Research in Developmental Disabilities 35 (2014) 2906–2920

Contents lists available at ScienceDirect

Research in Developmental Disabilities

Identification of children with mathematics learning

disabilities (MLDs) using latent class growth analysis
Terry T.-Y. Wong a,b,*, Connie S.-H. Ho b, Joey Tang c
a
Department of Psychological Studies, The Hong Kong Institute of Education, Hong Kong
b
Department of Psychology, The University of Hong Kong, Hong Kong
c
Society for the Promotion of Hospice Care, Hong Kong

A R T I C L E I N F O A B S T R A C T

Article history: The traditional way of identifying children with mathematics learning disabilities (MLDs)
Received 9 April 2014 using the low-achievement method with one-off assessment suffers from several
Received in revised form 3 July 2014 limitations (e.g., arbitrary cutoff, measurement error, lacking consideration of growth).
Accepted 7 July 2014 The present study attempted to identify children with MLD using the latent growth
Available online modelling approach, which minimizes the above potential problems. Two hundred and
ten Chinese-speaking children were classified into five classes based on their arithmetic
Keywords: performance over 3 years. Their performance on various number-related cognitive
Numerical cognition
measures was also assessed. A potential MLD class was identified, which demonstrated
Mathematics learning disability
poor achievement over the 3 years and showed smaller improvement over time compared
Latent class growth analysis
Approximate number system
with the average-achieving class. This class had deficits in all number-related cognitive
skills, hence supporting the number sense deficit hypothesis. On the other hand, another
low-achieving class, which showed little improvement in arithmetic skills over time, was
also identified. This class had an average cognitive profile but a low SES. Interventions
should be provided to both low-achieving classes according to their needs.
ß 2014 Elsevier Ltd. All rights reserved.

1. Introduction

1.1. What is mathematics learning disability (MLD)?

Mathematics learning disability (MLD), or developmental dyscalculia (DD), has been traditionally described as a disorder
of arithmetic skills that reflects a discrepancy between one’s low arithmetical abilities compared with his/her overall
intelligence level and chronological age (Mazzocco & Räsänen, 2013). Approximately 7% of the population suffer from this
disability (Shalev, 2007). Although numeracy skills are found to influence a person’s educational, financial, and even health
status to a similar extent as reading skills (Parsons & Bynner, 2005), there are far fewer research studies on MLD than on
reading disabilities (Bishop, 2010; Gersten, Clarke, & Mazzocco, 2007). Because of the limited studies carried out on MLD,
there is little consensus on the definition (Kaufmann et al., 2013) and the cognitive profiles of children with MLD (Mazzocco,
2007).

* Corresponding author at: Room 19, 1/F, B4, The Hong Kong Institute of Education, 10 Lo Ping Road, Tai Po, New Territories, Hong Kong. Tel.: +852 9878 6870.
E-mail address: terrytywong@gmail.com (Terry T.-Y. Wong).

http://dx.doi.org/10.1016/j.ridd.2014.07.015
0891-4222/ß 2014 Elsevier Ltd. All rights reserved.
 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920 2907

1.2. The low-achievement method

Previously, the majority of the studies on MLD employed the low-achievement method of identifying children with MLD
(e.g., Mazzocco & Thompson, 2005; Ostad, 1997; Passolunghi & Siegel, 2001). Children are identified as having MLD if their
mathematics achievement scores below a certain percentile while their intellectual functioning falls within the normal
range. However, there is a large variation in the cutoff values used in different MLD studies. A recent study summarized the
latest MLD studies and found that the cutoff values employed in different studies ranged from the 5th to the 45th percentile
on standardized mathematics achievement tests (Murphy, Mazzocco, Hanich, & Early, 2007). Clearly, with such a huge
discrepancy in the cutoff values used, these studies usually identify different groups of children. It has been shown that the
use of different cutoff values results in different cognitive profiles of the MLD samples identified (Murphy et al., 2007). The
resulting inconsistency may hinder us from understanding the cognitive profiles of children with MLD.
To minimize the effect of the different cutoff values used, most recent studies used the tripartite method and identified
three groups of children according to their mathematics achievement (e.g., Cowan & Powell, 2014; De Smedt & Gilmore,
2011; Geary, Bailey, & Hoard, 2009; Mazzocco, Feigenson, & Halberda, 2011b; Murphy et al., 2007). Children who show the
most severe deficits in mathematics (e.g., lowest 10th percentile or 1.5 S.D. below the mean) are identified as MLD, while
those who show milder deficits in mathematics (e.g., between the 11th and 25th percentile or between 1 to 1.5 S.D. below
the mean) are identified as low-achieving (LA). The cognitive profiles of these two groups are then compared with their
normally achieving (NA) peers. Studies using the tripartite method usually found significant differences in cognitive profiles
between the MLD group and the other two groups (e.g., Cowan & Powell, 2014; Mazzocco, Feigenson, & Halberda, 2011a;
Mazzocco et al., 2011b; Murphy et al., 2007). With more stringent criteria set for the MLD group, we can be more confident
that the MLD group identified represents the true MLD population instead of containing some children who only have a mild
degree of difficulties. The chance of having false positives is reduced.
Although the low achievement method (with either one or two cutoff values) has been used in most of the studies on
MLD, there are several limitations that come along with this method. First, the cutoff values involved in the low-achievement
model are arbitrary. These cutoff values may be used because they have been commonly used in previous studies instead of
because they reflect something truly meaningful. A more important concern is that, as mentioned above, there is a huge
range of cutoff values used in previous studies, and this variation results in different samples of MLD children identified
(Murphy et al., 2007). Although the use of the tripartite approach has improved the situation slightly by identifying children
with different degrees of difficulties, the issue of arbitrariness is not eliminated. It is possible that children within one group
may be more different than the children who score around the cutoff boundaries but are classified into different groups. The
use of an arbitrary cutoff in identifying children with MLD therefore warrants concerns.
Second, when using the low-achievement method, some studies made the classification based on information from a
single time point (e.g., Chan & Ho, 2010; De Smedt & Gilmore, 2011; Landerl, Fussenegger, Moll, & Willburger, 2009). The
use of information from a single time point is also problematic because of both the measurement error as well as the lack
of consideration of children’s learning process. Due to measurement error, children who score around the cutoff value
may be classified as MLD in one time point while not in the other time point. In Mazzocco and Myers’s (2003) study, for
example, approximately 30% of the children being identified as MLD in one grade were no longer identified as MLD in
another grade. Classifying children as MLD based on a one-off assessment therefore results in substantial classification
error.
The lack of consideration of the learning process is another issue due to the classification of MLD with data from a single
time point. According to the Response-to-Intervention (RTI) approach of identifying learning disabilities, students who are at
risk for learning disabilities are those who are not able to learn at a reasonable rate and are unresponsive to suitable
interventions (Fletcher, Coulter, Reschly, & Vaughn, 2004). The rate of learning is therefore an important piece of information
in identifying learning disabilities that cannot be measured by any measurement from a single time point. Two previous
studies suggest the importance of incorporating growth in the identification of children with learning disabilities. In Morgan,
Farkas, and Wu’s (2009) study, for example, children who had persistent mathematics difficulties in kindergarten showed a
slower growth rate in mathematics in elementary school and were more likely to be identified as having MLD in elementary
school compared with those who showed only transient difficulties or no difficulties. Having persistent difficulties in
mathematics in kindergarten implies that these children have slow initial growth in mathematics, thus suggesting the
importance of growth in predicting later outcome. Evidence also comes from the field of reading. For example, in a group of
kindergarteners who participated in a kindergarten phonemic intervention programme, those who turned out to be children
with dyslexia later on were found to require a longer time to master the phonemic knowledge taught in the programme. On
top of that, the number of sessions required for a child to master phonemic skills outperformed the child’s post-intervention
phonemic skills in predicting their reading level at later ages (Byrne, Fielding-Barnsley, & Ashley, 2000). The usual practice of
identifying children with MLD based on the assessment results from a single time point, which ignores the development of
skills with time, therefore warrants cautions.

1.3. The latent class growth analysis

In light of the above limitations of the traditional approach (i.e., arbitrary cutoff, measurement error, lacking
consideration of growth), a new latent growth modelling approach of identifying children with learning disabilities has
2908 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920

recently been proposed (Boscardin, Muthén, Francis, & Baker, 2008). This approach clusters children into different groups
according to their achievement at multiple time points. Latent class growth analysis (LCGA) and growth mixture modelling
(GMM) are the two major statistical techniques under this approach (Jung & Wickrama, 2008), with the former assuming
that the variance and covariance estimates for the growth factors within each class are zero, resulting in a more
parsimonious model.
This approach has several advantages over the traditional approach. First, the classification is based on data instead of the
arbitrary cutoff points set by researchers. The classification may therefore reflect more genuine differences among the
participants and may therefore be more meaningful. Second, the situation in which the participant is misclassified as MLD
due to unstable environmental factors (such as being sick during the testing) can be minimized when the results of several
assessments are taken into account. Third, this approach also considers the growth in children’s achievement, which is
another useful indicator of children’s learning disability status.
To demonstrate the reliability of the latent growth modelling approach of classification of learning disabilities,
Boscardin et al. (2008) employed GMM to classify children into different classes based on their development of
phonological awareness in kindergarten and their reading development in primary school. It was found that the
classification based on kindergarten data was highly predictive of their later reading status, with a 100% correspondence
between the learning disability status based on the kindergarten and the elementary school data. This group of children
showed worse performance in both phonological awareness and reading throughout the assessment period, with little
improvement with time, which matched well with our expectation of how typical individuals with dyslexia behaved. The
latent growth modelling approach of identifying children with learning disabilities therefore provides a possible
alternative for identifying children with MLD. However, to our knowledge, only one study attempted to identify children
with MLD using this approach (Geary, Bailey, Littlefield, et al., 2009). The present study therefore attempted to fill this gap
by identifying children with MLD based on their development in arithmetic skills over a period of 3 years. The cognitive
profiles of this group of children were also examined.

