COGNITIVE PSYCHOLOGY 16, 145-176 (1984)

“Mental Abacus”: The Effect of Abacus Training on

Chinese Children’s Mental Calculation
JAMES W. STIGLER

The University of Chicago

Skilled abacus operators report visualizing a mental image of the abacus, and
performing rapid mental calculation by manipulating the beads on their “mental
abacus.” Eleven-year-old Chinese children at three levels of abacus expertise
were observed performing both abacus and mental addition. Response times and
errors were examined as a function of problem type and mode of computation.
Chinese performance was compared with a group of American adults performing
the same mental additions. Chinese were also given a task which required them
to access intermediate states of mental abacus calculation. Results indicate that
abacus training has both quantitative and qualitative effects on children’s mental
calculation skill, and that a “mental abacus” is used by experts.

INTRODUCTION
Computation and mental computation are basic skills that are used
across a wide variety of cultures. How people learn and cognitively rep-
resent these skills, however, differs greatly as a function of the cultural
context in which they are learned and applied. In the United States, for
example, mental computation is not seen as a particularly important skill.
The skill is not routinely taught in American schools, and expertise in
mental calculation is acquired only by a small number of individuals,
using idiosyncratic techniques and working independently.
This work is based on dissertation research submitted to the University of Michigan in
partial fulfillment of the requirements for the PhD. It was funded in part by a University
of Michigan Rackham Dissertation Grant. I thank :he members of my dissertation com-
mittee: Harold Stevenson (Chairman), Keith Holyoak, Scott Paris, and Dan Bums, as well
as Craig R. Barclay, who was very involved with the planning of this study. The following
persons provided valuable help in expediting the collection of data in Taiwan: Lu Hui-then,
who acted as my research assistant; Dr. Mao Lian-wen, Commissioner of Taipei City
Schools; Professor Wu Wu-tien of National Taiwan Normal University; Dr. Hsu Chen-chin
of National Taiwan University Hospital; and Lai Sheng-shiung, abacus teacher at Dongyuan
Elementary School. I thank Tom Trabasso, Jon Baron, and Kevin F. Miller for careful and
constructive responses to earlier versions of this manuscript. And I acknowledge the helpful
suggestions of Dan Wagner, Richard Newman, Diane Enerson, Michelle Perry-Barras,
Susan Goldin-Meadow, Giyoo Hatano, Steve Shevell, Bill Goldstein, and Nancy Stein.
Correspondence and requests for reprints should be addressed to Jim Stigler, Committee
on Human Development, University of Chicago, 5730 S. Woodlawn Ave., Chicago, IL
60637.

145
OOlO-0285/84$7.50
Copyright 0 1984 by Academic Press, Inc.
. . . .~c_- -.- ..-.- I ._^. I^_ :_ ^_.. ‘-“-... _~-_ _.,- ,j
146 JAMES W. STIGLER

Asian cultures provide a strong contrast. In Japan and Taiwan, for

example, one finds organized institutional efforts to train children in
mental computation, both in school and in extracutricular programs. The
training is uniform across individuals, and revolves around a powerful
piece of technology: the abacus. Persons skilled at abacus calculation
report that they are able to construct a mental image of an abacus. They
then perform mental calculation by moving the “beads” on their “mental
abacus” as they would on a real abacus, and reading off the answer when
they are done. Abacus experts using this method are capable of as-
tounding feats. Mentally they can find the product of two S-digit numbers
in under 10 sec. They can look at a column of five 3-digit numbers and
figure the sum in about 3 set (Stigler, 1982).
Research into the nature of abacus calculation skills, and procedures
for mental calculation that directly result from training in abacus calcu-
lation, can be motivated within the context of several traditions of cog-
nitive research. One such tradition is that involving the nature and ac-
quisition of domain-specific expertise. This tradition is often traced to a
well-known study of the acquisition of telegraphic skill. In introducing
that study, Bryan and Hatter (1899) argued strongly for the development
of a “psychology of occupations.” “Occupation” was defined broadly
to include skill in many domains; acquiring skill in one or another of these
domains, they argued, is the chief engagement of everyone. Interest in
the psychology of expertise has increased recently (e.g., Chi, Glaser, &
Rees, 1982)due to a growing realization that much of cognitive processing
is inextricably tied to specific domains or knowledge bases.
Converging with the interest in expertise is an interest in the role cul-
ture-specific organizations of experience play in cognitive development
(e.g., Cole & Scribner, 1977; Vygotsky, 1978). Clearly, if the nature and
acquisition of domain-specific knowledge has a major impact on cognitive
processes, then new importance is ascribed to understanding the ways
different cultures structure both the knowledge that is acquired by its
members, and the contexts in which that knowledge is acquired.
What aspects of the environment become incorporated into human cog-
nitive skills? Abacus skill is a promising domain in which to examine this
question, for methodological as well as conceptual reasons. Abacus ex-
perts report that they internalize both a representation of the abacus, and
procedures for manipulating this representation for the purpose of cal-
culation. Of interest are the ways in which motor abacus skill and mental
abacus skill resemble each other. Motor abacus skill is easily described.
This description provides a solid basis for investigating which aspects of
motor abacus skill are retained in the mental skill. Because abacus skill
is culturally unique, it is possible to make comparisons with people using
different methods of mental calculation. Because becoming an abacus
 MENTAL ABACUS 147

expert requires extensive training and practice, it is possible to study the

developmental changes in representation that result from increased ex-
perience .
Although the phenomenon of mental abacus calculation has been re-
ported in the psychological literature (Ezaki, 1980; Hatano, Miyake, &
Binks, 1977), the number of studies is small. Recent studies by Hatano
and his colleagues at Dokkyo University in Japan (Hatano, 1981; Hatano
& Osawa, 1981) have investigated the effect of mental abacus training on
other cognitive skills such as digit memory and paper and pencil calcu-
lation. As yet, however, the nature of the representation involved in
mental abacus calculation, as well as implications of the mental abacus
for cognitive psychology remain largely unexplored.
The present research investigated abacus and mental abacus calcula-
tion skill among 1l-year-old children in Taiwan at various levels of ex-
pertise. The research had two specific purposes. The first was to docu-
ment the achievements of these children by gathering basic data on speed
and accuracy of abacus and mental calculation, as well as information on
the relation between levels of training and practice and resulting exper-
tise. The second purpose was to describe the nature of mental abacus
skill to determine what aspects of the physical abacus and its operation
are represented in the mental skill.
Three studies are described. Before presenting the three studies, how-
ever, it is first necessary to provide a brief introduction to the abacus
and abacus calculation. This description of abacus skill provides a basis
on which it will then be possible to outline some of the predictions and
hypotheses that motivated the empirical work.
The Abacus and Abacus Addition
The abacus most widely used today, and thus the subject of the present
study, is the Japanese abacus, or soroban. It is a modern adaptation of
the Chinese abacus, and took its present form about 1920 (Kojima,
1954a). The soroban is a wooden framed instrument made up of 23 col-
umns of beads, a subsection of which is illustrated in Fig. 1.
The abacus uses place value and the base-10 system of numeration to
represent numbers. Each column of beads has a place value, corre-
sponding to the ones, tens, hundreds, thousands, and so on. Which spe-
cific column of beads is defined as the “ones” is a matter of choice and
convenience, but once it has been defined, all other columns are valued
relative to it. The columns on an abacus thus correspond to the columns
of a base-10 numeral: 1452, for example, is represented on four adjoining
columns of abacus beads.
As can be observed in Fig. 1, each column of beads is divided into an
upper and a lower section. The bead in the upper section is equal to 5
 Actud Size
0

