The participants include 28 native speakers of Spanish (16 females; age range: 18–38; mean age: 27). Data from 27 of these participants (from a different study) were reported in Alemán Bañón and Rothman (2016). All participants indicated being right-handed, and this was confirmed via the Edinburgh Handedness Inventory (Oldfield, 1971). In addition, they all reported having no history of cognitive or neurological damage/diseases. They all spoke one or more foreign languages (mainly English) to varying levels of proficiency, and four of them identified themselves as speakers of another one of Spain’s co-official languages (Catalan, Galician) or Spanish Sign Language. They all received financial compensation for their time.

The materials comprise 160 single-clause sentences assigned to one of the four conditions in Table 1. All sentences follow the structure: subject + temporal adverb a menudo “often” + verb in the simple present + continuation (i.e., direct object or prepositional phrase). Half of the sentences (see conditions 1–2 in Table 1) include a lexical DP subject (e.g., el cazador “the hunter”), which corresponds to the default person (third person). In the grammatical version (condition 1), the verb is in the third-person singular. In the ungrammatical version (condition 2), the verb is incorrectly inflected as first-person singular, which is marked for person. In the other 80 sentences (conditions 3–4), the subject is the first-person singular pronoun yo (marked person: speaker). In the correct version (condition 3), the verb carries first-person singular inflection. In the ungrammatical version (condition 4), the verb shows third-person singular features and is, therefore, incorrectly underspecified for person. We chose the first as opposed to the second person as the marked subject for two reasons. First, only the first person allowed us to match the target verbs for length (e.g., lloro “cry-_{1ST PERSON-SG}” vs. llora “cry-_{3RD PERSON-SG}”; compare to lloras “cry-_{2ND PERSON-SG}”). Second, there is substantial variability with respect to the use of the second person across varieties of Spanish, even within European Spanish (e.g., Green, 1988).

In sum, markedness was manipulated via the speech participant status of the subject and its corresponding verb (el cazador…caza “the hunter-_{3RD PERSON-SG} hunt-_{3RD PERSON-SG},” yo…cazo “I-_{1ST PERSON-SG} hunt-_{1ST PERSON-SG}”) and agreement was manipulated by pairing up first-person subjects with third-person verbs, and third-person subjects with first-person verbs. The adverb a menudo “often” intervened between the subject and verb in order to create some linear distance between the agreeing elements. We reasoned that this might give the parser a better opportunity to engage in predictive processing, since additional time is available for prediction generation (e.g., Chow et al., 2016, 2018b). Thus, if subject–verb agreement is ever predictive, we thought that this would be an appropriate set-up to explore such a possibility.

For the conditions with third-person subjects, we used lexical subjects (as opposed to third-person singular pronouns) for two reasons. First, it allowed us to diversify the stimuli as much as possible. Most importantly, as discussed in Section “Introduction,” there is disagreement in the literature regarding whether third-person pronouns carry any person specification (e.g., Harley and Ritter, 2002 argue that they do not; Nevins, 2007 argues the reverse). In contrast, there seems to be agreement that lexical DPs are underspecified for person (e.g., Den Dikken, 2011; Nevins, 2011). Since the same could not be done in the conditions with first-person subjects, the fillers were designed so as to mitigate the salience of the first-person singular pronoun yo, which participants saw in 80 sentences. Therefore, the fillers involved 40 instances of the second-person singular pronoun tú “you,” 40 instances of the first-person plural pronouns nosotros/nosotras “we-_MASC/FEM,” and 80 instances of the third-person plural pronouns ellos/ellas “they-_MASC/FEM”). All materials are provided in Supplementary File 1.

