Cognitive Psychology Essay

Journal of Memory and Language 97 (2017) 135–153

Contents lists available at ScienceDirect

Journal of Memory and Language

journal homepage: www.elsevier.com/locate/jml

Comprehension in proficient readers: The nature of individual variation

http://dx.doi.org/10.1016/j.jml.2017.07.008 0749-596X/� 2017 Elsevier Inc. All rights reserved.

⇑ Corresponding author at: Department of Psychology, University of California, Davis, One Shields Avenue, Davis, CA 95616, United States.

E-mail address: [email protected] (D.L. Long).

Erin M. Freed, Stephen T. Hamilton, Debra L. Long ⇑ Department of Psychology, University of California, Davis, United States

a r t i c l e i n f o a b s t r a c t

Article history: Received 25 July 2016 Revision received 11 July 2017 Available online 25 August 2017

Keywords: Reading comprehension Individual differences Working memory Lexical processing SEM

Individual-difference research on reading comprehension is challenging because reader characteristics are as correlated with each other as they are with comprehension. This study was conducted to deter- mine which abilities are central to explaining comprehension and which are secondary to other abilities. A battery of psycholinguistic and cognitive tests was administered to community college and university students. Seven constructs were identified: word decoding, working-memory capacity (WMC), general reasoning, verbal fluency, perceptual speed, inhibition, and language experience. Only general reasoning and language experience had direct effects; these two variables accounted for as much variance in com- prehension as did the complete set. Direct effects of WMC and decoding were found only when general reasoning and language experience were deleted from the models. The authors question the need to include WMC in our theories of variability in adult reading comprehension and highlight the need to understand precisely how vocabulary facilitates comprehension.

� 2017 Elsevier Inc. All rights reserved.

Introduction

Reading is the primary means of knowledge acquisition in many domains; thus, the ability to construct accurate and comprehen- sive representations of texts has significant implications for aca- demic performance, occupational success, and physical well- being. Reading is a complex skill, involving both domain-general and language-specific abilities. Variation in reading skill among individuals is considerable, even among university students. Understanding this variation is important for both practical and theoretical reasons. With respect to practice, it can help in the identification of individuals who struggle to comprehend texts and in the design of effective reading instruction. With respect to theory, individual-difference research is an important means of specifying the cognitive and linguistic processes that underlie reading. For example, contemporary theories of working memory (WM) and its role in language processing, both written and spoken, have their foundation in studies showing that reading span, a task that involves processing sentences while holding a set of words in memory, is linked to language processing as assessed by compre- hension tests, eye-tracking, reading time, and event-related poten- tials (ERPs).

Although individual-difference research is critical to our under- standing of variation in language processing, it can be difficult to

conduct. A significant obstacle is that performance across a variety of tasks tends to correlate (cf. Deary, 2000). Researchers develop tasks that are intended to assess participants’ performance on a specific linguistic or cognitive process, such as working-memory capacity (WMC) or word-identification skill. The problem is that no task is ‘‘process pure;” they all involve multiple component pro- cesses. A particular task may be affected by one process more than another, but it is never an assessment of a single one. To the extent that processes overlap across tasks, performance on them will be correlated. Indeed, research on individual differences in reading comprehension shows that performance on individual-difference tasks correlate with each other as much as they do with reading comprehension itself. These correlations make it difficult to deter- mine whether a particular individual characteristic (e.g., WMC) is uniquely predictive of language processing skill or if it correlates only because of its relation to some other individual-difference variable. As we discuss in the following sections, several individual-difference measures, such as WMC and word decoding, have dominated research on language processing and comprehen- sion, even though there is scant evidence that these measures are uniquely predictive of performance. This raises significant con- cerns about theoretical interpretations of this research.

Our goals in the current study were (1) to understand how reader characteristics are related to each other and to text compre- hension, with a particular focus on determining the extent to which WMC, vocabulary, and word knowledge are uniquely pre- dictive of comprehension, and (2) to examine how these relations change depending on which reader characteristics are included in

136 E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153

the analysis. We used a structural equation modeling framework in a group of proficient adult readers to determine which reader char- acteristics are central in predicting comprehension and which are related by means of shared variance with other characteristics. Few large-scale studies of comprehension in proficient adult read- ers have been conducted even though several theories about the nature of individual variation in reading ability have been devel- oped based on empirical findings in this population (e.g., Ericsson & Kintsch, 1995; Gernsbacher, 1990; Just & Carpenter, 1992; MacDonald & Christiansen, 2002). In the sections below, we review the literature on three classes of variables: domain-general cogni- tive abilities, language-specific abilities, and background knowl- edge/reading experience. We focus our review on large-scale multiple-regression and SEM studies of proficient adult readers whenever possible.

Domain-general cognitive abilities

No single cognitive ability has received as much empirical and theoretical attention as WMC. Many researchers have argued that the WM system is integral to maintaining activated representa- tions and computing semantic and syntactic relations among them. Moreover, they have argued that individuals vary in the amount of information that they can maintain in memory as they perform computations to complete a complex cognitive task. These claims are supported by hundreds of studies showing a positive correla- tion between complex span tasks, such as reading span, and tests of verbal ability such as the verbal SAT and the Nelson-Denny Reading Test. Daneman and Merikle (1996) conducted a meta- analysis of the relation between complex span tasks and verbal ability and reported correlations across studies that ranged from .20 to .52. Studies in which WMC has been assessed as a latent variable have also found a significant relation between WMC and comprehension (Engle, Tuholski, Laughlin, & Conway, 1999; McVay & Kane, 2012).

Several explanations of the correlation between complex span and comprehension have been developed in the context of sentence-processing research to explain why some readers have greater difficulty processing sentences with complex syntactic structures than do other readers. These explanations are all grounded in the assumption that difficulty in processing sentences has consequences for comprehension overall. Explanations for the correlation between span and language processing generally fall into two classes: (1) those in which the relation between span and processing is a direct one in that limitations in the ability to simultaneously maintain and process information affect the types of relations that readers are able to construct during comprehen- sion and (2) those in which the relation between span and process- ing is indirect in that the correlation reflects shared variance between span tasks and other variables, in particular, language experience.

A direct relation between complex span and comprehension is predicted in two models: the Capacity Theory of Comprehension (Just & Carpenter, 1992) and the Separate-Sentence-Interpreta tion-Resource Theory (SSIR) (Waters & Caplan, 1996). According to Capacity theory, WM consists of a finite pool of cognitive resources that supports both storage and processing of informa- tion. The total amount of activation that is available in WM varies across individuals. When the amount of activation that is needed for storage and processing exceeds the total activation that is avail- able, one or both functions are impaired and information is lost. Thus, individuals who are low span have insufficient resources to execute necessary comprehension processes when storage and processing demands are high as, for example, when readers encounter syntactically difficult sentences (Just & Carpenter, 1992; King & Just, 1991). The SSIR Theory differs from Capacity

Theory with respect to predictions about the role of WMC in sen- tence processing, but not with respect to predictions about text comprehension. According to this view, WMC is modular with a dedicated module devoted to syntactic parsing and a second mod- ule that is devoted to post-parsing processes involved in the inte- gration and elaboration of ideas in comprehension. Individuals show little variation in the capacity of the first module, but vary substantially in the capacity of the second. Thus, SSIR Theory is similar to Capacity Theory in attributing individual differences at the discourse level to a limited-capacity system.

In contrast, an indirect relation between WM span and compre- hension is predicted in a connectionist-based framework proposed by MacDonald and Christiansen (2002) and the Long-Term Working-Memory (LTWM) Theory proposed by Ericsson and Kintsch (1995). According to the connectionist-based framework, the capacity of a system arises from its architecture (e.g., the num- ber of processing units, how activation passes through the weights) and the system’s experience (e.g., how often it has processed sim- ilar input in the past). Thus, capacity is not a separate pool of resources; it is a property of the processing system. The relation between complex span and comprehension arises from variation in two factors. First, individuals vary with respect to basic sen- sory/perceptual abilities, primarily the ability to represent and pro- cess phonological information. The ability to discriminate phonemes quickly and represent them accurately in short-term memory is important in grapheme-to-phoneme mapping during reading and for performing well on verbal span tasks. Second, indi- viduals vary in reading experience, giving rise to individual differ- ences in practice with linguistic stimuli. Poor comprehenders read less frequently than do good ones; thus, they are less likely to encounter low-frequency linguistic stimuli (e.g., uncommon syn- tactic structures). Consequently, variation in the processing of low frequency input is due to differences in practice (Long & Prat, 2008; Wells, Christiansen, Race, Acheson, & MacDonald, 2009). The LTWM Theory also emphasizes the role of experience in explaining the relation between complex span and comprehen- sion. According to the theory, skilled readers develop mechanisms for encoding and retrieving information from long-term memory that meet the demands of the task. Thus, individuals who read fre- quently are skilled at encoding linguistic input into structures that can be quickly and easily retrieved when needed.

Although WM span has been used in hundreds of studies to pre- dict language processing in proficient adult readers, only a handful of them have examined whether or not span is uniquely predictive of comprehension when other linguistic and cognitive variables are considered. Does span have a direct relation on comprehension as predicted by the Capacity and SSIR Theories or is the relation indi- rect as predicted by the connectionist-based and LTWM Theories? In one study, Hannon (2012) found that WMC had a significant effect on comprehension using an SEM approach. She assessed high-level skills (e.g., knowledge access, knowledge integration) and low-level skills (e.g., lexical decision, phonological decision). The model, the Cognitive Components and Resource Model of Reading Comprehension (CC-R), was restricted such that low- level skills (e.g., word decoding) directly predicted both reading speed and reading comprehension; reading speed directly pre- dicted reading comprehension; WMC, text-based processing, and knowledge access directly predicted knowledge integration; and knowledge integration directly predicted reading comprehension. A variant of the model, the CC-R2, allowed WMC to have a direct effect on reading comprehension and was found to be the best fit- ting one. Hannon concluded that high-level and low-level skills are dissociable and that high-level skills have a greater impact on com- prehension than do low-level ones. A critical drawback of the study in assessing the role of WMC in comprehension, however, is that latent variables in the model were not allowed to covary although

E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153 137

many of the measures were significantly correlated. Thus, WMC may have had a direct effect only because shared variance was not assessed.

In contrast, two large studies have found no unique effect of complex span on comprehension. Macaruso and Shankweiler (2010) used multiple regression to examine the unique contribu- tions of several variables to comprehension. They found significant effects of listening comprehension and word decoding. Impor- tantly, they found no unique effect of WMC. Britton and colleagues also found no unique effect of WMC in a study using SEM (Britton, Stimson, Stennett, & Gülgöz, 1998). They assessed reading compre- hension, metacognitive ability, inference ability, domain knowl- edge, and WM span. The model was constrained such that metacognitive ability predicted inference ability; inference ability predicted both domain knowledge and WM span; and domain knowledge predicted comprehension. All of the paths were signif- icant and positive. An alternative model allowing for a direct effect of WM span on comprehension did not result in a better fit and span had no significant effect on comprehension. In a recent meta-analysis of comprehension performance in struggling adult readers, Tighe and Schatschneider (2016) found no support for an effect of WMC when linguistic variables, such as word decoding and oral vocabulary, were included in the studies under review.

In summary, we have a huge literature in psycholinguistics that has focused on the role of WMC in language processing, based on the assumption that WMC has a unique and direct effect on com- prehension. However, only one major study has found such an effect (Hannon, 2012). Given the hundreds of studies that include WM span measures in investigations of language processing in proficient adults, the lack of large-scale studies that support WMC as a unique predictor of comprehension is striking. A major question to be addressed in this study is whether WM span has a direct effect on comprehension when it is allowed to covary with other variables.

A second cognitive variable, inhibition/suppression, has received modest attention in research on language processing in proficient adult readers. Readers appear to differ in the extent to which they can suppress or inhibit activated, but context- irrelevant, information. Suppression plays a prominent role in Gernsbacher’s (1990) Structure Building Framework. According to the framework, readers begin the process of constructing a text representation by establishing a foundation based on initial input. They add incoming information to the structure when it is meaningfully related. When information is unrelated, readers shift to initiate a new substructure. Two processes play an impor- tant role in creating text structures: enhancement and suppres- sion. Enhancement increases the activation of memory traces when their content is relevant to the developing text representa- tion; suppression dampens activation when their content is unre- lated. According to the framework, good and poor comprehenders are similar in their ability to enhance relevant information, but differ significantly in their ability to suppress activated, but irrel- evant, information (Gernsbacher, 1993, 1997; Gernsbacher, Robertson, Palladino, & Werner, 2004; Gernsbacher, Varner, & Faust, 1990).

