Mental Abacus 1

Please note: this manuscript is for noncommercial use only. It is the pre-publication version, and as such,
may contain errors not present in the final publication.

Learning mathematics in a visuospatial format: A randomized, controlled trial of mental abacus

instruction

David Barner1, George Alvarez2, Jessica Sullivan3, Neon Brooks2, Mahesh Srinivasan4, and

Michael C. Frank5
1
University of California, San Diego
2
Harvard University
3
Skidmore College
4
University of California, Berkeley
5
Stanford University

Acknowledgements: We would like to thank the children, families, and staff of Zenith School in

Vadodara, India for their patience and generosity over the past several years. In particular, we

would like to thank Abbasi Barodawala and Mary Joseph of Zenith School, and Snehal Karia and

Anand Karia of UCMAS India, for their invaluable contributions. Thank you also to Sean Barner,

Eleanor Chestnut, Jonathan Gill, Ali Horowitz, Talia Konkle, Ally Kraus, Molly Lewis, Bria

Long, Ann Nordmeyer, Viola Strmer, Jordan Suchow, and Katharine Tillman for help with data

collection. This work was funded by a grant to D.B. and G.A. from NSF REESE grant #0910206

and by an NSF GRFP to J.S.

Contact Information

David Barner, Department of Psychology, UCSD

9500 Gilman Drive, La Jolla, CA, 92093-0109; 858-246-0874; barner@ucsd.edu

 Mental Abacus 2

Abstract: Mental abacus (MA) is a technique for performing fast and accurate arithmetic using a

mental image of a physical abacus. Expert users exhibit astonishing calculation abilities. In a 3-

year randomized, controlled trial of 204 elementary-school students, we investigated the nature

of MA expertise, and whether it can be acquired by large groups of children in standard

classroom settings. We asked whether MA improves students mathematical abilities, and

whether expertise which requires sustained practice of mental imagery is driven by changes

to basic cognitive capacities like working memory. MA students improved on arithmetic tasks

relative to controls, but training was not associated with changes to basic cognitive abilities.

Instead, differences in spatial working memory at the beginning of the study mediated MA

learning. We conclude that MA expertise can be achieved by many children in standard

classrooms and results from efficient use of pre-existing abilities.

 Mental Abacus 3

Introduction

Mathematics instruction typically begins by introducing children to a system of numerals

and a set of arithmetic routines that operate on these numerals. For many children around the

world, early math lessons are supplemented by the use of an abacus, a physical manipulative

designed for representing exact quantities via the positions of counters, whose historical origins

date to 1200 AD or earlier.1 Extending the use of the physical abacus, children in some countries

also learn a technique known as mental abacus (MA). Using MA, users create and manipulate a

mental image of the physical device to perform arithmetic operations (see Figure 1 for details of

how MA represents number). Expert users of MA exhibit abilities that are nothing short of

astonishing: MA compares favorably to the use of electronic calculators (Kojima, 1957), it

enables rapid calculation even when users are speaking concurrently (Hatano, Miyake, & Binks,

1977), and it allows its users to dominate international calculation competitions like the Mental

Calculation World Cup. In the present study, we explored the nature of MA expertise.

Specifically, we asked whether the extraordinary levels of achievement witnessed in experts can

be attained by students in large K-12 classroom settings. In doing so, we asked a more general

question regarding the nature of expertise, and whether attaining unusual levels of performance

requires changes to basic cognitive capacities, or can instead arise via the exploitation of existing

cognitive resources (see Ericson & Smith, 1991, for review).

MA abilities appear to rely primarily on non-linguistic representations, especially

visuospatial working memory, as well as motor procedures that are learned during initial

physical abacus training. Although arithmetic computations of untrained college students are

1
Systems with similar physical structure perhaps direct predecessors of the abacus emerged
much earlier, and included the Roman counting board and the Babylonian counting tablet
(Menninger, 1969).
 Mental Abacus 4

Figure 1. The Japanese soroban-style abacus used by participants in this study, shown here
representing the value 123,456,789. A physical abacus represents number via the arrangement
of beads into columns, each of which represents a place value (e.g., ones, tens, hundreds,
thousands, etc...), with values increasing from right to left. To become proficient at MA, users
of the physical abacus learn to create a mental image of the device and to manipulate this
image to perform computations.

highly disrupted by verbal interference (e.g., concurrent speaking), MA users are less affected by

concurrent linguistic tasks, and much more affected by motor interference (Frank & Barner,

2011; Hatano, Miyake, & Binks, 1977). Consistent with this, while standard arithmetic routines

recruit brain regions related to verbal processing and verbal working memory, MA computations

selectively activate regions associated with vision and spatial working memory (Chen et al.,

2006; Hu et al., 2011; Tanaka, Michimata, Kaminaga, Honda, & Sadato, 2002). Finally, MA

users computational limits and their patterns of computation errors are consistent with known

limits to visual working memory, and suggest that each abacus column is represented as a

distinct object in visuospatial working memory (Frank & Barner, 2011).