Each inflected verb (e.g., llora “cry-_{3RD PERSON-SG},” lloro “cry-_{1ST PERSON-SG}”) was used twice, once with a third-person singular subject and once with a first-person singular subject (e.g., La viuda a menudo llora/^∗lloro en la iglesia “the widow often cry-_{3RD PERSON-SG}/^∗cry-_{1ST PERSON-SG} in church”; Yo a menudo lloro/^∗llora en las películas “I often cry-_{1ST PERSON-SG}/^∗cry-_{3RD PERSON-SG} at the movies”). This was done to ensure that all properties associated with a given verb (e.g., meaning, argument structure, lexical aspect, etc.) would be held constant across the two markedness conditions. With the exception of the subject, all sentences across the two markedness conditions were therefore identical up to the critical verb. Since the testing took place in two separate sessions, we distributed the materials in such a way that participants would only see one token of each verb per session.

Since the verbs were the same across markedness conditions, they were controlled with respect to number of characters [mean length of verbs inflected as third-person singular: 6.56; mean length of verbs inflected as first-person singular: 6.57; t(79) = 0.445, p = 0.658]. Mean length was, however, not exactly the same, due to five verbs showing certain conjugational or orthographic idiosyncrasies (e.g., conduce “drive-_{3RD PERSON-SG}” vs. conduzco “drive-_{1ST PERSON-SG}”; sigue “follow-_{3RD PERSON-SG}” vs. sigo “follow-_{1ST PERSON-SG}”). It was not possible to match the critical verbs with respect to frequency of use. We calculated the log frequency of each form with the EsPal database (Duchon et al., 2013), and found that third-person singular forms were significantly more frequent than first-person singular ones. This is unsurprising, given that default forms (i.e., third-person singular) have a wider syntactic distribution. Notice that a similar issue arose in Mancini et al.’s (2011a) study and that information about frequency is not provided in most other ERP studies on person agreement (e.g., Rossi et al., 2005; Nevins et al., 2007; Silva-Pereyra and Carreiras, 2007; Zawiszewski et al., 2016) ⁴. Finally, the position of the critical verb was always mid-sentence, and it was similar across markedness conditions (conditions 1–2: word #5; conditions 3–4: word #4).

These materials were intermixed with 240 sentences (160 ungrammatical) from a separate study that examines noun–adjective number and gender agreement, but does not manipulate subject–verb agreement (reported in Alemán Bañón and Rothman, 2016). All 80 fillers were grammatical, which brought the ratio of grammatical to ungrammatical sentences to 1/1. A sample of each filler type is provided in Table 1.

The testing was divided into two 3-hour sessions (e.g., O’Rourke and Van Petten, 2011; Alemán Bañón et al., 2012). Each EEG recording included 240 sentences (with an equal number of items per condition, including the fillers) and took approximately 1 h. Participants read the sentences quietly. The sentences were presented one word at a time, in random order. After each sentence, participants provided a grammaticality judgment, similar to previous ERP studies on person agreement (e.g., Rossi et al., 2005; Nevins et al., 2007; Silva-Pereyra and Carreiras, 2007; Mancini et al., 2011a, 2018; Zawiszewski et al., 2016). Participants received instructions to favor accuracy over speed while judging the sentences, to avoid blinks and muscle movements while reading them, and to rest their eyes between trials. At the beginning of each session, participants completed an eight-trial practice set (four ungrammatical) so that they would become acquainted with the task. None of the practice trials involved agreement errors or nouns/verbs from the experimental stimuli. Participants received feedback for the first three practice trials. The experiment began right after. Each session comprised six 40-sentence blocks, separated by five short breaks. Sentence presentation was carried out in Paradigm, by Perception Research Systems Inc. (Tagliaferri, 2005).

Each trial began with a fixation cross, which remained in the center of the screen for 500 ms. Then, the presentation of the sentence began, one word at a time, using the Rapid Serial Visual Presentation method. Each word remained on the screen for 450 ms, followed by a 300 ms pause (e.g., Alemán Bañón et al., 2012; see Molinaro et al., 2011a). Upon presentation of the last word (marked with a period), there was a 1000 ms pause. Right after, participants saw the prompts for the Grammaticality Judgment Task (GJT), the words Bien “good” and Mal “bad” for grammatical and ungrammatical sentences, respectively. The prompts remained visible until participants provided a response, which they did with their left hand (middle and index fingers, respectively). After the behavioral response, we added an inter-trial interval ranging between 500 and 1000 ms, pseudo-randomly varied at 50 ms increments.

