Beruflich Dokumente
Kultur Dokumente
II WHAT IS MEASUREMENT?
Foundations
A. A PHYSICAL ANALOGY
of
B. RELATIONS Measurement
C. DEFINITION OF Theory
MEASUREMENT
(Copyright, 2008)
Middlebury College divides its academic year into three major components:
two 12-week semesters sandwiching a four-week "Winter Term." During the Winter
Term each faculty member offers, and each student enrolls in, one course. Because of
the experimental nature of many of the courses offered, enrollment is often
restricted to 20 or 25 students in each class. Since a typical Winter Term will find
1800 students on campus and only 70 courses, it is clear that not every student will
be able to take the course she most desires.
When a student registers for Winter Term, then, she lists five courses in
descending order of preference. The registrar assigns each student to a course, using
these preferences as a guide. At the present time, the registrar uses a procedure
based on the desire to maximize the number of students who receive their first
choice. There has been considerable discussion lately about the fairness and
desirability of this particular priority scheme. An alternative method of assigning
students to courses has been devised which has gained some support. The
philosophy behind this scheme is not to maximize the number of first choices, but to
maximize the total amount of happiness among the students towards the courses
they are assigned. This assignment procedure can be given a rather tidy
mathematical formulation.
Denote the students by i = 1,2, ..., n and the courses by j = 1,2, ..., m. Let rij
denote how happy student i would be if she is assigned course j. Define the variable
xij to be equal to 1 if student i is placed in course j and 0 otherwise. The total
amount of happiness would then be represented by
r11x11 + r21x21 +... + rn1xn1 + r12x12 + ...+ rnmxnm
or, in more compact form,
m n
rijxij .
j=1 i=1
There are several restrictions on the registrar which must be taken into
account. First of all, every student must be assigned to some course and only to one
course. Second, no course should be assigned more students than the instructor is
willing to admit. Denoting the enrollment limit on the jth course by Cj, the
Registrar's Problem is formulated as follows:
August 28, 2008 Chapter 7: Foundations of Measurement Theory Page 3
m n
Maximize rijxij .
j=1 i=1
m
(2) xij = 1 for all i, and
j=1
.
n
(3 ) xij Cj, for all j
i=1
In this chapter and the succeeding one, we want to focus more sharply on the
aspect of the Registrar's Problem which remains somewhat vague in this
presentation: what precisely is rij and how is it determined?
A. A Physical Analogy
We let rij denote "how happy student i would be if she is assigned course j." In
the mathematical formulation of the Registrar's Problem, it is clear that we are
presuming that each rij is a real number which measures this happiness. Is it clear,
however, that it is always possible to measure such psychological attributes by
numbers? What is meant by "measuring" an attribute? Is there more than one way
to do it? What inferences, if any, can be made from a measurement scale? How can
you construct such a scale?
These questions form the basic problems of measurement theory. In the mid-
1960's four distinguished social scientists [David Krantz, R. Duncan Luce, Patrick
Suppes, and Amos Tversky, 1972] began a collaborative study of the foundations of
August 28, 2008 Chapter 7: Foundations of Measurement Theory Page 4
measurement theory that resulted in three large volumes. Early in their first book,
the authors discuss the roles of theories of measurement in science:
". ..The measurability of the variables of interest in physics is taken for
granted and the actual measurements are reduced, via the elaborate
superstructure of physical theory, to comparatively indirect observations.
Other sciences, especially those having to do with human beings, approach
measurement with considerably less confidence. In the behavioral and social
sciences we are not entirely certain which variables can be measured nor
which theories really apply to those we believe to be measurable; and we do
not have a superstructure of well-established theory that can be used to
devise practical schemes of measurement. ...A recurrent temptation when we
need to measure an attribute of interest is to try to avoid the difficult
theoretical and empirical issues posed by fundamental measurement by
substituting some easily measured physical quantity that is believed to be
strongly correlated with the attribute in question: hours of deprivation in
lieu of hunger; skin resistance in lieu of anxiety; milliamperes of current in
lieu of aversiveness, etc."
It should not be surprising then that our first insights into measurement will
come from considerations of measurement in the physical sciences. The question of
what is meant by "measuring an attribute" may perhaps best be answered by
examining first a physical attribute, weight. A provisional definition of measuring an
object's weight might be "assign some number to that object." This is a very poor
definition, since the same number might be assigned to every object.
This added restriction rules out the possibility of assigning all objects the
same number. It relies on the fact that the concept of the "weight" of an object is
intimately connected with a relation between objects, the relation "at least as heavy
as.'' This relation can be established empirically by placing any pair of objects on the
separate pans of a balance and observing which pan descends.