1.4. Number-specific cognitive measures

1.4.1. Approximate number system

The approximate number system (ANS) is one of the core systems that allows humans to represent numerosity
nonsymbolically (Feigenson, Dehaene, & Spelke, 2004). The ANS has been found right at birth (Izard, Sann, Spelke, & Streri,
2009) and its precision improves with age (Halberda, Ly, Wilmer, Naiman, & Germine, 2012; Piazza et al., 2010). Neurological
models suggest that the ANS encodes numerosity in an approximate manner along a logarithmically compressed mental
number line (Dehaene, Piazza, Pinel, & Cohen, 2003; Piazza, Izard, Pinel, Le Bihan, & Dehaene, 2004). The approximate nature
of the ANS means that it allows us to discriminate two numerosities only when the ratio between them is large enough. The
number acuity, or the precision of the ANS representation of numerosity, is reflected by the finest ratio between two
numerosities that can be reliably distinguished (i.e., the Weber’s fraction, w).
The discovery of the ANS has stimulated the field of mathematical cognition. Many attempts have been made to
investigate the relationship between this innate, nonsymbolic ‘number sense’ and humans’ symbolic mathematical skills.
The earliest support for the relationship between the two comes from Halberda, Mazzocco, and Feigenson’s (2008) study,
which reported a significant relationship between the number acuity of 14-year-old adolescents and their previous
mathematics achievement. Later on, more studies verified the above relationship (e.g., Bonny & Lourenco, 2013; Libertus,
Feigenson, & Halberda, 2013; Lonnemann, Linkersdörfer, Hasselhorn, & Lindberg, 2013; Wong, Ho, & Tang, under review b;
but see Smedt, Noël, Gilmore, & Ansari, 2013 for a summary of studies that did not find a significant relationship between the
two). The relationship between the ANS and humans’ mathematical skills has also been supported by the studies that found
ANS deficits among children with MLD (Landerl et al., 2009; Mazzocco et al., 2011a, 2011b; Wong, Ho, & Tang, under review
a; but see De Smedt & Gilmore, 2011; Rousselle & Noël, 2007 for the null findings).

1.4.2. Number–numerosity mapping

Number–numerosity mapping has been proposed to be the connecting step between humans’ number acuity and their
understanding of the number system (Geary, 2013). It can therefore be perceived as the mechanism of how the ANS
influences our mathematical skills. Number–numerosity mapping has been measured by counting and estimation tasks in
previous studies (Mejias, Mussolin, Rousselle, Grégoire, & Noël, 2012; Wong et al., under review b). Both types of tasks
require the participants to associate a numerosity with a number symbol.
Counting can be thought of as the first systematic way for children to learn about number symbols. It allows children to
learn the number symbols by associating these symbols to numerosities. Counting skills also serve as the foundation for
children to learn arithmetic. Before children are able to directly retrieve the answers to arithmetic problems, they solve basic
arithmetic through counting. Even after they are able to directly retrieve answers for arithmetic problems, they sometimes
fall back on counting as a back-up strategy. Because of the above reasons, children’s counting skills have been found to
correlate with their mathematics achievement (e.g., Aunio & Niemivirta, 2010; Reigosa-Crespo et al., 2012). On the other
hand, poor mathematics and poor counting usually go together. Children with MLD have been found to count more slowly
than their peers (Landerl, Bevan, & Butterworth, 2004), and children who count more slowly also perform worse in
arithmetic (Reeve, Reynolds, Humberstone, & Butterworth, 2012).
 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920 2909

Estimation is another process in which humans associate a number symbol with a magnitude, such as a numerosity or a
length, without the use of real-world knowledge (Booth & Siegler, 2006). Examples of estimation include estimating the
number of candies in a bottle and the position of 38 on a 0–100 number line. The process of associating a magnitude with a
number symbol is similar to the proposed neurological model in which the magnitude of a number is being represented on a
mental number line (Dehaene et al., 2003). This resemblance may explain why children’s estimation skills are strongly
correlated with their mathematics achievement (e.g., Ashcraft & Moore, 2012; Booth & Siegler, 2006; Fuchs et al., 2010) and
why children (Geary, Hoard, Nugent, & Byrd-Craven, 2008; Mejias, Mussolin, et al., 2012) and adults (Mejias, Grégoire, &
Noël, 2012) with MLD tend to perform more poorly in estimation tasks.

1.4.3. Symbolic number processing

The ability to compare the numerical values of numerals has been proposed to be one of the basic numeracy skills
(Reigosa-Crespo et al., 2012), which are thought to represent how well the number symbols are being associated to the
numerosity representation (De Smedt & Gilmore, 2011; Rousselle & Noël, 2007). The efficiency of comparing numbers has
been found to correlate with one’s mathematics achievement (see De Smedt et al., 2013 for a review), and children with MLD
tend to be slower in this task (De Smedt & Gilmore, 2011; Rousselle & Noël, 2007). The performance in the number
comparison task is one of the defining criteria of developmental dyscalculia in the Dyscalculia Screener developed by
Butterworth (2003).

1.5. Number sense deficit vs. access deficit hypothesis

By examining the cognitive profiles of children with MLD, we could, at the same time, verify the two major hypotheses
concerning the major deficits among children with MLD. These two hypotheses are the number sense deficit hypothesis
(Wilson & Dehaene, 2007) and the access deficit hypothesis (Rousselle & Noël, 2007). While the former suggests that the
deficits among children with MLD originates from a deficit in the number sense (i.e., ANS) that affects all other basic
numeracy skills and hence one’s mathematics achievement, the latter suggests that the deficit causing MLD lies in the
inability to access numerosity information from number symbols, leaving the number sense intact. The current literature
reports mixed findings of the two hypotheses (supporting number sense deficit: Landerl et al., 2009; Mazzocco et al., 2011a,
2011b; Piazza et al., 2010; supporting access deficit: De Smedt & Gilmore, 2011; Rousselle & Noël, 2007); it is therefore
unclear whether difficulties experienced by those with MLD are originated from a defective number sense or impaired
access. Both hypotheses predict deficits in tasks that involve number symbols among children with MLD. If a deficit in
number acuity is found among children with MLD, then the number sense deficit is supported. Otherwise, the access deficit is
supported.

1.6. The current study

The current study aimed to identify children with MLD using the latent growth modelling approach. LCGA was conducted
to classify participants into different groups according to their arithmetic performance over a period of 3 years. The class that
showed the worst performance over the three-year period would be the potential MLD candidates. The cognitive profiles of
this class were compared with those of the other classes so that the cognitive deficits of this class of children could be
identified. The number sense deficit hypothesis and the access deficit hypothesis could therefore be compared. A reading
measure was also included to check if children in different classes also differed in their reading performance.

2. Method

2.1. Participants

Two hundred ten participants (52.4% boys) participated in the present study. They were recruited from 17 different
kindergartens in Hong Kong. Hong Kong children spend 3 years in kindergartens from the age of three. All of the participants
were Chinese-speaking children, with a mean age of 6 years, 1 month (S.D. = 4 months) in the first wave of assessment. They
were in the third year of kindergarten during the first wave of assessment. Due to attrition, the sample size dropped to 154 in
the fourth wave of assessment (see Table 2 for the sample sizes at each time point).

2.2. Measures

2.2.1. Arithmetic
The arithmetic task, which was designed based on the local mathematics curriculum, consisted of 30 arithmetic items,
ranging from single-digit addition to three-digit subtraction to mixture problems (addition and subtraction). With the
development in participants’ arithmetic skills, we included 20 extra items in Time 3 and 4, resulting in a total of 50 items. The
added items included multiplication, division, and mixture problems of the four operations (see Appendix 1 for the items).
The task had adequate internal reliability (Cronbach’s a > 75) and external validity (correlated with the Learning and
Achievement Measurement Kit 2.0, a locally normed achievement test, at 0.4 or above).
2910 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920

2.2.2. Number-related cognitive measures

All of the number-related cognitive measures were conducted on computer.

