOB 4–6 (n = 22) = 41.73, SD 4, 34–48 AOB 7–9 (n = 13) = 40.38, SD 5.06, 31–49	47.94, SD 1.52, 44–50
Dominance Score (BLP, −218 to +218)	All (n = 53) = 29.95, SD 30.43, −76.18–84.73 AOB 0–3 (n = 18) = 38.05, SD 25.23, 2.45–83.73 AOB 4–6 (n = 22) = 25.61, SD 3.66, −76–84.73 AOB 7–9 (n = 13) = 26.08, SD 29.34, −17.42–80.1	−86.88 SD 25.78, −135.668–−21.516

.

3.2. Materials

A sentence-level reading task was designed in SR Research Experiment Builder (version 2.3.1; SR Research, 2020). Thirty-two experimental sentences were constructed to replicate Keating (2022, 2025) by using the same 32 critical nouns (half feminine, half masculine, all gender nontransparent, and inanimate) and 16 adjectives (each appearing once with a feminine and once with a masculine noun, once in each condition) from the original study. These nouns were initially selected from the SUBTLEX-ESP corpus (Cuetos et al., 2011) and were comparable in word length and log10 frequency (Keating, 2022, p. 1183). Adjectives ranged in length from three to five syllables (M = 2.44, SD = 0.51, 95% CI [2.16, 2.71]), with log10 frequencies ranging from 3.18 to 4.58 (M = 3.92, SD = 0.36, 95% CI [3.72, 4.11]), though they displayed unequal variances in the analysis of length in characters (Keating, 2022, p. 1183). Using these nouns and adjectives, a new set of stimulus sentences was created to target the research questions in the present study. First, we constructed a condition in which the adjective was preceded by an intensifier (muy), intervening between the noun and adjective (e.g., árboles muy altos/*altas), as in the non-adjacent condition in Keating (2022, 2025). Unlike in the original study, which alternated between “muy” and “más”, we opted to use “muy” across all items and conditions to maintain consistency in structure and lexical content across the board. This constituted our “No Cue” condition. In the “Cue” condition, noun phrases were headed by a gender-marked indefinite determiner (e.g., unos árboles muy altos/*altas). Importantly, we did not manipulate the grammaticality of the determiner. Because adjectives displayed unequal variance in length, these were controlled across conditions, so that the same unequal variance was present in both Cue and No Cue conditions as adjectives are used twice, once per each condition.

Indefinite determiners were preferred over definite ones for several reasons. First, native speaker consultants rated sentences with “unos/as” and “muy” as more natural and grammatical than alternatives with “los/las” and “más.” Second, “unos/as” allowed us to preserve the syntactic integrity of transitive constructions across both conditions (with and without determiners) without requiring structural changes or compromising naturalness. For example, constructions like “el granjero cortó (unos) árboles muy altos” were consistently judged to be more acceptable than those using “más” or definite determiners. This approach ensured that the only manipulated variables were cue presence and gender agreement. Native speakers were consulted throughout the norming process, and final items reflect their feedback on grammaticality and naturalness. All experimental sentences were between 9 and 14 words in length. As shown in

Table 2. Sample item for the sentence-level reading task.

	Gender Match	Gender Mismatch
No Cue	El granjero cortó árboles muy altos en la cima de la montaña.	El granjero cortó árboles muy *altas en la cima de la montaña.
Cue	El granjero cortó unos árboles muy altos en la cima de la montaña.	El granjero cortó unos árboles muy *altas en la cima de la montaña.
The Farmer cut (the) very tall trees in the top of the mountain

, our 2 × 2 design manipulated grammaticality (match vs. mismatch) and cue presence (No Cue vs. Cue) via gender agreement in the adjective. As in Keating (2022, 2025), all critical nouns were gender non-transparent and appeared in plural form, with adjectives appearing once per gender. Appendix A provides the full list of the 32 experimental items across conditions.

