You are on page 1of 15

DOI: 10.1111/cdev.

13884

EMPIRICAL ARTICLE

Subtypes of mathematical learning disability and their antecedents:


A cognitive diagnostic approach

Xiangzi Ouyang1 | Xiao Zhang1 | Pekka Räsänen2 | Tuire Koponen3 |


Marja-­Kristiina Lerkkanen3

1
Faculty of Education, The University of Abstract
Hong Kong, Hong Kong, China
2
Using cognitive diagnostic modeling (CDM), this study identified subtypes of
Faculty of Science, University of Turku,
Turku, Finland
mathematics learning disability (MLD) based on children's numerical skills and
3
Faculty of Psychology and Education,
examined the language and spatial precursors of these subtypes. Participants were
University of Jyväskylä, Jyväskylä, Finland 99 MLD children and 420 low achievers identified from 1839 Finnish children (966
boys) who were followed from preschool (age 6) to fourth grade (2007–­2011). Five
Correspondence
Xiao Zhang, Faculty of Education, The
subtypes were identified: the arithmetic fluency deficit only subtype, the counting
University of Hong Kong, Pokfulam Road, deficit subtype, the pervasive deficit subtype, the symbolic deficit subtype, and the
Hong Kong. counting and concept deficit subtype. Different subtypes depended on different
Email: xzhang1@hku.hk
constellations of language and spatial deficits. Findings highlight the effectiveness
Funding information of CDM in identifying MLD subtypes and underscore the importance of
Academy of Finland, Grant/Award understanding the specific deficits and antecedents of the subtypes.
Number: 213486, 268586 and 292466

The need for people to understand and use data is steadily Why are there MLD subtypes?
increasing with the development of modern society, yet
5%–­8% of students struggle with mathematics learning Before investigating MLD, it is necessary to determine
(Geary, 2004). These students are typically referred to the competencies that contribute to children's mathemat-
as students with mathematics learning disability (MLD), ics learning. Mathematics learning is a complex process
which is defined as significant mathematical under- involving a variety of domain-­specific numerical com-
achievement compared with the level at which individ- petencies (i.e., competencies that predict mathematical
uals are supposed to be based on their intelligence, age ability only, such as counting skills) and domain-­general
and education (American Psychiatric Association, 2013). cognitive-­linguistic competencies (i.e., competencies that
In view of the complexity of mathematics learning, MLD predict not only mathematics ability but also other abili-
students typically show significant heterogeneity in their ties, such as language skills; Desoete & Grégoire, 2006;
numerical deficits (Geary, 2004), which leads to different Geary, 1993, 2004; Zhang et al., 2020). Due to the com-
subtypes. Identifying subtypes of MLD can contribute plexity of mathematics learning, students with MLD
to a better understanding of the origins of MLD and fa- show heterogeneity in their deficits, leading to different
cilitate the design of specific intervention programs that subtypes. Geary's (2004) model of MLD provides a com-
benefit students with each subtype. Using cognitive di- prehensive framework and suggests that children's defi-
agnostic modeling (CDM), this study aimed to identify cits in mathematical learning are manifested as deficits
the subtypes of MLD and examine the language and spa- in domain-­specific numerical skills, which are further
tial skill precursors of these subtypes. affected by underlying cognitive-­linguistic antecedents

Abbreviations: ANS, approximate number system; CDM, cognitive diagnostic modeling; DINA, Deterministic-­Input, Noisy “And” Gate; DINO, Deterministic
Input, Noisy “Or” Gate; G-­DINA, Generalized Deterministic, Noisy “And” Gate; LA, low achievement; LR, likelihood ratio; MLD, mathematics learning
disability; PISA, Program for International Student Assessment; RAN, rapid automatized naming; RD, reading disability; RMSEA, root mean square error of
approximation.
This article is based on Chapter 3 of Xiangzi Ouyang's PhD dissertation under the supervision of Xiao Zhang.

© 2022 The Authors. Child Development © 2022 Society for Research in Child Development.

Child Development. 2022;00:1–15.  wileyonlinelibrary.com/journal/cdev | 1


2 |    OUYANG et al.

(e.g., central executive, language system, spatial system). to the ability to map symbolic numerals (e.g., Arabic
In contrast to previous studies that used both domain-­ numbers) onto the quantities they represent (Schneider
general and domain-­specific skills to identify MLD sub- et al., 2016). The access deficit hypothesis of MLD as-
types (e.g., Bartelet et al., 2014; Chan & Wong, 2020), the sumes that deficits in number-­numerosity mapping lead
present study, drawing on Geary's model, aimed to iden- to difficulties in mathematics learning (Rousselle &
tify MLD subtypes based on domain-­specific numerical Noël, 2007). Children with this deficit view the number
deficits and to investigate the domain-­general anteced- sequence as meaningless random symbols, and they are
ents of these subtypes. expected to perform poorly in mathematical tasks con-
taining symbolic numbers rather than non-­ symbolic
quantities (e.g., concrete objects). The access deficit
Early domain-­specific numerical deficits of hypothesis of MLD has received empirical support.
MLD subtypes Rousselle and Noël (2007) showed that MLD children
were slower in numerical tasks involving Arabic num-
Preschool is critical for children in building the founda- bers, such as comparing Arabic digits, but were no dif-
tion for future mathematics learning through acquiring ferent in non-­symbolic number comparisons to typically
informal number knowledge, and deficits in number achieving children. Salminen et al. (2018) also showed
knowledge may lead to later difficulties in mathematics that number-­numerosity mapping ability differentiated
learning (Geary, 2013). Purpura et al. (2013) proposed MLD children from children with low achievement (LA)
a theoretical model of informal number knowledge in mathematics. Consistent with the access deficit hy-
learned in preschool, which consisted of three numeri- pothesis, MLD subtypes with Arabic deficits have been
cal domains: numbering, relation, and arithmetic op- identified. For instance, von Aster (2000) identified an
erations. Numbering refers to children's ability to count Arabic subtype of MLD, which was manifested by poor
and determine the quantity of a series of objects; relation performance in number comparison and number iden-
is defined as children's understanding of the semantic tification. Bartelet et al. (2014) also identified an access
meanings of numbers and the relation between quantities deficit subtype of MLD, characterized by deficits in
and includes skills, such as number-­numerosity mapping Arabic number comparison and mapping between ver-
and set comparison; arithmetic operations are defined as bal number words and Arabic digits.
basic knowledge of the principles and concepts of addi- An understanding of basic arithmetic concepts (e.g.,
tion and subtraction (Jordan et al., 2006). understanding “more,” “less,” and “equal;” Siegel, 1971)
Previous studies have found that MLD children show is also an important numerical skill. A poor understand-
deficits in early numerical skills, such as counting skills ing of these concepts in preschool may contribute to later
(e.g., Desoete & Grégoire, 2006; Zhang et al., 2020), difficulties with arithmetic (Geary, 2004). Specifically,
number-­ numerosity mapping (e.g., Rousselle & such a poor understanding affects children's learning of
Noël, 2007; Salminen et al., 2018), and basic arithmetic more complicated arithmetic procedures and detection of
concepts (e.g., Desoete & Grégoire, 2006), each of which errors when performing arithmetic tasks (Geary, 2004).
corresponds to a domain of informal number knowledge In line with this notion, Desoete and Grégoire (2006)
as differentiated in Purpura et al.'s (2013) framework. found that approximately 67% of MLD students showed
Verbal counting, as a critical skill learned in preschool, deficits in understanding arithmetic concepts in pre-
provides the basis for linking verbal codes to their mean- school years. Notably, however, MLD subtypes with
ing (e.g., quantities) in long-­term memory and thus pre- conceptual knowledge deficits were rarely identified.
dicts the later development of arithmetic skills (Koponen
et al., 2019). Landerl et al. (2004) found that MLD chil-
dren were slow in counting objects, especially when the The linguistic and spatial correlates of
number of objects was larger than four. Similarly, Geary MLD subtypes
et al. (2000) showed that MLD children committed
more errors in counting and used less mature strategies In the literature, the antecedents underlying different
(e.g., using fingers) while counting than their non-­MLD MLD subtypes often vary. According to Geary's (2013)
counterparts. Such deficits lead to later difficulties in model, domain-­ specific numerical skills are affected
understanding cardinal concepts and remembering by underlying cognitive-­linguistic antecedents, such as
mathematical facts and retrieval errors (Geary, 1993). executive function, the language system and the spatial
MLD subtypes with counting deficits have been identi- system. Executive function typically serves as a general
fied in the literature. Specifically, von Aster (2000) iden- deficit underlying MLD (Geary, 2013), but language and
tified a verbal deficit subtype, characterized by deficits spatial skills may contribute to the differences between
in counting, using counting strategies when solving cal- MLD subtypes as mathematical information can be pro-
culation problems, and in storing mathematical facts. cessed by either the language system or the spatial system
Number-­numerosity mapping, as a crucial correlate of (von Aster, 2000). To illustrate, von Aster (2000) dem-
mathematical achievement (Purpura et al., 2013), refers onstrated that different MLD subtypes were related to
SUBTYPES OF MATHEMATICS DISABILITY    | 3

