This article was downloaded by: [University of Newcastle (Australia)] On: 26 August 2014, At: 18:16 Publisher: Routledge Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK Reading Psychology Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/urpy20 The Relationship Between Receptive and Expressive Subskills of Academic L2 Proficiency in Nonnative Speakers of English: A Multigroup Approach Hye K. Pae a & Daphne Greenberg b a School of Education , University of Cincinnati , Cincinnati, Ohio b Department of Educational Psychology and Special Education , Georgia State University , Atlanta , Georgia Published online: 23 Jan 2014. To cite this article: Hye K. Pae & Daphne Greenberg (2014) The Relationship Between Receptive and Expressive Subskills of Academic L2 Proficiency in Nonnative Speakers of English: A Multigroup Approach, Reading Psychology, 35:3, 221-259, DOI: 10.1080/02702711.2012.684425 To link to this article: http://dx.doi.org/10.1080/02702711.2012.684425 PLEASE SCROLL DOWN FOR ARTICLE Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of the
Transcript
This article was downloaded by: [University of Newcastle (Australia)]On: 26 August 2014, At: 18:16Publisher: RoutledgeInforma Ltd Registered in England and Wales Registered Number: 1072954Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH,UK
Reading PsychologyPublication details, including instructions forauthors and subscription information:http://www.tandfonline.com/loi/urpy20
The Relationship BetweenReceptive and ExpressiveSubskills of Academic L2Proficiency in NonnativeSpeakers of English: AMultigroup ApproachHye K. Pae a & Daphne Greenberg ba School of Education , University of Cincinnati ,Cincinnati, Ohiob Department of Educational Psychology and SpecialEducation , Georgia State University , Atlanta ,GeorgiaPublished online: 23 Jan 2014.
To cite this article: Hye K. Pae & Daphne Greenberg (2014) The RelationshipBetween Receptive and Expressive Subskills of Academic L2 Proficiency in NonnativeSpeakers of English: A Multigroup Approach, Reading Psychology, 35:3, 221-259, DOI:10.1080/02702711.2012.684425
To link to this article: http://dx.doi.org/10.1080/02702711.2012.684425
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all theinformation (the “Content”) contained in the publications on our platform.However, Taylor & Francis, our agents, and our licensors make norepresentations or warranties whatsoever as to the accuracy, completeness,or suitability for any purpose of the Content. Any opinions and viewsexpressed in this publication are the opinions and views of the authors, andare not the views of or endorsed by Taylor & Francis. The accuracy of the
Content should not be relied upon and should be independently verified withprimary sources of information. Taylor and Francis shall not be liable for anylosses, actions, claims, proceedings, demands, costs, expenses, damages,and other liabilities whatsoever or howsoever caused arising directly orindirectly in connection with, in relation to or arising out of the use of theContent.
This article may be used for research, teaching, and private study purposes.Any substantial or systematic reproduction, redistribution, reselling, loan,sub-licensing, systematic supply, or distribution in any form to anyone isexpressly forbidden. Terms & Conditions of access and use can be found athttp://www.tandfonline.com/page/terms-and-conditions
THE RELATIONSHIP BETWEEN RECEPTIVE ANDEXPRESSIVE SUBSKILLS OF ACADEMIC L2 PROFICIENCY
IN NONNATIVE SPEAKERS OF ENGLISH:A MULTIGROUP APPROACH
HYE K. PAESchool of Education, University of Cincinnati, Cincinnati, Ohio
DAPHNE GREENBERGDepartment of Educational Psychology and Special Education, Georgia State
University, Atlanta, Georgia
The purpose of this study was to examine the relationship between receptive andexpressive language skills characterized by the performance of nonnative speakers(NNSs) of English in the academic context. Test scores of 585 adult NNSs wereselected from Form 2 of the Pearson Test of English Academic’s field-test database.A correlated two-factor model was chosen as a baseline model and was tested formultigroup invariance. The results of this study suggest that a similar factorstructure underpins adult NNSs’ receptive and expressive linguistic performanceacross gender, but a different factor structure exists for high- and low-abilitygroups. Limitations and implications are discussed.
Introduction
Language skills require diverse componential skills, includingphonological, orthographic, semantic, and syntactic skills as wellas the social aspects of linguistic performance. Since linguisticabilities involve perception, comprehension, and production ofspeech as well as the written language, the most systematic reso-nance in language acquisition would be balanced proficiency inlistening, speaking, reading, and writing. Language acquisitionin a native language (first language; L1) primarily involves theassociation between spoken words and meanings, and, as liter-acy skills are attained, both spoken and print lexicons get con-nected to semantic properties (Adams, 1990). Unlike this sequen-tial development from spoken to print in L1 acquisition, second
Address correspondence to Hye K. Pae, School of Education, University ofCincinnati, 2610 McMicken Circle, Cincinnati, OH 45221-0022. E-mail: [email protected]
221
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
222 H. K. Pae and D. Greenberg
language (L2) learning typically involves spoken and print modal-ities at the same time. This concurrent feature of L2 learning maylead to complex and multifaceted linguistic activities. Receptiveand expressive language skills are instrumental in L2 proficiency.An examination of the complexity and multiplicity of L2 profi-ciency with respect to receptive and expressive skills will shed lighton a deeper understanding of the underlying constructs of L2skills. In spite of a wealth of research studies in L2 language andreading skills, little is known about the factor structure of recep-tive and expressive skills performed by nonnative speakers (NNSs)of English in the academic context. This study investigated therelations among linguistic indicators and between receptive andexpressive academic English skills using structural equation mod-eling (SEM).
Receptive and Expressive Language Skills in L2 Proficiency
Adults’ L2 learning is much more complex than that of L1 inpart because of the simultaneous learning of the four rudimen-tary skills, such as listening, speaking, reading, and writing, andbecause adults have already passed the critical period of naturallanguage acquisition. Due to these characteristics, L2 or foreignlanguage (FL) acquisition is more likely to take place in classroomenvironments rather than naturalistic settings, which results indifferences in the patterns of exposure, function, medium, andsocial interaction (Ellis, 2004). As opposed to L1 in which indi-viduals learn to speak effortlessly, intuitively, and automatically,L2 learning requires time, effort, motivation, input, use, and in-struction. Beyond time, effort, motivation, and instruction, L2input and its usage have constituted a viable research agendumfor decades in the field of second language acquisition (SLA).However, the role of L2 input and output in SLA has not beenfully addressed (VanPatten, 2004). A critical question on SLAlies in whether L2 skills are input dependent or L2 output sen-sitive. The question can also be expanded to the relationshipbetween the two modalities. L2 input is indispensable in thedevelopment of underlying linguistic competence such that itallows NNSs to make use of the knowledge base to compre-hend L2 words, phrases, and sentences, along with the context.This linguistic competence consists of mental representations of
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 223
phonological, syntactic, semantic, and sociopragmatic compo-nents of lexicons (VanPatten, 2004). Input which encompassesthese components serves as a contributory variable to advance-ment in L2 skills.
In order to gain proficiency in L2, input needs to be com-prehensible and meaningful to the learner to the degree that itbecomes intake (i.e., retained input), which allows for enhancednoticing of phonotactics, lexical recognition, morphosyntax, for-mulaic language, and semantics for comprehension (Ellis, 2004).Input does not automatically become intake unless attention andcognitive rehearsal are involved. Once input is retained in a com-prehensible and meaningful way, it is translated into receptiveskills in language learning. Receptive language skills are crucialin the pathway to oral language and literacy proficiency, alongwith expressive language skills (Barnett, Yarosz, Thomas, Jung, &Blanco, 2007; Bishop & Adams, 1990). Receptive language skillscomprise the comprehension of what is said and what is read asdecoding operations. Expressive language skills are related notonly to the ability to retrieve ideas and lexicons but also to theability to express ideas and thoughts in response to the given de-mand in an oral or written manner.
In the natural process of language acquisition, it is optimalto have symmetric skills in receptive and expressive skills in theforms of listening, speaking, reading, and writing. Vocabulary is aprime element of comprehension in that skilled readers need tohave receptive mastery of 95% or more of the words in the textfor full comprehension, coupled with automaticity of word pro-cessing (Adams, 1990; Grabe & Stroller, 2002). Webb (2008) hasinvestigated the relationship between receptive and productivevocabulary size of L2 learners and concluded that total recep-tive vocabulary size is larger than productive vocabulary partlybecause receptive mastery does not require receptive words to beproduced in speaking and writing. He suggested that receptivevocabulary size may serve as an indicator of productive vocabularysize. The disparity between receptive and expressive languageskills is found to be the result of variability of individuals’ repre-sentations of linguistic repertoire. The organization of receptivelanguage skills and the retrieval rate of lexicons stored in long-term memory may vary across individuals depending on theirprior knowledge and abilities to make use of the connections
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
224 H. K. Pae and D. Greenberg
among lexical and other linguistic entries. Research demonstratesthat receptive vocabulary gains precede expressive language skills(Barnett et al., 2007). A discrepancy between receptive andexpressive vocabulary growth rates has also been documentedin monolingual lexical acquisition (Bates, Bretheron, & Snyder,1988) and in dual-language learning (Barnett et al., 2007).