Although previous studies have documented impressive abilities in MA experts, they do

not shed light on whether MA training can produce benefits for a broad range of students in a

standard classroom setting.2 The present study tested this, and asked whether MA expertise

2
Most previous studies of MA have investigated experts who have elected to receive extensive
training outside of school, raising the possibility that expertise is possible only in individuals
who have particular mathematical talents or interests. Similarly, studies that have tested MA
 Mental Abacus 5

results from changes in the users ability to create and manipulate structures in visual working

memory (a hypothesis we refer to as cognitive transfer) or if instead MA simply exploits pre-

existing abilities such that expertise arises only in individuals with relatively strong spatial

working memory abilities (a hypothesis we call cognitive moderation, because these cognitive

abilities would serve as moderators of the techniques efficacy; Baron & Kenny, 1986).

On the cognitive transfer hypothesis, learning MA may provide mathematical benefits

for large groups of children in a K-12 setting, and these benefits may be due to gains in basic

cognitive abilities like imagery, working memory, and attention that appear to be important

for MA computation. Examples of this kind of transfer come from executive function and spatial

cognition interventions, which are both argued to provide a cognitive route for improving

academic performance (Diamond & Lee, 2011; Uttal et al., 2013). In the case of MA, mastery

requires hours of intensive practice with mental imagery, attentional allocation, and cross-

domain integration of information (e.g., from numerical symbols to beads and back again), hence

these skills might benefit from MA practice.

On the cognitive moderation hypothesis, MA may be most beneficial to a particular

subset of students, such that expertise is best predicted by performance on cognitive batteries at

the beginning of MA training (Frank & Barner, 2011). Rather than stemming from changes to an

individuals basic cognitive capacities, MA expertise may result when MA is learned and

practiced by individuals who are particularly able to perform complex computations in working

memory, and to manage the attentional demands required by the method.

To explore the nature of MA and its utility in large classroom settings, we randomly

assigned children from the same school to receive training in either MA or a standard

training in non-experts have not randomly assigned participants to training conditions,
complicating their interpretation (e.g., Stigler, Chalip, & Miller, 1986).
 Mental Abacus 6

mathematics curriculum. We conducted our study at a school located in Vadodara, India, where a

short (1 hour) weekly MA training had previously supplemented the standard curriculum for

students in the second grade and above, and thus instructors and appropriate training

infrastructure were already in place. Starting at the beginning of the second grade, half of the

children in our study were assigned to study MA for three hours per week (MA group). The

remaining half of students were assigned to receive no abacus training, and instead perform three

hours of supplementary practice using a secondary mathematics curriculum that complemented

their existing mathematics coursework (Control group). Thus, the total amount of mathematics

training received by each group was identical. We followed children over the course of three

years, and assessed outcomes using a battery of mathematical and cognitive assessments,

including measures of mental rotation, approximate number, and spatial and verbal working

memory. These tasks were administered both prior to intervention and at the end of each school

year so as to probe the extent of any possible cognitive mediation or transfer effects.

Methods

Participants

We enrolled an entire cohort of students attending a charitable school for low-income

children in Vadodara, India. At the initiation of the study, over 80% of children attending the

school came from families who earned less than $2,000 US per annum (~$5.50/day). In our

sample, 59% of children came from Hindu families and 41% from Muslim families. The total

population of the school was approximately 2100 students, ranging from pre-K to high school.

At the time of enrollment (which we refer to as Year 0), the participants were 204

children aged 5 7 years old who were beginning their 2nd grade year. We randomly assigned

these students to two groups, MA and Control. We then further randomly assigned children into
 Mental Abacus 7

three classrooms of approximately 65 70 children each (differing from their classroom

assignments in the previous year), with one classroom comprised of MA students, one of Control

students, and one split half and half.