either left or right hemisphere dysfunction based on their working memory than typically developing children.
deficits in language or spatial skills. Specifically, MLD Murphy et al. (2007) documented that spatial-­perceptual
subtypes with deficits in retrieving arithmetic facts have ability differentiated LA children from their average
been found to be associated with language dysfunction, achieving peers but not from MLD children. Zhang
while subtypes with deficits in spatial representation of et al. (2020) found that MLD children had poorer spa-
numbers have been found to relate to dysfunction in spa- tial visualization skills than LA and average achieving
tial skills. Similarly, Rourke's (1993) study also supported children. However, few studies have examined whether
the differential roles of language and spatial skills in two spatial skills contribute to MLD subtypes. Geary (1993)
subtypes of MLD (a reading spelling subtype—­an MLD proposed that a spatial subtype of MLD, manifested by
subtype relating to language deficits with intact spatial deficits in numerical skills involving spatial information
skills, and a nonverbal subtype—­an MLD subtype relat- such as place value, was due to children's poor spatial
ing to visuo-­spatial deficits with intact language skills). skills.
Although it is important to examine whether various
cognitive-­ linguistic factors contribute differentially to
MLD subtypes, the focus of the present study is on the Approaches to classifying MLD subtypes
linguistic and spatial antecedents of MLD subtypes.
Previous studies have documented the relation be- In the few empirical studies concerning the classifi-
tween language skills and numerical skills; these lan- cation of MLD subtypes, researchers have typically
guage skills include, but are not limited to, phonological adopted cluster analysis (Bartelet et al., 2014; Chan &
awareness (Simmons & Singleton, 2008), receptive vo- Wong, 2020; von Aster, 2000). In cluster analysis, the
cabulary (Purpura & Ganley, 2014), letter knowledge best fitting model determines the number of clusters,
(Zhang et al., 2014), and rapid naming (Cirino, 2011). after which the average scores of measured variables are
Phonological awareness, which enables people to de- computed and used to define the characteristics of each
code the meaning of a spoken language, has been found cluster. Cluster analysis can not only provide a simple
to be related to children's capacity to recall the phono- and direct report of cluster profiles but also accommo-
logical representations of numbers (i.e., number words) date multiple measures with different difficulty levels
from long-­term memory (see a review in Simmons & (Leonard & Droege, 2008). However, the use of cluster
Singleton, 2008). Poor receptive vocabulary contributes analysis for classification has a few limitations. First, in
to children's deficits in understanding mathematical lan- cluster analysis there are no prior assumptions about the
guage, such as words describing the relations between number and characteristics of the subtypes, and thus it
quantities (more, less, equal; Purpura & Ganley, 2014). may be difficult to interpret some subtypes from a theo-
Apart from spoken language skills, written language retical perspective (Templin & Henson, 2006). Second,
skills such as letter knowledge, also serve as a prereq- the data-­driven nature of cluster analysis may contribute
uisite for understanding symbolic numbers (e.g., Arabic to inconsistent findings about MLD subtypes even with
digits; Zhang et al., 2014). Moreover, rapid automatized similar tasks (e.g., Bartelet et al., 2014; von Aster, 2000).
naming (RAN) has been found to be involved in object Third, cluster analysis does not allow for variables with
recognition and recognizing written symbols, such as high correlations (Tabachnick et al., 2007), a condition
Arabic digits (Cirino, 2011). MLD children in general that is unlikely to be satisfied in empirical studies. As a
have been found to have poorer language skills than typ- result, previous studies have often combined highly cor-
ically developing children or children with LA, including related variables into one factor (e.g., Bartelet et al., 2014).
RAN (Zhang et al., 2020), receptive vocabulary (Cowan Lastly, cluster analysis is highly sensitive to the initiali-
& Powell, 2014), and phonological awareness (Murphy zation phase in data analysis and outliers because the
et al., 2007). Although the relations between language profile of each cluster is computed by the average scores
skills and children's numerical skills and the contribu- of the measured variables (Bhagat et al., 2016).
tion of language skills to general MLD have been well To address the limitations of traditional classifica-
established, only a few studies have investigated how tion approaches (i.e., cluster analysis), the current study
language skills contribute to MLD subtypes. Willburger adopted a new classification approach, CDM, to iden-
et al. (2008) showed that children with comorbid MLD tify MLD subtypes. CDM, as a type of confirmatory
and reading disability (RD), which is similar to the se- latent class model, classifies participants based on their
mantic memory subtype in Geary (1993), had poorer mastery status (mastery or non-­mastery) of a series of
RAN performance than the other MLD children and competencies (also referred to as “attributes”) measured
typically developing children. in a test (de la Torre, 2011). These competencies are pre-
Previous studies have also documented the correla- defined by specifying a Q-­matrix, the role of which is
tion between spatial skills and numerical skills in typi- similar to that of the assumed latent construct in confir-
cally developing children (Ouyang et al., 2022) and MLD matory factor analysis. To illustrate, the Q-­matrix speci-
children (Rourke, 1993). Specifically, Geary et al. (2012) fies the competencies measured in each item of a test (see
demonstrated that MLD children showed poorer spatial Table 1 for an example, in which “1” means that the item
4 |    OUYANG et al.

TA BL E 1 The initial Q-­matrix


performance in the Program for International Student
Item A1 A2 A3 Assessment (PISA). However, no studies, to our knowl-
Item 1 1 0 0
edge, have used CDM to identify subtypes of learning
disabilities or MLD.
Item 2 1 0 0
Item 3 1 0 0
Item 4 1 0 0 The present study
Item 5 0 1 0
Item 6 0 1 0 Our review of the literature revealed several limitations
Item 7 0 1 0 of previous studies. First, MLD students might show
Item 8 0 0a 1
early signs of difficulties in mathematics learning before
formal mathematics education, but little is known about
Item 9 0 0 1
whether MLD students show heterogeneity in domain-­
Item 10 0 0 1 specific skills as early as in preschool. Second, the data-­
Note: A1 represents counting sequence knowledge, A2 represents basic driven nature of the traditional classification approach
arithmetic concepts, A3 represents number-­numerosity mapping. “1” = the
(i.e., cluster analysis) to identifying MLD subtypes might
attribute was measured by the item, “0” = the attribute was not measured by
the item. lead to difficulties in determining the number of sub-
a
This cell was revised to “1” after Q-­m atrix verification. types, interpreting the identified clusters, and general-
izing the findings (Templin & Henson, 2006). Third, the
language and spatial correlates of MLD subtypes are
measures the attribute, and “0” means that the item does understudied.
not measure the attribute) according to the assumptions In order to fill these research gaps, the present study
or theories about the specific skills involved when an- aimed to address the following research questions (RQs):
swering each item (Sorrel et al., 2016). Importantly, the
assumed Q matrix can be verified based on examinee re- RQ1: What subtypes of MLD can be classified based
sponses to the tasks, which indicates whenever a q-­vector on children's deficits in early domain-­specific numer-
of an item is over-­specified (i.e., a “0” is supposed to be a ical skills using CDM?
“1”) or under-­specified (i.e., a “1” is supposed to be a “0”; Hypothesis 1: Based on the findings of previous stud-
de la Torre & Chiu, 2015). The main output of CDM is ies (Bartelet et al., 2014; von Aster, 2000), five sub-
the mastery status of the attributes involved in the test of types may be identified: a counting deficit subtype, a
each participant (also referred to as an attribute profile). number-­numerosity mapping deficit subtype, a basic
For instance, if a test measures three competencies, the arithmetic concepts deficit subtype, a symbolic defi-
possible attribute profiles of the participants would be cit subtype (with deficits in both counting and map-
“000,” “001,” “010,” “100,” “011,” “101,” “110,” and “111,” ping), and a pervasive deficit subtype (with deficits in
in which 0 represents the non-­mastery status of the attri- all the measured number skills).
bute, and 1 represents mastery. Therefore, according to RQ2: To what extent is CDM reliable and valid in
the attribute profiles, at most 8 latent classes of partici- identifying MLD subtypes?
pants can be identified by the test. Hypothesis 2: CDM shows good reliability in classi-
Given the characteristics of CDM, there are a few fying MLD subtypes. Hypothesis 3: With respect to
advantages of using the CDM approach to identify sub- internal validity, children with a certain MLD sub-
types of MLD over cluster analysis. First, CDM, unlike type perform worse in the domain-­specific numerical
data-­driven methods, considers previous theoretical as- skills that are identified with deficits compared with
sumptions about the deficits of MLD subtypes by spec- a control group and the MLD children without such
ifying the attributes and Q-­matrix in advance. Second, deficits. We used LA children rather than typically
the results of CDM provide the attribute pattern for each achieving children as the control group to reduce the
participant, which renders the interpretation of latent false-­positive error of the diagnosis and make our re-
classes easy. Third, CDM allows for high correlations sults more stringent. Hypothesis 4: With respect to
between attributes. In view of the advantages of CDM, predictive ability, children with different MLD sub-
it has been applied to identify latent classes in many re- types differ in arithmetic reasoning in fourth grade.
search areas, such as digital literacy (Liang et al., 2021) Specifically, children with subtypes involving deficits
and psychological disorder (Templin & Henson, 2006), in counting and number-­numerosity mapping show
and these studies have shown CDM's feasibility and ad- lower arithmetic reasoning ability in fourth grade
vantages in the classification of latent classes. Moreover, than their peers without such deficits.
CDM has been used in educational assessments, espe- RQ3: What are the language and spatial correlates of
cially in mathematical assessments (e.g., Wu et al., 2021). different MLD subtypes?
For example, Wu et al. (2021) have used CDM to identify Hypothesis 5a: Children with MLD subtypes involv-
different cognitive profiles of children's mathematical ing deficits in counting, basic arithmetic concepts, and
SUBTYPES OF MATHEMATICS DISABILITY    | 5