The importance of receptive and expressive language skillshas been highlighted mostly in children’s vocabulary researchstudies. Deficiencies in the use of expressive language inpreschool children have been found to be a cause of subsequentacademic difficulties (Bishop & Adams, 1990). The importance ofexpressive language has been echoed by other researchers. Specif-ically, Chiappe, Chiappe, and Gottardo (2004) have found, in acomparative study between poor readers and proficient readers,that expressive vocabulary knowledge has a stronger associationwith overall language skills than receptive vocabulary when con-trolling for age. The results also demonstrated that poor read-ers produced significantly lower scores in expressive vocabularyknowledge than their counterparts, while struggling and profi-cient readers did not differ in receptive vocabulary scores. Thisfinding suggests that expressive vocabulary skills play a more crit-ical role in reading than receptive vocabulary knowledge becauseexpressive skills, which are oral and productive in nature, re-quire holistically specified and organized mental representations(Chiappe, Chiappe, & Gottardo, 2004). Metsala (1997) has alsonoted the interplay of receptive and expressive vocabulary skills,suggesting a more pivotal role of expressive skills than receptiveskills in predicting reading outcomes.
Academic L2 Skills
In the field of SLA, social and academic language uses have beendifferentiated. As a foundational ability, basic interpersonal com-munication skills (BICS) are needed to perform day-to-day activ-ities in social contexts, while cognitive academic language profi-ciency (CALP) is required in formal academic settings in whichhigher-level cognitive tasks are demanded (Cummins, 2008). Con-versational fluency in BICS is expected to be acquired in ap-proximately 2 to 3 years of L2 use and exposure, while CALPtakes longer than BICS and is expected to be attained after
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 225
5 to 7 years of L2 exposure and use (Cummins, 2008). Academictextbook reading requires higher-order problem solving, analyti-cal evaluation, critical reasoning, and complex language skills, inpart because academic language is context reduced and cogni-tively demanding (Cummins, 2008).
As opposed to BICS for social purposes, CALP for academicpurposes often uses low-frequency academic content words. Poorreceptive abilities in relation to CALP hardly lead to a proper un-derstanding of the course content and lectures which are vitalfor successful academic achievement, leaving gaps in a knowledgebase. Poor expressive language can also result in lower academicachievement and the inability to work effectively with others in thesetting in which students are expected to give oral presentations,engage in oral discussion, and interact with their classmates ona daily basis. Receptive and expressive language skills are neededconsistently throughout schooling and in the work environment.
The components that underpin L2 proficiency and the na-ture of mapping processes involving SLA have been explainedin terms of the form-meaning connection (VanPatten, Williams,Rott, & Overstreet, 2004). The form-meaning connection ad-dresses learners’ internal response to their exposure to linguis-tic input by encoding the form’s referential meaning. There areseveral correspondences in the form-meaning relationship: one-to-one, one-to-many, and many-to-one form-meaning connections(VanPatten et al., 2004). Words are often characterized by phono-logical properties to determine their meanings. For example, theintricacies of sounds and graphemes in homophones (e.g., wholevs. hole), homographs (e.g., bat, a baseball equipment vs. bat, ananimal), and heterographs (e.g., present, verb vs. present, noun)can be resolved through the context. The quality of the connec-tion among form, meaning, and use is also dependent upon syn-tactic classes, frequencies, and lengths. It is useful to look at themechanism including all components at a glance. All the key ele-ments involved in L2 skills can be summarized as seen in Figure 1.
L2 learning begins with input, which grows into recep-tive skills through retained input (i.e., intake). Receptive skillsare characterized by listening and reading. At the core are L2skills marked by form (orthographic and syntactic properties),meaning (semantic properties), and use (contextual pragmatics).
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
226 H. K. Pae and D. Greenberg
FIGURE 1 A model of L2 skills.
L2 proficiency is also manifested though expressive modalitiesincluding speaking and writing. Expressive skills make use of theinteractive function of form, meaning, and use as well. All the el-ements are not mutually exclusive but interconnected. Unlike L1acquisition, affective factors play a part in the process.
Receptive knowledge is a necessary condition for languagelearning. At the same time, expressive skills are also essentialfor accuracy and fluency beyond internalized linguistic compe-tence, as expressive skills utilize underlying linguistic compe-tence in operation through self-correction and feedback. Despitea wealth of discussion on L2 proficiency, what is unclear is theextent to which global L2 skills are defined by receptive skills or
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 227
expressive skills. The question of whether L2 proficiency isreceptive-skill dependent, expressive-skill sensitive, or integrated-skill dependent can be examined by looking at the structure ofunderlying constructs and their relationships. Among many ele-ments involved in a model of L2 skills as seen in Figure 1, only re-ceptive and expressive modes of L2 skills were taken into accountin this study, as input is related to receptive skills and output toexpressive skills.
Unitary-Trait L2 Skills vs. Multitrait L2 Skills
Although receptive and expressive skills have been differentiated,a question may arise as to whether receptive and expressive lan-guage skills are interrelated traits or distinctive, independent en-tities. For representative linguistic skills of the receptive mode,Reves and Levine (1988) indicated that listening and reading aredistinct but similar skills sharing commonalties in integrated andholistic comprehension of the message. Further research studieshave suggested crossover and overlap between listening and read-ing abilities beyond the unique trait of skill operation and func-tion (Bae & Bachman, 1998; Buck, 1992, 2001). The interface be-tween speaking and writing as expressive modes is expected toplay out in a similar way to that of listening and reading, despitethe modality difference.
Of several hypotheses which have been proposed in relationto latent constructs underlying L2 proficiency, two lines of re-search findings have overshadowed the understanding of the re-lationship of linguistic-component skills. One line of research up-holds a single component hypothesis which assumes that L2 skillsconsist of a single canonical entity. For example, Oller (1979)proposed a unitary trait hypothesis through expectancy grammar,indicating that various modalities of language skills would stemfrom a single, general language ability. Oller’s expectancy gram-mar suggests that the receptive processing of information is basedon the ability to expect elements in sequence through sequentialencoding and decoding. In a similar vein, Rost (1993) found, us-ing factor analysis, that one broad factor, general reading competence,accounted for 85% of the variance in reading comprehension.
The other line of research findings has emphasized multi-componential language skills (Bachman, Davidson, & Foulkes,
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
228 H. K. Pae and D. Greenberg
1990; Bachman & Palmer, 1981, 1982; Sasaki, 1996; Shin, 2005),testing multiple factor models such as single-factor, two-factor, andsecond-order factor models. Evidence converges on multitrait ormulti-componential L2 abilities with different factor structures,challenging the unitary trait hypothesis. When explaining a multi-componential L2 ability, some researchers have found that corre-lated first-order factors are associated with L2 skills (Bachman &Palmer, 1981; Kunnan, 1995), whereas others have found that ahigher-order factor structure explained first-order factors (Bach-man & Palmer, 1982; Sasaki, 1996; Sawaki, Stricker, & Oranje,2009; Shin, 2005). This multiple-component theory posits thatmultiple entities for L2 are encoded and stored in a mental stor-age in the form of detailed perceptual constituents.
Although converging evidence documents multi-dimensional components involved in L2 performance, thereis little consensus on how many subskills are engaged in L2proficiency. The subskills that emerged have a wide range in thenumber and factors. For example, identified were 3 to 36 subskills(Alderson & Lukmani, 1989); four components of reading skills,such as word knowledge, comprehension of explicit meaning,comprehension of implicit/inferential meaning, and apprecia-tion (Lennon, 1962); four common factors in reading, includinggeneral comprehension, specific comprehension, decoding, andreading rate (Carroll, 1993); and two processes, such as decod-ing and comprehension (Lund, 1991). Recently, Song (2008)conducted a series of confirmatory factor analyses (CFA) toexamine divisible subskills in L2 listening and reading, and foundthat listening and reading assess two or three different kinds ofsubskills. Song (2008) suggested that the divisible subskills in L2comprehension tests might be dependent upon the test takers’L2 proficiency and the task characteristics of the test.