It is natural to define" A has the same weight as B" to mean that A * B and B
* A. As an easy exercise, the reader should show that w(A) = w(B) if and only if A has
the same weight as B.
August 28, 2008 Chapter 7: Foundations of Measurement Theory Page 5
Next consider the set of remaining objects which have yet to be assigned a
weight. Find a lightest element in this set. Assign weight 1 to this object and to any
object of the same weight. Repeat the process on the set of remaining objects
(assigning weight 2 to its lightest element) and continue in this manner until all
objects have been assigned a weight.
B. Relations
EXAMPLE 1 Let S be a set with four elements, S = {a, b, c, d}. The Cartesian
product S S consists of 16 ordered pairs,
S S = {(a,a),...,(a,d),(b,a),...,(b,d),...,(d,a),...,(d,d)}.
A relation on S consists of some subset of these sixteen ordered pairs. One such
example is a relation with three elements, R = {(a, c), (a, d), (b, d)}. We have aRc,
aRd, and bRd and for no other pair i and j is it true that iRj.
EXAMPLE 2 Let S be the set of all positive integers and consider the relation
R defined by x R y if and only if the difference x - y is even. Thus (2,4) is an element of
R, while (3,2) is not. The relation R consists of all pairs (x, y) such that either both x
and y are even or both x and y are odd.
EXAMPLE 3 Let S be the set of all real numbers and let R be the set of all ordered
pairs (x, y) such that x y. Note that (5, 3) is an element of R, but (3, 5) is not. Since
the Cartesian product S S consists of all ordered pairs of real numbers, it can be
represented geometrically by the points in the plane. Any relation on S then
August 28, 2008 Chapter 7: Foundations of Measurement Theory Page 6
corresponds to some subset of the plane. The relation R defined here is shown
graphically in Fig. 7.1.
x
Fig. 7.1 The shaded region R consists of all pairs (x,y) of real numbers such that x
y.
EXAMPLE 4 Let S be the set of all people in Georgia and let R be the relation
defined by x R y if and only if x knows y.
EXAMPLE 5 Let S be the set of all men in the United States Navy and let R
be the relation defined by x R y if and only if y must obey an order given by x.
EXAMPLE 6 Let S be the set of all automobiles in Honest Harry's Used Car
Lot. Define a relation x R y if and only if x costs more than y.
EXAMPLE 7 Let S be the set of all objects in your attic and let R be the
relation defined by x R y if and only if x is at least as old as y.
EXAMPLE 8 Let S be the set of all words in the English language and let R be
the relation defined by x R y if and only if x precedes y in the dictionary.
EXAMPLE 9 Let S be the set of all ordered pairs of real numbers. Define a
relation (x, y) R (x', y') if and only if x < x' or ( x = x' and y < y'). For example, we have
August 28, 2008 Chapter 7: Foundations of Measurement Theory Page 7
(3,20) R (5,11) and (3,20) R (3, 21). This relation is called the lexicographic or
dictionary order.
EXAMPLE 10 Let S be the set of all courses offered by your college and let R be
the relation x R y if and only if you like course x at least as much as course y.
EXAMPLE 11 Let S be the set of all courses offered by your college and let R be
the relation x R y if and only if course x is a prerequisite for course y.
C. Definition of Measurement
Note first that it is not always possible to find a measure for a given
relational system.
EXAMPLE 12 Let S be the set of three elements {x, y, z} and let R be the
relation {(x, y), (y, z), (z, x)}. The relational system < S,R > has no measure. Suppose,
to the contrary, that there is a measure m. Since xRy and yRz, we must have m(x)
m(y) and m(y) m(z). But m(x), m(y), and m(z) are real numbers so it follows that
m(x) m(z). Since m is a measure, the definition implies that xRz or (x, z) is an
element of R. The ordered pair (x,z), however, does not belong to R. The assumption
that <S,R> has a measure leads to a contradiction.
Theorem 1 says that one necessary condition for a relational system to have a
measure is that the relation be transitive. It is easy to establish a second necessary
condition; namely, the relation must be connected.
Proof If x and y are any two elements of S, then m(x) and m(y) are defined and are
real numbers. It must be true that either m(x) m(y) or m(y) m(x). In the former
case, x R y and in the latter, y R x. Thus either (x, y) R or (y,x) R.