123456789
FIG. 1. The number 123,456,789as represented on the Japanese abacus.
 MENTAL ABACUS 149

times the units value of the column (5, 50, 500, etc., depending on the
place value of the column) when it is pushed down toward the dividing
bar that separates the two sections. If it is away from the dividing bar,
then it is equal to 0. Each of the four lower beads is equal to 1 times the
units value of the column (1, 10, 100, etc.) when pushed up toward the
dividing bar, and 0 when pushed away from the bar.
By pushing different combinations of beads toward the dividing bar, it
is possible to represent the numbers 0 through 9 on any single column.
For example, if the upper bead was pushed down and the four lower
beads all pushed up, the number 9 would be represented. If the upper
bead was pushed back up, but the lower beads left in place, the number
4 would be represented. Figure 1 shows the abacus representation of each
of the digits 1 through 9 in order, or a representation of the number
123,456,789.
Addition on the abacus can be described both in terms of bead move-
ments and in terms of the finger movements that serve to move the beads.
For my purposes here it will be sufficient to explain abacus addition only
in terms of the movement of beads.
The basic units of abacus addition are the single-digit additions. For
each single-digit number 0 through 9 that could be represented on a single
abacus column, one could add any of the single-digit numbers 1 through
9. There are thus 90 different possible single-digit additions. These single-
digit additions fall into four distinct groups: those that require no carrying,
those that require carrying within a column, those that require carrying
to the next column, and those that involve both kinds of carrying.
Additions that are carried out within a single column are those yielding
sums of less than or equal to 9. The simplest are those that do not involve
carrying: starting with the first addend on the abacus, the beads corre-
sponding to the second addend are pushed toward the dividing bar, and
the addition is complete. For example, adding 1 + 3 is a matter of pushing
one of the lower beads up to represent the 1, and then pushing three more
beads up to get the total of 4. A slightly more complicated example would
be the addition of 2 + 7. Two of the bottom beads are first pushed up to
represent the 2; then 7 is added by simultaneously pushing two more
lower beads up and the upper bead down.
Slightly more complex are those additions that require carrying within
a single column. Take, for example, the addition of 4 + 1. First the 4 is
represented by pushing all four lower beads up toward the dividing bar.
Since there are no remaining unit beads that can be pushed up, the ad-
dition of 1 requires using the upper bead (that is worth 5). In this case,
the four lower beads are pushed back down (ie. subtracted), and the upper
bead is pushed down to yield the answer of 5. (This all is carried out by
a single downward brush of the index finger.) The general rule for addi-
150 JAMES W. STIGLER

tions requiring a within-column carry is that when the second addend

cannot be directly added, then one simply subtracts the complement of
that addend to 5, and adds 5 by pushing down the upper bead. Or more
generally: if (X + y) cannot be added directly, one simply translates it
into (x - (5 - y) + 5).
Additions that yield sums greater than 9 require carrying onto the next
adjacent column. The general rule for these is that if (X + y) cannot be
added on a single column, one subtracts the complement of y to 10, and
then adds 1 to the next column (or (x - (10 - y) + 10)). A simple
example is adding 9 + 3: first 9 is represented by pushing the 4 lower
beads up and the upper bead down. Because 3 cannot be added directly,
7 is subtracted (by pushing the upper bead back up and two of the lower
beads back down), and 1 is added to the next (or 10s) column. The most
complicated additions require carrying both within a column and to the
next column. For example, in adding 5 + 6 one runs into a problem in
trying to subtract 4 (the complement of 6 to 10) since the 4 lower beads
are already in their zero position. To subtract 4, one must first subtract
5 (by pushing the upper bead up) and then add 1 with a lower bead. Then
1 can be added to the next column to complete the addition. A problem
of this complexity could be represented as (X + (5 - (10 - y)) - 5),
which in fact is all carried out with two simple finger movements.
As described thus far, it would appear that there is a large cognitive
component in what could be viewed as primarily a motor skill. I have
depicted the abacus operator as making a judgment about the type of
addition problem, mentally computing complements to 5 and to 10, plot-
ting the necessary bead movements, and planning the linger movements
that will realize the bead movements. It is important, however, to distin-
guish this theoretical description of abacus operation from the theory of
abacus employed by the operator. It appears that the actual cognitive
component of abacus operation varies greatly with both level of expertise
and the method by which one has been taught to use the abacus.
In many respects, the acquisition of abacus expertise appears to re-
semble the acquisition of expertise in many domains of cognition (see,
e.g., Bryan & Harter, 1899; Chi et al., 1982). With increasing expertise
comes a transition from conscious to automatic processing. As explained
earlier, there are 90 possible single-digit additions. The beginner, by un-
derstanding the theory of abacus operation as I have described it, might
proceed through a series of cognitive steps to determine the correct rule
in a given situation, and then apply the rule. The expert, on the other
hand, has a detailed knowledge of each of the 90 situations, and auto-
matically executes the appropriate moves. The expert does not need a
theory of abacus operation; the beginner may utilize such a theory to
guide his practice.
 MENTAL ABACUS 151
The extent to which a conscious, theoretical approach is employed by
the beginner will depend upon the philosophy of his teacher. Some
teachers teach a general, abstract theory, like the one I have outlined,
and encourage beginners to “reason out” the correct move when they
are uncertain of how to proceed. Other teachers may teach 90 rules to
correspond to the 90 single-digit addition situations, and expect students
to commit these rules to memory. Obviously, the cognitive processing
that goes on during learning will differ according to which approach is
being used.
Having discussed in some detail the single-digit additions, I will turn
briefly to the addition of multiple-digit numbers. Any addition problem,
no matter how large, is accomplished by executing sequences of the
single-digit additions. Numbers are added a digit at a time, moving from
left to right through each successive addend. For example, the addition
of 123 + 456 proceeds as follows: first the 1 is set on the abacus, followed
by the 2 and then the 3 to yield the first addend. Then the next addend
is processed from left to right: first the 4 is added to the 1; then the 5 is
added to the 2; and finally the 6 is added to the 3. At this point the
answer, 579, appears on the abacus. If there were a third addend, it would
be processed from left to right also, and added into the 579.
As is apparent from the preceding example, it is possible to take any
addition problem and generate the sequence of single-digit additions that
would be used to solve it on the abacus. Thus, the problem 123 + 456
would generate this sequence: 0 + 1, 0 + 2, 0 + 3, 1 + 4, 2 + 5, 3 +
6. One could also generate a list of the intermediate states the abacus
passes through in the solution of a problem, describing states with the
number represented on the abacus after each of the additions in the se-
quence. Again in the case of 123 + 456, the abacus would pass through
the following states: 000, 100, 120, 123, 523, 573, 579. (If carrying is
involved in the problem, an extra intermediate state occurs between the
initial subtraction from one column and the addition to the next column.)
Once the procedures of abacus operation are understood, it is relatively
straightforward to write a computer program that will generate the se-
quence of moves (single-digit additions) and intermediate states, given
any addition problem. Such a program has been written, and was used
as a tool in conducting the present research. This program takes as its
input any addition problem, and generates the sequence of moves and
intermediate states required for the solution of the problem on the abacus.
An example of the output of this computer program is shown in Fig. 2.
A more detailed explanation of this figure is presented in Study 2.
Abacus-based mental addition, as described by its practitioners, pro-
ceeds exactly the way abacus addition does. A mental image of the abacus
is visualized, and the beads are then manipulated exactly as they would
152 JAMES W. STIGLER

PROBLEM
232
779
648
511
MOVEMENT RESULT
*aL2
02 0200
Rl
03 0230
RI
02 0232
wlL2
27 0932
Rl
37 0902
c9 0002
co* 1002
R2
29 1001
CO% 1011
SL2
06 1611
Rl
14 1651
Rl
18 1659
=rL2
65 1159
ClR 2159
Rl
51 2169
Rl
91 2160
C6* 2170

FIG. 2. Output of computer program that generates the sequence of moves and interme-
diate states for solving addition problems on the abacus.

be on the actual abacus. The beginner at mental abacus calculation moves

his fingers to aid in moving the mentally imaged beads. The expert is able
to perform mental abacus calculations without moving the fingers.
Children in ‘E&wan are first introduced to the abacus as part of the
fourth-grade elementary school mathematics curriculum. (Training in con-
ventional methods of computation begins in first grade as it does in the
United States.) Children who wish to become experts undertake more
intensive training by enrolling in programs that meet after school. These
after school programs are usually associated with an elementary school,
and the participants engage in interschool competitions involving both
abacus and mental calculation. They also take certifying exams admin-
istered by the government-run Chinese Abacus Association which rates
participants according to their expertise in abacus and mental computa-
tion.
Abacus teachers in Taiwan disagree as to the best method for teaching
abacus-based mental calculation. Some believe that mental calculation
should not be introduced until the student is fairly advanced in motor
 MENTAL ABACUS 153
abacus skill. Others believe that mental abacus calculation can be intro-
duced early, and taught in parallel with motor abacus training. (See
Enerson & Stigler, 1983 for a discussion of this transition.) Whatever the
view, however, everyone seems to agree that mental abacus calculation
must be taught and practiced, and will not just arise as a natural conse-
quence of abacus training. (Readers desiring further information about
the abacus and its use may refer to Kojima, 1954a, 1954b; Moon, 1971;
Stigler, 1982; or Tani, 1964.)
Introduction to the Three Studies
Now that the workings of the abacus have been explained, it is possible
to give a brief, but specific, introduction to the hypotheses that guided
and motivated the studies to be presented. The research was aimed at
investigating what aspects of the physical abacus are incorporated into
mental abacus skill. The basic strategy was to describe abacus calcula-
tion, and then use this description as a basis for making predictions about
mental abacus performance. Of interest was the extent to which motor
and mental abacus skills would resemble each other, yet differ from the
performance of persons using some other method of mental calculation.
In the first study, children at three different levels of abacus expertise
were timed solving addition problems of varying widths (i.e., numbers of
digits per addend) and lengths (i.e., numbers of addends). Each child
solved the problems both mentally and using the abacus. The purpose of
this study was to gather general information that might shed light on the
similarities and differences between abacus and mental abacus calcula-
tion, and on the changes in skill that accompany the development of
expertise.
Several predictions were made. First of all, it was hypothesized that
the amount of time required to solve a problem using the abacus should
be a linear function of the number of total digits (i.e., digits x addends)
in the addition problem. The description of abacus addition shows that
any addition problem is simply a series of single-digit additions, each
requiring a single, stereotypical movement of the fingers. Once these
movements are automated, there is no reason to believe any one move-
ment would take longer than any other, and thus the time added per
additional digit should be a constant. According to this description, it
should not matter how the digits are arranged in terms of rows and col-
umns; the time per digit should still be constant.
A slight modification of this prediction about abacus calculation was
made in thinking about the characteristics of “mental abacus.” Insofar
as the mental abacus is an analog of the actual abacus, it would be ex-
pected to produce the same linear relationship between response time
and total number of digits in the problem. However, the fact that the state
154 JAMES W. STIGLER