Invariance of Factor Structure across Groups
An examination of group differences has been one of the pop-ular analysis methods employed in the literature. With respectto factor differentiation depending on ability groups, there havebeen conflicting results reported. Some research findings havesuggested an indistinct subskill segregation between the low-and high-proficiency levels (Alderson, 1991; Oltman, Stricker, &
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 229
Barrows, 1988), while others indicated the increased degree ofskill differentiation (Swinton & Powers, 1980). Oltman et al.(1988) has proposed the former hypothesis in a three-way multi-dimensional scaling analysis of the test of English as a foreign lan-guage (TOEFL), stating that different modes of L2 skills becomeless differentiated as NNSs’ proficiency improves. Swinton andPowers (1980) have found, in a factor analysis of TOEFL items,that the score of the highest proficiency group accounts for thegreatest amount of factor separation and that of the lowest profi-ciency group explains the least amount of factor differentiation.
A number of research studies have shown statistical dif-ferences across gender. For example, significant differences ofperformance by gender were found in computer-based tests(Gallagher, Bridgeman, & Cahalan, 2002); item differential esti-mates in the tests of English proficiency as an FL (Ryan & Bach-man, 1992); L2 comprehension and vocabulary learning in avideo-based, computer-assisted language learning program (Lin,2011); and task performance in a tape-mediated assessment ofspeaking (Lumley & O’Sullivan, 2005).
As used in Stricker et al.’s (2005) study, an item parcelingpractice has been common in SEM (Bandalos & Finney, 2001).Item parceling involves a usage of a composite score resulted fromsumming or averaging two or more items as a unit of analysis.Once the unidimensionality assumption of the items are met, theadoption of item parceling not only produces increased reliability,
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
230 H. K. Pae and D. Greenberg
more normal distribution, and more stable parameter estimates;it also reduces the influence of idiosyncratic features of the items,diminishes biased parameter estimates, and simplifies the inter-pretation of model parameters (Bandalos & Finney, 2001). Re-search shows that item parceling is more powerful than item-levelanalysis with respect to the low χ2/df ratio and smaller error rates.For example, Yuan, Bentler, and Kano (1997) have indicated thatitem parceling yields greater power and smaller mean square er-ror than using individual items as indicators. Item parceling re-duces the influence of idiosyncratic features of the items and sim-plifies the interpretation of model parameters.
The Purpose of the Study
As stated earlier, the role of receptive and expressive languageskills has been investigated primarily in the area of children’sspeech-language impairments and reading disabilities. Moreover,research studies on receptive and expressive language skills havebeen limited to vocabulary size. A multitude of research stud-ies have provided evidence of a significant relationship betweenreceptive and expressive skills. However, there is a dearth ofresearch studies which investigate the underlying constructs ofreceptive and expressive skills in adults’ L2 performance.
A theoretical interest which motivated this study was totest whether adult NNSs’ observed data support the converg-ing evidence of a multicomponential model of English skills andwhether this model determines the extent to which the measure-ment and structural equation models fit the subsamples of genderand ability in terms of the degree of invariance in fit indices, pa-rameter estimates, and standard errors. The objective of this studywas two-fold. The first objective was to select the best fit modelfor the theoretical L21 English skills measured using the Pear-son Test of English Academic (PTE Academic) by testing mea-surement models for the underlying constructs of adult NNSs’ L2performance. The second objective was to investigate the extentto which the prediction model of L2 skills was similar across twoindependent subgroups (i.e., gender and ability subgroups). Thismultigroup CFA permitted an examination of the specified modelfor group differences in specific parameter estimates by imposingconstraints.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 231
This study was conducted as a secondary data analysis usingPTE Academic’s archival dataset. Three research questions wereaddressed in this study to investigate the relations of receptive andexpressive language skills by adult NNSs.
1. What is the most plausible factor structure in relation to adultNNSs’ English proficiency indicators?
2. To what extent do the covariances predicted by the model cor-respond to the observed covariances in the NNSs’ data?
3. To what extent does the multiple-grouping variable meetthe assumption of invariant paths or equal latent variablemodels?
Research Questions 1 and 2 addressed the first objective byanalyzing the mean and covariance structures of the observedvariables. Research Question 3 evaluated the extent to which thetrue model corresponded to the implied theoretical model acrosstwo subsamples of gender and ability groups (i.e., high- and low-ability groups) and the extent to which the theoretical model issupported by the obtained sample data.
Method
Participants
The participants were 585 adult NNSs from 62 countries. Indiahad the largest portion of the participant pool, followed by China,Israel, Taiwan, France, and Korea. Their mean age was 25 years,ranging from 17 to 59 years of age. Females accounted for 54.2%(317 test takers) and males 45.8% (268 test takers). Fifty-threepercent of the participants had studied English for more than10 years. Fifty-seven percent of them had lived in English-speakingcountries. Although the participants were recruited from 62 coun-tries, sample homogeneity was found to be achieved via differen-tial item functioning analyses (see Pae, 2012 for details), whichshowed no specific item advantages or disadvantages across gen-der and ability subgroups.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
232 H. K. Pae and D. Greenberg
Measures
The database utilized for this study was part of the larger fieldtest administered by PTE Academic. PTE Academic is a computer-based, international, academic English measure assessing NNSs’global English skills, which are required for English-speakingacademic settings. The use of the test results is not limitedto higher education institutions but also applicable to govern-ment departments and other organizations which require aca-demic English proficiency (Pearson, 2009). The contents of theexam cover a wide range of English skills, including listen-ing, reading, speaking, and writing skills as well as skill-specificdomains, such as fluency, grammar, pronunciation, spelling, vo-cabulary, and written discourse skills (Pearson, 2009). These var-ious scores are grouped into the scores of two traits: commu-nicative skills (listening, reading, speaking, and writing) andenabling skills (fluency, grammar, pronunciation, spelling, vo-cabulary, and written discourse). The measure included 20item types2 which were composed of a total of 87 items onthe basis of authentic task-based examples from academic set-tings (Pearson, 2009). The 20 item types assessed different skilldomains, including four discrete, independent domains (listen-ing, speaking, reading, and writing) and five interconnected-skill domain measures (listening/speaking, listening/reading, listen-ing/writing, reading/writing, and reading/speaking). Receptive skillswere defined as listening, reading, and listening/reading , whereas ex-pressive skills were indicated by speaking, writing, listening/writing,reading/writing, and reading/speaking. Although it may be question-able whether or not the shared-skill domain can be designatedinto one of the latent variables, a decision rule was made basedon the output modality. Specifically, as seen in the expressiveconstruct, reading/writing constitutes combined, interactive skillsbut reading was instrumental to output a writing product. There-fore, the reading/writing indicator was considered an expressiveconstruct.
coefficients falling in the acceptable range and no conspicuousdistortion from the Rasch model (Pae, 2012).
Observed-Variable Building
PTE Academic test items that were classified under the same skillswere combined to create an observed variable. Despite concernsabout item parceling (i.e., item parceling can obscure the factorstructure of the data; West, Finch, & Curran, 1995), the use ofitem parcels has been a common practice in SEM and has beenconsidered more powerful than item use (Bandalos & Finney,2001). It has been documented that item-parceling is more re-liable than using individual items in SEM when items share thesame traits or are conceptually related skill domains (Kishton &Widaman, 1994). Moreover, item parceling was recommendedwhen items share a unidimensional psychometrical nature(Bandalos & Finney, 2001; Kishton & Widaman, 1994) and whenthe composite score is normally distributed because item parcelsconform more closely to the assumptions of theory-based esti-mation methods (Enders & Bandalos, 1999; Bandalos & Finney,2001). Prior to item parceling made for this study, unidimen-sionality of the items was checked. Substantial evidence of unidi-mensionality was found through an item analysis. Using the itemparceling method, nine observed variables were created and en-tered into the equation of SEM. Since the items showed congru-ence within unidimensional domains, no distortions of the factorstructure and biased parameter estimate were expected.
With regard to the ability grouping, a median-split methodwas used to categorize the ability groups (i.e., high and low) inthis study. There has been some criticism on using median splits,but the use of median splits has provided differential effects acrossgroups in other studies and has been a common practice (Pae,Greenberg, & Morris, 2012; Qingquan, Chatupote, & Teo, 2008).
Data Screening
An SEM analysis of PTE Academic skill measures was carriedout using the LISREL 8.80 statistical program (Joreskog &Sorbom, 2007). SEM was chosen as a statistical technique due to
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
234 H. K. Pae and D. Greenberg
its advantages over regression modeling, such as its more flexi-ble assumptions, use of CFA to reduce measurement error by hav-ing multiple indicators per latent variable, the ability to model er-ror terms, the ability to test coefficients across multiple between-subjects groups, and the desirability of its strategy of comparingalternative models to assess relative model fit (Garson, 2011).
