Individual differences in syntactic processing: Is there evidence for reader-text interactions?

Abstract

There remains little consensus about whether there exist meaningful individual differences in syntactic processing and, if so, what explains them. We argue that this partially reflects the fact that few psycholinguistic studies of individual differences include multiple constructs, multiple measures per construct, or tests for reliable measures. Here, we replicated three major syntactic phenomena in the psycholinguistic literature: use of verb distributional statistics, difficulty of object-versus subject-extracted relative clauses, and resolution of relative clause attachment ambiguities. We examine whether any individual differences in these phenomena could be predicted by language experience or general cognitive abilities (phonological ability, verbal working memory capacity, inhibitory control, perceptual speed). We find correlations between individual differences and offline, but not online, syntactic phenomena. Condition effects on reading time were not consistent within individuals, limiting their ability to correlate with other measures. We suggest that this might explain controversy over individual differences in language processing.

In Lee Cronbach’s famous presidential address to the American Psychological Association Annual Convention in 1957, he described an optimistic vision of the future of psychology in which the best of the correlational and experimental traditions joined forces as the united discipline. A complete theory of human behavior, he argued, requires the modeling of individual variability along with the prediction of an individual’s response to varying conditions. The usefulness of such a united approach is especially clear in the domains of applied psychology: It would be best to provide an intervention that is tuned to the particular needs of each individual (Pellegrino, Baxter, & Glaser, 1999). Since that 1957 address, psychologists have taken up the challenge of the united discipline. In their 1999 review, Pellegrino, Baxter, and Glaser chart the progress of the field, focusing on the intersections of cognitive psychology and psychometrics that follow directly from Cronbach’s initial interests, focusing first on “aptitude-treatment interactions”, or the relationship between a student’s intellectual abilities and expertise on one hand, and educational materials and instructional methods on the other.

A specific case of this type of investigation is what we will call “reader-text interactions”. Substantial prior work has revealed that the time required to read a sentence or text is a function of both the individual reader and the text being read: Researchers in individual differences and educational psychology have identified important sources of variation in reading and comprehension skill (e.g., Kuperman & Van Dyke, 2011; Perfetti & Hart, 2002), and work in cognitive psychology and psycholinguistics has identified the types of words, sentences, and texts that are more difficult for comprehenders (e.g., some syntactic structures are more difficult to process; Just & Carpenter, 1992; Waters & Caplan, 1996; Gibson, 1998). What is less clear is whether and how reader and text characteristics interact: Are difficult sentences equally challenging for all readers? And, conversely, does variation in reading skill affect the comprehension of all linguistic materials, or just especially difficult ones?

Here, we investigate reader-text interactions in the domain of syntactic processing. We first consider whether there is evidence for such interactions—that is, are some types of sentences consistently more difficult for specific types of readers? Then, to the extent that we observe such differences, we consider what might explain them.

The Relevance of Reader-Text Interactions

The potential for reader-text interactions in syntactic processing is relevant to several broader issues in psychology. First, if variability in readers interacts with properties of texts, that can provide insights into the underlying mechanisms of language processing (for a discussion of how individual differences contribute to more general theoretical development, see Vogel & Awh, 2008). As we review below, theories of language processing make different claims about why some texts are more difficult to process. Consequently, they also imply different hypotheses about which individual differences are likely to modulate syntactic processing. For instance, theories that attribute the difficulty of some syntactic structures to comprehenders’ relative inexperience with them predict that individual differences in language experience might drive differences in syntactic processing. By contrast, in accounts in which some syntactic structures are difficult because of the demands they place on memory, it is individual differences in memory capabilities that are most likely to relate to individual differences in syntactic processing.

Second, whether and which individual differences exist in syntactic processing speak to broader, fundamental questions about the architecture of the mind and language processing system. For example, as we review in greater detail below, some theories (e.g., Waters & Caplan, 2003) propose that language processing is divided into initial, automatic stages and later, interpretive stages, with only the latter subject to individual differences in working memory and other cognitive abilities. Studying individual differences in both online and offline processing allows us to test this theoretical claim. Similarly, another central question in psychology is the extent to which cognitive systems are modular rather than driven by domain-general systems (see Fodor, 1983). By understanding whether variability in the capacity of domain general systems like working memory and executive function is associated with syntactic processing, we can better understand the overall architecture of the mind: To what degree is language (and other motor and perceptual systems) modular, and to what degree does it recruit domain-general systems? It also provides an opportunity to understand why characteristics like high working memory are associated with positive outcomes in more complex domains like reading comprehension.

Finally, and most broadly, reader-text interactions exemplify one of the central questions of the united discipline envisioned by Cronbach: How do the skills and abilities identified by psychometricians intersect with the cognitive-processing effects discovered by experimentalists? Aptitude-treatment interactions have been reported in some educational domains. For instance, learners with greater prior knowledge learn better from different types of texts (McNamara, Kintsch, Songer, & Kintsch, 1996) and feedback (Hausmann, Vuong, Towle, Fraundorf, Murray, & Connelly, 2013) than do low-knowledge learners. Indeed, several reader-text interactions have been reported within the language processing literature. For instance, slower overall readers show larger effects of word frequency (Seidenberg, 1985), and readers with greater linguistic experience may be less sensitive to word difficulty and correspondingly more sensitive to discourse-level factors (e.g., the introduction of new concepts; Stine-Morrow, Soederberg Miller, Gagne, & Hertzog, 2008). Most relevant for the present paper, readers with greater linguistic knowledge are also more efficient at resolving syntactic ambiguity (Traxler & Tooley, 2007). On the other hand, a review of the learning-styles hypothesis—that certain learners do best under one instructional method and other learners do best with a different method—has found little evidence to date in favor of such an interaction; instead, the most well-established mnemonic effects appear to apply across learners (Pashler, McDaniel, Rohrer, & Bjork, 2008). Thus, there is a need to investigate in other domains whether the cognitive-processing effects discovered by experimentalists are consistent across individuals, and whether the important skills and abilities identified in psychometrics apply across tasks and materials.

Assessing Reader-Text Interactions

Given the applied and theoretical relevance of potential reader-text interactions, it is not surprising that they have been studied by both educational psychologists and cognitive psychologists. While educational psychologists have investigated reader-text interactions with the goal of promoting learning in young readers (e.g. Coté, Goldman, & Saul, 1998) and comprehension among students (e.g., McNamara et al., 1996), a complementary literature grew in cognitive psychology as theories of reading began to include ideas about individual differences in cognitive abilities. An influential example is Just and Carpenter (1992), who proposed, and reviewed evidence, that differences in capacity between individuals correlate with differences in reading ability. Since then, psycholinguists have employed individual differences to promote both memory-capacity theories of language comprehension (e.g., Fedorenko, Gibson, & Rohde, 2006; 2007; Gibson, 1998; 2000), competing experience-based theories (MacDonald & Christiansen, 2002, discussed in greater detail below), and a number of other explanations that combine language-specific and domain-general mechanisms (e.g., Farmer, Fine, Misyak, & Christiansen, 2017; Novick, Trueswell, & Thompson-Schill, 2010; Payne, Grison, Gao, Christianson, Morrow & Stine-Morrow, 2014; Swets, Desmet, Hambrick, & Ferreira, 2007; Van Dyke, Johns, & Kukona, 2014; Engelhardt, Nigg, & Ferreira, 2017).

As the individual differences approach in psycholinguistics has continued to grow in popularity in recent years, it is important to take a step back and assess its progress toward the united discipline. These psycholinguistic investigations are nested within the experimental approach, investigating language-processing effects that have been previously shown across subjects using controlled linguistic stimuli. So, the question is whether these investigations live up to the ideals of the correlational approach. Here, we describe several methodological demands identified by the correlational approach and discuss how these constraints may have contributed to a lack of consensus regarding individual differences in syntactic processing.

First, a critical insight from measurement theory is that two variables can be observed to correlate only to the degree that there is meaningful variation in those individual variables and to the degree that such variation is reliably measured (Spearman, 1904). If there are genuine, stable individual differences in syntactic processing, those individuals who show large syntactic-processing effects on one subset of items should also show large effects on another, similar subset. By contrast, a failure to observe such correlations would suggest that either (a) there are not consistent individual differences in syntactic processing or (b) such differences exist, but our methods cannot reliably detect them.

For instance, consider a scenario in which all readers read a syntactically complex sentence 300 ms more slowly than a syntactically simple sentence. In this case, there is clearly a text effect—one sentence is more difficult than another—but there is no reader-text interaction because all readers found the complex sentence more difficult than the simple sentence to the same degree. In this scenario, it would be impossible for any other construct (such as verbal working memory) to explain individual differences in syntactic processing because such variation was not observed to begin with.

Unfortunately, while past investigations of individual differences in syntactic processing have sometimes used measures of working memory and other cognitive abilities that have been normed for their reliability, researchers have only rarely assessed whether we observe meaningful variation across individuals in the syntactic processing effects themselves (but see Swets, et al., 2007 for one application of psychometric principles to syntactic processing). Thus, before we ask why individuals might differ in syntactic processing, it is first necessary to establish that such individual differences exist at all. If we cannot observe consistent individual differences in syntactic processing to begin with, differences in online syntactic processing cannot be expected to relate to any other measure.

Second, individual differences are best assessed with multiple measures. “Perhaps the most valuable trading of goods the correlator can offer,” Cronbach (1957) states, “…is his multivariate conception of the world. No experimenter would deny that situations and responses are multifaceted, but rarely are his procedures designed for a systematic multivariate analysis” (p. 676). A strength of the multivariate approach is that it deals explicitly with measurement error: Observed performance on almost any single task reflects not only the construct of interest but also measurement error, which includes both random error and non-random error from other constructs (Bollen, 1989). Consider reading span (Daneman & Carpenter, 1980), which has been used as the single measure of verbal working memory capacity in several influential psycholinguistic studies of individual differences (e.g., Just & Carpenter, 1992; MacDonald, Just, & Carpenter, 1992; Pearlmutter & MacDonald, 1995). The reading span task purports to measure verbal working memory capacity because it requires participants to remember particular words while reading sentences, but it might also be influenced by participants’ knowledge of specific lexical items (Engle, Nations, & Cantor, 1990; MacDonald & Christiansen, 2002). These confounds make it difficult to interpret a high or low score on any single measure. But, including multiple measures of a single construct allows researchers to assess the degree of common variance between them and use composite scores within a construct; for instance, a composite score for verbal working memory can be created by administering both a reading span and an operation span task. Unfortunately, not all psycholinguistic studies have used multiple measures (or indicators) for any given factor.

Further, the psychometric approach implies that multiple constructs should be measured simultaneously in order to tease apart their effects. A challenge for studying individual differences is that many potential explanatory constructs, such as verbal working memory and linguistic experience, might be intercorrelated (e.g., MacDonald & Christiansen, 2002), making it more challenging to attribute effects to any one construct in particular. In order to demonstrate that a specific construct—say, linguistic experience—is the one that drives differences in online syntactic processing, it is important to also measure other competing constructs and to show that it is specifically linguistic experience, and not (for example) verbal working memory or inhibitory control, that relates to individual differences in processing. However, many psycholinguistic studies have examined only one or two of these constructs within a single investigation; for instance, a study may measure verbal working memory but not reading experience, or vice versa.

In the current study, we aim to address these three issues by (1) assessing multiple constructs—both domain-general and language-specific—within individuals, (2) including multiple measures of each construct (e.g., multiple span tasks to create a composite measure of verbal working memory), and (3) assessing whether our data include consistent individual differences in the predictor variables and in the syntactic processing effects, both in initial processing and in subsequent comprehension. And, to the extent we find such differences, we attempt to tease apart competing theoretical explanations of them by examining multiple constructs that have been separately proposed to account for individual differences in syntactic processing: language experience, phonological ability, verbal working memory, inhibitory control, and processing speed. We review these proposed explanations below.

What Might Account for Individual Differences in Syntactic Processing?

Language experience

Experience-based accounts propose that individual differences in syntactic processing, even those that are correlated with domain-general abilities, are better explained as differences in exposure to various structures (MacDonald & Christiansen, 2002). This claim about individual differences is consistent with broader theories of language comprehension that posit a strong influence of experience on syntactic processing more generally. For example, constraint-based theories of language comprehension (Altmann & Steedman, 1988; MacDonald, 1994; MacDonald, Pearlmutter, & Seidenberg, 1994; Spivey-Knowlton, Trueswell, & Tanenhaus, 1993) propose that language comprehension is fast and accurate because it incorporates numerous probabilistic constraints, including syntactic ones, that comprehenders have learned through their experience with language. Experience-based theories are supported by demonstrations that syntactic structures are read more quickly when they are more frequent or predictable, as determined from either global statistics or those of particular verbs (Garnsey, Pearlmutter, Myers, & Lotocky, 1997; MacDonald & Christiansen, 2002), and even when memory demands are equated (e.g., Levy, Fedorenko, Breen, & Gibson, 2012). Experience-based accounts are further supported by evidence that online processing of initially difficult structures can be facilitated on the basis of recent laboratory-provided experience with the structures, including both trial-to-trial changes (Arai, van Gompel, & Scheepers, 2007; Thothathiri & Snedeker, 2008; Tooley, Traxler, & Swaab, 2009; Traxler, 2008) and changes over the course of one or more experimental sessions (Farmer, Fine, Yan, Cheimariou, & Jaeger, 2014; Fine, Qian, Jaeger, & Jacobs, 2010; Fine, Jaeger, Farmer, & Qian, 2013; Wells, Christiansen, Race, Acheson, & MacDonald, 2009), and even for structures that are only marginally grammatical (Luka & Barsalou, 2005; Luka & Choi, 2012) or that were previously unfamiliar to the comprehender (Kaschak, 2006; Kaschak & Glenberg, 2004; Fraundorf & Jaeger, 2016).

The examples thus far discuss differences between syntactic structures but within individuals. But, the claim that syntactic processing is guided by relative experience with different structures also suggests that processing could be influenced by differences among individuals in their relevant linguistic experience: Some individuals may come into the reading task with substantially more or less of the experience that was experimentally manipulated in some of the experiments described above. Thus, for instance, computational simulations suggest that rare, difficult structures are less disruptive for more experienced readers, who have more experience with these uncommon structures (MacDonald & Christiansen, 2002). This prediction has been supported by recent studies in the spoken language processing domain, which have found that individuals with higher vocabulary or higher literacy show facilitation in online, anticipatory language processing in the visual world (e.g. Borovsky, Elman, & Fernald, 2012; Mishra, Singh, Pandey, & Huettig, 2012; Rommers, Meyer, & Huettig, 2015). There is still relatively little comparable work in the written modality, but Traxler and Tooley (2007) found that individuals with greater knowledge were less affected by temporary syntactic ambiguity in their online processing.

Phonological ability

Phonological abilities have long been hypothesized to be a major factor in determining reading ability, particularly in acquisition or among poor readers (e.g. Byrne & Letz, 1983; Read & Ruyter, 1985; Sawyer & Fox, 1991; Wagner, Torgesen, & Rashotte, 1999). Experimental manipulations of phonological interference in text (Baddeley, Eldrige, & Lewis, 1981; Keller, Carpenter, & Just, 2003; Kennison, 2004; McCutchen, Bell, France, & Perfetti, 1991, to name a few) also suggest a role of phonology in offline syntactic comprehension even among skilled adult readers.

However, fewer studies have investigated effects of phonology during initial, on-line syntactic processing, and those that have yielded mixed evidence. Acheson and MacDonald (2011) found that sentences with embedded relative clauses were made more difficult by phonological overlap between the head noun of the relative clause and a noun embedded within it (e.g. baker and banker) and between the relative clause verb and main clause verb (e.g. sought and bought). This overlap effect was larger for object-extracted relative clauses (ORCs; 1a), which are typically more difficult in general, relative to subject-extracted relative clauses (SRCs; 1b), perhaps because in some theoretical accounts, phonological representations could be used to maintain the non-canonical ordering of agent and patient in the ORC (that is, the sought baker precedes the seeking banker; MacDonald & Christiansen, 2002).

• (1a)The baker that the banker sought bought the house.
• (1b)The baker that sought the banker bought the house.

