A Basis for Scaling Qualitative Data

Author(s): Louis Guttman

Source: American Sociological Review, Vol. 9, No. 2 (Apr., 1944), pp. 139-150
Published by: American Sociological Association
Stable URL: http://www.jstor.org/stable/2086306 .
Accessed: 11/02/2014 14:59

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].

American Sociological Association is collaborating with JSTOR to digitize, preserve and extend access to
American Sociological Review.

http://www.jstor.org

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
 A BASIS FOR SCALING QUALITATIVE DATA*
Louis GUTTMAN
Cornell University

I. INTRODUCTION investigators (such as critical ratios, biserial

IN A GREAT deal of researchin the social correlations,factor analysis, etc.), and which
and psychological sciences, interest lies give a complete picture of the data not af-
in certain large classes of qualitative forded by these other techniques. The word
observations. For example, research in mar- "picture"might be interpretedhere literally,
riage is concernedwith a class of qualitative for the results of the analysis are presented
behavior called marital adjustment, which and easily assimilated in the form of a
includes an indefinitelylarge numberof inter- "scalogram," which at a glance gives the
actions between husband and wife. Public configuration of the qualitative data.
opinion research is concerned with large Description of the practical procedures,
classes of behaviorlike expressionsof opinion as well as the mathematical analysis, must
by Americans about the fighting ability of be postponed to other papers. The present
the British. Educational psychology deals paper is devoted to a non-technical discus-
with large classes of behavior like achieve- sion of what we mean by a scale.
ment tests.
2. THE NOTIONS OF VARIABLE, FUNCTION,
It is often desired in such areas to be able
AND SIMPLE FUNCTION
to summarize data by saying, for example,
that one marital couple is better adjusted First, a word about what is meant by a
than another marital couple, or that one variable, whether qualitative or quantitative.
person has a better opinion of the British We use the term in its conventional logical
than has another person, or that one student or mathematical sense, as denoting a set of
has a greater knowledge of arithmetic than values. These values may be numerical
has another student. There has been con- (quantitative) or non-numerical (qualita-
siderable discussion concerning the utility of tive).' We shall use the term "attribute"
such orderings of persons. It is not our
conventionalcourses in undergraduatecol-
intention in this paper to review such dis- lege1In mathematicsit is not ordinarilypointed out
cussions, but instead to present a rather new that a great deal of mathematicsdeals with purely
approach to the problem which seems to qualitativevariables.Notions of metrics and quan-
afford an adequate basis for quantifying titative variablescan be arrivedat by sequencesof
qualitative data. qualitativeclassifications.In fact, this is the manner
in which our approachto scaling derives a scale
This approach has been used successfully ordering.
for the past year or so in investigating morale The reader who is interestedmight look- at a
and other problems in the United States recent departurein textbooks for an introductory
Army by the Research Branch of the Morale course in college mathematics (M. Richardson,
ServicesDivision of the Army ServiceForces. Fundamentals of Mathematics,Macmillan, I941).
This book gives a simple, entertaining,and mature
While this approach to quantification leads introductionto the foundationsof mathematics.Its
to some interesting mathematics, no knowl- emphasisis on understanding,rather than on ma-
edge of this mathematics is required in ac- nipulation.It covers fundamentaltopics like point
tually analyzing data. Simple routines have sets, the concept of number, and others that are
rarely mentioned in ordinary undergraduatecur-
been established which requireno knowledge riculaand yet are mainstaysof mathematicaltheory.
of statistics, which take less time than the It is only that most of us have been exposed ex-
various manipulations now used by various clusively to certainalgebraicmanipulationsthat we
conceive such manipulationsto be the essence of
* Presentedto the Thirty-eighthAnnualMeeting mathematics.A more sophisticatedview is to regard
of the American Sociological Society, New York mathematics as unveiling necessary relationships
December4, I943. that arise from classifications.Much uselessdiscus-
'39

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
I40 AMERICAN SOCIOLOGICALREVIEW

interchangeably with "qualitative variable." will have a one-to-one correspondencewith