2.2.2.1. Approximate number system (ANS). The ANS of the participants was assessed using the nonsymbolic comparison task
and the nonsymbolic multiplication task. The nonsymbolic comparison task was adapted from Piazza et al. (2004), in which
the participants were asked to compare the two arrays of dots presented and to decide on the array containing more dots
without counting. There were always 16 dots in one of the arrays, while the other array consisted of 10–22 dots. The ratio of
dots involved ranged from 1.0625 to 1.60. Dots were presented in varying dot sizes (diameter ranging from 6 to 58 pixels).
The average dot size was directly proportional to the numerosity in half of the trials, while the relationship was reversed in
the other half. There were five practice trials followed by 50 experimental trials. Both the accuracy and the Weber’s fraction
(w) were used in the analyses.
The nonsymbolic multiplication task was modified from the nonsymbolic addition task by Gilmore, McCarthy, and Spelke
(2010). In each item, participants first saw an array of dots being covered by a shade. The shade then multiplied itself by a
factor of 2, 3, or 4, indicating that the array of dots being covered was multiplied by that factor. Finally, another array
appeared on the other side of the screen, and the participants had to decide which array (the product vs. the comparison)
contained more dots. Dots were represented in varying dot sizes (diameter ranging from 6 to 42 pixels). The average dot size
was directly proportional to the numerosity in half of the trials, while inversely proportional to the numerosity in the other
half. A PowerPoint animation was first shown to the participants to familiarize them with the task. There were four practice
items followed by 24 experimental items.

2.2.2.2. Number–numerosity mapping. The precision of mapping between number symbols and numerosity was assessed
using counting and estimation tasks.
The counting skills of the participants were assessed using the dot–number matching and dot–dot matching tasks. The
dot–number matching task was adapted from Butterworth’s (2003) Dyscalculia Screener. Participants saw an Arabic
numeral on one side of the screen and an array of dots on the other side, and their task was to decide whether the numerical
value of the numeral matched the number of dots. The dot–dot matching was similar to the dot–number matching, except
that the participants saw two arrays of dots this time and they had to decide whether the two arrays were identical in
number. The time limit was 8 s for both tasks. The numerosity involved in both tasks ranged from one to nine. There were
four practice items followed by 36 experimental items in both tasks.
Participants’ estimation skills were assessed using the numerosity production task and the number line estimation task. The
numerosity production task was adapted from Crollen, Castronovo, and Seron (2011). Participants were first shown an array of
100 dots in hardcopy, and they were told about the numerosity. Afterwards, an Arabic numeral appeared on the computer
screen, and the participants’ task was to produce that number of dots by pressing certain keys (e.g., ‘‘J’’ for producing dots and ‘‘F’’
for deleting dots). Because this was an estimation task, participants were told to press the keys so that they had to estimate,
instead of count, the number of dots produced. The 100-dot hardcopy remained beside the computer so that they could refer to
it. Dots were presented in varying sizes (diameter ranging from 6 to 82 pixels). In half of the trials, the average dot size was held
constant so that the total occupied area was directly related to the target numerosity. In the other half of the trials, the total
occupied area was held constant so that the average dot size was inversely related to the target numerosity. There were three
practice trials (numerosities of 3, 5, and 7) followed by 20 experimental trials (numerosity ranging from 10 to 98).
The number line estimation task was adapted from Siegler and Booth’s (2004) study. Participants were shown a number
line with 0 and 100 on the two ends. An Arabic numeral was shown above the number line, and the participants were asked
to locate that numeral on the correct position on the number line. The time limit was 30 s. There were three practice trials
(50, 25, and 75) preceding the 24 experimental trials. Due to the ceiling effect observed in Time 3, the number range in Time 4
was changed to 0–1000 instead. For both estimation tasks, linearity, or the fit towards the best-fit linear model, was taken as
the indicator of participants’ estimation skills.

2.2.2.3. Symbolic number processing. The number comparison task from the Dyscalculia Screener (Butterworth, 2003) was
used to assess children’s symbolic number processing skills. Two Arabic numerals were presented on the computer screen
each time, and the participants had to decide which numeral had a greater numerical value. The numerals involved ranged
from 1 to 9, with a numerical distance between one and four. There were four practice trials and 36 experimental trials. Both
accuracy and reaction time (considering only the correct trials) were analyzed.

2.2.3. Control measures

2.2.3.1. Nonverbal intelligence. Participants’ nonverbal intelligence was assessed using the short form of Raven’s Standard
Progressive Matrices (Raven, 1956). In each item, there was a pattern with a missing piece. Participants had to choose from
among six to eight pieces a piece that could complete the pattern. A total of 36 items were involved in the short form. The raw
scores were then converted to scaled scores based on the local norm.

2.2.3.2. Working memory. The working memory capacity of the participants was assessed by the backward digit span task.
Participants were verbally presented with a sequence of digits at a rate of one per second. After listening, they had to recall
 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920 2911

the digit in a backward manner. The recalled sequence had to be completely correct in order to score one mark. There were
seven levels in the task, with two items in each level. Starting with a level of two digits, each successive level had one more
digit to be recalled. The task would be terminated when the participant failed both items in a level. Two practice trials were
given beforehand to ensure that the participants understood the task.

2.2.4. Reading
The reading achievement of the participants was assessed using the Chinese word reading subtest of the Hong Kong Test
of Specific Learning Difficulties in Reading and Writing for Primary School Students – Second Edition, or HKT-P(II) (Ho et al.,
2007). The HKT-P(II) is a standardized diagnostic tool for identifying children with dyslexia. The Chinese word reading
subtest consists of 150 two-character Chinese words in ascending difficulty. Participants were asked to read aloud the words.
Each correctly read word yielded one mark. The test would be discontinued if the participants failed to get any marks in 15
consecutive items. The raw scores were converted into the standardized scores based on the local norm.

2.3. Procedures

A total of 17 kindergartens participated in the present study, with a total sample of 210 recruited initially. Testing was
conducted in the kindergartens in the first wave of assessment, while the following waves of testing were conducted in
participants’ homes. While the cognitive and arithmetic tasks were conducted in all waves of assessment, the working memory,
nonverbal intelligence, and reading measures were conducted in one of the waves only (the first, the second, and the fourth
wave, respectively). Waves 2 and 3 were conducted half a year after the previous wave, and wave 4 was one year after wave 3.
Each assessment session lasted for approximately 2 h. All of the testing was conducted either by the first author or by trained
experimenters who were psychology undergraduates. Souvenirs were given to the participants after each testing.

3. Results

3.1. The latent class growth model

Participants were first classified into different classes by the LCGA (Jung & Wickrama, 2008) according to their arithmetic
performance across the four time points using Mplus, Version 6.0 (Muthén & Muthén, 2011). The number of classes in the
final solution was determined based on both the interpretability of the class structure and the model fit indices. The model fit
indices of two- to six-class models are presented in Table 1. While the four-class solution obtained the lowest Bayesian
Information Criterion (BIC) value, the p-value of the bootstrapped likelihood ratio test (BLRT) suggested the five-class
solution. Because it has been suggested that the BLRT performs best in selecting models (Nylund, Asparouhov, & Muthén,
2007), we followed this criterion and selected the five-class solution. This five-class solution, with adequate sample size in
the smallest group and distinct profiles in different classes, was also interpretable. The entropy of the five-class model was
0.736, with the values on the diagonal of the posterior probability matrix (indicating the distinctness of different classes)
ranging from 0.725 to 0.964.
The classes in the five-class solution were labelled according to their performance in arithmetic. The top class (n = 14,
6.7%) had the highest intercept and obtained the highest scores throughout the four time points. The high class (n = 49, 23.3%)

Table 1
Fit indices of models with different numbers of classes.

No. of classes BIC LMR p-value BLRT p-value Entropy

2 4287.151 0.0001 0.0000 0.835

3 4244.829 0.0379 0.0000 0.760
4 4225.585 0.0197 0.0000 0.732
5 4227.255 0.1708 0.0206 0.736
6 4235.296 0.4081 0.1500 0.758

Note. BIC, Bayesian information criterion; LMR, Lo–Mendell–Rubin; BLRT, bootstrapped likelihood ratio test. The row in boldface indicates the final model
selected.

Table 2
Sample sizes in the four waves of assessment.

Time Class Total N

Top (1) High (2) Average (3) Low (4) Bottom (5)

1 14 49 109 17 21 210
2 14 43 85 17 20 179
3 12 43 77 17 19 168
4 12 41 69 17 15 154
[(Fig._1)TD$IG]
2912 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920

50
Arithmetic performance of different classes
45

40

35

Arithmetic (raw score)

30

25

20

15 Top

High
10
Average

Low
5
Bottom (MLD)

0
1 2 3 4
Phase

Fig. 1. Arithmetic performance of different classes. Error bars indicate the standard errors.

had a slightly lower intercept than the top class and always ranked second in terms of arithmetic performance. The average
class (n = 109, 51.9%) had a relatively low intercept, but they showed the greatest improvement across time, catching up with
the high class at Time 4. The low class (n = 17, 8.1%) had an average intercept, but they showed the slowest growth, with the
arithmetic performance at Time 4 being similar to that of the bottom class (n = 21, 10%), whose performance remained the
worst across all four time points. Given their consistently poor performance in arithmetic, it was likely that this bottom class
represented the MLD population (see Fig. 1).