of the abacifs must be constantly maintained in short-term memory

throughout the calculation adds an additional factor to the prediction of
mental response time. It was hypothesized that along with the linear
relationship of response time to total number of digits expected in abacus
calculation, mental calculation would be influenced as well by the width
of the image (in terms of the number of abacus columns required to carry
out the computation) that must be maintained during the calculation.
Width of a problem can be made independent of total number of digits
in the problem by varying the number of addends. For example, a
problem with a total of 10 digits could be constructed to be wide (as in
the addition of two S-digit addends), or narrow (as in the addition of five
2-digit addends). Kosslyn (1980) has shown that the more parts an image
contains, the more time is required simply to maintain the image in
memory. Therefore, other things being equal, it was predicted that wider
problems would taken longer than those requiring fewer columns of the
abacus.
Response times and frequency of errors for both abacus and mental
calculation were expected to decrease with level of expertise. Although
no specific predictions were made, the relative speed of abacus and
mental calculation was also of interest, both in general and as a function
of increasing expertise.
The second study explored the degree to which abacus operators could
access the intermediate stages of an abacus or mental abacus calculation.
It was hypothesized that if the mental abacus was in fact analogous to
the actual abacus, it should not only arrive at the right answer to the
addition problems, but should also pass through the same intermediate
states as the abacus does in the solution of a problem. This was tested
by designing a task which required subjects to answer questions about
intermediate states in a calculation that are unique to the abacus solution
of the problem. ‘Iwo predictions were made. First, it was predicted that
questions about intermediate states should be answered correctly, both
using the abacus and mentally. Second, it was predicted that the amount
of time required to answer such questions should be a linear function of
the position of the intermediate state requested in the sequence of states
leading to the solution of the problem. This latter prediction assumes that
intermediate states are accessed by actually carrying out each successive
manipulation of the abacus, and then reading the resulting state off either
the mental or actual abacus.
The third study compared Chinese abacus operators with Americans
mentally solving the same set of addition problems. Since none of the
Americans had any knowledge of abacus calculation, their performance
was not necessarily expected to conform to predictions based on a de-
scription of abacus skill.
 MENTAL ABACUS 155
It was predicted that the Chinese errors in mental calculation could be
shown to result from specific aspects of abacus representation. American
errors were not expected to show the same characteristics. Because the
abacus contains beads worth 5 as well as beads worth 1, it seems likely
that errors could occur that would put the answer off by exactly 5 more
often among the Chinese than would be expected among Americans using
a simple base-10 system of numeration. Another error expected to be
more frequent among the Chinese was the error of “omission.” Nu-
merous studies (Trabasso & Riley, 1975) have found that when linear
arrays of various types are represented in memory, discriminability is
poorer for items in the middle of the array than for items on either end.
For this reason it was predicted that Chinese employing the mental abacus
would tend to lose track of and omit digits from the middle of the answer,
whereas the Americans would not tend to produce this error. Positions
of errors (i.e., in which column of the answer they occurred) should be
similar between abacus and mental abacus calculation, but the positional
distribution of Chinese errors should not necessarily resemble that of the
Americans.
With respect to response times, it was predicted that the Americans
would be slower than the Chinese, since they were not considered experts
at mental calculation. While it was predicted that the Americans too might
produce a linear relationship between response time and total number of
digits in the problem, there was no reason to expect that the width vs
length of the problem would have any differential effect on American
response times. Most Americans do not report using a visual/spatial
image while engaged in mental calculation, and so the width of the
problem was seen to present no predictable difficulties apart from the
total size of the problem.
A Study of Abacus and Mental Addition Skill
This first study gathered basic data about abacus and mental addition
skills. Chinese children at different levels of expertise in abacus operation
were tested individually with a wide range of addition problems, both
with and without using an abacus. Accuracy of both abacus and mental
solutions. was examined as a function of type of problem and level of
expertise.
Method
Sf?tting. The study was conducted at Dongyuan Elementary School in Taipei, Taiwan.
Dongyuan is a large urban school with about 4500 students in grades 1 through 6. The
school itself is typical of public schools in Taipei, but it has an unusually good abacus
program (according to the Commissioner of ‘Ihipei schools).
Complementing the abacus program at the school is Dongyuan Buxiban, an after school
program directed by the abacus teacher from the elementary school. Dongyuan Buxiban is
156 JAMES W. STIGLER