The EEG was recorded with the Brain Vision Recorder software (Brain Products, GmbH, Germany) from 64 sintered Ag/AgCl electrodes mounted in an elastic cap (Easycap, Brain Products, GmbH, Germany). The placement of the electrodes followed the 10% system (midline: FPz, Fz, Cz, CPz, Pz, POz, Oz; hemispheres: FP1/2, AF3/4, AF7/8, F1/2, F3/4, F5/6, F7/8, FC1/2, FC3/4, FC5/6, FT7/8, FT9/10, C1/2, C3/4, C5/6, T7/8, CP1/2, CP3/4, CP5/6, TP7/8, TP9/10, P1/2, P3/4, P5/6, P7/8, PO3/4, PO7/8, O1/2). Electrode AFz served as the ground electrode and FCz as the online reference. The recordings were then re-referenced offline to the average of near-mastoid electrodes (TP7/8). Electrodes FP1/2, located above the eye-brows, were used to monitor blinks. Electrode IO was placed on the outer canthus of the right eye to capture horizontal eye movements. Electrode impedances were kept below 10 kΩ for all electrodes. The recordings were amplified by a BrainAmp MR Plus amplifier (Brain Products, GmbH, Germany) with a bandpass filter of 0.016–200 Hz, and digitized at a sampling rate of 1 kHz.

We analyzed the EEG data with the Brain Vision Analyzer 2.0 software (Brain Products, GmbH, Germany). After re-referencing the EEG, it was segmented into epochs relative to the critical verb. Epochs started 300 ms before the critical verb (i.e., the pre-stimulus baseline) and ended 1200 ms post-onset. Trials with blinks, horizontal eye movements, excessive alpha waves, or excessive muscle movement were manually rejected before analysis (based on visual inspection). We also discarded trials associated with incorrect responses in the GJT. This resulted in approximately 10% of data loss. After cleaning the data, the mean number of trials per condition ranged between 33 and 37 out of 40 (Condition 1: 37; Condition 2: 33; Condition 3: 36; Condition 4: 36), and this difference was significant, F(2.01,54.31) = 11.049, p < 0.01. Follow-up tests showed that the number of artifact-free trials in Condition 2 was lower than in all other conditions [Condition 2 vs. Condition 1: F(1,27) = 18.973, p < 0.001, q^∗ = 0.008; Condition 2 vs. Condition 3: F(1,27) = 14.415, p = 0.001, q^∗ = 0.017; Condition 2 vs. Condition 4: F(1,27) = 11.758, p < 0.01, q^∗ = 0.025], which did not differ from one another. Although this is not ideal, it should not be problematic for mean amplitude analyses (as opposed to peak analyses, which we did not conduct). As explained by Luck (2014, supplement, chapter 8, pp. 4–5), when measuring mean amplitudes, different numbers of trials per condition will not yield a spurious effect and should not be considered a confound. Following artifact rejection, data were baseline-corrected relative to the pre-stimulus baseline and averaged per condition and per subject. Finally, we applied a 30-Hz low-pass filter to the waveforms.