Noncritical items included 32 distractors that were part of another experiment (not reported here), which manipulated the grammaticality of verbal agreement, as well as the frequency of the verb (see 4 for a sample sentence). In addition, 64 sentences, all grammatical, served as fillers (see 5 for a sample sentence). The experimental items and distractors were divided into four lists and pseudorandomized within each list. This was done to ensure that a given participant only saw one version of each experimental and distractor item, while all participants read all fillers, consistent with standards for sentence processing research in bilingualism using the anomaly detection paradigm (see Godfroid, 2020; Keating & Jegerski, 2015). Additionally, because adjectives were repeated, they were divided by blocks to ensure that each adjective was read only once per block. Each participant read a total of 132 sentences (four practice, 32 experimental, 32 distractors, and 64 fillers) divided into three blocks (a practice block with four sentences followed by two blocks with 64 sentences each). A binary choice comprehension question followed all stimulus sentences (practice, experimental, distractors, and fillers). For instance, a question like (6) would follow a sentence like in our sample experimental sentence (Table 2), and would ask participants to select between options A or B. None of the questions specifically targeted the critical areas of the experimental sentences; rather, they assessed overall comprehension of the sentence.

(4)	La pregunta/Las preguntas que hizo la maestra comenzó/entabló el debate entre los alumnos. The question/questions that the teacher asked started the debate among students
(5)	El edificio más alto de Latinoamérica es la Gran Torre Santiago. The tallest building in Latin America is the Gran Torre Santiago.
(6)	¿Quién hizo esto?
	(A) El granjero.	(B) El gobierno.

3.3. General Procedure

Participants completed all tasks individually in a lab setting. Consent was obtained upon arrival. Participants proceeded to the reading task first, for which they were told that the purpose was to test their reading comprehension via questions after each sentence. Performing the reading task first allowed for participants to not be primed by the content of the subsequent tasks. A desktop-mounted Eyelink 1000 eye tracker was used (SR Research, 2005), with chin and forehead rests, a sampling rate of 1000 Hz, and tracking of only the right eye. Participants sat 39 inches from the display (22-inch ViewSonic monitor). A practice trial of four sentences preceded two blocks of experimental sentences (64 sentences each block), with a 10 min break offered to participants between experimental blocks. Calibration was performed before each block, after breaks, and as needed when drifting occurred (e.g., when eye gaze recording deviates its location upwards or downwards from the sentence). A nine-point grid was used as calibration, which was validated to ensure 0.5 degrees maximum error. Stimulus sentences and comprehension questions were presented in black, 24-point Monospace font on a white background, with each sentence presented as a single line of text. Participants proceeded from one screen to the next by pressing a button on a Microsoft Sidewinder game controller, which was also used to respond to all comprehension questions (buttons were marked “A” and “B”). The reading task took about 35–45 min to complete.

After the reading task, participants were debriefed about the experiments and their experiences with Spanish and English, followed by a working memory task (not reported here). Next, participants completed a language background questionnaire, the DELE proficiency test, and the Bilingual Language Profile for dominance (Birdsong et al., 2012). Finally, participants completed a vocabulary task to assess their knowledge of the meaning and grammatical gender of the target nouns and adjectives. All testing took approximately 2 h to complete in a single session. All participants were compensated for their time.

3.4. Data Analysis

As in Keating (2022), we used eye tracking to investigate moment-by-moment processing at the word level via a combination of early and late reading measures. These measures were extracted at the adjective to assess the time course of gender agreement processing in our study. Early measures are thought to reflect more automatic processing (e.g., lexical access, word recognition) while late measures are thought to reflect more cognitive-demanding processes (e.g., reanalysis, rereading). Early measures included First fixation duration (time in ms spent the first time a word is fixated), and gaze duration (sum of all durations of all fixations made on the target word from the first time the word is fixated until exiting to the right). Late measures included rereading (sum of all fixations made on the target word after gaze duration was complete), go-past time (first fixation duration plus the total duration of all other fixations made until exiting to the right, including refixations of words prior to the critical word), and total time (summed duration of all fixations made on the critical word).

