Professional Documents
Culture Documents
3 MAY, 1960
example, when an individual is asked The fourth axiom has already been
which of two statements he prefers to discussed, so we conclude this section
endorse. So it will be convenient to with a discussion of the fifth axiom and
construct a subset of the set A, consist- its accompanying definition. On a
ing of those pairs of points (ct, qfi given trial or moment (h) when the
where i is fixed, such a subset is labelled behavior is interpreted as a relation be-
At. The subset Ai then consists of tween a pair of points from distinct sets
pairs of elements, one a fixed individual (conveniently referred to as an individ-
*, and the other a stimulus j. ual, i, and a stimulus, /) we conceive of
For some models the members of a the behavior as being generated by
pair of points are drawn from the same some but not necessarily all of their
set. It makes no difference whether attributes. Hence, the distance between
such a pair is regarded as being drawn the pair of points is a distance in a
from the set C or the set Q but inas- subspace of the total space called the
much as such pairs of points are usually relevant dimensions, D'.
identified with stimuli they will be re- In a similar manner, if the behavior
garded as drawn from the set Q and of an individual (»'), on a given trial
the set of such pairs of points drawn or moment (h) is interpreted as a rela-
from the same set is called B. A typical tion on a pair of points from the same
example is in the scaling by pair com- set (the stimuli / and k ) , we conceive
parison of lifted weights or brightness of the behavior as being generated by
of lights. some but not necessarily all of the
It is interesting to note that when attributes in the space K. Hence, the
models deal with observations on pairs distance between the pair of points j
of pairs of points the pairs of points are and k is a distance in the subspace
always drawn from the same set, A or called the relevant dimensions, D".
B, never is one drawn from A and the The discussion of these axioms and
other from B. There is nothing log- their accompanying definitions complete
ically necessary about this, of course, it the description of the basis of the theory
is just that there are no psychological of data. We proceed next to construct
measurement models specifically con- the eight types of data and the definition
structed for such data. of the information in each. The objec-
The second axiom postulates the tive here is to illustrate the mapping
existence of a "distance" function in the between this abstract model on the one
space K—that is, between every pair of hand and the types of observations
points in the space K there is a "dis- made by psychologists.
tance" between them. It is to be noted
that nothing is said about this distance THE CLASSIFICATION OF PSYCHOLOG-
concept other than that it satisfies the ICAL MEASUREMENT MODELS
minimum conditions for a distance func- AND METHODOLOGIES
tion. It is not required, for example, Taking the conjunction or cross-par-
that the space be Euclidean or any tition of the three dichotomies given in
other particular geometry. Some mod- Axiom 4 yields eight classes—a cube
els do not require a metric space at all which is 2 X 2 X 2. For simplicity of
and may be spoken of as "recovering" portrayal it may be drawn in a plane
the space at a lower level. with one of the three dichotomies pro-
The third axiom merely links the jected onto the other two. I have
first two together and defines a positive chosen to project the third because this
direction for each dimension Ka. dichotomy is collasped when an inter-
A THEORY OF DATA 147
mediate category of response of the pr*. of pis. pr>. of prs.of pts.
nique of multiple factor analysis can be amples of it will be given. One illus-
used for the multidimensional analysis tration of this type of data is mental
of this same data. test behavior—for example, the response
of an individual to an arithmetic item.
Quadrant Ib An individual passing an item implies
An illustration of such data as this is that he has more of the ability involved
the behavior of an individual when than is required by the difficulty of the
asked whether or not he can choose be- item. The phenotypic behavior, pass
tween a pair of alternatives. For every or fail, is interpreted as an order rela-
pair of alternatives he answers merely tion on a pair of points—one the in-
that he would or would not prefer one dividual's ability and the other the dif-
alternative more than the other. Such ficulty of the item.
data are proximity relations on pairs of In another context, the study by Janis
distances. There are no models spe- (1949) on psychophysiological corre-
cifically designed to analyze such data, lates of fear is interpreted as this type
perhaps because the amount to be of data. When the individual said
learned from such data is small com- "yes" or "no" to having experienced a
pared with Quadrant la data. The particular fear symptom in combat, this
primary purpose in mentioning it here, was interpreted as an order relation on
aside from formal completeness, is its a pair of points, one corresponding to
relevance to the interpretation of pref- the amount of fear he had experienced
erential choice data when an inter- and the other corresponding to the
mediate category of judgment is used. amount implied by the symptom.