number-­numerosity mapping have poorer language Measures of cognitive-­linguistic antecedents, which were
skills than LA children (Geary, 2004). Hypothesis administered at T1 (preschool), consisted of phonologi-
5b: MLD children with subtypes involving deficits cal awareness (10 initial phoneme identification tasks;
in number-­numerosity mapping and basic arithmetic Lerkkanen et al., 2007), letter knowledge (29 Finnish
concepts have poorer spatial visualization skills than letters; Lerkkanen et al., 2006), spatial visualization (31
LA children. items; Woodcock & Johnson, 1977), receptive vocabulary
(the shortened version of Peabody Picture Vocabulary
Test—­ Revised; Dunn & Dunn, 1981), and RAN (50
M ET HOD object-­naming tasks; Ahonen et al., 1999). The numeri-
cal, language, and spatial tasks were administered indi-
Participants and procedure vidually in a quiet room at T1. The Arithmetic Reasoning
Test (Räsänen, 2000) was used for validation and it was
This study was not preregistered and was part of a lon- administered in a group setting at Grade 4 (T6). Fluid in-
gitudinal project which followed 1839 Finnish children telligence was tested in groups using the short version of
(966 boys, mean age at kindergarten entry = 74 months, Raven's Colored Progressive Matrices (Raven et al., 1983),
SD = 3.60 months) for a total of six times from kinder- and parental education was coded from 1 = no vocational
garten (April 2007; Time 1 [T1]; n = 1839) to fourth grade education to 7 = Licentiate or Doctoral degree.
(April 2011; Time 6 [T6], Grade 4; n = 1596). The par-
ticipants were recruited from four municipalities located
in Eastern, Central, and Western Finland. Written con- Data analysis
sent was obtained from participants' parents. The pro-
ject received approval from the Human Research Ethics Identifying subtypes of MLD through CDM
Committee of the University of Jyväskylä.
Among the 1839 children, 99 were identified with Initial determination of the Q-­matrix
MLD and 420 with LA through latent class growth mod- Based on the content of the numerical tasks (e.g., Koponen
eling based on their development trajectory of arithme- et al., 2016; Zhang et al., 2020), an initial Q-­matrix was
tic fluency from first to fourth grade in a previous study constructed by specifying each item corresponding to
(Zhang et al., 2020; see Supporting Information for more an attribute (see Table 1). A confirmatory factor analy-
details). Among the 99 students with MLD, 53 were boys sis supported the structure of the initial Q-­matrix. The
and 46 were girls. Among the 420 low achievers, 206 were three-­factor model showed an acceptable model fit
boys and 214 were girls. The average age was 74 months (χ2 = 209.72, p < .01, comparative fit index = .95, Tucker–­
for both MLD and LA children, and the SDs were 4.0 Lewis index = .93, root mean square error of approxima-
and 3.5 months for MLD and LA children respectively. tion [RMSEA] = .055, standardized root mean squared
Students with MLD had substantially lower scores in residual = .045), with factor loadings larger than .30.
fourth grade and a slower growth rate, which indicated
that they benefited less from formal mathematics edu- Q-­matrix validation using empirical data
cation than their peers. In contrast, students with LA The validation of the Q matrix was evaluated using a dis-
showed a mild but persistent difficulty in mathematics crimination index (de la Torre & Chiu, 2015) computed
from first to fourth grade. Among the MLD and LA from examinee responses to the tasks. The discrimination
children, 18 children did not participate in fourth grade index indicates whenever a q-­vector of an item is over-­
due to their school transfer or moving out of the school specified (i.e., when a “0” is supposed to be a “1”) or under-­
district. Attrition analysis showed no significant differ- specified (i.e., when a “1” is supposed to be an “0”). In the
ence in all the variables measured at T1 between children present study, changes to the Q-­matrix suggested by the
who left the project and those who were retained. model were adopted if (a) the proportion of the discrimina-
tion index of a particular q-­vector relative to the maximum
(PVAF; de la Torre & Chiu, 2015) after modification was
Measures larger than 0.8, and (b) the modification of the q-­vector
made theoretical sense. We included all 1839 participants
A battery of cognitive and numerical tests was adminis- tested at T1 in the Q-­matrix validation and CDM estima-
tered to students (see Appendix S1). The measures used tion since the sample size required for an accurate esti-
to identify subtypes of MLD consisted of three numeri- mation in CDM is typically larger than 500 or 1000 (de la
cal tasks: a counting task (e.g., “counting forward from Torre, 2011). However, follow-­up analyses were conducted
6 to 13”; Items 1–­4 in Table 1), a basic arithmetic con- for MLD and LA children (N = 519).
cepts task (e.g., “please draw one more ball than the balls
in the picture”; Items 5–­7 in Table 1), and a number-­ Selection of the appropriate CDM
numerosity mapping task (e.g., “draw as many balls as According to different assumptions about the combina-
the Arabic numeral represents”; Items 8–­10 in Table 1). tion of multiple skills measured in one item, different
6 |    OUYANG et al.

CDMs have been proposed. The Deterministic-­Input, in counting. Second, predictive validity was evaluated
Noisy “And” Gate (DINA) model (Junker & Sijtsma, 2001) by examining whether MLD children with different sub-
is representative of non-­compensatory models that assume types also differed in arithmetic reasoning in Grade 4.
that only if participants master all the measured attributes The scores of arithmetic reasoning were compared be-
of the item can they correctly answer it. The Deterministic tween the MLD subtypes using an ANOVA.
Input, Noisy “Or” Gate (DINO) model (Templin &
Henson, 2006) is representative of compensatory mod-
els that assume that participants can correctly answer an The linguistic and spatial correlates of
item if they master at least one measured attribute of the MLD subtypes
item. The Generalized Deterministic, Noisy “And” Gate
(G-­DINA) model (de la Torre, 2011) is a saturated model We investigated whether the language and spatial ante-
that considers both kinds of relations (Sorrel et al., 2016) cedents (i.e., phonological awareness, letter knowledge,
and assumes that the probability of participants success- receptive vocabulary, rapid naming, and spatial visuali-
fully answering an item increases if they master more at- zation) predicted the odds of belonging to each MLD
tributes. The three models will be the same when the item subtype versus LA children using multiple logistic re-
only measures one attribute. These models were compared gression. The sex, age, parental education, and fluid in-
to test different assumptions about items that measured telligence of the children were controlled for. The data,
multiple attributes. The likelihood ratio (LR) was used to study materials, and data analysis codes are available
compare the fitness of the reduced models and the satu- upon request.
rated G-­DINA model. RMSEA was adopted to evaluate
the absolute model fit.
R E SU LT S
Estimating attribute profiles
The main output of CDM is the { attribute profile} of each Descriptive statistics
student represented by 𝛼 i = 𝛼 i1 , 𝛼 i2 , … , 𝛼 iK , which
specifies the mastery status of each attribute, with K de- The descriptive statistics and correlations of measured
noting the number of attributes. 𝛼 ik = 1 or 0 means that variables are presented in Table 2. The three numerical
this attribute is mastered or not been mastered by the tasks showed small but significant correlations between
participant, and there are at most 2K latent classes. each other (rs > .097, ps < .05). Counting was correlated
with all the domain-­general variables (|rs| > .126, ps < .01).
Reliability evaluation in CDM Basic arithmetic concepts and number-­numerosity map-
Common indicators of reliability in the CDM frame- ping were associated with all the domain-­general vari-
work are pattern-­and attribute-­level classification accu- ables (rs > .87, ps < .05) except RAN, which provided the
racy indices, which indicate the agreement between the statistical basis for later regression analyses of the ante-
expected classification and the observed classification of cedents. Children's arithmetic reasoning skill at T6 was
the participants (Wang et al., 2015). Although there are positively associated with counting (r = .252, p < .01) and
no gold standards for the accuracy indices, simulation basic arithmetic concepts at T1 (r = .214, p < .01).
studies have indicated that the pattern-­level accuracies
of a high-­quality test with 10 items and 5 attributes—­
similar to that in the present study—­are about 0.55, and Cognitive diagnostic modeling results
attribute-­level accuracies larger than 0.88 indicate good
reliability for the test (Wang et al., 2015). Based on the original Q-­matrix (Table 1), the validation
results suggested that the q-­vector of Item 8 (i.e., “draw
as many balls as the number below represents”) needed
Validation of the MLD subtype classification revision (PVAF = .64 < .80). The model suggested that
Item 8 measured not only number-­numerosity mapping
Validation of the classification consisted of two steps. but also basic arithmetic concepts. Theoretically, Item 8
First, internal validity was evaluated by examining (“draw as many balls as the Arabic numeral represents”)
whether the raw scores of the three domain-­specific skills measured children's understanding of an important
predicted the odds of belonging to each MLD subtype arithmetic concept, namely, equivalence. After including
versus the control group (LA children) using logistic re- the attribute (basic arithmetic concepts) in the q-­vector
gression. In addition, the differences in the raw scores of Item 8, the model resulted in a better PVAF = 0.93,
of the three numerical tasks between subtypes were which was larger than 0.8. Therefore, we modified the
examined via an ANOVA. For example, we examined q-­vector of Item 8 to “011.”
whether MLD children with the counting deficit sub- The model fit indices of the DINA, DINO and G-­DINA
type and those with the pervasive deficit subtype (both models are shown in Table S1. The RMSEA indicated an
showed a deficit in counting) had equivalent raw scores excellent model fit for all three models (RMSEA <0.03,
SUBTYPES OF MATHEMATICS DISABILITY    | 7

Liu et al., 2016). The relative model fit showed that the

15.88
3.17
G-­DINA model fitted better than the DINA (LR = 16.84,

.120 (.007)** —­
13
p < .001) and DINO models (LR = 19.29, p < .001). This re-
sult suggested that the more parsimonious models (DINA

19.57
and DINO) reduced the model fit significantly, so we

3.25
.233 (<.001)** −.092 (.037)* —­
12

chose the G-­DINA model for further analysis. The results


demonstrated that the relation between the two attributes

−.080 (.074)
measured by Item 8 was neither compensatory (DINO)

74.70
17.40
nor non-­compensatory (DINA); instead, the probability

−.192 (<.001)** −.138 (.002)** —­


11

of correctly answering Item 8 increased when partici-

.119 (.008)**
pants had mastered more attributes (G-­DINA).