Linear mixed-effects models were fit using the lme4 package (D. Bates et al., 2015) in R (R Core Team, 2023) to examine the effects of grammaticality (gender agreement match/mismatch) at the adjective word for each reading measure. Following Keating (2022), the group variable was specified as a four-level factor reflecting AOB cohorts: 0–3 (simultaneous), 4–6 (early sequential), 7–9 (late sequential), and >10 (comparison group). Importantly, this variable received reversed Helmert contrast coding to capture theoretically motivated comparisons: Group 1 compared the comparison group vs. all heritage speakers. Group 2 compared late sequential bilinguals (AOB 7–9) vs. all early bilinguals (AOB 0–6), and Group 3 compared early sequential (AOB 4–6) vs. simultaneous bilinguals (AOB 0–3). Initial models employed a maximal random effects structure (Barr et al., 2013), including random intercepts for subjects and items, as well as random slopes for grammaticality and group where justified. However, like in Keating (2022), some of these maximal models failed to converge or resulted in singular fits. As a result, model simplification was implemented in a stepwise manner, first reducing random slopes by item, then by subject. At each step, model comparisons were evaluated using the Akaike Information Criterion (AIC), with the best-fitting, convergent model retained for interpretation. All fixed effects included grammaticality (match vs. mismatch), group (reversed Helmert-coded), and centered continuous covariates: adjective length, trial order, and DELE proficiency score. Final models were fit using maximum likelihood to enable direct model comparisons. Because adjective length displayed unequal variances in the analysis of length in characters (Keating, 2022, p. 1183), and following Keating (2022, 2025), adjective length was included as a covariate in all final statistical models, although the effects of adjective length are not interpreted here as they are beyond the scope of this paper. Pairwise comparisons were conducted using the emmeans package, with Bonferroni-corrected p-values to control for multiple comparisons. Appendix B provides all model outputs for each reading measure per condition. All R codes can be found in Supplementary Materials.

Prior to statistical analysis, fixations shorter than 80 ms were merged with a neighboring fixation that measured at least 80 ms if the two were located within 1° of visual angle. Otherwise, fixations that were shorter than 80 ms or greater than 800 ms were discarded, affecting 8.2% of the data. Both procedures were performed in EyeLink Data Viewer (SR Research Ltd. Version 4.4.1). Next, winsorization was applied to the remaining data for each measure by group and condition, meaning that values 3 SDs greater than respective group means were trimmed to the cutoff (3 SDs above the mean) value. This affected 1.48% of data in all conditions/measures. Skips accounted for loss of 6.18% of the data. Trials that corresponded to incorrect gender assignment during the vocabulary test (selecting “el” or “la” for each noun post-testing) were excluded and accounted for 1.28% of trials. All in all, 84.34% of the data remained for statistical analysis. Log transformation was applied to each eye movement measure to avoid positive skew.

4. Results

4.1. No Cue Condition

Descriptive statistics for first fixation duration, gaze duration, total time, rereading, and go-past at the critical adjective in the No Cue condition are displayed in Figure 2, Figure 3, Figure 4, Figure 5 and Figure 6, respectively.

Recall that we used reversed Helmert contrast coding for all Groups in our analysis (Section 3.4, second paragraph). Group 3 compares simultaneous (AOB 0–3) to early sequential (AOB 4–6) only. Group 2 compares late AOB (7–9) to all earlier AOB (0–6), in the same way that Group 1 compares all late AOB (monolingually raised) to all HSs (AOB 0–9).

In early measures, statistical modeling (Appendix B) revealed no significant effects of Grammaticality (or relevant interactions) in any of the early measures (first fixation duration or gaze duration), indicating that participants did not detect gender agreement mismatches in early stages of agreement processing in the No Cue condition.