Of the 100 students in the MA group and the 104 students in the Control group, 88 (88%)

and 99 (95%) provided some data in every year of testing, respectively. Missing data were

sometimes due to sickness or absenteeism, but primarily to changes of school. We analyzed data

only from complete cases.

Intervention Procedure

Children in both the MA and Control groups studied the schools standard (non-abacus)

mathematics curriculum over the duration of the study. Additionally, both groups received three

hours per week of additional mathematics instruction as follows. In the MA group, children were

given three hours per week of instruction in the use of the physical and mental abacus by an

experienced MA teacher outside the childrens home classroom (such that control group children

were not exposed to MA technique). Abacus instruction was broken into two 90-minute sessions

per week and followed a common international curriculum that begins with use of the physical

abacus for addition and subtraction, and then moves to mental abacus computations. The first

year of training focused primarily on the physical abacus, with greater emphasis placed on MA

in subsequent years. Common activities in the MA training program included worksheet practice

of addition and subtraction, practice translating abacus configurations into Arabic numerals, and

practice doing speeded arithmetic using MA.

Control students were provided with three hours per week of supplemental mathematics

training using international curriculum materials. Activities from the supplemental curriculum
 Mental Abacus 8

included extra practice on topics from the standard math curriculum, clock reading, and the use

of geometric manipulatives such as peg boards.

Assessment Procedure

The study spanned three years of the participants elementary education, and began with

a baseline test before training began. In each of four annual assessments, children received a

large battery of computerized and paper-based tasks. Year 0 assessments were given at the

beginning of 2nd grade; Year 1 3 assessments were given at the end of 2nd, 3rd, and 4th grade,

respectively. All assessments included both measures of mathematics and more general cognitive

measures. A small number of other tasks were included but are not discussed in the current

manuscript (see Supplemental Online Material for detailed description of all measures).

Mathematics measures. Children completed the Calculation subtest of the Woodcock

Johnson Tests of Achievement (WJ-III) and the Math Fluency subtest of the Wechsler Individual

Achievement Test (WIAT-III). We also administered two in-house assessments of mathematics

skill that, unlike the standardized tests, were designed to specifically target arithmetic skills

acquired between 2nd and 4th grade (see SOM). The first measured childrens arithmetic abilities

by testing performance in single- and multi-digit addition, subtraction, division, and

multiplication problems. The second measured conceptual understanding of place value by

asking children to complete number-decomposition problems (e.g. 436 = 6 + ___ + 30).

Cognitive measures. At each assessment point, children completed 1-2 subsets of 10

problems from Ravens Progressive Matrices (Raven, 1998), as well as two paper-based tests of

speeded mental rotation (one using letters and one using shapes). They also completed three

computerized tasks: (1) an adaptive test of spatial working memory (SWM); (2) an adaptive test

of verbal working memory (VWM); (3) a number comparison task, in which children were asked
 Mental Abacus 9

to indicate the larger of two dot arrays (see SOM Section 1 for detailed task descriptions and

related citations). For the working memory tasks, we report estimates of span i.e., the average

number of items on which a child was successful. For the number comparison task, we report

Weber fractions (a measure of approximate number acuity, estimated from our task via the

method of Halberda, Mazzocco, & Feigenson, 2008).

Grades. For each year, we obtained childrens grades in English, Math, Science, and

Computer classes, as well as in Music, Art, and Physical Education.

Abacus Only Measures. For each year after the intervention began, students in the MA

group completed a set of three paper and pencil tasks to assess their ability to use an abacus. All

three were administered after the end of all other testing. The first two, Abacus Sums (addition)

and Abacus Arithmetic (addition and subtraction), tested the ability to do computations using a

physical abacus, while the third, Abacus Decoding, tested the ability to decode abacus images

into standard Arabic numerals. These tasks allowed us to verify that any differences between the

MA and Control groups were indeed mediated by changes to abacus skills.

Attitude measures. In Year 3, we administered two measures that explored whether the

intervention had changed childrens attitudes towards mathematics, and thus whether training

effects might be mediated by differences in motivational level and engagement with mathematics.

First, we administered a growth-mindset questionnaire (adapted from Dweck, 1999), which

probed childrens attitudes regarding the malleability of their own intelligence. Second, we

administered a math-anxiety questionnaire (adapted from Ramirez et al., 2013), which measured

the anxiety that children experienced when solving different kinds of math problems and

participating in math class.