devoted solely to abacus instruction for children from Dongyuan Elementary School. Chil-
dren who want to pursue additional abacus training can attend classes three afternoons per
week at the Buxiban for 1 to 1% hrs on each afternoon. Of the approximately 750 students
at each of grades 4 through 6, about 100from each grade elect to attend these abacus classes
after school. A small number of students in first, second, and third grade also attend classes
at the Buxiban. Members of the abacus team that will represent Dongyuan Elementary
School in competition are chosen from among those children who attend classes at the
Buxibnn. Anyone is welcome to join the Buxiban program, and a large cross-section of
children attend.
Subjects. Twelve Sth-grade students from Dongyuan Elementary School participated
in the study. They ranged in age from 10 years 9 months to 11 years 6 months. All were
considered good students, with average school grades (on a lOO-point scale) ranging from
85 to 95. The students were selected in consultation with the abacus teacher to represent
three levels of expertise in abacus skill: four experts, four intermediates, and four novices.
The four expert abacus operators were rated as duan wei by the Chinese Abacus Asso-
ciation, which is the highest possible rating (Zhusuan xuehui, 1980). All participate in the
Dongyuan Buxiban program; two have been in the program continuously since the beginning
of first grade, and the other two have been in the program since the beginning of the third
grade. All four reported first learning to add on the abacus when they were in the first
grade. Two were boys and two were girls. They were among 22 students at Dongyuan
Elementary School who had a duan wei rating.
The four intermediate abacus operators were rated as Grade 5 by the Chinese Abacus
Association. They also were participants in the Buxiban program, but had not been at-
tending for as long as the experts had. One of the four started going to the Buxiban during
the second half of the third grade; the other three started in fourth grade. They also reported
first learning to add on the abacus later than did the experts: two said they first learned in
third grade, and two in fourth grade. All of the intermediate group were boys.
The four novices had no rating from the Chinese Abacus Association. Three of the four
had attended Buxiban in the past, but had stopped going for lack of interest. Two reported
first learning to add on the abacus in fourth grade, one in third grade, and one in second
grade. Two were boys and two were girls.
Subject-selection factors. Whenever novices and experts are compared, the pos-
sibility of subject-selection factors must be assessed. In the current study, expert abacus
operators are shown to be capable of astounding feats of mental calculation. Is it solely
due to training, or do these experts start out with supernormal skills?
The nature of subject-selection factors was assessed by means of a comprehensive survey
of the 714 fdth-grade students in Dongyuan Elementary School. For each student the fol-
lowing information was collected from school records: mother’s and father’s education
level, sex, score on Raven’s Colored Progressive Matrices when in first grade, average
grade in mathematics while in first and second grade, number of hours spent in extracur-
ricular abacus training up through the fifth grade, and current level of expertise achieved
in abacus and mental abacus calculation (as rated by the abacus teacher).
Of the 714 students in the fifth grade, 104 had enrolled at some point in the after school
abacus training program. The binary variable of whether or not training had been elected
was regressed on five predictor variables: sex, mother’s education, father’s education,
average early math grade, and the Raven’s Progressive Matrices score. The general lack of
a strong selection bias was demonstrated by the fact that for the overall analysis, R2 = .04.
Of the independent variables, only one was statistically significant: early math grades pro-
duced a beta weight of .16, t = 3.80, p < .OOl.
One further analysis was carried out only on those students who had attended the abacus
training program. Their level of abacus and mental abacus expertise (measured on a l- to
 MENTAL ABACUS 157
lo-point scale) was regressed on six predictor variables: the first five were identical to those
used in the first analysis, the sixth was total number of hours spent in abacus training. This
time the RZ statistic was .70. However, most of the variance was explained by the single
predictor that was statistically significant, total hours of training. The B weight for total
hours was equal to .78, t = 13.04, p < .OOl.
It is thus reasonable to conclude that subject-selection factors played little if any role in
the results to be reported here.
Problem ser. A total of 256 addition problems were constructed as follows. A matrix
of 16 cells was formed by crossing two factors: number of digits per addend (2, 3, 4, or 5
digits), and number of addends per problem (2, 3, 4, or 5 addends). Sixteen problems were
constructed for each cell of the matrix. Addends were randomly generated (with replace-
ment) from the population of nonnegative integers with the required number of digits. These
256 problems were then formed into eight sets. Each of the sets of problems was constructed
by randomly choosing two problems from each cell of the matrix, and randomly ordering
the resulting 32 problems. The eight sets were presented to each subject in the same order,
one through eight consecutively, for both abacus and mental calculation.
Procedure. Experts and intermediates performed both abacus and mental addition on
the same sets of problems. Order of presentation- abacus versus mental-was balanced
so that half the sets of problems were first calculated using the abacus, and half were first
calculated mentally. Abacus and mental calculations using the same problem set were sep-
arated in time by at least 1 day. Testing was divided into three sessions on different days
within a l-week period. All of the experts and two of the intermediates did all eight of the
problem sets (or all 256 problems) both using the abacus and mentally. The other two
intermediates did only six of the sets (or 192problems), both using the abacus and mentally.
These subjects did fewer problems because they were slower, and found the work more
difficult than did the other subjects. It was decided to reduce the number of problems in
order to reduce fatigue and frustration.
The novice group did only abacus calculation. They did two of the probelm sets (or 64
problems) in each of three individual testing sessions on 3 different days within a l-week
period of time. In all, they were presented with 192 problems each, 12 of each of the 16
problem types. They were asked to try the mental addition problems, but it quickly became
apparent in each case that many of the problems were too difficult, and to complete even
one set would be a frustrating and painful experience.
All problems were presented using an Apple II computer and a Sanyo DM5 112CX green
video display. The procedure for mental calculation was as follows: when the subject was
ready for the next problem, he pressed the space bar on the Apple keyboard. The press of
the space bar simultaneously presented the problem on the display and started a millisecond
timer. The subject was instructed to solve the addition problem mentally “as quickly as
possible without making too many mistakes,” and to press the space bar again as soon as
he had the answer. Subjects were instructed to begin saying the answer verbally as they
were pressing the space bar. The second press of the space bar simultaneously stopped the
millisecond timer, recorded the solution time, and erased the problem from the screen. The
experimenter used an answer sheet to check whether the solutions were correct, and re-
corded all incorrect answers on a protocol for each subject. Subjects set their own pace
through each of the problem sets, and were given a rest between problem sets.
The procedure for abacus calculation was only slightly different from that for mental
calculation. After the first press of the space bar, there was a 2-set delay before the problem
appeared on the screen and the timer was started. This was to allow time for the subject
to get his hand positioned correctly on the abacus before the problem was presented. As
soon as the abacus calculation was completed, the subject was instructed to press the space
bar as quickly as possible, and then read the answer from the abacus.
158 JAMES W. STIGLER

Subjects were given as much practice on simple problems as they needed to become
proficient at using the apparatus. The procedure was easily mastered by all subjects, and
none required more than 10 practice problems each for the abacus and mental calcuiation
sets.

Results
Intraindividual analyses. Depending upon the number of problem sets
done, each subject generated between 192 and 256 data points for each
mode of calculation, abacus and mental. Because of this large amount of
data, it was possible to carry out the first analyses separately for each
child. Since the number of subjects was small, it seemed sensible to do
some preliminary analyses that would validate the combining of subjects
into novice, intermediate, and expert groups.
Practice (or fatigue) effects were first evaluated by analyses of variance
of response times across problem sets, separately for each subject and
mode of computation. As no significant effects were found, problem set
was dropped as a factor from all further analyses.
In a study of a single adult abacus expert which was conducted as a
preliminary to the present study (Stigler, Barclay, & Aiello, 1982), re-
sponse time was shown, as predicted, to have a strong linear relationship
with the total number of digits in the addition problem. Number of digits
was computed by multiplying the number of digits (per addend) by the
number of addends. The first step in analyzing the present data was to
assess the replicability of this earlier finding for each of the 12 subjects.
In the present data set, total digits in the problem ranged from 4 (two
addends of two digits) to 25 (five addends of five digits).
Regression analyses were conducted in which response times were
regressed on the total number of digits, separately for each subject.
Abacus calculation times for each subject were strongly and linearly re-
lated to total number of digits in the problem. The 12 R* statistics ranged
from .71 to .91, with the median R* = .84.
Mental calculation times for the eight intermediates and experts also
showed strong linear relationships with total number of digits in the
problem. However, the percentage of variance accounted for was not as
great as with the abacus calculation times. R* statistics for mental cal-
culation ranged from .39 to .77, with the median R* = .65. Further anal-
yses revealed that adding higher order components to the regression so-
lutions did not add significantly to the percentage of variance explained
for either abacus or mental calculation times.
Because the response times were a linear function of total number of
digits in a problem, each subject’s overall speed of computation can be
characterized in terms of the slope of the predicted regression line (msec
per digit). The y-intercepts of the regression solutions are assumed to
 MENTAL ABACUS 159

TABLE 1
Slopes from Individual Regression Analyses of Response Time (msec) on Total Number
of Digits in Problem
Mode of calculation
Level of expertise Abacus Mental
Experts 376 264
428 312
370 321
596 294
Intermediates 663 580
877 918
1051 1277
836 1018
Novices 670 -
1572 -
1527 -
1062 -

include such components as preparation for responding, which would not

necessarily be equivalent for the abacus and mental versions of the task.
The slopes of the regression lines are presented in Table 1, separately for
each subject and for each mode (abacus vs mental) of calculation. As
seen in the table, each of the experts was faster (per digit) at abacus
calculation than was each of the intermediates, and each of the inter-
mediates was faster than three of the four novices. (One of the novices
was actually quite good at abacus calculation, yet this same subject
showed great difficulty in mental calculation, and was not able to com-
plete even one problem set.) All of the experts were faster at mental
calculation than were each of the four intermediates. The validity of ca-
tegorizing the children into novice, intermediate, and expert groups was
thus confirmed.
Error rate. Based on the preceding analyses, it was decided to carry
out further analyses at a group level, averaging across individuals within
each level of expertise. Error rates are presented in Fig. 3, broken down
by problem type, level of expertise, and abacus vs mental calculation.
The graph consists of four panels, with number of digits per addend
constant within each panel at 2, 3, 4, or 5 digits. Solid lines represent
abacus calculation, and dotted lines mental calculation.
The error rate for experts is quite low both using the abacus and men-
tally for 2- and 3-digit addends, with percentage incorrect less than 5%.
The rate rises for 4- and 5-digit addends, and increases more markedly
for mental computation than it does for abacus. Intermediates show
 JAMES W. STIGLER

TWO THREE fOUff

IOOr D/G/KS r D/G/?3 ,- D/G/KS

NUMBER OF ADDENDS

FIG. 3. Error rates as a function of digits and addends for Chinese abacus operators.