Such data constitute what might be Responses of individuals to neurotic
called Quadrant I in that there are data inventory type items are typically inter-
in each of the two octants. preted as this kind of data also. An
individual answers "yes" or "no" to the
Quadrant I question "Do you wet your bed fre-
If an individual is asked which of quently?" The answer is interpreted
two alternatives he prefers and is per- as an order relation between a pair of
mitted to answer "I don't know" or points, one corresponding to the indi-
"I can't say" we may speak of this as vidual's perception of how frequently
using an intermediate category of judg- he wets his bed and the other corre-
ment. sponding to his decision as to how often
The result would be that when an one must wet his bed to do it "fre-
individual made a choice the data would quently."
be Quadrant la and when he failed to Psychophysical threshold studies deal
make a choice it would be Quadrant Ib. with identically equivalent data in a
No models exist for the analysis of such formal sense. When an individual is
data to yield a joint space; one would asked whether or not he perceives a
be compelled to neglect responses using stimulus, the behavior is interpreted as
the intermediate category of judgment an order relation on a pair of points one
or to make assumptions which would corresponding to his threshold and the
map it into Quadrant la data for which other corresponding to the magnitude
models are available. of the stimulus. In the case of deter-
mining a difference limen, the same
Quadrant Ha thing holds if the stimulus is defined to
This is the most prevalent kind of be a difference between two stimuli.
psychological data so a number of ex- Clearly the data of psychophysical
A THEORY OF DATA 149
studies, mental tests, and neurotic in- no matter how the individuals behave,
ventories all have the same formal and then measures of reliability and
character of consisting of order rela- homogeneity are constructed which are
tions on pairs of points from distinct related to how good a fit is obtained
sets. But certain further differences by the one dimension. Another model,
remain. The great advantage that psy- like scalogram analysis, yields one di-
chophysics has over mental testing is mension but is highly sensitive to im-
the apparent possibility of experimen- precision or lack of unidimensional
tally independent replication of the in- homogeneity. This method is essen-
dividual-stimulus pair. The great ad- tially a method for testing whether
vantage that mental testing has over certain conditions for unidimensionality
that of the neurotic inventory is that in the genotypic structure generating
the individual is not able to decide for the behavior are met. One has one's
himself how difficult an arithmetic prob- choice here between models which are
lem is. The difficulty of an arithmetic deterministic or probabilistic and which
problem is presumably more compelling are unidimensional or multidimensional.
than how shy you have to be to say yes. Torgerson's law of categorical judg-
In abstract terms, the point correspond- ment is designed to analyze the data
ing to the difficulty of an arithmetic obtained by having individuals sort
item is, presumably, relatively stable stimuli into ordered classes and the oc-
over different individuals compared currence of a stimulus in a particular
with the point corresponding to the class is interpreted as an order relation
item, "Are you shy ?" But in a formal between a point corresponding to the
sense the data from these separate areas stimulus and a point corresponding to
are all equivalent, the same basic meas- one of the boundaries of the class. Here
urement models apply, and the con- the two distinct sets of real world ob-
cepts and problems of one suggest jects are the stimuli being sorted and
equivalences in the other. the boundaries of the classes.