 Mental Abacus 10

Data Analysis3

We used a longitudinal mixed models approach to quantify the statistical reliability of the

effects of randomization to training group on our outcome measures. For each outcome measure,

we fit a baseline model that included a growth term for each student over time and an overall

main effect of intervention-group (to control for differences between groups at study initiation).

We then tested whether the fit of this model was improved significantly by an interaction term

capturing the effect of the intervention over time.

Because we did not have any a priori hypotheses about the shape of the dose-response

function between the intervention and particular measures of interest, we fit models using three

types of growth terms: (1) simple linear growth over time, (2) quadratic growth over time, and

(3) independent growth for each year after baseline. This last type of model allows for the

possibility of non-monotonic growth patterns. We tested for interactions between group

assignment and growth in each of these models. All p values are reported from likelihood ratio

tests (see SOM).

Results

Mathematics outcomes. As seen in Figure 2, MA training produced significant gains in

mathematical abilities relative to the control group. Consistent with this, in Year 3, we observed

numerical differences between the two groups on three out of the four mathematics tasks, with

effect sizes of Cohens d = .60 (95% CI: .30 - .89) for arithmetic, .24 (-.05 - .52) for WJ-III,

3
All data and code for the analyses reported here are available at
http://github.com/langcog/mentalabacus
 Mental Abacus 11

Figure 2. Mathematics outcome measures for the two intervention conditions, plotted by study
year (with 0 being pre-intervention). Error bars show 95% confidence intervals computed by
non-parametric bootstrap.

and .28 (.00 - .57) for place value. 4 We observed only a small numerical difference for WIAT-III,

however (d =.13; -.15 - .42).

We used longitudinal growth models to assess whether advantages observed in the MA

group were in fact driven by additional MA training (Figure 2). All three of these models (i.e.,

4 The confidence intervals for these effect sizes represent confidence intervals on pairwise tests

for Year 3 alone; the statistical significance of the broader patterns we observed are best captured
by the longitudinal models described below.
 Mental Abacus 12

linear, quadratic, and non-monotonic) showed strong time by condition interactions for both the

arithmetic and WJ-III measures, suggesting that performance on each of these tasks did improve

with additional MA training (likelihood-ratio tests for adding the time by condition interaction

term to the growth model were 2linear(1) > 6.33, 2quadratic(2) > 11.56, and 2independent(3) > 12.51,

with ps < .01 in all cases). Consistent with the small numerical difference observed between

groups on the WIAT-III, this measure did not approach significance in any of the three growth

models. The smaller effects observed on the standardized measures are perhaps not surprising,

given the smaller number of arithmetic-focused items on these measures and hence the

likelihood of them having lower sensitivity to individual differences between children (see SOM

Section 4.7 for further analysis). More surprising, however, was that the place-value measure did

not approach significance in the growth models, especially given that we observed substantial

numerical differences between the groups (e.g., performance in year 3 differed significantly

between groups in a univariate analysis, t(185) = 1.96, p = .05). We speculate that we did not

observe consistent growth with this measure due to its low reliability from Year 0 to Year 1 (r

= .22). Together, these analyses suggest that the MA intervention was more effective in building

students arithmetic skill than an equivalent amount of supplemental training in standard

mathematics techniques. Effects of MA training on arithmetic ability were observed not only in

our in-house measure, which included many arithmetic problems tailored to the level of

elementary school students, but also on the WJ-III, a widely-used standardized measure that

includes a range of problem types and formats. While the evidence for differential gains in

conceptual understanding of place value was more limited, MA students did not fall behind

students in the Control group, despite the fact that the MA curriculum primarily stresses rote

calculation rather than conceptual understanding.

 Mental Abacus 13

Figure 3. Cognitive outcome measures for the two intervention conditions, plotted by study
year. Top axes show mean items correct in working memory span tasks, while bottom axes
show proportion correct across trials in number comparison and mental rotation tasks. Error
bars show 95% confidence intervals computed by non-parametric bootstrap.