3.2. Class comparisons

After classifying the participants into different classes, the cognitive profiles of the different classes were compared. The
class comparisons included only those participants who remained in the final sample at Time 4. Outliers (3 or more S.D.s
beyond the class means) were deleted before the analyses. Class identity was the independent variable in all of the analyses
below. First, the class classification was validated by examining the class difference in arithmetic performance using
repeated-measures ANOVA. Secondly, performances on the control measures (i.e., nonverbal intelligence and working
memory) were compared among the classes using ANOVA. Third, the class differences in number-related cognitive
capacities were examined using repeated-measures ANCOVA with the effects of working memory and intelligence being
controlled. Finally, an ANCOVA was used to examine whether there were significant class differences in reading
achievement, again controlling for the effects of working memory and intelligence.
Given the complexity of the analyses, the focus of the repeated-measures analyses was on the between-subjects effects.
For the comparisons with unequal variance in any of the time points involved (i.e., the ratio of the largest group variance to
the smallest group variance > 2), a critical alpha value of 0.01 was used (Tabachnick & Fidell, 2006). Although this happened
in most of our analyses, our conclusions were not affected at all because all class comparisons were highly significant (i.e.,
p < 0.01, except for reading). All of the significant between-subjects comparisons were followed up by post hoc analyses with
Bonferroni adjustment.

3.2.1. Arithmetic
The repeated-measures ANOVA on the class difference in arithmetic performance showed highly significant results,
F(4,149) = 190.264, p < 0.001, h2p ¼ 0:836. The large effect size suggested that more than 80% of the variance in arithmetic
performance was captured by the class membership, thereby validating the latent class classification. The post hoc analyses
suggested the best performance in the top class, followed by the high class. Both the average and low classes ranked next,
while the two did not differ significantly. Both classes scored significantly higher than the bottom (MLD) class. All of the
significant comparisons were significant at p < 0.001 (see Table 3).
Another ANOVA was conducted to examine whether the classes also differed in terms of their growth in arithmetic
performance. The class comparison was highly significant, F(4,149) = 309.448, p < 0.001, h2p ¼ 0:893, and the post hoc
Table 3
Class differences in age, arithmetic, domain-general cognitive measures, and reading achievement.

T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920

Measures Analysisa Timeb Performance of the classes Between-subjects effect

Top (1) High (2) Average (3) Low (4) Bottom (5) F h2p Post hoc
M (SD) M (SD) M (SD) M (SD) M (SD)

Age Repeated- 1 73.083 (3.059) 73.220 (3.831) 72.841 (3.760) 72.471 (3.502) 72.813 (6.058) F(4,150) = 0.150 0.004 NA
measures
ANOVA
2 78.417 (2.906) 78.268 (3.742) 78.316 (4.024) 77.824 (4.419) 78.188 (5.764)
3 85.250 (2.989) 85.220 (3.752) 84.870 (3.808) 84.118 (3.533) 84.625 (5.898)
4 97.417 (3.059) 97.317 (3.731) 96.971 (3.722) 96.353 (3.534) 96.813 (6.134)
Arithmetic Repeated- 1 22.417 (4.055) 14.902 (3.064) 8.565 (2.723) 11.059 (2.076) 4.067 (1.710) F(4,149) = 190.264** 0.836 1>2>3=4>5
measures
ANOVA
2 25.417 (3.030) 19.927 (3.601) 13.101 (3.618) 15.294 (1.829) 9.800 (3.590)
3 36.750 (5.562) 27.537 (5.604) 22.725 (4.724) 19.706 (3.274) 14.800 (3.098)
4 43.417 (2.353) 38.732 (2.470) 37.464 (2.041) 30.706 (2.519) 30.667 (2.193)
Nonverbal intelligence ANOVA 2 120.250 (15.130) 116.146 (11.968) 109.412 (12.563) 108.706 (8.206) 100.625 (7.293) F(4,149) = 7.371** 0.165 1,2 > 3,5
Working memory ANOVA 1 5.917 (2.235) 5.854 (1.174) 5.162 (1.101) 5.647 (1.169) 4.313 (1.014) F(4,149) = 5.609** 0.131 1,2,4 > 5
2 > 3+
Reading ANCOVA 4 13.333 (2.103) 13.342 (2.394) 10.971 (3.278) 11.118 (3.039) 10.800 (3.406) F(4,146) = 3.311* 0.083 2>3
a
Depending on the variable, different types of ANOVA were conducted (single vs. multiple time points, with vs. without covariates).
b
The time point in which the measure was conducted.
+
p < 0.07.
* p < 0.05.
** p < 0.001.

2913
2914 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920

analyses suggested that the growth rate was highest for the average class, followed by the bottom (MLD) class (p < 0.001)
and then the high class (p = 0.005). Both the top class and the low class showed significantly slower growth than the high
class (ps < 0.001), while the two classes did not differ significantly in terms of growth rate (p = 0.256).

3.2.2. Control measures

The class differences in nonverbal intelligence and working memory were analyzed using ANOVA. Both analyses yielded
significant results: F(4,149) = 7.371, p < 0.001, h2p ¼ 0:165 for nonverbal intelligence; F(4, 149) = 5.609, p < 0.001, h2p ¼ 0:131
for working memory. Post hoc analyses suggested that the top class and the high class did not differ in terms of nonverbal
intelligence, with both scoring significantly higher than the average class (ps = 0.038 and 0.044, respectively) and the bottom
class (ps < 0.001). No other class difference in terms of nonverbal intelligence was significant. For working memory, the post
hoc analyses suggested significantly higher working memory capacities among the top class, the high class, and the low class
compared with the bottom (MLD) class (ps = 0.010, 0.001, and 0.027, respectively). The high class also had marginally
significantly higher working memory capacity than the average class (p = 0.060). No other class comparisons were significant
(see Table 3).

3.2.3. Number-related cognitive measures

The class differences in the number-related cognitive capacities (i.e., ANS, symbolic–nonsymbolic mapping, and symbolic
number processing) were analyzed using repeated-measures ANCOVA, with the effects of nonverbal intelligence and
working memory being controlled for. Details of the analyses are shown in Table 4.

3.2.3.1. Approximate number system. The number acuity of the participants was reflected by three indices: accuracy and w in
nonsymbolic comparison and accuracy in nonsymbolic multiplication. For accuracy in nonsymbolic comparison, the
ANCOVA suggested a significant between-subjects effect, F(4,147) = 3.728, p = 0.006, h2p ¼ 0:92. The post hoc analyses
suggested that the bottom (MLD) class scored significantly lower than the top and the high classes (p = 0.036 and 0.002,
respectively) and marginally significantly lower than the average class (p = 0.057) in this task. No other class comparisons
were significant. The analyses on w also yielded significant findings, F(4,136) = 8.370, p < 0.001, h2p ¼ 0:198. The post hoc
analyses suggested that the bottom (MLD) class had significantly higher w (indicating worse performance) than all other
classes (p = 0.001 for the comparison with the top class, ps < 0.001 for the comparisons with the other classes), while the w of
these four classes did not differ significantly from one another. The results seemed to suggest a specific deficit in w among the
bottom (MLD) class.
The repeated-measures ANCOVA on the accuracy on the nonsymbolic multiplication task also yielded a significant
between-subjects effect, F(4,146) = 7.011, p < 0.001, h2p ¼ 0:161. The post hoc analyses suggested that the performance in
this task divided the classes into two major groups. The first group (consisting of the top and the high classes) scored
significantly higher than the second group (consisting of the average, the low, and the bottom classes), with no intra-group
differences being significant (ps < 0.04 for all significant comparisons).

3.2.3.2. Number–numerosity mapping. Participants’ performance in number–numerosity mapping was indicated by two
major skills (counting and estimation), each measured by two indices (accuracy in dot–number and dot–dot matching for
counting, linearity in numerosity production and number line for estimation). For counting, the repeated-measures ANCOVA
suggested a significant between-subjects effect in the accuracy of dot–number matching, F(4,142) = 10.366, p < 0.001,
h2p ¼ 0:226. The post hoc analyses revealed that both the top and the high class counted significantly more accurately than
did the average class (p = 0.012 and p < 0.001, respectively), which also counted significantly more accurately than the
bottom (MLD) class (p = 0.019). The low class also counted significantly more accurately than the bottom (MLD) class
(p = 0.005), although their performance did not differ from the other classes. The findings from the dot–dot matching task
were similar, F(4,146) = 11.975, p < 0.001, h2p ¼ 0:247, with an almost identical pattern in the post hoc analyses (i.e.,
Top = High > Average [p = 0.012 and p < 0.001, respectively] > Bottom [p = 0.007], Low > Bottom [p = 0.035]), except that the
low class also counted significantly less accurately than the high class (p = 0.023).
Significant class differences were found in the linearity in both estimation tasks: F(4,146) = 14.277, p < 0.001, h2p ¼ 0:281
for numerosity production; F(4,141) = 10.745, p < 0.001, h2p ¼ 0:234 for number line. The post hoc analyses on the
numerosity production task suggested that the top and the high classes had significantly higher linearity than the average
class (p = 0.006 and p < 0.001, respectively), which significantly outperformed the bottom (MLD) class (p = 0.002). The high
class also significantly outperformed the low class (p = 0.007), which further outperformed the bottom (MLD) class
significantly (p = 0.014). Again, the findings were similar for the number line task (Top = High > Average [p = 0.041 and .006,
respectively] > Bottom [p < 0.001]), except that the performance of the low class no longer differed significantly from that of
the high class. Yet, the contrast between the low class’s and the bottom (MLD) class’s performance remained significant
(p = 0.002).

3.2.3.3. Symbolic number processing. The number comparison task was used to assess participants’ symbolic number
processing. Both accuracy and reaction time were analyzed. The repeated-measures ANCOVA on accuracy suggested a
significant between-subjects effect, F(4,143) = 5.924, p < 0.001, h2p ¼ 0:142, and the post hoc analyses suggested a specific
deficit among the bottom (MLD) class (ps = 0.001 for the comparisons with the top and high classes, p < 0.001 for the
Table 4
Class differences in domain-specific cognitive abilities examined by repeated-measures ANCOVA.