13.66
2.37
—­
10

Identified MLD subtypes


.172 (<.001)** .189 (<.001)** .229 (<.001)** .220 (<.001)**
.208 (<.001)** .210 (<.001)**

.142 (.001)**
Table 3 lists the attribute pattern of each subtype group
21.27
as well as its prevalence. We focused on the subtypes that
6.95
.163 (<.001)** .231 (<.001)** .492 (<.001)** —­

accounted for more than 5% of the MLD children in the


9

follow-­up analyses, which is a criterion typically used in


−.143 (.001)**

.153 (.001)**

latent class analyses to determine the number of classes


(He & Fan, 2018). Among the 99 MLD children, 40
8.72
1.75
—­

(40%) had no deficits in any of the three attributes (111 in


8

Table 3). However, they were identified as having MLD


.281 (<.001)** .147 (.001)**

.157 (<.001)** .087 (.048)*


−.062 (.161)

.214 (<.001)** .073 (.103)

due to their deficits in arithmetic fluency; therefore, this


subtype was labeled the arithmetic fluency deficit only
2.66
0.71
.325 (<.001)** —­

subtype, or the fluency deficit subtype for short. A total of


7

22 MLD children (22%) had a deficit in counting but not


−.085 (.054)

in the other attributes (011), so this subtype was labeled


the counting deficit subtype. Twelve MLD children (12%)
0.57
2.78
—­

had deficits in all three attributes, so this subtype was


6

−.126 (.004)**
.166 (<.001)** .144 (.001)** .464 (<.001)**
.138 (.002)** .249 (<.001)**
.113 (.010)** .221 (<.001)**

.252 (<.001)**

labeled the pervasive deficit subtype. Eight MLD chil-


.137 (.002)** .211 (<.001)**
.119 (.007)** .318 (<.001)**

.159 (<.001)**

dren (8%) were characterized by deficits in counting and


The descriptive statistic and correlations between the measured variables

number-­numerosity mapping (010), so this subtype was


2.24
1.28
—­

labeled the symbolic deficit subtype. Eight MLD children


5

(8%) had difficulties in counting and basic arithmetic


Note: N = 519. T1: preschool, T6: fourth grade. The p values are presented in parenthesis.
.097 (.027)*

.172 (<.001)** .093 (.037)*


.139 (.002)** .184 (<.001)** .108* (.014)
.011 (.795)

concepts (001), so this subtype was labeled the counting


16.89
1.03

and concept deficit subtype.


—­
4

−.120 (.010)*
.109 (.020)*
.108 (.021)*

.040 (.388)
.022 (.630)

.073 (.120)
.024 (.614)

Reliability and validity of the classification


1.45
4.11
—­
3

Reliability of the classification


−.086 (.066)

−.158 (<.001)** −.084 (.061)


−.017 (.701)
−.172 (<.001)** .104 (.018)*

.101 (.021)*
.041 (.348)

.023 (.601)
.085 (.053)
.057 (.199)
.065 (.141)

Table 3 shows that the pattern-­level accuracies of all


73.53
3.63
—­

subtypes were larger than 0.55. Moreover, the attribute


2

level classification accuracies of counting, basic arith-


.202 (<.001)**
8 T1 phonological awareness .133 (.002)**
.131 (.003)**
.141 (.001)**

−.051 (.242)
−.058 (.184)

−.022 (.621)
−.091 (.051)

−.014 (.753)

metic concepts and number-­numerosity mapping were


.021 (.630)

0.93, 0.97, and 0.99 respectively. Both the pattern-­level


0.50
0.50

and attribute-­level accuracies indicated a good level of


—­
1

reliability of the classification.


12 T1 receptive vocabulary
13 T6 arithmetic reasoning
10 T1 spatial visualization
6 T1 arithmetic concepts

9 T1 letter knowledge
3 Parental education

11 T1 rapid naming

Validation of the domain-­specific deficits of


*p < .05; **p < .01.

MLD subtypes
5 T1 counting

7 T1 mapping
TA BL E 2

4 T5 IQ

We first examined whether children with each subtype


2 Age
1 Sex

SD
M

were more likely to have lower scores in the non-­mastered


8 |    OUYANG et al.

TA BL E 3 The attribute profiles and their prevalence among mathematics learning disability

Classification
Subtype Attribute profilea N Prevalence (%) accuracy

1. Fluency deficit 111 40 40 0.97


2. Counting deficit 011 22 22 0.70
3. Pervasive deficits 000 12 12 0.81
4. Symbolic deficit 010 8 8 0.85
5. Counting and concepts deficits 001 8 8 0.72
6. Mapping deficit 110 4 4 0.81
7. Concept deficit 101 3 3 0.55
8. Concepts and mapping deficits 100 2 2 0.55
a
The attribute profile indicated the mastery status of each of the three attributes respectively: counting, basic arithmetic concepts, and number-­numerosity
mapping. 1 = the attribute was mastered, 0 = the attribute was not mastered. The subtypes that accounted for more than 5% of the MLD children were bolded.

T A B L E 4 Estimated odds ratios (OR) of each subtype and low achievement (LA) in relation to three numerical variables in logistic
regression models

LA versus 111 LA versus 011 LA versus 000 LA versus 010 LA versus 001

Variables OR p-­Value OR p-­Value OR p-­Value OR p-­Value OR p-­Value

Counting 1.55** .01 0.32*** <.001 .44* .04 0.49* .04 0.17** .01
a
Arithmetic —­ —­ 1.42 .61 .28** .01 —­ —­ 0.09*** <.001
concepts
Mapping 4.75 .09 4.79 .09 .34** .01 0.21*** <.001 5.18* .03
Note: 111 = The fluency deficit subtype, 011 = The counting deficit subtype, 000 = The pervasive deficit subtype, 010 = The symbolic deficit subtype, 001 = The
counting and concept deficit subtype.
a
The empty cells were due to that the Wald test cannot be computed as the variance of the raw scores of the variables is 0 for children with a certain subtype.
*p < .05; **p < .01; ***p < .001.

domain-­specific skills than LA children. The logistic re- they did not have weaker arithmetic concepts than LA
gressions are shown in Table 4, with each column rep- children. Children with the counting and concept deficit
resenting a regression model. We coded the LA group subtype were more likely to show lower scores in count-
as “0” and each subtype group as “1,” so an odds ratio ing (OR = .17, p < .01) and arithmetic concepts (OR = .09,
smaller than 1 would mean that children with a certain p < .001), and they also showed higher scores in number-­
subtype were more likely to have lower scores in the numerosity mapping (OR = 5.18, p < .05) than the LA
predictor than the LA group, and vice versa. The first group. To recap, the results from logistic regression were
column in Table 4 shows that children with the fluency consistent with the attribute patterns of the MLD sub-
deficit subtype were more likely to have higher scores in types, which validated the attribute patterns estimated
counting than the LA group (OR = 1.55, p < .01). Although by the CDMs. The differences in the three numerical
this finding was unexpected, it validated the attributes skills among MLD children with different subtypes also
profile of the subtype group, which did not have signifi- supported their attribute patterns in most cases (see
cantly lower scores in the three domain-­specific skills Table S2). In other words, MLD subgroups identified
compared with the LA group. MLD children with the with a deficit in a certain attribute showed significantly
counting deficit subtype were more likely to have lower lower scores in the corresponding task than did the sub-
scores in counting than the LA group (OR = .32, p < .001), groups identified as having mastered this attribute.
and they did not show significantly lower scores in basic
arithmetic concepts or mapping tasks. MLD children
with the pervasive deficit subtype were more likely to have Differences in arithmetic reasoning between
lower scores in counting (OR = .44, p < .05), basic arithme- children with different MLD subtypes
tic concepts (OR = .28, p < .01), and number-­numerosity
mapping (OR = .34, p < .01) than the LA group. MLD The differences in fourth-­ g rade arithmetic reasoning
children with the symbolic deficit subtype were more between children with different subtypes were further
likely to have lower scores in counting (OR = .49, p < .05) investigated. Results indicated significant differences
and number-­ numerosity mapping than the LA group in arithmetic reasoning between children with the five
(OR = .19, p < .001). Children with this subtype had maxi- main subtypes (F[4, 81] = 4.39, p < .01). Further post hoc
mum scores in arithmetic concepts, which indicated that analyses (Bonferroni) indicated that children with the
SUBTYPES OF MATHEMATICS DISABILITY    | 9

fluency deficit subtype showed higher scores in arithme- (1.04–­1, p < .05) when controlling for demographic and
tic reasoning than their peers with the symbolic deficit other cognitive-­linguistic factors.
subtype (mean difference = 3.68, p < .05) and the pervasive
deficit subtype (mean difference = 3.00, p < .05).
DI SC US SION