Univariate normality was checked using skewness and kur-tosis statistics. The variables were approximately normally dis-tributed (skewness and kurtosis values <1.0, except the kurto-sis estimate of speaking [value: −1.10]). There were no missingvalues because when examinees skipped items, the items weregiven scores of 0. No conspicuous outliers were observed. Thetest of multivariate normality for continuous variables showed thatthe assumption of a multivariate normal data distribution mightbe violated. Hence, two different analyses were performed us-ing normal scores and the maximum likelihood estimation tech-nique to estimate the factor loadings and test the statistical signifi-cance of the correlation coefficient between the constructs underconsideration.
Model Determination and Model Testing
To find a statistically significant theoretical model, several fit cri-teria were utilized for the statistical significance and substantivemeaning of a theoretical model. The statistical nonsignificanceof the chi-square likelihood test has been widely used to assess agoodness-of-fit in SEM. However, criticism has arisen for its un-satisfactory and problematic estimation (Schumacker & Lomax,2004). Given this limitation, as recommended by Schumacker andLomax (2004), three criteria were used in determining a theoret-ical model: (a) the nonsignificance of the χ2 statistics, the χ2/dfratio (3.0 or below, as suggested by Kline, 1998 ), the comparativefit index (CFI ; 0.95 or above), the goodness of fit index (GFI ; 0.95or above), the root mean-square error of approximation (RMSEA;0.05 or below), and expected cross-validation index (ECVI ); (b)the statistical significance of individual parameter estimates forthe paths in the model, compared to a tabled t value of 1.96 at the0.05 level of significance; and (c) the magnitude and the directionof the parameter estimates.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 235
Based on the theoretical basis on L2 skills, four competingmodels were tested in order to choose the best fit model for thedata. Below are the four competing models.
1. Model A. Correlated Two-Factor Model (Figure 2A): This model hy-pothesized that indicator variables had two underlying traits,receptive and expressive skills and that these two constructswere freely correlated with each other. The theoretical consid-eration for this model was based on the multi-componentialmodel of L2 skills (Bachman & Palmer, 1981, 1982; Sasaki,1996; Shin, 2005).
2. Model B. Single Factor Model (Figure 2B): This model hypothe-sized that a single first-order factor was related to all observedindicators. A theoretical relevance to this model stemmed froma unitary structure of language skills (Alderson, 1991; Oller,1979).
3. Model C. Second-Order Factor Model (Figure 2C): This modelhypothesized that the first-order factors of two separate la-tent traits (that is, receptive and expressive skills) were ex-plained by a higher-order factor structure (i.e., L2 English over-all skills). This model was considered based on Sawaki et al.’sstudy (2009).
4. Model D. Independent Two-Factor Model (Figure 2D): This modelhypothesized that discrete skills and integrated skills were inde-pendent from each other. A theoretical reason for this modelwas drawn from the findings that L2 skills were composedof multifaceted dimensions which were independent of eachother (Bachman & Palmer, 1982; Sasaki, 1996).
On the basis of the best model identified, a multiple-groupmodel was examined for gender (male vs. female) and abil-ity (high vs. low) subgroups in order to estimate separatelythe parameters for each independent sample as well as to testwhether specified parameters’ matrices were equivalent acrossthese groups. Model testing was involved in two steps in termsof group testing. First, primary single-group analyses were per-formed to establish a tenable measurement model. The chi-square value from an unequal-parameter model was compared tothat from an equal-parameter model with imposed constraints in
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
236 H. K. Pae and D. Greenberg
FIGURE 2 Conceptual models. (Continued)
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 237
FIGURE 2 Conceptual models. (Continued)
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
238 H. K. Pae and D. Greenberg
TABLE 1 Descriptive Statistics Using Raw Scores
Skewness Kurtosis
Variable Mean SD Range Statistics St. Error Statistics St. Error
order to test for partial measurement invariance across groups.The hypothesis for the different parameter estimates was that thetwo groups had equal parameter estimates in the path model.Second, with a tenable baseline model, between-group equiva-lence was investigated by imposing between-group equality con-straints on the parameters for invariance testing. In other words,the model was tested to evaluate whether the two groups shareda common path model. In this analysis, the parameters specifiedin the equation command were set equal between the two groups.As a result, the covariances among the observed variables werethe only parameters that were free to vary (Schumacker & Lomax,2004). The hypothesis of the second analysis was that both sets ofdata had similar path coefficients in the path model.
Results
The means, standard deviations, skewness, and kurtosis statisticsof the variables are shown in Table 1. There was a wide rangeof variability across the variables, suggesting that the data werefree from a range-restriction problem. The univariate skew-ness and kurtosis were within ±1, except for speaking skewness(−1.1), suggesting that the univariate distributions were normally
Note. All correlations are significant at the 0.01 level.
distributed. Multivariate normality was also tested, and theresults suggested the assumption of a multivariate normalitywas not sustained. However, a scatter-plot inspection indicatedno curvilinearity or heteroscedasticity. Concerning the slightdeviation from multivariate normality, two separate analyseswere performed using square-root-transformed scores and rawscores. The comparison between the results of transformed- anduntransformed-score analyses showed no evident difference.Hence, original untransformed scores were retained, and SEMwas estimated using maximum likelihood techniques instead ofperforming linear transformation.
An important issue in CFA is the reliability of a set of itemsthat defines a construct or factor (Schumacker & Lomax, 2004).The Cronbach α internal consistency coefficient, which tests theextent to which multiple indicators for a latent variable belongtogether, was 0.835. Since the observed variables are continuousscores and not scale scores, Raykov’s reliability rho was not ob-tained.
Bivariate correlation coefficients are presented in Table 2.There were significant, positive correlations between the vari-ables, ranging from r = 0.40 to r = 0.77. The highest correlationwas found between listening/speaking and listening/writing variables
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
240 H. K. Pae and D. Greenberg
(r = 0.77), followed by listening/writing and reading/writing skills(r = 0.73). The speaking and reading/speaking indicators showedthe lowest correlations (r = 0.40).
SEM data analyses were conducted in two steps. The first stepwas to perform CFA to test whether the constructs under consid-eration were adequately measured in a model. The second stepwas to evaluate for relationships among the latent variables in astructural model. Prior to testing multiple-group models, overall-group analyses were performed to identify a baseline model bytesting four competing conceptual models. Both global-type om-nibus test of the fit of the entire model and the fit of individ-ual parameters of the model were examined to check a modelfit. For the global test, the aforementioned fit indices were evalu-ated for the comparison of the model-implied covariance matrix� to the sample covariance matrix S. For the individual parame-ters of the model, the parameter estimates were evaluated with re-spect to the expected direction, statistically different critical valuefrom zero, and parameter estimates within an expected range ofvalues.
Model A. Correlated Two-Factor Model
As a necessary condition for model identification (Schumacker& Lomax, 2004), the order condition was assessed by comparingthe number of free parameters and the number of distinct valuesin the matrix S. There were nine factor loadings, nine measure-ment error variances, seven measurement error covariances, andone correlation among the latent variables, resulting in 26 freeparameters to be estimated. The number of distinct values in thematrix S was equal to 45 [(p (p + 1)/2, where p is the number ofobserved variables]. According to the order condition, the modelwas over identified, meaning that the degrees of freedom werepositive.
Since previous research has documented that receptiveskills and expressive skills are critical factors in English com-prehension in L1 (Bates, Bretheron, & Snyder, 1988) and L2(Barnett et al., 2007; Lugo-Neris, Jackson, & Goldstein, 2010),a two-factor model (i.e. receptive and expressive latent con-structs) was tested using a CFA technique. Receptive skills were
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 241
indicated by three observed variables, and the expressive skill fac-tor was indicated by six indicators, as seen in Figure 1. More-over, an inter-factor correlation was specified. The initial modeldid not fit well [χ2 (26, N = 585) = 216.95, p = 0.000, CFI =0.97, NFI = 0.97, RMSEA = 0.11]. For model modification,the modification indices (MI), which was related to the La-grange Multiplier test (Schumacker & Lomax, 2004), and thestandardized residual matrix were used. The MI suggested thaterror covariances be added to observed variables. Through amodel modification by adding seven error-covariance correlations(Listening/Speaking–Speaking; Reading/Writing–Reading; Rea-ding/Speaking–Listening/Speaking; Reading/Speaking–Listen-ing/Writing; Reading–Listening; Reading/Speaking–Speaking;Reading/Writing–Listening), the two-factor model became ten-able [χ2 (19, N = 585) = 30.21, p > 0.05, CFI = 1.00, RMSEA =0.032]. This indicated that the sample covariance matrix S was suf-ficiently reproduced by this theoretical model. Despite possiblecriticism on having error terms covary, it was felt that the additionof the error covariance to the model would be better than addingnew paths to the model because an addition of new paths to themodel would create different conceptual and theoretical models.Moreover, the addition of error covariances did not deviate fromthe theoretical relevance and consideration.