Proof Theorems 1 and 2 establish the "only if" part of the conclusion. It
remains to show that if R is connected and transitive, then it is always possible to
find a measure. The idea behind the proof is essentially the same as the one used in
describing how to assign numerical weights to a set of objects.
Denote the elements of the set S by xl, x2, ..., xn. Since the relation R is
connected and transitive, we can find, by checking all possible pairs of elements, an
element xj such that xi R xj for all i j. Define m(x) to be O. If there is any element x i
so that xjRxi as well as xiRxj, then define m(xj) to be 0 also.
At this point, at least one and possibly more elements of S have been
assigned measure 0. Consider the subset S' of remaining elements. Find an element
xk of S' so that xm R xk for all xm xk in S'. Define m(xk) = 1. If there is any other
element xm of S' with xk R xm as well as xm R xk, then also define m(xm) = 1.
August 28, 2008 Chapter 7: Foundations of Measurement Theory Page 10
Repeat the entire process on the set S" of elements which have not yet been
assigned measures, using a measure of 2 to distinguish one or more special
members of S". Continue in the indicated manner until each element of S has been
assigned a measure. This will take at most n steps.
We used the finiteness of the set S at several crucial steps in the proof of
Theorem 3. The theorem remains true if S is a countably infinite set. but may fail if
S is uncountable; see the Exercises for the relevant definitions and examples.
Note that the proof of Theorem 3 not only establishes the existence of a
measure, but provides an effective method of constructing one.
EXAMPLE 13 Let S be the set of three elements {x,y,z} and R the relation
{(x,y),(y,z),(x,z)}. Since this relation is connected and transitive, the elements can be
represented numerically by a measure, according to Theorem 3. If the procedure
outlined in the proof of that theorem is followed, the result is
m(z) = 0, m(y) = 1, and m(x) = 2.
These values are not determined by the measurement model. We could set
m(z) = -17, m(y) = 23 , and m(x) = 10
and still satisfy the definition of a measure. In fact, any three numbers m(x), m(y),
m(z) satisfying the inequalities m(z) < m(y) < m(x) would be an admissible set of
scale values.
Such scales are called ordinal scales. Any transformation of the scale numbers
that preserves their original order yields another admissible scale. A
transformation that changes the order in any way would give a set of scale values
which is not admissible. The resulting numbers, m(x), would define a function which
is not a measure.
The reason this model gives only a partial solution to the question originally
asked will become apparent in the next section of this chapter.
Write (x, y) R* (z, w) to denote the student's judgment that the difference in
happiness between courses z and w does not exceed the difference in happiness
between courses x and y. Note that R* defines a relation on the set S S. This type
of relation, which is a subset of the set (S S) (S S), is called a quaternary
relation on S as opposed to a binary relation, which is a subset of S S.
The problem is to find a measure which preserves both R and R*. More
precisely, does there exist a real-valued function u defined on the set S such that for
all x, y, z, w in S,
1. u(x) u(y) if and only if x R y, and
2. u(x) - u(y) u(z) -u(w) if and only if (x, y) R* (z, w)?
u(y). The question can then be posed this way: Is there a real valued function u on S
which preserves R* such that the induced relation R' is identical to the relation R?
Proof If x, y, z, and w are any elements of S, then the real numbers u(x) - u( y)
= A and u(z) -u(w) = B must satisfy the inequality A B or the inequality B A. In
the former case, (x, y) R* (z, w), and in the latter, (z, w) R* (x, y). Thus R* is
connected.
Next, suppose (x, y) R* (z, w) and (z, w) R* (a, b), where x, y, z, w, a, b are
arbitrary elements of S. We have the inequalities
u(x) - u(y) u(z) - u(w)
and
u(z) - u(w) u(a) - u(b),
which imply by transitivity
u(x) -u(y) u(a) -u(b),
The four conditions of Theorem 4 are necessary for the existence of the
required measure u, but unlike the case for binary relations, they turn out not to be
sufficient. There exists a finite set and a quaternary relation R* on it that satisfies
the four conditions but for which it is not possible to construct a numerical scale
preserving R*.
In a 1958 paper in the Journal of Symbolic Logic, Dana Scott and Patrick
Suppes prove an even stronger result: If S is a finite set and R* a quaternary
relation on S, then there is no finite list of axioms that provides necessary and
sufficient conditions for the existence of a real-valued function u preserving R*.
Scott and Suppes cite an example, essentially due to Herman Rubin, that
indicates the kind of difficulty that arises in trying to construct a set of necessary
and sufficient conditions.