There are a variety of methods and
models for the analysis of this kind of Quadrant lib
data. These include Guttman's scalo- Typical of the kind of behavior
gram analysis (Green, 1954), Lazars- mapped into this octant is that of an
feld's latent distance model (Green, individual agreeing or disagreeing to
1954), test theory (Gulliksen, 1950), endorse a statement of opinion. Such
the law of categorical judgment (Tor- behavior is interpreted as a proximity
gerson, 1958), Coombs-Kao nonmetric relation on a pair of points, one cor-
factor analysis models (1955), and responding to the attitude of the in-
multiple factor analysis (Thurstone, dividual, the other corresponding to the
1947). attitude expressed by the statement of
These models all yield a joint space opinion. More broadly, the interpreta-
in which the elements of two distinct tion is that there are two sets of ele-
sets are located, typically individuals ments, and the elements of one set are
and stimuli, except in the case of the being matched with the elements of
law of categorical judgment, as ex- another set. Clearly, in the most gen-
plained below. They vary in their as- eral case, a labelling process. Member-
sumptions and hence correspond to dif- ship in an organization, marriage, and
ferent theories of how the behavior is clinical diagnosis are all further ex-
generated. Some models, like test amples of matching between elements
theory yield a one-dimensional / scale of distinct sets.
150 C. H. COOMBS
Galanter (1956) has since begun the the data are Quadrant Ilia; if the
construction of a model for analyzing matrix is symmetric, as in whom do
such data. Even more recent is the you go to the movies with, the data
work of Hefner (1958) on the con- are Quadrant Illb; and if the matrix
struction of a model for this octant for is mixed, as in whom do you like, the
data obtained by degrading stimuli data are Quadrant III. Thus once
through brief time exposures. again we see relations between models
The potential significance of this type apparently constructed for quite dif-
of data (Quadrant Illb) resides in ferent purposes but now become sug-
the fact that the symmetric predicate gestive for other real world content.
(do the stimuli match or not) yields
information about the distance between Quadrant IVa
a pair of stimuli in the stimulus space The behavior of individuals when
and hence may lead to the exploration presented with two pairs of stimuli and
of multidimensional stimulus spaces. asked which pair is more alike is rep-
The data of Quadrant Ilia, in con- resentative of behavior typically mapped
trast, is derived from an asymmetric into Quadrant IVa. The basic ob-
predicate (which stimulus is more of servations are the comparative simi-
something) and there is serious ques- larities of pairs of stimuli. The indi-
tion as to whether such data could lead vidual is presumed to be responding to
to anything more than a one-dimen- the distance between the members of
sional stimulus scale. These issues are a pair of stimuli. Such data may lead
discussed fully by Goodman. via the unfolding technique to a one-
Clearly, if an experimenter wanted dimensional ordered metric stimulus
to construct a one-dimensional stimulus scale (Coombs, 1954a) if certain con-
scale, an asymmetric predicate should ditions are satisfied by the data. If
be used. Whereas, if he wants to one dimension will not satisfy the data
explore the cognitive space in which then a multidimensional solution may
the perceptions of the stimuli are im- be sought.
bedded, the symmetric predicate of Hays, in a paper being prepared, has
Quadrant Illb is superior. adapted his multidimensional unfolding
However, the construction of multi- solution of Quadrant la data to the data
dimensional spaces is also possible with of Quadrant IVa and the result is a
the asymmetric predicate of Quadrant nonmetric model for multidimensional
IVa and will be discussed below. psychophysics which requires only an
order relation on the distances between
Quadrant HI pairs of points in the space (assumed
Finally, as in the case of the pre- to be Euclidean). This model yields
ceding quadrants, if an individual in the stimulus space recovered only at
judging which of two stimuli is greater the level of a product of simple orders.
were permitted to respond "I can't An example is contained in Coombs
decide" the behavior constitutes Quad- (1958).
rant III data and a finite e correspond- Torgerson's model (1958) assumes
ing to a threshold for decision would be more and yields more: (a) the order
involved. relations on distances are transformed
It is of interest to note that socio- into measures of the distances on a
metric matrices contain data belonging ratio scale by a Quadrant Ilia model
in this quadrant. If the matrix is (b) these distances between pairs of
asymmetric, as in who bosses whom, points are transformed into scalar prod-
A THEORY OF DATA 153
ucts (c) which are then factor analyzed. convenience, these four quadrants may
The result is the recovery of a real be given names to signify descriptively
Euclidean space. the type of behavior that is mapped into
each.