In later measures, total time revealed an effect of Grammaticality (b = −0.021, 95% CI [0.001, 0.013], SE = 0.052, t = 4.02, p < 0.001), and Proficiency (b = −0.024, 95% CI [–0.041, −0.006], SE = 0.009, t = 101.8, p = 0.011). A Group 3 × Grammaticality interaction was significant (Group 3: the contrast between AOB 4–6 and AOB 0–3 HSs; b = 0.130, 95% CI [0.034, 0.227], SE = 0.049, t = 2.66, p = 0.008). Pairwise comparisons confirmed that the early sequential heritage speakers (AOB 4–6) exhibited a significant difference between reading match and mismatch adjectives (b = −0.367, 95% CI [−0.5840, −0.1500], SE = 0.070, t = 87.5, p = 0.0001) but the simultaneous heritage speakers (AOB 0–3) did not (b = −0.109, 95% CI [−0.3375, 0.1186], SE = 0.073, t = −1.492, p = 0.809). A Grammaticality x Proficiency interaction was also significant (Group 3: the contrast between AOB 4–6 and AOB 0–3 HSs, b = 0.256, 95% CI [0.030, 0.482], SE = 0.103, t = 2.49, p = 0.027), which indicated that the grammaticality effect was increased with proficiency. A Grammaticality × Group 3 × Proficiency three-way interaction also emerged (Group 3: the contrast between AOB 4–6 and AOB 0–3 HSs, b = 0.230, 95% CI [0.045, 0.416], SE = 0.092, t = 2.43, p = 0.016). To explore the three-way interaction within each group, two LMMs were fitted with Grammaticality and Proficiency as fixed factors predicting log total times for the two AOB groups in question (0–3 and 4–6). While the interaction was not significant for the AOB 0–3 group (b = 0.009, 95% CI [−0.010, 0.029], SE = 0.010, t = 0.97, p = 0.118), it was significant for the AOB 4–6 group (b = 0.058, 95% CI [0.024, 0.091], SE = 0.019, t = 3.374, p = 0.003). Figure 7 further displays the interaction: As proficiency increased, so did sensitivity to agreement mismatches for early sequential heritage speakers (AOB 4–6).

In rereading, an effect of Grammaticality emerged (b = 0.138, 95% CI [0.006, 0.271], SE = 0.067, t = 0.058, p = 0.042). Higher proficiency was associated with reduced rereading times (b = −0.027, 95% CI [–0.048, −0.005], SE = 0.011, t = −0.026, p = 0.005). A Grammaticality × Proficiency interaction was significant (b = 0.029, 95% CI [0.002, 0.056], SE = 0.014, t = 0.058, p = 0.039), which indicated that the grammaticality effect was increased with proficiency.

Finally, in go-past, Proficiency was significant, with higher proficiency associated with faster reading times (b = −0.020, 95% CI [–0.035, −0.004], SE = 0.009, t = 105.7, p = 0.007). A Grammaticality × Proficiency interaction (b = 0.025, 95% CI [0.005, 0.047], SE = 0.011, t = 243.0, p = 0.020) indicated that the proficiency effect was stronger in the mismatch condition. A Grammaticality × Proficiency × Group 3 three-way interaction was also significant (Group 3: the contrast between AOB 4–6 and AOB 0–3 HSs, b = 0.023, 95% CI [0.004, 0.043], SE = 0.096, t = 2.31, p = 0.021). To explore the three-way interaction within each group, two LMMs were fitted with Grammaticality and Proficiency as fixed factors predicting log go-past times for each AOB group (0–3 and 4–6). The interaction was not significant for the AOB 0–3 group (b = 0.008, 95% CI [−0.025, 0.023], SE = 0.012, t = −0.067, p = 0.947), but it was significant for the AOB 4–6 group (b = 0.047, 95% CI [0.005, 0.047], SE = 0.017, t = 2.695, p = 0.015). As displayed in Figure 8, sensitivity to agreement mismatches increased with proficiency for the latter group in go-past.

4.2. Cue Condition

Descriptive statistics for first fixation duration, gaze duration, total time, rereading, and go-past at the critical adjective in the Cue condition can be found in Figure 9, Figure 10, Figure 11, Figure 12 and Figure 13, respectively.

Recall that we used reversed Helmert contrast coding for all groups in our analysis (Section 3.4, second paragraph). Group 3 compares simultaneous (AOB 0–3) to early sequential (AOB 4–6) only. Group 2 compares late AOB (7–9) to all earlier AOB (0–6), in the same way that Group 1 compares all late AOB (monolingually raised) to all HSs (AOB 0–9).

For early measures, statistical modeling (Appendix B) revealed that the presence of a gender-marked determiner resulted in earlier detection of mismatched adjectives for all groups. Although no effects were observed in first fixation duration, agreement mismatches led to longer gaze durations (b = 0.166, 95% CI [0.017, 0.316], SE = 0.076, t = 186.0, p = 0.030). Higher proficiency was associated with shorter gaze durations, indicating more efficient early processing (b = −0.024, 95% CI [–0.040, −0.008], SE = 0.008, t = 254.1, p = 0.003). However, no further group effects or interactions were observed.