However, Kush, Johns, and Van Dyke (2015) present data that suggest that these effects are the result of encoding interference rather than interference with syntactic integration. Indeed, some theories (e.g., McElree, Foraker, & Dyer, 2003; Martin & McElree, 2008) propose that maintaining serial order is not necessary for comprehension because previous constituents can be directly accessed in memory. Thus, while Van Dyke et al. (2014) found that reading times were related both to vocabulary and to non-verbal memory for serial order, they found no effects of phonological ability.

Whether variation between individuals in phonological ability plays a role in processing is a point of controversy, but it is possible that individual differences between individuals in phonological ability could also influence syntactic processing ability—especially for structures where it may be important to maintain serial order to arrive at the correct meaning of the sentence.

Verbal working memory capacity

As we introduced briefly above, capacity constraints in verbal working memory have figured prominently in research on reader-text interactions. Some theories have proposed that syntactic structures are difficult to process to the extent that they impose greater demands on memory (Fedorenko et al., 2006, 2007; Gibson, 1998, 2000; Just & Carpenter, 1992; King & Just, 1991). For instance, in both the ORC (1a) and SRC (1b) above, the relative pronoun that introduces a dependency in which the relative pronoun must eventually be co-indexed with a syntactic gap in the relative clause. In the ORC, this integration occurs later (at sought, the reader must recall it was the baker who was sought) and requires a longer-distance memory retrieval than in the SRC, in which the gap occurs immediately after that. It has been argued (Gibson, 1998, 2000) that these memory demands explain why ORCs are understood more slowly and less accurately. Thus, differences between individuals in their ability to store and retrieve these dependencies may be associated with how much more difficult they find ORCs.

Other theories suggest a second reason that memory abilities may be important to online language processing. Just and Carpenter (1992) propose that individuals differ in their total capacity to consider multiple sources of information; as a result, individuals with lower memory capacity may also be less able to use additional constraints such as semantic plausibility or referential contexts to help resolve a syntactic ambiguity.

Many studies have evaluated both of these predictions by directly relating syntactic processing to individual differences in measures of verbal working memory. These studies have often used complex span tasks in which participants receive sets of items to store and remember while completing a concurrent or interleaved processing task. For instance, participants may read sentences while remembering particular words from the sentences (Daneman & Carpenter, 1980). It has sometimes been reported that readers with lower scores on complex span tasks have greater difficulty with online processing of challenging syntactic structures, such as the object-extracted relative clauses described above (King & Just, 1991). However, Waters and Caplan (1996) point out that low-span readers in these studies performed worse overall and were not differentially more affected by syntactic difficulty. Moreover, studies have revealed inconsistent results as to whether low-span participants are actually more or less influenced by semantic and pragmatic information; some results suggest that low-but not high-span subjects see a benefit in online processing when helpful pragmatic cues are present (King & Just, 1991), and others suggest exactly the reverse (Just & Carpenter, 1992; Long & Prat, 2008; Pearlmutter & MacDonald, 1995; Traxler, Williams, Blozis, & Morris, 2005).

As a result, Caplan and Waters (1999) propose that online, automatic language processing and later interpretive processes tap separate resources and that only later, post-interpretive processes are assessed by complex span tasks and other working memory measures. For instance, differences in verbal working memory significantly relate to performance on object-extracted relative clauses in offline comprehension accuracy but not in online reading time, even when the measures come from the same participants reading the same sentences (Caplan, DeDe, Waters, Michaud, & Tripodis, 2011; Waters & Caplan, 2005). Indeed, although it is unclear whether such measures correspond to online reading, complex span performance correlates with offline syntactic processing, as well as reading comprehension more generally (Daneman & Merikle, 1996). For instance, Swets et al. (2007) found that working memory— even when measured using non-verbal complex span tasks—was significantly associated with how participants would interpret a syntactically ambiguous relative clause in offline comprehension questions (see also Payne et al., 2014).

Inhibitory control

Differences in working memory relate closely to another construct that has been proposed to drive individual differences in language processing: attentional control. Recent work (Novick et al., 2010) has examined syntactic processing as a function of domain-general inhibitory control, or the ability to resolve conflict between competing internal representations. Inhibitory control may be necessary for syntactic processing because the interpretation that comprehenders initially favor sometimes turns out to be wholly wrong and needs revision. This possibility is suggested by evidence that an initial misparse, even when later ruled out syntactically (Christianson, Hollingworth, Halliwell, & Ferreira, 2001) or revised by a speaker (Lau & Ferreira, 2005), is not always fully suppressed and may continue to influence readers’ eventual, offline interpretations. Indeed, online competition may even arise from syntactic structures that are never supported globally but that are coherent in the local syntactic context (Tabor, Galantucci, & Richardson, 2004).

In addition to the demands of revising the syntactic structure of a sentence, inhibitory control may be necessary for resolving competition between similar constituents online as the sentence unfolds. For example, the online processing difficulty of object-extracted relative clauses may be amplified by semantic (Gordon, Hendrick, & Johnson, 2001) or phonological (Acheson & MacDonald, 2011) similarity between the referents in the sentence. These findings are consistent with theories, both of language comprehension specifically (Lewis, Vasishth, & Van Dyke, 2006) and of memory more generally (Nairne, 2002), in which the primary determinant of short-term remembering is not a fixed storage capacity but rather the degree of interference between items to be remembered.

Thus, differences in the ability to suppress irrelevant information and resolve competition might lead to differences in the speed and accuracy of comprehension, and such correlations have been observed (Novick, Trueswell, & Thompson-Schill, 2005). More generally, the ability to suppress incorrect or irrelevant information has been argued to contribute to many aspects of language comprehension ability (Gernsbacher, 1993). Differences in inhibitory control might even account for effects previously attributed to working memory capacity—measures of inhibitory control often correlate with complex span task performance, and individual differences in performance on such tasks have sometimes been attributed in whole (Engle, 2002) or in part (Unsworth & Engle, 2007) to differences in inhibitory control. Indeed, it has been proposed that working memory span performance correlates with language comprehension and other complex activities because each of these activities relies on general attentional control processes (for review, see Kane, Conway, Hambrick, & Engle, 2007).

Perceptual speed

The final construct explored here is perceptual speed, or how quickly one is able to process perceptual stimuli (in the visual domain, within the current study), an ability that falls under the more general construct of processing speed (Salthouse, 1996). The inclusion of this basic ability is intended to capture and control for shared aspects of the reading task and other cognitive tasks that result from rapid visual processing of on-screen stimuli. For instance, perceptual speed has been proposed as one of the core abilities that support working memory (see Jarrold & Towse, 2006, for review), so controlling for perceptual speed would allow us to examine other aspects of working memory that may relate more to sentence processing. In addition, perceptual speed itself has been implicated in individual differences in language processing, although most frequently as an explanation for age-related changes in cognition (e.g., Salthouse, 1996; Caplan et al., 2011). Nevertheless, individual differences in processing speed may also explain some of the variability within an age group.

Current study

In the current study, we examined the contributions of both domain-specific and domain-general mechanisms to online and offline syntactic processing. We selected three syntactic constructions that have been relevant in the psycholinguistic literature in motivating both general theories of language processing and specifically those of individual differences. Our choice of constructions also allowed us to measure both online processing and offline comprehension, which provide insight into potential differences between interpretative and post-interpretive mechanisms. Critically, we also measured the internal consistency of each of these measures: Do we, in fact, observe consistent individual differences such that (for instance) some subjects consistently find ORCs easier to read than do other subjects? As we note above, although many studies have sought to relate verbal working memory and other such constructs to online sentence processing, researchers have not always assessed directly tested whether there are genuine individual differences in sentence processing to begin with.

To the extent that we do observe individual differences in syntactic processing, we also assessed individual differences in five other constructs (language experience, phonological ability, verbal working memory, inhibitory control, and perceptual speed) that might explain those differences in syntactic processing. All constructs were measured within the same set of participants, allowing for their effects to be distinguished and compared. Further, each of these constructs was assessed with multiple tasks, which allows us to create composite measures and mitigate task-specific effects.

Finally, we applied linear mixed-effects regression to relate these individual differences to syntactic processing. One potential challenge in distinguishing the influences of, say, verbal working memory and language experience is that, with a relatively large number of predictors and too few observations, regression models tend to capitalize on chance aspects of the data rather than yield generalizable results (the problem of overfitting; Babyak, 2004). Linear mixed-effects models reduce this problem because the unit of analysis is the reading time on an individual word or the response to an individual comprehension question, rather than an average of all of a participant’s reading times or responses. Thus, thousands of observations are available to the regression model. (For further discussion of linear mixed-effects models and other solutions to the study of individual differences in reading, see Matsuki, Kuperman, & Van Dyke, 2016).

Below, we detail each of these three syntactic structures and their corresponding processing measures.

Structures of interest

Relative clause extraction

First, we tested differences in reading and comprehending object-extracted versus subject-extracted relative clauses, a hallmark syntactic phenomenon that has contributed to numerous theories of syntactic processing. As reviewed above, within a participant, ORCs are typically more demanding and are read more slowly than SRCs within the relative clause; to preview, we replicate this well-established effect in our own data. Our interest, however, was whether there were differences across participants in the degree to which ORCs were relatively more difficult than SRCs. Thus, we took as a measure of individual differences the degree to which each participant read the syntactically difficult ORCs more slowly than the syntactically simpler SRCs.

Verb bias

We also examined a second widely-studied phenomenon in syntactic processing: the online use of verb distributional statistics in the sentential complement structure. In sentence (2), a temporary ambiguity between a direct object and sentential complement reading is introduced. In (2a), the ambiguity is resolved early: The complementizer that signals that the main verb accepted takes a sentential complement in which the contract is the subject. In (2b), removing the complementizer makes the contract temporarily ambiguous between the subject of the sentential complement (the player accepted some fact about the contract) and the direct object of accepted (the contract is what the player accepted).

• (2a)The basketball player accepted that the contract required him to play every game.
• (2b)The basketball player accepted the contract required him to play every game.

In general, the verb accepted is more likely to take a direct object than a sentential complement. Correspondingly, in the ambiguous version, readers slow down when the sentence is disambiguated to the sentential complement structure (at the verb required), suggesting they had initially favored the direct object interpretation that is consistent with the distributional statistics of accepted. However, other verbs, such as acknowledged, take a sentential complement more than a direct object; for these verbs, there is no benefit to disambiguating the structure with that, suggesting that readers already favor the sentential complement interpretation (Fine et al., 2010; Garnsey et al., 1997; Wilson & Garnsey, 2009; but see Kennison, 2001). Thus, our dependent measure of interest was individual differences in magnitude of this verb bias x ambiguity interaction, which indexes the influence of these distributional statistics on online syntactic processing. The use of verb bias is of interest not only because it is another cue that is available during online processing, but also because the learning of these biases provides evidence for how processing is shaped through experience with the language environment (for further discussion, see Ryskin, Qi, Duff, & Brown-Schmidt, 2017).

Attachment ambiguity

Finally, we examined the resolution of globally ambiguous relative clause attachments, such as (3) below:

• (3)The maid of the princess who scratched herself in public was terribly embarrassed.

The relative clause who scratched herself in public could modify either the maid or the princess. No syntactic information within the sentence resolves this ambiguity, but attaching the relative clause to the second noun (low attachment) is more common than attaching to the first noun (high attachment) in English (Rayner, Carlson, & Frazier, 1983), though not in all languages (Cuetos & Mitchell, 1988).

For these items, our interest was purely in participants’ offline syntactic processing (in contrast to Payne et al., 2014). Specifically, we queried whether participants arrived at the low attachment or high attachment reading, as revealed by offline probe questions, such as Did the princess scratch herself? Note that a “yes” answer to this question, taken alone, might reflect either a genuine low-attachment preference or a simple bias to affirm whichever interpretation is presented. However, as detailed in the Method and Results sections, we varied the question type across items, which allowed us to obtain a measure of participants’ low-attachment preference that was independent of a bias to respond “yes”; this measure of low-attachment preference then served as the key individual-difference variable for these items.

Research Questions

For each of these structures, we considered three questions. Our first question was simply whether we in fact observe consistent individual variation in each of the syntactic processing effects described above. That is, are there some individuals who are consistently advantaged at reading ORCs relative to other individuals? Do some individuals consistently show a stronger low-attachment preference than others? As we note above, a critical first step is to establish that individual differences exist and have been reliably measured before considering what other constructs might explain those differences.

Where we found that individuals do vary significantly in their syntactic processing, our second question was determining which individual differences, if any, relate to this variability: Are they domain-specific influences such as linguistic experience, or are they more domain-general abilities such as verbal working memory or executive function?

Finally, we considered whether the relationship between sentence processing and any of the individual differences here is present only in online processing, only in offline comprehension, or in both. Caplan and Waters (1999) propose that there are different constraints on online versus post-interpretive processing, and that only the latter is sensitive to differences in capacity between individuals; however, direct tests of this claim have still been relatively sparse in the literature.

Method

Participants

One hundred and thirty-three subjects participated for course credit or a cash honorarium. The study was advertised to the campus community and was thus biased toward younger adults and university students. Of the 133 participants, 10 did not provide any demographic information: Nine did not show up for the second session, in which the questionnaire was given, and one declined to complete the questionnaire. Of the 123 participants with demographic information, 78 (63%) were female. Participants’ ages ranged from 18 to 67 years (M = 20.94 years; SD = 5.37; median = 20 years; 94.3% under age 30). Our sample had only slightly more years of formal education than the nationwide mean (M =13.3 years completed; SD = 1.91; median = 12 years; range = [12, 19]; versus a nationwide mean of 12.9 years according to the United Nations Development Programme, 2014). Most participants (87%) indicated that they had completed at least “some college,” and of the 16 remaining responses, 10 came from University students participating for course credit, who presumably did in fact have some college education.

All participants reported that they were native speakers of English who had not been exposed to any other languages before the age of 5 and that they had normal or corrected-to-normal vision and hearing.

Of the total 133 participants that took part in the study, two were excluded entirely for not following directions on the processing speed tasks. An additional 14 participants were excluded from the regression analyses for having incomplete data (running out of time during a session or not arriving for the second session), leaving a total of 117 participants included in the analyses.

Materials

Critical stimuli for the self-paced moving window task consisted of 80 sentences with DO- or SC-bias verbs, 32 unambiguous subject-modifying relative clause sentences manipulated for extraction type, and 20 globally ambiguous relative clause sentences. We describe each of these stimulus types in detail below.

Use of verb bias

The online use of verb bias was tested using 80 critical sentences taken from Lee, Lu, and Garnsey (2013). Each sentence included a matrix subject, followed either by a DO-bias verb (40 sentences) or by a SC-bias verb (40 sentences), and then followed by a sentential complement. Each sentence had 2 versions that differed from each other solely in whether the sentential complement was headed by the complementizer that. Example sentences are presented in (4) below; they were otherwise identical in word frequency, length, and order. (Emphasis is added here for illustration purposes only and was not presented to participants.).

• (4a)DO-biased verb: The club members understood (that) the bylaws would be applied to everyone.
• (4b)SC-biased verb: The ticket agent admitted (that) the mistake might be hard to correct.

In the version without that, the role of the post-verbal noun was temporarily ambiguous between the direct object of the verb and the subject of a sentential complement. This ambiguity persisted until the next word (e.g., would in 4a or might in 4b), which disambiguated the sentence towards a sentential complement structure. In the version with the complementizer, the post-verbal noun was unambiguously the subject of a sentential complement.

Lee et al. (2013) controlled the character length and Francis-Kucera log word frequency of the post-verbal noun across verb type. Although the post-verbal noun was intended in all cases to be highly plausible as a direct object of the verb, plausibility as a direct object was rated as slightly higher after DO-bias verbs than after SC-bias verbs in a norming study conducted on a 7-point scale (6.4 and 6.1 respectively; 1: highly implausible, 7: highly plausible). For details of these norms, see Lee et al. (2013).

We used self-paced reading times to measure participants’ online processing of the verb bias items. Reaction times under 200 milliseconds were dropped and remaining times were log transformed. Only reaction times for correct trials were included in further analyses. For both sentence types, the critical region of analysis consisted of the embedded verb and the word immediately afterward, such as would be or might be, underlined in (4a) and (4b) above (“the disambiguation region” following Garnsey et al., 1997).

To measure offline comprehension, we created a YES-NO comprehension question for each sentence measuring participants’ understanding of its general meaning (e.g., Did the ticket agent think the mistake would be a problem?). The questions did not probe whether the participant arrived at the direct object or sentential complement interpretation.