Event-related potentials were then quantified as mean amplitudes in two time windows: 250–450 ms, which corresponds to the LAN/N400, and 500–1000 ms, which corresponds to the P600. Both time-windows are consistent with previous reports on agreement. Importantly, they are the same time windows that we examined in Alemán Bañón and Rothman (2016). Thus, both time windows are the best estimates of where effects of agreement/markedness should emerge. For statistical analysis, we also used the same nine regions of interest (ROI) as in Alemán Bañón and Rothman (2016). Each ROI was calculated by averaging across the mean amplitudes of all electrodes in the region (left anterior: F1, F3, F5, FC1, FC3, FC5; right anterior: F2, F4, F6, FC2, FC4, FC6; left medial: C1, C3, C5, CP1, CP3, CP5; right medial: C2, C4, C6, CP2, CP4, CP6; left posterior: P1, P3, P5, P7, PO3, PO7; right posterior: P2, P4, P6, P8, PO4, PO8; midline anterior: Fz, FCz; midline medial: Cz, CPz; midline posterior: Pz, POz). The resulting values were then submitted to a repeated-measures ANOVA with Markedness (first-person singular subject, third-person singular subject), Agreement (grammatical, ungrammatical), Anterior–Posterior (anterior, medial, posterior), and Hemisphere (left, right) as the repeated factors. Since the hemisphere and midline regions comprise different numbers of electrodes, they were analyzed separately. For the analyses on the midline regions, Anterior–Posterior was the only topographical factor in the model. The Geisser and Greenhouse correction was applied in cases where sphericity could not be assumed. In such cases, we report corrected degrees of freedom (Field, 2005). A false discovery rate correction (Benjamini and Hochberg, 1995) was applied to all follow-up tests, to avoid an inflated Type I error. For all follow-up tests, we provide both the raw p-value and the adjusted significance level (q^∗), that is, the significance level below which we consider effects significant.

Table 2 provides the percentage of accurate responses in the GJT for each of the four experimental conditions (together with standard deviations). D-prime scores are also provided in the rightmost column. As can be seen, accuracy was generally very high (above 90% across the board), although participants were less accurate rejecting “unmarked subject violations.” A repeated-measures ANOVA with Markedness (first-person, third-person singular subject) and Agreement (grammatical, ungrammatical) as the repeated factors revealed a main effect of Markedness, F(1,27) = 9.051, p < 0.01, a main effect of Agreement, F(1,27) = 10.731, p < 0.01, and a Markedness by Agreement interaction, F(1,27) = 10.662, p < 0.01. Follow-up tests to the interaction revealed that the main effect of Agreement was only significant in the conditions with third-person singular subjects, F(1,27) = 13.316, p = 0.001, q^∗ = 0.025, driven by the fact that participants were less accurate rejecting ungrammatical sentences than accepting grammatical ones.

Figure 1 plots ERPs for all four experimental conditions in the six ROIs computed for analysis. As can be seen, approximately 500 ms after presentation of the critical verb, both types of person violations yielded a positivity relative to their grammatical counterparts. In both cases, the positivity shows a central-posterior distribution and a slight right hemisphere bias, consistent with the P600 (e.g., Barber and Carreiras, 2005). In addition, the positivity does not go back to baseline before the end of the epoch (at 1200 ms). The positivity appears more robust for “marked subject violations,” as it almost completely engulfs the positivity for the opposite error type, especially between 700 and 900 ms. This is also visible in Figure 2, which plots the magnitude of the violation effects for both types of person dependencies in four time windows of interest.

Also at approximately 700 ms, both types of person violations become more negative than grammatical sentences in the left anterior region, an effect that also remains visible until the end of the epoch (see Figure 1, 2). This late left anterior negativity also appears larger for “marked subject violations.” Preceding the P600, no evidence for a LAN or an N400 is apparent in Figure 1 or 2 for either type of person violation (e.g., Nevins et al., 2007; Silva-Pereyra and Carreiras, 2007). The following statistical analyses were conducted in the 250–450 ms time window (i.e., LAN effects should emerge in left anterior; N400 effects should emerge primarily in central-parietal regions) and the 500–1000 ms time window (i.e., P600 effects should emerge in central-posterior regions, possibly spanning over frontal regions for “marked subject violations”).