Linguistic and spatial correlates of the Regarding RQ1, five main subtypes were identified
MLD subtypes among the 99 MlD children, which accounted for more
than 90% of the MLD children. These subtypes are the
Table 5 shows the logistic regression, in which the six an- fluency deficit subtype, the counting deficit subtype, the
tecedents were used to predict the odds of belonging to pervasive deficit subtype, the symbolic deficit subtype,
each MLD subtype compared with the LA group. The and the counting and concept deficit subtype. With re-
results indicated that children with the counting deficit spect to RQ2, CDM showed good reliability and validity
subtype were more likely to have lower scores in spa- in the classification of the MLD subtypes. Concerning
tial visualization (OR = .78, p < .01) than the LA group. RQ3, this study demonstrated that spatial visualization
Specifically, for each point increase in spatial visualiza- uniquely contributed to the identification of the count-
tion, the odds of children being identified with the ver- ing deficit subtype; language skills (i.e., phonological
bal deficit subtype rather than the LA group decreased awareness, letter knowledge, and receptive vocabulary)
by 22% (1–­0.78) when controlling for child's age, sex, IQ, uniquely contributed to the identification of the pervasive
parental education, and the other cognitive-­ linguistic deficit subtype; letter knowledge uniquely contributed to
antecedents. Moreover, for each point increase in the the symbolic deficit subtype; and RAN uniquely contrib-
receptive vocabulary, phonological awareness, and let- uted to the identification of the counting and concept defi-
ter knowledge scores, the odds of children being identi- cit subtype after including control variables in each case.
fied with the pervasive subtype rather than the LA group
decreased by 21% (1–­0.79, p < .05), 37% (1–­0.63, p < .05),
and 15% (1–­ 0.85, p < .05) respectively, after including The identification of MLD subtypes
control variables and the other cognitive-­linguistic an-
tecedents. Additionally, for each point increase in the Among the five main subtypes identified in this study, the
letter knowledge score, the odds of children being identi- fluency deficit subtype accounted for 40% of MLD chil-
fied with the symbolic deficit subtype rather than the LA dren. The identification of this subtype was inconsistent
group decreased by 14% (1–­0.86, p < .01) when controlling with our hypothesis, but the subtype was similar to the
for demographic and other cognitive-­linguistic factors. no numerical deficit subtype (8.4% of the MLD children)
Lastly, for each second increase in RAN time, the odds identified in Bartelet et al. (2014) and the mild deficit
of children being identified with the verbal and concept subtype (50% of the MLD children) identified in Chan
deficit subtype rather than the LA group increased by 4% and Wong (2020). Children with this subtype appeared

TA BL E 5 Estimated odds ratios (OR) of each subtype and low achievement (LA) in relation to predictors in logistic regression model

LA versus 111 LA versus 011 LA versus 000 LA versus 010 LA versus 001

Variables OR p-­Value OR p-­Value OR p-­Value OR p-­Value OR p-­Value

Gender 0.60 .16 2.32 .12 2.52 .29 0.81 .83 3.37 .29
Age 1.06 .21 1.02 .77 1.06 .64 0.90 .45 1.12 .37
Parental education 1.07 .60 0.99 .93 1.29 .39 0.85 .59 0.86 .68
T6 fluid intelligence 0.92 .62 0.65 .06 0.48 .07 0.51 .10 0.94 .91
T1 phonological 1.16 .34 1.13 .47 0.63* .02 1.75 .10 1.25 .52
awareness
T1 receptive 1.03 .69 1.09 .32 0.79* .03 1.18 .30 0.83 .27
vocabulary
T1 letter knowledge 1.03 .44 0.94 .09 0.85* .02 0.86* .01 0.86 .07
T1 spatial 0.92 .28 0.78** .01 0.97 .82 0.75 .09 0.93 .65
visualization
T1 RAN 1.01 .60 1.01 .45 1.00 .91 1.00 .90 1.04* .02
Note: 111 = the fluency deficit subtype, 011 = the counting deficit subtype, 000 = the pervasive deficit subtype, 010 = the symbolic deficit subtype, 001 = the counting
and concept deficit subtype.
Abbreviation: RAN, rapid automatized naming.
*p < .05.; **p < .01.
10 |    OUYANG et al.

to have the fewest difficulties in the three numerical performance in counting and Arabic number knowl-
skills and cognitive-­ linguistic antecedents among all edge) identified in Bartelet et al. (2014) and the symbolic
subtypes of MLD children. Nevertheless, children with deficit group identified in Chan and Wong (2020). It is
the fluency deficit subtype might have deficits in other nu- not surprising that children with this subtype have defi-
merical skills that were not measured in this study, such cits in verbal number processing and number-­numerosity
as the approximate number system (ANS). Mazzocco mapping, as the process of learning Arabic numerals
et al. (2011) showed that MLD children were less accu- has been found to be based on the process of mapping
rate when approximately representing magnitude com- number words onto quantities (Lipton & Spelke, 2005).
pared with typically achieving children. Notably, despite This also explains why very few children were identified
the importance of the ANS, it is not yet clear whether with a deficit only in mapping (4%). The identification
different variants of the ANS tasks really catch some- of this subtype also supports the hypothesis that MLD
thing important in MLD children (Szűcs et al., 2013). is at least partially caused by difficulties in processing
Moreover, based on De Smedt et al. (2013), the ANS symbolic numbers, including number words and Arabic
seems to differentiate between MLD and LA only after digits (Bartelet et al., 2014).
the ages of 9–­10 but not before that. Therefore, it will be Finally, the present study also identified children with
interesting for future studies to examine whether MLD the counting and concept deficit subtype, which accounted
children with the fluency deficit subtype show a deficit in for about 8% of MLD children. To our knowledge, simi-
the ANS using different ANS tasks and participants of lar subtypes have not been classified in previous studies,
different ages. Another possibility is that children with and it is also inconsistent with our hypothesis. However,
the fluency deficit subtype have a deficit in the speed, but the identification of this subtype supports Geary's theory
not the accuracy, of their processing of numerical tasks (Geary, 2004) that a poor understanding of conceptual
(Chan & Wong, 2020; Rousselle & Noël, 2007). knowledge leads to difficulties in learning mathematics.
The second subtype, the counting deficit subtype, ac- Furthermore, very few children were identified with a
counted for 20% of the MLD children, was consistent deficit only in basic arithmetic concepts (3%), indicating
with our hypothesis. This subtype is similar to the “ver- that the concept deficit was likely to be accompanied
bal subtype” in von Aster's (2000) study, which was by a deficit in counting. The finding supports the mul-
characterized by frequent errors in counting routines. tiple deficits hypothesis (Desoete & Grégoire, 2006) and
To illustrate, the counting deficit found in children with extends previous studies by indicating that the two nu-
this subtype might result in difficulties in understand- merical deficits may co-­occur and jointly contribute to
ing cardinal concepts, remembering mathematical facts, children's MLD. This can be explained by the notion that
and avoiding retrieval errors, thus leading to MLD counting often serves as an arithmetic strategy that fa-
(Geary, 1993). Moreover, consistent with their attribute cilitates children's learning of basic arithmetic concepts
pattern, children with this subtype also showed lower (Sievert et al., 2021).
scores in counting than both LA children and children
with the fluency deficit subtype.
The pervasive deficit subtype, which accounted for Reliability and validity of the MLD subtype
12% of the MLD children, was characterized by defi- classification
cits in all three numerical competencies measured in
the present study. The identification of this subtype was Consistent with the research hypotheses, the present
consistent with our hypothesis and von Aster's (2000) study demonstrated that CDM showed good reliability
and Bartelet et al.'s (2014) studies. As von Aster (2000) and good internal and predictive validity in the classi-
suggested, children with this subtype may have deficits fication of MLD subtypes. MLD children with differ-
in analogue magnitude representation, which is related ent subtypes differed in arithmetic reasoning in fourth
to the malfunction of the brain (e.g., bilateral horizontal grade: MLD children with the symbolic deficit subtype
intraparietal sulci), or to genetic influences. Such deficits and the pervasive deficit subtype had poorer performance
may appear at an early age and impede children's acqui- in arithmetic reasoning than their counterparts with the
sition of symbolic number skills as well as basic arith- fluency deficit subtype. This finding is consistent with
metic concepts. Another possibility is that children with previous studies indicating that counting skills (Zhang
this MLD subtype have deficits in working memory. et al., 2017) and number-­ numerosity mapping skills
Deficits in working memory have been found to be com- (Jang & Cho, 2018) contribute to arithmetic reasoning
mon in MLD children (Geary, 2013) and to contribute to skills. Moreover, interestingly, a deficit only in counting
poor performance in multiple numerical tasks, such as skills did not lead to lower arithmetic reasoning ability
number-­numerosity mapping and explicit number pro- in fourth grade compared with that of children with the
cessing (Geary et al., 2004). other subtypes. One explanation may be that arithmetic
The symbolic deficit subtype, which accounted for 8% reasoning is a composite and complex skill, which de-
of the MLD children, is consistent with our hypothesis pends on a number of cognitive skills (Zhang et al., 2017).
and the access deficit subtype (characterized by poor The influence of a verbal deficit on arithmetic reasoning
SUBTYPES OF MATHEMATICS DISABILITY    | 11