The values of an attribute (or of a quantita- the values of y, then we shall say the at-
tive variable, too, for that matter) may be tribute y is a simple function of x. For
called its subcategories,or simply categories. example, suppose x takes on the ten values
An example of an attribute is religion. A O, I, 2, 3, 4, 5, 6, 7, 8, 9. Then the corre-
person may have the value "Catholic," spondence table might be as follows:
"Buddhist,""Jewish,""Mormon,""atheist,"
or some other value of this variable. There X 0 I 2 3 4 5 6 7 8 9
is no particular intrinsic ordering among
Y Y' Y1 Y' Y3 Y3 Y2 Y2 Y2 Y2 32
these values. Another example is expression
of an opinion. A person may say, "I like Or we might show this graphicallyby plotting
the British," "I don't like the British," or the x values on a straight line, and cutting
"I don't know whether or not I like the it into intervals:
British." Another example is, a person may
YI Ys Y2
be observed to smile at another person upon
meeting him, or he may be observed not to .. I i I I .I ~~~~~~~~~~I
-I I I I
smile. 0. 2 3 ' 5 6 7 8 9
Quantitative variables are readily recog-
FIGURE I
nized and need no discussion here.
A variable y is said to be a single-valued For statistical variables, another representa-
function of a variable x if to each value of tion is in terms of a bar chart of frequencies,
x there corresponds a single value of y. and this is what we use for convenience in
Thus, if y has the distinct values Y1,Y2,. . *, ? Io and ? I I below.
yi, and if x has the distinct values, xl, x,
. . ., X., where m and n may be different, y 3. THE DEFINITION OF SCALE
is called a single-valued function of x if a For a given population of objects, the
table of correspondencecan be set up like, for multivariate frequency distribution of a uni-
example, the following: verse of attributes will be called a scale if
it is possible to derive from the distribution
X X1 X2 X3 ... X
a quantitative variable with which to charac-
terize the objects such that each attribute
Y Y3 Y5 Ym-2 Y2
is a simple function of that quantitative
For each value of x there is one and only variable. Such a quantitativevariableis called
one value of y. (The converse need not hold: a scale variable.
for the same value of y there may be two Perfect scales are not to be expected in
or more values of x.) Obviously, if y is to be practice. The deviation from perfection is
a single-valued function of x, then we must measured by a coefficient of reproducibility,
have m < n. which is simply the empirical relative fre-
In particular, suppose y is an attribute, quency with which the values of the at-
say like the above attribute about expression tributes do correspondto the properintervals
of liking for the British. Then m= 3, and of a quantitative variable. In practice, 85
we may denote by y1 the statement, "I like percent perfect scales or better have been
the British"; by y2, the statement, "I don't used as efficient approximations to perfect
like the British"; and by y., "I don't know scales.
whether or not I like the British." If x is a A value of a scale variable will be called
quantitative variable which takes on more a scale score, or simply a score. The ordering
than m values (n > m), and if we can of objects according to the numerical order
divide the x values into m intervals which of their scale scores will be called their scale
order.
sion of mathematicsas a "tool" in social research Obviously, any quantitative variable that
could be saved by recognitionof the fact that quali-
tative lead
classifications to just as rigorousimplica- is an increasing (or decreasing) function of
tions as do quantitative. a scale variable is also a scale variable. For

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
 A BASIS FOR SCALING QUALITATIVE DATA 14I

example, in the illustration in ?2, consider define the concept. Another way of describ-
x to be a scale variable. Any constant could ing the universe is to say it consists of all the
be subtracted from or added to each of the attributes of interest to the investigation
x scores, and y would remain a simple func- which have a common content, so that they
tion of the transformed x. Thus, the scores are classified under a single heading which
0, I, 2, 3, 4, 5, 6, 7, 8, 9 could be replaced indicates that content.
by the respective scores -5, -4, -3, -2, For ease in focusing, let us take an example
-I, O. I, 2, 3, 4. Or the x scores could be from opinion research where it is desired
multiplied by any constant, or their square to observe the population of individuals in
roots or logarithms could be taken-any a standardizedmannerby a checklist of ques-
transformation,continuous or discontinuous, tions. The behavior of interest to the investi-
could be used, as long as the rank order gation is responses of individuals to such
correlation between the original x and the questions. Suppose the universe of attributes
transformed variable remained perfect. All consists of all possible questions which could
such transformations will yield scale vari- be asked in such a list concerningthe fighting
ables, each of which is equally good at ability of the British. Such questions might
reproducingthe attributes. be: "Do you think the British Army is as
Therefore, the problem of metric is of no tough as the German Army?"; "Do you
particular importance here for scaling. For think the R.A.F. is superior to the Luft-
certain problems like predicting outside vari- waffe?"; etc. (We do not pause here for
ables from the universe of attributes, it may problems of wording, interpretation,and the
be convenient to adopt a particular metric like. The reader is urged rather to focus on
like a least squares metric, which has con- the general outline we are trying to es-
venient properties for helping analyze mul- tablish.) There may be an indefinitely large
tiple correlations. The interesting mathe- number of such questions which belong in
matics involved here will be discussed in the universe; and in a particular investiga-
another paper. However, it must be stressed tion, ordinarilyonly a sample of the universe
that such a choice of metric is a matter of is used.
convenience; any metric will predict an out- An attribute belongs to the universe by
side variable as accurately as will any other. virtue of its content. The investigator indi-
In practice, the rank order has been used cates the content of interest by the title he
as a scale variable. (It is in fact a least chooses for the universe, and all attributes
squares metric for a rectangulardistribution with that content belong in the universe.
of scale scores.) There will, of course, arise borderline cases
in practice where it will be hard to decide
4. THE UNIVERSE OF ATTRIBUTES2
whether or not an item belongs in the uni-
A basic concept of the theory of scales is verse. The evaluation of the content thus
that of the universe of attributes. In social far remains a matter that may be decided
research, a universe is usually a large class by consensus of judges or by some other
of behavior such as described in the intro- means. This has been recognized before, al-
duction above. The universe is the concept though it need not be regarded as a "sin
whose scalability is being investigated, like against the Holy Ghost of pure operational-
marital adjustment, opinion of British fight- ism."3 It may well be that the formal
ing ability, knowledge of arithmetic,etc. The analysis for scalability may help clarify un-
universe consists of all the attributes that certain areas of content. However, we have
2The words population and universe are ordi- found it most useful at present to utilize
narily used interchangeably in statistical literature. informal experience and consensus to the
For scales, it is necessary to refer both to a complete fullest extent in defining the universe.
set of objects and to a complete set of attributes, so
it will be convenient to reserve population for the 3 Clifford Kirkpatrick, "A Methodological Anal-
former, and universe for the latter. In social re- ysis of Feminism in Relation to Marital Adjust-
search, the objects are usually people, so that popu- ment," American Sociological Review, June I939,
lation is appropriate for them. 4: 325-334.