In later measures, total time revealed a significant effect of Grammaticality (b = 0.038, 95% CI [0.275, 0–049], SE = 0.052, t = 6.951, p = 0.001) and Proficiency (b = −0.031, 95% CI [–0.041, −0.005], SE = 0.009, t = −3.394, p = 0.001). There was also a significant Grammaticality × Proficiency interaction (b = 0.027, 95% CI [0.008, 0.047], SE = 0.011, t = 2.699, p = 0.008) as well as a three-way interaction of Grammaticality × Proficiency × Group 2 (Group 2: the contrast between AOB 7–9 with all earlier AOB 0–3 and 4–6 HSs; b = 0.012, 95% CI [−0.023, −0.004], SE = 0–006, t = −2.037, p = 0.04). To explore the three-way interaction, two LMMs were fitted with Grammaticality and Proficiency as fixed factors predicting sensitivity in total time between the two AOB groups in question (AOB 7–9 vs. all earlier 0–6 HSs). The models confirmed that the interaction was significant for both AOB groups (0–6 b = 0.031, 95% CI [0.014, 0.052], SE = 0.009, t = 3.425, p = 0.001; 7–9 b = 0.054, 95% CI [−0.023, −0.004], SE = 0.015, t = 3.587, p = 0.001), suggesting that sensitivity to agreement mismatches in total time increased as proficiency increased for all heritage speakers, as displayed in Figure 14.

In rereading, higher Spanish proficiency was associated with shorter rereading times (b = −0.036, 95% CI [–0.058, −0.014], SE = 0.019, t = −1.87, p = 0.012). A significant main effect of Group 1—the contrast between monolingually raised speakers and all heritage speakers—was observed (b = 0.114, 95% CI [0.034, 0.194], SE = 0.042, t = 2.69, p = 0.008), indicating that the comparison group exhibited longer rereading times overall. A significant Grammaticality × Group 1 interaction (Group 1: the contrast between monolingually raised speakers and all heritage speakers, b = −0.124, 95% CI [–0.209, −0.039], SE = 0.044, t = −2.85, p = 0.004) further revealed that the comparison group processed agreement differently to the heritage speakers. However, pairwise comparisons within the comparison group showed no significant difference between match and mismatch adjectives (b = 0.292, 95% CI [–0.225, 0.808], SE = 0.170, t = 1.720, p = 0.674) while the heritage speakers in the 4–6 and 7–9 AOB groups did (AOB 4–6: b = −0.230, 95% CI [–0.452, −0.008], SE = 0.073, t = −3.15, p = 0.037; AOB 7–9: b = −0.337, 95% CI [–0.667, −0.006], SE = 0.109, t = −3.10, p = 0.042). The simultaneous AOB 0–3 heritage speakers did not exhibit a significant difference between match and mismatch conditions in rereading (b = −0.049, 95% CI [−0.289, 0.190], SE = 0.078, t = −0.626, p = 0.998).

For go-past, group differences neutralized: All participants showed a significantly longer go-past time when reading agreement regardless of AOB (b = 0.137, 95% CI [0.028, 0.246], SE = 0.042, t = 3.25, p = 0.001), although higher Spanish proficiency was associated with a significantly shorter go-past time (b = −0.025, 95% CI [–0.041, −0.008], SE = 0.009, t = −2.79, p = 0.006).

Taken together, results from the No Cue condition revealed an overall delayed detection of gender agreement mismatches for all groups. Late detection was modulated by Spanish proficiency and AOB. Specifically, early sequential HSs (AOB 4–6) exhibited greater sensitivity to mismatches than simultaneous HSs (AOB 0–3), as evidenced by a significant interaction between Grammaticality and Group, and confirmed through pairwise comparisons. Importantly, proficiency consistently enhanced sensitivity to mismatches in this group, as revealed by significant Grammaticality × Proficiency interactions in total time and go-past, and a three-way interaction confirming that proficiency modulated sensitivity more strongly for the 4–6 group than for the 0–3 group. In the Cue condition, the inclusion of a gender-marked determiner facilitated earlier detection of gender agreement mismatches for all participants, as evidenced by longer gaze durations in the mismatch condition. Additionally, in total time, reading times increased with proficiency for both early and late AOB groups (0–6 and 7–9), suggesting that cue-based facilitation can also ease up processing demands across AOB backgrounds when proficiency is high. In rereading, monolingually raised speakers exhibited longer durations overall, yet only heritage speakers with AOB 4–6 and 7–9, but not 0–3, showed reliable sensitivity to agreement mismatches in this measure. Notably, for go-past, group differences converged: All participants showed increased processing costs for mismatches, regardless of AOB, while higher proficiency predicted shorter reading times overall. This pattern suggests that while earlier cues can accelerate detection, later integration and reanalysis remain sensitive to both AOB and proficiency in the heritage language. We discuss our interpretations of these findings in the next section.