in the early years may be compensated for by other un- Inconsistent with our hypothesis (i.e. children with
impaired numerical skills in later mathematical learning. deficits in number-­numerosity mapping and basic arith-
This explanation is supported by previous findings on metic concepts would exhibit lower spatial ability com-
the compensatory relations between cognitive skills: for pared with LA children), spatial visualization was found
example, Fassbender and Schweitzer (2006) found that to contribute to the identification of the counting deficit
children with attention deficit hyperactivity disorder subtype. However, this finding is consistent with previ-
and difficulties in engaging in verbal strategies to solve ous studies indicating that spatial visualization skills
arithmetic problems tended to rely on spatial, visual, and predict counting skills in typically achieving children
motor strategies. However, this speculation is tentative (Liu & Zhang, 2022; Zhang et al., 2014). It also joins ev-
and requires further investigation. idence showing spatial representations of numbers and
time in the human mind (i.e., numbers and time are spa-
tially represented from left to right on a horizontal line;
Linguistic and spatial correlates of Gibson & Maurer, 2016). A possible explanation for this
MLD subtypes is that children typically practice their counting skills by
pointing to objects from left to right, and such a counting
The present study found that language skills (i.e., pho- practice may be connected with spatial direction (Liu &
nological awareness, receptive vocabulary, and letter Zhang, 2022).
knowledge) were critical correlates of the pervasive
deficit subtype, which supports our hypothesis (H5a).
We speculate that children with this subtype may have CDM as a promising approach to identifying
RD as well, and their deficits in domain-­specific nu- MLD subtypes
merical skills may be due to their deficits in general
language competency (Geary, 1993). Previous find- This study explored whether the CDM approach could
ings have shown that students with comorbid RD and be applied in the classification of MLD subtypes and
MLD display severe difficulties in multiple domain-­ how reliable and valid the method is. There are at least
specific skills, such as retrieving arithmetic facts and three reasons why CDM is a promising approach to
understanding arithmetic concepts (Mazzocco & identifying MLD subtypes. First, the attribute pattern
Räsänen, 2013). of each participant estimated by CDM can provide finer-­
The results regarding the cognitive-­linguistic anteced- grained information about the deficits of each MLD
ents of the symbolic deficit subtype are also consistent subtype than that obtained by cluster analysis. Cluster
with the research hypotheses (H5a). Specifically, this analysis estimates the cognitive profiles of subtypes by
study showed that letter knowledge made a unique con- computing the average scores in the measured variables,
tribution to the probability of being identified with the in which the similarity between participants is empha-
symbolic deficit subtype compared with the LA group sized, but individual differences may not be precisely
after including control variables. This finding corrobo- captured. In contrast, using CDM to estimate the cogni-
rates Zhang et al.'s (2014) study regarding the predictive tive profile of each participant ensures that subtypes with
role of letter knowledge in children's numerical skills. cognitive profiles that are not very common (i.e., those
A possible explanation is that the acquisition of letter comprising very few participants) are not “averaged.”
knowledge may facilitate understanding, decoding, and Moreover, finer-­g rained information about the subtypes
manipulating number symbols (e.g., number words and renders the interpretation easier and is also helpful for
Arabic digits; Zhang et al., 2017). the development of interventions to target the different
In addition, consistent with the research hypotheses, deficits of children with MLD subtypes. Second, CDM
RAN was found to contribute uniquely to the identi- allows for the evaluation of the reliability of the classifi-
fication of the counting and concept deficit subtype. A cation, which cannot be evaluated using cluster analysis.
possible explanation for this is that RAN, as Torgesen CDM provides reliability indices of the classification
et al. (1994) described, reflects how rapidly and easily at attribute and pattern levels. If the classification ac-
an individual retrieves phonological information from curacies of certain patterns are low, researchers should
long-­term memory. If the quality of the storage of pho- exercise caution when interpreting and generalizing the
nological information (e.g., number words) in long-­term subtypes. Furthermore, researchers can assess whether
memory is poor, it could be difficult for children to re- any of the attributes were measured with low reliability
trieve and manipulate these phonological codes, thus by checking the attribute-­level classification accuracies
leading to more errors in counting or being stuck when provided by CDM. If this is the case, additional items or
counting (Simmons & Singleton, 2008). Furthermore, tasks could be developed to improve the reliability of the
RAN may be associated with basic arithmetic concepts attribute. Third, compared with cluster analysis, CDM
indirectly through counting skills because counting is can be used to analyze tasks and attributes that have
an important strategy when children start to learn basic moderate or high correlations, which is in accord with
arithmetic concepts (Geary, 1993). the reality of mathematical research (Wang et al., 2015).
12 |    OUYANG et al.

However, the costs and limitations of CDM should of MLD. Ultimately, we hope that this study can provide
be acknowledged. First, CDM assumes that the latent a guide for applying CDM in the identification of MLD
variables are discrete (e.g., mastery or non-­mastery) and subtypes. Future studies can include more numerical
thus cannot be used to analyze continuous data, such skills to identify more possible subtypes of MLD using
as participants' reaction time. As previous studies have CDM. Second, the MLD subtypes were identified based
suggested (e.g., Geary et al., 2007), some MLD children on three numerical tasks, each of which was measured
may not show deficits in accuracy in numerical tasks by three to four items. Although CDM requires a shorter
but may do so in speed. Future studies could use timed test length than traditional psychometric models (Sorrel
numerical tasks or directly measure children's speed of et al., 2016), the small number of numerical items may
processing as an antecedent. Second, the fine-­g rained have affected the reliability of the tasks and may have
cognitive profiles provided by CDM could be a “double-­ limited the generalizability of the findings. Future re-
edged sword,” because as more attributes are included search should involve more items for each task. Third,
in CDM, the sensitivity of CDM increases, probably re- a critical cognitive-­linguistic antecedent, executive func-
sulting in more subtypes being identified and an overly tioning (e.g., working memory), was not included in the
complex classification. However, this does not mean that current study. For example, working memory has been
all the possible subtypes will be identified and valid be- found to affect children's numerical skills such as basic
cause there might be very few (less than 5%) or no partic- arithmetic concepts (Desoete & Grégoire, 2006), count-
ipants in some subtypes. For example, Liang et al. (2021) ing (Geary, 1993), and number-­ numerosity mapping
used CDM to identify latent classes of digital literacy. (Geary, 1993, 2013). It will be worthwhile for future re-
Although they included 5 attributes (at most 25 = 32 la- search to examine the role of working memory in con-
tent classes could be identified), 14 of the latent classes tributing to MLD subtypes. Fourth, the present study
included no participants, and only three latent classes only included one mathematical outcome variable (i.e.,
had prevalence larger than 5%. The classification results arithmetic reasoning) when examining the predictive va-
depend not only on how many attributes are included lidity of the classification. More mathematical outcome
but also on whether participants differ significantly on variables, such as word problem-­ solving and number
the included variables. Therefore, all the variables in- fact retrieval, should be included in future studies to val-
cluded should be empirically and theoretically justified. idate the MLD subtypes more effectively in different do-
Based on this notion, the initial phase of CDM analyses mains of mathematics. Lastly, while there is stability in
(e.g., the selection of attributes and the construction of a MLD, the profiles of MLD subtypes may change due to
Q-­matrix) is important. Put differently, if the assumed development and learning (Mazzocco & Räsänen, 2013).
factor structure and the number of factors are not well Future studies could measure numerical skills several
founded at the beginning in confirmatory factor analy- times from preschool to early primary grades to investi-
sis, the results will not be trustworthy. Constructing a Q-­ gate whether the subtypes are stable over time.
matrix based on a theoretical model, using think-­aloud
protocols, and validating the Q-­matrix using empirical
data can be helpful in this process. Moreover, validation Implications
is also important after MLD subtyping. Further studies
could validate the classification of MLD subtypes from The present study has made important contributions
several perspectives: examining the stability of the clas- to identifying and theorizing about MLD subtypes.
sification over time, investigating whether there are dif- The finding that MLD children with different subtypes
ferences in the effectiveness of targeted interventions for showed differences in numerical deficits supports the
MLD children with different subtypes, and replication heterogeneity hypothesis of MLD and the proposition
studies with other samples, age groups, and different that MLD children should not be treated as a homoge-
tasks. neous group. Furthermore, this study found that MLD
children with different subtypes differed in mathematical
outcomes (i.e., arithmetic reasoning), indicating that the
Limitations and directions for future research heterogeneity of MLD was not temporary but had a lon-
gitudinal effect on mathematical learning. The finding
The current study has a few limitations. First, the iden- underscores the necessity of identifying MLD subtypes
tified MLD subtypes were based on three numerical and their numerical deficits in early grades. In addition,
skills, while other critical numerical skills, such as the this study examined the cognitive-­linguistic antecedents
ANS, the number line, and subitizing, were not included. underlying MLD subtypes, namely, language and spatial
Moreover, this study only measured children's under- skills. The results extend previous findings and provide a
standing of Arabic symbols, while other symbols such foundation for custom-­built treatment for MLD children
as operative symbols (e.g., “+”) may also affect children's with different subtypes.
arithmetic performance. Therefore, we do not intend to The present study also made important method-
claim that this study identified all the possible subtypes ological contributions to the research field of MLD
SUBTYPES OF MATHEMATICS DISABILITY    | 13