Measures Timea Performance of the classes Between-subjects effect

Top (1) High (2) Average (3) Low (4) Bottom (5) F h2p Post hoc
M (SD) M (SD) M (SD) M (SD) M (SD)

Nonsymbolic comparison 1 35.750 (4.434) 37.707 (4.149) 35.855 (4.074) 37.813 (2.949) 33.000 (3.327) F(4,147) = 3.728* 0.092 1,2 > 5
3 > 5+
2 39.500 (2.812) 37.561 (3.406) 37.609 (3.687) 37.375 (3.364) 35.750 (5.196)
3 38.750 (4.093) 38.195 (3.393) 36.884 (3.496) 35.938 (2.977) 35.000 (4.590)
4 39.167 (4.303) 39.171 (4.224) 37.768 (4.106) 38.000 (4.719) 35.563 (4.718)
Nonsymbolic comparison (w) 1 0.254 (0.132) 0.195 (0.104) 0.239 (0.121) 0.188 (0.051) 0.388 (0.165) F(4,136) = 8.370** 0.198 1,2,3,4 > 5

T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920

2 0.195 (0.078) 0.193 (0.079) 0.198 (0.087) 0.262 (0.141)
3 0.173 (0.086) 0.184 (0.083) 0.208 (0.082) 0.226 (0.060) 0.331 (0.251)
4 0.176 (0.102) 0.170 (0.088) 0.203 (0.100) 0.156 (0.070) 0.329 (0.332)
Nonsymbolic multiplication 1 15.917 (2.275) 15.293 (1.952) 14.597 (1.978) 14.177 (1.845) 13.563 (1.504) F(4,146) = 7.011** 0.161 1,2 > 3,4,5
2 16.833 (2.250) 15.756 (2.047) 15.448 (1.608) 15.177 (2.270) 14.688 (1.740)
3 17.583 (2.937) 17.854 (2.286) 16.239 (2.001) 16.353 (2.805) 15.250 (1.571)
4 18.833 (1.115) 18.244 (2.498) 17.134 (2.328) 16.882 (2.342) 15.875 (1.746)
Dot–number matching 1 32.750 (1.815) 33.077 (2.366) 30.687 (3.299) 31.267 (3.240) 28.875 (3.914) F(4,142) = 10.366** 0.226 1,2 > 3 > 5
4>5
2 33.917 (2.811) 33.282 (2.339) 31.776 (2.964) 32.667 (1.839) 30.750 (2.049)
3 34.917 (.900) 34.051 (2.025) 32.836 (2.403) 33.933 (2.017) 31.063 (3.255)
4 34.167 (2.082) 34.410 (1.650) 33.900 (1.653) 33.533 (1.506) 32.813 (3.250)
Dot–dot matching 1 28.000 (2.892) 26.675 (2.606) 24.145 (3.478) 26.188 (3.600) 22.938 (3.108) F(4,146) = 11.975** 0.247 1,2 > 3 > 5
2>4>5
2 28.000 (3.384) 27.525 (2.909) 26.130 (2.717) 26.000 (3.033) 24.063 (1.914)
3 27.917 (3.528) 28.925 (3.285) 27.015 (2.903) 27.875 (2.579) 24.563 (2.732)
4 31.750 (2.958) 31.775 (2.713) 29.275 (3.014) 27.313 (2.726) 26.688 (2.651)
Numerosity production 1 0.644 (0.205) 0.677 (0.144) 0.399 (0.233) 0.501 (0.249) 0.297 (0.254) F(4,146) = 14.277** 0.281 1,2 > 3 > 5
2>4>5
2 0.765 (0.158) 0.686 (0.148) 0.546 (0.195) 0.574 (0.176) 0.380 (0.191)
3 0.743 (0.142) 0.716 (0.148) 0.649 (0.177) 0.665 (0.141) 0.483 (0.172)
4 0.749 (0.188) 0.805 (0.091) 0.734 (0.146) 0.645 (0.202) 0.606 (0.172)
Number line 1 0.927 (0.048) 0.879 (0.085) 0.702 (0.239) 0.786 (0.172) 0.578 (0.241) F(4,141) = 10.745** 0.234 1,2,>3 > 5
4>5
2 0.927 (0.068) 0.915 (0.059) 0.866 (0.088) 0.830 (0.169) 0.735 (0.193)
3 0.963 (0.026) 0.946 (0.031) 0.911 (0.050) 0.904 (0.045) 0.794 (0.123)
4 0.949 (0.035) 0.926 (0.049) 0.900 (0.064) 0.899 (0.068) 0.810 (0.158)
Number comparison (Accuracy) 1 34.667 (0.888) 34.000 (1.569) 33.769 (1.684) 34.294 (1.448) 29.875 (6.217) F(4,143) = 5.924** 0.142 1,2,3,4 > 5
2 34.333 (1.723) 34.475 (1.320) 34.754 (1.031) 34.588 (1.502) 33.438 (1.590)
3 34.833 (1.403) 34.375 (1.462) 34.492 (1.174) 33.588 (1.583) 33.500 (2.556)
4 35.250 (1.357) 34.675 (1.248) 34.508 (1.226) 34.471 (1.463) 34.563 (1.153)
Number comparison (Reaction time) 1 1224.671 (412.010) 1276.866 (322.186) 1603.400 (380.594) 1680.607 (402.915) 2060.538 (476.487) F(4,141) = 10.721** 0.233 2 > 3,4 > 5
1>5
2 1086.166 (306.882) 1089.601 (291.210) 1360.495 (285.017) 1326.019 (505.776) 1618.371 (408.720)
3 936.952 (214.136) 904.967 (235.401) 1061.292 (213.535) 1057.206 (245.231) 1350.229 (517.944)
4 880.737 (245.164) 797.547 (181.421) 875.475 (163.404) 901.551 (128.150) 1018.964 (313.122)
a
The time point in which the measure was conducted.
+
p < 0.07.
* p < 0.01.

2915
** p < 0.001.
2916 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920

Table 5
Class differences in SES measures.

Measures Class F h2p Post hoc

Top (1) High (2) Average (3) Low (4) Bottom (5)

Father’s education 3.333 (0.985) 2.811 (1.126) 2.774 (0.895) 2.357 (0.497) 2.571 (0.852) F(4,134) = 1.936 0.055 NA
levela
**
Mother’s education 3.250 (1.055) 3.054 (0.999) 2.714 (0.831) 2.214 (0.699) 2.286 (0.726) F(4,135) = 4.382 0.115 1,2 > 4; 1,2 > 5+
levela
Family incomeb 4.667 (1.670) 4.737 (1.884) 4.129 (1.903) 3.267 (1.580) 4.286 (2.301) F(4,136) = 1.851 0.052 NA
SES compositec 1.340 (2.679) 0.582 (2.741) 0.080 (2.316) 0.709 (2.574) 1.713 (1.598) F(4,130) = 3.191* 0.089 1 > 4; 2 > 4+
a
For parental education level, 1 = primary, 2 = junior secondary, 3 = senior secondary, 4 = college, 5 = university or above.
b
For family income, 1 = <$9000, 2 = $9000–$12,000, 3 = $12,000–$16,000, 4 = $16,000–$20,000, 5 = $20,000–$24,000, 6 = $24,000–$30,000, 7 = >$30,000.
All are monthly family income in Hong Kong Dollars.
c
SES composite refers to the sum of the standardized scores of the three SES measures.
+
p < 0.07.
* p < 0.05.
** p < 0.01.

comparison with the average class, and p = 0.011 for the comparison with the low class). The analyses on reaction time
yielded a similar but larger between-subjects effect, F(4,141) = 10.721, p < 0.001, h2p ¼ 0:233, and a similar pattern in the
post hoc analyses (all classes significantly outperformed the bottom [MLD] class: ps < 0.001 for the comparisons with the top
and high classes, p = 0.001 for the comparison with the average class, and p = 0.036 for the comparison with the low class),
but this time, the high class also significantly outperformed the average class (p = 0.001) and the low class (p = 0.019).

3.2.4. Reading
The reading achievement of the classes was also compared using ANCOVA. The results were significant, but with a smaller
effect size, F(4,146) = 3.311, p = 0.013, h2p ¼ 0:83. The post hoc analyses suggested that the only significant class difference lay
between the high class and the average class, in which the high class did better (p = 0.01). This suggested that neither the low
class nor the bottom (MLD) class had difficulties in reading. The class classification was therefore arithmetic-specific (see
Table 3).