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
 142 AMERICAN SOCIOLOGICALREVIEW

An important emphasis of our present ap- yields the data of interest to the investiga-
proach is that a criterion for an attribute tion may be used. Such techniques for the
to belong in the universe is not the magnitude social and psychological sciences might be
of the correlations of that item with other case histories, interviews, introspection, and
attributes known to belong in the universe. any other technique from which observations
It will be seen (in ?io below) that attributes may be recorded.The important thing is not
of the same type of content may have any how the observationswere obtained, but that
size of intercorrelations,varying from prac- the observationsbe of central interest to the
tically zero to unity.4 investigation.
Use of a questionnaire implies that the
5. THE POPULATION OF OBJECTS investigator is interested in a certain type of
Defining the universe of attributes is a universe of verbal behavior. Participant ob-
problem similar to the standard problem of servation may imply that the investigator is
defining the population of objects or indi- interested in a certain type of universe of
viduals5 of interest to the investigation. An non-verbal behavior. Such distinct universes
investigator must always delimit the popula- may each be investigated separately. It may
tion with which he is working. For example, often be of interest to see how well one uni-
in the case of opinion about the British as verse correlates with another, but such a
fighters, he must decide whose opinions he correlationcannot be investigated until each
wishes to ascertain. Is he interested in every- universe is defined and observed in its own
one in the world, or just in everyone in the right.
United States? Is he interested in everyone The examples of scales to be given later
in the United States, or just in adults? If in this paper happen to comprise observa-
just in adults, how is an adult to be defined? tions made by means of questionnaires. It
Here, too, decisions will sometimes be diffi- should not be inferred, however, that scaling
cult as to whether a particular individual refers only to that technique.Scaling analysis
belongs in a population or not, and decisions is a formal analysis, and hence applies to any
must be made somehow before the investiga- universe of qualitative data of any science,
tion begins, else the investigator will not obtained by any manner of observation.
know whom to observe.
7. THE PURPOSE OF SCALING
6.
METHODS OF OBSERVATION Obviously it is very clumsy to record the
Let us assume that somehow we have a large number of observations ordinarily in-
universe of attributes and a population of volved in a universe of attributes for a pop-
individuals defined. Next, observations are ulation of individuals.The recordingrequires
made as to the behavior of the population a table with one row for each individual and
with respect to the universe. (In practice one column for each attribute. (The table
this will often be done only with samples. may theoretically be indefinitely large.) It
A sample of individuals from the population would be convenient if we could represent
will have their behaviorobservedon a sample the observations in a more compact manner
of attributes from the universe.) How the which would enable us to reproduce such a
observationsare to be made is of no concern table whenever desired. A compact repre-
to us here. In opinion research and other sentation, if it could be obtained, would have
fields, questionnairesand scheduleshave been two great advantages: first, a mnemonic ad-
used. But any techniqueof observationwhich vantage, for a compact representationwould
4That correlationsare no criterionfor content be easier to rememberthan would be a large
has been quite well known. See, for example,R. F. table; and second, if it were desired to relate
Sletto, Construction of Scales by the Criterion of the universe to other variables it would be
Internal Consistency, Sociological Press, Hanover, easier to do so by means of the compact
N.H., I937.
' For convenience,since the examplesin this pa- representation than by using the large mul-
per concernpopulationsof human beings, we shall tivariate distribution of the attributes in the
talk entirelyin terms of such populations. universe. From these are derived other ad-