Quadrant IVb In Quadrant I, the relation observed
The type of behavior which would be is on a pair of distances where each
mapped into Quadrant IVb would be distance is between a pair of points
the response of an individual to two from distinct sets, usually an individual
pairs of stimuli which was interpreted and a stimulus. The real world con-
to mean that one pair was no more text in which this kind of data is most
alike (or different) than the other pair. commonly obtained is in observing the
Again, as in Quadrant IVa, the in- preferential choices of an individual
dividual is presumed to be responding over a set of stimuli. The data may be
to the distances between the members referred to as individual-stimulus dif-
of a pair and the response is interpreted ferences comparison or, more meaning-
as a proximity relation on these dis- fully, Preferential Choice data.
tances. Observations collected by the In Quadrant II, the relation observed
methods of equisection in psychophysics is on a pair of points which are from
are representative of behavior mapped distinct sets, typically an individual and
into this octant. a stimulus. Such data may be called
The data of this quadrant seem to be individual-stimulus comparison data,
of only slightly more interest than that more commonly known in psychology
of Quadrant Ib. as Single Stimulus data. It is im-
portant to note that this includes not
Quadrant IV only mental test data, endorsing state-
If an individual were permitted to ments of opinion, and psychophysical
judge that one pair is more alike than threshold data, but also rating scales,
the other or that he can't decide, the the method of successive intervals,
behavior would fall into both Quadrant magnitude estimation, and, in general
IVa and IVb and would constitute what is known as absolute judgment
Quadrant IV data again introducing a data.
finite E as a parameter. In Quadrant III, the relation ob-
served is on a pair of points which are
GENERAL DISCUSSION from the same set, called stimuli. So
such data may be called Stimuli Com-
The Four Basic Kinds of Psychological
parison data. One is tempted to call
Data methods for collecting such data psy-
From the preceding discussion of the chophysical methods, but this would
information in data, it is apparent that lead to confusion with psychophysical
the dichotomy of whether an order rela- studies of thresholds which belong in
tion or a proximity relation is observed Quadrant II. Also, such a name is
is a subordinate dichotomy to the others content-bound and there are many ex-
in the sense that it is not satisfied when amples outside of conventional psy-
an intermediate category of response is chophysics which deal with identically
used by the experimenter. Hence the the same kind of data—as for example,
four quadrants rather than the eight sociometric matrices.
octants may be seen as representing In Quadrant IV, the relation ob-
four primary kinds of psychological dis- served is on a pair of distances where
tance observations. As a mnemonic each distance is between a pair of stim-
154 C. H. COOMBS
A BRIEF GLOSSARY
Individuals = undefined. to objects in the real world (stimuli
Stimuli = undefined. and/or individuals) and the relations
Behavior = any potentially observable among points reflect the observed rela-
relation among individuals and stimuli. tions among real world objects.
Raw Data = observed relations among A Psychological Measurement Model
real world objects, for example, pass- = a set of assumptions from which is
fail, preferential choice, yes-no, accept- derived a calculus to construct a psy-
reject, amplitude, latency, inconsistency, chological space from the data matrix.
etc. A Method of Collecting Data = the
Data = raw data mapped into relations rules and the lore for arriving at raw
between points. data.
A Psychological Space = an abstract A Behavior Theory = the mapping from
space in which lie points corresponding raw data into data.
APPENDIX
The following sets are given : Axiom 3. Given two vectors differing
only in one component, the sign of p is
determined by that one component.
H= {l,2,- Axiom 4. Every measurement model
I = {l,2,- - i - - - m satisfies the following three conditions:
j = { l , 2 , - -j,k,l.-- 1. A relation exists on a pair of
Axiom 1. There exists distinct sets points, or a pair of pairs of points.
, d e D, where each JTW) is a segment 2. The elements of the pair of points
are drawn from two distinct sets as
of the real line. in A or from one set as in B.