5. Discussion

The present study investigated how English-dominant HSs of Spanish and a comparison group of monolingually raised Spanish-dominant speakers process gender agreement morphology at a distance, focusing on (1) the time course of gender agreement processing in contexts of linear distance, (2) the role of gender-marked determiners as morphosyntactic cues for agreement processing in contexts of linear distance, and (3) the role of Age of Onset of Bilingualism (AOB) in this process. What follows is a discussion of the findings of the present study.

5.1. Delayed Sensitivity at a Distance

In the No Cue condition, all participants failed to detect gender agreement mismatches in early measures (first fixation and gaze duration) but they did so in later measures (rereading, total time, and go-past), extending previous research on morphosyntactic processing at a distance (e.g., Alemán Bañón et al., 2014; Keating, 2022, 2025). The effects of grammaticality in later measures suggest that the detection of gender agreement mismatches was not automatic, but rather required the parser to revisit the word at some point (rereading), to spend longer time at this word compared to its grammatical version (total time), and to regress to earlier parts of the sentence (go-past) after initially processing the word.

We interpret this delayed sensitivity through the lens of cue-based retrieval and activation decay (Van Dyke & Lewis, 2003). Processing relies on cue-based retrieval mechanisms in memory, where agreement dependencies are established via retrieval cues that must successfully identify relevant constituents based on features. Successful agreement processing requires that the gender feature of the head noun be retrieved and maintained in memory until the eyes encounter the adjective word. The presence of an additional word between the noun and the adjective, e.g., the intensifier “muy” in our design, may cause the activation of the gender feature of the previously read noun to decay, which causes the parser to revisit earlier segments of the sentence. Importantly, because the intensifier “muy” does not carry gender features, such a word is unlikely to interfere with cue matching. Rather, our data suggest feature activation decay, where distance alone—without interference—can weaken cue matching success, delaying the detection of agreement mismatches online.

5.2. Determiners as Cues for Agreement Processing

Results from the Cue condition show that gender-marked determiners facilitated earlier detection of agreement mismatches for all speakers. Significant grammaticality effects were found in gaze duration, marking an earlier sensitivity point than in the No Cue condition. This finding contributes to previous research on the role of determiners in morphosyntactic processing (Lew-Williams & Fernald, 2010; Grüter et al., 2012; Hopp, 2013; Fuchs, 2022a, 2022b). Our data provide evidence that gender-marked determiners can enhance morphosyntactic processing during reading, even when another word intervenes between agreeing elements. This aligns with previous auditory research showing that heritage speakers can use gender-marked determiners to aid comprehension (Fuchs, 2022a, 2022b; Vergara & Socarrás, 2020) while also aligning with previous research showing that gender agreement is sensitive to earlier cue matching attraction during reading (González Alonso et al., 2021).

The Cue condition also supports cue-retrieval models of sentence processing (Van Dyke & Lewis, 2003), in which gender cues must be retrieved and matched at the adjective position. When a cue is present, the detection of gender agreement mismatches emerged earlier in the time course of agreement processing (in gaze duration) in addition to later reanalysis (rereading, total time, and go-past). One possible interpretation of this finding is that the determiner preactivates gender expectations, allowing the parser to generate predictions about upcoming morphosyntactic elements. Upon encountering a gender-marked determiner, readers may activate a set of grammatical constraints that narrow the range of expected forms for subsequent elements, such as adjectives. This predictive processing likely facilitates rapid integration of gender features, making agreement mismatches more immediately salient and triggering earlier detection when mismatches occur. In other words, the determiner provides an anticipatory cue that enables the syntactic system to monitor agreement relations more efficiently and to initiate anomaly detection processes sooner. Consequently, determiners may function not only as retrieval cues for morphosyntactic integration at the point of cue matching, but also as forward-looking signals that guide the parser’s expectations in real time. This dual role—both retrospective and prospective—may explain the robust grammaticality effects observed in early measures (gaze duration) and highlights the importance of early morphosyntactic cues in sentence processing.