As a second criterion for model fit, the statistical significanceof individual parameter estimates for the paths in the model wasexamined. The critical values computed by dividing the parame-ter estimates by their respective standard errors were significantlydifferent from zero. Table 3 displays the parameter estimates, stan-dard error (SE), t values, and R2 for each of the paths in the two-factor model. The magnitude and the direction of the parameterestimates were within the expected range.
Model B. Single Factor Model
The second model hypothesized that a single first-order trait wasindicated by all observed variables under consideration in whichthe unitary trait hypothesis (Oller, 1979) was tested. The concep-tual model did not support the hypothesis. Through the modelmodification procedure using 10 correlations of error covarian-ces (Listening/Speaking–Speaking; Reading/Writing–Reading;
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
242 H. K. Pae and D. Greenberg
TABLE 3 The Parameter Estimates, Standard Errors (SE), and t Values forEach of the Paths in the Two-Factor Model
Reading/Speaking–Listening/Speaking; Reading/Writing–Liste-ning/Speaking; Reading/Speaking–Listening/Writing; Reading–Listening; Listening/Writing–Reading; Reading–Writing; Read-ing/Writing–Listening; Listening/Speaking–Reading), however,the single-trait model became plausible [χ2 (17, N = 585) =24.03, p = 0.12, CFI = 1.00, RMSEA = 0.027].
Model C. Second-Order Factor Model
The third model tested was a second-order factor model inwhich the receptive and expressive latent traits were explainedby the overall L2-skill construct. This method was problematicand failed model identification, having a convergence prob-lem. The Phi matrix was non-positive definite. Consequently,latent variable variances were constrained to 1.0, and read-ing and speaking error variances were set equal to 1 so as toprevent negative error variance. The hypothesized second-order model produced a misfit model [χ2 (24, N = 585) =445.69, p = 0.000, CFI = 0.94, RMSEA = 0.147], indicatingthat this model deviated from the implied theoretical model.Further steps for model specification and modification bysetting admissibility check off to obtain significant parameterestimates were not taken because the purpose of the competingmodel testing was to identify the best-fit model. It is consistentwith Shin’s (2005) finding that a second-order factor model didnot support his hypothesized structure of L2 proficiency.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 243
TABLE 4 Summary of Fit Indices for the Four CFA Models
Models df χ2 χ2/df GFI CFI RMSEA ECVI
Model A 19 30.21 1.59 0.99 1.00 0.032 0.14Model B 17 24.03 1.41 0.99 1.00 0.027 0.14Model C 24 445.69 18.57 0.88 .94 0.147 0.63Model D 16 22.43 1.40 0.99 1.00 0.026 0.14
Note. Model A = Correlated Two-Factor Model; Model B = Single Factor Model; ModelC = Second-Order Factor Model; Model D = Independent Two-Factor Model.GFI = Goodness of Fit Index; CFI = Comparative Fit Index; RMSEA = Root Mean SquareError of Approximation; ECVI = Expected Cross-Validation Index; Model AIC = ModelAkaike information criterion.
Model D. Uncorrelated Two-Factor Model
An uncorrelated two-factor model was tested in which two la-tent constructs were not to covary each other. This model wasevaluated based on a theoretical basis that L2 proficiency iscomposed of a set of independent dimensions (Sasaki, 1996).The hypothesis of this model was that (a) the discrete-skillfactor was associated with the four observed variables (listening,speaking, reading, and writing) and (b) shared skills were indi-cated by five combined skills (listening/speaking, listening/reading,listening/writing, reading/writing, and reading/speaking). The finaluncorrelated two-factor model with eight correlations of errorcovariances (Listening/Speaking–Speaking; Reading/Writing–Reading; Reading/Speaking–Listening/Speaking; Listening/Speaking–Reading; Reading/Speaking–Listening/Writing; Read-ing–Listening; Listening/Writing–Reading; Listening–Reading/Writing) was a good fit [χ2 (18, N = 585) = 25.91, p = 0.10,CFI = 1.00, RMSEA = 0.027].
Baseline Model Selection
Table 4 exhibits the fit indices of the four models, includingthe chi-square statistic, χ2/df , GFI, CFI, RMSEA, ECVI, andmodel Akaike information criterion (AIC). The ECVI was com-puted in LISREL as ECVI = (c/n) + 2(p/n), where c is thechi-square value for the overall fitted model, p is the numberof independent parameters estimated, and n = N –1 (sample
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
244 H. K. Pae and D. Greenberg
size) (Schumacher & Loma, 2004). The ECVI for the sam-ple was 0.14, indicating a measurement model was expectedto cross-validate, with 0.12 to 0.17 as the 90% confidenceinterval for ECVI. Parsimony was achieved, as indicated bythe parsimony normed fit index (PNFI = 0.53), which sought theminimum number of estimated coefficients required to achieve aspecific level of model fit (Schumacker & Lomax, 2004). Mod-els A, B, and D show lower ECVI than Model B, indicatingthe former three models fit the data better than Model B. Ex-cept the second-order model, Heywood cases, multicollinearity,or a non-positive definite matrix were not encountered in modelestimation. Nor were any convergence problems or inadmissi-ble solutions encountered in the three valid model tests exceptModel C. The GFI for Model C was 0.876, whereas the otherthree models showed 0.99, meaning that 99% of the sample vari-ances and covariances were sufficiently reproduced by the modelestimates.
In order to select a best-fit baseline model for a multiple-group model comparison, the fit indices of the three competingmodels, except the second-order factor model, which had a con-vergence problem, were evaluated with respect to model parsi-mony and substantive meaningfulness of the theoretical model.The fit indices of Models A, B, and D were strikingly similaracross the models. Of Models A and B, Model A was chosen asthe baseline model for three reasons. First, Model A’s correlatedtwo-factor model was consistent with both theoretical and empir-ical bases that linguistic abilities were complex and multifaceted,which was congruent with the findings of previous studies (Bach-man & Palmer, 1981, 1982; Sawaki et al., 2009; Shin, 2005). Sec-ondly, Model A fulfilled the parsimonious principle because ittook the smallest number of error-covariance correlations for thetenable model (7 vs. 10 and 8 for the single factor model andthe independent two-factor model, respectively). Double loadingswere not allowed because they would generate different models.Although Models B and D were not chosen, their implications arediscussed in the discussion section. Lastly, because this study didnot run the model separately for the discrete-skill indicators andcombined-skill indicators, it was still unclear whether or not thecombined skills played a role in the estimates of the single-factormodel.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 245
TABLE 5 Standardized Parameter Estimates for Factor Loadings
Note. All loadings are significant (|t| > 1.96).SMC = Squared multiple correlation.
Table 5 displays the baseline model’s standardized param-eter estimates for factor loadings as well as squared multiplecorrelation coefficients for each observed variable by the twolatent constructs. The squared multiple correlation coefficientsindicated that the observed variables served well as the measuresof the latent variables. All standardized factor loadings werestatistically significant for both receptive-skill (values ranged from0.71 to 0.79) and expressive-skill factors (values ranged from0.56 to 0.94). The inter-factor correlation was also large andstatistically significant (r = 0.98, p < 0.001).
Multiple-Group Model
The group was broken down into two male and female subgroupsas well as high- and poor-performer subsamples. The general pro-cedure was to test for measurement invariance between the un-constrained models for all groups combined, then for a modelwhere certain parameters are constrained to be equal between thegroups. The measurement model was used to assess whether ornot the factor loadings of the measurement model were invariantacross the two gender and ability subgroups. These assessmentsconsisted of testing the null hypothesis that the factor loadingswere identical across gender and ability subsamples against thealternative hypothesis that the factor loadings were not identicalacross gender and ability subgroups.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
246 H. K. Pae and D. Greenberg
First, the gender subsamples were tested with the basemodel.3 The results indicated different parameter estimates forthe female and male data when applied to the path model. Thefemale and male data fit the path model as indicated by the non-significant chi-square values (χ2 = 47.16, χ2 = 19.65, ps > 0.05,respectively). The global fit statistics indicated a chi-square forthe hypothesis of unequal (separate) parameter estimates to bea good fit to the path model (χ2 = 66.81, df = 48, p > 0.05).