Definition. Let K = { x \ x = (xm,xm, 3. The relation is either a proximity
• • -xw,- • -XM] where xw f Kw, in which relation (0) or an order rela-
the elements x are vectors in r-dimen- tion (>).
sional space. Let, in addition,
Axiom 5. To each triple (h,i,j) and to
:, c= each quadruple (M,jk) corresponds a sub-
set D' = D'(h,i,j) or D" = D"(hi,jk) of
A = C X Q, A = (ci,g.i)\ D, that is, D' C D, D" C D. D' or D",
(d,qj) Infixed} as the case may be, will be called the set
of relevant dimensions.
B = QXQ, B=
Definition.
The sets A, At, and B are sets whose
elements are pairs of points. We will 1. qhu is the projection of the vector
also have a need for sets whose elements g,- in the set of relevant dimensions,
consist of pairs of such pairs of points, in D' or D".
particular we construct the sets: 2. Cuj is the projection of the vector
Ci in the set of relevant dimen-
AX A, Ai X Ait BX B sions, D'.
Axiom 2. There exists a function p on 3. phu = p(chij,q.Mi) is the image of
the ordered pair (chii,qhii), (the
K X K into the real line, which satisfies "distance" between the pair of
the following conditions for a distance points) in the set of relevant di-
function: mensions, D'.
4. pu.ik — p(q.Mi,q.Mk) is the image of
p(a,b) = 0 {=} a = 6 the function p in the set of relevant
\P(a,b}\<\p(a,c)\+\p(b,c)\ dimensions D".
A THEORY OF DATA 159
Definitions of the information in the where : i JlJ j signifies responses of the
behavior mapped into the various quad- form the individual i says yes, agrees,
rants (see Fig. 1). endorses, etc., the stimulus j. More
generally, an element i of one set is
Quadrant la matched with an element j of another set.
\Pw\-\phik\ <0{=)j?>k Quadrant II with an intermediate cate-
gory of responses
where : jf> k signifies responses of the
form "j preferred to k." phu > ehij {=} i > j
\phii\ ^ tha(=)iMj
Quadrant Ib
phu < ««)(=) j > i
\\Phii\-\PMk\\ < tM.ik(=)j Mk
Quadrant Ilia
where: e is a nonnegative number and
j M K signifies responses of the form "I
cannot choose between j and k" or "I do where : j > k signifies responses of the
not prefer one more than the other." form "j is greater than k,"
(The symbol Mis used to signify "matches
in preference.") Quadrant Illb
Quadrant I with an intermediate cate-
\Pki,ik\< f M . j k (=)j M k
gory of responses
where : j M k signifies responses of the
\PMi\- \Phik\ < — fhi.ik (=)j>. k form "these stimuli j and k are not
\\Phij\-\Phik\\ < tM.ik(=)j Mk different, they match."
\pkii\- \phik\>tki.ik(=)k>j Quadrant III with an intermediate
Quadrant Ila category of response
Phi.ik > fhi.jk (=) j > k
\PM.jk\< thi.jk (=)j Mk
where : i > j signifies responses of the phi.jk < —fhi.ik (=} k > j
form the individual, i, passes, accepts,
etc., the stimulus j. More generally, an Quadrant IV a
element * of one set dominates an element
j of another set. \phi.ik\~ \pu.*\< 0<=> 0',*) < M
Quadrant IVb
\PM.K\- \Phi.lm\\ < «M,,-*,Ii» (=> 0'.*) M (l,m)
where: (j,k) M (l,m) signifies responses of the form "the pair of stimuli (j,k) are no
more different than are the pair of stimuli (l,m)."
\Phi.ik\~ \Phi,lm\ < —fhi.ik.lm <=> (j,k) < (l,m)
\\Phi.ik\~ \Phi,lm\\ < fhi.jkjm {=> 0'.*) M (l>m~)
\pM,jk\— \pM,lm\ > thi,jk,lm <=> (I,™) < (j,k)