Unfortunately, large-scale studies of comprehension in profi- cient adult readers have not included measures of suppression/ inhibition ability. This is surprising given that several researchers have suggested that the ability to inhibit task-irrelevant informa- tion is an important component of the WM system (Conway & Engle, 1994; Engle, Conway, Tuholski, & Shisler, 1995; Engle, Kane, & Tuholski, 1999; Engle et al., 1999; Kane, Bleckley, Conway, & Engle, 2001; Kane & Engle, 2000; Rosen & Engle, 1997). Thus, the relation between performance on complex span tasks and comprehension may be secondary to the relation between span and suppression ability.

Two cognitive variables that have been important in studies of language processing in children—processing speed and general reasoning—have been largely ignored in studies of proficient adult reading. Processing speed is believed to be important because reading is a sequential and speed-dependent activity; words are received one at a time and must be integrated into a sentence rep- resentation before the verbal trace of preceding words begins to decay. Numerous studies in children have found that performance on speed-of-processing tasks, such as pattern comparison and let- ter comparison, are uniquely predictive of comprehension (Borella, Carretti, & Pelegrina, 2010; Borella & de Ribaupierre, 2014; Joshi & Aaron, 2000; Peter, Matsuishita, & Raskind, 2011; Swanson, 1996; Swanson, Howard, & Saez, 2006; Tiu, Thompson, & Lewis, 2003). Processing speed has also been investigated in older adults (Caplan, DeDe, Waters, Michaud, & Tripodis, 2011; Payne & Stine-Morrow, 2014). For example, Payne and Stine-Morrow (2014) found that individual variation in processing speed was pre- dictive of sentence wrap-up effects. However, investigations of proficient adult readers are absent from the literature.

Similarly, measures of general reasoning, such as IQ, are com- mon in large-scale studies of reading in children and low-literacy adults (Swanson et al., 2006; Tiu et al., 2003), but are seldom included in comprehension studies of proficient adult readers. This is unfortunate because the purpose of including such measures is to identify the cognitive variables that influence a complex task separate from the overall influence of intelligence. If a construct such as WMC fails to influence comprehension when general rea- soning is included as a variable, then the theoretical importance the construct is undermined.

In summary, few large-scale studies have examined the unique influence of different domain-general cognitive variables on com- prehension in proficient adults. A handful of large studies have included measures of WMC, but they have produced contradictory results. In addition, measures of inhibition/suppression, processing speed, and general reasoning have seldom been examined concur- rently with WMC. In the current study, we include latent variables of all these constructs. We allow the variables to covary and exam- ine their direct and indirect effects.

Language-specific abilities

In order to understand a text, a reader must map orthographic representations to phonological ones, use these representations to access word meanings, and integrate these meanings with pre- ceding information. The dominant theory of reading comprehen- sion in children is the Simple View of Reading (Gough & Tumner, 1986). According to this view, reading comprehension is the pro- duct of linguistic comprehension—all of the skills and capacities that are necessary to understand discourse in its oral form—and word decoding. Word decoding is a limiting factor in reading com- prehension because the product of word decoding and linguistic comprehension will be relatively low whenever word decoding is poor. Most studies of the Simple View of Reading have focused on reading comprehension in children and have found consider- able support for the view (Adams, 1990; Snow, Burns, & Griffin, 1998). Decoding skill is also important in predicting comprehen- sion in struggling adult readers (for a review, see Tighe & Schatschneider, 2016), although the effects of decoding seem to be smaller in low-literacy adults than in children.

Word decoding skill plays a somewhat less prominent role in Perfetti’s (2007) Lexical Quality Hypothesis, a framework for understanding the role of lexical ability in comprehension that emphasizes the rich and detailed representations of words that skilled readers know (Taylor & Perfetti, 2016). He argues that com- prehension skill requires high-quality lexical representations, that is, representations that consist of detailed orthographic-to-

138 E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153

phonological mappings and rich semantic meanings. Readers who have poor quality representations are slow at identifying words, have a poor understanding of the words that they read, and may activate incorrect word meanings.

The relative contributions of word knowledge (e.g., vocabulary) and word decoding (e.g., word-identification speed) have been studied extensively in children and both characteristics are strongly predictive of comprehension (Gough & Tumner, 1986; Hoover & Gough, 1990; Joshi, 2005; Perfetti, 1985; Perfetti & Hogaboam, 1978; Verhoeven & Van Leeuwe, 2008). The role of word-decoding skill in proficient adult readers is less clear than it is in children; fewer studies have included measures of both decoding and word knowledge. Those that have done so report mixed results with respect to the significance of word decoding in proficient readers. We briefly describe a few studies that are rep- resentative of those that have included both word-decoding and word knowledge in large samples of proficient adult readers.

Macaruso and Shankweiler (2010) used multiple regression to assess reading comprehension, listening comprehension, word decoding, receptive vocabulary, WMC, and general reasoning in community college students. With all predictors included, the model accounted for 48% of the variance in reading comprehen- sion. Word decoding and listening comprehension alone accounted for 34% of the variance; listening comprehension uniquely con- tributed 13%. Although word decoding was a significant predictor, it only accounted for 6% of unique variance. Similarly, Landi (2010) found a relatively small effect of word decoding ability on reading comprehension. She used hierarchical regression analyses with measures of word knowledge and experience, spelling, and general reasoning. Regardless of order of entry, vocabulary accounted for the largest portion of variance, ranging from 39% to 45%. Word decoding was a significant predictor, but accounted for only a small portion of the variance, ranging from 0.3% to 0.8%.

Braze, Tabor, Shankweiler, and Mencl (2007) found no unique effect of word decoding on comprehension in a group of adoles- cents and young adults when vocabulary was included as a pre- dictor. They assessed listening comprehension, word decoding, vocabulary, and general reasoning. When reading comprehension was regressed on listening comprehension and word decoding with age as a covariate, both measures uniquely and significantly predicted reading comprehension. However, when vocabulary was added to the analysis, listening comprehension and word decoding no longer made individually identifiable contributions. Braze et al. concluded that their results challenged the Simple View of Reading, suggesting that decoding and listening compre- hension was secondary to vocabulary in adolescents and young adults.

Very few studies have used an SEM framework to examine reading comprehension in adults and even fewer have included measures of both word decoding and word knowledge. Cromley, Snyder-Hogan, and Luciw-Dubas (2010) assessed reading compre- hension as a function of variation in domain knowledge, inference ability, reading strategy use, vocabulary, and word decoding. Domain knowledge was the most important predictor of compre- hension, with significant direct and indirect effects. Vocabulary also had direct and indirect effects, whereas strategy use had only indirect effects. In contrast, they found no significant effect of word decoding on comprehension. Braze et al. (2016) examined compre- hension in a group of adolescent readers and community college students. They assessed listening comprehension, word decoding, and vocabulary. They found that the listening comprehension and vocabulary measures loaded on a single factor. In the SEM, this factor covaried with word decoding and both it and word decoding had direct effects on comprehension. This result was in conflict with the earlier study by Braze and his colleagues in which decod- ing had no unique effect on comprehension (Braze et al., 2007).

In summary, studies that have examined the unique effects of word decoding and word knowledge in proficient readers have generally found that word knowledge accounts for significant vari- ation in comprehension (Bell & Perfetti, 1994; Braze et al., 2007, 2016; Cromley et al., 2010; Landi, 2010). A few studies have found unique effects of word decoding on comprehension, although the effects have been relatively small or limited to particular genres of text (Bell & Perfetti, 1994; Braze et al., 2016; Landi, 2010; Macaruso & Shankweiler, 2010).

Domain knowledge/print exposure

Individuals enjoy reading to different extents and engage in a wide variety of reading practices. Research has established a posi- tive correlation between print exposure, the frequency with which individuals read, and comprehension in both children and adults (Cunningham & Stanovich, 1990, 1991; Stanovich & Cunningham, 1992; Stanovich & West, 1989; West & Stanovich, 1991; West, Stanovich, & Mitchell, 1993). Print exposure is likely to influence comprehension in at least two ways. First, individuals who read often are more likely to learn about rare words and low frequency syntactic structures than are individuals who read less, primarily because these stimuli appear more often in print than they do in speech (Carroll, Davies, & Richman, 1971; Hayes & Ahrens, 1988). Second, individuals who read often are likely to acquire more world knowledge than individuals who read less often as reading is an important means of knowledge acquisition in many domains.

Just as readers vary in their reading practices, they vary in their knowledge about the domain in which they are reading. High- knowledge readers access a rich, interconnected network of con- cepts and ideas when they read a text in their domain of expertise (Chi, Feltovich, & Glaser, 1981; Chiesi, Spilich, & Voss, 1979; Means & Voss, 1985). Moreover, they employ more effective reading strategies than do less knowledgeable readers (Afflerbach, 1986; Lundeberg, 1987) and are faster and more efficient at retrieving information from their knowledge domain (Ericsson & Smith, 1991). The comprehension advantage associated with background knowledge has been well documented in both recall (Alba, Alexander, Hasher, & Caniglia, 1981; Bransford & Johnson, 1972; Schneider, Körkel, & Weinert, 1990; Spilich, Vesonder, Chiesi, & Voss, 1979; Sulin & Dooling, 1974; Summers, Horton, & Diehl, 1985) and recognition (Long, Johns, & Jonathan, 2012; Long & Prat, 2002; Long, Prat, Johns, Morris, & Jonathan, 2008). Two of the studies that we described in the previous section included domain knowledge among other reader characteristics to examine its unique effect on comprehension: Cromley et al. (2010) and Britton et al. (1998). Both studies found that domain knowledge was the most significant of their predictors, with both reliable direct and indirect effects.

In summary, few large-scale studies have included print expo- sure and domain knowledge as predictors of reading comprehen- sion in proficient adult readers. This is unfortunate given two studies reporting that domain knowledge was more important than other predictors in accounting for variance in reading com- prehension (Britton et al., 1998; Cromley et al., 2010).

The current study

The two aims of this study were to understand how reader char- acteristics are related to each other and to text comprehension and to examine how these relations change depending on which reader characteristics are included in the analysis. We used structural equation modeling to identify the unique predictors of reading comprehension in a large group of community college and univer- sity students. SEM has important advantages over other techniques for analyzing correlational data. First, variables in SEM can be both

E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153 139

predictors and outcomes simultaneously. Thus, it goes beyond techniques that identify unique sets of predictors, such as multiple regression, by specifying the relations among the predictor vari- ables. Mediation of one variable by another can be identified, pro- viding information about the primary or secondary role of a particular variable in explaining the outcome. Second, predictors can be measured as latent variables in SEM, derived from shared variance among a set of measures that is intended to assess a the- oretical construct. This allows the construct to be identified sepa- rately from the measurement error that is associated with individual tasks. Finally, SEM estimates the error that is associated with measuring a construct of interest separately from unex- plained variance in the outcome variable. Thus, it can be used to analyze dependencies among psychological constructs without measurement error.

The list of language-specific and domain-general abilities that could be investigated is long. Given the constraints of any single study, decisions must be made about which abilities to investigate. Our decisions were guided by three criteria. First, we selected abil- ities that have played the largest role in theories of reading com- prehension in adult readers: word decoding, vocabulary, and WMC. Second, we included a set of variables that have been involved in debates about the extent to which the Simple View of Reading (Gough & Tumner, 1986) should be augmented with additional abilities. These variables include processing speed (Joshi & Aaron, 2000; Tiu et al., 2003), verbal fluency (Adolf, Catts, & Little, 2006; Silverman, Speece, Harring, & Ritchey, 2013), and general reasoning (Tiu et al., 2003; Van Dyke, Johns, & Kukona, 2014). Third, we selected several variables that have been linked specifically to variation in comprehension among proficient adults; these variables include inhibition/suppression (Gernsbacher, 1990), print exposure (Cunningham & Stanovich, 1990, 1991; Stanovich & Cunningham, 1992; Stanovich & West, 1989; West & Stanovich, 1991; West et al., 1993), and domain knowledge (Long & Prat, 2002; Long et al., 2008, 2012). In selecting the abilities that we included in this study, we chose those that we believed had the greatest likelihood of direct, rather than indirect, effects. Many high-level abilities have been linked to adult reading comprehension, such as meta-cognitive ability, comprehension monitoring, and inference ability. Given limitations on the number of abilities that can be investigated in any single study, we thought in best to focus on abilities that have been closely linked to intel- ligence, WMC, general reasoning, perceptual speed, and inhibi- tion/suppression. In addition, we chose not to include oral measures in this study. The difference in performance between oral and written tasks diminishes greatly across development and tends not to be significant in proficient readers (Braze et al., 2016; Gernsbacher, 1990).