Cognitive outcomes. Whereas MA training produced consistent gains in arithmetic ability, it did

not produce consistent gains in the cognitive abilities we measured (Figure 3). Higher math

performance in the MA group was therefore not the result of improved cognitive capacity due to

MA training. For example, in Year 3, we observed between-group effect sizes of -0.16 (95% CI:

-0.45 - 0.12) on our number comparison measure (note that smaller Weber fractions indicate

more accurate estimations). Also, we found an effect size of -0.14 (-0.43 0.15) for Ravens
 Mental Abacus 14

progressive matrices, of -.06 for mental rotation (-0.34 0.23), and of .05 (-.24 - .34) for spatial

working memory. Only one cognitive measure verbal working memory showed an advantage

in Year 3 for the MA group (.26; -.03 - .55).

For these cognitive measures, as with the arithmetic tasks, we used longitudinal growth

models to assess whether advantages for the MA group were driven by training in MA. Because

we used different sets of Ravens problems for each year, we could not fit growth models, but t-

tests showed no reliable effects of MA training for any year (all ts < .96, ps > .34). Similarly,

longitudinal models (linear, quadratic, and independent) confirmed that none of the cognitive

tasks (numerical comparison, mental rotation, verbal working memory, or spatial working

memory), showed significant time by condition interactions, with one exception. For verbal

working memory, the non-independent growth model showed a significant time by condition

interaction (p < .01), though both linear and quadratic growth models showed no significant time

by condition interaction (SOM Section 4.4). Thus, this result appears to have been driven by the

fast growth in verbal working memory span in Year 1 exhibited by the MA group, relative to the

control group (see Figure 3).

The large effect of MA training on verbal working memory in Year 1 is mirrored in a

similar trend observed in spatial working memory in Years 1 and 2 (significant or close to

significant in individual t-tests, t(185) = 2.36, p = .02 and t(184) = 1.84, p = .07, but not in any

longitudinal model). In both cases, the overall shape of the developmental curve is asymptotic,

with working memory spans approaching approximately four items by Year 2 in the MA group.

This pattern could be interpreted as evidence that differences in working memory between the

MA and control groups do exist, but are expressed only in the rate of growth to asymptote, rather

than in the absolute level of the asymptote itself. Against this hypothesis, however, additional
 Mental Abacus 15

analyses (SOM Section 4.5) find that (1) our spatial working memory task did not exhibit ceiling

effects and (2) data from 20 American college undergraduates and 67 high socio-economic status

(SES) Indian children from the same region of India show that children in our study had overall

lower spatial working memory than higher-SES children, and were far from being at adult levels

of performance. Most important, these Year 1 effects surfaced before children began to receive

training on the mental component of MA and were still learning the physical technique. We

therefore do not believe that this result is likely to be related to the ultimate gains we see in MA

across the study.

Academic outcomes. MA did not produce large, consistent changes in students grades

across academic subjects, although we saw some small trends towards better math, science, and

computer grades in the MA group in some models. These differences are subject to teacher bias,

however, since teachers were of course knowledgeable about the intervention. Thus, we do not

believe they should be weighted heavily in evaluating performance, especially since our own

standardized measures of mathematical competence were available for analysis (for additional

analysis, see SOM and Figure S1).

Attitude Measures. There were no differences between groups on either childrens self-

reported mathematics anxiety (t(184) = 1.05, p = .29) or their endorsement of a growth mindset

(t(184) = -.61, p = .54). Thus, it is unlikely that differences we observed in mathematics

measures were due to differential effects of our intervention on childrens anxiety about

mathematics or on their general mindset towards learning.

 Mental Abacus 16

Figure 4. Performance on the arithmetic task, split by both intervention condition and median
spatial working memory performance in Year 0. Error bars show 95% confidence intervals;
lines show best fitting quadratic curves.

Mediators of intervention effects. Given that MA training produced gains in math

outcomes, we next asked which factors mediated these gains, and thus whether individual

differences between children at the beginning of the study predicted MA achievement. As

already noted, MA training did not augment cognitive abilities, so the math advantages in the

MA group could not have been driven by enhanced working memory, mental imagery, or

approximate number acuity that resulted from MA training. However, it is possible that

individual cognitive differences between children in the MA group (prior to their entry into the

study) were responsible for how well they learned and benefitted from MA. To explore this

possibility, we conducted post-hoc analyses using moderator variables.