Subject-versus object-extracted relative clauses

Processing of subject-versus object-extracted relative clauses was examined using 32 critical items taken from Gibson, Desmet, Grodner, Watson, and Ko (2005). Critical items began with a subject noun phrase, which was modified by a relative clause, and then continued with the verb phrase of the main clause of the sentence. Each item was manipulated for relative clause extraction site as shown in (5) below. The antecedent noun (in this case, reporter) was the subject of the relative clause in the SRC condition, and it was the object of the relative clause in the ORC condition.

• (5a)SRC: The reporter who attacked the senator on Tuesday ignored the president.
• (5b)ORC: The reporter who the senator attacked on Tuesday ignored the president.

Because the order of the words in the relative clause differed across extraction type, self-paced reading times were analyzed for a combined region including all of the relevant words (following Gibson et al., 2005). This region is underlined above and consisted of the relative pronoun who, the noun phrase, and the verb. Note that the conditions varied only in the order of these words; thus, word frequency and word length was controlled across conditions.

For each item, a YES-NO comprehension question was also created to assess offline comprehension. In half of the items, the questions required identifying the subject and object of the relative clause correctly (e.g., Did the reporter attack the senator?/Did the senator attack the reporter?). In the other half, the questions asked about main clauses (e.g., Did the reporter ignore the president?/Did the senator ignore the president?). This distinction allowed us to probe whether any difficulties in interpreting the ORCs were driven by difficulty in interpreting the relative clause in particular as opposed to the sentence more broadly.

Offline resolution of relative clause attachment ambiguities

To test offline judgments of relative clause attachment, we used 20 relative clause sentences taken from Swets et al. (2007). Each sentence contained a complex noun phrase modified by a relative clause, which was followed by the verb phrase of the main clause. The complex noun phrase included two animate nouns that were linked by the preposition of. Relative clauses contained a reflexive pronoun that could refer to either noun of the complex noun phrase, thus creating an attachment ambiguity. An example sentence is presented in (3), reproduced below.

• (3)The maid of the princess who scratched herself in public was terribly embarrassed.

For each item, we created a YES-NO question asking explicitly about relative clause attachment. In half of the items, a YES response indicated a low attachment interpretation (e.g., Did the princess scratch herself?); in the other half, a YES response indicated a high attachment interpretation (e.g., Did the maid scratch herself?). This design allowed us to apply signal-detection analyses (Green & Swets, 1996; Macmillan & Creelman, 2005; Murayama, Sakaki, Yan, & Smith, 2014) to separate participants’ potential response bias (any overall tendency to answer yes to all questions) from their low-attachment preference (an increase in yes responses specifically when that response indicates a low-attachment reading, termed sensitivity in the signal-detection framework).

List construction

Two lists were constructed by counterbalancing the complementizer presence-absence pairings for each of the 80 verb bias sentences and the SRC-ORC pairings for each of the 32 unambiguous relative clause sentences across lists. The stimuli for the 20 globally ambiguous relative sentences were identical across lists. In addition to these 132 experimental sentences, each list contained 80 filler sentences of various structures. The filler sentences were constructed to include a variety of grammatical structures and thereby disguise the structures of interest. Seventeen fillers were passive sentences (e.g., The terrifying monster was killed by the heroic knight), twenty-three were simple transitive sentences (The motivational speaker fixed the projector before her lecture), six included infinitive clauses (The game show contestant expected to win), four were simple intransitive sentences (The four kids shrieked when the monster appeared on screen), three were ditransitive sentences (The friendly man lent sugar to the neighbor next door), eight were conjoined sentences (Tania was accepted to graduate school and Steve passed the bar exam), sixteen used the sentential-complement structure but with a post-verbal noun phrase that was implausible as a direct object of the verb (eleven with the complementizer that and five without; e.g., The housewife hoped the antiques were valuable), two used the past progressive (The experienced flight attendant was giving instructions to a group of trainees), and one was an existential (There is an old house on the street whose roof was fixed).

Because the filler sentences were not constructed to be syntactically difficult or confusing, the comprehension questions did not specifically probe the syntax of the sentences but rather their general semantic content (e.g., Did Steve fail the bar exam?). For half of the fillers, the correct answer to the comprehension question was true; for the other half, it was false.

All participants saw the experimental and distracter sentences in the same, pseudo-randomized order. This design was motivated by our goal of measuring differences between individuals in their language processing, which requires minimizing extraneous sources of variability between participants. Differences in the experimental procedure (e.g., item ordering) across participants introduce additional, irrelevant between-participants variance that cannot be explained by the constructs of interest. By contrast, presenting items in the same order to all participants, although it confounds variance in item properties with serial position, crucially reduces the variance between participants in their experience in the experiment, and the goal of the present study was to explain variance between individuals rather than between items. (See Swets et al., 2007, for another example of an application of this principle to language processing studies.)

Procedure

All tasks, including the self-paced reading task, were completed on a Macintosh desktop computer running MATLAB software and the Psychophysics Toolbox (Brainard, 1997; Kleiner, Brainard, & Pelli, 2007; Pelli, 1997) and CogToolbox (Fraundorf et al., 2014). Participants sat approximately 750 mm from the screen.

Participants completed a total of 16 tasks (described individually in detail below) over two experimental sessions 24 hours apart. All participants completed the tasks in the same order to minimize experimental variability between individuals. First, participants completed a self-paced moving-window reading task designed to measure syntactic processing. Participants then completed a battery of tasks measuring the other individual differences of interest. On the first day, these tasks included, in order, three measures of verbal working memory (Reading Span, Listening Span, and Operation Span), two measures of perceptual speed (Letter Comparison and Pattern Comparison), three measures of inhibitory control (Antisaccade, Stroop, and Flanker), and two of five measures of language experience (vocabulary and Author Recognition Test). On the second day, participants completed a third language experience task (North American Adult Reading Test), three measures of phonological ability (Pseudoword Repetition, Phoneme Reversal, and Blending Nonwords), and finally the two remaining language experience measures (Comparative Reading Habits and Reading Time Estimates questionnaires). Between tasks, the list of tasks was displayed on the screen with checkmarks beside the completed tasks to indicate subjects’ progress. Participants were encouraged to take breaks between tasks as needed.

Self-paced moving window

Syntactic processing was assessed through a self-paced moving-window reading task (Just, Carpenter, & Woolley, 1982). The first word of a sentence was displayed on the screen, with each remaining word in the sentence replaced by a number of dashes equal to the character length of the word (e.g., chair would be replaced with -----). When the participant pressed the space bar, the next word was displayed and the previous word was replaced by dashes. Sentences were aligned with the left edge of the screen and displayed equidistant from the top and bottom of the display. All sentences occupied only a single line of text on the screen.

After participants read the last word of a sentence, the sentence disappeared, and a comprehension question was presented in its entirety. Participants answered yes or no by pressing one of two keys on the keyboard.

Between trials, the serial position of the upcoming trial was displayed for 750 ms in the same screen position as the first word of each sentence. Participants were given a rest period every 40 trials. This task lasted approximately forty-five minutes.

Reading Span

As in all variants of the Reading Span task (Daneman & Carpenter, 1980), participants read sentences while remembering material for a memory test. In the reading portion of the task, participants saw a sentence defining a common noun either truthfully, as in (6a), or falsely, as in (6b). Sentences were taken from Stine and Hindman (1994). Approximately half of the sentences were true and half were false.

• (6a)An article of clothing that is worn on the foot is a sock.
• (6b)A part of the body that is attached to the shoulder is the toe.

Each sentence was displayed in its entirety in the center of the screen. Participants read the sentence aloud, and then pressed the space bar. The sentence disappeared and was replaced with the prompt “Is this true?” Participants pressed one of two keys on the keyboard to judge the sentence as true or false.

One goal was to obtain measures of complex span performance that were less influenced by participants’ linguistic experience, which otherwise might explain any potential relation between verbal working memory and sentence processing (Engle et al., 1990; Macdonald & Christiansen, 2002). For instance, one way that language experience could influence span scores is by speeding the processing (sentence-reading) component of the task: If all participants saw the sentences for the same amount of time, those participants who could read the sentences more quickly would have more time remaining to implement rehearsal strategies (Friedman & Miyake, 2004). Indeed, allowing participants time to implement strategies in this way reduces the predictive power of complex span tasks (Friedman & Miyake, 2004; McCabe, 2010; Unsworth, Redick, Heitz, Broadway, & Engle, 2009). Thus, we followed the procedure of Unsworth, Heitz, Schrock, and Engle (2005) to reduce the influence of language processing speed by introducing an initial calibration phase to the task. During the initial calibration phase, participants performed only the processing (semantic judgment) task on 15 sentences and did not perform the memory storage task described below. Participants had unlimited time to read each sentence and make the judgment, and they received feedback on their accuracy afterwards. This procedure was designed to assess each participant’s reading speed. We then controlled for reading speed in the main task by giving participants a response deadline that was based on their speed in the calibration phase. In the Results section, we provide evidence that these procedures successfully deconfounded Reading Span scores from language experience.

A second way that language experience might influence complex-span performance is by facilitating processing of the to-be-remembered items. In some versions of the Reading Span task (such as the original version by Daneman & Carpenter, 1980), the to-be-remembered items are the final words of the sentences in the processing task. However, participants’ ability to remember such words is influenced by their familiarity or experience with the lexical items themselves (Engle et al., 1990). We thus instead adopted the procedure of Unsworth and colleagues by asking participants to remember letters, which all participants should find highly familiar and easy to process. The letters were randomly chosen from the set F, H, J, K, L, N, P, Q, R, S, T, Y, with the constraint that no letter ever appeared twice within the same trial. After each sentence in the main task, the to-be-remembered letter was displayed in caps in the center of the screen for 800 ms.

We also took two other steps to reduce participants’ ability to implement strategies. First, participants were required to read the sentence aloud and to press the space bar immediately after doing so; the program displayed a warning if participants were too slow at reading the sentences. Past work has established that stricter pacing of complex span tasks increases their predictive power (Friedman & Miyake, 2004; McCabe, 2010; Unsworth et al., 2009). Second, to prevent participants from neglecting the reading task in favor of rehearsing the to-be-remembered items, participants were instructed that their primary goal was to maintain at least 85% accuracy on the reading portion of the task. After each test phase, participants saw their cumulative accuracy on the processing task (i.e., their accuracy in judging the sentences as true or false) and received a warning whenever it dropped below 85% (Unsworth et al., 2005).

After completing the calibration procedure, participants proceeded to the main task. Participants continued to read sentences and judge them as true or false, but the maximum time allowed to read a sentence and make the semantic judgment was now set as the participant’s mean reading time in the calibration phase plus 2.5 standard deviations (Unsworth et al., 2005). If participants took longer than this time, “TOO SLOW!” displayed on the screen for 1000 ms, the sentence was counted as an error, and the computer proceeded to the next sentence. Participants did not receive feedback on their processing accuracy during the main task. After a predetermined number of sentences and letters, participants proceeded to the test phase of each trial, in which they were required to type the to-be-remembered letters in the order in which they had been presented.

Within the main task, participants first completed two practice trials at span length two (that is, two sentences and a total of two to-be-remembered letters). The critical trials consisted of two trials each at span lengths two to six, for a total of ten trials. A common procedure for complex span tasks has been for participants to start at the shortest span length and progress towards the longest span length, with the task ending if participants do not meet some criterion level of performance. However, researchers have raised several concerns with this procedure. First, performance typically decreases over repeated memory tests (the phenomenon of proactive interference). Presenting spans in ascending order confounds span length with the amount of proactive interference, and so variability in complex span performance could actually reflect variability in susceptibility to proactive interference (Lustig, May, & Hasher, 2001, but see Salthouse & Pink, 2008). Second, concluding the task early reduces the data collected from each participant. Participants may succeed or fail at a particular span lengths for reasons other than their putative verbal working memory abilities, such as the idiosyncratic difficulty of particular sentences (Conway, Kane, Bunting, Hambrick, Wilhelm, & Engle, 2005). Thus, even if a participant does not completely succeed at a given span length, performance at longer spans can still be revealing of their verbal working memory ability. Consequently, we presented the spans in a random order and required all participants to complete all spans.

Scoring was performed according to the partial-credit unit scoring procedure recommended by Conway and colleagues (2005). Trials on which participants remembered all of the items were scored as 1 point. Trials on which participants remembered some but not all items were scored as the proportion of items the participants did remember. This procedure makes use of all of the information available about participants’ performance and incorporates the fact that, for instance, remembering five out of six items represents somewhat better performance than remembering one out of six items. In a comparison of several scoring systems, Conway and colleagues found this procedure to produce the most normal distribution of scores.

Operation Span

The Operation Span task (Turner & Engle, 1989; Unsworth et al., 2005) was also intended to measure verbal memory and generally followed a similar procedure to the reading span task, except that the processing component of the task involved verifying the solutions to equations such as (7).

• (7)(6 × 4) − 2 = ?

In the processing portion of the Operation Span task, participants silently read the equation and pressed the space bar when finished. The equation was erased and a probe (such as 22) displayed on the screen; participants pressed one of two keys to judge whether or not the probe was the correct answer to the equation. Equations were generated according to the procedure of Unsworth et al. (2005). Specifically, the three numbers were always digits between 1 and 9. The first two digits were multiplied or divided together; then, a third digit was added or subtracted. These digits were selected semi-randomly such that the final answer was always a positive integer. Approximately half the test probes were true, and half were false. False probes were generated from the true answer by adding or subtracting a random number between one and nine, with the constraint that the resulting probe was always a positive integer.

As in the Reading Span procedure described above, participants first completed 15 equations in a calibration phase, which involved only the processing component of the task, in order to set the response deadline of the main task. The to-be-remembered items in the main task were the same set of letters used in the reading span task. Participants completed one practice trial at span length two and one at span length three, followed by three critical trials each at span lengths three to seven (for a total of 15 critical trials). As in the Reading Span task, the critical trials were presented in random order.

Listening Span

The Listening Span task generally followed the same procedure as the Reading Span task. However, rather than reading printed sentences aloud, participants listened to pre-recorded sentences spoken by a female native speaker of American English. The prompt to judge the sentence as true or false appeared immediately after the recorded sentence ended. Because the recorded sentence had an identical duration for all participants, calibration of the response deadline was based only on the latency to respond to the prompt. The to-be-remembered letters were also spoken aloud by the same recorded speaker. The task followed the same procedure as the Reading Span task in all other aspects.

Stimulus sentences were also taken from Stine and Hindman (1994) but comprised a different set of sentences than used in the Reading Span task. There were two practice trials at span length two, followed by two critical trials each at span lengths two to six, again presented in random order.

Letter Comparison

The Letter Comparison task followed Salthouse and Babcock (1991). Participants judged, as quickly as possible, whether two arrays of consonant letters were identical. Trials were presented in six blocks: two blocks comparing three-letter arrays, two blocks comparing six-letter arrays, and two blocks comparing nine-letter arrays. For practice, participants first completed two trials with three-letter arrays, in which one trial contained a match and the other contained a mismatch. Then, during each block, participants were given 20 seconds to complete as many comparison trials as possible, pressing one key for matching arrays and another for mismatching arrays. On mismatching trials, only one letter differed between the arrays. The dependent measure was the total number of correct answers provided within the duration of the critical blocks.

Pattern Comparison

The procedure of the Pattern Comparison task was the same as Letter Comparison, except that participants compared arrays of line segments rather than letters (Salthouse & Babcock, 1991). Blocks of three-, six-, and nine-segment arrays were presented in an order identical to that in the letter comparison task, with the dependent measure being the number of correct answers provided within this time.

Vocabulary

One word was displayed at the top of the screen in capital letters, followed by five other words (in lower case) and DON’T KNOW. Participants pressed one of the keys 1-5 on the keyboard to indicate which word was closest in meaning to the word at the top, or they pressed 6 if they did not know. There was one practice item, followed by two critical blocks of 24 items each. Participants had six minutes to complete each block; all participants completed the task within this time limit. All items were taken from the Extended Range Vocabulary Test of the Kit of Factor-Referenced Cognitive Tests (Ekstrom, French, Harman, & Dermen, 1976). Following the procedure recommended by Ekstrom et al. (1976), the dependent measure was the number of correct responses minus a penalty of 0.25 for each incorrect guess. Responses of DON’T KNOW were not penalized.