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
 A BASIS FOR SCALING QUALITATIVE DATA

vantages which will become apparent as the right but the second question wrong, would
reader's familiarity with scales grows. not occur. In such a case, it is possible to
A particularlysimple representationof the assign to the population a set of numerical
data would be to assign to each individual a values like 3, 2, i, o. Each member of the
numerical value and to each category of population will have one of these values as-
each attribute a numerical value such that, signed to him. This numerical value will be
given the value of the individual and the called the person's score. From a person's
values of the categories of an attribute, we score we would then know precisely to which
could reproduce the observations of the in- problems he knows the answersand to which
dividual on the attribute. This will be pos- he does not know the answer. Thus a score
sible only for restricted types of data, where of 2 does not mean simply that the person
each attribute in the universe can be ex- got two questions right, but that he got two
pressed as a simple function of the same particular questions right, namely, the first
quantitative variable, that is, where the uni- and second. A person's behavior on the prob-
verse of attributes forms a scale for the pop- lems is reproducible from his score. More
ulation of individuals. specifically, each question is a simple func-
tion of the score, as is shown in ?io below.
8. AN EXAMPLE OF A DICHOTOMOUS SCALE
9. THE MEANING OF "MORE"2 AND "LESS"2
As may be expected, the universe of at-
tributes must form a rather specialized con- Notice that there is a very definite mean-
figuration for the population of individuals ing to saying that one person knows more
if it is to be scalable. Before describing a mathematics than another with respect to
more general case, let us give a little exam- this sample. For example, a score of 3 means
ple. (A sociological interpretation of this more than a score of 2 because the person
apparently mathematical example is given with a score of 3 knows everything a person
in ?I5 below.) Consider a mathematics test with a score of 2 does, and more.
composed of the following problems: There is also a definite meaning to saying
(a) If r is the radius of a circle, then what that getting a question right indicates more
is its area? knowledge than getting the same question
(b) What are the values of x satisfying wrong, the importance of which may not be
the equation too obvious. People who get a question right
ax2 + bx + c = o? all have higher scale scores than do people
(c) What is de-/dx? who get the question wrong. As a matter of
If this test were given to the population fact, we need no knowledge of which is a
of membersof the American SociologicalSo- right answer and which is a wrong answer
ciety, we would perhaps find it to form a beforehand to establish a proper order
scale for that population. The responses to among the individuals. For convenience,sup-
each of these questions might be reported pose the questions were given in a "true-
as a dichotomy, right or wrong. There are false" form,6 with suggested answers 2IIr,
2 X 2 X 2 = 8 possible types for three ( b \/ b2-4ac) / 2a, and xe-1 for the
dichotomies. Actually, for this population of respective questions. Each person records
sociologists we would probably find only either a T or an F after each question, ac-
four of the possible types occurring. There cording as he believes the suggested answers
would be the type which would get all three to be true or false. If the responses of the
questions right, the type which would get the population form a scale, then we do not have
first and second questions right, the type to know which are the correct answers in
which would get only the first question right, order to rank the respondents (only we will
and the type which would get none of the
questions right. Let us assume that this is We shall assume that no one gets an answer
what would actually happen. That is, we right by guessing.In a later paperit will be shown
how scale analysis can actually pick out responses
shall assume the other four types, such as the that were correctmerely by guessing.But for this,
type getting the first and the third questions much more than three itemsare necessary.

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
I44 AMERICAN SOCIOLOGICALREVIEW

not know whether we are ranking them from ordering is a purely formal consequence of
high to low or from low to high). By the the configurationof the behavior of the pop-
scale analysis, which essentially is based on ulation with respect to the items. The impor-
sorting out the joint occurrencesof the three tance of this fact will become more apparent
items simultaneously, we would find only 4 in more complicated cases where the attri-
types of persons occurring. One type would butes are not dichotomous but have more
be F1T2F3, where the subscripts indicate than two categories. We do not take the
the questions; that is, this type says F to space here to expand on this point, but
question I, T to question 2, and F to ques- merely state that the scale analysis auto-
tion 3. The other three types would be matically decides, for example, where an
F1T2T3, F1F2T3, and T1F2T3. These types "undecided" response to a public opinion
could be shown in a chart (a "scalogram") poll questionnaire belongs, whether it is
where there is one row for each type of per- above "yes," below "no," in between, equiv-
son and one column for each category of alent to "yes," or equivalent to "no."
each attribute. Without going into details,
the scale analysis would establish an order I 0. THE BAR CHART REPRESENTATION
among the rows and among the columns Another way of picturing the dichotomous
which would finally look like this: scale of the sample of three items would be
F3 T2 F1 T3 F2 T, as follows: suppose that 8o percent of the
. . . . ~~~~~~~///
X
Wrong Right
(c)
(90%) (10[%