5.3. The Role of AOB

All heritage speakers in our study detected gender agreement mismatches in the early (Cue condition) and late (No Cue and Cue conditions) stages of gender agreement processing. However, the timing of sensitivity to gender agreement processing between groups and across conditions was modulated by Age of Onset of Bilingualism (AOB). Let us start with the No Cue condition.

AOB, as indicated by group effects that received reversed Helmert contrast coding, revealed that the age of exposure to majority language input can affect the total time spent reading mismatched adjectives (total time), and the time spend regressing to earlier segments of the sentence after reading mismatched adjectives (go-past) in the No Cue condition. Early sequential heritage speakers (AOB 4–6) spent a longer time reading gender agreement mismatches at the adjective than simultaneous heritage speakers (AOB 0–3). Also, the amount of time revisiting earlier parts of the sentence after having processed gender agreement mismatches at the adjective (go-past) increased with Spanish proficiency for early sequential heritage speakers (AOB 4–6), but not for simultaneous heritage speakers (AOB 0–3), even when proficiency was not significantly different between these groups, (t = −0.12, p = 0.904, 95% CI [–3.06, 2.71]). Our findings support previous accounts that predict differences in the morphosyntactic system of heritage speakers as a function of timing of majority language input (Montrul, 2008; Tsimpli, 2014), but contrast with the findings of Keating (2022, 2025), which did not reflect AOB effects in the comparable nonadjacent condition. We speculate that the differences in these results may stem from individual differences with experience with heritage language schooling within our sample participants. Earlier schooling experience could have provided participants in the AOB 4–6 subgroup with greater exposure to the heritage language, resulting in greater online sensitivity to morphosyntactic mismatches (see Armstrong & Montrul, 2025). As displayed in

Table 3. Heritage speakers divided by AOB and schooling modality.

	Age of Onset of Bilingualism (AOB)
Schooling Modality	0–3	4–6	7–9
English Only	n = 9	n = 7	–
ESL Elementary	n = 4	n = 6	n = 7
Dual Elementary	n = 4	n = 7	n = 5
Dual Middle School	–	n = 1	–
Dual High School	n = 1	n = 1	–
Spanish Interrupted	–	–	n = 1
Total	n = 18	n = 22	n = 13

, the heritage speakers in the AOB 4–6 subgroup were not only the larger subgroup (n = 22), but they also displayed the most variety in their schooling backgrounds. While most participants reported English-only education in the AOB 0–3 group (n = 9), most participants in the AOB 4–6 group (n = 9) received some sort of dual immersion schooling (elementary, n = 7; middle, n = 1; high school, n = 1).

Recall that Montrul and Potowski (2007) found that sequential bilingual heritage speakers attending a dual immersion elementary school made fewer gender agreement errors than simultaneous bilinguals in a picture description task aimed at eliciting determiner–noun–adjective sequences orally. In contrast, Cuza and Pérez-Tattam’s (2016) study on heritage speaker children attending majority-language schooling reported no AOB effect. Dual immersion schooling in childhood may afford heritage speakers greater morphosyntactic representation that results in stronger sensitivity in adulthood in our study (Armstrong, 2024; Armstrong & Montrul, 2025). Microparametric features (Tsimpli, 2014), or the gender agreement system, may be strengthened by sustained exposure to the heritage language during school-age years. Although not the focus of our study, future research should investigate the issue of schooling experience as a modulator for morphosyntactic processing in adult heritage speakers with the aim of investigating the role of early overall exposure to the heritage language vs. the role of early heritage language schooling.