Our analytic approach relied on the same longitudinal modeling approach described

above. For each math outcome variable, we fit models that included participants Year 0

performance on each cognitive predictor (for simplicity and to avoid over-parameterizing our

models, we used linear and quadratic models only). The coefficient of interest was a three-way
 Mental Abacus 17

interaction of time, condition, and initial performance on the cognitive predictor of interest. This

three-way interaction term captures the intuition that growth in performance on a task for MA

participants is affected by their baseline abilities on a particular cognitive task. As before, we

used likelihood ratio tests to assess whether these interaction terms improved model fit.5

A median split of children according to SWM prior to MA training resulted in a low

SWM group with an average threshold of 1.9 items, and a high SWM group with an average

threshold of 3.7 items. Previous studies of SWM thresholds for middle to high SES 5- to 7-year-

olds find thresholds of approximately 3.5 4 items (Pickering, 2010; Logie & Pearson, 1997).

Thus, a subset of children in our study exhibited especially low SWM capacity. Related to this,

spatial working memory was a reliable moderator in both linear and quadratic models of

arithmetic (Figure 4). Those children who began the study with relatively strong spatial working

memory skills and who were randomly assigned to MA training showed significantly stronger

growth in our arithmetic assessment (2linear(1) = 4.63, p < .05; 2quadratic(2) = 5.93, p = .05).

Relative to receiving equal amounts of standard math training, children with weaker spatial

working memory did not appear to benefit differentially from the MA intervention, and instead

performed at an equivalent level to the students in the control group. Since the MA technique

relies on visuo-spatial resources for storage of the abacus image during computation (Frank &

Barner, 2011), it seems likely that those children with relatively lower spatial working memory

spans struggled to learn to perform computations accurately using MA.

5
Although with greater numbers of longitudinal measurements we could potentially have
detected interactive growth patterns (e.g. gains in working memory driving later gains in
mathematics), our current study did not have the temporal resolution for these analyses. We thus
restrict our analysis to testing for mediation in mathematics outcomes on the basis of each of the
cognitive variables measured at Year 0.
 Mental Abacus 18

Figure 5. Performance on the Physical Abacus Sums, Decoding, and Arithmetic tasks
(administered in years 1 3), plotted by a median split on spatial working memory. Error bars
show 95% confidence intervals computed by non-parametric bootstrap.

There was no comparable mediation effect with verbal working memory (additional

analysis in SOM Section 4.4), but a small number of additional moderation effects did approach

significance in one of the two models. There was a trend towards an effect of spatial working

memory on WJ-III (the other mathematics measure that showed strong MA effects; 2quadratic(2) =

5.59, p = .06, Figure S3). In addition, there were trends towards effects of Year 0 mental rotation

performance on arithmetic (2linear(1) = 2.71, p = .10) and place value (2linear(1) = 3.27, p = .07),

and an effect of number comparison acuity on WJ-III (2quadratic(2) = 8.29, p = .02, Figure S4).

These effects, though more tentatively supported, are nevertheless consistent with the hypothesis

that the MA intervention was most effective for children with greater visuo-spatial abilities at the

beginning of instruction.

Abacus only measures. Confirming that children in the MA group learned to use an

abacus, we found consistently high performance in Abacus Decoding for the MA group (> 80%

correct for all years). Also, performance on the abacus arithmetic and sums tasks rose

substantially from year to year, suggesting that childrens abacus computation abilities improved
 Mental Abacus 19

over the course of their training (Figure 5). Performance on these tests of physical abacus

arithmetic were significantly correlated with performance on our other math measures (In-house

arithmetic: r = .69, .74, and .81 for Years 1 3 respectively, all ps < .0001; WIAT: r = .57, .54, .73, all ps

< .0001; WJ: r = .45, .51, .64, all ps < .0001).

Critically, spatial working memory span in Year 0 was also related to intervention uptake,

as measured by the Abacus Only tasks, which were accomplished using a physical abacus.

Because we did not have Abacus Only data for Year 0, we could not directly test whether spatial

working memory moderated growth, but we did find a main effect of spatial working memory on

all three measures of abacus uptake for both linear and quadratic growth models (all 2(1) > 3.97,

all ps < .05; Figure 5).