Since testing a hypothesis as to whether the two groupsshared a common path model was of interest, a subsequent SEMwas run by setting the parameters specified in the equations’ com-mand to be equal between the two groups. In this analysis, the co-variance among the observed variables was free to vary by placingconstraints on the others. The parameter estimates were the samein both groups. The individual chi-square values for each groupalso summed up to the global chi-square statistic for this commonmodel. The chi-squares for the female and male groups were χ2 =43.09 and χ2 = 27.97, ps > 0.05, respectively, which yielded theglobal chi-square value of 71.06, df = 61, p > 0.05. These resultsindicated that the two sets of data fit the path model based on thehypothesis of similar path coefficients in the baseline path model.
Computed next was a chi-square difference test (a.k.a. a like-lihood ratio test) between the two path model analyses. A chi-square difference between the unequal parameter estimates andthe equal parameter estimates was calculated using an EXCELspreadsheet chi-square difference calculator program provided byLISREL 8.80 (Joreskog & Sorbom, 2007). The difference in thechi-square values and associated p-values was obtained using theglobal chi-square value from the analysis of equal parameter es-timates and the global chi-square value from the analysis of un-equal parameter estimates. There was no significant chi-squaredifference between the two model analyses (χ2 = 4.25, df = 13).This implies that the female and male data separately fit the pathmodel as well as both datasets fit a common path model. Table 6displays the results of the multiple-group CFA models.
Figure 3 exhibits standardized parameter estimates for thegender subgroups. The top values are for the female group, whilethe bottom values are for the male group.
The group equivalence of models for paths and latent vari-ables for the two ability subgroups (i.e., high- and low-ability
groups) were compared as it was done with the gender subgroups.The same model was first applied to the pooled covariance ma-trices of the two groups in which parameter estimates were aver-aged across the two groups. Next, the imposed constraints wereplaced so that parameters could be identical across groups, andthen compared to see whether the constrained models with equalparameters fit the data.
The chi-square values of unequal parameter estimates forthe low- and high-ability groups were χ2 = 76.52 and χ2 = 98.23,p < 0.001, respectively, which yielded the global chi-square valueof 174.75, p = 48, p < 0.001. These results indicated that the twosets of data did not fit the path model based on the hypothesisof similar path coefficients in the baseline path model. With con-straints imposed for equal parameter estimates, the low- and high-ability groups’ data did not fit the path model as indicated by thesignificant chi-square values (χ2 = 91.05, χ2 = 146.86, p < 0.001,respectively). The global fit statistics indicated a chi-square for thehypothesis of equal parameter estimates to be misfit to the pathmodel (χ2 = 237.91, df = 62, p < 0.001).
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 249
The results of the chi-square difference indicated a differ-ence between the two model analyses (χ2 = 63.16, df = 14,p > 0.05). The large p value (p = 3.251) suggests that there wassufficient evidence that the cross validation of the measurementmodel for the nine observed variables across the ability subsam-ples was not supported by the data. Since the ability multigroupCFA produced a misfit model, no further reports are provided.
Discussion
Structural Model for Multitrait Skills: Receptive and Expressive Skills
In order to determine which set of observed variables shared com-mon variance-covariance characteristics which defined theoreticalconstructs, this study had two goals. The first goal was to find theunderlying covariance structure for the observed indicators whichfit best for the model, and the second was to examine whether thegrouping variables had any influence on the CFA results of the ob-served variables.
The relationship of adult NNSs’ receptive and expressiveskills was examined using the four discrete-skill item types(i.e. listening, speaking, reading, and writing) and five shared-skill item types (i.e. listening/speaking, listening/reading, listen-ing/writing, reading/writing, and reading/speaking). These nine ob-served indicators mapped well into the pre-established two latentconstructs—receptive skills and expressive skills. Of the four apriori-specified competing theoretical models, a correlated two-factor model was adopted as a baseline model, because (a) it wasconsistent with more recent and stronger theoretical and empir-ical considerations as indicated by a series of research evidenceof multicomponential subskills (Bachman & Palmer, 1982; Sasaki,1996; Shin, 2005), and (b) it was the most parsimonious modelamong the three competing candidates. The correlated two-factormodel suggested that the two latent constructs, receptive and ex-pressive skills, accounted for adult NNSs’ observed test scores. It isimportant to note that a single factor model and an independenttwo-factor model were also eligible candidates for the baselinemodel given the tenable fit indices. However, it is still unknownwhether the indicators of combined skills (i.e. listening/speaking,listening/reading, etc.) masked the factor structure in the single
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
250 H. K. Pae and D. Greenberg
factor model because the indicators of both discrete skills andcombined skills were entered in the model. A subsequent studyis needed to address this issue. Oller’s (1979) unitary trait hypoth-esis has fallen out of favor and is largely unsupported by empiri-cal evidence over time. However, the results of this study suggestthat adults’ receptive and expressive skills of academic English areclosely interconnected and converge on general linguistic profi-ciency. The four rudimentary linguistic skills, such as listening,speaking, reading, and writing, seem to be characterized by thediscrete skills independently from shared skills in which recep-tive and expressive skills are integrated (i.e. listening/speaking,listening/reading, listening/writing, reading/writing, and read-ing/speaking). However, the independent two-factor model wasless convincing than the correlated two-factor model because lin-guistic components skills are interconnected.
The indicators which were loaded on the receptive factorshowed more homogeneous factor loadings in magnitude thanthose loaded on the expressive construct. This indicates that dif-ferent indicators of receptive skills function in a similar way com-pared to those of the expressive-skill construct. When it comesto expressive skills, the range of factor loadings produced by theindicators was wide, and the speaking indicator showed the low-est standardized factor loading, followed by the reading/speakingvariable.
The correlated two-factor model was congruent with the mul-ticomponential model of language skills documented in previousresearch (Sawaki et al., 2009; Shin, 2005) in that the observedvariables were explained by multiple latent constructs. Unlike thefindings of previous research (Sawaki et al., 2009; Shin, 2005),however, the data did not support a second-order factor model. Itseems that the second-order factor has a relatively weak influenceon the receptive and expressive skills characterized by the NNSs’performance scores.
Contrary to previous research findings that expressive lan-guage skills were more predictive of reading skills (Bishop &Adams, 1990; Chiappe et al., 2004), the magnitudes of the vari-ances explained by the expressive skills were smaller and theranges were wider than those of the receptive skills. The differ-ence between the previous findings which used children’s L1 andL2 data and the findings of this study with adults’ L2 data casts
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 251
important implications in the field of SLA because it provides ev-idence that adult NNSs’ performance on academic English is dif-ferent from that of children’s L1 and L2.
The greater variance explained by the receptive skills, in com-parison to previous studies (Chiappe et al., 2004; Metsala, 1997),can be explained in several ways. First, one explanation has to dowith the different learning trajectory between children and adults.As stated earlier, children’s L1 and L2 acquisition and adults’ L2learning have different learning trajectories. Children’s languageacquisition does not typically begin with sophisticated and analyt-ical knowledge about print until literacy instruction takes place.However, adults’ L2 learning involves multiple linguistic modali-ties at the same time, including oral and written modes as well asreceptive and productive forms. Secondly, it might have stemmedfrom speaking characteristics of L2 because L2 speaking is influ-enced by both linguistic and affective factors. Although this studydid not take affective variables into account, it is possible thatthe performance of speaking by the adult participants was sus-ceptible to affective factors given its salient effect on L2 learning(Gardner, 1985). As Song (2008) notes, it is possible that inter-national adult NNSs demonstrate a mismatch between receptiveand expressive language skills, typically showing higher written-language proficiency than spoken-language skills. Next, it mighthave to do with the characteristics of academic language skills.Academic English involves cognitively demanding tasks that re-quire higher-order problem-solving skills and critical reasoningand use low-frequency vocabulary. According to Pearson (2009),the item contents of PTE Academic were drawn from real-life ex-amples of authentic lectures and textbook contents used in anacademic setting. There might be considerable overlaps be-tween academic content and task demands in CALP. Finally, itmight have to do with the modality difference. Receptive skillsand expressive skills may be represented differently because ofdifferent processing routes in which primary language inputprocessing becomes intake (i.e., retained input) for linguistic com-petence repository and finally for the efficient mechanism ofproduction (VanPatten, 2004). There might be greater linguisticdependency among listening , reading , and listening/reading tasks,as these skills entail the cognitive processes which focus on inter-nal recovery of conveyed messages in oral and written form. In
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
252 H. K. Pae and D. Greenberg
contrast, the expressive modality is active and productive in na-ture. As a result, the participants’ academic English skills mighthave resulted in disparities between receptive and expressive lan-guage skills. A higher variability seems to exist in the participants’expressive functioning because production requires fluent worduse and meaning that match the intended message.