subtypes. The study is the first of its kind to use the Cirino, P. T. (2011). The interrelationships of mathematical precursors
in kindergarten. Journal of Experimental Child Psychology, 108,
CDM approach to identify the subtypes of MLD and
713–­733. https://doi.org/10.1016/j.jecp.2010.11.004
has filled an important research gap in this field. It Cowan, R., & Powell, D. (2014). The contributions of domain-general
demonstrates the feasibility and advantages of this and numerical factors to third-grade arithmetic skills and math-
approach compared with traditional methods and il- ematical learning disability. Journal of Educational Psychology,
lustrates how to use CDM to classify MLD subtypes, 106(1), 214–­229. https://doi.org/10.1037/a0034097
de la Torre, J. (2011). The generalized DINA model frame-
which can inform future studies.
work. Psychometrika, 76, 179–­199. https://doi.org/10.1007/
The present study also has practical implications for s11336-­011-­9207-­7
intervening with children at risk for MLD. First, this de la Torre, J., & Chiu, C. Y. (2015). A general method of empirical
study demonstrated that MLD children showed early q-­matrix validation. Psychometrika, 81, 253–­273. https://doi.
numerical deficits. If these children receive appropri- org/10.1007/s1133​6 -­015-­9467-­8
De Smedt, B., Noël, M. P., Gilmore, C., & Ansari, D. (2013). How do
ate treatment in early years, it may reduce the possi-
symbolic and non-­symbolic numerical magnitude processing skills
bility of their being identified as having MLD in later relate to individual differences in children's mathematical skills? A
years. Second, not only should the treatment be con- review of evidence from brain and behavior. Trends in Neuroscience
ducted at a young age, but the interventions should be and Education, 2(2), 48–­55. https://doi.org/10.1016/j.tine.2013.06.001
“custom-­built.” The straightforward reports provided Desoete, A., & Grégoire, J. (2006). Numerical competence in young
children and in children with mathematics learning disabilities.
by CDM can help researchers to develop interventions
Learning and Individual Differences, 16, 351–­367. https://doi.
for different MLD subtypes. For instance, for MLD org/10.1016/j.lindif.2006.12.006
children with the symbolic deficit subtype, teachers can Dunn, L. M., & Dunn, L. M. (1981). Manual for the Peabody Picture
encourage children to access the meaning of symbolic Vocabulary Test—­Revised. American Guidance Service.
numbers in the learning process, such as illustrating Fassbender, C., & Schweitzer, J. B. (2006). Is there evidence for neural
compensation in attention deficit hyperactivity disorder? A review
the cardinal or ordinal meanings when learning Arabic
of the functional neuroimaging literature. Clinical Psychology
digits. Finally, this research demonstrates the import- Review, 26, 445–­465. https://doi.org/10.1016/j.cpr.2006.01.003
ant role of underlying cognitive-­linguistic antecedents Geary, D. C. (1993). Mathematical disabilities: Cognitive, neuropsy-
in contributing to MLD subtypes. Interventions aimed chological, and genetic components. Psychological Bulletin, 114,
at improving MLD children's cognitive-­l inguistic com- 345–­362. https://doi.org/10.1037/0033-­2909.114.2.345
Geary, D. C. (2004). Mathematics and learning disabilities. Journal
petencies may improve their mathematical ability by
of Learning Disabilities, 37, 4–­15. https://doi.org/10.1177/00222​
addressing root causes. For example, researchers and 19404​03700​10201
educators should pay more attention to the language Geary, D. C. (2013). Early foundations for mathematics learning and
difficulties of children with the pervasive subtype. their relations to learning disabilities. Current Directions in
Training in spatial skills may be helpful for children Psychological Science, 22, 23–­27. https://doi.org/10.1177/09637​
21412​469398
with the counting deficit subtype.
Geary, D. C., Hamson, C. O., & Hoard, M. K. (2000). Numerical
and arithmetical cognition: A longitudinal study of process and
F U N DI NG I N F OR M AT ION concept deficits in children with learning disability. Journal
This study was funded by three grants from the Academy of Experimental Child Psychology, 77(3), 236–­263. https://doi.
of Finland (Nos. 213486, 268586, and 292466). org/10.1006/jecp.2000.2561
Geary, D. C., Hoard, M. K., Byrd-­Craven, J., & DeSoto, M. C. (2004).
Strategy choices in simple and complex addition: Contributions
ORC I D of working memory and counting knowledge for children
Xiangzi Ouyang https://orcid.org/0000-0002-5399-6820 with mathematical disability. Journal of Experimental Child
Xiao Zhang https://orcid.org/0000-0001-6572-4411 Psychology, 88, 121–­151. https://doi.org/10.1016/j.jecp.2004.03.002
Tuire Koponen https://orcid.org/0000-0003-0039-1016 Geary, D. C., Hoard, M. K., Byrd-Craven, J., Nugent, L., &
Numtee, C. (2007). Cognitive mechanisms underlying achieve-
ment deficits in children with mathematical learning dis-
R EF ER ENCE S ability. Child Development, 78(4), 1343–­1359. https://doi.
Ahonen, T., Tuovinen, S., & Leppäsaari, T. (1999). Nopean sarjallisen org/10.1111/j.1467-8624.2007.01069.x
nimeämisen testi [The test of rapid serial naming]. Niilo Mäki Geary, D. C., Hoard, M. K., & Nugent, L. (2012). Independent con-
Instituutti & Haukkarannan koulu. tributions of the central executive, intelligence, and in-­ class
American Psychiatric Association. (2013). DSM 5. American attentive behavior to developmental change in the strategies
Psychiatric Association. used to solve addition problems. Journal of Experimental Child
Bartelet, D., Ansari, D., Vaessen, A., & Blomert, L. (2014). Cognitive Psychology, 113, 49–­65. https://doi.org/10.1016/j.jecp.2012.03.003
subtypes of mathematics learning difficulties in primary educa- Gibson, L. C., & Maurer, D. (2016). Development of SNARC and dis-
tion. Research in Developmental Disabilities, 35, 657–­670. https:// tance effects and their relation to mathematical and visuospatial
doi.org/10.1016/j.ridd.2013.12.010 abilities. Journal of Experimental Child Psychology, 150, 301–­313.
Bhagat, A., Kshirsagar, N., Khodke, P., Dongre, K., & Ali, S. (2016). https://doi.org/10.1016/j.jecp.2016.05.009
Penalty parameter aelection for hierarchical data stream clus- He, J., & Fan, X. (2018). Latent class analysis. In V. Zeigler-­Hill &
tering. Procedia Computer Science, 79, 24–­31. https://doi. T. K. Shackelford (Eds.), Encyclopedia of personality and individ-
org/10.1016/j.procs.2016.03.005 ual differences (pp. 1–­4). Springer.
Chan, W. W. L., & Wong, T. T. Y. (2020). Subtypes of mathematical Jang, S., & Cho, S. (2018). The mediating role of number-­to-­magnitude
difficulties and their stability. Journal of Educational Psychology, mapping precision in the relationship between approximate
112, 649–­666. https://doi.org/10.1037/edu00​0 0383 number sense and math achievement depends on the domain of
14 |    OUYANG et al.