Discussion

Our study investigated the nature of mental abacus (MA) expertise, and whether MA is

an effective tool for improving math outcomes in a standard classroom setting. To do this, we

conducted a three-year longitudinal study of mental abacus training. We found that MA training

led to substantial gains in students ability to perform accurate arithmetic computations. These

gains began to emerge after a single year of training suggesting that simply learning to use the

physical abacus had some effects on students mathematics aptitude, even prior to learning the

MA technique and became more pronounced with time. Consistent with this explanation and

with a role for abacus expertise in explaining the intervention effect we found that physical

abacus expertise at the end of the study was significantly correlated with arithmetic performance

across all math measures within our experimental group. Although there were signs of early

gains in cognitive capacities like spatial working memory in the MA group, such effects did not

persist to the end of the study, and could not explain gains in mathematics achievement. Also,
 Mental Abacus 20

differences in mathematics achievement did not appear to be driven by motivational or

attitudinal differences towards mathematics or intelligence more broadly.

Our findings are consistent with previous suggestions that cognitive transfer is rare.

Although performance on basic measures of attention and memory can be improved via direct

training on those measures (Diamond & Lee, 2011; Gathercole, Dunning, Holmes, & Wass,

under review; Melby-Lervg, & Hulme, 2012; Noack, Lvdn, Schmiedek, & Lindenberger,

2009), it may be difficult to achieve far transfer from training on unrelated tasks, even with

hours of focused practice (Dunning, Holmes, & Gathercole, 2013; Owen et al., 2010; Redick et

al., 2013). However, our findings suggest that although cognitive capacities are not importantly

altered by MA, they may predict which children will benefit most from MA training. MA

students who began our study with low SWM abilities did not differ in their math performance

from Control students, while those above the median made large gains on our arithmetic measure

(similar effects were not seen for verbal working memory).

Because MA relies on visuo-spatial resources for the storage and maintenance of abacus

images during computation (Frank & Barner, 2011), children with especially weak SWM may

have attained only basic MA abilities enough to reap benefits equal to additional hours of

standard math, but not to acquire unusual expertise. This finding suggests that the development

of MA expertise is mediated by childrens pre-existing cognitive abilities, and thus that MA may

not be suitable for all K-12 classroom environments, especially in groups of children who have

low spatial working memory or attentional capacities (which may have been the case in our

study). Critically, this does not mean that MA expertise depends on unusually strong cognitive

abilities. Perhaps because we studied children from relatively disadvantaged backgrounds, few
 Mental Abacus 21

children in our sample had SWM capacities comparable to those seen among typical children in

the United States.

Our study leaves open several questions about MA as an educational intervention. First, it

remains uncertain how much training is necessary to benefit from MA. In our study, children

received over 100 hours of MA instruction over three years; however, we found evidence that

children began to show improvements after a smaller amount of training (within one year of

exposure to the technique). Future studies should investigate the efficacy of MA training in

smaller, more focused sessions, and also whether the required number of training hours is

smaller in different populations (e.g., in middle or high-SES groups). Second, our results speak

to the efficacy of concrete manipulative systems in the classroom, and suggest that relatively

extensive training may be required to yield benefits. Previous studies have found mixed results

regarding the effectiveness of manipulatives for teaching mathematics (Ball, 1992; Uttal,

Scudder, & Deloache, 1997). However, MA may be unlike other manipulative systems.

Although the abacus is a concrete representation of numerosity that can be used to reinforce

abstract concepts, MA is unique in requiring the use of highly routinized procedures for

arithmetic calculation. Thus, additional research is needed to understand how MA differs from

other manipulatives. Finally, and most importantly, our study did not investigate the long-term

consequences of improved arithmetic fluency in MA users (e.g., on learning concepts like

negative numbers or fractions).

In sum, we find evidence that mental abacus a system rooted in a centuries-old

technology for arithmetic and accounting affords some children a significant advantage in

arithmetic calculation compared with additional hours of standard math training. Our evidence

also suggests that MA provides this benefit by building on childrens pre-existing cognitive
 Mental Abacus 22

capacities. Future studies should explore the long-term benefits of enhanced arithmetic abilities

using MA and the generalizability of this technique to other groups and cultural contexts.
 Mental Abacus 23

Author Contributions

All authors contributed to the design of the study and to data collection. Data analyses were

prepared chiefly by MF. All authors contributed to writing the manuscript. All authors approved

the final version of the manuscript for submission.

 Mental Abacus 24

References

Ball, D. L. (1992). Magical hopes: Manipulatives and the reform of math education. American

Educator, 16, 14-18.

Baron, R., & Kenny, D. (1986). The moderator-mediator variable distinction in social

psychological research: Conceptual, strategic, and statistical considerations. Journal of

Personality and Social Psychology, 51, 1173-1182.