higher error rates than do experts, and higher rates mentally than using
the abacus in all panels of the graph. The novices show surprisingly
accurate performance, which may be partially due to their slow speed of
calculation, indicating more caution.
Before moving to further analyses of response times, the question may
be raised as to whether the pattern of results would be biased if latencies
for cot-rent and incorrect responses were averaged and analyzed in a
single analysis. Given the high error rate in some cells of the design, there
are obvious practical reasons for combining times for correct and incor-
rect responses. It is reasonable to argue that subjects are engaged in
similar processes on correct and incorrect trials. After all, only one step
need be done incorrectly in order to get the wrong answer. Long addition
problems that require the addition of 25 total digits offer many more
opportunities for error than a shorter problem would, which could ac-
count for the high error rates on these problems. An empirical test of the
effect of combining correct and incorrect response times was conducted
as follows. Response times were standardized within subject, mode of
calculation (abacus or mental), and each of the 16 types of problems.
Analyses of variance of these standardized times showed no significant
differences between correct and incorrect responses for any of the three
groups for either abacus or mental calculation, all p’s > .05. Because of
these findings, correct and incorrect response times are not separated in
any of the subsequent analyses.
Analyses of response times by group. Response times for each type of
problem, averaged across the four subjects in each group, are plotted in
Fig. 4 for both abacus and mental calculation. The graph in Fig. 4 fohows
the same format as the one in Fig. 3.
There are several things worth noting in Fig. 4. As expected, with
number of digits constant, response times are close to a linear function
of number of addends. The three levels of expertise are clearly differ-
 MENTAL ABACUS 161
FIVE
DIGITS

,
/ ,p’

.**
/ ,*“-

1
2345 2345 2 3 45 2345
NUMBER OF ADDENDS

FIG. 4. Response time as a function of digits and addends for Chinese abacus operators.

entiated from each other on the graph. The experts display a constant
advantage of mental over abacus calculation across all problem types.
For intermediates, the relation of abacus to mental calculation is mixed:
for narrow problems (i.e., 2 digits per addend), the mental response times
appeared faster, while for wider problems the two modes of computation
did not appear to differ appreciably.
Alternative plot of response times. As has been demonstrated, a large
part of the variance in response time is accounted for by the total number
of digits in the problem, regardless of how these digits are arranged. In
this section the effect of different arrangements of the digits in a problem
is examined. Specifically, the question is addressed as to whether the
width of the problem (i.e., digits per addend) has any effect over and
above the effect of total digits. It was hypothesized that such a width
effect may occur in mental abacus calculation as a consequence of the
unique form of the mental representation used to store intermediate states
in the solution process.
In Fig. 5 response times are plotted separately for abacus and mental
calculation. Time (on the y-axis) is plotted against the total number of
digits in the problem (on the x-axis). Each problem width (2, 3, 4, or 5
digits per addend) is represented by a separate line. If the width of the
problem did not have any effect on response time (controlling for total
digits in the problem), then the four lines for each width, within each
group of subjects, should coincide on this graph.
It is apparent from Fig. 5 that the plots for abacus calculation show a
markedly different pattern than those for mental calculation, The separate
lines for each width of problem coincide to a high degree for both inter-
mediates and experts performing abacus calculation, implying that the
form of the problem (width vs length) does not effect response time in-
dependently of the total size of the problem. For mental calculation a
different picture emerges. The slopes of the lines representing the four
162 JAMES W. STIGLER

20 ABACUS MENTAL
T T

4 6 B IO 12 14 16 IS 20 4 6 8 IO 12 14 I6 18 20
TOTAL DIOITS IN PROBLEM

FIG. 5. Response time as a function of total digits in problem, separately for problems of
varying widths and abacus vs mental calculation.

problem widths appear to differ, causing the lines to fan out on the graph
for both experts and intermediates, although more for the latter than the
former. This different picture for mental calculation than for abacus cal-
culation appears to support the prediction of a width effect imposed by
the nature of the mental abacus.
A statistical test for what has been graphically demonstrated was car-
ried out by use of a repeated measures analysis of covariance. Total
number of digits in the problem was covaried out separately (with slopes
allowed to vary) for each subject and mode of calculation. SUBJECT was
a random factor nested within two levels of expertise (GROUP, inter-
mediate and expert) and crossed with two repeated measures factors,
MODE (abacus vs mental) and number of digits per addend (WIDTH).
The MODE x WIDTH interaction provided the appropriate test, and
was significant, F(3,18) = 4.88, p < .025. The higher order MODE x
WIDTH x GROUP interaction was not significant, F(3,18) = 1.11,
p > .lO.
A Study of Access to Intermediate States
Solving any addition problem on the abacus will cause the abacus to
pass through a sequence of intermediate states specific to the problem
being solved. If, in fact, abacus experts use a mental image of an abacus
exactly as they would an actual abacus, the mental image should pass
through the same states an abacus does when the same problem is being
solved. The correct answer to an addition problem could be arrived at
by using many different methods. If, however, subjects can answer ques-
tions about intermediate states unique to the abacus solution of a
 MENTAL ABACUS 163

problem, it would provide clear evidence that they are in fact manipu-
lating a mental image as they would an actual abacus. The study that
follows provides a test for these ideas.
Method
Subjects. The subjects used in this study were the eight Sth-grade children in Taiwan
who comprised the intermediate and expert groups from the study reported above. Data
for this study were collected during the week following the previous study, in a single
additional testing session for each subject. Novice abacus operators were not included,
since they had essentially no “mental abacus” skill.
Construction of the task. Two sets of 35 addition problems were randomly gen-
erated for the task. The first set was used for an abacus version of the task, the second for
a mental abacus version. All problems were comprised of four 3-digit numbers. After the
70 problems were generated, each was analyzed into a sequence of intermediate states using
the computer program described above.
For the purpose of designing the task for this study, the intermediate states passed through
in solving an addition problem on the abacus were categorized as either major or minor.
Both types of intermediate states can be found in Fig. 2. Major states are those that appear
after the addition of each complete addend. For example, in adding 5 + 3 + 1, the major
states would be 5,8, and 9 (which is the answer). Major states are preceded on the computer
output by a double asterisk (**) in the “move” column; in Fig. 2, the major states are 232,
1011, 1659, and 2170. Since all problems had four addends, they would all have four major
states, including the answer. Major states are of limited usefulness in the present study,
since they could be determined by any method of addition that added one addend at a time.
In other words, major states are not unique to abacus calculation.
Whereas major states follow the addition of a complete addend, minor states occur after
the addition of each digit of an addend. Most of the numbers in Fig. 2 that lie between the
major states represent minor states. Minor states are unique to abacus addition, and would
not result from any other common method of calculation. It is minor states, therefore, that
form the basis of the task in the present study.
A probe card was constructed for each of the 70 problems by making a 3 x 3-in. black-
and-white photograph of four adjacent abacus columns, with some number represented on
the abacus. Photographs were pasted on white 4 x 6-in. index cards. Within each set of
35 problems, 30 probes depicted minor states in the solution of a problem, and 5 were
distractor probes that depicted a number that was not an intermediate state, but was near
in value to an actual intermediate state. The 30 probes depicting minor states were divided
into three groups of 10 states each from the beginning, middle, and final thirds of the list
of intermediate states for each problem. The 35 problem/probe pairs in each set were placed
in random order.
Procedure. Each problem was presented as follows: the subject was fust shown the
probe card depicting an intermediate state, and allowed to study it as long as he wished.
When the subject was ready, he was instructed to present the accompanying addition
problem on the video screen by pressing the space bar on the Apple keyboard. As soon as
the problem appeared, the subject began to add, searching for the intermediate state de-
picted in the probe (which remained visible throughout the problem). As soon as the subject
could determine whether or not the probe state was an intermediate state in the solution
of the problem, the subject pressed the space bar again, simultaneously answering “yes”
or “no.” This second press of the space bar caused the problem to disappear from the
screen, and recorded the response time. The experimenter recorded whether or not the
response was correct, and then presented the probe for the next problem. The subject
164 JAMES W. STIGLER

TABLE 2
Average Percentage of “Yes” Responses
Probe type
Minor Distractor*
Experts
Abacus 90.9 5.0
Mental 92.5 5.0
Intermediates
Abacus 85.8 20.0
Mental 93.4 50.0
* Correct answer for distractor probes is “no.”

studied the probe at his leisure, and when he was ready to continue pressed the space bar
again to present the next problem.
All subjects did both sets of 35 problems. The first set was always done using the abacus,
the second set manually. Other than mode of solution, the procedure was the same for each
set. Each set was preceded by three practice problems, and none of the subjects had
difftculty understanding the task after doing the practice problems. All practice problems
used probes for minor states.