Performance on the battery of psycholinguistic and cognitive tests was factor-analyzed to determine its underlying structure and to confirm that the measures in our study assessed the theo- retical constructs of interest. We hypothesized 9 factors: WMC, general reasoning, perceptual speed, inhibition/suppression, word decoding, vocabulary, verbal fluency, print exposure, and back- ground knowledge. We constructed an SEM in which the predic- tors were latent variables corresponding to the constructs that were identified in the factor analysis. Comprehension was also a latent variable and each predictor was allowed a unique regression path to it. In addition, all predictors were allowed to covary. Paths were removed from the model following a backward-building procedure.

Once we had the full model constructed, we examined its stabil- ity as a function of the latent variables that were included in it. The paths in a SEM can vary in significance depending on the inclusion or exclusion of variables, just as the significance of a predictor vari- able can change in a regression analysis depending on what other

predictors are included. We conducted two types of analyses. In one type, we compared the amount of explained variance in a large model (Models 1, 2, 3, and 4) to a model in which only those latent variables with direct paths were included (Models 1-Reduced, 2- Reduced, 3-Reduced and 4-Reduced). Our goal in these analyses was to determine how much explanatory power was lost when variables with indirect effects were eliminated. In the second type of analysis, we examined the pattern of direct and indirect effects on comprehension when we deleted a latent variable that had a significant direct path to comprehension in the full model (Model 1). Our goal in these analyses was to reconcile discrepancies in the previous literature. Together, these analyses allowed us to determine the stability of our direct effects and information about why previous studies have obtained mixed results about the pre- dictive power of variables such as WMC and word decoding.

Method

Participants

Participants were 357 young adults (243 female, 26 declined to respond) ranging in age from 17 to 29 (M = 18.48, SD = 0.87) who were paid $10/h. to attend �16 h of testing across 11 sessions. Par- ticipants were University of California, Davis undergraduates (58% of participants) and local community college students. All partici- pants were fluent English speakers, with normal or corrected-to- normal vision, and none reported neurological or cognitive impairments.

Materials and procedure

Participants received a battery of tests to assess the constructs that we discussed in the introduction: WMC, suppression/inhibi- tion, processing speed, general reasoning, word decoding, vocabu- lary, print exposure, verbal fluency, and background knowledge. One goal in selecting measures was to identify those that were standardized or had good reliability in previous studies. A second goal was to choose measures that were likely to give us a factor structure in which separate latent variables could be identified for the constructs of interest. One source of measures was the Ekstrom Battery of Factor-Referenced Tests (Ekstrom, French, Harman, & Dermen, 1976). The Kit consists of 72 tests; each test loads strongly on one of 23 factors. We selected 12 tests from the Kit; these tests loaded on 4 factors: vocabulary, verbal fluency, general reasoning, and processing speed.

We anticipated a significant dropout rate given that tasks were administered across a large number of sessions. Thus, we random- ized the order of tasks so that data would be missing at random if participants failed to complete one or more sessions. This was important because missing data can be estimated in a SEM only if the data are missing at random. A random order of tasks was generated for each participant and then adjusted such that the most difficult and time-consuming tasks (e.g., Raven’s matrices, span tasks) were administered in different sessions. The tasks were administered to participants individually and each session was �1.5–2 h long. Participants received breaks between tasks.

Comprehension measures

Participants read 10 full-length texts in a dual-Purkinje eye- tracker. The texts were selected to represent a range of genres: lit- erature (‘‘The Oval Portrait” by Edgar Allen Poe; ‘‘Harrison Berg- eron” by Kurt Vonnegut, Jr.; 1288 and 2228 words, respectively), contemporary fiction (‘‘The Secret Life of Walter Mitty” by James Thurber; ‘‘I am Bigfoot” by Ron Carlson; 2110 and 725 words,

140 E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153

respectively), procedural (‘‘What you can do to Minimize Online Risks,” from Consumer Reports; 912 words), biography (‘‘The Bass, the River, and Sheila Mant” by W.D. Wetherell; 2720 words), and expositions about science and history (‘‘Four Score and Seven Lat- tes Ago: How a coffee shortage killed the Confederacy” by David A. Norris; ‘‘Professors: Email diminishes formality, respect students display towards teachers” by Dyanna Quizon; ‘‘One-hit Wonder” from the Science section of the New York Times; ‘‘Cell phone or pheromone? New props for the mating game” by Natalie Angier; 693, 574, 777, and 1227 words, respectively). The focus in this study was on the comprehension data; thus, the eye-tracking data were not included in the analyses.

A comprehension test (Investigator-Generated Test) was devel- oped for each text to assess participants’ memory for explicit information and their understanding of main ideas and themes. The test consisted of 10 multiple-choice questions (4-response options) for each text. Main ideas were identified by having a group of 25 pilot participants identify the 10 most important ideas in each text. We then chose five of the text ideas with the highest rate of identification. All of these ideas were selected by more than half of the participants (M = 19). The investigators wrote an addi- tional five questions for each text that were intended to assess inferences that were important in comprehending the texts. Com- pletion of the test was untimed. Participants also received the comprehension section of the Nelson-Denny Reading Test, Forms F (Brown, Bennett, & Hanna, 1980) or G (Brown, Fischo, & Hanna, 1993). Participants read short passages and answered a total of 36 or 38 questions (Forms F and G, respectively). Partici- pants were given 20 min to complete the test.

Individual-difference measures

Working-memory capacity Four WM span tasks were included: (1) Reading Span (Just &

Carpenter, 1980; Unsworth, Heitz, Schrock, & Engle, 2005) – Partic- ipants read a series of sentences and made a sense/nonsense judg- ment to each one (e.g., On warm sunny afternoons, I like to walk in the park; Most people agree that Monday is the worst stick of the week). After each judgment, they received a single word that was unre- lated to the previous sentence and were asked to remember it. The sentences were presented in 15 sets, varying in size from two to seven sentences. Set size was presented randomly. At the end of each set, participants recalled the target words in order of pre- sentation. They received 60 target words across all trials; (2) Alpha- bet Span (Craik, 1986) – Participants received 25 lists of words, one word at a time (e.g., jam, dog, eel, book). Each list was preceded by a fixation cross for 1000 ms and each item was presented for 1000 ms. The number of words in a set varied from two to six. At the end of each set, participants were asked to recall the words in alphabetical order; (3) Minus Span (Salthouse, 1988) – Participants received 35 sets of random numbers, one number at a time. The size of the set varied from two to eight. Participants were asked to sub- tract 2 from each number in the set and then to recall the differ- ences in numerical order. Participants received 175 numbers across all trials; (4) Visual Number Span (from the Ekstrom Battery) – Participants received digits in 24 sets of varying lengths from 4 to 13 at a rate of one digit per second. At the end of a set, participants recalled the digits in the reverse order of presentation.

Suppression/inhibition ability A set of tasks was selected to assess participants’ ability to

respond to a target while inhibiting a pre-potent response1: (1)

1 The Eriksen-Flaker Task was included as one of our inhibition/suppression task; however, a programming error led to significant loss of data and the task was dropped from the study

Go/No–Go (https://www.sacklerinstitute.org/cornell/assays_ and_tools/) – Participants received letters one at a time on the com- puter screen, each for 1500 ms. They were asked to press the ‘z’ key (go) in response to any letter except the letter X. When the letter X appeared, participants were told not to respond (no-go). Participants received a set of practice trials and then 300 test trials, 240 of which were ‘go’ trials. All responses were recorded; (2) Stroop Interference – Participants received 105 letter strings one at a time on a computer screen. The strings consisted of a series of X’s or the name of a color (e.g., red). All of the letter strings were presented in colored font and participants were asked to name the color of the font. Included were compatible and incompatible word trials. Compatible trials consisted of words in a font that matched the meanings of the words (e.g., the word ‘‘red” in red font). Incompatible trials consisted of words in a font that did not match the meanings of the words (e.g., the word ‘‘red” in green font). Accuracy and onset latencies were recorded for each letter string.

General reasoning A set of general reasoning tasks was selected to assess readers’

ability to solve novel problems. We included a standardized test, Raven’s Advanced Progressive Matrices (Raven, 1962), and tests from the Ekstrom Battery that loaded on a general reasoning factor. Participants received the first two sets of Raven’s Matrices. There were a total of 48 matrix problems in which participants had to choose the missing element that completed each pattern. The tests from the Ekstrom Battery included (1) Arithmetic Aptitude – the test consisted of arithmetic word problems, two sections of 15 items each. Participants had 10 min to complete each section; (2) Mathematic Aptitude – the test consisted of multiple-choice word problems that required the use of arithmetic and algebraic opera- tions, two sections of 15 items each. Participants had 10 min to complete each section; (3) Necessary Arithmetic Operations – the test consisted of arithmetic word problems. Participants had to determine which numerical operations were required to solve each problem, but were not required to perform the computations. There were two sections; each had 15 items. Participants had five minutes to complete each section.

Perceptual speed tasks A set of tasks was selected to assess the speed with which par-

ticipants could perform perceptual processes. These included: (1) Letter Comparison (Salthouse & Babcock, 1991) – The task con- sisted of pairs of letter strings (e.g., MVX—MXV; QFLJEO—QFLJEO). The strings were three, six, or nine letters in length. String length was presented in random order. Participants were asked to indi- cate whether the letter strings in each pair were the same or differ- ent. There were two lists of 21 pairs and each list had a 30 s time limit; (2) Pattern Comparison (Salthouse & Babcock, 1991) – The task consisted of patterns that were presented in pairs. Each pat- tern contained three, six, or nine line segments. Participants were asked to indicate whether the patterns in each pair were the same or different. There were two lists of 15 pairs and each list had a 30 s time limit; (3) Finding As (Ekstrom Battery) – Participants received a list of words that were arranged in columns. Each column con- tained 41 words. Participants were asked to scan the list and to mark five words in each column that contained the letter ‘‘a”. There were 25 columns to be scanned in a total of two minutes; (4) Num- ber Comparison (Ekstrom Battery) – Participants inspected pairs of multi-digit numbers and indicated whether or not the numbers in each pair were the same. There were 48 pairs of numbers in a list and participants had 90 s to complete the list. Participants com- pleted two lists; (4) Identical Pictures (Ekstrom Battery) – Partici- pants were shown six numbered geometrical figures in a row. They were asked to use the first item as their standard and to determine

E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153 141

which of the remaining five were identical to the first. There were 48 rows in a list and two lists. Each list had a time limit of 90 s.

Word decoding tasks and phonological awareness Our primary interest was in the decoding tasks, but we included

two phonological awareness tasks given how important phonolog- ical awareness has been in the study of reading comprehension in children. The decoding tasks included (1) Phonological Decision, sometimes called the pseudohomophone choice task (Bell & Perfetti, 1994; Landi, 2010; Olson, Kliegl, Davidson, & Foltz, 1985) – Participants received two letter strings side-by-side on a computer screen, both of which were non-words. One of the letter strings sounded like a real word if pronounced aloud (e.g., HOWLKE vs. HOWSE). Participants indicated, as quickly as possible, which of the two letter strings sounded like a real word. Accuracy and reac- tion times were recorded for each of 75 trials. Thus, only the latency data were analyzed; (2) Non-word Naming (Bell & Perfetti, 1994) – Participants received pronounceable letter strings that were presented in the center of a computer screen (e.g., phlambust). They pronounced the string aloud as quickly and as clearly as possible. A voice key was used to record responses and voice onset times for each of 100 trials; (3) Orthographic Decision task (Bell & Perfetti, 1994; Olson et al., 1985). Participants received two letter strings side-by-side on a computer screen, one of which was a correctly spelled word. The other was a non-word, although it sounded like a word if pronounced aloud (e.g., DEAL vs. DEEL). Participants indicated, as quickly as possible, which of the two strings was a correctly spelled word. Accuracy and reaction times were recorded for each of 75 trials.

One phonological awareness task was administered2: Phoneme Transposition (Olson, Wise, Conners, Rack, & Fulker, 1989) – Partic- ipants heard words one at a time. They were asked to remove the first phoneme from the word, move it to the end, and add a long ‘‘a” sound (e.g., ‘‘lock” would become ‘‘ocklay”). Accuracy was recorded for each of 45 trials.