The P600 effects for both types of person violations are consistent with a large literature on agreement processing (e.g., Osterhout and Mobley, 1995; Barber and Carreiras, 2005; Alemán Bañón et al., 2012), including all previous studies on person agreement (Rossi et al., 2005; Nevins et al., 2007; Silva-Pereyra and Carreiras, 2007; Mancini et al., 2011a, 2018; Zawiszewski et al., 2016). As previously discussed, the functional significance of the P600 is still a matter of debate. Initial proposals viewed the P600 as an index of syntactic reanalysis and repair (or syntactic difficulty, more generally) (e.g., Osterhout and Holcomb, 1992; Hagoort et al., 1993). Subsequent ones have posited that the P600 reflects reanalysis processes in general (i.e., not exclusively morphosyntactic) (e.g., Kuperberg, 2007; Bornkessel-Schlesewsky and Schlesewsky, 2008), or conflict monitoring (van de Meerendonk et al., 2010). Our results do not adjudicate between these proposals (nor was it the purpose of the study), but they are consistent with them. That is, the P600 effects for person violations here might reflect the reprocessing costs associated with trying to reconcile conflicting information (i.e., morphosyntactic and discourse information) in light of top-down expectations.

The lack of an N400 effect for both types of person violations deserves some discussion. An N400 effect was reported by Mancini et al. (2011a) for person violations in Spanish, and for person errors in Basque by both Zawiszewski et al. (2016) and Mancini et al. (2018). Mancini et al. (2011a) interpret this effect as evidence that person violations disrupt the assignment of a discourse role to the subject, due to the failure to map morphosyntactic and discourse information (i.e., person inflection on the verb + speech participant role). We agree that this is indeed possible, but we remain skeptical about how generalizable this account is, since our results did not reveal N400 effects for either type of person error (consistent with Nevins et al., 2007 and Silva-Pereyra and Carreiras, 2007).

Finally, person violations in the present study also elicited a late anterior negativity in the same time window where the P600 emerged. This effect has been reported in previous studies on agreement that required participants to provide a sentence-final judgment (e.g., Sabourin and Stowe, 2004; Gillon-Dowens et al., 2010; Alemán Bañón et al., 2012; Alemán Bañón and Rothman, 2016; Zawiszewski et al., 2016). One position in the literature is that this late negativity reflects the cost of keeping the ungrammaticalities in working memory until the end of the sentence. This interpretation is consistent with our results. It also explains why the negativity was less robust for “unmarked subject violations” relative to the opposite error type, as participants were less accurate rejecting the former in the GJT (92 vs. 98% accuracy, respectively). One possibility is that the parser can better maintain the feature specification of the subject in the focus of attention when the subject is marked, which would explain why our participants were more accurate rejecting “marked subject violations” in the GJT (e.g., Wagers and McElree, unpublished). Another possibility is that, Spanish being a null-subject language, the salience of an overt personal pronoun facilitated the detection of the ungrammaticalities at the verb. We come back to this possibility below.

In Alemán Bañón and Rothman (2016), we hinted that this late anterior negativity might be a phase reversal of the P600 (Nunez and Srinivasan, 2006), since both effects showed similar latency, but the reverse scalp distribution (right posterior vs. left anterior). The same is true of the late anterior negativity in the present study (see Figure 2). That both components were impacted by markedness in a similar way makes us wonder the extent to which these two components are independent from one another (although Osterhout and Hagoort, 1999 point out that two different ERPs can be impacted by the same factor). We, therefore, remain cautious in interpreting this effect.