Chen, F., Hu, Z., Zhao, X., Wang, R., Yang, Z., Wang, X., et al. (2006). Neural correlates

of serial abacus mental calculation in children: a functional MRI study. Neuroscience

Letters, 403 (1-2), 46-51.

Diamond, A. & Lee, K. (2011). Interventions shown to Aid Executive Function Development in

Children 4-12 Years Old. Science, 333, 959-964.

Dunning, D.L., Holmes, J., & Gathercole, S. (2013) Does working memory training lead to

generalised improvements in children with low working memory? A randomised

controlled trial. Developmental Science, 16(6), 915-926.

Dweck, C. S. (1999). Self-theories: Their role in motivation, personality, and development.

Philadelphia: Psychology Press.

Ericsson, K. A., & Smith, J. (Eds.). (1991). Toward a general theory of expertise: Prospects and

limits. Cambridge University Press.

Frank, M. C & Barner, D. (2011). Representing exact number visually using mental abacus.

Journal of Experimental Psychology: General, 141, 134149.

Halberda J., Mazzocco, MM. & Feigenson, L. (2008). Individual differences in non-verbal

number acuity correlate with maths achievement. Nature, 455, 665668.

Hatano, G., Miyake, Y. & Binks, M. G. (1977). Performance of expert abacus operators.
 Mental Abacus 25

Cognition, 5, 57-71.

Hu, Y., Geng, F., Tao, L., Hu, N., Du, F., Fu, K., & Chen, F. (2011). Enhanced white matter tract

integrity in children with abacus training. Human Brain Mapping, 32, 1021.

Kojima, T. (1954). The Japanese abacus: Its use and theory. Tokyo, Japan: Tuttle.

Logie, R., & Pearson, D. (1997). The inner eye and the inner scribe of visuo-spatial working

memory: Evidence from developmental fractionation. European Journal of Cognitive

Psychology, 9, 241-257.

Menninger, K. (1969). Number words and number symbols: A cultural history of numbers.

Cambridge, MA: MIT Press.

Melby-Lervg, M. & Hulme, C. (2012). Is working memory training effective? A meta-analytic

review. Developmental Psychology, 49, 270291. doi: 10.1037/a0028228.

Noack, H., Lvdn, M., Schmiedek, F., Lindenberger, U. (2013). Age-related differences in

temporal and spatial dimensions of episodic memory performance before and after

hundred days of practice. Psychology and Aging, 28,, 467-480.

Owen, A. M., Hampshire, A., Grahn, J. A., Stenton, R., Dajani, S., Burns, A. S., & Ballard, C.G.

(2010). Putting brain training to the test. Nature, 465(7299), 775U776.

Pickering, S. (2010). The development of visuo-spatial working memory. Memory, 9, 423-432.

Ramirez, G., Gunderson, E. A., Levine, S. C., & Beilock, S. L. (2013). Math anxiety, working

memory, and math achievement in early elementary school. Journal of Cognition and

Development, 14(2), 187-202.

Raven, J. (1998). Manual for Ravens progressive matrices and vocabulary scales. Oxford

Psychologists Press. Oxford, UK.

Redick, T. S., Shipstead, Z., Harrison, T. L., Hicks, K. L. Fried, D. E., Hambrick, D. Z., Kane, M.

J., & Engle, R. W. (2013). No evidence of intelligence improvement after working

 Mental Abacus 26

memory training: A randomized, placebo-controlled study. Journal of Experimental

Psychology: General, 142(2), 359-379.

Stigler, J. W., Chalip, L., & Miller, K. (1986). Consequences of skill: The case of abacus training

in Taiwan. American Journal of Education, 94(4), 447479.

Tanaka, S., Michimata, C., Kaminaga, T., Honda, M., & Sadato, N. (2002). Superior digit

memory of abacus experts: an event-related functional MRI study. Neuroreport,

13(17), 2187.

Uttal, D. H., Scudder, K. V., & DeLoache, J. S. (1997). Manipulatives as symbols: A new

perspective on the use of concrete objects to teach mathematics. Journal of Applied

Developmental Psychology, 18, 3754.

Uttal, D. H., Meadow, N. G., Tipton, E., Hand, L. L., Alden, A. R., Warren, C., and Newcombe,

N. (2013). The Malleability of Spatial Skills: A Meta-Analysis of Training Studies.

Psychological Bulletin, 139, 352-402.