3.2.5. SES
In all of the above analyses, the low class did not seem to be particularly distinctive in any of the cognitive measures. The
reasons for their slow development in arithmetic therefore remained unclear. An attempt was made to compare the SES of the
classes to see if the SES measures differentiated the low class from the other classes. Three SES measures had been collected,
namely father’s and mother’s education level (1 = primary, 2 = junior secondary, 3 = senior secondary, 4 = college, 5 = university
or above) and family income (1 = <$9000, 2 = $9000–$12,000, 3 = $12,000–$16,000, 4 = $16,000–$20,000, 5 = $20,000–$24,000,
6 = $24,000–$30,000, 7 = >$30,000, all in Hong Kong Dollars). The results indicated that this class scored the lowest on all three
SES measures, although the differences did not reach statistical significance except for mother’s education, F(4,135) = 4.382,
p < 0.01, h2p ¼ 0:115 (see Table 5). The standard scores of the above SES measures were then summed up to obtain a composite
SES score. The analysis on the composite SES score suggested that this low-achieving class obtained a standard score of 1.713
compared with 1.340, 0.582, 0.080, and 0.709 in the top, high, average, and bottom classes, respectively. The between-group
comparison turned out to be significant, F(4,130) = 3.191, p = 0.015, h2p ¼ 0:89, with the low class having significantly and
marginally significantly lower SES than the top and high achieving classes, respectively.

4. Discussion

In the present study, a large sample of kindergarteners was assessed on their arithmetic skills four times over a three-year
period. An LCGA was conducted to classify the participants according to their arithmetic performance over the four time
points. A five-class solution emerged, with the lowest performing class being the potential candidates for having MLD.
Further analyses of class differences on number-related cognitive measures suggested that this bottom class showed deficits
in most number-related cognitive measures (except for accuracy in nonsymbolic multiplication) compared with the
average-performing class. Their performances on domain-general cognitive measures and reading, however, were
comparable to those of the average class. On the other hand, the LCGA also identified another class of low arithmetic
performers, who had average initial arithmetic scores but showed the smallest improvement over the 3 years. The cognitive
profile of this class was not distinctive, but they tended to have lower SES. Theoretical and practical implications are
discussed in the following sections.

4.1. The use of a latent growth modelling approach in identifying children with MLD

The present study was one of the few studies that employed the latent growth modelling approach in identifying children
with MLD. Participants were classified into five classes according to their arithmetic performance over 3 years. This approach
 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920 2917

has several advantages over the traditional approach. First, it reduces the likelihood that children are classified into the
wrong group due to an arbitrary cutoff (Boscardin et al., 2008). Secondly, taking multiple measures of achievement into
account minimizes the misclassification rate due to measurement error, resulting in higher predictive power (Morgan et al.,
2009). Third, the growth of children’s mathematical skills has been taken into account. Previous studies in reading suggested
that the rate of learning of phonemic skills was more predictive of children’s later reading achievement than the end-of-
intervention measure of phonemic skills (Byrne et al., 2000). The same logic may also apply to the case of mathematics
learning. As discussed below, the various aspects of the present findings suggested that the latent growth modelling
approach provides a possible alternative for identifying children with MLD.
First, the present study identified 10% of the participants as having MLD. The prevalence rate is within the range of 3.4% to
13.8% obtained in previous studies (see Shalev, 2007 for a summary) that employed an arbitrary cutoff to identify children
with MLD. Although a completely different way of identifying MLD was used in the present study, the similarity in
prevalence rates obtained across different methods cross-validates each other and suggests that the true prevalence rate
may be within this range.
Secondly, the growth of the MLD class matched well with the theoretical expectations and previous findings. According to
the RTI model, children who are at risk of learning disabilities are those who fail to learn at a reasonable rate and are not
responsive to intervention (Fletcher et al., 2004). Previous studies also suggested that children with MLD showed a shallower
growth rate compared with their low-achieving and normally achieving peers (Geary, Bailey, Littlefield, et al., 2009; Morgan
et al., 2009; Murphy et al., 2007). The present study echoed the previous studies and found a shallower growth rate in
arithmetic among the MLD class compared with the average class. Although the growth rates of the top and high classes were
also outperformed by the average class, their restricted growth rates could be a result of a ceiling effect, which certainly did
not apply to the MLD class. These converging results suggest that children with MLD not only show a performance lag
compared with their normally achieving peers but also show that this performance lag tends to widen with time.
Third and most importantly, the MLD class identified showed deficits in almost all of the number-related cognitive
measures in the present study, including the ANS, number–numerosity mapping, and symbolic processing of numbers. The
deficits in these skills could not be a result of individual differences in general intelligence or working memory capacity,
which had been controlled for. Furthermore, the MLD class showed comparable performance with other classes in reading,
suggesting that the difficulties of this class are specifically in mathematics. Children with MLD have been shown to have
deficits in the above domains in previous studies (ANS: Landerl et al., 2009; Mazzocco et al., 2011a, 2011b; Piazza et al., 2010;
number–numerosity mapping: Geary et al., 2008; Landerl et al., 2004; Mejias, Mussolin et al., 2012; symbolic processing of
numbers: De Smedt & Gilmore, 2011; Rousselle & Noël, 2007). The current findings therefore converge with those of previous
studies and suggest that the core deficits of MLD may lie in the above-mentioned skills.
In fact, the presence of the above-mentioned deficits might have contributed to the shallower learning curve observed in the
MLD class. According to Butterworth, Varma, and Laurillard (2011), dyscalculia, or MLD, can be perceived as a disability arises
from the cognitive deficits in numerical processing. Because of these deficits, children with MLD may have difficulties even in
basic processing of numbers, as oppose to the automatic processing of numbers among normally achieving children. These
difficulties may result in under-achievement in mathematics either directly or indirectly (e.g., by distracting them from key
elements in the mathematical tasks). With the increasing complexity of mathematical tasks across grades, the situation could
only get worse. Therefore, unless these basic numerical processing skills have been improved, or compensatory mechanisms
have been developed, these deficits are expected to influence children’s learning throughout their lifetime. An increased
performance gap between children with MLD and their normally achieving peers is therefore expected.

4.2. Number sense deficit vs. access deficit

The present study also provided us with an opportunity to verify whether the core deficit of MLD lies in the number sense,
as predicted by the number sense deficit hypothesis (Wilson & Dehaene, 2007), or whether it originates from a deficit in
accessing numerosity information from number symbols, as predicted by the access deficit hypothesis (Rousselle & Noël,
2007). The present findings suggest that, on top of the deficits in number–numerosity mapping and symbolic number
processing, the MLD class also showed a deficit in their number sense. The number sense deficit hypothesis is therefore
supported. The findings converge with other studies that found significant ANS impairment among children with MLD
(Landerl et al., 2009; Mazzocco et al., 2011a, 2011b; Piazza et al., 2010; Wong et al., under review a) and further confirm the
link between the ANS and our symbolic mathematical skills. The impaired number sense may influence the number–
numerosity mapping process, which then results in poor symbolic number processing and hence an under-achievement in
mathematics (Wong et al., under review b)

4.3. Interventions for children with MLD

Based on the above arguments, a logical way to help children with MLD is to design interventions targeting both number
sense and number–numerosity mapping skills. Some previous attempts have been successful. For example, by training
adults on some nonsymbolic arithmetic tasks, Park and Brannon (2013) successfully improved participants’ efficiency in
doing symbolic arithmetic. A similar training programme for children has also been found to be effective (Hyde, Khanum, &
Spelke, 2014). On the other hand, using a number line training game, Kucian et al. (2011) improved the arithmetic
2918 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920

performance of both children with MLD and their normally achieving peers. These successful attempts have provided
researchers, educators, and parents new hope in helping children with MLD.

4.4. The low-achieving class

On top of identifying the MLD class, the present study also identified another low-achieving class. This class of children
performed reasonably well in kindergarten, scoring slightly higher than the average class in arithmetic initially. However,
they showed the smallest growth rate among the five classes, and their arithmetic scores were caught up to by the MLD class
in Time 4. None of the cognitive skills measured in the present study seemed to distinguish this class from the better classes,
although it did outperform the MLD class in most domains. However, this class was found to have the lowest SES among the
five classes. The differences in SES between this class and the other classes were substantial, with the composite average SES
score being 1 S.D. below the mean of the second-lowest SES class (i.e., the MLD class). The low SES may explain the slow
arithmetic development among this class of children. The results converge with those of other studies that suggest that
children from lower SES families suffer from lower mathematics achievement and develop their mathematics skills more
slowly (Jordan & Kaplan, 2006; Morgan et al., 2009). The lack of parental involvement in learning among the low SES families
could be one of the reasons for the slow development among this low-achieving group. Given that the growth curve of the
low-achieving class was even worse than that of the MLD class, interventions should also target this group of children.
However, the intervention they need may differ from that of the MLD children. While children with MLD may need
interventions that target their cognitive skills, simple behavioural interventions such as peer tutoring and increasing
homework supervision may be helpful for the low-achieving class.
It should also be noted that the low-achieving class would not be identifiable using the low-achievement method alone as
it was the class’s growth rate, but not its achievement at any time point, that distinguished it from other classes. LCGA
therefore provides a way for educators to identify other groups of children who may fall behind in their learning, which
would have been ignored using the low-achievement method alone.

4.5. Limitations and future directions

The major limitation of the current study was the relatively small sample size. Given the relatively small initial sample
size, fewer than 20 children with MLD were identified in the final sample, which might limit the generalizability of the
findings. However, the patterns revealed by the class comparisons were clear, and the effect sizes were relatively large (all
except one comparison in the number-specific cognitive factors yielded medium or larger effect sizes). Furthermore, the
findings from the current study converged well with those obtained in other studies (e.g., Mazzocco et al., 2011a; Piazza
et al., 2010). All of these elements suggested that results of the current study were reliable. By increasing the initial sample
size, future studies may yield a larger MLD sample for further comparisons.