Vocabulary We administered the vocabulary section of the Nelson-Denny

Reading Test, Form F (Brown et al., 1980) and Form G (Brown et al., 1993). Form F consists of 100 items; each item is a sentence with the final word missing. Participants are asked to complete the sentence with an appropriate word from among five response options. Form G consists of 80 items, structured in the same fash- ion. Participants had 15 min to complete each test. All scores were converted to scale scores to ensure the forms were equivalent. We also administered the Extended Range Vocabulary and Advanced Vocabulary sections of the Ekstrom Battery. Both are multiple- choice tests of participants’ knowledge about synonyms. The Extended Range test consists of two sections of 24 items each and participants had four minutes to complete each section. The Advanced Vocabulary test consists of two sections of 18 items each and participants had four minutes to complete each section.

Verbal fluency The verbal fluency tasks were selected from the Ekstrom Bat-

tery to assess the speed with which participants were able to gen- erate words in specific categories. These included (1) Word Beginnings – Participants were asked to write as many words as possible beginning with a target letter (e.g., ‘‘Write as many words as you can that begin with C”). They were given two different prompts and allowed to work for three minutes on each prompt; (2) Word Endings – Participants were asked to write as many

2 A second phonological awareness task, phoneme-deletion, was also administered. A error in task administration made the data unusable.

words as possible ending with a specific letter (e.g., ‘‘Write as many words as you can that end in T”). They were given two different prompts and allowed to work for three minutes on each prompt; (3) Word Beginnings and Endings – Participants were asked to write as many words as possible beginning with one target letter and ending with another target letter (e.g., ‘‘Write as many words as you can that begin with S and end with N”). They were given two different prompts and allowed to work for three minutes on each prompt.

Background knowledge and print exposure Most measures of background knowledge are constructed to

assess the depth of an individual’s expertise in a single domain. The texts that were used in this study, however, spanned multiple genres and knowledge domains. Thus, we decided to include mea- sures that would assess knowledge generally. Unfortunately, our choices were limited by the absence of general knowledge mea- sures in previous reading research. Thus, we examined the litera- ture on cultural literacy and selected a test from a popular book

on the topic: Test-Prep your IQ with the Essentials of Cultural Lit-

eracy (Zahler & Zahler, 2003). Questions on the test assessed knowledge about American History, Geography, Myth and Reli- gion, Science, and Art.

Several of the texts that we used in the study were about topics in science; thus, we also constructed a ‘‘Scientist Knowledge Test” that was modeled after the Author Recognition Test (Stanovich & West, 1989). Participants were asked to distinguish the names of scientists from foils. Fifty scientists were drawn from a list of those who had received a Nobel Prize in science. Fifty foils were created by drawing names from the National Academy of Sciences, at ran- dom. The foils were scientists, but we did not expect them to have the same name-recognition as the Nobel Prize winners. Similar measures have been constructed by Long and colleagues to assess knowledge in science-fiction domains (Long & Prat, 2002; Long, Wilson, Hurley, & Prat, 2006). Participants were told to respond only if they were sure that the name belonged to a scientist in order to discourage guessing. Two measures of reading frequency were included: (1) the Author Recognition Test (Stanovich & West, 1989) – Participants were asked to distinguish real author names from foils. Fifty authors (80% fiction, 20% non-fiction) were drawn from a list of those who had appeared on the New York Times Best Seller List for at least one month during the preceding 12 months. Fifty foils were created by drawing names from the edi- torial boards of experimental psychology journals. Participants were told to respond to the name only if they were sure that the name was an author in order to discourage guessing; (2) The Read- ing Habits Questionnaire (Scales & Rhee, 2001) – Participants com- pleted a questionnaire consisting of 37 items that assessed how often participants read, what genres that they prefer to read, their self-perceived level of reading skill, their behavior while reading (e.g., ‘‘How often do you look up the definition of a word that you don’t know?”). We included participants’ responses to the item ‘‘How often do you read?” (5-point scale from ‘‘Never” to ‘‘Very Often”) as a measure of reading frequency.

Results

Two-hundred and seventy-four participants (77%) completed all sessions. Almost all of the remaining participants completed at least half of the sessions. The analyses below confirmed that the data were missing at random (i.e., there was no systematic pat- tern to the data that were missing). For those tasks in which we collected latency data, outliers were identified by computing a mean and SD for each participant. Outliers were defined as laten- cies that were three SDs above or below the participant’s mean

142 E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153

and were deleted from the data set. Accuracy on the reaction time tasks was very high, ranging from 92% to 97% across tasks. Accu- racy on these tasks did not correlate with performance on any other task; thus, only the reaction-time data were analyzed fur- ther. Descriptive statistics and correlations among the variables are reported in Tables 1 and 2, respectively. Accuracy data is reported in proportions to facilitate comparison across tasks. It is important to note that many of the measures were significantly correlated with each other and with one or both of the comprehen- sion tests.

Factor analysis

All measures except the comprehension ones were factor ana- lyzed using a principal components analysis (PCA) method of extraction, with orthogonal rotation (Varimax). The purpose of the factor analysis was to identify the individual-difference mea- sures that corresponded to the theoretical constructs that were described in the introduction and to reduce the number of con- structs to a set that would be manageable in the SEMs. In our initial analysis, we identified nine factors. The measures loaded onto our hypothesized factors with three exceptions: the Go/No-Go mea- sure loaded with the word-decoding measures, the phoneme transposition measure failed to load highly on any factor, and the

Table 1 Descriptive statistics for the 29 individual difference variables and the two comprehensio

Variables N Minimum

Decoding & Phono. Awareness Orthographic Decision (ms) 261 523.207 Phonological Decision (ms) 252 1157.615 Non-word Naming (ms) 243 139.750 Phoneme Transpositiona 224 0.040

Working-Memory Capacity Reading Spana 285 0.270 Alphabet Spana 284 0.130 Minus Spana 281 0.100 Visual Number Spana 278 0.000

Suppression/Inhibition Go-No-Goa 256 0.000 Stroop Interference (ms) 245 �257.809 Print Exp. & Background Know. Author Recognition Testa 279 0.000 Reading Questionnaire 330 1.000 Cultural Intelligencea 262 0.120 Scientist Recognition Testa 279 0.000

Vocabulary Extended Range Vocabularya 278 �0.090 Advanced Vocabularya 270 0.000 Nelson-Denny Vocabularya 283 0.630

General Reasoning Raven’s Progressive Matricesa 280 0.290 Arithmetic Aptitude Testa 278 0.070 Mathematic Aptitude Testa 261 �0.150 Necessary Arithmetic Ops.a 274 �0.040 Perceptual Speed Letter Comparisona 274 0.040 Finding Asa 274 0.080 Number Comparisona 275 0.100 Pattern Comparisona 262 0.570 Identical Picturesa 260 0.320

Verbal Fluency Word Beginnings 276 9.000 Word Endings 262 4.000 Word Beginnings & Endings 277 0.000

Comprehension Nelson-Dennya 263 0.222 Investigator-Generateda 303 0.460

a Reported as proportions.

pattern comparison task did not load with other measures of per- ceptual speed. We then dropped these measures from the dataset and reanalyzed the data.

The component matrix for the remaining measures is reported in Table 3. The Kaiser-Meyer-Olkin measure verified the sampling adequacy of the analysis, KMO = 0.825. Bartlett’s test of sphericity, v2 (325) = 1791.110, p < 0.001, indicated that correlations among variables were appropriate for PCA. Seven components were extracted, accounting for 61.73% of variance in the data. The factor loadings generally corresponded to hypothesized relations among the variables in our battery of tests. All span tasks loaded on a fac-

tor that we labeled WMC. Raven’s Matrices and the general reason- ing tasks from the Ekstrom Battery loaded onto a factor that we

labeled reasoning. All perceptual speed tests, except pattern com-

parison (see above), loaded onto a factor that we labeled percep-

tual speed. The phonological awareness measures were dropped for the reasons described above and all of the word decoding mea-

sures loaded on a single factor that we called decoding. The mea- sures of verbal fluency from the Ekstrom Kit loaded on a factor

that we called fluency. Stroop Interference loaded onto its own fac- tor. As we mentioned above, the Go/No-Go Task was dropped because it loaded with the decoding measures. As a consequence, Stroop Interference was entered into the SEM as a single measure

n measures.

Maximum Mean Standard Deviation

157.653 762.638 141.630 8084.414 2659.039 1106.412 6479.044 1053.328 483.228 1.000 0.876 0.131

0.930 0.640 0.135 0.970 0.733 0.121 1.000 0.834 0.136 0.710 0.345 0.073

0.700 0.138 0.136 1052.417 96.981 102.886

0.500 0.175 0.104 5.000 3.580 0.962 0.910 0.573 0.138 0.640 0.247 0.109

0.720 0.319 0.148 0.690 0.345 0.143 0.980 0.866 0.079

1.000 0.723 0.129 0.990 0.498 0.194 0.960 0.366 0.154 0.960 0.516 0.179

0.740 0.191 0.054 0.690 0.312 0.083 0.530 0.284 0.063 1.000 0.962 0.061 1.000 0.707 0.138

57.000 26.360 7.916 53.000 31.500 8.128 58.000 19.230 6.944

1.000 0.788 0.157 0.922 0.773 0.091

Table 2 Bivariate correlations among the individual-difference and comprehension measures.

1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. 29. 30. 31.

1. Orthographic Decision 2. Phonological Decision 0.52 3. Non-word Naming 0.26 0.20 4. Phoneme Transposition �0.19 �0.23 �0.10 5. Reading Span 0.05 �0.07 �0.12 0.17 6. Alphabet Span �0.14 �0.11 �0.09 0.14 0.45 7. Minus Span �0.12 �0.10 0.00 0.23 0.41 0.27 8. Visual Number Span �0.03 0.10 �0.02 0.13 0.21 0.23 0.33 9. Go-No-Go �0.17 �0.09 �0.15 0.07 �0.11 �0.06 0.12 �0.00 10. Stroop Interference 0.05 0.13 0.12 �0.09 0.01 0.02 0.00 �0.03 �0.10 11. Author Recognition Test �0.18 �0.22 �0.08 0.31 0.12 0.25 0.08 0.02 �0.06 0.00 12. Reading Questionnaire 0.09 0.03 �0.00 0.20 0.13 �0.00 0.04 0.04 �0.02 0.07 0.35 13. Cultural Intelligence �0.17 �0.32 �0.21 0.28 0.26 0.28 0.16 �0.03 0.04 �0.10 0.51 0.36 14. Scientist Recognition Test �0.10 �0.18 �0.07 0.10 0.15 0.20 0.10 �0.01 �0.03 �0.08 0.36 0.17 0.48 15. Extended Range Vocabulary �0.14 �0.25 �0.16 0.26 0.29 0.31 0.14 �0.01 �0.05 �0.06 0.58 0.26 0.64 0.43 16. Advanced Vocabulary �0.09 �0.20 �0.13 0.22 0.19 0.25 0.05 �0.01 �0.05 �0.15 0.46 0.28 0.57 0.36 0.67 17. Nelson-Denny Vocabulary �0.22 �0.33 �0.21 0.35 0.27 0.32 0.13 �0.08 �0.06 �0.15 0.58 0.37 0.71 0.37 0.73 0.61 18. Raven’s Progressive Matrices �0.05 0.07 0.01 0.19 0.24 0.32 0.33 0.31 �0.08 0.04 0.14 �0.02 0.12 0.14 0.22 0.14 0.18 19. Arithmetic Aptitude Test �0.16 �0.06 �0.11 0.21 0.28 0.31 0.26 0.22 �0.02 �0.11 0.13 �0.10 0.27 0.19 0.29 0.24 0.29 0.29 20. Mathematic Aptitude Test �0.11 �0.06 �0.09 0.21 0.22 0.26 0.29 0.16 0.03 �0.02 0.16 �0.02 0.39 0.21 0.35 0.26 0.32 0.28 0.71 21. Necessary Arithmetic

Operations �0.05 �0.09 �0.11 0.25 0.31 0.33 0.32 0.11 �0.02 �0.08 0.24 0.03 0.37 0.17 0.37 0.28 0.42 0.29 0.63 0.61