Now let us turn to the Cue condition, where AOB effects were more nuanced. Firstly, total time increased with proficiency, and this was true for all heritage speakers regardless of AOB (Figure 14). However, a greater time revisiting the adjective (rereading) was found for heritage speakers with AOB 4–9, but not for AOB 0–3 or the monolingually raised comparisons. This was an unexpected yet intriguing result because it suggests that reanalysis was triggered by later AOB (4–9), but neither simultaneous bilinguals nor monolingually raised speakers showed significantly more rereading of the mismatched adjective (vs. the matched one), a finding that also deviates from Keating (2022, 2025). Because proficiency and dominance were not significantly different between subgroups, we speculate that qualitative differences in reanalysis may also stem from experience with heritage language schooling. Both AOB 4–6 and 7–9 (Table 3) reported more dual immersion schooling in the language questionnaire, possibly suggesting that proficiency and dominance alone is insufficient to predict processing outcomes and must be considered alongside language experience variables—especially input quality, quantity, and literacy exposure. Heritage speakers with more exposure to Spanish in school may engage in deeper and more efficient morphosyntactic processing and feature tracking, especially in structurally demanding contexts.

Additionally, metalinguistic awareness may play a crucial role in this delayed processing. Participants with a higher awareness of morphological features may be better equipped to monitor and revise their parsing strategies when encountering incongruent gender cues. Inspection of our qualitative debriefing data revealed a somewhat conscious focus on agreement. Nearly 60% of the monolingually raised participants reported noticing mismatches when being debriefed. In contrast, only seven heritage speakers reported noticing gender-related issues, and they all had mid-to-high DELE scores. These observations align with the statistical findings, where proficiency predicted sensitivity to agreement mismatches in late measures but only among those participants with AOB 4–9.

5.4. Theoretical Implications and Future Directions

These findings contribute to cue-based parsing accounts by showing that heritage speakers integrate cues to activate gender features during reading. They also provide new evidence that the time course of gender agreement processing is delayed in distant contexts, but that the availability of cues (e.g., determiners) and bilingual experience (AOB) can jointly modulate sensitivity to gender agreement mismatches during sentence processing—a view consistent with models of language comprehension that emphasize integration of bottom-up and top-down information (Van Dyke & Lewis, 2003). Moreover, this study highlights the importance of fine-grained group comparisons among heritage speakers, as effects of bilingual experience (such as AOB) do not always manifest uniformly across measures or studies.

Future sentence processing work with heritage speakers may benefit from incorporating experience with heritage language schooling more systematically, in addition to reading skills and speed (e.g., Jegerski & Keating, 2023), to capture more individualized variation in heritage language processing. Another avenue to pursue would be the role of morphosyntactic features in guiding overall reading comprehension online, as previous research in the auditory modality (Vergara & Socarrás, 2020; Fuchs, 2022a, 2022b) suggests that morphosyntactic cues can enhance comprehension. Finally, our data show that gender agreement processing can trigger the parser to engage in sentence reanalysis in later reading measures. Future work should employ scanpath analysis (such as Parshina et al., 2022) to explore processing strategies during the detection of agreement mismatches in heritage language processing.

6. Conclusions

The present findings demonstrated that sensitivity to gender agreement mismatches in contexts of linear distance—such as when an adjective is separated from its head noun by an intervening word—is delayed when agreement is dependent on the retrieval of the gender feature of a word. Detection reliably emerged in later processing measures, indicating that gender agreement mismatches at a distance imposes reanalysis costs that are not captured by early reading measures. Importantly, both Age of Onset of Bilingualism (AOB) and Spanish proficiency shaped these delayed effects. Heritage speakers with an early sequential AOB (4–6) showed greater sensitivity to mismatches than those with simultaneous exposure (AOB 0–3), particularly at higher proficiency levels and in later measures. This suggests that early bilingual exposure provides a foundation for morphosyntactic processing, but that sustained engagement with the heritage language strengthens the retrieval and integration of features over time, as background data show that early sequential AOB (4–6) HSs showed a larger number of participants attending dual immersion schooling. Moreover, the presence of an earlier gender-marked determiner facilitated earlier detection of agreement mismatches, supporting the view that earlier gender cues can enhance agreement processing, even when agreement relations span across structural distance. Moreover, the later integration of these cues was modulated by proficiency, and the extent to which reanalysis occurred predicted sensitivity in heritage speakers with a later AOB (4–6) and (7–9). In sum, the time course of gender agreement processing in heritage speakers of Spanish was shaped by linear distance, cue retrieval, and bilingual experience. These findings highlight the need for more fine-grained approaches in bilingualism research, emphasizing the dynamic interplay between language experience, cues, and morphosyntactic processing.