The findings of this study suggest that the viable theoreti-cal basis lies in the closely interconnected networks in which L2receptive and expressive subskills are not constrained to one di-rection but bi-directional interactions, regardless of the process-ing modality (i.e., receptive vs. expressive) or perceptual modality(i.e., visual vs. auditory or print vs. speech). However, this struc-ture does not predict symmetry with respect to how receptive skillsinteract with expressive abilities and how these skills interact withoverall L2 performance. Since the path structure among the vari-ables does not specify the locus of the causal effect, the asymmet-ric associations of the receptive and expressive skills to the indica-tors need to be further investigated.
Multiple-Group SEM
After the baseline model was confirmed for the whole group, themodel was examined across subsamples to determine the degreeof invariance in fit indices, parameter estimates, and standarderrors. Multiple-group analyses in covariance-based SEM provideuseful information about invariance across subgroups. To addressthis, a multiple-group model, which relied on separate covariancematrices for each group of gender and ability subgroups, was em-ployed to test the theoretical model with two subgroups of datasimultaneously. Factor structures between the subgroups werecompared using separate covariance matrices for each group butestimating each group’s parameters simultaneously. The baselinemodel first determined the extent to which the measurement andSEMs fit across the gender subgroups and then examined whethermodel differences exist between the two subgroups.
With respect to the gender-group effect, both unequal andequal parameter estimates in the path baseline model were equiv-alent across the two gender subgroups, indicating an invariantpattern of factor loadings for the group. The chi-square differ-ence test showed that gender subgroups did not differ in the fac-
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 253
tor structure in relation to the receptive and expressive latentconstructs, indicating that the gender-grouping variable did nothave an influence on the factor loadings for the observed vari-ables of the data. This compatibility of the data with the modelconfirms that the subskill factor structure of receptive and expres-sive linguistic components are similar across gender. This findingis important in the field of adult NNSs’ SLA. Previous researchhas shown a gender difference in computer-based assessments(Gallagher et al., 2002) and L2 comprehension and vocabularylearning in a video-based computer-assisted language learningprogram (Lin, 2011), and task performance in tape-mediated as-sessment of speaking (Lumley & O’Sullivan, 2005). When it comesto NNSs’ academic English skills, however, the latent constructs ofmale and female achievement seem to be comparable.
Although the models resulted in approximately equal factorcorrelations for gender, the two models differed with respect tomultiple-group ability analyses investigating the invariance of fac-tor structure. There was sufficient evidence that factor structurefor the high-proficient and low-proficient groups was different.Therefore, a decreasing or increasing differentiation effect as a re-sult of the proficiency level could not be specified, as Shin (2005)noted. This finding suggests that there might be a different latentconstruct for different ability groups; that is, the factor structurefor a highly functioning group is different from that for a poorlyperforming counterpart. Hence, caution needs to be paid in theinterpretation of test scores according to the skill level.
This finding deviates from the results of Shin’s (2005)study, which examined the factor structure of three proficiencylevels. One explanation has to do with the difference of groupnumbers. Shin (2005) had three proficiency groups, includinglow-, intermediate-, and advanced-level groups, while this studyhad two groups (i.e., low- and high-ability groups). It was stillunclear whether the arbitrary nature of group assignment yieldeda spurious effect on the results. However, the possibility of adifferent factor structure across low- and high-ability groups isconsistent with the finding of previous research. Song (2008)noted that the divisible subskills in L2 skills are related to testtakers’ L2 proficiency level. The influence of asymmetry inproficiency levels needs to be further investigated.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
254 H. K. Pae and D. Greenberg
Overall, the findings of this study provide additional supportfor the view that L2 language skills are multicomponential. Therepresentation of L2 linguistic ability encompasses a receptive de-scription of L2 as well as information about expressive proficiency.It should be noted that although the one-factor and independenttwo-factor models were candidates for the base model, the dis-crimination between the discrete-skill and combined-skill indica-tors was not made because it was beyond the scope of this study.Hence, a subsequent study to address this issue is needed. Thisstudy did not show evidence of distortions of the factor struc-ture, biased parameter estimates, or possible differential effectsof the dimensions on other variables in the model. The principaltheoretical issue revolves around the factor structure of receptiveand expressive skills characterized by adult NNSs’ academic En-glish performance. Despite the possibilities of randomness of thedata, it seems that the structure of latent constructs accounts forthe L2 performance of different ability groups (i.e., high- vs. low-performing individuals).
Although this study contributes to the body of knowledge re-garding the relationship and the factor structure of receptive andexpressive skills, limitations, which are also related to future direc-tions, should be noted. First, the adult NNSs’ L1 proficiency datawere not available in the PTE Academic database. L1 characteris-tics and L1 ability level might have an influence on the adult learn-ers’ L2 performance, but no associations can be noted due to theunavailability of L1 data. Further studies addressing L1 skill levelsof test takers are recommended. Secondly, in relation to the firstpoint, all the participants were combined to form a subject poolregardless of their L1 differences. It is possible that different L1groups demonstrate different problem-solving strategies and lan-guage profiles. Although the sample homogeneity was achieved(Pae, 2012), the characteristics of the language family of the test-takers’ L1s might have affected L2 academic English performancedifferently because some languages share commonalities with En-glish in terms of vocabulary, phonology, orthography, and seman-tics. Therefore, the unavailability of L1 data imposes limitationsin the interpretation of L1 effects on L2 English learning. Furtherresearch is warranted to investigate an L1 influence on L2 attain-ment as well as its possibility of masking the true factor structure
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 255
of the data. Thirdly, the data utilized was cross-sectional, whichallows for only one-time evaluation and interpretation. Longitu-dinal data will allow for repeated observations of the same itemtypes over time and make more accurate observations of changesin time. Fourthly, the interpretation of the results is limited be-cause factor structures do not reflect perfect simple structure andstructural consistency. Lastly, a practical implication is related tothe use of test scores. As Alderson (2000) pointed out, there is adoubt as to whether a single score successfully measures intendedsub-skills. Test scores are mere indicators of individuals’ under-lying abilities, and the indicators are based on the number ofcorrect responses to test items. This raises a question of whetherNNSs use the specific skills or processes that the test intends tomeasure when they answer test questions and whether NNSs fullyutilize their latent skills and expectations about the test based ontheir L2 competence. A “deep description” achieved through aqualitative study can address this question in a richer and moreefficient way than a quantitative study.
The theoretical implication has to do with the identificationof underpinning factor structures of receptive and expressive L2skills demonstrated by adult NNSs across gender and ability sub-groups. Since the correlated two-factor model was tenable, the re-sults enlighten pedagogical implications with respect to the inter-related constructs of input and output modalities. The findingsof this study provide evidence for establishing the foundation foreffective instructional programs that take the indispensable twoskills (i.e., receptive and expressive skills) into account in order topromote adult NNSs’ English skills.
Funding
This work was supported by Pearson Language Test Academic[grant number: Pearson2010-A-001] and the University ResearchCouncil of the University of Cincinnati.
Notes
1. Although English may not necessarily be an L2 for all the participants, an L2was used instead of a foreign language for the sake of consistency with theliterature and convenience throughout this article.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
256 H. K. Pae and D. Greenberg
2. The 20 items types include a wide range of question types as follows: Read-ing Multiple choice, choose single answer; Reading Multiple choice, choosemultiple answers; Re-order paragraphs; Reading: Fill in the blanks; Highlightcorrect summary; Read aloud; Summarize written text; Listening Multiplechoice, choose single answer; Listening Multiple choice, choose multiple an-swers; Listening: Fill in the blanks; Highlight incorrect words; Select missingwords; Write from dictation; Summarize spoken text; Repeat sentence; Writeessay; Reading & writing: Fill in the blanks; Describe image; Re-tell lecture;and Answer short question. For more details on the test types, see Zheng &De Jong (2011).
3. In response to a reviewer’s suggestion, we ran the same analyses with the othertwo models (i.e., single-factor model and independent two-factor model). Theresults were similar to those of the base model. Hence, the results with thebase model are reported here.
References
Adams, M. J. (1990). Beginning to read: Thinking and learning about print. Boston,MA: MIT Press.
Alderson, C. (1991). Language testing in the 1990s: How far have we come?How much further have we to go? In S. Anivan (Ed.), Current developments inlanguage testing (pp. 1–26). Singapore: SEAMEO Regional Language Centre.
Alderson, J. C. (2000). Assessing reading. Cambridge, UK: Cambridge UniversityPress.
Alderson, J. C., & Lukmani, Y. (1989). Cognition and reading: Cognitive levels asembodied in test questions. Reading in a Foreign Language, 5, 253–270.
Bachman, L. F., & Palmer, A. S. (1981). The construct validation of the FSI oralinterview. Language Testing, 31, 67–86.
Bachman, L. F., & Palmer, A. S. (1982). The construct validation of some com-ponents of communicative proficiency. TESOL Quarterly, 16, 449–465.