mathematics and age. Learning and Individual Differences, 64, by numeral knowledge. Journal of Educational Psychology, 105,
113–­124. https://doi.org/10.1016/j.lindif.2018.05.005 453–­464. https://doi.org/10.1037/a0031753
Jordan, N. C., Kaplan, D., Nabors Olah, L., & Locuniak, Purpura, D. J., & Ganley, C. M. (2014). Working memory and lan-
M. N. (2006). Number sense growth in kindergarten: A longi- guage: Skill-­specific or domain-­general relations to mathemat-
tudinal investigation of children at risk for mathematics diffi- ics? Journal of Experimental Child Psychology, 122, 104–­121.
culties. Child Development, 77, 153–­175. https://doi.org/10.1111/ https://doi.org/10.1016/j.jecp.2013.12.009
j.1467-­8624.​2006.00862.x Räsänen, P. (2000). NMART—­Arithmetic reasoning test. Niilo Mäki
Junker, B. W., & Sijtsma, K. (2001). Cognitive assessment models with Institute.
few assumptions, and connections with nonparametric item re- Raven, J. C., Court, J. H., & Raven, J. (1983). Manual for Raven's pro-
sponse theory. Applied Psychological Measurement, 25, 258–­272. gressive matrices and vocabulary scales. Lewis.
https://doi.org/10.1177/01466​21012​2032064 Rourke, B. P. (1993). Arithmetic disabilities, specific and other-
Koponen, T., Aunola, K., & Nurmi, J. E. (2019). Verbal counting wise: A neuropsychological perspective. Journal of Learning
skill predicts later math performance and difficulties in mid- Disabilities, 26, 214–­226. https://doi.org/10.1177/00222​19493​
dle school. Contemporary Educational Psychology, 59, 101803. 02600402
https://doi.org/10.1016/j.cedps​ych.2019.101803 Rousselle, L., & Noël, M. P. (2007). Basic domain-­specific skills in
Koponen, T., Salmi, P., Torppa, M., Eklund, K., Aro, T., Aro, M., children with mathematics learning disabilities: A compari-
Poikkeus, A. M., Lerkkanen, M. K., & Nurmi, J. E. (2016). son of symbolic vs non-­symbolic number magnitude process-
Counting and rapid naming predict the fluency of arithmetic ing. Cognition, 102, 361–­395. https://doi.org/10.1016/j.cogni​tion.​
and reading skills. Contemporary Educational Psychology, 44–­ 2006.01.005
45, 83–­94. https://doi.org/10.1016/j.cedps​ych.2016.02.004 Salminen, J. B., Koponen, T. K., & Tolvanen, A. J. (2018). Individuality
Landerl, K., Bevan, A., & Butterworth, B. (2004). Developmental dy- in the early number skill components underlying basic arith-
scalculia and basic numerical capacities: A study of 8–­9-­year-­old metic skills. Frontiers in Psychology, 9. https://doi.org/10.3389/
students. Cognition, 93, 99–­125. https://doi.org/10.1016/j.cogni​ fpsyg.2018.01056
tion.2003.11.004 Schneider, M., Beeres, K., Coban, L., Merz, S., Susan Schmidt, S.,
Leonard, S. T., & Droege, M. (2008). The uses and benefits of clus- Stricker, J., & De Smedt, B. (2016). Associations of non-­symbolic
ter analysis in pharmacy research. Research in Social and and symbolic numerical magnitude processing with mathemat-
Administrative Pharmacy, 4, 1–­11. https://doi.org/10.1016/j.sapha​ ical competence: A meta-­analysis. Developmental Science, 20,
rm.2007.02.001 e12372. https://doi.org/10.1111/desc.12372
Lerkkanen, M.-­K., Poikkeus, A.-­M., & Ketonen, R. (2006). ARMI—­A Siegel, L. S. (1971). The development of the understanding of certain
tool for assessing reading and writing skills in first grade. WSOY. number concepts. Developmental Psychology, 5, 362–­363. https://
Lerkkanen, M.-­K., Poikkeus, A.-­M., & Ketonen, R. (2007). ARMI—­ doi.org/10.1037/h0031425
Luku-­ja kirjoitustaidon arviointimateriaali l. luokalle. NMI-­ Sievert, H., van den Ham, A. K., & Heinze, A. (2021). Are first graders'
bulletin. WSOY. arithmetic skills related to the quality of mathematics textbooks?
Liang, Q., de la Torre, J., & Law, N. (2021). Do background charac- A study on students' use of arithmetic principles. Learning
teristics matter in Children's mastery of digital literacy? A cog- and Instruction, 71, 101401. https://doi.org/10.1016/j.learn​i nstr​
nitive diagnosis model analysis. Computers in Human Behavior, uc.2020.101401
122, 106850. https://doi.org/10.1016/j.chb.2021.106850 Simmons, F. R., & Singleton, C. (2008). Do weak phonological rep-
Lipton, J. S., & Spelke, E. S. (2005). Preschool children's mapping of resentations impact on arithmetic development? A review of re-
number words to nonsymbolic numerosities. Child Development, search into arithmetic and dyslexia. Dyslexia, 14, 77–­94. https://
76, 978–­988. https://doi.org/10.1111/j.1467-­8624.2005.00891.x doi.org/10.1002/dys.341
Liu, Y., Tian, W., & Xin, T. (2016). An application of M2 statis- Sorrel, M. A., Olea, J., Abad, F. J., de la Torre, J., Aguado, D., &
tic to evaluate the fit of cognitive diagnostic models. Journal Lievens, F. (2016). Validity and reliability of situational judge-
of Educational and Behavioral Statistics, 41, 3–­26. https://doi. ment test scores. Organizational Research Methods, 19, 506–­532.
org/10.3102/10769​98615​621293 https://doi.org/10.1177/10944​28116​630065
Liu, Y., & Zhang, X. (2022). Spatial skills and counting sequence Szűcs, D., Nobes, A., Devine, A., Gabriel, F. C., & Gebuis, T. (2013).
knowledge: Investigating reciprocal longitudinal relations in Visual stimulus parameters seriously compromise the measure-
early years. Early Childhood Research Quarterly, 59, 1–­11. https:// ment of approximate number system acuity and comparative
doi.org/10.1016/j.ecresq.2021.09.013 effects between adults and children. Frontiers in Psychology, 4,
Mazzocco, M. M., & Räsänen, P. (2013). Contributions of 444. https://doi.org/10.3389/fpsyg.2013.00444
longitudinal studies to evolving definitions and knowledge of de- Tabachnick, B. G., Fidell, L. S., & Ullman, J. B. (2007). Using multi-
velopmental dyscalculia. Trends in Neuroscience and Education, variate statistics (Vol. 5). Pearson.
2, 65–­73. https://doi.org/10.1016/j.tine.2013.05.001 Templin, J. L., & Henson, R. A. (2006). Measurement of psycholog-
Mazzocco, M. M. M., Feigenson, L., & Halberda, J. (2011). Impaired ical disorders using cognitive diagnosis models. Psychological
acuity of the approximate number system underlies mathemat- Methods, 11, 287–­305. https://doi.org/10.1037/1082-­989x.11.3.287
ical learning disability (dyscalculia). Child Development, 82, Torgesen, J. K., Wagner, R. K., & Rashotte, C. A. (1994). Longitudinal
1224–­1237. https://doi.org/10.1111/j.1467-­8624.2011.01608.x studies of phonological processing and reading. Journal of
Murphy, M. M., Mazzocco, M. M. M., Hanich, L. B., & Early, M. C. Learning Disabilities, 27, 276–­286. https://doi.org/10.1177/00222​
(2007). Cognitive characteristics of children with mathematics 19494​02700503
learning disability (MLD) vary as a function of the cutoff cri- von Aster, M. (2000). Developmental cognitive neuropsychology of
terion used to define MLD. Journal of Learning Disabilities, 40, number processing and calculation: Varieties of developmental
458–­478. https://doi.org/10.1177/00222​19407​04000​50901 dyscalculia. European Child & Adolescent Psychiatry, 9, S41–­S57.
Ouyang, X., Zhang, X., & Zhang, Q. (2022). Spatial skills and num- https://doi.org/10.1007/s0078​70070008
ber skills in preschool children: The moderating role of spatial Wang, W., Song, L., Chen, P., Meng, Y., & Ding, S. (2015). Attribute-­
anxiety. Cognition, 225, 105165. https://doi.org/10.1016/j.cogni​ level and pattern-­ l evel classification consistency and accu-
tion.2022.105165 racy indices for cognitive diagnostic assessment. Journal of
Purpura, D. J., Baroody, A. J., & Lonigan, C. J. (2013). The transition Educational Measurement, 52, 457–­476. https://doi.org/10.1111/
from informal to formal mathematical knowledge: Mediation jedm.12096
SUBTYPES OF MATHEMATICS DISABILITY    | 15

Willburger, E., Fussenegger, B., Moll, K., Wood, G., & Landerl, K. (2008). skills in multiple domains of arithmetic learning. Developmental
Naming speed in dyslexia and dyscalculia. Learning and Individual Psychology, 53, 2304–­2318. https://doi.org/10.1037/dev00​0 0432
Differences, 18, 224–­236. https://doi.org/10.1016/j.lindif.2008.01.003
Woodcock, R. W., & Johnson, M. B. (1977). Woodcock-­Johnson
psycho-­educational battery. DLM Teacher Resources. S U PP ORT I NG I N F OR M AT ION
Wu, X., Zhang, Y., Wu, R., & Chang, H. H. (2021). A comparative
Additional supporting information can be found online
study on cognitive diagnostic assessment of mathematical key
competencies and learning trajectories. Current Psychology, 41, in the Supporting Information section at the end of this
7854–­7866. https://doi.org/10.1007/s1214​4 -­020-­01230​- ­0 article.
Zhang, X., Koponen, T., Räsänen, P., Aunola, K., Lerkkanen, M. K.,
& Nurmi, J. E. (2014). Linguistic and spatial skills predict early
arithmetic development via counting sequence knowledge. Child
Development, 85, 1091–­1107. https://doi.org/10.1111/cdev.12173 How to cite this article: Ouyang, X., Zhang, X.,
Zhang, X., Räsänen, P., Koponen, T., Aunola, K., Lerkkanen, M.,
Räsänen, P., Koponen, T., & Lerkkanen, M.-K.
& Nurmi, J. (2020). Early cognitive precursors of children's
mathematics learning disability and persistent low achievement: (2022). Subtypes of mathematical learning
A 5-­year longitudinal study. Child Development, 91, 7–­27. https:// disability and their antecedents: A cognitive
doi.org/10.1111/cdev.13123 diagnostic approach. Child Development, 00, 1–15.
Zhang, X., Räsänen, P., Koponen, T., Aunola, K., Lerkkanen, https://doi.org/10.1111/cdev.13884
M. K., & Nurmi, J. E. (2017). Knowing, applying, and reasoning
about arithmetic: Roles of domain-­general and domain-­specific

You might also like