FIGURE 2

Wrong
// Rih 7,
Or, alternatively, both rows and columns
might be completely reversed in order. Each (60%)('% 7/
response to a question is indicated by a
check mark. Each row has three checkmarks
because each question is answered, either
correctly or incorrectly.The "parallelogram" Wrong Right
pattern in the chart7 is necessary and suffi.
cient for a set of dichotomous attributes to
(a) [(20%) (//// /A
be expressibleas simple functions of a single FIGuRE 3
quantitative variable.
From this chart we can deduce that F1, population got the first question right, 40
T2, and F3 are all correct answers, or are all percent got the second question right, and
incorrect answers. That is, if we were now IO percent got the third question right. The
told that F1 is a correct answer, we would univariate distributions of the three respec-
immediately know that T2 and F3 are also tive items could be shown by the bar chart
correct answers. This means that we can in Figure 3.
order the men according to their knowledge The bars show the percentagedistributions
even if we do not know which are the correct for the respective questions. The multivari-
answers and which are the incorrectanswers, ate distributionfor the three questions, given
only we do not know whether we are order- that they form a scale for the population,
ing them from highest to lowest or from can also be indicated on the same chart,
lowest to highest. Except for direction, the since all those who are included in the group
getting a harder question right are also in-
'Such a chart,whereone columnis used for each cluded in the
category of each attribute,we call a scalogram. The
group getting an easier question
scalogramboards used in practicalproceduresare right., Thus, we could draw the bar chart
simply devices for shifting rows and columns to over again, but connect the bars with dashed
find a scale patternif it exists. lines in the fashion shown in Figure 4.

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
 A BASIS FOR SCALING QUALITATIVE DATA I45

Here we can see how the three questions axis. However, the point correlationsbetween
are simple functions of the scores. From the the items are not at all perfect. For exam-
marginal frequencies of the separate items, ple, the four-fold table between the second
together with the fact that the items form a and third items is as follows:
scale, we are enabled to deduce that io per-
cent of the people got a score of 3. The io Question (b)
percent who got the hardest question right
Right Wrong
are included in those who got the easier
questions right. This is indicated by the
dashed line on the right, between the scores Right Io 0 Io
2 and 3, which carries the same io percent
Question (c)
of the people (those with a score of 3)
through the three bars. The 40 percent who Wrong 30 6o 90
got the second question right include the io
percent who got the hardest question right
and 30 percent out of those who got the 40 6o I00
hardest question wrong, but all 40 percent
got the easiest question right. This leaves us The point correlationbetween the two items
30 percent who got just the first and second is .4I. As a matter of fact, the point correla-
questions right. And so on. Thus we can tion between two dichotomousitems may be
anything from practically zero to unity, and
(c) WoWrong R
yet they may both be perfect functions of
the same quantitative variable. That this
may be paradoxical might be explained by
inadequate treatment of qualitative variables
in conventional courses and textbooks on
statistics.8
An important feature of this four-fold
Wron
)Rroght table is the zero frequencyin the upper right-
(60%)
hand corner cell. Nobody who got the third
question right got the second question wrong.
(a) (20%)
| /////w/////+
Such a zero cell must always occur in a
four-fold table between two dichotomous
Frequency 20% 40% 30% 10%
Score 0 1 2 3
items which are simple functions of the same
quantitative variable.
FIGURE 4
8 Technical Footnote. A tetrachoric coefficient for
think of an ordering of the persons along a the four-fold table above, assuming a bivariate nor-
horizontal axis, and each item can be thought mal distribution, would be unity. However, this is
of as a cut on that axis. All those above the not the correlation between the items. It does not
tell how well one can predict one item from the
cutting point get the question right and all other. The tetrachoric coefficient expresses instead
those below the cutting point get the ques- the correlation between two quantitative variables of
tion wrong. Thus there is a one-to-one cor- which the items are functions, provided the assump-
respondence between the categories of an tions of normality are true. The reason the tetra-
item and segments of the axis. Or we can say choric is unity in this case is that the quantitative
variables of which the items are functions are one
that each attribute is a simple function of and the same variable, namely, the scale variable.
the rank order along the axis. Notice, however, that the distribution of the scale
It is because all the items in the sample variable according to the rank order is not at all
can be expressed as simple functions of the normal. One of the contributions of scaling theory is
same ordering of persons that they form a to do away with untested and unnecessary hy-
potheses about normal distributions. It is the point
scale. Each item is perfectly correlated with correlation that is involved in the mathematical
or reproduciblefrom the ordering along the analysis of scaling, not the tetrachoric.

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
I46 AMERICAN SOCIOLOGICALREVIEW

I I. ANOTHER EXAMPLE OF A SCALE 3. If you could get no job at all, what would
you do?
Now let us give an example of a more (a) I wouldnot go back to school
complicated scale. Suppose we were inter- (b) If the governmentwould aid me, I
ested in finding out how much desire soldiers wouldgo back to school
may express now about going back to school (c) I would go back to school even with-
after the war is over. Suppose that out of out governmentaid

J~o~b W
o1 u I .d t a k e a g ol
I
o d jo b : (7:~~~~~~~~~~~~~
0 % )w|uldw
? . '
turn tur
godown

, l , ~~~~~~~'job
(20%) t 10%)

I I got aided
I job
(10%)20% (20%)

Some i govtt I . aided Would

turn down
oKib Would t tke some kind f job (6'0% ) wouldturndowno tme kind of obn

Som Ifgoe mn aife govtu

ieldWudtundw

iNd I fWould not go back to school o If job (6:0% would r dwn Id go bick to
Job go back to school school even without
At Al l (35%) (40Z)
sgjaid | | (25%)

No, would not like to go to school (50(0) Yes, would %)iketo go to' school t50,)
TofGo
To School