Author Recognition Test

The Author Recognition Test (ART) was developed as a measure of exposure to print materials (Stanovich & West, 1989). We used an updated and slightly lengthened version of the task developed by Acheson, Wells, and MacDonald (2008), which included the names of 65 authors’ names and 65 foil names, and adapted that version of the task for the computer. Participants saw names presented one at a time in a random order. For each name, the participant clicked one of two response buttons that appeared at the bottom of the screen reading Author and Don’t know. Participants were told that there was a penalty for guessing, so they were encouraged to only respond with Author if they were sure, and to otherwise choose Don’t know. Participants received one point for each correctly identified author, they lost one point for each foil name that they identified as an author, and there was no change to the score if they selected Don’t know.

North American Adult Reading Test

The North American Adult Reading Test (NAART) was developed as a way to estimate pre-morbid IQ in brain trauma patients (Blair & Spreen, 1989). Participants received a list of 61 words with irregular spellings, presented one at a time at increasing difficulty. The participants’ task was to correctly pronounce each word aloud. Correct pronunciations, determined by Merriam-Webster’s online dictionary, were given one point. Any incorrect response was given zero points with no partial credit. Table 1 displays inter-rater reliability for the NAART and for the other tasks discussed below that require manual scoring.

Table 1.

Inter-rater reliability for tasks requiring subjective scoring.

Task	Nsubjects	Nratings	Match proportion	Cohen’s κ
NAART	100	6100	0.884	0.765
Stroop accuracy	36	7200	0.980	0.816
Pseudoword Repetition	12	1056	0.863	0.848
Blending Nonwords	96	2304	0.954	0.900
Phoneme Reversal	107	2354	0.992	0.981

Note. N_subjects is the number of subjects that were scored by two raters, and N_ratings is the total number of trials with two ratings for these subjects. Match proportion is the proportion of N_ratings that were the same across both raters. Cohen’s κ provides a correction for chance agreement among raters (Cohen, 1960).

Comparative Reading Habits (CRH) survey

Participants answered five questions comparing their own self-reported reading habits to what they perceive to be the norm for their fellow college students (Acheson et al., 2008).

Reading Time Estimate (RTE) survey

Participants estimated how many hours in a typical week they read various types of materials, including fiction, newspapers, and online materials (Acheson et al., 2008).

Stroop

Following Stroop (1935) and Brown-Schmidt (2009), the Stroop task consisted of two phases. In the first, no-conflict phase, participants named the color of squares displayed one at a time on the screen. The possible colors were red, blue, green, yellow, purple, and orange. Before beginning the task, participants viewed a screen that displayed all six of the possible colors and their names. During the task, participants spoke aloud the name of the color of the square and then pressed a key to advance to the next trial; the key press was used to record participants’ response time for the trial.

In the second, conflict phase, participants performed the same task, except that the colored squares were replaced by the English names of colors (e.g., red printed in blue). Again, participants’ task was to name the color that the word appeared in, rather than read the word aloud.

Each phase contained 100 trials. There was no practice block in either phase, but the first and last trials in each phase were excluded from analysis to account for extreme reaction times attributed to beginning and ending the task.

Participants’ responses were recorded and coded for accuracy. Trials were coded as errors if the participant produced the incorrect color name, did not name a color at all, produced a filled pause such as uh or um (Fraundorf & Watson, 2013; Maclay & Osgood, 1959), or began speaking an incorrect color name before correcting themselves (e.g. gree-blue). Accuracy was generally high even in the conflict phase (M = 94%), and all participants obtained accuracy of 74% or greater.

The dependent measure was the difference in median response time between the conflict (second) phase and the no-conflict (first) phase. Because response times were positively skewed, as is typical in response time tasks (e.g., Van Zandt, 2000), response times were first log-transformed before conflict scores were calculated. Only correct trials were analyzed.

Antisaccade

Following the procedures of Kane, Bleckley, Conway, and Engle (2001), participants needed to look in the opposite direction of an anti-predictive cue in order to identify a letter briefly flashed on the opposite side of the screen. Each trial began with a fixation cross that lasted 200, 600, 1000, 1400, 1800, or 2200 ms; this duration varied across trials in order to prevent participants from anticipating the onset of the target. A cue (the equality sign =) then flashed one line of text below the fixation point, at either 11.3 degrees of visual angle to the left or to the right. The cue was visible for 100 ms, disappeared for 50 ms, and reappeared for 100 ms. The target display was then presented at the opposite location (e.g., if the cue appeared on the left, the target appeared on the right) on the same line of text as the fixation point. The target display consisted of a forward mask (the letter H) for 50 ms, then the target letter itself (B, P, or R) for 100 ms, and then a backward mask (the numeral 8). The backward mask remained on the screen until participants indicated the identity of the target by pressing the 1, 2, or 3 key on the keyboard. All of the characters subtended 2 degrees of visual angle vertically on the screen. There was a 400 ms interval between trials.

Participants first completed 18 trials in a response-mapping phase to practice the mapping between letters and response keys. In this phase, no cue appeared, and the masks and target appeared in the center of the screen. The response mapping was followed by 52 practice trials of the full task. During the practice trials only, participants received feedback in the form of a 175 Hz tone for 500 ms in response to incorrect responses. There was no feedback for a correct response. The practice trials were followed by 72 critical trials. Each possible combination of target identity (B, P, R), target location, and fixation duration was represented twice, and the trials were presented in random order. The dependent measure was the proportion of trials in the critical block on which participants responded correctly.

Flanker

Participants completed a version of the “flankers” response competition paradigm (Eriksen & Eriksen, 1974; see Eriksen, 1995 for review) in which a visually-presented target item is flanked either by congruent items that facilitate correct responding or by incongruent items that inhibit correct responding. In this particular implementation, participants indicated the direction of an arrow that was flanked by four arrows of the same (< < < < <) or different (> > < > >) direction. The incongruent items are thought to activate the incorrect response, making selecting the correct response more difficult, as reflected in longer response latencies (Eriksen, 1995). Similar to the Stroop analysis, the dependent measure was the difference between the median of log-transformed reaction times in the incongruent versus congruent trials.

Pseudoword Repetition

Following Gupta (2003), participants listened to recordings of pseudowords that were phonotactically legal in English (e.g., ginstabular), spoken by a female native speaker of American English. After each recording ended, a green dot appeared on the center of the screen and participants attempted to repeat the pseudoword they had just heard. When participants had finished repeating the word, they pressed a key and, after a 100 ms delay, the next trial began. To ensure that participants attempted to produce each word, participants could not end the trial before at least 1000 ms had elapsed; this time point was signaled by the dot on the screen turning blue. There were four critical blocks, each with 18 words: six two-syllable words, six four-syllable words, and six seven-syllable words. Before the main task, participants also completed six practice trials, two at each syllable length. Materials were taken from Gupta (2003).

Participants were awarded one point for each correctly repeated syllable from the onset of the word; correctly repeated syllables that occurred after an erroneous syllable did not earn points. For example, repeating ginstabular as ginstabcular would score two points; the first two syllables were repeated correctly, but the fourth syllable, while correct, occurred after an error in the third syllable. Some trials (7%) could not be coded because of problems in the recordings, usually because the participant pressed the key before completing the word; for this reason, the dependent measure used was the proportion of points earned out of the points possible on the coded trials only.

Blending Nonwords

Blending Nonwords is a task from the Comprehensive Test of Phonological Processing (CTOPP; Wagner et al., 1999). On each trial, participants heard a list of phonemes or syllables and were asked to combine these elements into one pseudoword, or “nonword”. For instance, if the participants heard/h/,/ε/, and/t/, they would need to produce/hεt/as one word. The number of elements ranged from two to eight. Participants were given six practice trials and eighteen critical trials, and the dependent measure was the proportion of correct responses. Following the CTOPP procedure, responses were scored as either fully correct or incorrect, with no partial credit.

Phoneme Reversal

In the Phoneme Reversal task (CTOPP; Wagner et al., 1999), participants heard a pseudoword and were asked to repeat the word and then pronounce it backwards, creating a real English word. For instance, if the participants heard/stu:b/, they would need to produce the word boots. Participants were given four practice trials and eighteen critical trials, and the dependent measure was the proportion of correct responses. Following the CTOPP procedure, responses were scored as either fully correct or incorrect, with no partial credit.

Results

As we reviewed above, interpreting any relationship between self-paced reading times and the other constructs requires establishing that the measures are reliable (consistent). It is also critical to demonstrate that the measures are valid (measuring what they intend to measure). We thus first discuss the reliability and validity of, in turn, (a) the measures of verbal working memory, perceptual speed, inhibitory control, language experience, and phonological ability and (b) individual differences in syntactic processing in the self-paced reading task. Finally, we turn to whether any individual differences in syntactic processing—should we observe any—can be explained by the other cognitive constructs.

Individual Differences

Mean performance on all 16 individual difference measures across the five domains is summarized in Table 2.

Table 2.

Summary of task performance on measures of individual differences.

Construct	Task	Measure	Min.	Mean	Max.	SD
Language experience	ART	Number correct with penalty	−9	10	47	11.570
	ERVT		1	17.180	36.750	7.740
	CRH	Sum of Likert responses	5	22	33	5.326
	NAART	Number correct	0.283	0.560	0.885	0.125
	RTE	Hours per week	5	20	63	10.695
Verbal	Listening	Score with	5.633	8.943	10	0.954
working	Operation	partial credit	1.852	10.588	15	3.435
memory span	Reading		2.367	6.749	10	1.757
Inhibitory control	Anti. Acc.	Proportion correct responses (all conflict trials)	0.264	0.717	0.986	0.192
	Anti. RT	Log median reaction time for correct responses (all conflict trials)	−2.043	−0.559	0.270	0.328
	Flanker	Difference in log	−0.254	0.151	0.344	0.069
	Stroop	median reaction time of correct responses for conflict and no-conflict trials	−0.153	0.193	0.514	0.135
Phonological ability	Pseudo. Rep.	Proportion correct	0.487	0.801	0.949	0.077
	BNW		0.167	0.646	1	0.176
	PR		0.182	0.687	1	0.176
Perceptual	Letter	Number correct	41	73	405	34.305
speed	Pattern	within time limit	46	84	398	31.378

Notes: Min. and Max. refer to the observed minimum and maximum scores, respectively. ART = Author Recognition Test (Stanovich & West, 1989; Acheson et al., 2008); ERVT = Extended Range Vocabulary Test (Ekstrom et al., 1976); CRH = Comparative Reading Habits (Acheson et al., 2008); NAART = North American Adult Reading Test (Blair & Spreen, 1989); RTE = Reading Time Estimate (Acheson et al., 2008); “Listening” = Listening Span task (Daneman & Carpenter, 1986; Stine & Hindman, 1994; Unsworth et al., 2005); “Operation” = Operation Span task (Turner & Engle, 1989; Unsworth et al., 2005); “Reading” = Reading Span task (Daneman & Carpenter, 1986; Stine & Hindman, 1994; Unsworth et al., 2005); “Anti.Acc” and “Anti.RT” refer to accuracy and median reaction time on the Antisaccade Task (Kane et al., 2001), respectively; “Flanker” = Flanker task (Eriksen & Eriksen, 1974); ; “Stroop” = Stroop task (Stroop, 1935; Brown-Schmidt, 2009); “Pseudo.Rep” = Pseudoword Repetition (Gupta, 2003); BNW = Blending Nonwords (Wagner et al., 1999); PR = Phoneme Reversal (Wagner et al., 1999); “Letter” = Letter Comparison (Salthouse & Babcock, 1991); “Pattern” = Pattern Comparison (Salthouse & Babcock, 1991).

The split-half correlations for each task are given in Table 3. These measures of internal consistency in individual differences on these tasks were generally on par with prior literature, indicating that we had successfully measured meaningful variation across individuals. However, the split-half correlations for Reading Span, NAART, Blending Nonwords, and Phoneme Reversal were noticeably lower than measures of internal consistency that have been reported in previous norms; this likely reflects the fact that our sample comprises a somewhat more restricted range of reading skills (Conway et al. 2005; Uttl, 2002; Wagner et al., 1999). The Eriksen flanker task had the lowest split-half correlations, which is perhaps not surprising given that the measure is a difference score and difference scores generally have lower consistency (e.g., Lord, 1963; Redick & Engle, 2006, but see Wöstmann, Aichert, Costa, Rubia, Möller, and Ettinger, 2013, for higher consistency of the flanker test in other reports).

Table 3.

Split half correlations of individual differences tasks

Construct	Task	Split half correlation
Language experience	ART	0.721
	ERVT	0.702
	NAART	0.827
Verbal working memory	Reading	0.623
	Listening	0.574
	Operation	0.807
Inhibitory control	Flanker	0.287
	Anti. Acc.	0.891
	Anti. RT	0.892
	Stroop	0.834
Phonological ability	Pseudo. Rep.	0.791
	BNW	0.531
	PR	0.459
Perceptual speed	Letter	0.893
	Pattern	0.916

Notes: Tasks were split into balanced halves, scores were calculated for each subject in each half in the same manner as in the main analyses, and the two lists of subject scores were correlated. ART = Author Recognition Test (Stanovich & West, 1989; Acheson et al., 2008); ERVT = Extended Range Vocabulary Test (Ekstrom et al., 1976); CRH = Comparative Reading Habits (Acheson et al., 2008); NAART = North American Adult Reading Test (Blair & Spreen, 1989); RTE = Reading Time Estimate (Acheson et al., 2008); “Listening” = Listening Span task (Daneman & Carpenter, 1986; Stine & Hindman, 1994; Unsworth et al., 2005); “Operation” = Operation Span task (Turner & Engle, 1989; Unsworth et al., 2005); “Reading” = Reading Span task (Daneman & Carpenter, 1986; Stine & Hindman, 1994; Unsworth et al., 2005); “Anti.Acc” and “Anti.RT” refer to accuracy and median reaction time on the Antisaccade Task (Kane et al., 2001), respectively; “Flanker” = Flanker task (Eriksen & Eriksen, 1974); ; “Stroop” = Stroop task (Stroop, 1935; Brown-Schmidt, 2009); “Pseudo.Rep” = Pseudoword Repetition (Gupta, 2003); BNW = Blending Nonwords (Wagner et al., 1999); PR = Phoneme Reversal (Wagner et al., 1999); “Letter” = Letter Comparison (Salthouse & Babcock, 1991); “Pattern” = Pattern Comparison (Salthouse & Babcock, 1991).

The next question was whether these individual differences reflect the underlying constructs that we expected them to. To assess this, we turned to the correlations between tasks. Ideally, tasks chosen to reflect the same underlying construct should exhibit moderately positive correlations, and tasks reflecting different constructs should be less correlated (e.g., Kane, Hambrick, Tuholski, Wilhelm, Payne, & Engle, 2004). Table 4 lists the correlations among all measures of individual differences of primary interest.

Table 4.