Results
Accuracy. The first results concern the accuracy with which children
were able to distinguish probes depicting actual intermediate states from
distractor probes. Table 2 presents the percentage of “yes” responses
(averaged across subjects within level of expertise separately for abacus
and mental sets) broken down according to probe type (i.e., minor or
distractor).
Experts experienced no difficulty distinguishing the probes depicting
minor states from the distractor probes. Both using the abacus and men-
tally, experts responded “yes” (correctly) to over 90% of the probes
depicting minor states, but to only 5% (incorrectly) of the distractor
probes. Chi-square tests comparing responses across the two probe types
were carried out separately for each subject, for both abacus and mental
sets. All of the eight tests were significant, with p < .OOl.
Intermediates performed more poorly than did the experts, and more
poorly mentally than they did when using the abacus. While they re-
sponded “yes” to the minor state probes almost as often as did the
experts, their “false alarm” rate (responding “yes” to distracters to
which the response should have been “no”) was much higher than the
experts, at 20% using the abacus and 50% mentally. Most subjects were
still, however, doing better than would be expected by chance. Separate
x*s for each subject revealed that using the abacus, three of the four
 MENTAL ABACUS 165

Intermediate
.- -. Abacus
o---o MantaI
Expert

POSITION OF PROBE IN PROBLEM

(THIRDS)

FIG. 6. Response time as a function of probed intermediate state.

intermediates were responding “yes” significantly more often to the

minor state probes than they were to the distracters (all p’s < .OOl).
Mentally, two of the four x*s achieved significance (p < .OOl) and one of
the remaining ones was marginally significant (p < .lO).
Response times. As described earlier, minor states were probed in the
first, middle, and bottom thirds of the sequence of intermediate states.
Each subject received 10 problems probing each of the three positions,
or 30 minor state probes. If subjects were searching for intermediate
states by going through each step of the abacus solution sequence, it
should take longer to find and verify states further along in the sequence.
Response times for minor state probes are shown in Fig. 6. Each point
represents the average time in seconds across the four subjects at each
level of expertise, calculated separately according to mode (abacus vs
mental) and position of the probe in the sequence of intermediate states
(first, second, or third). Experts make the judgments faster than do the
intermediates; judgments are made more quickly when solving the task
mentally than when using the abacus; and it takes longer to respond when
the probe is positioned further along in the sequence required to solve
the problem on an abacus.
These results were confirmed by a 2 (group) x 2 (mode) x 3 (position
of probe) analysis of variance. Average times for each subject (across the
166 JAMES W. STIGLER

10 observations in each cell) were entered into the analysis. Main effects
for group, F(1,6) = 11.06, p < .02, mode, F(1,6) = 86.32,~ < .OOl, and
position of probe, F(1,12) = 14.37, p < .OOl, were all significant. None
of the two- or three-way interactions was significant (all p’s > .20).
Abacus operators are clearly able to make correct judgments about
intermediate states unique to abacus addition, whether using the abacus
or a mental abacus to solve the problem. They make these judgments
fairly quickly, and even more quickly when they are solving the problem
mentally than when using the abacus. Both mentally and using the
abacus, the time required to verify an intermediate state is a function of
how far down in the solution sequence the state lies. And as expected,
the intermediates were not as accurate or as fast in their judgments as
were the experts.
A Comparative Study of Chinese and American Mental Addition
The previous experiment provided some direct evidence of the abacus-
like nature of mental addition processes in Chinese abacus operators.
The intermediate states task, however, is somewhat unnatural in its de-
mands. Indeed, it may be the first time these children had ever been
asked to respond with some byproduct of the solution process rather than
with the answer. In fact, the response time for answering questions about
intermediate states in the final third of a problem with four 3-digit addends
was 2-3 set slower than just giving the answer to a similar problem would
be. One might argue that while the intermediate states task shows these
abacus operators are capable of manipulating a mental image of an
abacus, it does not show that they do in fact do so when adding in a more
natural situation.
The purpose of the next study was to provide evidence for abacus-like
addition processes in a mental addition task conducted under more nat-
ural conditions. The strategy was to compare the performance of Chinese
abacus operators with that of Americans who had no experience with the
abacus. Ideally, one would expect that mental calculation performance
would differ markedly across the two groups, and that furthermore the
mental calculation performance of the Chinese would bear striking re-
semblance to their motor abacus calculation performance.
While cognitive psychological models of western mental addition have
been reported in the literature (Ilitch, 1978), the logic of the present study
did not require a model of the American’s mental addition processes. All
that was required was that the Americans be shown to differ from the
Chinese in as many ways as possible. Similarities in mental calculation
performance across the two groups would have been uninterpretable,
since they could result either from similarities in the task (i.e., both
groups are adding the same arabic numerals presented on a video dis-
 MENTAL ABACUS 167

play), or similarities in mental processes used in solving the task. Differ-

ences, on the other hand, could be assumed to reflect differences in
cognitive processes. Once differences were demonstrated, the next step
was to assess the extent to which the physical abacus and procedures for
its operation could provide a model (albeit an informal model) for pro-
cesses used by the Chinese in mental calculation. To summarize, two
predictions were made: first, that the Chinese would differ from the
Americans, and second that the nature of the differences could be ac-
counted for using a description of abacus skill.

Method
Subjects. Chinese subjects were the eight Sth-grade children from Taiwan who com-
prised the intermediate and expert groups from the studies reported above.
Four American subjects were recruited to participate in the study. ‘IIvo were mathematics
and accounting majors from the University of Michigan, both in their senior years. One
was a four-year graduate student in psychology, and one was an administrative assistant.
All considered themselves fairly good at mental computation, although none should be
considered experts in the sense that the Chinese experts were.
Comparisons are made, therefore, between Chinese 1I-year-olds and American adults.
The age difference between these two groups was considered less important than the fact
that different methods of mental addition were used by the Americans and by the Chinese.
All the American subjects reported using the same method of mental addition: working
from right to left, they added the numbers of a column at a time, remembering each suc-
cessive digit of the answer while working on the next column. This method follows the way
most people would add with paper and pencil. None of the American subjects had any
experience with abacus calculation.
Procedure and problem set. The ,data for the Chinese group were the same as
those reported in Study 1. That is, the data were derived from the mental addition of the
eight sets of problems of various types.
American subjects were given the first six sets of addition problems (192 problems) from
Study 1, and were asked to calculate the solutions mentally. Testing was carried out indi-
vidually, using the procedure for presentation previously described. Subjects received two
sets of problems (a total of 64 problems) at each of three sessions held on separate days
within a single week.

Results
Speed and accuracy. Response times for mental addition by Chinese
abacus experts, intermediates, and Americans are presented in Fig. 7.
Each point represents the average solution time for a type of problem,
averaged across the four subjects in each group. The four successive
panels report the addition of 2-, 3-, 4-, and Sdigit addends, respectively.
Two characteristics of this graph are worth mentioning. First, it is clear
that the American adults are slower in mental addition than are the
Chinese fifth graders. Second, it is apparent that given a constant number
of digits per addend, the addition of each additional addend adds a fairly
constant amount of time to the total for all three groups. This confirms
168 JAMES W. STIGLER

rwo THREE FOUR F/YE

46 -DIG/TS _ DIGITS r D/G/TS r DIG/?-S
46 - I
44 - :
42 :
-Chinese Experts
40 M Chinoe Intmsdiates
38 c-e Amsricons - i
36 - :
34 -
32 -
30 -
26 -
26 -
24 -
22 -
20 -
IS -
16-
14 -
12-
IO -
a-
?’

II,, ,,,I ,I,,

2345 2345 2345 2345
NUMBER OF ADDENDS

FIG. 7. Response time as a function of digits and addends for Chinese and American mental
addition.

for the Americans what already has been demonstrated among the abacus
operators: the total solution time is a linear function of the total number
of digits in the problem. Thus, the time advantage of the Chinese abacus
experts over the Americans becomes greater as the number of digits in
the problem increases. For example, the experts add two 2-digit numbers
in 1.3 set, the Americans add these digits in 2.7 sec. Adding five j-digit
numbers, however, takes the Americans 47.0 set, and the Chinese ex-
perts just 7.4 sec.
Figure 8 presents the percentage of solutions that were incorrect, using
the same format as Fig. 7. Each point represents the average proportion
incorrect for a problem type, averaged across the four subjects in each
group. The great speed of solution demonstrated by the abacus experts
is accompanied by a rather low error rate. For 2- and 3-digit addends,
the proportion incorrect is negligible for the experts. While on the Sdigit
five addend problems the experts get the wrong answer on nearly 40%
of the problems, they still are doing notably better than are the Americans
or intermediate abacus operators.
The time advantage of the intermediate abacus operators over the
Americans does not lead to a corresponding advantage in terms of error
rate. In fact, as problems become more complex, the Americans seem to
gain an advantage over the intermediate group in accuracy. This may be
 MENTAL ABACUS 169

FIVE

a---a Americans

NUMBER OF ADDENDS

FIG. 8. Error rates as a function of digits and addends for Chinese and American mental
addition.