4.6. Conclusions

The present study attempted to identify children into different classes according to their arithmetic performance over 3
years. Using LCGA, a five-class structure was obtained, with the lowest performing class being the potential candidates for
having a MLD. This MLD class, which accounted for 10% of the sample, scored the worst in arithmetic throughout the
assessment period, with a shallower growth curve compared with the average class. They also showed deficits in all number-
related cognitive processing, including number sense, hence supporting the number sense deficit hypothesis. On the other
hand, another low-achieving class was also identified that had no observable deficits in cognitive processing but had a lower
SES in general. The latent growth modelling approach therefore provides us with a possible alternative for identifying
children who had difficulties learning mathematics. Interventions should be provided to these children according to their
needs.

Appendix 1. Arithmetic task used in the present study

Item no. Item Item no. Item

1 1+2 31 24
2 3+4 32 43
3 2+7 33 73
4 10 + 4 34 58
5 11 + 6 35 67
6 6 + 13 36 98
7 17 + 6 37 14  7
8 28 + 37 38 5  32
9 54 + 38 39 19  23
10 138 + 129 40 41  26
11 232 + 173 41 18/3
12 695 + 227 42 15/5
 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920 2919

Appendix 1 (Continued )

Item no. Item Item no. Item

13 3–1 43 42/6
14 5–2 44 56/7
15 9–5 45 135/9
16 13–2 46 456/12
17 19–7 47 15 + 6  8
18 18–12 48 (68–4)/4
19 14–9 49 (28  7 + 12)/8
20 22–4 50 (207–108)  6/33
21 40–27
22 102–59
23 227–135
24 741–498
25 5+8+9
26 24 + 17 + 41
27 20–7–5
28 78–24–19
29 17 + 9–6
30 70–37 + 23

References

Ashcraft, M. H., & Moore, A. M. (2012). Cognitive processes of numerical estimation in children. Journal of Experimental Child Psychology, 111(2), 246–267. http://
dx.doi.org/10.1016/j.jecp.2011.08.005
Aunio, P., & Niemivirta, M. (2010). Predicting children’s mathematical performance in grade one by early numeracy. Learning and Individual Differences, 20(5), 427–
435. http://dx.doi.org/10.1016/j.lindif.2010.06.003
Bishop, D. V. M. (2010). Which neurodevelopmental disorders get researched and why? PLOS ONE, 5(11), e15112. http://dx.doi.org/10.1371/journal.-
pone.0015112
Bonny, J. W., & Lourenco, S. F. (2013). The approximate number system and its relation to early math achievement: Evidence from the preschool years. Journal of
Experimental Child Psychology, 114(3), 375–388. http://dx.doi.org/10.1016/j.jecp.2012.09.015
Booth, J. L., & Siegler, R. S. (2006). Developmental and individual differences in pure numerical estimation. Developmental Psychology, 42(1), 189–201. http://
dx.doi.org/10.1037/0012-1649.41.6.189
Boscardin, C. K., Muthén, B., Francis, D. J., & Baker, E. L. (2008). Early identification of reading difficulties using heterogeneous developmental trajectories. Journal of
Educational Psychology, 100(1), 192–208. http://dx.doi.org/10.1037/0022-0663.100.1.192
Butterworth, B. (2003). Dyscalculia screener. London: NFER Nelson Publishing Company Ltd.
Butterworth, B., Varma, S., & Laurillard, D. (2011). Dyscalculia: From brain to education. Science, 332(6033), 1049–1053. http://dx.doi.org/10.1126/sci-
ence.1201536
Byrne, B., Fielding-Barnsley, R., & Ashley, L. (2000). Effects of preschool phoneme identity training after six years: Outcome level distinguished from rate of
response. Journal of Educational Psychology, 92(4), 659–667. http://dx.doi.org/10.1037/002
Chan, B. M., & Ho, C. S. (2010). The cognitive profile of Chinese children with mathematics difficulties. Journal of Experimental Child Psychology, 107(3), 260–279.
http://dx.doi.org/10.1016/j.jecp.2010.04.016
Cowan, R., & Powell, D. (2014). The contributions of domain-general and numerical factors to third-grade arithmetic skills and mathematical learning disability.
Journal of Educational Psychology, 106(1), 214–229. http://dx.doi.org/10.1037/a0034097
Crollen, V., Castronovo, J., & Seron, X. (2011). Under- and over-estimation: A bi-directional mapping process between symbolic and non-symbolic representations
of number? Experimental Psychology, 58(1), 39–49. http://dx.doi.org/10.1027/1618-3169/a000064
De Smedt, B., & Gilmore, C. K. (2011). Defective number module or impaired access? Numerical magnitude processing in first graders with mathematical
difficulties. Journal of Experimental Child Psychology, 108(2), 278–292. http://dx.doi.org/10.1016/j.jecp.2010.09.003
De Smedt, B., Noël, M.-P., Gilmore, C., & Ansari, D. (2013). How do symbolic and non-symbolic numerical magnitude processing skills relate to individual
differences in children’s mathematical skills? A review of evidence from brain and behavior. Trends in Neuroscience and Education, 1–8. http://dx.doi.org/
10.1016/j.tine.2013.06.001
Dehaene, S., Piazza, M., Pinel, P., & Cohen, L. (2003). Three parietal circuits for number processing. Cognitive Neuropsychology, 20(3-6), 487–506. http://dx.doi.org/
10.1080/02643290244000239
Feigenson, L., Dehaene, S., & Spelke, E. (2004). Core systems of number. Trends in Cognitive Sciences, 8(7), 307–314. http://dx.doi.org/10.1016/j.tics.2004.05.002
Fletcher, J. M., Coulter, W. A., Reschly, D. J., & Vaughn, S. (2004). Alternative approaches to the definition and identification of learning disabilities: Some questions
and answers. Annals of Dyslexia, 54(2), 304–331. http://dx.doi.org/10.1007/s11881-004-0015-y
Fuchs, L. S., Geary, D. C., Compton, D. L., Fuchs, D., Hamlett, C. L., Seethaler, P. M., et al. (2010). Do different types of school mathematics development depend on
different constellations of numerical versus general cognitive abilities? Developmental Psychology, 46(6), 1731–1746. http://dx.doi.org/10.1037/a0020662
Geary, D. C. (2013). Early foundations for mathematics learning and their relations to learning disabilities. Current Directions in Psychological Science, 22(1), 23–27.
http://dx.doi.org/10.1177/0963721412469398
Geary, D. C., Bailey, D. H., & Hoard, M. K. (2009). Predicting mathematical achievement and mathematical learning disability with a simple screening tool: The
number sets test. Journal of Psychoeducational Assessment, 27(3), 265–279. http://dx.doi.org/10.1177/0734282908330592
Geary, D. C., Bailey, D. H., Littlefield, A., Wood, P., Hoard, M. K., & Nugent, L. (2009). First-grade predictors of mathematical learning disability: A latent class
trajectory analysis. Cognitive Development, 24(4), 411–429. http://dx.doi.org/10.1016/j.cogdev.2009.10.001
Geary, D. C., Hoard, M. K., Nugent, L., & Byrd-Craven, J. (2008). Development of number line representations in children with mathematical learning disability.
Developmental Neuropsychology, 33(3), 277–299. http://dx.doi.org/10.1080/87565640801982361
Gersten, R., Clarke, B., & Mazzocco, M. M. M. (2007). Historical and contemporary perspectives on mathematical learning disabilities. In D. B. Berch & M. M. M.
Mazzocco (Eds.), Why is math so hard for some children? (pp. 7–27). Baltimore: Paul H. Brookes Publishing Company.
Gilmore, C. K., McCarthy, S. E., & Spelke, E. S. (2010). Non-symbolic arithmetic abilities and mathematics achievement in the first year of formal schooling.
Cognition, 115(3), 394–406. http://dx.doi.org/10.1016/j.cognition.2010.02.002
Halberda, J., Ly, R., Wilmer, J. B., Naiman, D. Q., & Germine, L. (2012). Number sense across the lifespan as revealed by a massive Internet-based sample. Proceedings
of the National Academy of Sciences of the United States of America, 109(28), 11116–11120. http://dx.doi.org/10.1073/pnas.1200196109
Halberda, J., Mazzocco, M. M. M., & Feigenson, L. (2008). Individual differences in non-verbal number acuity correlate with maths achievement. Nature, 455, 665–
668. http://dx.doi.org/10.1038/nature07246
Ho, C.S.-H., Chan, D.W.-O., Chung, K. K., Tsang, S.-M., Lee, S.-H., & Cheng, R.W.-Y. (2007). The Hong Kong test of specific learning difficulties in reading and writing for
primary school students-second edition [HKT-P(II)]. Hong Kong: Hong Kong Specific Learning Difficulties Research Team.
2920 T.T.-Y. Wong et al. / Research in Developmental Disabilities 35 (2014) 2906–2920