Correlations among measures of individual differences

Speed		Language experience					Verbal Working Memory				Inhibitory Control			Phon. Ability
LComp	PComp	NAART	RTE	CRH	Vocab	ART	OSpan	LSpan	RSpan		Anti.Acc	Flanker	Stroop	PR	BNW	Gupta
−0.027	0.015	0.184	−0.137	0.166	0.063	−0.036	0.511***	0.398***	1.000	RSpan	0.301***	0.130	0.027	0.199*	0.166	0.197*
0.176	0.177	0.181	−0.001	0.180	0.204*	0.131	0.493***	1.000	0.398***	LSpan	0.207*	0.027	−0.269**	0.129	0.069	0.148
0.076	0.128	0.025	−0.181	0.069	0.060	0.050	1.000	0.493***	0.511***	OSpan	0.319***	−0.020	−0.197	0.172	0.032	0.142
−0.038	−0.047	0.386***	0.194*	0.248**	0.453***	1.000	0.050	0.131	−0.036	ART	−0.061	−0.103	−0.081	0.223*	0.028	0.277**
0.191*	0.202*	0.682***	0.104	0.349***	1.000	0.453***	0.060	0.204*	0.063	Vocab	0.264**	0.112	−0.099	0.329***	0.198*	0.451***
0.019	0.016	0.263**	0.387***	1.000	0.349***	0.248**	0.069	0.180	0.166	CRH	0.051	−0.042	0.007	−0.044	0.153	0.201*
−0.106	−0.039	0.182	1.000	0.387***	0.104	0.194*	−0.181	−0.001	−0.137	RTE	0.333***	0.068	0.123	−0.167	−0.079	0.010
0.221*	0.211*	1.000	0.182	0.263**	0.682***	0.386***	0.025	0.181	0.184	NAART	0.161	0.055	−0.225*	0.299**	0.336***	0.504***
0.682***	1.000	0.211*	−0.039	0.016	0.202*	−0.047	0.128	0.177	0.015	PComp	0.328***	0.163	−0.110	0.186*	0.139	0.175
1.000	0.682***	0.221*	−0.106	0.019	0.191*	−0.038	0.076	0.176	−0.027	LComp	0.217*	0.049	−0.182*	0.137	0.084	0.101
0.101	0.175	0.504***	0.010	0.201*	0.451***	0.277**	0.142	0.148	0.197*	Gupta	0.233*	0.096	−0.169	0.322***	0.459***	1.000
0.084	0.139	0.336***	−0.079	0.153	0.198*	0.028	0.032	0.069	0.166	BNW	0.290**	0.128	−0.104	0.345***	1.000	0.459***
0.137	0.186*	0.299**	−0.167	−0.044	0.329***	0.223*	0.172	0.129	0.199*	PR	0.401***	0.107	−0.135	1.000	0.345***	0.322***
−0.182*	−0.110	−0.225*	0.123	0.007	−0.099	−0.081	−0.197	−0.269**	0.027	Stroop	−0.204*	−0.049	1.000	−0.135	−0.104	−0.169
0.049	0.163	0.055	0.068	−0.042	0.112	−0.103	−0.020	0.027	0.130	Flanker	0.258**	1.000	−0.049	0.107	0.128	0.096
0.217*	0.328***	0.161	0.333***	0.051	0.264**	−0.061	0.319***	0.207*	0.301***	Anti.Ac	1.000	0.258**	−0.204*	0.401***	0.290**	0.233*

Notes:

^***p < 0.001;

^**p < 0.01;

^*p < 0.05

Additionally, in Table 5, we devote special attention (given past controversy on this point; e.g., MacDonald & Christiansen, 2002) to the correlations between the language experience composite measure and three aspects of the span measures: accuracy on the processing component of the task, the maximum time allotted to the processing component according to the calibration phase, and the actual span measure. Notably, language experience was significantly correlated with the response deadline set by the calibration phase in the Reading Span task and with accuracy in the processing task in the Listening and Reading Span tasks, but it was crucially not correlated with the actual memory span scores for both Reading Span and Operation Span. This pattern implies that the calibration procedure was successful in separating the aspects of the Reading Span task that reflect linguistic experience (i.e., processing task speed and accuracy) from verbal working memory capacity per se. However, language experience was still correlated with Listening Span scores; This may be because the calibration procedure for Listening Span altered only the time taken to answer the questions and could not alter the presentation rate of the stimuli themselves.

Table 5.

Correlations among components of memory span tasks and language composite score

	Processing accuracy		Calibrated time limit		Span score
Task	Pearson’s r	t-value	Pearson’s r	t-value	Pearson’s r	t-value
Reading	0.397***	4.899	−0.237**	−2.762	0.092	1.027
Listening	0.312***	3.715	−0.142	−1.618	0.214*	2.375
Operation	0.178*	2.039	0.012	0.134	0.054	0.558

Notes:

^***p < 0.001;

^**p < 0.01;

^*p < 0.05

In general, however, measurement properties of the individual differences tasks were mixed: Although many tasks behaved as expected, some showed only modest internal consistency, and the pattern of correlations among individual tasks did not align neatly with a priori constructs (i.e., some significant correlations between constructs and weaker correlations within constructs). As the inhibitory control battery was particularly problematic (consistent with the low convergent validity of this construct in other work; Duckworth & Kern, 2011), we also ran a version of the primary regression analyses below without these scores. Withholding inhibitory control had no effect on the overall pattern of results, except in one regression (see notes of Table 11). Therefore, we have chosen to retain all measures in the following analyses, but leave further exploration of these issues for future work.

Table 11.

Fixed effects in model of comprehension accuracy in Relative Clause Extraction sentences

	Fixed effect	Estimate	SE	p-value
	(Intercept)	1.850	0.186	<0.001
Individual differences	VWM	0.378	0.096	<0.001
	Inhib.	−0.025	0.140	0.860
	Phon.	0.194	0.120	0.104
	Lang.	0.437	0.123	<0.001
	Speed	0.103	0.089	0.246
Condition effects	ORC	−0.464	0.155	0.003
	RCQ	−0.595	0.353	0.092
	ORC × RCQ	−0.472	0.309	0.127
Individual difference × Condition interactions	VWM × ORC	0.102	0.115	0.376
	VWM × RCQ	0.058	0.113	0.605
	Inhib. × ORC	0.147	0.169	0.386
	Inhib. × RCQ	−0.245	0.166	0.139
	Phon. × ORC	−0.180	0.147	0.221
	Phon. × RCQa	−0.249	0.143	0.082
	Lang. × ORC	−0.020	0.158	0.898
	Lang. × RCQ	−0.190	0.155	0.222
	Speed × ORCa	0.195	0.109	0.074
	Speed × RCQ	0.091	0.107	0.395
	VWM × ORC × RCQ	−0.013	0.227	0.955
	Inhib. × ORC × RCQ	−0.012	0.333	0.971
	Phon. × ORC × RCQ	0.363	0.289	0.209
	Lang. × ORC × RCQ	−0.653	0.314	0.037
	Speed × ORC × RCQ	−0.325	0.216	0.131

Notes: Effects coding was used for the condition effects. The condition effects here refers to the change in the log odds of correct responding when given an object-extracted relative clause (0.5) opposed to subject-extracted (−0.5) and receiving a comprehension question that probed the relative clause region of the sentences (0.5) opposed to the main clause (−0.5). Random intercepts and slopes for all condition effects for both subjects and items were also included in the model. “VWM” = Verbal working memory span; “Inhib” = Inhibitory control; “Speed” = Perceptual speed; “Phon.” = Phonological ability; “Lang.” = Language experience.

^aThese effects reached significance in the model if inhibitory control and its interactions were dropped.

Composite scores for each construct were devised by first standardizing all task scores (creating z-scores) and then averaging the standardized scores by subject within each domain (following Stine-Morrow et al., 2008). Table 6 lists the correlations among the composite scores from each domain.

Table 6.

Correlations among composite scores

	Lang.	WM	Inhib.	Speed
VWM	0.161
Inhib.	0.078	0.327***
Speed	0.134	0.088	0.241**
Phon.	0.333***	0.299***	0.299***	0.194*

Notes:

^***p < 0.001;

^**p < 0.01;

^*p < 0.05. “VWM” = Verbal working memory span; “Inhib” = Inhibitory control; “Speed” = Perceptual speed; “Phon.” = Phonological ability; “Lang.” = Language experience.

Creating composite scores in this way, while intuitive, makes the assumption that all task scores should be given equal weight within a construct. The low internal consistency of some of the measures (Table 3) and an “eyeball” factor analysis of the inter-task correlations in Table 4 suggests that this assumption is not justified; some scores are less stable than others and vary in their correlation strength to other within-construct tasks. A confirmatory factor analysis (see Appendix A) is consistent with these observations. While the data suggest that it would be useful to differentially weight the standardized task scores before combining them (e.g. by computing factor scores), this approach has its own limitations. First, as Bollen (1989, p. 306) points out, using factor scores in a regression model does not remove measurement error; the CFA model estimate itself contains error. Second, the CFA model should be validated on a new dataset, as the model was respecified based on initial fit; data-driven model respecification risks exploiting chance variation (MacCallum, Roznowski, & Necowitz, 1992).

Self-Paced Reading Measures

Prior to all other analyses, reading times were first corrected for word length by residualizing per-word reading times on word length (Ferreira & Clifton, 1986). Specifically, the dependent measure in the following reading time analyses was the residual of a linear regression model predicting log reading times from word length only (i.e., random slopes for subjects were not included; this was done to preserve subject-based variation for the individual differences analysis).

Residual reading times were then included as the dependent measure in a series of linear mixed-effects models to examine each of the syntactic phenomena of interest. Effects coding was used for the sentence-type factors, producing main effect estimates comparable to those of an ANOVA. All models included random intercepts and random slopes for condition effects for subjects and items. Complete equations for the models can be found in Appendix B.

We first assessed whether we replicated the standard patterns in reading times across individuals (e.g., that verb bias interacts with ambiguity) for each of the three syntactic phenomena of interest. Then, we turn to whether we observed consistent individual differences in these phenomena, such that (for instance) some subjects consistently had larger verb bias effects than others.

Verb bias effects

For online processing of the verb-bias sentences, the critical region of analysis was defined as the embedded verb plus the following word (spillover). We constructed a mixed-effects regression examining length-residualized reading times in the critical region as a function of ambiguity, verb bias, their interaction, and random effects for subjects and items.

The model yielded a significant ambiguity effect such that sentences without the complementizer that took longer to read ( $\widehat{\mathrm{\beta }}=0.074$, SE = 0.0107, p < 0.001), and it yielded a significant interaction with verb bias such that the ambiguity effect was larger for DO-biased sentences ( $\widehat{\mathrm{\beta }}=0.099$, SE = 0.023, p < 0.001). Together, these findings replicate the verb bias effect in the literature. Reading times across DO- and SC-biased sentences across both ambiguity conditions are plotted in Figure 1.

Figure 1.

Mean residual reading times across sentence regions for DO-biased (a) and SC-biased (b) sentences. Error bars represent the standard errors of the mean residual reading times with the correction for within-subject factors given in Morey (2008).

In offline comprehension, accuracy was high across all conditions (unambiguous SC = 92.8%; unambiguous DO = 92.0%; ambiguous SC= 93.1%; ambiguous DO = 92.0%; see Figure 2). As expected from the design of the comprehension questions, which probed general comprehension of the sentence rather than the DO/SC ambiguity specifically, there were no significant condition effects on accuracy ( $\text{unambiguous}:\widehat{\mathrm{\beta }}=0.050$, SE = 0.115, p = 0.65; $\text{SC-bias}:\widehat{\mathrm{\beta }}=-0.157$, SE = 0.240, p = 0.511; $\text{interaction}:\widehat{\mathrm{\beta }}=-0.212$, SE = 0.180, p = 0.52).

Figure 2.

Mean comprehension question accuracy on ambiguous and unambiguous direct object (DO)- and sentential complement (SC)-biased sentences. Error bars represent the standard errors of the mean residual reading times with the correction for within-subject factors given in Morey (2008).

Extraction effects

Online reading times in the relative clause region (after the relativizer and before the main verb) were modeled as a function of relative clause condition (object- or subject-extracted) and random effects for subjects and items. Reading times were significantly longer in the ORC sentences relative to the SRC sentences ( $\widehat{\mathrm{\beta }}=0.222$, SE = 0.050, p < 0.001), replicating the standard finding in the literature. Reading times across both sentence types are plotted in Figure 3.

Figure 3.

Mean residual reading times across sentence regions for object-(ORC) and subject-extracted (SRC) relative clause sentences. Error bars represent the standard errors of the mean residual reading times with the correction for within-subject factors given in Morey (2008). The critical region includes the entire relative clause region, beginning with either the embedded subject (ORC) or embedded verb (SRC) and ending before the main verb.

Offline comprehension accuracy across conditions is shown in Figure 4. Overall accuracy was highest when subjects were asked about the main clauses of SRC sentences (83.7%) and lowest when they were asked about the relative clauses of ORC sentences (71.5%); accuracy for the other trial types was closer to the high end (relative clause/SRC = 80.9%, main clause/ORC = 81.2%; see Figure 4). Accuracy on a given trial was modeled as a function of relative clause type, question type (whether than main clause or relative clause was probed), and random effects for subjects and items. Accuracy was significantly lower for ORC sentences ( $\widehat{\mathrm{\beta }}=-0.446$, SE = 0.143, p < 0.01). While the condition difference was numerically larger when the relative clause was probed rather than the main clause, this interaction was not significant ( $\widehat{\mathrm{\beta }}=-0.386$, SE = 0.286, p = 0.18).

Figure 4.

Mean comprehension question accuracy for object-(ORC) and subject-extracted (SRC) relative clause sentences, by question type (whether the main clause or the relative clause was probed). Error bars represent the standard errors of the mean residual reading times with the correction for within-subject factors given in Morey (2008).

Attachment preferences

As noted above, to distinguish participants’ attachment preferences from any overall bias to affirm the readings provided in the comprehension questions, we used detection-theoretic analyses (Green & Swets, 1966; Macmillan & Creelman, 2005; Murayama et al., 2014; for applications to language processing, see Fraundorf, Watson, & Benjamin, 2010; Fraundorf, Benjamin, & Watson, 2013; Lee & Fraundorf, 2017; Tokowicz & MacWhinney, 2005). In these models, the dependent variable is whether participants made a yes or no response to each comprehension question. Thus, a general response bias—across question types—to affirm the presented reading would be reflected in a significant intercept term whereas an effect of Condition (i.e., whether yes indicates a low attachment reading or a high attachment reading) on the odds of a yes response indicates whether participants preferred one type of attachment. Further, main effects of the individual-difference measures on the odds of a yes response would represent effects on the overall response bias whereas an interaction of the individual differences with the question type indicates effects on attachment preference.

The intercept term was significantly greater than 0 ( $\widehat{\mathrm{\beta }}=0.349$, SE = 0.142, p < .05), indicating that participants did indeed have some overall preference to respond yes. However, this effect was small compared to significant main effect of question type: Participants gave far more yes responses when a yes response indicated a low attachment reading ( $\widehat{\mathrm{\beta }}=2.246$, SE = 0.338, p < 0.001), consistent with prior results for attachment preferences in English. Figure 5 shows that subjects were more likely to endorse paraphrases consistent with the low attachment reading (75%) than the high attachment reading (36%).

Figure 5.

Mean proportion of yes responses to comprehension questions after sentences containing a global relative clause attachment ambiguity, according to whether the question suggested a high attachment reading (“Yes=High”) or a low attachment reading (“Yes=Low”). Error bars represent the standard errors of the mean residual reading times with the correction for within-subject factors given in Morey (2008).

Consistency of self-paced reading effects

The above analyses replicated the standard, across-participant effects from the language processing literature; for instance, ORCs were read more slowly than SRCs. However, critical for examining individual differences in syntactic processing is whether these effects consistently vary from subject to subject; that is, are there some subjects who consistently have more difficulty with ORCs than other subjects?

To assess the within-subject consistency of each effect, the data were randomly split into halves that were balanced on item and subject variables. Then, a regression model of the condition effects with random intercepts and slopes for subjects and items was run with each half of the data. Finally, the random effects for subjects from each of the two models were correlated as a measure of subject-level consistency. This entire process was repeated 100 times for each model, and the average correlation was taken as the final measure.

The results of this procedure are given in Table 7. The random intercepts for overall reading time showed a very high correlation between halves (r = 0.96 in verb bias sentences and r = 0.95 for relative-clause extraction sentences), indicating that overall individual differences in reading time differences were reliable. That is, some people were consistently faster readers than others, and we reliably measured this variability. Reliable differences in overall reading speed are expected and validate our analytical procedure as one that is capable of detecting individual differences that are known to exist. We also found that differences in overall comprehension accuracy were relatively consistent (r = .57 in verb-bias sentences and r = .94 in relative-clause extraction sentences).

Table 7.

Mean correlation of random subject effects in split halves

Model	Random effect	r	95% CI
Verb bias reading time	Intercept	0.956	[0.954, 0.958]
	Ambig	−0.051	[−0.089, −0.012]
	Bias	−0.103	[−0.133, −0.073]
	Interaction	0.237	[0.211, 0.263]
Verb bias accuracy	Intercept	0.573	[0.565, 0.582]
	Ambig	−0.176	[−0.223, −0.13]
	Bias	0.426	[0.382, 0.471]
	Interaction	0.216	[0.152, 0.281]
RC-extraction reading time	Intercept	0.948	[0.947, 0.95]
	Slope	0.053	[0.016, 0.09]
RC-extraction accuracy	Intercept	0.941	[0.941, 0.941]
	RC	−0.005	[−0.084, 0.074]
	Question	0.229	[0.163, 0.295]
	Interaction	0.198	[0.13, 0.265]
Attachment preference	Intercept	0.033	[−0.031, 0.096]
	Slope	0.678	[0.67, 0.685]

Note: Correlations were calculated by running each of the five models of condition effects on randomly-generated halves of the data and correlating the random effects. Means were calculated by repeating the correlation procedure 100 times and averaging over the results. The 95% confidence interval (CI) is also given.