Our results revealed that person agreement violations realized at the verb yielded P600 effects of different magnitude in the 700–900 ms time window, depending on the speech participant status of the subject. More specifically, violations with a first-person singular subject, which corresponds to the speaker role, yielded a larger positivity than errors with a third-person (lexical) subject, which is underspecified for person (e.g., Harley and Ritter, 2002). This pattern of results is consistent with the proposal that, upon encountering a subject with marked features, feature activation allows the parser to generate a stronger prediction regarding the upcoming verb (e.g., Nevins et al., 2007). We do not argue that the larger P600 reflects prediction disconfirmation itself, since the effect was not frontally distributed (e.g., DeLong et al., 2011; Van Petten and Luka, 2012) (see Figure 2). The larger P600 for person violations with a marked subject might index the reanalysis process that the parser initiates when there is a conflict between a highly expected verbal form (i.e., more so than in the conditions with an unmarked subject) and the form that was actually encountered (e.g., van de Meerendonk et al., 2010).

These results are not consistent with our previous investigation on the role of markedness in the processing of noun–adjective number and gender agreement in Spanish, which involved the same participants (Alemán Bañón and Rothman, 2016). In that study, we found that violations realized on marked adjectives (plural for number; feminine for gender) yielded earlier and, in the case of number, larger P600 effects than violations realized on unmarked adjectives. Here, we found the reverse. It is possible that differences between the target structures where we examined agreement in each study explain this discrepancy. As we discussed above, the configuration where we examined noun–adjective agreement (e.g., una catedral que parecía inmensa “a cathedral-_FEM-SG that looked huge-_FEM-SG”) might not have been sufficiently constraining to allow the parser to generate strong predictions regarding upcoming adjectives, since other continuations were possible (e.g., una catedral que parecía desafiar la gravedad “a cathedral that seemed to defy gravity”). In fact, adjective phrases are always optional, although some structures might make adjectives more predictable (e.g., una fruta muy jugosa “a fruit-_FEM-SG very juicy-_FEM-SG,” where the adverb muy “very” makes it very likely that an adjective will follow; see Alemán Bañón et al., 2012). The same is not true of subject–verb agreement, where the presence of a subject DP allows for the strong prediction that a verb phrase (VP), headed by a verb, will appear in order to satisfy the phrase structure rule for sentence building (e.g., Chomsky, 1957, 1995). It is, therefore, possible that markedness influences agreement processing in different ways at different stages, depending on the nature of the computation itself (see Dillon et al., 2013, who suggested agreement attraction in comprehension to be sensitive to the predictability of the dependency).

The results of the present study differ from those by Mancini et al. (2018) in a number of ways, although there are certain similarities. Unlike Mancini et al. (2018), our results did not reveal reliable N400 effects for either type of person violation, although this is consistent with previous studies (e.g., Nevins et al., 2007; Silva-Pereyra and Carreiras, 2007). With respect to the P600, the present study found that “marked subject violations” yielded a larger P600 than the reverse configuration. A similar asymmetry between violations with marked vs. unmarked plural subjects emerged in Mancini et al.’s study (2018), except that, in their study, only violations with first-person plural subjects yielded a P600. Recall, however, that the ungrammatical status of “third-person plural subject + first-person plural verb” errors in Mancini et al.’s study (2018) was uncertain, given that participants accepted them at a rate of 42% in the judgment task, consistent with theoretical accounts of person agreement in Basque (e.g., Torrego and Laka, 2015; see Mancini et al., 2018 for counterarguments). In addition, the authors did not discard incorrectly judged trials from analysis. The same was not true of our study, where “unmarked subject violations” were unambiguously ungrammatical. In fact, our participants only accepted them at a rate of 8% (and we discarded incorrectly judged trials from analysis). This might explain, partly, why a P600 did not emerge for errors with unmarked subjects in Mancini et al.’s study.