Hyde, D. C., Khanum, S., & Spelke, E. S. (2014). Brief non-symbolic, approximate number practice enhances subsequent exact symbolic arithmetic in children.
Cognition, 131(1), 92–107. http://dx.doi.org/10.1016/j.cognition.2013.12.007
Izard, V., Sann, C., Spelke, E. S., & Streri, A. (2009). Newborn infants perceive abstract numbers. Proceedings of the National Academy of Sciences of the United States of
America, 106(25), 10382–10385. http://dx.doi.org/10.1073/pnas.0812142106
Jordan, N. C., & Kaplan, D. (2006). Number sense growth in kindergarten: A longitudinal investigation of children at risk for mathematics difficulties. Child
Development, 77(1), 153–175. http://dx.doi.org/10.1111/j. 1467-8624.2006.00862.x
Jung, T., & Wickrama, K. A. S. (2008). An introduction to latent class growth analysis and growth mixture modeling. Social and Personality Psychology Compass, 2(1),
302–317. http://dx.doi.org/10.1111/j. 1751-9004.2007.00054.x
Kaufmann, L., Mazzocco, M. M. M., Dowker, A., von Aster, M., Göbel, S. M., Grabner, R. H., et al. (2013). Dyscalculia from a developmental and differential
perspective. Frontiers in Psychology, 4(August), 516. http://dx.doi.org/10.3389/fpsyg.2013.00516
Kucian, K., Grond, U., Rotzer, S., Henzi, B., Schönmann, C., Plangger, F., et al. (2011). Mental number line training in children with developmental dyscalculia.
Neuroimage, 57(3), 782–795. http://dx.doi.org/10.1016/j.neuroimage.2011.01.070
Landerl, K., Bevan, A., & Butterworth, B. (2004). Developmental dyscalculia and basic numerical capacities: A study of 8–9-year-old students. Cognition, 93(2), 99–
125. http://dx.doi.org/10.1016/j.cognition.2003.11.004
Landerl, K., Fussenegger, B., Moll, K., & Willburger, E. (2009). Dyslexia and dyscalculia: two learning disorders with different cognitive profiles. Journal of
Experimental Child Psychology, 103(3), 309–324. http://dx.doi.org/10.1016/j.jecp.2009.03.006
Libertus, M. E., Feigenson, L., & Halberda, J. (2013). Is approximate number precision a stable predictor of math ability? Learning and Individual Differences, 25, 126–
133. http://dx.doi.org/10.1016/j.lindif.2013.02.001
Lonnemann, J., Linkersdörfer, J., Hasselhorn, M., & Lindberg, S. (2013). Developmental changes in the association between approximate number representations
and addition skills in elementary school children. Frontiers in Psychology, 4(October), 783. http://dx.doi.org/10.3389/fpsyg.2013.00783
Mazzocco, M. M. M. (2007). Defining and differentiating mathematical learning disabilities and difficulties. In D. B. Berch & M. M. M. Mazzocco (Eds.), Why is math
so hard for some children? (pp. 29–47). Baltimore: Paul H. Brookes Publishing Company.
Mazzocco, M. M. M., Feigenson, L., & Halberda, J. (2011a). Impaired acuity of the approximate number system underlies mathematical learning disability
(dyscalculia). Child Development, 82(4), 1224–1237. http://dx.doi.org/10.1111/j. 1467-8624.2011.01608.x
Mazzocco, M. M. M., Feigenson, L., & Halberda, J. (2011b). Preschoolers’ precision of the approximate number system predicts later school mathematics
performance. PLOS ONE, 6(9), e23749. http://dx.doi.org/10.1371/journal.pone.0023749
Mazzocco, M. M. M., & Myers, G. F. (2003). Complexities in identifying and defining mathematics learning disability in the primary school-age years. Annuals of
Dyslexia, 53(1), 218–253. http://dx.doi.org/10.1007/s11881-003-0011-7.Complexities
Mazzocco, M. M. M., & Räsänen, P. (2013). Contributions of longitudinal studies to evolving definitions and knowledge of developmental dyscalculia. Trends in
Neuroscience and Education, 2(2), 65–73. http://dx.doi.org/10.1016/j.tine.2013.05.001
Mazzocco, M., & Thompson, R. (2005). Kindergarten predictors of math learning disability. Learning Disabilities Research and Practice, 20(3), 142–155. http://
dx.doi.org/10.1111/j. 1540-5826.2005.00129.x.Kindergarten
Mejias, S., Grégoire, J., & Noël, M.-P. (2012). Numerical estimation in adults with and without developmental dyscalculia. Learning and Individual Differences, 22(1),
164–170. http://dx.doi.org/10.1016/j.lindif.2011.09.013
Mejias, S., Mussolin, C., Rousselle, L., Grégoire, J., & Noël, M.-P. (2012). Numerical and nonnumerical estimation in children with and without mathematical
learning disabilities. Child Neuropsychology: A Journal on Normal and Abnormal Development in Childhood and Adolescence, 18(6), 550–575. http://dx.doi.org/
10.1080/09297049.2011.625355
Morgan, P., Farkas, G., & Wu, Q. (2009). Five-year growth trajectories of kindergarten children with learning difficulties in mathematics. Journal of Learning
Disabilities, 42, 306–322. http://dx.doi.org/10.1177/0022219408331037
Murphy, M., Mazzocco, M. M. M., Hanich, L. B., & Early, M. C. (2007). Cognitive characteristics of children with mathematics learning disability (MLD) vary as a
function of the cutoff criterion used to define MLD. Journal of Learning Disabilities, 40(5), 458–478. http://dx.doi.org/10.1177/00222194070400050901
Muthén, L. K., & Muthén, B. O. (2011). Mplus user’s guide (6th ed.). Los Angeles, CA: Muthén & Muthén.
Nylund, K. L., Asparouhov, T., & Muthén, B. O. (2007). Deciding on the number of classes in latent class analysis and growth mixture modeling: A Monte Carlo
simulation study. Structural Equation Modeling: A Multidisciplinary Journal, 14(4), 535–569. http://dx.doi.org/10.1080/10705510701575396
Ostad, S. A. (1997). Developmental differences in addition strategies: A comparison of mathematically disabled and mathematically normal children. The British
Journal of Educational Psychology, 67(Pt 3), 345–357. http://dx.doi.org/10.1111/j. 2044-8279.1997.tb01249.x
Park, J., & Brannon, E. M. (2013). Training the approximate number system improves math proficiency. Psychological Science, 24(10), 2013–2019. http://dx.doi.org/
10.1177/0956797613482944
Parsons, S., & Bynner, J. (2005). Does numeracy matter more? London: National Research and Development Centre for Adult Literacy and Numeracy.
Passolunghi, M. C., & Siegel, L. S. (2001). Short-term memory, working memory, and inhibitory control in children with difficulties in arithmetic problem solving.
Journal of Experimental Child Psychology, 80(1), 44–57. http://dx.doi.org/10.1006/jecp.2000.2626
Piazza, M., Facoetti, A., Trussardi, A. N., Berteletti, I., Conte, S., Lucangeli, D., et al. (2010). Developmental trajectory of number acuity reveals a severe impairment
in developmental dyscalculia. Cognition, 116(1), 33–41. http://dx.doi.org/10.1016/j.cognition.2010.03.012
Piazza, M., Izard, V., Pinel, P., Le Bihan, D., & Dehaene, S. (2004). Tuning curves for approximate numerosity in the human intraparietal sulcus. Neuron, 44(3), 547–
555. http://dx.doi.org/10.1016/j.neuron.2004.10.014
Raven, J. C. (1956). Standard progressive matrices sets A, B, C, D & E. Oxford: Psychologists Press.
Reeve, R., Reynolds, F., Humberstone, J., & Butterworth, B. (2012). Stability and change in markers of core numerical competencies. Journal of Experimental
Psychology General. http://dx.doi.org/10.1037/a0027520
Reigosa-Crespo, V., Valdés-Sosa, M., Butterworth, B., Estévez, N., Rodrı́guez, M., Santos, E., et al. (2012). Basic numerical capacities and prevalence of
developmental dyscalculia: the Havana Survey. Developmental Psychology, 48(1), 123–135. http://dx.doi.org/10.1037/a0025356
Rousselle, L., & Noël, M.-P. (2007). Basic numerical skills in children with mathematics learning disabilities: A comparison of symbolic vs non-symbolic number
magnitude processing. Cognition, 102(3), 361–395. http://dx.doi.org/10.1016/j.cognition.2006.01.005
Shalev, R. S. (2007). Prevalence of developmental dyscalculia. In D. B. Berch & M. M. M. Mazzocco (Eds.), Why is math so hard for some children? The nature and
origins of mathematical learning difficulties and disabilities (pp. 49–60). Baltimore: Paul H. Brookes Publishing Company.
Siegler, R. S., & Booth, J. L. (2004). Development of numerical estimation in young children. Child Development, 75(2), 428–444. http://dx.doi.org/10.1111/j. 1467-
8624.2004.00684.x
Tabachnick, B., & Fidell, L. (2006). Using multivariate statistics (6th ed.). Needham Heights, MA: Allyn & Bacon, Inc.
Wilson, A. J., & Dehaene, S. (2007). Number sense and developmental dyscalculia. In D. Coch, G. Dawson, & K. W. Fischer (Eds.), Human behavior, learning, and the
developing brain: Atypical development (pp. 212–238). New York: Guilford Press.
Wong, T. T.-Y., Ho, C. S.-H., & Tang, J. (under review-a). Defective number sense or impaired access: Differential impairment in different subgroups of children with
mathematics difficulties..
Wong, T. T.-Y., Ho, C. S.-H., & Tang, J. (under review-b). The relationship between ANS and arithmetic achievement: The mediating role of number–numerosity
mappings..