due to the fact that the Americans are adults, who tend to become more
cautious as the problems become more difficult.
Analysis of errors. Thus far a quantitative difference between the per-
formance of Americans and that of the Chinese abacus operators has
been demonstrated. In this section, qualitative differences in the types of
errors made by Chinese versus Americans are discussed, and these dif-
ferences are related to differences in method of calculation. An analysis
of the demands of abacus-based mental calculation led to three different
predictions concerning types of computation errors that might be ex-
pected from the Chinese subjects. Each of these predictions, along with
the results, is now presented.
1. Answers arrived at by abacus computation should be off by a unit
of five more often than would answers resulting from a different method
of calculation. The reasoning behind this prediction is as follows. As
described earlier, each column of the abacus contains a five bead and
four unit beads. In adding numbers on an abacus, it is often necessary
to carry not only across columns, but within a column as well, exchanging
units for fives. This feature of the abacus could lead to several possible
sources of error. First, in executing a within-column carry, one might
forget to enter the five bead after removing the unit beads. Second, it
seems likely that random beads could be accidentally knocked or moved
in the course of abacus computation. One would expect that a certain
percentage of accidental moves would involve the five bead. Third, in
reading the answer from the abacus (mental or physical) it seems likely
that configurations having the same lower bead positions but a different
upper bead position would be particularly easy to confuse visually. For
example, 2 and 7 both have two lower beads pushed up, and differ only
in the position of the upper (five) bead. It would thus seem easier to
confuse 2 with 7 than, say, 2 with 8.
170 JAMES W. STIGLER

TABLE 3
Analysis of Errors
Chinese American
Abacus Mental Mental

Total number of errors analyzed* 208 481 244

Percentage confusions of five 22.6 20.4 3.7
Percentage omissions - 11.9 3.7
Percentage occurring in position:
Left 10.2 10.5 13.3
Middle 58.3 59.9 67.9
Right 31.5 29.7 18.8
* Total number of errors should not be compared, since each is based on a different
number of problems attempted.

All of these postulated sources of error are unique to abacus addition,

and would lead to a digit in the answer being off by five. To test this
prediction, each error was coded for the presence or absence of a “five
confusion.” The coding criteria were as follows: an incorrect answer, to
be judged as evidencing a confusion of five, had to have the same number
of digits as the correct answer, and be off in any column by exactly 5.
For example, if the correct answer to a problem was 4138, incorrect an-
swers such as 4638, 9638, or 4133 would all be coded as having a con-
fusion of five. An answer like 41338, on the other hand, would not be so
coded, even though there might be some element of five confusion in the
answer.
The percentage of errors showing confusions of five is presented in the
second row of Table 3, separately for Chinese abacus calculation, Chinese
mental calculation, and American mental calculation. The results strongly
support the prediction outlined above. Confusions of five occurred about
six times as often in the Chinese group as they did among the Americans,
and, importantly, the percentage was almost the same whether the
Chinese were using the abacus or the mental abacus. The overall x2 was
highly significant, x2(2) = 39.91, p < .OOl; the x2 for Chinese only (abacus
vs mental) was not significant.
2. While confusions of five were predicted based on how the abacus
works, the next prediction was based on the memory demands one might
expect when using a mental abacus for mental addition. The Chinese
reported representing the intermediate results in a constant visual image.
It was predicted that because the columns of an abacus look very similar
to each other, adjacent columns in the image might tend to blur together,
and thus lead to more omissions of digits in the answer (e.g., 156, instead
of 1526) by Chinese than by Americans. This difference was supported
 MENTAL ABACUS 171
by observations of Americans versus Chinese while engaged in the task.
The Americans seemed to know how many digits should be in the answer
even when they were not sure what all the digits were; they might answer,
for example, “two, three, five, something, eight,” filling in the missing
place with a marker (“something”). The Chinese never answered in this
way, and showed no awareness of anything being wrong if they left out
a digit in the answer.
Each incorrect answer was coded for presence or absence of omissions.
Errors were coded as omissions if (1) they had fewer digits than the
answer should have, and (2) the remaining digits formed part of the cor-
rect answer. An example of an omission would be answering 3289, when
the correct answer should be 32589.
The third row of Table 3 presents the percentage of mental calculation
errors coded as omissions, separately for Chinese and Americans. The
prediction was born out: the Chinese made three times as many errors
of omission as did the Americans, x*(l) = 13.03, p < .OOl.
3. The final categorization of errors was according to the position of
the error in the answer, i.e., whether the answer given was incorrect in
the left, middle, or right digits. Addition by both abacus and nonabacus
methods is carried out a step at a time, and making an error at an early
stage of a calculation could easily throw off all of the remaining steps.
Because the Americans were working from right to left, and adding a
column at a time, it would seem reasonable that they would make fewer
errors in the rightmost column of the answer, since as the earliest result
obtained, it would be less likely to fall prey to errors compounded from
earlier steps. The Chinese, working left to right and constantly updating
each digit of the answer, should make more errors than the Americans
in the rightmost column.
Errors were coded as to position in the following way: If the leftmost
digit of an answer was incorrect, it was coded as a left position error. If
the rightmost digit of an answer was incorrect, it was coded as a right
position error. If any of the intervening digits was incorrect, a middle
position error was coded. If an answer was incorrect in more than one
position, it was coded more than once. The distribution of errors ac-
cording to position is presented in the bottom three rows of Table 3.
As predicted, the Americans did in fact show fewer errors in the right
position than did the Chinese. The most striking thing about the table,
however, is the obvious similarity in the distribution of errors by position
between the Chinese using the abacus and the Chinese using the mental
abacus. A x2 confirmed that there was no difference in distribution across
the two modes of addition among the Chinese, x2(2) = .349, p = .84.
The Chinese distribution did, however, differ significantly from the Amer-
172 JAMES W. STIGLER

40-
t
Digits par&Mend
-*
a 32
36:: ......... 3 , /h ,i
:, ---
-.-. 4 // .’
p 2*.. r-1 5 / /
// .i

4 6 a IO 12 14 I6 18 20
TOTAL DIOITS IN PROBLEM

FIG. 9. Americans’ response time as a function of total digits in problem, separately for
problems of varying widths.

icans; the overall analysis produced x2(2) = 16.00, p < .Ol, a finding
which again is quite obvious in the table.
A “width” effect for Americans? One additional analysis of response
times supports the notion that American mental calculation is qualita-
tively, as well as quantitatively, different from Chinese abacus-based
mental calculation. In the earlier study of Chinese abacus and mental
addition (reported as Study I), it was found that replotting response times
as a function of total digits in the problem revealed an interesting differ-
ence between abacus and mental calculation. For Chinese mental cal-
culation, increasing the number of digits per addend, or “width” of the
problem, increased response time (holding constant total number of digits
in the problem). “Width” did not, however, effect the times for abacus
calculation. The graph that demonstrated this finding was presented as
Fig. 5.
In Fig. 9 the corresponding graph for the American subjects’ mental
calculation times is presented. Comparing this graph with the one in Fig.
5 reveals a striking difference: the four lines representing the different
problem “widths” appear to approach coincidence for the Americans, as
would be predicted if response time was completely a function of total
number of digits in the problem. This is similar to the findings for Chinese
abacus calculation, but quite different from the Chinese mental calcula-
tion findings.
A statistical test for what is apparent in the graphs was carried out as
it was in the earlier study by use of a repeated measures analysis of
covariance. Mental calculation times were analyzed for both Americans
and Chinese in a single analysis. Total number of digits in the problem
was covaried out separately (with slopes allowed to vary) for each sub-
 MENTAL ABACUS 173