Bachman, L. F., Davidson, F. G., & Foulkes, J. (1990). A comparison of the abil-ities measured by the Cambridge and Educational Testing Service EFL TestBatteries. Issues in Applied Linguistics, 1, 30–55.
Bae, J., & Bachman, L. F. (1998). A latent variable approach to listeningand reading: Testing factorial invariance across two groups of children inthe Korean/English two-way immersion program. Language Testing, 15, 380–414.
Bandalos, D. L., & Finney, S. J. (2001). Item parceling issues in structural equa-tion modeling. In G. A. Marcoulides & R. E. Schumacker (Eds.), New develop-ments and techniques in structural equation modeling (pp. 269–296). Mahwah, NJ:Lawrence Erlbaum.
Barnett, W. S., Yarosz, D. J., Thomas, J. H., Jung, K., & Blanco, D. (2007). Two-waymonolingual English immersion in preschool education: An experimentalcomparison. Early Childhood Research Quarterly, 22, 277–293.
Bates, E., Bretheron, I., & Snyder, I. (1988). From first words to grammar. New York,NY: Academic Press.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 257
Bishop, D., & Adams, C. (1990). A prospective study of the relationship betweenspecific language impairment, phonological disorders, and reading retarda-tion. Journal of Child Psychology and Psychiatry, 21, 1027–1050.
Buck, G. (1992). Listening comprehension: Construct validity and trait charac-teristics. Language Learning, 42, 313–357.
Buck, G. (2001). Assessing listening. Cambridge, UK: Cambridge University Press.Carroll, J. B. (1993). Human cognitive abilities. Cambridge, UK: Cambridge Uni-
versity Press.Chiappe, P., Chiappe, D. L., & Gottardo, A. (2004). Vocabulary, context, and
speech perception among good and poor readers. Educational Psychology, 24,825–843.
Cummins, J. (2008). BICS and CALP: Empirical and theoretical status of the dis-tinction. In B. Street & N. H. Hornberger (Eds.), Encyclopedia of Language andEducation, 2nd Edition, Volume 2: Literacy (pp. 71–83). New York, NY: Springer.
Ellis, N. C. (2004). The processes of second language acquisition. In B. VanPat-ten, J. Williams, S. Rott, & M. Overstreet (Eds.), Form-meaning connections insecond language acquisition (pp. 49–76). Mahwah, NJ: Lawrence Erlbaum.
Enders, K. C., & Bandalos, L. D. (1999). The effects of heterogeneous item dis-tribution on reliability. Applied Measurement in Education, 12, 133–150.
Gallagher, A., Bridgeman, B., & Cahalan, C. (2002). The effect of computer-based tests on racial-ethnic and gender groups. Journal of Educational Measure-ment, 39, 133–147.
Gardner, R. (1985). Social psychology and second language learning: The role of atti-tudes and motivation. London, UK: Arnold.
Garson, G. D. (2011). Structural Equation Modeling. Retrieved from http://faculty.chass.ncsu.edu/garson/PA765/structur.htm
Grabe, W., & Stroller, F. L. (2002). Teaching and researching reading. Harlow, UK:Longman.
Joreskog, K. G., & Sorbom, D. (2007). LISREL 8.80. Hillsdale, NJ: Scientific Soft-ware International.
Kishton, J. M., & Widaman, K. F. (1994). Unidimensional versus domain repre-sentative parceling of questionnaire items: An empirical example. Educationaland Psychological Measurement, 54, 757–765.
Kline, R. B. (1998). Principles and practice of structural equation modeling. New York,NY: Guilford.
Kunnan, A. J. (1995). Test taker characteristics and test performance: A structural equa-tion modeling approach. Cambridge, UK: Cambridge University Press.
Lennon, R. T. (1962). What can be measured? The Reading Teacher, 15, 326–337.Lin, L.-F. (2011). Gender differences in L2 comprehension and vocabulary learn-
ing in the video-based CALL program. Journal of Language Teaching and Re-search, 2, 295–301.
Lugo-Neris, M. J., Jackson, C. W., & Goldstein, H. (2010). Facilitating vocabularyacquisition of young English language learners. Language, Speech, and HearingServices in Schools, 41, 314–327.
Lumley, T., & O’Sullivan, B. (2005). The effect of test-taker, audience and topicon task performance in tape-mediated assessment of speaking. Language Test-ing, 22, 415–437.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
258 H. K. Pae and D. Greenberg
Lund, R. J. (1991). A comparison of second language listening and reading com-prehension. Modern Language Journal, 75, 196–204.
Metsala, J. L. (1997). Spoken word recognition in reading disabled children.Journal of Educational Psychology, 89, 159–169.
Oller, J. (1979). Language tests at school: A pragmatic approach. London, UK: Long-man.
Oltman, P. K., Stricker, L. J. & Barrows, T. S. (1988). Native language, Englishproficiency, and the structure of the Test of English as a Foreign Language.TOEFL Research Report Series. Princeton, NJ: Educational Testing Service.
Pae, H. K. (2012). A psychometric measurement model for adult English lan-guage learners: Pearson Test of English Academic. Educational Research andEvaluation, 18, 211–229.
Pae, H. K., Greenberg, D., & Morris, R. D. (2012). Construct validity and mea-surement invariance of the PPVT in the performance of struggling adult read-ers. Language Assessment Quarterly, 9, 152–171.
Pearson Test of English Academic (2009). Research. Retrieved from http://pearsonpte.com/research/Pages/ResearchSummaries.aspx.
Qingquan, N., Chatupote, M., & Teo, A. (2008). A deep look into learning strat-egy use by successful and unsuccessful students in the Chinese EFL learningcontext. REK Journal, 39, 338–358.
Reves, T., & Levine, A. (1988). The FL receptive skills: Same or different? System,16, 327–336.
Rost, D. (1993). Assessing the difference components of reading comprehen-sion: Fact or fiction? Language Testing, 10, 79–92.
Ryan, K., & Bachman, L. F. (1992). Differential item functioning on two tests ofEFL proficiency. Language Testing, 9, 12–29.
Sasaki, M. (1996). Second language proficiency, foreign language aptitude, and intelli-gence: Quantitative and qualitative analyses. New York, NY: Peter Lang.
Sawaki, Y., Stricker, L. J., & Oranje, A. H. (2009). Factor structure of the TOEFLInternet-based test. Language Testing, 26, 5–30.
Schumacker, R. E., & Lomax, R. G. (2004). A beginner’s guide to structural equationmodeling (2nd ed.). New York, NY: Taylor & Francis.
Shin, S.-K. (2005). Did they take the same test? Examinee language proficiencyand the structure of language tests. Language Testing, 22, 31–57.
Song, M.-Y. (2008). Do divisible subskills exist in second language (L2) com-prehension? A structural equation modeling approach. Language Testing, 25,435–464.
Stricker, L. J., Rock, D. A., & Lee, Y.-W. (2005). Factor structure of the LanguEdgetest across language groups (TOEFL Monograph Series MS-32). Princeton, NJ:Educational Testing Service.
Swinton, S. S., & Powers, D. E. (1980). Factor analysis of the Test of English as aForeign Language for several language groups (TOEFL Research Report No. 6).Princeton, NJ: Educational Testing Service.
VanPatten, B. (2004). Input and output in establishing form-meaning connec-tions. In B. VanPatten, J. Williams, S. Rott, & M. Overstreet (Eds.), Form-meaning connections in second language acquisition (pp. 29–47). Mahwah, NJ:Lawrence Erlbaum.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
ewca
stle
(A
ustr
alia
)] a
t 18:
16 2
6 A
ugus
t 201
4
Receptive and Expressive Subskills 259
VanPatten, B., Williams, J., Rott, S. & Overstreet, M. (Eds.). (2004). Form-meaningconnections in second language acquisition. Mahwah, NJ: Lawrence Erlbaum.
Webb, S. (2008). Receptive and productive vocabulary sizes of L2 learners. Studiesin Second Language Acquisition, 30, 79–95
West, S. G., Finch, J. F., & Curran, P. J. (1995). Structural equation models withnonnormal variables: Problems and remedies. In R. H. Hoyle (Ed.), Struc-tural equation modeling: Concepts, issues, and applications (pp. 56–75). ThousandOaks, CA: Sage.
Yuan, K.-H., Bentler, P. M., & Kano, Y. (1997). On averaging variables in a confir-matory factor analysis model. Behaviormetrika, 24, 71–83.
Zheng, Y., & De Jong, J. H. A. L. (2011). Establishing construct and con-current validity of Pearson Test of English Academic. Retrieved from http://pearsonpte.com/research/Documents/PTEAcademicValidityPaper.pdf.