Frequency 35,% 15% 10% 10% 5% 5% 10,% 10%

Score 0 1 2 3 5 _ 6 7

FIGURE5

the universe of attributes which define this 4. If you could do what you like after the war
desire we select the following sample of four is over, wouldyou go back to school?
questions to be presented on a questionnaire. (a) Yes
(b) No
i. If you were offereda good job, what would
you do? Let us suppose the responses of the men
(a) I wouldtake the job to these questions form a scale in the man-
(b) I would turn it down if the govern- ner shown in Figure 5.
ment wouldhelp me go to school We now know how to read such a chart.
(c) I would turn it down and go back to io percent of the men said they would turn
school regardless down a good job to go back to school; 20
2. If you wereofferedsomekindof job, but not
percent said they would turn down a good
a good one, whatwouldyou do?
(a) I wouldtake the job job only if the government aided them; 70
(b) I would turn it down if the govern- percent said they would take a good job; and
ment wouldhelp me go to school so on. The io percent who said they would
(c) I would turn it down and go back to turn down a good job are included in the
school regardless 20 percent who said they would turn down

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
 A BASIS FOR SCALING QUALITATIVE DATA I47

some kind of a job, and the 20 percent are I 2. ON SAMPLING THE UNIVERSE
included in the 25 percent who said they OF ATTRIBUTES
would go back to school if they got no job An important property of a scalable uni-
at all, and these 25 percent are included in verse is that the ordering of persons based
the 50 percent who said they would like to on a sample of items will be essentially that
go back to school. based on the universe. If the universe is a
For three trichotomous and one dichot- scale, what the addition of further items
omous questions there are 3 X 3 X 3 X would do would be merely to break up each
2 = 54 possible types. In order for these to type given by the sample into more differ-
form a scale, it can be shown that at most entiated types. But it would not interchange
eight types can occur. The chart shows the the order of the types already in the sample.
eight types, which have been scored from For example, in Figure 5 above, type 6
o through 7. The chart shows the character- would always have a higher rank order than
istics of each type. For example, the type type 5. People in type 6 might be ordered
with the score 3 includes all men with the within the type into more subcategories;
following four values: they say that they people within type 5 might be ordered into
would take a good job if it were offered to more subcategories; but all subcategories
them rather than go back to school; that within 6 would remain of higher rank than
they would turn down some kind of job if all those in type 5. This may be seen in
the government would aid them to go back reverse, for example, by deleting one of the
to school; that they would go back to school questions and noticing that all that is ac-
if the government would aid them if they complishedis to collapse the number of types
could get no job at all; and that they would to a smaller number so that two neighboring
like to go back to school. Thus, by reading types may now become indistinguishable;but
the categories crossed by the dashed lines any types two steps apart would still remain
which enclose each type, we can read off the in the same order with respect to each other.
characteristics of the type. Hence, we are assured that if a person
Notice that each of the four attributes is ranks higher than another person in a sam-
a simple function of the scale scores. For ple of items, he will rank higher in the
example, the "good job" question has its universe of items. This is an important prop-
categories correspond with the following erty of scales, that from a sample of attri-
three intervals of the scale scores: 0-3, 4-6, 7. butes we can draw inferences about the
The question might be raised as to how universe of attributes.
often will scales be found in practice. Isn't One of the criteria for selecting a sample
even a fair approximationto a structure like of items is to choose a sample with enough
that in the above chart too much to hope categories to provide a desired amount of
for to be found in real life? Towards an differentiation between individuals. Thus if
answer to this, we can only cite thus far our individuals are desired to be differentiated
experience with research in the Army. Lit- say only into IO groups, items should be
erally dozens of sufficiently perfect scales chosen which will yield IO types.9 The shape
have been found in various areas of attitude, of the distribution of the rank orders in a
opinion, and knowledge. The example given sample of attributes will of course depend
above of desire to go to school is a fictitious upon the sample. One sample of attributes
version of a set of similar questions that have may give one shape distribution; another
actually proved scalable for the Army. Many sample may give another shape distribution.
varieties of data have been found scalable, This need not be a matter of concern, since
and many have not. Those data which our primary interest lies in the ordering of
proved scalable could then be related to
'We are of course not consideringproblemsof
other variables very easily. Those that were reliabilityin the sense of repeated observationsof
not scalable required a more complicated the same attributes.For convenience,we are tacitly
analysis to handle them properly. assumingperfectreliability.