22. Letter Comparison �0.16 �0.13 �0.10 0.13 0.10 0.12 0.09 0.06 0.01 �0.05 0.14 0.12 0.14 0.02 0.17 0.14 0.10 0.00 0.21 0.19 0.21 23. Finding As �0.19 �0.13 �0.06 0.17 0.05 0.08 0.13 0.11 0.00 �0.17 0.16 0.11 0.13 0.00 0.12 0.06 0.17 0.03 0.13 0.15 0.18 0.33 24. Number Comparison �0.26 �0.05 �0.14 0.07 0.01 0.04 0.15 0.16 0.04 �0.11 0.03 �0.08 �0.05 �0.05 �0.12 �0.04 �0.05 0.04 0.25 0.21 0.17 0.23 0.38 25. Pattern Comparison �0.15 �0.07 �0.05 0.06 0.12 0.04 0.08 0.03 �0.04 �0.04 0.15 �0.06 0.12 0.14 0.18 0.12 0.15 0.06 0.03 0.09 0.11 0.00 0.11 0.04 26. Identical Pictures �0.16 �0.15 �0.06 0.06 0.08 0.13 0.10 0.17 0.08 0.02 0.15 0.04 0.22 0.05 0.17 0.14 0.26 0.12 0.17 0.20 0.22 0.18 0.26 0.20 0.08 27. Word Beginnings �0.24 �0.31 �0.11 0.22 0.18 0.31 0.20 0.11 �0.05 0.16 0.24 0.06 0.33 0.18 0.36 0.32 0.42 0.08 0.25 0.26 0.24 0.17 0.27 0.12 0.17 0.19 28. Word Endings �0.24 �0.30 0.00 0.09 0.15 0.33 0.15 0.09 �0.08 �0.02 0.23 �0.04 0.23 0.13 0.28 0.24 0.29 0.16 0.16 0.20 0.27 0.11 0.18 0.14 0.12 0.20 0.48 29. Word Beginnings & Endings �0.25 �0.29 �0.13 0.18 0.19 0.28 0.14 0.09 �0.04 �0.03 0.26 �0.06 0.29 0.21 0.37 0.28 0.39 0.15 0.24 0.22 0.22 0.01 0.16 0.12 0.18 0.16 0.44 0.44 30. Nelson-Denny

Comprehension �0.30 �0.34 �0.19 0.29 0.17 0.16 0.18 �0.03 0.04 �0.22 0.36 0.13 0.49 0.22 0.36 0.38 0.55 0.24 0.27 0.31 0.38 0.18 0.18 0.24 0.08 0.16 0.26 0.20 0.21

31. Investigator-Generated Comprehension

�0.06 �0.22 �0.11 0.26 0.35 0.34 0.27 �0.04 �0.05 �0.04 0.37 0.21 0.50 0.31 0.59 0.47 0.57 0.27 0.29 0.35 0.36 0.10 0.12 �0.03 0.08 0.13 0.29 0.19 0.25 0.48

Note: Significant correlations (p < 0.05) are shown in bold.

E .M

. Freed

et a l./Jo

u rn a l o f M em

o ry

a n d La

n gu

a ge

9 7 (2 0 1 7 ) 1 3 5 – 1 5 3

1 4 3

Table 3 Principal components analysis of the remaining 26 individual difference variables.

Factor Variables 1 2 3 4 5 6 7

Decoding Orthographic Decision �0.022 �0.023 �0.319 0.028 �0.244 0.691 �0.032 Phonological Decision �0.227 0.102 �0.397 0.043 �0.080 0.626 0.056 Non-word Naming �0.133 �0.069 0.183 �0.107 �0.029 0.713 0.067

WMC Reading Span 0.240 0.143 �0.001 0.732 �0.093 �0.085 0.018 Alphabet Span 0.230 0.210 0.307 0.560 �0.059 �0.085 0.047 Minus Span 0.021 0.192 0.078 0.692 0.106 �0.070 0.024 Visual Number Span �0.137 0.084 0.054 0.646 0.215 0.138 �0.061

Inhibition Stroop Interference �0.109 �0.012 0.040 0.006 �0.086 0.075 0.938 Language Author Recognition Test 0.703 �0.005 0.146 0.038 0.182 �0.013 0.080 Experience Reading Questionnaire 0.591 �0.272 �0.319 0.148 0.258 0.182 0.197

Cultural Intelligence 0.794 0.187 0.078 0.073 0.057 �0.189 �0.001 Scientist Recognition Test 0.549 0.144 0.094 0.048 �0.144 �0.077 �0.096 Extended Range Vocabulary 0.803 0.210 0.192 0.076 0.003 �0.075 0.016 Advanced Vocabulary 0.747 0.142 0.151 0.007 0.013 �0.017 �0.121 Nelson-Denny Vocabulary 0.819 0.169 0.209 0.053 0.079 �0.158 �0.037

Reasoning Raven’s Progressive Matrices 0.051 0.638 0.134 0.316 0.131 0.174 0.082 Arithmetic Aptitude Test 0.124 0.835 0.072 0.163 0.106 �0.109 �0.078 Mathematic Aptitude Test 0.210 0.831 0.061 0.095 0.136 �0.049 0.035 Necessary Arithmetic Operations 0.282 0.718 0.083 0.180 0.143 �0.015 �0.038

Perceptual Letter Comparison 0.130 0.140 �0.130 0.041 0.607 �0.180 0.103 Speed Finding As 0.084 �0.012 0.157 0.068 0.759 0.004 �0.196

Number Comparison �0.231 0.217 0.095 0.064 0.617 �0.177 �0.166 Identical Pictures 0.131 0.172 0.197 0.035 0.489 0.005 0.132

Fluency Word Beginnings 0.283 0.097 0.596 0.147 0.213 �0.137 0.278 Word Endings 0.160 0.081 0.763 0.118 0.144 0.003 �0.004 Word Beginnings & Endings 0.242 0.146 0.686 0.114 �0.003 �0.118 �0.075

Note: The highest factor loading for each measure is shown in bold font.

144 E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153

instead of a latent variable. We labeled this variable inhibition. We had anticipated separate factors for measures of vocabulary, back- ground knowledge, and print exposure; however, all of these mea- sures loaded on a single factor. The factor appears to assess knowledge about words and the world that correlates with general

language experience. We labeled this factor language experience.

Structural equation model

Full model (Model 1) We constructed a model with the seven predictor variables that

we identified in the factor analysis. Comprehension was a latent variable consisting of the Nelson-Denny and Investigator- Generated Tests. The factor loadings are shown in Fig. 1. Paths were removed from the model following a backward-building pro- cedure. Each predictor was allowed a unique regression path to comprehension. All predictors were allowed to covary. The final model, Model 1, appears in Fig. 2. Model testing statistics for the procedure appear in the supplementary materials, Tables 4–7. The tables include change in model misfit, measurement-model factor loadings, factor variances and unique variances, and stan- dardized covariance path weights. The model was an adequate fit to the data (N = 346 with 134 patterns of missing data; v2(333) = 542.882, p < 0.001; CFI = 0.913; TLI = 0.901; RMSEA = 0.043; 90% confidence interval [0.036–0.049]). The data were missing at ran- dom in all models; thus, missing data were estimated in the analyses.

The latent variables accounted for 76.68% of variance in com- prehension. Only two of the variables had direct effects: language experience (b = 8.44, SE = 0.002, p < 0.001) and reasoning (b = 3.43, SE = 0.001, p = 0.001). Higher reasoning and greater language expe- rience were associated with better comprehension. All other vari- ables had indirect effects on comprehension via their covariance with reasoning and/or language experience. As we predicted, the covariances were substantial and in the expected directions. Faster

decoding was associated with greater language experience, faster perceptual speed, and greater fluency. Decoding covaried posi- tively with inhibition, indicating that slower decoding was associ- ated with more Stroop interference when participants had to suppress a word meaning to name an incongruent font color. Higher WMC was associated with greater language experience, higher reasoning, faster perceptual speed, and greater fluency. Lan- guage experience was positively associated with reasoning, per- ceptual speed, and fluency. Language experience covaried negatively with inhibition, such that greater language experience was associated with less Stroop interference. Higher reasoning was associated with faster perceptual speed and greater fluency. Perceptual speed also covaried positively with fluency such that faster speed was associated with greater fluency. Perceptual speed covaried negatively with inhibition, such that faster speed was associated with less inhibition.

Models of path stability and alternative sets of latent variables

An important goal in this study was to examine how changes in the dataset affect direct and indirect relations between predictors and comprehension. This goal is critical for two reasons: (1) to ensure that the direct paths in the model are robust to changes in the particular set of predictor variables that were included in the study and (2) to understand how explained variance shifts from one latent variable to another when measures are deleted from the models. We addressed this goal in two types of analyses. In one type (model-reduced analyses), we examined how the direct paths changed as we deleted the latent variables that had indirect effects. This enabled us to determine the extent to which the indi- rect effects contributed to the amount of variance explained by the variables with direct effects and to determine whether the direct effects remained stable in the absence of other variables. In the second type of analysis, we examined the pattern of direct and indirect effects on comprehension when a latent variable with a

Fig. 1. Measurement model with comprehension as a latent variable.

Fig. 2. SEM with all latent variables (Model 1).

E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153 145

direct path to it (language experience, reasoning, or both) was removed from the model. This analysis allowed us to examine the importance of those variables with indirect effects and pro- vided potential explanations for why previous studies have reported inconsistent results with respect to relations among decoding, vocabulary, WMC, and comprehension.

Stability of the direct paths in Model 1 (Model 1-Reduced) Model 1-Reduced was constructed to test random path coeffi-

cients against fixed path coefficients by including only those fac- tors from Model 1 that had a direct effect on comprehension: language experience and reasoning. The covariance paths among

individual-difference variables were included as per their relations in Model 1. Their path weights were allowed to vary at random in each model. Latent variables that indirectly affected comprehen- sion were removed systematically from the data set according to the number and weights of their covariance paths. As each latent variable was removed, the regression paths were backward-built until only the significant paths remained. In all iterations, the only direct paths were from reasoning and language experience to com- prehension. As the regression weights were allowed to vary at ran- dom in the models, we ran each model again and fixed the regression weights for reasoning and language experience to those that we had obtained in Model 1. The change in model misfit

Fig. 3. Models 1–4 reduced to only significant and stable components. Panel A: Model 1-Reduced, Panel B: Model 2-Reduced, Panel C: Model 3-Reduced, Panel D: Model 4- Reduced.

146 E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153

between the random- and fixed-coefficient models was tested to determine the stability of the paths. The results revealed that the paths remained stable with no increase in model misfit across the model-reduction process; thus, we report only the fixed regres- sion paths. Model testing statistics appear in the supplementary materials, Tables 5–8. The tables include measurement-model fac- tor loadings, factor variances and unique variances, standardized covariance path weights, and change in model misfit. The final model, with only reasoning and language experience, is depicted in Fig. 3, Panel A. The paths accounted for 76.01% of the variance in comprehension. The model was an good fit to the data (N = 345, with 76 patterns of missing data; v2(64) = 114.482, p < 0.001; CFI = 0.967; TLI = 0.960; RMSEA = 0.048; 90% confidence interval [0.033–0.062]).

SEM model without language experience: Model 2 Model 2 was constructed to determine how variance in compre-

hension was apportioned by removing language experience from the model. This allowed us to determine the significance of decod- ing in the absence of covariation with knowledge about words. The SEM was backward-built following the same procedure as in Model 1. The final model appears in Fig. 4. Model testing statistics for the procedure appear in the supplementary materials, Tables 9–12. The tables include change in model misfit, measurement-model

factor loadings, factor variances and unique variances, and stan- dardized covariance path weights. The covariance paths among the remaining latent variables generally replicated those found in Model 1. In the absence of language experience, decoding had a direct effect on comprehension (b = �4.81, SE = 0.06, p < 0.001). The path between reasoning and comprehension remained signif- icant (b = 5.28, SE = 0.002, p < 0.001). The latent variables accounted for 52.82% of the variance. The model fit the data ade- quately (N = 330, with 99 patterns of missing data; v2(178) = 295.049, p < 0.001; CFI - 0.913; TLI = 0.897; RMSEA = 0.045; 90% confidence interval [0.035–0.054]).

Stability of the direct paths in Model 2 (Model 2-Reduced) Model 2-Reduced was constructed to test random path coeffi-

cients against fixed path coefficients by reducing the factors from Model 2 to those that had direct effects on comprehension, as we did in Model 1-Reduced. The change in model misfit between the random- and fixed-coefficient models was tested to determine the stability of the paths. The paths remained stable with no increase in model misfit across the factor-reduction process. Model testing statistics appear in the supplementary materials, Tables 10–13. The tables include measurement-model factor loadings, factor variances and unique variances, standardized covariance path weights, and changes in model misfit. Decoding and reasoning

Fig. 4. SEM with Language Experience deleted from the model (Model 2).

Fig. 5. SEM with Reasoning deleted from the model (Model 3).