Mancini et al. (2018) interpret their results as evidence that, when the subject carries no person specification (i.e., third-person plural), encountering a verb with first-person plural features (i.e., marked for person) allows the parser to extend the verb’s person specification to the subject. The authors point out that such a process only applies to plural subjects, which include more than one entity. For example, first-person plural includes the speaker + associates, and second-person plural includes the addressee + associates. In contrast, singular subjects are atomic entities that can only take their canonical speech role. What this means is that Mancini et al.’s proposal cannot explain our findings, since we found an asymmetry in the same direction as they did, but for person errors with singular subjects that differed with respect to markedness. However, Mancini et al.’s results can be explained in terms of an interplay between markedness and top-down expectations. That is, it is possible that the marked status of the first/second-person plural suffix –ok, relative to the third-person plural suffix –ek, allowed the parser to generate a stronger prediction regarding the upcoming verb. Future studies should explore this possibility, for example, by looking at person dependencies with plural subjects in non null-subject languages, where “third-person plural subject + first-person plural verb” configurations are more categorically disallowed (Höhn, 2016)⁵.

We must point out, however, that first- and third-person subjects in our study differed with respect to more than just feature specification. While the first-person conditions involved a personal pronoun, the third-person conditions involved referential DPs, and the reader might rightfully wonder how this could have affected our results. Recall that we opted for lexical DPs (as opposed to third-person pronouns) because there is consensus in the literature that they carry no person specification. Therefore, only first-person subjects should have allowed for prediction generation with respect to person morphology at the verb⁶.

One possibility is that sentences with first-person subjects were more salient than sentences with referential subjects because Spanish licenses pro drop and personal pronouns are often null. While this is indeed possible, we point out that overt pronouns are syntactically licensed and pragmatically appropriate as subjects in Spanish. Null pronouns are preferred as subjects if their referent can be inferred from context (topic maintenance), whereas overt pronouns tend to be used when there is a discourse switch to another referent (topic shift) (e.g., Lubbers-Quesada and Blackwell, 2009) or for contrastive focus (e.g., Rothman, 2009). This division of labor clearly emerges in cases of anaphora resolution such as the man pushed the boy when he/Ø... Here, null pronouns have been found to prefer subjects (the man) (e.g., Alonso-Ovalle et al., 2002; Carminati, 2005; Filiaci et al., 2014) and overt pronouns, objects (the boy) (e.g., Alonso-Ovalle et al., 2002; cf. Filiaci et al., 2014). Our materials, however, did not require anaphoric resolution. In fact, since each sentence was presented with no prior context (one that would determine topic maintenance or shift), the use of an overt pronoun does not seem overtly salient. In addition, we are skeptical that the use of third-person singular pronouns would have ameliorated this issue (even beyond theoretical considerations). Such a strategy would have made third-person pronouns more salient, because of their lower proportion in the language overall. For example, Morales (1997) shows that the proportion of overt first-person singular pronouns in European Spanish (our participants’ variety) is 28%, compared to 8% for third-person pronouns (see similar results in Duarte and Soares da Silva, 2016).

Another possibility is that the parser might have extracted feature information more easily from personal pronouns than lexical DPs, which encode lexical information that can slow down processing. While we cannot rule out this possibility, we point out that the adverb a menudo “often” intervened between the subject and the verb. Thus, since we used a 750 ms stimulus onset asynchrony, participants had 1800 ms to extract person information from the subject before encountering the verb (la viuda a menudo VERB). This time interval should have allowed participants to generate predictions (e.g., Chow et al., 2018b). Alternatively, the semantic features of lexical DPs might have impacted processing at the verb, either by allowing the parser to predict the type of event encoded by the verb, or by allowing combinatorial processing with the verb’s semantic features (even if the verb itself was not predicted). While this is also possible, we point out that the verb was held constant in the grammatical and ungrammatical conditions (la viuda…llora/^∗lloro). Thus, this should not have impacted the violation effect. We examined the possibility that lexical DPs might have allowed the parser to predict the event described by the verb by calculating the cloze probability of the target verbs in these conditions. The results of this cloze test (N = 33) show that mean cloze probability (across items) was very low (mean = 0.03; SD: 0.1), and that only one item had a cloze probability over 0.67, which corresponds to high probability (e.g., Block and Baldwin, 2010). Thus, the target verbs in the conditions with DP subjects were, overall, not predictable. Future research should investigate how markedness modulates person agreement while controlling for these differences, for example, by introducing the two subjects in a previous context, in order to reduce the salience of yo in the sentence where agreement is manipulated, and by using demonstratives in lieu of lexical DPs or third-person pronouns (see 9).