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
148 AMERICAN SOCIOLOGICALREVIEW

people, not the relative frequency of each student five years after he graduates from
position. college, from his present knowledgeof mathe-
It might be asked how can one know the matics. To do this, an experimental sample
universe forms a scale if all one knows is a would have to be obtained where salaries five
sample from the universe. At present it years after college are known for each per-
seems quite clear that in general the prob- son and where responses to each item on the
ability of finding a sample of attributes to mathematics test are known. If the criterion
form a scale by chance for a sample of in- of least-squares is adopted, then the best
dividuals is quite negligible, even if there prediction on the basis of the sample would
are as few as three dichotomousitems in the be the multiple regression of income on the
sample and as many as one hundred individ- three items in the sample. The multivariate
uals.'0 It seems quite safe to infer in gen- distribution of the three items and the out-
eral that if a sample of attributes is selected side variable would give the necessary data
without knowledge of their empirical inter- for computing the regression, curvilinear or
relationships and is found to form a scale linear, which would be best for predicting
for any sizeable random sample of individ- the outside variable. If we wished to predict
uals, then the universe from which the at- some other outside variable from the same
tributes are selected is scalable for the entire items, a new multiple regressionwould have
population of individuals. to be worked out from the multivariate dis-
tribution of the three items and the new
I3. SCALING AND PREDICTION outside variable. In general, the first of these
It is important to distinguish between two regressions would ordinarily be expected to
closely related topics, scaling and prediction. differ from the second. In general, weights
Finding that a universe of attributes is to be used to predict one outside variable
scalable for a population means that it is from a set of attributes will differ from those
possible to derive a quantitative variable used to predict another outside variable; a
from the multivariate distribution such that new multiple regressionmust be worked out
each attribute is a simple function of that for each outside variable.
variable. We might phrase this otherwise by This emphasizes an important property of
saying that each attribute is (perfectly) scales. If the items have a multivariate dis-
predictablefrom the quantitative variable. tribution that is scalable, it can easily be
This is the converse of the ordinary prob- seen that no matter what the outside variable
lem of prediction. In an ordinary problem may be, the same prediction weights may be
of prediction, there is an outside variable, given to the items. The correlation of any
independently defined, that is to be pre- outside variable with the scale scores is pre-
dicted from the attributes. For example, it cisely the same as the multiple correlation
might be desired to predict the income of a of that outside variable with the items in
the scale. Thus we have an outstandingprop-
10 Technical Footnote. To work out the complete erty of scaling, namely, that it provides an
probability theory would require two things: first, invariant quantification of the attributes for
a definition of a sampling process for selecting items,
predicting any outside variable. No matter
and second, a definition of what is meant by a scale
not existing. A definition of the sampling process is what prediction purpose is to be served by
difficult because items are ordinarily developed in- the attributes, the scale scores will serve that
tuitively. Stating a null hypothesis that a scale does purpose.
not exist leads to many possible analytical formula-
tions, for different limiting conditions may be im- 14. ON "ITEM ANALYSIS"
posed upon the multivariate distribution of the
items. For example, should the marginal frequencies Let us repeat the distinction just made.
be considered fixed in all samples, should the bivari- In scaling we reproduce the attributes from
ate frequencies be considered fixed, etc.? These are a quantitative variable. In prediction, we
questions which may become clearer as the theory
of scaling develops, and in return may clarify our
predict a variable from the attributes. This
conceptions of what observation of social phenomena is a sharp difference which enables us to
implies. avoid much of the confusion that seems to

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
 A BASIS FOR SCALING QUALITATIVE DATA

prevail in the previous literature on scale then the scores are merely a compact frame-
construction. It seems to have been felt that work with which to represent them.
items in a universe are merely stepping If a compact framework is found, it has
stones from which to obtain scores. It seems the additional important property of being
to have been felt that it was an embarrassing an efficientdevice for predicting any outside
deficiency to lack a particular variable to variable in the best manner possible from
predict from the items-that as a necessary the given universe of attributes.
evil one had to resort to methods of internal
I5. THE RELATIVITY OF SCALES
consistency to derive scores.
This accounts for current "item analysis" An interesting problem associated with
approaches to scaling. These use procedures scales is: why does a universe form a scale
that are typically as follows. A trial set of for a given population? For example, take
weights is assigned the categories, yielding the sample of three mathematics questions
a trial set of scores. Then each item is ex- given above. Why should these three ques-
amined to see how well it by itself discrimi- tions be scalable? There is no necessary
nates between these scores, that is, how well logical reason why a person must know the
the scores can be predicted from the item. area of a circle before he can know what a
Those items which individually discriminate derivative is, and in particular the deriva-
best are retained, and the others eliminated. tive of e1. The reason for a scale emerging in
The misleading character of such proce- this case seems largely cultural. Our educa-
dures can be seen by inspection of the ex- tional system is such that the sequence with
amples of scales in ? io and ? i i above. We which we learn our mathematics in our high
have pointed out that the intercorrelations schools and colleges is first to get things such
between attributes in a scale can be as close as areas of circles, then algebra, and then
to zero as one pleases. It can also easily be calculus. And the amount of drill that we
seen that the correlation ratio of the scale have on each of these topics is probably also
scores with any single item can also be as in that order. It would be quite possible,
close to zero as one pleases. The predictabil- however, for the proverbial "man from
ity of the scale variable from an attribute Mars" to come to this earth and study cal-
does not tell whether or not the attribute is culus without having to learn the area of a
predictable from the scale variable. circle, so that he might not be a scale type,
The use of the "item analysis" procedures according to the scale presented above; or
in connection with scales seems to be an a student may have had some personal inci-
unfortunate carry-over from the problem of dent which somehow impressed upon him
ordinary prediction of an outside variable. with great force the derivative of e-, but
In such a prediction problem, the items are in the ordinary course of circumstances
truly but stepping stones to enable predic- would have forgotten it even more readily
tions to be made. It is known"l that item than he forgot the area of a circle.
analysis affordsa first approximationto mul- The scale analysis will pick out such de-
tiple correlation (or the discriminant func- viants or non-scale types. Of course, if these
tion), and an item is of interest only insofar non-scale types are too numerous, we shall
as it aids in the multiple regression. not say that a scale exists. In practice we
Our emphasis for scaling is quite different. find scales, although never perfect scales.
In scaling, we are interested in each and only because there has been sufficient uni-
every attribute in the universe on its own formity of experience for the population of
merits. If we were not, we would not work individualsso that the attributes mean essen-
with the universe. The attributes are the tially the same thing to the different indi-
important things; and if they are scalable, viduals. As a matter of fact a study of the
" See, for example, Louis Guttman, "An Outline deviants is an interesting by-product of the
scale analysis. Scale analysis actually picks
of the Statistical Theory of Prediction," in Paul
Horst, et al., The Predictionof PersonalAdjustment, out individuals for case studies.
Social Science Research Council, 1941. A universe may form a scale for a popula-