But, the consistency of subject slopes for the syntactic variables (i.e., the individual differences in the syntactic-processing phenomena) was much lower; all split-half correlations had a magnitude of .24 or lower for the online measures. Thus, for instance, while we replicated the overall verb bias effect, we did not observe consistent individual differences in the size of this effect. Correlations for the offline comprehension effects were of somewhat greater magnitude, but still relatively low.

Relation of Individual Differences to Language Comprehension

The previous sections have described (a) the consistency of total scores for individual differences measures, (b) models of trial-level reading data across participants, and (c) the consistency of individual differences in the reading time effects. Next, we construct additional mixed-effects models to determine whether trial-level reading effects (i.e., level-1 effects) can be predicted from individual differences at the subject level (i.e., level 2), above and beyond the overall condition effects reported above. There is reason to think this will be unlikely: In the section above, we did not observe strong evidence that some subjects consistently showed larger syntactic processing effects than others (as well, some of the other individual-difference measures were only moderately consistent). Since individual differences in syntactic processing either largely do not exist or could not be reliably measured, it does not make sense to expect that such differences could be related to the other constructs. Nevertheless, we considered on an exploratory basis whether individual differences in syntactic processing might be associated with the five other individual-difference constructs. In doing so, we are guided by the fact that relatively few other studies have assessed syntactic processing in conjunction with multiple other individual-difference constructs and across multiple syntactic phenomena; therefore, a gap in the literature could be filled by examining what relations might be observed in such data despite the limited reliability of the syntactic processing effects.

Reading times were examined as a simultaneous function of (a) syntactic condition variables for the relevant sentence type, (b) composite scores of all five cognitive domains measured in our individual differences battery, and (c) the interaction of the individual-difference scores with the syntactic condition variable(s). Individual differences that affect syntactic processing (e.g., individual differences that differentially affect ORCs as opposed to SRCs) should be realized as an interaction between one of the individual-difference variables and syntactic condition. The purpose of including all five individual difference domains simultaneously in each regression was to allow us to interpret effects of one domain as accounting for a share of the variance independent of the other domains: As in other multiple regression models, parameter estimates in a mixed-effects regression reflect the effect of varying one variable (e.g., verbal working memory) while holding others constant (Baayen, 2008, p. 192; Wurm & Fisicaro, 2014). We included the composite individual-difference scores as continuous predictors to reflect the full range of these variables across individuals. Including continuous variation is more powerful than a median split (Cohen, 1983) and also yields more accurate estimates of effect size and lower rates of Type I error (MacCallum, Zhang, Preacher, & Rucker, 2002; Preacher, Rucker, MacCallum, & Nicewander, 2005). Appendix C presents the complete equations for these models.

For verb bias, although we replicated the overall verb bias effect in online reading (Table 8), we did not find that individual differences in the size of this effect were related to any of the other. Higher perceptual speed composite scores predicted faster reading times overall ( $\widehat{\mathrm{\beta }}=-0.094$, SE = 0.036, p < 0.01), but none of the individual difference measures significantly interacted with the syntactic condition effects. In offline performance (Table 9), there was a significant main effect of language experience, with higher scores leading to higher overall accuracy ( $\widehat{\mathrm{\beta }}=0.508$, SE = 0.115, p < 0.001). Phonological ability also significantly interacted with ambiguity; it more strongly benefited accuracy in the ambiguous condition ( $\widehat{\mathrm{\beta }}=-0.314$, SE = 0.138, p < 0.05).

Table 8.

Fixed effects in model of residual reading times in Verb Bias sentences

	Fixed effect	Estimate	SE	p-value
	(Intercept)	−0.102	0.042	0.015
Individual differences	VWM	−0.003	0.040	0.938
	Inhib.	−0.051	0.057	0.375
	Lang.	−0.010	0.048	0.835
	Phon.	−0.002	0.048	0.971
	Speed	−0.094	0.036	0.010
Condition effects	Ambiguous	0.071	0.016	<0.001
	DO Bias	0.033	0.057	0.567
	Ambiguous × DO Bias	0.096	0.031	0.003
Individual difference × Condition effect interactions	VWM × Ambiguous	−0.005	0.014	0.710
	VWM × DO Bias	0.015	0.014	0.298
	Inhib. × Ambiguous	−0.001	0.021	0.959
	Inhib. × DO Bias	−0.011	0.020	0.575
	Lang. × Ambiguous	−0.005	0.017	0.783
	Lang. × DO Bias	−0.004	0.017	0.815
	Phon. × Ambiguous	0.023	0.018	0.200
	Phon. × DO Bias	0.017	0.017	0.332
	Speed × Ambiguous	0.001	0.013	0.910
	Speed × DO Bias	−0.021	0.013	0.095
	VWM × Ambiguous × DO Bias	−0.019	0.029	0.509
	Inhib. × Ambiguous × DO Bias	−0.023	0.041	0.573
	Lang. × Ambiguous × DO Bias	−0.022	0.034	0.518
	Phon. × Ambiguous × DO Bias	0.020	0.035	0.564
	Speed × Ambiguous × DO Bias	−0.045	0.026	0.089

Notes: Effects coding was used for condition effects. Condition effects here refer to the change in residual reading time when sentences were ambiguous (0.5) opposed to unambiguous (−0.5) and DO-biased (0.5) opposed to SC-biased (−0.5). Random intercepts and slopes for all condition effects for both subjects and items were also included in the model. “VWM” = Verbal working memory span; “Inhib” = Inhibitory control; “Speed” = Perceptual speed; “Phon.” = Phonological ability; “Lang.” = Language experience.

Table 9.

Fixed effects in model of comprehension accuracy in Verb Bias sentences

	Fixed effect	Estimate	SE	p-value
	(Intercept)	3.274	0.142	<0.001
Individual differences	VWM	0.052	0.087	0.552
	Inhib.	0.223	0.125	0.075
	Phon.	0.086	0.106	0.416
	Lang.	0.508	0.115	<0.001
	Speed	0.000	0.078	0.999
Condition effects	Unambiguous	0.109	0.128	0.392
	SC Bias	−0.231	0.261	0.377
	Unambiguous × SC Bias	−0.238	0.201	0.236
Individual difference × Condition effect interactions	VWM × Unambiguous	0.057	0.113	0.614
	VWM × SC Bias	−0.057	0.110	0.601
	Inhib. × Unambiguous	−0.092	0.163	0.574
	Inhib. × SC Bias	−0.137	0.159	0.388
	Phon. × Unambiguous	−0.314	0.138	0.023
	Phon. × SC Bias	0.104	0.134	0.438
	Lang. × Unambiguous	0.312	0.162	0.054
	Lang. × SC Bias	−0.270	0.159	0.089
	Speed × Unambiguous	−0.016	0.101	0.875
	Speed × SC Bias	0.059	0.098	0.545
	VWM × Unambiguous × SC	0.165	0.219	0.453
	Bias
	Inhib. × Unambiguous × SC	−0.076	0.318	0.812
	Bias
	Phon. × Unambiguous × SC	−0.059	0.269	0.826
	Bias
	Lang. × Unambiguous × SC	−0.555	0.318	0.081
	Bias
	Speed × Unambiguous × SC	−0.347	0.197	0.079
	Bias

Notes: Effects coding was used for condition effects. Condition effects here refer to the change in log odds of a correct response when the sentences were ambiguous (0.5) opposed to unambiguous (−0.5) and DO-biased (0.5) opposed to SC-biased (−0.5). Random intercepts and slopes for all condition effects for both subjects and items were also included in the model. “VWM” = Verbal working memory span; “Inhib” = Inhibitory control; “Speed” = Perceptual speed; “Phon.” = Phonological ability; “Lang.” = Language experience.

Similarly, for the extraction effects, although we replicated the overall difference between ORCs and SRCs in online reading (Table 10), the individual difference measures did not reveal any significant interactions with RC type. Again, however, there was an effect of perceptual speed on overall reading speed ( $\widehat{\mathrm{\beta }}=-0.323$, SE = 0.121, p < 0.01). There were also individual differences in offline comprehension (Table 11): Overall accuracy was significantly associated with higher scores in verbal working memory ( $\widehat{\mathrm{\beta }}=0.378$, SE = 0.096, p < 0.001) and language experience ( $\widehat{\mathrm{\beta }}=0.437$, SE = 0.123, p < 0.001). There was also a significant three-way interaction among language experience and the two condition effects; how much language experience benefits accuracy depends on both clause type and question type, such that benefit is most positive when the condition interaction is least extreme (i.e., ORCs and main clause questions, or SRCs and relative clause questions) ( $\widehat{\mathrm{\beta }}=-0.653$, SE = 0.314, p < 0.05).

Table 10.

Fixed effects in model of residual reading times in Relative Clause Extraction sentences

	Fixed effect	Estimate	SE	p-value
	(Intercept)	−0.222	0.164	0.179
Individual differences	VWM	0.069	0.135	0.611
	Inhib.	−0.106	0.194	0.586
	Phon.	0.070	0.164	0.671
	Lang.	−0.099	0.162	0.542
	Speed	−0.323	0.121	0.009
Condition effect Individual difference × Condition interactions	ORC	0.217	0.056	<0.001
	VWM × ORC	0.000	0.062	0.994
	Inhib. × ORC	−0.117	0.088	0.189
	Phon. × ORC	0.031	0.075	0.683
	Lang. × ORC	−0.114	0.073	0.121
	Speed × ORC	0.075	0.055	0.173

Notes: Effects coding was used for the condition effect, and contrasts were weighed to correct for the slight imbalance of observations in each condition. The condition effect here refers to the change in reading time when given an object-extracted relative clause (0.52) opposed to subject-extracted (−0.48). Random intercepts and slopes for all condition effects for both subjects and items were also included in the model. “VWM” = Verbal working memory span; “Inhib” = Inhibitory control; “Speed” = Perceptual speed; “Phon.” = Phonological ability; “Lang.” = Language experience.

Finally, for the attachment-preference items (Table 12), we did observe that several constructs related to individual differences in offline interpretation of these ambiguous items. Specifically, lower verbal working memory was associated with a stronger preference for high attachment ( $\widehat{\mathrm{\beta }}=1.020$, SE = 0.292, p < 0.001), consistent with Swets et al. (2007) and Payne et al. (2014). Lower processing speed was also related to a high attachment preference ( $\widehat{\mathrm{\beta }}=0.656$, SE = 0.255, p < 0.05). Additionally, both verbal working memory and language experience had significant effects on the overall response bias such that lower scores were associated with a higher yes bias ( $\text{VWM}:\widehat{\mathrm{\beta }}=-0.471$, SE = 0.156, p < 0.01; $\text{language}\text{experience}:\widehat{\mathrm{\beta }}=-0.415$, SE = 0.192, p < 0.05).

Table 12.

Fixed effects in model of relative clause attachment preferences

	Fixed effect	Estimate	SE	p-value
	(Intercept)	0.339	0.150	0.023
Condition effect Individual differences	Low Attach.	2.464	0.344	<0.001
	VWM	−0.003	0.084	0.969
	Inhib.	−0.036	0.122	0.770
	Phon.	−0.002	0.103	0.985
	Lang.	−0.246	0.103	0.016
	Speed	0.049	0.080	0.539
Individual difference × Condition interactions	Low Attach. × VWM	0.937	0.275	0.001
	Low Attach. × Inhib.	−0.042	0.399	0.915
	Low Attach. × Phon.	0.251	0.338	0.458
	Low Attach. × Lang.	0.338	0.335	0.313
	Low Attach. × Speed	0.655	0.255	0.010

Notes: Effects coding was used for the condition effects. The condition effect here refers to the likelihood of answering yes to a comprehension question when it promoted the low attachment reading (0.5) opposed to the high attachment reading (−0.5). Random intercepts and slopes for all condition effects for both subjects and items were also included in the model. “VWM” = Verbal working memory span; “Inhib” = Inhibitory control; “Speed” = Perceptual speed; “Phon.” = Phonological ability; “Lang.” = Language experience.

Discussion

Although we replicated across-participant online effects of verb bias and relative clause extraction type, individual differences in the magnitude of these effects were seen only offline. Verbal working memory and language experience were both shown to significantly relate to overall offline accuracy for different sentence types, and each also interacted with characteristics of the sentences: Language experience interacted with verb bias in offline comprehension of the verb bias sentences, while lower verbal working memory and slower perceptual speed were associated with a stronger high-attachment preference for RC attachment ambiguities. Below, we first discuss the absence of online effects, then the theoretical and practical implications of the offline findings that we did observe.

Effects of Individual Differences Are Offline, Not Online

Despite theories of sentence comprehension that predict online effects of verbal working memory, language experience, and other cognitive abilities, our study did not yield any interactions between the reading time effects and the individual differences assessed with our battery of tasks.

Why did we observe no relation between these individual differences and online syntactic processing? One possibility is that we simply did not measure the right individual-difference construct. However, any relation between an individual-difference construct and online syntactic processing would have actually been unlikely given that individual differences in syntactic processing—as well of some of the other individual-difference measures—had only moderate to low consistency to begin with. Thus, for instance, although the overall verb bias effect was robust across subjects, we did not observe strong evidence that some individuals consistently showed a larger verb bias effect than others. Because we did not observe strong individual differences in syntactic processing to begin with, no other construct could be expected to explain them. Similarly, an absence of relations between online syntactic processing and other individual differences could in principle reflect low statistical power owing to an insufficient sample size. However, the consistency of individual differences in syntactic processing was so low that it is unlikely that even a larger sample would have related them to other individual differences.

It is important to emphasize that these limitations are specific to the consistency of individual differences in the syntactic phenomena of interest. We do not claim that the self-paced reading task fails to reliably measure reading time in general or even that it fails to measure individual differences in reading time. In fact, as noted above, the split-half correlations for individual differences in overall reading speed, as assessed by subject-level random intercepts for reading time, were quite high (all rs > .9). Further, these differences in overall reading time correlated with individual differences in perceptual speed. Rather, it was specifically individual differences in the magnitude of the syntactic processing effects that were not consistent. Nor do we claim that the general (across-participant) syntactic processing phenomena of interest cannot be reliably observed. Indeed, when averaging across subjects, we replicated the standard findings from the literature (e.g., that ORCs are read more slowly than SRCs) for all of the sentence types presented here. That is, we observed both clear reader differences (differences in baseline reading speed and comprehension accuracy) and clear text differences (effects of verb bias and of relative clause extraction type). What we did not observe, at least at the level of syntactic processing, were consistent reader-text interactions whereby certain syntactic structures were particularly challenging for some readers.

In fact, the consistency across subjects of these syntactic-processing effects likely contributes to the absence of consistent individual differences. In general, effects that are robust and consistent across participants often make poorer individual-difference measures precisely because everyone exhibits the effect to similar degrees and there are few individual differences (see, for instance, Salthouse, Sieldlecki & Krueger, 2006, for similar results in memory control). These points were recently articulated and labeled by Hedge, Powell, and Sumner (2017) as the reliability paradox. The authors suggest that “experimental designs have been developed and naturally selected for providing robust effects, which means low between-participant variance” (p. 17-18).

More broadly, the distinction between robust experimental effects and effective individual-difference measures reflects what are often differing goals in the two research traditions described by Cronbach (1957) and summarized in our introduction here: Experimental research most frequently seeks general principles of cognition that generalize across persons whereas individual-differences research generally seeks those characteristics that differentiate individuals.