ject. SUBJECT was a random factor nested within GROUP (Chinese vs

American) and crossed with one repeated measures factor, number of
digits per addend (WIDTH). The GROUP x WIDTH interaction was
significant, F(3,30) = 4.86, p < .Ol.
DISCUSSION
This research investigated the ways in which calculation using the men-
tally imagined abacus resembles calculation using the actual abacus. Sev-
eral pieces of converging evidence suggest that the “mental abacus”
incorporates important features of the actual abacus.
On the intermediate states task both experts and intermediates at
abacus calculation were able to manipulate a representation of the abacus
so that it passed through the same states an actual abacus would pass
through in solving an addition problem. This was evidenced in several
ways. First, subjects were able to answer correctly-and without using
an actual abacus-questions about intermediate states unique to the
abacus-based solution of a problem. Second, response times for an-
swering such questions proved to be a function of the position of the
probed state in the sequence required for solving the problem. Finally,
the pattern of response times did not differ according to whether the task
was being solved using the abacus or mentally.
Additional evidence for the mental abacus as an analog representation
of the physical abacus comes from the comparative study of Chinese
abacus operators and Americans performing mental addition. An analysis
of the errors produced by the two groups showed that Chinese errors
differed from American errors in ways predicted according to character-
istics of the abacus. The remarkable similarity in percent of errors
showing confusions of five for abacus and mental abacus calculation
among the Chinese, and the markedly different result for the Americans,
is strong evidence for the abacus-like nature of the “mental abacus.”
Perhaps even more striking are the distributions of errors according to
position that are displayed in Table 3. Again, the distribution of errors
differs between Chinese and Americans, but not between abacus and
mental abacus calculation.
A comparative analysis of response times also showed a qualitative
difference between the two groups. When total number of digits in the
problem was held constant, Chinese mental calculation times increased
with the number of digits per addend. This was not true for abacus cal-
culation times or for American mental calculation times. One explanation
for this finding can be derived from the work of Kosslyn (1980). Thinking
in terms of a “mental abacus,” it seems that the width of the image of
the abacus (in number of columns) that must be maintained is a factor in
the time it takes to solve a problem. Kosslyn reports that visual images
174 JAMES W. STIGLER

tend to fade rapidly, and so must be constantly refreshed. Further, it

appears (on the basis of several experiments by Kosslyn) that images are
refreshed a part at a time, and that it takes longer to refresh an image
the more parts it contains. In the case of the mental abacus, it seems
reasonable that the more columns of the abacus that are required to
provide the “workspace” for a given problem, the more time must be
taken in constantly refreshing each column on the mental image.
An alternative, simpler explanation (suggested by Hatano, personal
communication) is that the mental abacus experts, when faced with a
problem which exceeds the width of two or three columns, may tend to
break the problem in half vertically and solve the two parts separately.
Presenting the problems visually made it possible for subjects to use such
a method. The time required to remember the first half result while cal-
culating the second, and then combining the two halves could account
for the extra time required for wider problems. This theory was checked
out informally on a subsequent trip to Taiwan, when abacus experts did
report, in an interview, that such a strategy is sometimes used. However,
it is important to note that both explanations for the “width” effect are
consistent with the general claim that there is a capacity limit for large
and detailed images. What is at issue is which strategy mental abacus
operators use for dealing with this limitation.
Along with the striking similarities between abacus and mental abacus
calculation were some interesting differences between the two modes.
Given that the mentally imagined abacus appears to pass through the
same intermediate states as the actual abacus would on the same problem,
how can the experts’ mental manipulation be accomplished much faster
than on the abacus? As in the well-known study of mental rotation by
Shepard and Metzler (1971), the relative response times among different
mental abacus transformations are in accord with the relative times
among the corresponding transformations that occur when an actual
abacus is used. Clearly, the absolute time required to complete a mental
manipulation is not constrained by the time it takes to manipulate the
actual abacus. The pervasive linear relationship found between response
time and total number of digits for both abacus and mental calculation
means that each additional digit adds a constant amount to the total time
for both abacus and mental calculation; what differs between the two
modes is the value of this constant. It has been suggested elsewhere
(Stigler, Barclay, & Aiello, 1982) that time for abacus calculation may be
composed of decision time plus movement time. If this is the case, then
it is reasonable to speculate that the time advantage for mental calculation
is primarily due to reduced movement time, since there is no reason to
believe that decision time would differ across the two modes of calcula-
tion. Further research would be needed to clarify such an interpretation.
 MENTAL ABACUS 175
I would like to conclude by raising an interesting point made by Kos-
slyn (1980). Kosslyn argues that much of what we know can be repre-
sented in more than one format (e.g., images or propositions), and that
for different tasks, different representational formats may lead to differ-
ences in the nature, speed, and efftciency of the processing they support.
Kosslyn demonstrates this point by comparing two external formats for
representing geographical information: a map and a chart of intercity
distances.
These two systems may be completely isomorphic to each other in all important
respects since either one can be generated from the other. Nevertheless, they
have obviously different properties . . . If we want to know quickly whether there
are any three cities in some region that fall in a straight line, we consult the map;
if we want to know the total distance of an air flight from New York to Los
Angeles to Miami, we consult the chart. (Kosslyn, 1980, p. 24)

In a recent article, Thorndyke and Hayes-Roth (1982) take Kosslyn’s idea

a step further, by demonstrating that the representational system by
which a person comes to know an environment (e.g., map reading vs
navigation experience) has a clear-cut impact on that person’s judgments
about characteristics of the environment. Different types of representa-
tion have differing “power” with respect to different criterion tasks.
What is the inherent “power” of the mental abacus as a way of rep-
resenting numbers for mental computation? It is true that amazing speed
is achieved by children who use the mental abacus. However, it is also
true that they spend a great deal of time practicing mental calculation,
regardless of the form of representation. One cannot help but wonder
what kinds of results could be achieved by Americans who use a different
form of representation, if they were to engage in the same amount of
practice.
This study of the mental representation used by Chinese children in
mental calculation highlights one of the ways in which culture may influ-
ence cognitive processes. Not only can culturally specific training alter
the strategies a child brings to bear on a cognitive problem, but it also
alters the content of the child’s thought. Indeed, perhaps the most pow-
erful tools a culture can provide to the developing child will come in the
form of specialized mental representations that are passed down through
education.

REFERENCES
Bryan, W. L., & Hatter, N. (1899). Studies in the telegraphic language: The acquisition of
a hierarchy of habits. Psychological Review, 6(4), 345-375.
Chi, M. T. H., Glaser, R., & Rees, E. (1982). Expertise in problem solving. In R. Stemberg
(Ed.), Advances in the psychology of human intelligence (Vol. 1). Hillsdale, NJ: Erl-
baum.
176 JAMES W. STIGLER

Cole, M., & Scribner, S. (1977). Cross-cultural studies of memory and cognition. In R. V,
Kail & J. W. Hagen (Eds.), Perspectives on the development of memory and cognition.
Hillsdale, NJ: Erlbaum.
Enerson, D., & Stigler, J. W. (1983). Acquiring the “Mental Abacus”: An introspective
account. Unpublished manuscript, University of Chicago.
Ezaki, S. (1980). On interiorized activity in abacus-derived mental arithmetic. Nihon-
Shuzun, 314, 2-5 [in Japanese].
Hatano, G., Miyake, Y., 62 Binks, M. G. (1977). Performance of expert abacus operators.
Cognition, 5, 47-55.
Hatano, G. (1981). Abacus learning, digit span and multi-digit addition/subtraction skills
among 3rd-graders. Paper presented at the annual meetings of the American Educa-
tional Research Association, Los Angeles.
Hatano, G., & Osawa, K. (1981). Digit memory of grand experts in abacus-derived mental
calculation. Paper presented at the annual meetings of the American Psychological
Association, Los Angeles.
Hitch, G. J. (1978). The role of short-term working memory in mental arithmetic. Cognitive
Psychology, 10, 302-323.
Kojima, T. (1954a). The Japanese abacus. Rutland, VT: Tuttle.
Kojima, T. (1954b). Advanced abacus. Rutland, VT: ‘Ibttle.
Kosslyn, S. M. (1980). Image and mind. Cambridge, MA: Harvard Univ. Press.
Moon, P. (1971). The abacus. New York: Gordon & Breach.
Shepard, R. N., & Cooper, L. A. (1982). Mental images and their transformations. Cam-
bridge, MA: MIT Press.
Shepard, R. N., & Metzler, J. (1971). Mental rotation of three-dimensional objects. Science,
171, 701-703.
Stigler, J. W. (1982). Abacus skill in Chinese children: Imagery in mental calculation. Doc-
toral dissertation, University of Michigan.
Stigler, J. W., Barclay, C. R., & Aiello, P. (1982). Motor and mental abacus skill: A prelim-
inary look at an expert. The Quarterly Newsletter of the Laboratory of Comparative
Human Cognition, 4(l), 12-14.
Tani, Y. (1964). The magic calculator. Rutland, VT: Japan Publ. Trading Co.
Thorndyke, P. W., L Hayes-Roth, B. (1982). Differences in spatial knowledge acquired
from maps and navigation. Cognitive Psychology, 14, 560-589.
Trabasso, T., & Riley, C. A. (1975). The construction and use of representations involving
linear order. In R. L. Solso (Ed.), Information processing and cognition: The Loyola
Symposia (pp. 381-410). Hillsdale, NJ: Erlbaum.
Vygotsky, L. S. (1978). Mind in society. Cambridge, MA: Harvard Univ. Press.
Zhusuan xuehui (1980). Chinese abacus association examination results record.
(Accepted October 6, 1983)