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions
 150 AMERICAN SOCIOLOGICALREVIEW

tion at a given time and may not at a later the orderings of objects and of categories
time. For example, the items in the scale of are in general unique (except for direction).
expression of desire of American soldiers to Both orderings emerge from analysis of the
go back to school after the war may not data, rather than from a priori considera-
prove to be scalable if they were asked once tions.
more at the close of the war. 5. The predictability of any outside variable
A universe may form a scale for one pop- from the scale scores is the same as the
ulation of individuals, but not for another. predictability from the multivariate distribu-
Or the attributes may form scales for two tion with the attributes. The zero order cor-
relation with the scale score is equivalent to
populations in different manners. For ex-
the multiple correlation with the universe.
ample, a sample of items of satisfaction with Hence, scale scores provide an invariant
Army life which formed a scale for combat quantification of the attributes for predicting
outfits in the Air Force did not form a scale any outside variable whatsoever.
for men in the technical schools of the Air
6. Scales are relative to time and to populations.
Force. The structure of camp life for these
a. For a given population of objects, a
two groups was too different for the same universe may be scalable at one time
items to have the same meaning in both sit- but not at another, or it may be scalable
uations. at two periods of time but with different
If a universe is scalable for one population orderings of objects and categories.
but not for another population, or forms a b. A universe may be scalable for one
scale in a different manner, we cannot com- population but not for another, or it
pare the two populations in degree and say may be scalable for two populations but
that one is higher or lower on the average with different orderings of objects and
than another with respect to the universe. categories.
c. Comparisons with respect to degree can
They differ in more than one dimension, or
be made only if the same scaling ob-
in kind rather than in degree. It is only if tains in both cases being compared.
two groups or two individuals fall into the
same scale that they can be ordered from 7. From the multivariate distribution of a sam-
higher to lower. A similar considerationholds ple of attributes for a sample of objects,
inferences can be drawn concerning the com-
for comparisonsin time. An important con-
plete distribution of the universe for the
tribution of the present theory of scaling is population.
to bring out this emphasis quite sharply. a. The hypothesis that the complete dis-
tribution is scalable can be adequately
i 6. SUMMARY tested with a sample distribution.
b. The rank order among objects accord-
i. The multivariatefrequencydistributionof a
ing to a sample scale is essentially that
universeof attributesfor a populationof ob-
in the complete scale.
jects is a scale if it is possibleto derivefrom c. The ordering of categories in a sample
the distributiona quantitativevariablewith
scale is essentially that in the complete
which to characterizethe objects such that
scale.
each attributeis a simple function of that
quantitativevariable. 8. Perfect scales are not found in practice.
a. The degree of approximation to perfec-
2. There is an unambiguousmeaning to the
tion is measured by a coefficient of re-
orderof scale scores.An objectwith a higher producibility, which is the empirical rel-
score than anotherobject is characterized by ative frequency with which values of
higher,or at least equivalent,values on each the attributes do correspond to intervals
attribute. of a scale variable.
3. There is an unambiguousmeaningto the or- b. In practice, 85 percent perfect scales or
der of attributevalues. One categoryof an better have been used as efficient ap-
attributeis higherthan anotherif it charac- proximations to perfect scales.
terizes objectshigheron the scale. 9. In imperfect scales, scale analysis picks out
4. It can be shownthat if the data are scalable, deviants or non-scale types for case studies.

This content downloaded from 66.77.17.54 on Tue, 11 Feb 2014 14:59:38 PM

All use subject to JSTOR Terms and Conditions