In light of these principles, one possible explanation for why we did not observe consistent differences in online syntactic processing is that there are little or no individual differences in syntactic processing to begin with—that is, all readers find (for instance) ORCs more difficult than SRCs to similar degrees. Under this hypothesis, improvements in reliability of the measures would not reveal a relationship between online syntactic processing performance and other constructs because the underlying relationship does not exist. This claim would be consistent with the general framework put forth by Caplan and Waters (1999) suggesting a distinction between interpretive and post-interpretive processing, with span measures relating only to the latter. Caplan and Waters argued that interpretive processes, including word recognition and syntactic parsing, require a different resource pool than post-interpretive processes, such as encoding and reasoning about the input. This theory is supported by other work showing that working memory span is unrelated to eye-tracking measures of differences between ORCs and SRCs in free reading (Traxler et al., 2005).

The other possibility, of course, is that there are individual differences in online syntactic processing, but the present study simply failed to measure them. Although we did find that the self-paced reading task reliably measured differences in overall reading speed, it is possible that features of the self-paced reading task itself may obscure individual differences in syntactic processing more specifically. Unlike in natural reading, the moving-window technique does not allow re-reading of prior material and requires readers to manually proceed through the sentence, making reading speed about twice as slow (Rayner, 1998, p. 391). These differences from natural reading may obscure typical individual tendencies in per-word reading times while still allowing variation in subsequent comprehension. Thus, it could be informative for future work to examine whether greater internal consistency in syntactic processing measures is obtained with eye-tracking of free reading. Some evidence suggests that overall subject-level differences in eye movements (e.g., individual differences in total reading time per word) can be reliably measured in free reading (Carter & Luke, 2016; Traxler et al., 2005). Further, at least one study (Traxler & Tooley, 2007) found that linguistic experience did correlate with individual differences in syntactic processing of DO/SC ambiguity items, unlike in the present study; this discrepancy might reflect the fact that Traxler and Tooley (2007) measured eye-tracking of free reading rather than the self-paced reading task.

Regardless of its cause, we argue that the lack of consistent subject-level variation in syntactic processing condition effects in self-paced reading is unlikely to be exclusive to the present dataset. The present task and materials were very similar to those used in other self-paced reading investigations of syntactic processing, suggesting that this lack of consistency is likely to extend to syntactic effects in self-paced reading time more broadly. In fact, low levels of internal consistency have been observed not just for online syntactic processing but for some other prospective individual differences derived from cognitive experiments, such as event-related potential measures of language comprehension (Tanner & Bulkes, 2015) and perspective-taking in comprehension (Brown-Schmidt & Fraundorf, 2015; Ryskin, Benjamin, Tullis, & Brown-Schmidt, 2015). The scope of these limitations is unclear: While standardized measures of domain-general cognitive ability and of linguistic experience have often been normed and show reliability, examining and reporting the consistency of individual differences in online language processing tasks themselves is less common (a problem that has also been noted elsewhere; Ryskin et al., 2015). It would be helpful for future investigations of individual differences in language processing to measure the internal consistency of individual differences in the online language processing measures themselves.

Offline Effects of Verbal Working Memory Capacity

Although we did not observe consistent individual differences in the measures of online syntactic processing, we did see one such effect in the offline measures. Namely, participants with lower verbal working memory more strongly preferred high attachment of ambiguous relative clauses in the current study, consistent with Swets et al. (2007) and Payne et al. (2014). Swets et al. (2007) present evidence that this relation obtains because individuals with limited memory resources adopt a chunking strategy that generates an implicit prosodic representation amenable to a high-attachment reading, analogous to the effects of explicit prosody (for a review of those effects, see Frazier, Carlson, & Clifton, 2006). These results contradict predictions that people with lower memory resources would prefer the low-attachment reading because it minimizes distance of the dependency (e.g., Gibson, 1998, 2000; Traxler, 2007). Rather, this finding supports the account that internal prosody may be an important strategy for readers with low working memory (see Swets et al., 2007).

Lower verbal working memory was also associated with lower overall accuracy on RC extraction sentences, consistent with other findings that verbal working memory relates to general reading comprehension accuracy (Daneman & Merikle, 1996, for meta-analysis). Note, however, that this effect of working memory was a main effect across all of the RC extraction sentences and did not differentially affect ORCs relative to SRCs. Thus, it can be best characterized as a reader effect—lower-span readers have poorer comprehension—rather than a reader-text interaction whereby lower-span readers are particularly disadvantaged with specific types of relative clauses.

Nevertheless, it is noteworthy that an effect of verbal working memory on reading comprehension emerged even when other factors, including language experience, were included in the model. It is worth noting that our span tasks included a calibration phase intended to help deconfound language experience and span, and the results presented above in Table 5 indicate that this procedure was generally successful. It is possible that this step allowed us to observe an independent contribution of verbal working memory span. However, while we appear to have deconfounded language experience and verbal working memory, there is still the possibility that the effects of language experience reflect some other construct we did not measure, such as IQ; recent evidence suggests that working memory is not related to reading comprehension ability once IQ is accounted for (Van Dyke et al., 2014). Finally, it should be noted that all our measures concerned exclusively verbal working memory; we can make no claims about how syntactic processing may be affected by any other possible forms of working memory (e.g., visuospatial working memory; Shah & Miyake, 1996).

Offline Effects of Language Experience

Other results from offline comprehension provided mixed evidence in favor of experience-based theories of language comprehension (e.g., Garnsey et al., 1997; Levy, 2008; MacDonald & Christiansen, 2002), which suggest that individual differences in the amount or quality of language experience might account for differences among individuals in their ability to process difficult syntactic structures.

Higher scores on the language experience measures were indeed related to higher overall comprehension accuracy on both verb bias and RC extraction sentences, consistent with several studies demonstrating that individuals with more exposure to language have higher reading comprehension skills (e.g., Stanovich, 1985). However, in the offline measures, there was only inconsistent evidence for the language experience by syntax interactions that would indicate language comprehension facilitated comprehension of specific syntactic structures, and we did not observe effects of language experience at all on self-paced reading times. As mentioned previously, it is unlikely that we would observe such relations given that we did not find consistent individual differences in the online syntactic effects to begin with. Further, it is unclear that more total language experience would necessarily improve processing of the dispreferred structure. Rather, it may be the relative exposure to these more difficult structures that is crucial. For instance, in the training studies by Wells et al. (2009) and Fine et al. (2013), comprehension of the uncommon structures may have improved with additional exposures because those specific structures were proportionally more frequent in the training input than in the typical distributional statistics participants had experienced previously. In fact, increased overall exposure to the typical distributional statistics of English may only strengthen biases against the statistically dispreferred structure, such as the high-attachment processing cost observed for high-print-exposure older adults in Payne and colleagues’ (2014) study.

Another possibility is that the differences in exposure within our educated adult sample are not great enough to modulate online effects. However, this possibility is unlikely because scores on the individual differences tasks, including the language experience tasks, were well distributed across the range of possible scores rather than being clustered at the ceiling (see Table 2). Further, the correlations across tasks are evidence against a ceiling effect; the fact that some subjects consistently score higher than others across tasks suggests that not all subjects are at ceiling, and that there are individual differences even within this restricted range. Nonetheless, including a wider range of prior language experience should be a goal of future research.

Conclusion

Each experience of reading brings together both a reader and a to-be-read text. Research in educational psychology and individual differences has made it clear that some people are more skilled readers than others, and psycholinguistics has revealed how some syntactic structures that may appear in a text are more difficult than others (e.g., subject-extracted relative clauses are more difficult than object-extracted relative clauses). What has been less clear, at least in the domain of syntactic processing, is whether there are reader-text interactions: Are there particular syntactic structures that are especially challenging for particular readers, and are there some readers who are especially advantaged at reading otherwise difficult structures?

We investigated whether such interactions exist in syntactic processing and, if so, what other individual differences might drive them. In doing so, we were guided by several insights from the correlational approach: We measured multiple individual-difference constructs, we obtained multiple measures of each construct, and we assessed the consistency of our measures. Our results suggest a possible reason for the lack of consensus across studies that have examined individual differences in syntactic processing: the relatively low consistency of those differences, especially in online (rather than offline) measures. Although we replicated well-studied syntactic phenomena overall (e.g., the effects of verb distributional statistics), we found low consistency of individual differences in those effects: That is, it was not the case that some subjects consistently showed (for instance) large verb bias effects and other subjects consistently showed small verb bias effects.

By contrast, we did observe both reliable differences among individuals in their overall reading speed and in baseline comprehension accuracy (i.e., reader effects). We also observed that some syntactic structures were more challenging than others (i.e., text effects). What we found little evidence of—within the domain of syntactic processing—were reader-text interactions whereby some syntactic structures were differentially more difficult for some readers than others. Rather, good readers were comparatively good with all syntactic structures, and syntactically challenging sentences were more difficult for all readers, and these two effects did not interact. This result is not unlike those observed in at least some other domains; for example, it has been argued that there is little evidence for interactions between instructional methods and an individual’s learning style (Pashler et al., 2008). However, reader-text interactions have been observed for some aspects of language processing other than syntactic processing, such as lexical or discourse-level processing (e.g., Seidenberg, 1985; Stine-Morrow et al., 2008).

Our results are thus consistent with psycholinguistic theories in which initial stages of syntactic processing are relatively automatic and not influenced by domain-general processes (Caplan & Waters, 1999). And, they imply that the skills and abilities that educational psychologists have identified as underlying reading success are likely relevant across a broad range of syntactic structures. Consequently, it may not be necessary to tailor texts or interventions to particular readers, at least at the level of syntax. Nevertheless, it will be important to bring the language-processing effects demonstrated by experimental psychologists together with the important individual differences identified by correlational approach in order to identify how they might or might not interact—and to progress towards a united discipline of correlational and experimental traditions in psychology.

Highlights.

• We replicate verb bias, relative clause extraction, and attachment ambiguity effects

• Online syntactic effects are unreliable as measures of individual differences

• High attachment for relative clauses is associated with lower working memory

• Language experience and working predict overall comprehension accuracy

• Perceptual speed predicts overall reading times

Appendix A

The current study was designed to assess five latent constructs with sixteen tasks. In the primary analysis, tasks scores were standardized and averaged within-construct to create five composite scores for each subject. This method of creating composite scores assumes that (1) the a priori grouping of tasks into constructs is valid and (2) each task score contributes equally to the constructs they purport to measure. While this method is intuitive and straightforward to replicate, the assumptions need to be tested.

The following is a confirmatory factor analysis (CFA) approach that tests these assumptions. Individual standardized task scores were entered into CFA models using the lavaan package (Rosseel, 2012) in R software (R Core Team, 2017). Three models are presented in this Appendix.

Model 1 (Figure A1) is structured according to the a priori study design: tasks are scored as described in the main text; “Language Experience” has five indicators (ERVT, ART, NAART, CRH, and RTE), “Working Memory” has three indicators (RSpan, LSpan, and OSpan), “Phonological Ability” has three indicators (Gupta, BNW, and PR), “Inhibitory Control” has three indicators (Anti-Saccade, Stroop, and Flanker), and “Perceptual Speed” has two indicators (Letter Comparison and Pattern Comparison).

Models 2 and 3 are updated models given the results of Model 1. In Model 2 (Figure A2), the factor loadings from “Inhibitory Control” to Stroop and Flanker are improved by changing task scores from the difference between conflict and neutral trial RTs to RTs from conflict trials only. Reliability of the measures is improved, and the use of conflict RTs is consistent with the Anti-Saccade measure (the Anti-Saccade task only includes conflict trials). However, all three measures of Inhibitory Control now conflate differences in inhibition ability with overall differences in speed.

Model 3 improves Model 2 by splitting the “Language Experience” factor into two separate factors; “Language Skill” is indicated by ERVT, ART, and NAART, while “Language Survey” is indicated by CRH and RTE.

The CFA models provide more information about the relationships among the observed variables and their relative strengths as indicators. Additionally, the better fit of the six-factor model suggests that future work would benefit from subdividing the multifaceted construct into more specific components, each with multiple indicators chosen a priori.

Overall, however, the results of this analysis are consistent with those of the regression analysis with composite scores provided in the main text, and is subject to similar limitations. Namely, task scores are derived by summarizing across the individual trials within-subject (e.g. as a span score, proportion correct, etc.) before entering the model, rather than creating a hierarchical model that treats items as indicators for the task score, which are in turn indicators of the latent variables. There is not sufficient data in the current sample to estimate a model of that complexity. As in the main analysis, certain tasks are clearly less stable than others. While the CFA analyses suggest that validity could be improved by regrouping task scores and giving them unequal weight before forming a composite. However, as post hoc changes are data-driven, the models derived from the current data set should ideally be validated against a new dataset. As the current study includes widely-used tasks of theoretically important constructs, validation of the model structure would be a valuable goal for future work.

Figure A.1.

Model 1: Confirmatory factor analytic model relating sixteen observed variables (represented in rectangles) to five latent variables (represented in ovals). Arrows between latent and observed variables label the factor loadings. Values in gray are loadings that are not significant. Curved arrows label correlations between latent variables. The fit indices for this model are as follows: p(χ²) = 0.003, SRMR = 0.082, RMSEA = 0.058, CFI = 0.899. The p(χ²) and CFI values indicate a poor model fit.

Figure A.2.

Model 2: A confirmatory factor analytic model relating sixteen observed variables (represented in rectangles) to five latent variables (represented in ovals). Arrows between latent and observed variables label the factor loadings. Values in gray are loadings that are not significant. Curved arrows label correlations between latent variables. The fit indices for this model are as follows: p(χ²) = 0.005, SRMR = 0.088, RMSEA = 0.057, CFI = 0.917.

Figure A.3.

Model 3: A six-factor confirmatory factor analytic model derived from Model 2 by splitting the “Language Experience” factor into “Language Skill” and “Language Survey” factors. Arrows between latent (ovals) and observed (rectangles) variables label the factor loadings. Curved arrows label correlations between latent variables. The fit indices for this model are as follows: p(χ²) = 0.036, SRMR = 0.081, RMSEA = 0.047, CFI = 0.947.

Appendix B

Mixed-effect model equations for models of syntactic effects with experimental conditions only.

Residual reading time for verb bias items was modeled as:

$$\begin{gathered} Y_{ij}={\gamma }_{000}+{\gamma }_{100}\ast \text{Ambiguity}+{\gamma }_{200}\ast \text{Bias}+{\gamma }_{1200}\ast \text{Ambiguity}\ast \text{Bias}+u_{0\mathrm{i}0}+u_{1\mathrm{i}0}\ast \text{Ambiguity}+ \\ {u}_{2\mathrm{i}0}\ast \text{Bias}+u_{12i0}\ast \text{Ambiguity}\ast \text{Bias}+v_{00\mathrm{j}}+{\epsilon }_{ij} \end{gathered}$$

where Y_ij is the residual reading time for subject i on item j, γ₀₀₀ is an intercept term representing grand mean residual reading time, γ₁₀₀, γ₂₀₀, and γ₁₂₀₀ are the fixed effects of the experimental conditions, u_0i0 is the error for the subject intercept for subject i, u_1i0, u2i0, and u_12i0 are the error terms for the random subject slopes for the experimental conditions for subject i, v_00j is the error for the item intercept for item j, and ε_ij is the error for subject i on item j.

Comprehension accuracy for verb bias items was modeled as:

$$\begin{gathered} \log \left(Y_{\text{ij}}\right)={\gamma }_{000}+{\gamma }_{100}\ast \text{Ambiguity}+{\gamma }_{200}\ast \text{Bias}+{\gamma }_{1200}\ast \text{Ambiguity}\ast \text{Bias}+u_{0\mathrm{i}0}+u_{1\mathrm{i}0}\ast \text{Ambiguity}+ \\ {u}_{2\mathrm{i}0}\ast \text{Bias}+u_{12i0}\ast \text{Ambiguity}\ast \text{Bias}+v_{00\mathrm{j}}+v_{10\mathrm{j}}\ast \text{Ambiguity} \end{gathered}$$

where Y_ij are the odds of subject i correctly responding to item j, γ₀₀₀ is an intercept term representing grand mean accuracy, γ₁₀₀, γ₂₀₀, and γ₁₂₀₀ are the fixed effects of the experimental conditions, u_0i0 is the error for the subject intercept for subject i, u_1i0, u2i0, and u_12i0 are the error terms for the random subject slopes for the experimental conditions for subject i, v_00j is the error for the item intercept for item j, and v_10j is the error for the random slope of ambiguity for item j, and ε_ij is the error for subject i on item j.