E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153 147

did not covary and both significantly predicted comprehension, accounting for 50.86% of the variance. The final model, depicted in Fig. 3, Panel B, was a good fit to the data (N = 327, with 66 pat- terns of missing data; v2(27) = 48.293, p = 0.007; CFI = 0.968; TLI = 0.958; RMSEA = 0.049; 90% confidence interval [0.025– 0.071]).

SEM model without reasoning: Model 3 In Model 3, Reasoning was removed from the data set. The final

model appears in Fig. 5. Model testing statistics for the procedure are located in the supplementary materials, Tables 14–17. The tables include change in model misfit, measurement-model factor loadings, factor variances and unique variances, and standardized

covariance path weights. The covariance paths among the remain- ing latent variables generally replicated those in Model 1. In the absence of reasoning, WMC (b = 2.85, SE = 0.001, p = 0.004) and language experience (b = 8.81, SE = 0.002, p < 0.001) had direct effects on comprehension, accounting for 73.77% of the variance. The model was an adequate fit to the data (N = 346, with 132 pat- terns of missing data; v2(239) = 390.634, p < 0.001; CFI = 0.917; TLI = 0.904; RMSEA = 0.043; 90% confidence interval [0.035– 0.050]).

Stability of the direct paths in Model 3: Model 3-Reduced Model 3-Reduced was constructed to test random path coeffi-

cients against fixed path coefficients by reducing the factors from

Fig. 6. SEM with Language Experience and Reasoning deleted from the model (Model 4).

148 E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153

Model 3 to those with direct effects on comprehension. The change in model misfit between the random- and fixed-coefficient models was tested to determine the stability of the paths. Model testing statistics appear in the supplementary materials, Tables 15–18. The tables include measurement-model factor loadings, factor variances and unique variances, standardized covariance path weights, and changes in model misfit. The paths remained stable with no increase in model misfit across the factor-reduction pro- cess. Language experience and WMC covaried positively and signif- icantly predicted 74.51% of the variance in comprehension. The final model, shown in Fig. 3, Panel C, was good fit to the data (N = 345, with 80 patterns of missing data; v2(64) = 126.315, p < 0.001; CFI = 0.950; TLI = 0.940; RMSEA = 0.053, 90% confidence interval [0.039–0.067]).

SEM model without language experience and reasoning: Model 4 In Model 4, both language experience and reasoning were

removed from the data set. The final model appears in Fig. 6. Model testing statistics for the procedure appear in the supplementary materials, Tables 19–22. The tables include changes in model mis- fit, measurement-model factor loadings, factor variances and unique variances, and standardized covariance path weights. The covariance paths among the remaining latent variables generally replicated those in Model 1. In the absence of language experience and reasoning, decoding had a significant effect on comprehension (b = �4.39, SE = 0.07, p < 0.001), replicating the pattern that we found in Model 2, and WMC had a direct effect (b = 4.02, SE = 0.001, p < 0.001), as we observed in Model 3. The latent vari- ables together accounted for 44.30% of the variance. The model was a barely adequate fit to the data (N = 329, with 84 patterns of missing data; v2(111) = 199.140, p < 0.001; CFI = 0.888; TLI = 0.863; RMSEA = 0.049; 90% confidence interval [0.038– 0.060]).

Stability of the direct paths in Model 4: Model 4-Reduced Model 4-Reduced was built to test random path coefficients

against fixed path coefficients by reducing the factors in Model 4 to those that had direct effects on comprehension. The change in

model misfit between the random- and fixed-coefficient models was tested to determine the stability of the paths. The paths remained stable with no increase in model misfit across the model-reduction process. Model testing statistics appear in the supplementary materials, Tables 20–23. The tables include measurement-model factor loadings, factor variances and unique variances, standardized covariance path weights, and changes in model misfit. In the final model, shown in Fig. 3, Panel D, decoding and WMC did not covary and accounted for 43.23% of the variance in Comprehension. The model was a poor fit to the data (N = 326, with 49 patterns of missing data; v2(27) = 90.357, p < 0.001; CFI = 0.847; TLI = 0.796; RMSEA = 0.085; 90% confidence interval [0.066–0.104]).

SEM model with only language experience and decoding: Model 5 Our final model was constructed to examine the relations

among vocabulary, decoding, and comprehension. Braze et al. (2007) conducted regression analyses of comprehension in young adults (16–24 years old) and found no effect of decoding on com- prehension when vocabulary was included in the analyses. In con- trast, Braze et al. (2016) used SEM to investigate the relations in a sample of adolescent and community college readers and found a direct effect of both decoding and vocabulary on comprehension. We conducted the same analysis in our sample of community col- lege and university students.

The model was constructed using the procedures described above. The measurement model is shown in Fig. 7, Panel A, and the final model is shown in Fig. 7, Panel B. Model testing statistics for the procedure appear in the supplementary materials, Tables 24–27. The model was a good fit to the data (N = 331 with 59 pat- terns of missing data; v2(18) = 46.497, p < 0.001; CFI = 0.96; TLI = 0.937; RMSEA = 0.069; 90% confidence interval [0.045– 0.094]). The latent variables accounted for 71.74% of variance in comprehension. Only vocabulary had a direct effect (b = 8.04, SE = 0.002, p < 0.001). Faster decoding was predictive of higher vocabulary and higher vocabulary was predictive of better comprehension.

Fig. 7. Measurement Model (Panel A) and SEM (Panel B) with only Decoding and Vocabulary (Model 5).

E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153 149

Discussion

Previous research has identified numerous variables that are predictive of comprehension in proficient readers. Most of this research has been relatively small in scale, examining only a few variables at a time. This is problematic because researchers often operate under the assumption that the variables they study are uniquely predictive of comprehension. This assumption can lead to the development of theories that overstate the contribution of a variable to comprehension, as we believe has been the case with WMC. We addressed this gap in the literature by examining direct and indirect influences of language-specific abilities (e.g., word decoding, vocabulary) and domain-general ones (e.g., WMC, gen- eral reasoning) on reading comprehension. We also examined how these relations change depending on the set of measures included in the analysis. We used an SEM framework that allowed us to assess the influence of latent variables on comprehension in light of their covariance with other variables. Our results have implications for evaluating theories about the roles of lexical abil- ities and WMC in comprehension, for understanding inconsistent outcomes in previous research, and for selecting measures of indi- vidual differences in future studies of variation in adult comprehension.

We selected tasks for the study from among those that have been used in previous studies of comprehension in proficient read- ers. In addition, we included some tasks that have received sub- stantial attention in empirical studies of cognition in children and older adults, but have received much less attention in the lit- erature on adult reading comprehension, such as general reasoning and perceptual speed. As expected, these measures were correlated with each other as strongly as they were with comprehension. Our factor analyses revealed a pattern of loadings that generally corre- sponded to the set of constructs that we discussed in the introduc-

tion. Importantly, we found that decoding measures loaded on a factor separate from our vocabulary measures, allowing us to assess the unique contribution of decoding to comprehension in proficient readers. In addition, our WMC measures loaded on a fac- tor separate from other domain-general abilities, such as general reasoning and perceptual speed, allowing us to examine predic- tions about the direct and indirect effects of WMC on comprehen- sion. Not all of our measures loaded as expected, however. We failed to find separate factors for vocabulary, print exposure, and background knowledge. Although we had hypothesized separate factors, we were not surprised to find that these measures loaded on a single one. Our vocabulary measures were normed for a uni- versity population and print exposure is the primary means by which an advanced vocabulary is acquired. In addition, our mea- sures of background knowledge—Cultural Literacy and Scientist Recognition—assessed general world knowledge more than they assessed knowledge about domain-specific topics that were rele- vant to our texts. General world knowledge, like vocabulary, is enhanced by substantial print exposure.

The focus of our SEM analyses was to identify those latent vari- ables that were directly predictive of comprehension. The seven individual-difference variables in Model 1 accounted for 77% of the variance in comprehension. Of the seven variables, only two had significant direct paths: language experience and reasoning. Comprehension scores increased as performance on language experience and reasoning tasks increased. The direct paths were stable and did not depend on the presence of other latent variables in the model. When we systematically deleted measures from the dataset, we found that the paths from language experience and reasoning remained significant. The fact that language experience and reasoning directly predicted comprehension is not surprising; we included vocabulary in the study because it is so strongly pre- dictive of comprehension in both children and adults and we

150 E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153

included general reasoning because all complex cognitive tasks correlate with reasoning measures, such as IQ, to a significant

extent. The surprise in this study is that only language experience and reasoning had direct effects. Given the substantial literature identifying other variables, such as WMC and decoding, as signifi- cant contributors to comprehension, we had expected direct paths from these variables as well. Indeed, language experience and rea- soning alone accounted for as much variance in comprehension (76%) as did the full set of measures (77%).

The pattern of direct and indirect effects in Model 1 has impor- tant implications for evaluating theories about the role of WMC in comprehension. As we described in the introduction, a direct path from WMC to comprehension is predicted by the Capacity Theory and the SSIR Theory, whereas the connectionist-based framework and the LTWM theory predict an indirect path via shared variance between measures of WMC and language experience. In Model 1, we found that WMC affected comprehension indirectly via shared variance with other latent variables, in particular, language experi- ence, reasoning, and fluency. The indirect effect is consistent with previous research using an SEM framework by Britton et al. (1998), but inconsistent with findings by Hannon (2012).

Subsequent analyses were informative about the nature of the relations among WMC language experience and comprehension. In Model 3 and Model 3-Reduced, measures of general reasoning were deleted from the dataset. In the absence of a reasoning vari- able, the path from WMC to comprehension was significant (Model 3). The path was stable in that removing additional latent variables from the dataset did not affect model fit. Language experience and WMC alone accounted for approximately the same amount of vari- ance that was explained in the full model (74%). The indirect rela- tion between WMC and comprehension and the strong covariance between WMC and language experience is consistent with the connectionist-based framework and LTWM theory. Interestingly, however, eliminating language experience measures from the dataset (Model 2) did not result in a significant path from WMC to comprehension as might be expected given predictions from these accounts. Instead, WMC had a significant direct path only when measures related to reasoning were eliminated from the dataset (Model 3). Thus, the covariance between WMC and reason- ing appears to be more important in explaining the correlation between WM span and comprehension than is the covariance between WMC and language experience, a finding that is some- what problematic for these latter two models. A more parsimo- nious account of our data is that WMC is mostly secondary to reasoning, even though it has indirect paths through both reason- ing and language experience. Individuals perform well on span tasks because they are high on reasoning and they perform well on comprehension tests because they are high on reasoning and language experience. This account goes somewhat further than the connectionist-based and LTWM in diminishing the theoretical importance of WMC in explaining individual variation in adult reading comprehension.

Our failure to find a direct effect of WMC on comprehension is significant given that WMC is among the most studied individual- difference variables in proficient adult readers. Much of this research, however, has been conducted using on-line measures of processing rather than off-line comprehension measures, as we did in this study. This raises an important question: does WMC have a direct effect on on-line processing measures such as reading time, eye fixations, and ERPs? A recent study by Van Dyke and colleagues suggest that the answer to this question may be no (Van Dyke et al., 2014). They analyzed reading times and recall as a function of sev- eral variables, including IQ, WMC, and vocabulary. Importantly, they showed that WMC had no effect on reading time in an analysis accounting for shared variance between WMC and IQ. Although Van

Dyke et al. found no unique effect of WMC on reading times, aspects of their design prevented them from entering all variables in a sin- gle model and their sample size was relatively small. Thus, an important goal of future research should be to examine the direct and indirect effects of WMC on on-line measures when general rea- soning is included in the models. The results of such studies may lead to a change in the emphasis placed on WMC in research on on-line language processing. This change may be warranted even if WMC is found to have a significant direct effect on on-line mea- sures. It is reasonable to ask whether WMC should be the focus of so much individual-difference research when it has no ultimate effect on comprehension. Rather, it may be more worthwhile to understand how low-capacity readers compensate such that their overall comprehension is unaffected. At minimum, researchers who use span measures to predict on-line processing should acknowledge that evidence for a direct effect of WMC on compre- hension is scant and should interpret their results accordingly.

Our results highlight the importance of understanding why WM span measures are correlated with general reasoning tasks (Shipstead, Redick, Hicks, & Engle, 2012). Several studies have examined the relation between WMC and measures of fluid intel- ligence, such as Raven’s Matrices and concluded that attentional control plays a significant role in the correlation (Engle & Kane, 2004), but is probably not solely responsible for it (Harrison, Shipstead, & Engle, 2014; Shipstead, Harrison, & Engle, 2015; Unsworth & Spillers, 2010). The strong correlation between WMC and fluid intelligence makes it important to find situations in which the measures .dissociate. Our results suggest that reading comprehension is one of these situations and that it may be fruitful to conduct studies to exploit this dissociation by manipulating text properties that differentially strengthen the relations among com- prehension, WMC, and general reasoning.