Two additional issues, however, might seem to undermine our claims. First, the mean number of trials for “unmarked subject violations” (Condition 2) was significantly lower than in the other three conditions. Thus, one could easily argue that the smaller P600 for person errors with an unmarked subject could be accounted for by signal-to-noise ratio differences across the conditions being compared. We can provide two counterarguments, one methodological and one theoretical. First, as discussed above, Luck (2014) points out that differences with respect to the mean number of trials per condition may affect analyses based on peak amplitudes, which we did not conduct, but not comparisons based on mean amplitudes, which are the basis for our conclusions. We therefore assume that the P600 size differences between the two error types are not epiphenomenal. Notice also that, albeit significant, the numerical differences in number of items across conditions were rather small (Condition 1: 37; Condition 2: 33; Condition 3: 36; Condition 4: 36) and the mean number of good items per condition was well above 30 across the board. Our second argument is that we only retained for analysis artifact-free trials that the participants had correctly judged in the GJT (unlike Mancini et al., 2018). As discussed in Section “Results,” participants were least accurate rejecting “unmarked subject violations” (Condition 2). Thus, the fact that Condition 2 encompassed fewer trials than the other conditions is not independent from how markedness impacts person agreement resolution online, which is our main research question.

The second issue concerns differences in lexical frequency between the critical verbs. Recall that first-person verbs were significantly less frequent than third-person ones. How could this have affected our results? There is evidence in the literature that lexical frequency is inversely related to the amplitude of the N400 (e.g., Neville et al., 1992; Kutas et al., 2006), a component associated with lexical access and retrieval. That is, less frequent words tend to show a larger N400. One possibility is that violations realized on first-person verbs (la viuda…^∗lloro “the widow-_{3RD PERSON-SG} cry-_{1ST PERSON-SG}”) yielded more negative effects than their grammatical counterparts (la viuda…llora “the widow-_{3RD PERSON-SG} cry-_{3RD PERSON-SG}”) in the N400 time-window, due to the fact that the verb was less frequent in the violation condition. In turn, this might have attenuated the following P600. Moreover, the reverse could have happened in the conditions with a marked subject. That is, violations on third-person singular verbs (yo…^∗llora “I-_{1ST PERSON-SG} cry-_{3RD PERSON-SG}”) might have elicited a smaller N400 relative to their grammatical counterparts (yo…lloro “I-_{1ST PERSON-SG} cry-_{1ST PERSON-SG}”), due to the fact that the verb was more frequent in the ungrammatical condition. In turn, this might have amplified the size of the subsequent P600. In fact, the results reported for the N400 time window are compatible with this scenario. Those analyses revealed a trend toward an N400 for “unmarked subject violations,” and a trend toward a positivity for “marked subject violations.” Crucially, however, the effects of markedness in our study emerged between 700 and 900 ms, in the late phase of the P600. If differences in the N400 time window (caused by differences in lexical frequency between the critical verbs) were responsible for the difference in P600 size across markedness conditions, those differences should have been largest in the early phase of the P600, right after the N400 (500–700 ms), which was not the case. To rule out this possibility, we recalculated effects in the 700–900 ms time window by using the N400 time window as a baseline (we used both the 250–450 and the 300–500 ms time windows) (e.g., Hagoort, 2003; Wicha et al., 2004; Martín-Loeches et al., 2006). These analyses revealed a similar pattern of results as with a pre-stimulus baseline. That is, the P600 was larger for marked subject violations, relative to violations with a third-person subject⁷. Thus, we can safely assume that the markedness effects that we found in the P600 time window are, at least to some extent, independent of baseline differences.