Residual reading time for relative-clause extraction items was modeled as:

$$Y_{ij}={\gamma }_{000}+{\gamma }_{100}\ast \text{RCType}+u_{0\mathrm{i}0}+u_{1\mathrm{i}0}\ast \text{RCType}+v_{00\mathrm{j}}+v_{10\mathrm{j}}\ast \text{RCType}+{\epsilon }_{ij}$$

where Y_ij is the residual reading time for subject i on item j, γ₀₀₀ is an intercept term representing grand mean residual reading time, γ₁₀₀ is the fixed effect of experimental condition, u_0i0 is the error for the subject intercept for subject i, u_1i0 is the error term for the random slope of condition for subject i, v_00j is the error for the item intercept for item j, v_10j is the error for the random slope of condition for item j, and ε_ij is the error for subject i on item j.

Comprehension accuracy for relative-clause extraction items was modeled as:

$$\begin{gathered} \log \left(Y_{ij}\right)={\gamma }_{000}+{\gamma }_{100}\ast \text{RCType}+{\gamma }_{200}\ast \text{QuestionType}+{\gamma }_{1200}\ast \text{RCType}\ast \text{QuestionType}+u_{0\mathrm{i}0}+ \\ {u}_{1\mathrm{i}0}\ast \text{RCType}+u_{2i0}\ast \text{QuestionType}+u_{12i0}\ast \text{RCType}\ast \text{QuestionType}+v_{00\mathrm{j}}+v_{10\mathrm{j}}\ast \text{RCType} \\ +v_{20\mathrm{j}}\ast \text{QuestionType}+v_{120\mathrm{j}}\ast \text{RCType}\ast \text{QuestionType} \end{gathered}$$

where Y_ij are the odds of subject i correctly responding to item j, γ₀₀₀ is an intercept term representing grand mean accuracy, γ₁₀₀, γ₂₀₀, and γ₁₂₀₀ are the fixed effects of the experimental conditions, u_0i0 is the error for the subject intercept for subject i, u_1i0, u_2i0, and u_12i0 are the error terms for the random subject slopes for the experimental conditions for subject i, v_00j is the error for the item intercept for item j, and v_10j, v_20j, and v_120j are the error terms for the random slopes for the experimental conditions for item j, and ε_ij is the error for subject i on item j.

Responses to the attachment ambiguity items were modeled as:

$$\log (Y_{\text{ij}})={\gamma }_{000}+{\gamma }_{100}\ast \text{QuestionType}+u_{0\mathrm{i}0}+u_{1\mathrm{i}0}\ast \text{QuestionType}+v_{00\mathrm{j}}+v_{10\mathrm{j}}\ast \text{QuestionType}$$

where Y_ij are the odds of subject i answering yes to item j, γ₀₀₀ is an intercept term representing the grand mean of answering yes (response bias), γ₁₀₀ is the fixed effect of the question type condition (low- or high-attachment) on yes responses (sensitivity), u_0i0 is the error for the subject intercept for subject i, u_1i0 is the error term for the random subject slope of condition for subject i, v_00j is the error for the item intercept for item j, and v_10j is the error term for the random slope of condition for item j.

Appendix C

Mixed-effect model equations for models of syntactic effects with experimental conditions and individual differences.

Residual reading time for verb bias items was modeled as:

$$\begin{gathered} Y_{ij}={\gamma }_{000}+{\gamma }_{100}\ast \text{Ambiguity}+{\gamma }_{200}\ast \text{Bias}+{\gamma }_{1200}\ast \text{Ambiguity}\ast \text{Bias}+{\gamma }_{300}\ast \text{VWM}+{\gamma }_{400}\ast \text{Inhib}+{\gamma }_{500} \\ \ast \text{Phon}+{\gamma }_{600}\ast \text{Speed}+{\gamma }_{700}\ast \text{Lang}+{\gamma }_{1300}\ast \text{VWM}\ast \text{Ambiguity}+{\gamma }_{1400}\ast \text{Inhib}\ast \text{Ambiguity}+ \\ {\gamma }_{1500}\ast \text{Phon}\ast \text{Ambiguity}+{\gamma }_{1600}\ast \text{Speed}\ast \text{Ambiguity}+{\gamma }_{1700}\ast \text{Lang}\ast \text{Ambiguity}+ \\ {\gamma }_{2300}\ast \text{VWM}\ast \text{Bias}+{\gamma }_{2400}\ast \text{Inhib}\ast \text{Bias}+{\gamma }_{1500}\ast \text{Phon}\ast \text{Bias}+{\gamma }_{2600}\ast \text{Speed}\ast \text{Ambiguity}+ \\ {\gamma }_{2700}\ast \text{Lang}\ast \text{Ambiguity}+{\gamma }_{12300}\ast \text{VWM}\ast \text{Ambiguity}\ast \text{Bias}+{\gamma }_{12400}\ast \text{Inhib}\ast \text{Ambiguity}\ast \text{Bias} \\ +{\gamma }_{12500}\ast \text{Phon}\ast \text{Ambiguity}+{\gamma }_{12600}\ast \text{Speed}\ast \text{Ambiguity}\ast \text{Bias}+ \\ {\gamma }_{12700}\ast \text{Lang}\ast \text{Ambiguity}\ast \text{Bias}+u_{0\mathrm{i}0}+u_{1\mathrm{i}0}\ast \text{Ambiguity}+u_{2\mathrm{i}0}\ast \text{Bias}+u_{12i0}\ast \text{Ambiguity}\ast \text{Bias}+v_{00\mathrm{j}}+{\epsilon }_{ij} \end{gathered}$$

where Y_ij is the residual reading time for subject i on item j, γ₀₀₀ is an intercept term representing grand mean residual reading time, γ₁₀₀, γ₂₀₀, and γ₁₂₀₀ are the fixed effects of the experimental conditions, γ₃₀₀ through γ₇₀₀ are the fixed effects of the individual-difference composites on overall reading time, γ₁₃₀₀ through γ₁₇₀₀ are the fixed effects of the individual-difference composites on the ambiguity effect, γ₂₃₀₀ through γ₂₇₀₀ are the fixed effects of the individual-difference composites on the verb bias effect, γ₁₂₃₀₀ through γ₁₂₇₀₀ are the fixed effects of the individual-difference composites on the ambiguity x verb bias interaction, u_0i0 is the error for the subject intercept for subject i, u_1i0, u2i0, and u_12i0 are the error terms for the random subject slopes for the experimental conditions for subject i, v_00j is the error for the item intercept for item j, and ε_ij is the error for subject i on item j.

Comprehension accuracy for verb bias items was modeled as:

$$\begin{gathered} \log \left(Y_{ij}\right)={\gamma }_{000}+{\gamma }_{100}\ast \text{Ambiguity}+{\gamma }_{200}\ast \text{Bias}+{\gamma }_{1200}\ast \text{Ambiguity}\ast \text{Bias}+{\gamma }_{300}\ast \text{VWM}+{\gamma }_{400}\ast \text{Inhib}+ \\ {\gamma }_{500}\ast \text{Phon}+{\gamma }_{600}\ast \text{Speed}+{\gamma }_{700}\ast \text{Lang}+{\gamma }_{1300}\ast \text{VWM}\ast \text{Ambiguity}+{\gamma }_{1400}\ast \text{Inhib}\ast \text{Ambiguity} \\ +{\gamma }_{1500}\ast \text{Phon}\ast \text{Ambiguity}+{\gamma }_{1600}\ast \text{Speed}\ast \text{Ambiguity}+{\gamma }_{1700}\ast \text{Lang}\ast \text{Ambiguity}+ \\ {\gamma }_{2300}\ast \text{VWM}\ast \text{Bias}+{\gamma }_{2400}\ast \text{Inhib}\ast \text{Bias}+{\gamma }_{1500}\ast \text{Phon}\ast \text{Bias}+{\gamma }_{2600}\ast \text{Speed}\ast \text{Ambiguity}+ \\ {\gamma }_{2700}\ast \text{Lang}\ast \text{Ambiguity}+{\gamma }_{12300}\ast \text{VWM}\ast \text{Ambiguity}\ast \text{Bias}+{\gamma }_{12400}\ast \text{Inhib}\ast \text{Ambiguity}\ast \text{Bias} \\ +{\gamma }_{12500}\ast \text{Phon}\ast \text{Ambiguity}+{\gamma }_{12600}\ast \text{Speed}\ast \text{Ambiguity}\ast \text{Bias}+ \\ {\gamma }_{12700}\ast \text{Lang}\ast \text{Ambiguity}\ast \text{Bias}+u_{0\mathrm{i}0}+u_{1\mathrm{i}0}\ast \text{Ambiguity}+u_{2\mathrm{i}0}\ast \text{Bias}+u_{12i0}\ast \text{Ambiguity}\ast \text{Bias} \\ +v_{00\mathrm{j}}+v_{10\mathrm{j}}\ast \text{Ambiguity} \end{gathered}$$

where Y_ij are the odds of subject i correctly responding to item j, γ₀₀₀ is an intercept term representing grand mean accuracy, γ₁₀₀, γ₂₀₀, and γ₁₂₀₀ are the fixed effects of the experimental conditions, γ₃₀₀ through γ₇₀₀ are the fixed effects of the individual-difference composites on overall reading time, γ₁₃₀₀ through γ₁₇₀₀ are the fixed effects of the individual-difference composites on the ambiguity effect, γ₂₃₀₀ through γ₂₇₀₀ are the fixed effects of the individual-difference composites on the verb bias effect, γ₁₂₃₀₀ through γ₁₂₇₀₀ are the fixed effects of the individual-difference composites on the ambiguity x verb bias interaction, u_0i0 is the error for the subject intercept for subject i, u_1i0, u_2i0, and u_12i0 are the error terms for the random subject slopes for the experimental conditions for subject i, v_00j is the error for the item intercept for item j, and v_10j is the error for the random slope of ambiguity for item j, and ε_ij is the error for subject i on item j.

Residual reading time for relative-clause extraction items was modeled as:

$$\begin{gathered} Y_{ij}={\gamma }_{000}+{\gamma }_{100}\ast \text{RCType}+{\gamma }_{200}\ast \text{VWM}+{\gamma }_{300}\ast \text{Inhib}+{\gamma }_{400}\ast \text{Phon}+{\gamma }_{500}\ast \text{Speed}+{\gamma }_{600}\ast \text{Lang}+ \\ {\gamma }_{1200}\ast \text{VWM}\ast \text{RCType}+{\gamma }_{1300}\ast \text{Inhib}\ast \text{RCType}+{\gamma }_{1400}\ast \text{Phon}\ast \text{RCType}+ \\ {\gamma }_{1500}\ast \text{Speed}\ast \text{RCType}+{\gamma }_{1600}\ast \text{Lang}\ast \text{RCType}+u_{0\mathrm{i}0}+u_{1\mathrm{i}0}\ast \text{RCType}+v_{00\mathrm{j}}+v_{10\mathrm{j}}\ast \text{RCType}+{\epsilon }_{ij} \end{gathered}$$

where Y_ij is the residual reading time for subject i on item j, γ₀₀₀ is an intercept term representing grand mean residual reading time, γ₁₀₀ is the fixed effect of experimental condition, γ₂₀₀ through γ₆₀₀ are the fixed effects of the individual-difference composites on overall reading time, γ₁₂₀₀ through γ₁₆₀₀ are the fixed effects of the individual-difference composites on the RC type effect, u_0i0 is the error for the subject intercept for subject i, u_1i0 is the error term for the random slope of condition for subject i, v_00j is the error for the item intercept for item j, v_10j is the error for the random slope of condition for item j, and ε_ij is the error for subject i on item j.

Comprehension accuracy for relative-clause extraction items was modeled as:

$$\begin{gathered} \log (Y_{ij})={\gamma }_{000}+{\gamma }_{100}\ast \text{RCType}+{\gamma }_{200}\ast \text{QuestionType}+{\gamma }_{1200}\ast \text{RCType}\ast \text{QuestionType}+u_{0\mathrm{i}0}+ \\ {u}_{1\mathrm{i}0}\ast \text{RCType}+u_{2i0}\ast \text{QuestionType}+u_{12i0}\ast \text{RCType}\ast \text{QuestionType}+{\gamma }_{300}\ast \text{VWM}+ \\ {\gamma }_{400}\ast \text{Inhib}+{\gamma }_{500}\ast \text{Phon}+{\gamma }_{600}\ast \text{Speed}+{\gamma }_{700}\ast \text{Lang}+{\gamma }_{1300}\ast \text{VWM}\ast \text{RCType}+ \\ {\gamma }_{1400}\ast \text{Inhib}\ast \text{RCType}+{\gamma }_{1500}\ast \text{Phon}\ast \text{RCType}+{\gamma }_{1600}\ast \text{Speed}\ast \text{RCType}+ \\ {\gamma }_{1700}\ast \text{Lang}\ast \text{RCType}+{\gamma }_{2300}\ast \text{VWM}\ast \text{QuestionType}+{\gamma }_{2400}\ast \text{Inhib}\ast \text{QuestionType}+{\gamma }_{1500} \\ \ast \text{Phon}\ast \text{RCType}\ast \text{QuestionType}+{\gamma }_{12600}\ast \text{Speed}\ast \text{RCType}\ast \text{QuestionType}+{\gamma }_{12700}\ast \text{Lang}\ast \\ \text{RCType}\ast \text{QuestionType}+v_{00\mathrm{j}}+v_{10\mathrm{j}}\ast \text{RCType}+v_{20j}\ast \text{QuestionType}+ \\ {v}_{120\mathrm{j}}\ast \text{RCType}\ast \text{QuestionType} \end{gathered}$$

where Y_ij are the odds of subject i correctly responding to item j, γ₀₀₀ is an intercept term representing grand mean accuracy, γ₁₀₀, γ₂₀₀, and γ₁₂₀₀ are the fixed effects of the experimental conditions, γ₃₀₀ through γ₇₀₀ are the fixed effects of the individual-difference composites on overall accuracy, γ₁₃₀₀ through γ₁₇₀₀ are the fixed effects of the individual-difference composites on the extraction-type effect, γ₂₃₀₀ through γ₂₇₀₀ are the fixed effects of the individual-difference composites on the question-type effect, γ₁₂₃₀₀ through γ₁₂₇₀₀ are the fixed effects of the individual-difference composites on the extraction type x question type interaction, u_0i0 is the error for the subject intercept for subject i, u_1i0, u_2i0, and u_12i0 are the error terms for the random subject slopes for the experimental conditions for subject i, v_00j is the error for the item intercept for item j, and v_10j, v_20j, and v_120j are the error terms for the random slopes for the experimental conditions for item j, and ε_ij is the error for subject i on item j.

Responses to the attachment ambiguity items were modeled as:

$$\begin{gathered} \log (Y_{ij})={\gamma }_{000}+{\gamma }_{100}\ast \text{QuestionType}+u_{0\mathrm{i}0}+u_{1\mathrm{i}0}\ast \text{QuestionType}+{\gamma }_{200}\ast \text{VWM}+{\gamma }_{300}\ast \text{Inhib}+{\gamma }_{400} \\ \ast \text{Phon}+{\gamma }_{500}\ast \text{Speed}+{\gamma }_{600}\ast \text{Lang}+{\gamma }_{1200}\ast \text{VWM}\ast \text{QuestionType}+ \\ {\gamma }_{300}\ast \text{Inhib}\ast \text{QuestionType}+{\gamma }_{1400}\ast \text{Phon}\ast \text{QuestionType}+{\gamma }_{500}\ast \text{Speed}\ast \text{QuestionType}+ \\ {\gamma }_{1600}\ast \text{Lang}\ast \text{QuestionType}+v_{00\mathrm{j}}+v_{10\mathrm{j}}\ast \text{QuestionType} \end{gathered}$$

where Y_ij are the odds of subject i answering yes to item j, γ₀₀₀ is an intercept term representing the grand mean of answering yes (response bias), γ₁₀₀ is the fixed effect of the question type condition (low- or high-attachment) on yes responses (sensitivity), γ₂₀₀ through γ₆₀₀ are the fixed effects of the individual-difference composites on response bias, γ₁₂₀₀ through γ₁₆₀₀ are the fixed effects of the individual-difference composites on sensitivity to the question type, u_0i0 is the error for the subject intercept for subject i, u_1i0 is the error term for the random subject slope of condition for subject i, v_00j is the error for the item intercept for item j, and v_10j is the error term for the random slope of condition for item j.