The pattern of direct and indirect effects also has implications for evaluating the role of decoding in comprehension among profi- cient readers and for explaining why word decoding effects have been small or non-significant in studies that have included mea- sures of both decoding and vocabulary (Bell & Perfetti, 1994; Braze et al., 2007; Cromley et al., 2010; Landi, 2010; Macaruso & Shankweiler, 2010). When measures of language experience were deleted from our dataset (Models 2 and 4), the decrease in explained variance was substantial, from 77% in the full model (Model 1) to 53% in Model 2 and 44% in Model 4. In Model 5, we examined the influence of decoding and vocabulary alone, similar to the analyses conducted by Braze et al. (2016) in their study of adolescent and community college participants. We found that only vocabulary had a direct effect; moreover, the amount of variance explained by vocabulary alone was 72% compared to the 76% that was explained by language experience and reasoning (Model 1- Reduced). This finding goes well beyond previous research in high- lighting the importance of word knowledge in explaining adult reading comprehension and is likely to explain why reading com- prehension in older adults is relatively unimpaired even when they exhibit significant deficits in general cognitive ability. Our finding is also consistent with Braze et al.’s suggestion that the influence of decoding on comprehension probably diminishes with reading expertise (see also, Protopapas, Mouzaki, Sideridis, Kotsolakou, & Simos, 2013). Thus, the Simple View of Reading in which decoding plays a large role is most appropriate as a model of comprehension in children; the direct influence of decoding on comprehension declines significantly in adolescent readers and disappears in the most proficient adult readers. Although proficient readers differ in their decoding abilities, slow decoding impairs comprehension pri- marily because it is associated with limitations in vocabulary.

The significance of language experience (and more specifically, vocabulary) in this study suggests that more research should be directed at understanding precisely how knowledge about words

E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153 151

affects comprehension. Obviously, comprehension will be impaired to the extent that a text contains words that a reader does not know. Among proficient readers, however, the influence of word knowledge is likely to be more subtle and more interesting (Adelman, Brown, & Quesada, 2006; Adolf, Frishkoff, Dandy, & Perfetti, 2016; Durso & Shore, 1991). Knowledge about words is not all-or-none and proficient readers are likely to have at least some knowledge about most of the words that they encounter. Subtle variations in word meanings, however, can have significant consequences for comprehension. Thus, it is important to under- stand how the depth of a reader’s vocabulary affects comprehen- sion. One possibility is that knowledge about words facilitates a reader’s ability to make predictions about upcoming words. Recent advances in computational psycholinguistics have contributed to the development of metrics (e.g., surprisal, entropy) that quantify the likelihood that words will appear in very specific contexts. High vocabulary readers are likely to have more knowledge than low vocabulary ones about how words are used across contexts and may be better at predicting subsequent words (Kuperman & Van Dyke, 2011, 2013). This knowledge may be important in pro- cessing incoming words even if the words themselves are not pre- dictable in context. Boudewyn, Long, and Swaab (2015) examined the processing of nouns (e.g., cake) that were plausible in context, but had low cloze probabilities in their specific contexts. They found that semantic features of the word that were relevant to the context (e.g., sweet) were activated before the word was received, even when that word could not be predicted. The ability to predict context-relevant features of subsequent words would mean that high vocabulary readers could integrate incoming words much faster than other readers.

The results of our study must be interpreted in light of several limitations. One limitation involves the large number of parame- ters that were estimated in our full model. This raises a concern about whether some paths in our model would have been signifi- cant with more power. This concern is partially addressed in our Model-Reduced analyses, in which we deleted measures from our data set, decreasing the number of latent variables. Removing these variables from the model reduced the number of parameters to be estimated, but did not reduce the number of participants. Importantly, these analyses did not yield an increase in the number of significant paths that we detected. A second limitation is that we were unable to create a latent variable for inhibition/suppression ability. We lost one measure of inhibition due to a programming error and another loaded with decoding measures in our factor analyses. Our remaining measure, Stroop Interference, covaried with our latent variables in the manner that we hypothesized, but the direct path to comprehension was not significant. This result may change if inhibition is assessed as a latent construct from multiple measures. Finally, the outcome of an SEM is strongly dependent on the inter-correlations among variables; thus, out- comes can change when variables are added or deleted. We demonstrated this when we conducted analyses in which we elim- inated measures of language experience and general reasoning. Our full model contained the variables that we hypothesized were most likely to be predictive of comprehension, but numerous other abilities are likely to play a role. Further research will be necessary to determine whether our pattern of findings hold when other lan- guage and cognitive abilities are assessed.

Conclusions

The results of the current study have both practical and theoret- ical significance. Practically, our results offer information about the identification of struggling readers and the design of effective remediation. Our results showed that language experience, as

assessed by vocabulary tests, is a robust means of identifying poor comprehenders. Importantly, research has shown that vocabulary knowledge is particularly amenable to training (Coyne, McCoach, & Kapp, 2007; Roberts, Torgesen, Boardman, & Scammacca, 2008; Scammacca, Roberts, Vaughn, & Stuebing, 2015). Further research should be directed in understanding how readers learn and use subtle distinctions of word meanings as they develop their mental representations of texts. Theoretically, our results contribute to understanding the complex correlations among individual- difference variables and reading comprehension. We have shown that proficient adult readers vary substantially in their decoding ability, but the variance in decoding is secondary to language expe- rience in predicting individual differences in reading comprehen- sion. In addition, our results raise significant concerns about the unique role of WMC in comprehension, suggesting that its contri- bution is secondary to general reasoning ability.

Acknowledgments

This research was supported by grants from the National Insti- tutes of Health to Debra Long (R01HD048914 and R01HD073948). We thank Matthew Traxler, Shelley Blozis, Keith Widaman, Emilio Ferrer, Chantel Prat, Clinton Johns, and a large number of graduate and undergraduate students for their valuable help on this project.

A. Supplementary material

Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.jml.2017.07.008.

References

Adams, M. J. (Ed.). (1990). Beginning to read: Thinking and learning about print. Cambridge, MA: MIT Press.

Adelman, J. S., Brown, G. D., & Quesada, J. F. (2006). Contextual diversity, not word frequency, determines word-naming and lexical decision times. Psychological Science, 17, 814–823.

Adolf, S. M., Catts, H. W., & Little, T. D. (2006). Should the simple view of reading include a fluency component? Reading and Writing, 19, 933–958.

Adolf, S., Frishkoff, G., Dandy, J., & Perfetti, C. (2016). Effects of induced orthographic and semantic knowledge on subsequent learning: A test of the partial knowledge hypothesis. Reading and Writing, 29, 475–500.

Afflerbach, P. (1986). The influence of prior knowledge on expert readers’ importance assignment processes. National Reading Conference Yearbook, 35, 30–40.

Alba, J. W., Alexander, S. G., Hasher, L., & Caniglia, K. (1981). The role of context in the encoding of information. Journal of Experimental Psychology: Human Learning and Memory, 7, 283–292.

Bell, L. C., & Perfetti, C. A. (1994). Reading skill: Some adult comparisons. Journal of Educational Psychology, 86, 244–255.

Borella, E., Carretti, C., & Pelegrina, S. L. (2010). The specific role of inhibitory efficacy in good and poor comprehenders. Journal of Learning Disabilities, 43, 541–552.

Borella, E., & de Ribaupierre, A. (2014). The role of working memory, inhibition, and processing speed in text comprehension in children. Learning and Individual Differences, 34, 86–92.

Boudewyn, M. A., Long, D. L., & Swaab, T. Y. (2015). Graded expectations: Predictive processing during spoken language comprehension. Cognitive, Affective, & Behavioral Neuroscience, 15, 607–624.

Bransford, J. D., & Johnson, M. K. (1972). Contextual prerequisites for understanding: Some investigations of comprehension and recall. Journal of Verbal Learning & Verbal Behavior, 11, 717–726.

Braze, D., Katz, L., Magnuson, J. S., Mencl, W. E., Tabor, W., Van Dyke, J. A., ... Shankweiler, D. P. (2016). Vocabulary does not complicate the simple view of reading. Reading and Writing, 29, 435–451.

Braze, D., Tabor, W., Shankweiler, D. P., & Mencl, W. E. (2007). Speaking up for vocabulary: Reading skill differences in young adults. Journal of Learning Disabilities, 40, 226–243.

Britton, B. K., Stimson, M., Stennett, B., & Gülgöz, S. (1998). Learning from instructional text: Test of an individual-differences model. Journal of Educational Psychology, 90(3), 476–491.

Brown, J., Bennett, J., & Hanna, G. (1980). The Nelson-Denny reading test. Boston: Houghton Mifflin.

Brown, J., Fischo, V., & Hanna, G. (1993). The Nelson-Denny reading test. Boston: Houghton Mifflin.

152 E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153

Caplan, D., DeDe, G., Waters, G., Michaud, J., & Tripodis, Y. (2011). Effects of age, speed of processing, and working memory on comprehension of sentences with relative clauses. Psychology and Aging, 26, 439–450.

Carroll, J. B., Davies, P., & Richman, B. (1971). The American heritage word frequency book. Boston: Houghton Mifflin.

Chi, M. T. H., Feltovich, P. J., & Glaser, R. (1981). Categorization and representation of physics problems by experts and novices. Cognitive Science, 5, 121–152.

Chiesi, H. L., Spilich, G. J., & Voss, J. F. (1979). Acquisition of domain-related information in relation to high and low domain knowledge. Journal of Verbal Learning & Verbal Behavior, 18, 257–273.

Conway, A. R. A., & Engle, R. W. (1994). Working memory and retrieval: A resource- dependent inhibition model. Journal of Experimental Psychology: General, 123(4), 354–373.

Coyne, M. D., McCoach, D. B., & Kapp, S. (2007). Vocabulary intervention for kindergarten students: Comparing extended instruction to embedded instruction and incidental exposure. Learning Disability Quarterly, 30, 74–88.

Craik, F. I. M. (1986). A functional account of age differences in memory. In F. Klix & H. Hagendorf (Eds.), Human memory and cognitive capabilities, mechanisms, and performance (pp. 409–422). Amsterdam: North-Holland.

Cromley, J. G., Snyder-Hogan, L. E., & Luciw-Dubas, U. A. (2010). Reading comprehension of scientific text: A domain-specific test of the direct and inferential mediation model of reading comprehension. Journal of Educational Psychology, 102, 687–700.

Cunningham, A. E., & Stanovich, K. E. (1990). Assessing print exposure and orthographic processing skill in children: A quick measure of reading experience. Journal of Educational Psychology, 82(4), 733–740.

Cunningham, A. E., & Stanovich, K. E. (1991). Tracking the unique effects of print exposure in children: Associations with vocabulary, general knowledge, and spelling. Journal of Educational Psychology, 83(2), 264–274.

Daneman, M., & Merikle, P. M. (1996). Working memory and language comprehension: A meta-analysis. Psychonomic Bulletin & Review, 3(4), 422–433.

Deary, J. J. (2000). Looking down on human intelligence: From psychometrics to the brain. Oxford, England: Oxford University Press.

Durso, F. T., & Shore, W. J. (1991). Partial knowledge of word meanings. Journal of Experimental Psychology: General, 120, 190–202.

Ekstrom, R. B., French, J. W., Harman, H. H., & Dermen, D. (1976). Manual for kit of factor-referenced cognitive tests. Princeton, NJ: Educational Testing Service.

Engle, R. W., Conway, A. R. A., Tuholski, S. W., & Shisler, R. J. (1995). A resource account of inhibition. Psychological Science, 6(2), 122–125.

Engle, R. W., & Kane, M. J. (2004). Executive attention, working memory capacity, and a two-factor theory of cognitive control. In B. Ross (Ed.), The psychology of learning and motivation (pp. 145–199). New York: Elsevier.

Engle, R. W., Kane, M. J., & Tuholski, S. W. (1999). Individual differences in working memory capacity and what they tell us about controlled attention, general fluid intelligence and functions of the prefrontal cortex. In A. Miyake & P. Shah (Eds.), Models of working memory: Mechanisms of active maintenance and executive control (pp. 102–134). London: Cambridge Press.

Engle, R. W., Tuholski, S. W., Laughlin, J. E., & Conway, A. R. A. (1999). Working memory, short-term memory, and general fluid intelligence: A latent-variable approach. Journal of Experimental Psychology: General, 128(3), 309–331.

Ericsson, K. A., & Kintsch, W. (1995). Long-term working memory. Psychological Review, 102(2), 211–245.

Ericsson, K. A., & Smith, J. (1991). Prospects and limits of the empirical study of expertise: An introduction. In K. A. Ericsson & J. Smith (Eds.), Toward a general theory of expertise: Prospects and limits (pp. 1–38). New York, NY: Cambridge University Press.

Gernsbacher, M. A. (1990). Language comprehension as structure building. Hillsdale, NJ: Lawrence Erlbaum Associates Inc.

Gernsbacher, M. A. (1993). Less skilled readers have less efficient suppression mechanisms. Psychological Science, 4, 294–298.

Gernsbacher, M. A. (1997). Two decades of structure building. Discourse Processes, 23, 265–304.

Gernsbacher, M. A., Robertson, R. R. W., Palladino, P., & Werner, N. K. (2004). Managing mental representations during narrative comprehension. Discourse Processes, 37, 145–164.

Gernsbacher, M. A., Varner, K. R., & Faust, M. (1990). Investigating differences in general comprehension skill. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 430–445.

Gough, P. B., & Tumner, W. E. (1986). Decoding, reading, and reading disability. RASE. Remedial & Special Education, 7, 6–10.

Hannon, B. (2012). Understanding the relative contributions of lower-level word and working memory to reading proficient adult readers. Reading Research Quarterly, 47, 125–152.

Harrison, T. L., Shipstead, Z., & Engle, R. W. (2014). Why is working memory capacity related to matrix reasoning tasks? Memory & Cognition, 43, 226–236.

Hayes, D. P., & Ahrens, M. G. (1988). Vocabulary simplification for children: A special case of ‘‘motherese?”. Journal of Child Language, 15, 395–410.

Hoover, W. A., & Gough, P. B. (1990). The simple view of reading. Reading and Writing, 2, 127–160.

Joshi, R. M. (2005). Vocabulary: A critical component of comprehension. Reading & Writing Quarterly: Overcoming Learning Difficulties, 21, 209–219.

Joshi, R. M., & Aaron, P. G. (2000). The component model of reading: Simple view of reading made a little more complex. Reading Psychology, 21, 85–97.

Just, M. A., & Carpenter, P. A. (1980). A theory of reading: From eye fixations to comprehension. Psychological Review, 87, 329–354.

Just, M. A., & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in working memory. Psychological Review, 99, 122–149.

Kane, M. J., Bleckley, M. K., Conway, A. R. A., & Engle, R. W. (2001). A controlled- attention view of working-memory capacity. Journal of Experimental Psychology: General, 130(2), 169–183.

Kane, M. J., & Engle, R. W. (2000). Working-memory capacity, proactive interference, and divided attention: Limits on long-term memory retrieval. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26(2), 336–358.

King, J., & Just, M. A. (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language, 30, 580–602.

Kuperman, V., & Van Dyke, J. A. (2011). Effects of individual differences in verbal skills on eye-movement patterns during sentence reading. Journal of Memory and Language, 65, 42–73.

Kuperman, V., & Van Dyke, J. A. (2013). Reassessing word frequency as a determinant of word recognition for skilled and unskilled readers. Journal of Experimental Psychology: Human Perception and Performance, 39, 802–823.

Landi, N. (2010). An examination of the relationship between reading comprehension, higher-level and lower-level reading sub-skills in adults. Reading and Writing, 23, 701–717.

Long, D. L., Johns, C. L., & Jonathan, E. (2012). A memory-retrieval view of discourse representation: The recollection and familiarity of text ideas. Language and Cognitive Processes, 27(6), 821–843.

Long, D. L., & Prat, C. S. (2002). Memory for Star Trek: The role of prior knowledge in recognition revisited. Journal of Experimental Psychology: Learning, Memory, & Cognition, 28, 1073–1082.

Long, D. L., & Prat, C. S. (2008). Individual differences in syntactic ambiguity resolution: Readers vary in their use of plausibility information. Memory & Cognition, 36(2), 375–391.

Long, D. L., Prat, C., Johns, C., Morris, P., & Jonathan, E. (2008). The importance of knowledge in vivid text memory: An individual-differences investigation of recollection and familiarity. Psychonomic Bulletin & Review, 15(3), 604–609.

Long, D. L., Wilson, J., Hurley, R., & Prat, C. S. (2006). Assessing text representations with recognition: The interaction of domain knowledge and text coherence. Journal of Experimental Psychology: Learning, Memory, and Cognition, 32, 816–827.

Lundeberg, M. A. (1987). Metacognitive aspects of reading comprehension: Studying understanding in legal case analysis. Reading Research Quarterly, 22, 407–432.

Macaruso, P., & Shankweiler, D. (2010). Expanding the simple view of reading in accounting for reading skills in community college students. Reading Psychology, 31(5), 454–471.

MacDonald, M. C., & Christiansen, M. H. (2002). Reassessing working memory: A comment on Just & Carpenter (1992) and Waters & Caplan (1996). Psychological Review, 109, 35–54.

McVay, J. C., & Kane, M. J. (2012). Why does working memory capacity predict variation in reading comprehension? On the influence of mind wandering and executive attention. Journal of Experimental Psychology: General, 141(2), 302–320.

Means, M. L., & Voss, J. F. (1985). Star wars: A developmental study of expert and novice knowledge structures. Journal of Memory and Language, 24, 746–757.

Olson, R. K., Kliegl, R., Davidson, B. J., & Foltz, G. (1985). Individual and developmental differences in reading disability. In G. E. MacKinnon & T. G. Waller (Eds.). Reading research: Advances in theory and practice (Vol. 4, pp. 1–64). New York, N.Y.: Academic Press.

Olson, R. K., Wise, B., Conners, F., Rack, J., & Fulker, D. (1989). Specific deficits in component reading and language skills: Genetic and environmental influences. Journal of Learning Disabilities, 22, 339–348.

Payne, B. R., & Stine-Morrow, E. A. L. (2014). Adult age differences in wrap-up during sentence comprehension: Evidence from Ex = Gaussian distributional analyses of reading time. Psychology and Aging, 29, 213–228.

Perfetti, C. A. (1985). Reading ability. New York, NY: Oxford University Press. Perfetti, C. A. (2007). Reading Ability: Lexical Quality to Comprehension. Scientific

Studies of Reading, 11, 357–383. Perfetti, T. W., & Hogaboam, T. (1978). Relationship between single word decoding

and reading comprehension skill. Journal of Educational Psychology, 67, 461–469. Peter, B., Matsuishita & Raskind, W. H. (2011). Global processing speed in children

with low reading ability and in children and adults with typical reading ability: Exploratory factor analytic models. Journal of Speech, Language, and Hearing Research, 54, 885–889.

Protopapas, A., Mouzaki, A., Sideridis, G. D., Kotsolakou & Simos, P. G. (2013). The role of vocabulary in the context of the simple view of reading. Reading & Writing Quarterly, 29, 168–202.

Raven, J. C. (1962). Advanced progressive matrices (set II). London: Lewis. Roberts, G., Torgesen, J. K., Boardman, A., & Scammacca, N. (2008). Evidence-based

strategies for reading instruction of older students with learning disabilities. Learning Disabilities Research & Practice, 23, 63–69.

Rosen, V. M., & Engle, R. W. (1997). The role of working memory capacity in retrieval. Journal of Experimental Psychology: General, 126(3), 211–227.

Salthouse, T. A., & Babcock, R. L. (1991). Decomposing adult age differences in working memory. Developmental Psychology, 27, 763–776.

Salthouse, T. A. (1988). The role of processing resources in cognitive aging. In M. L. Howe & C. J. Brainerd (Eds.), Cognitive development in adulthood. New York, N.Y.: Springer-Verlag.

Scales, A. M., & Rhee, O. (2001). Adult reading habits and patterns. Reading Psychology, 22, 175–203.

E.M. Freed et al. / Journal of Memory and Language 97 (2017) 135–153 153

Scammacca, N., Roberts, G., Vaughn, S., & Stuebing, K. K. (2015). A meta-analysis of interventions for struggling readers in grades 4–12: 1980–2011. Journal of Learning Disabilities, 48, 369–390.

Schneider, W., Körkel, J., & Weinert, F. E. (1990). Expert knowledge, general abilities, and text processing. In W. Schneider & F. E. Weinert (Eds.), Interactions among aptitudes, strategies, and knowledge in cognitive performance (pp. 235–251). New York: Springer-Verlag.

Shipstead, Z., Harrison, T. L., & Engle, R. W. (2015). Working memory capacity and the scope and control of attention. Attention, Perception, & Psychophysics, 77, 1863–1880.

Shipstead, Z., Redick, T. S., Hicks, K. L., & Engle, R. W. (2012). The scope and control of attention as separate aspects of working memory. Memory, 20, 608–628.

Silverman, R. D., Speece, D. L., Harring, J. R., & Ritchey, K. D. (2013). Fluency has a role in the simple view of reading. Scientific Studies of Reading, 17, 108–133.

Snow, C. E., Burns, M. S., & Griffin, P. (Eds.). (1998). Preventing reading difficulties in young children. Washington, DC: National Academy Press.

Spilich, G. J., Vesonder, G. T., Chiesi, H. L., & Voss, J. F. (1979). Text processing of domain-related information for individuals with high and low domain knowledge. Journal of Verbal Learning & Verbal Behavior, 18, 275–290.

Stanovich, K. E., & Cunningham, A. E. (1992). Studying the consequences of literacy within a literate society: The cognitive correlates of print exposure. Memory & Cognition, 20, 51–68.

Stanovich, K. E., & West, R. F. (1989). Exposure to print and orthographic processing. Reading Research Quarterly, 24, 402–433.

Sulin, R. A., & Dooling, D. J. (1974). Intrusion of thematic idea in retention of prose. Journal of Experimental Psychology, 103, 255–262.

Summers, W. V., Horton, D. L., & Diehl, V. A. (1985). Contextual knowledge during encoding influences sentence recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 11, 771–779.

Swanson, H. L. (1996). Individual and age-related differences in children’s working memory. Memory & Cognition, 24, 70–82.

Swanson, H. L., Howard, C. B., & Saez, L. (2006). Do different components of working memory underlie different subgroups of reading disabilities? Journal of Learning Disabilities, 39, 252–269.

Taylor, J. N., & Perfetti, C. A. (2016). Eye movements reveal readers’ lexical quality and reading experience. Reading & Writing, 29, 1069–1103.

Tighe, E. L., & Schatschneider, C. (2016). Modeling the relations among morphological awareness dimensions, vocabulary knowledge, and reading comprehension in adult basic education students. Frontiers in Psychology. http://dx.doi.org/10.3389/fpsyg2016.00086.

Tiu, R. D., Thompson, L. A., & Lewis, B. A. (2003). The role of IQ in a component model of reading. Journal of Learning Disabilities, 36, 424–436.

Unsworth, N., Heitz, R. P., Schrock, J. C., & Engle, R. W. (2005). An automated version of the operation span task. Behavior Research Methods, 37, 498–505.

Unsworth, N., & Spillers, G. J. (2010). Working memory capacity: Attention control, secondary memory, or both? A direct test of the dual-component model. Journal of Memory and Language, 62, 392–406.

Van Dyke, J. A., Johns, C. L., & Kukona, A. (2014). Low working memory capacity is only spuriously related to poor reading comprehension. Cognition, 131(3), 373–403.

Verhoeven, L., & Van Leeuwe, J. (2008). Prediction of the development of reading comprehension: A longitudinal study. Applied Cognitive Psychology, 22, 407–423.

Waters, G. S., & Caplan, D. (1996). The capacity theory of sentence comprehension: Critique of Just and Carpenter (1992). Psychological Review, 103(4), 761–772.

Wells, J. B., Christiansen, M. H., Race, D. S., Acheson, D. J., & MacDonald, M. C. (2009). Experience and sentence comprehension: Statistical learning and relative clause comprehension. Cognitive Psychology, 58, 250–271.

West, R. F., & Stanovich, K. E. (1991). The incidental acquisition of information from reading. Psychological Science, 2, 325–330.

West, R. F., Stanovich, K. E., & Mitchell, H. R. (1993). Reading in the real world and its correlates. Reading Research Quarterly, 28, 34–50.

Zahler, D., & Zahler, K. (2003). Test prep your IQ cultural literacy (1st ed.). Lawrenceville, NJ: Peterson’s.