In each pair, the first course is preferred to the second. Suppose the student
perceives A,B,C,D as equal in difference to E,F,G,H, respectively, that the difference
between courses in K is greater than the difference in courses in J and that the
difference in I is greater than the difference in K. Then the relations between the
remaining pairs may be chosen so that any subset of nine courses can be represented
by a measure u which preserves R*, but the full set of 10 courses cannot! The
interested reader may wish to work out the details of this example.
V. AN ALTERNATIVE APPROACH
k-1
u(xk) = u(xk-1) + m(xk-1, xk) = m(xj, xj+ 1) ,
j=1
n-1
u(xn) = u(xn-1) + m(xn-1,xn) = m(xj,xj+1)
j=1
Hence the measure u also preserves R. This completes the proof of sufficiency. For
necessity, see Exercise 29.
To illustrate the procedure outlined in the proof, suppose we have a set of five
elements with standard order x1,x2,x3,x4,x5. Then the set T consists of four pairs,
T = {(x1,x2), (x2,x3), (x3,x4), (x4,x5)}
(see Fig. 7.2).
Fig. .7.2 The heavy dots indicate the elements of S @ S. The members of T are
circled.
Suppose that examination of the relation R indicates that the standard ordering on
T is
(x3, x4), (x2, x3), (x4, x5), (x1, x2),
August 28, 2008 Chapter 7: Foundations of Measurement Theory Page 16
so that m(x3, x4) m(x2, x3) m(x4, x5) m(x1, x2) for every measure m on the
system < T, R). If scale values of 3, 5, 6, 7 are chosen for m, then the proof of
Theorem 5 defines a measure u by
Now the measure u can be used to define a relation Ru on the full set S S.
Define (x, y) Ru (z, w) if and only if u(y) -u(x) u(w) -u(z). Note that R is a subset of
Ru, so we might say that Ru extends R. As an example, note that since
u(x4) -u(x2) = 15 -7 = 8
while
,
u(x5)-u(x4) = 21 - 5 =6
we have
(x2,x4) R (x4,x5).
The choice of scale values for m is, as we have seen earlier, unique only up to
an order-preserving transformation. We might have chosen, with equal validity,
scale values of 1,2,4,8. With these values for m, we obtain a measure v on S with
v(x1) = 0, v(x2) = 8, v(x3) = 10, v(x4) = 11, v(x5) = 15.
One set of scale values for m is consistent with the student's judgment that
there is a greater difference between x2 and x4 than between x4 and x5, while a
different set of scale values is not. In this alternative approach, we have agreed not
to ask the subject to make comparisons between pairs (x2, x4) and (x4, x5). This
example shows that we cannot determine what judgment the student would make
on these pairs solely on the information we have concerning the pairs in the set T.
The Second Representation Theorem (Theorem 5) then gives a measure u which is a
better reflection of the student's attitude toward the courses in the winter term than
the ordinal measure m, but it does not completely answer the objections raised by
the student complaint of Section IV.
1. Given any two objects, it is always possible to decide which one "possesses"
more of the quantity than the other;
R. Duncan Luce
http://today.uci.edu/image_library/profile/luce
Patrick Suppes
http://www.stanford.edu/~psuppes/
http://www-history.mcs.st-andrews.ac.uk/Mathematicians/Holder.html
EXERCISES
2. How would you describe, using the "*" notation, the fact that A is heavier than B?
3. Let A, B, and C be any three objects. Show that the following statements are all
true:
a) A * A
b) If A * B and B * C, then A * C.
c) Either A * B or B * A or both.
4. If w(A) = 1 and two copies of object A exactly balance one copy of B in a pan
balance, does it follow from the procedure outlined in II.A that w(B) = 2w(A)? Why?
How would you modify the procedure to insure this?
August 28, 2008 Chapter 7: Foundations of Measurement Theory Page 19
5. Determine which of the Examples 1-12 are reflexive, symmetric, transitive, and
connected.
6. Let C be the set of all ordered pairs (a, b) of real numbers, Define (a, b) R (c, d) if
and only if a > c and b > d. Is this relation transitive? Is it connected?
10. Consider the relations on the set of real numbers determined by the concepts of
>, , , , and . Which are weak orders? strong orders? equivalence relations?
12. Let S be the set of all adults in New England. Define a relation R by x R y if and
only if x lives with y. Is R an equivalence relation?
14. Can you carry out the process of creating equivalence classes defined in Exercise
13 if R is not an equivalence relation? Why?
15. A semi-order is a relation P on a set S satisfying the following three axioms for
all x,y,z,w in S:
i) x P/ x
ii) If x P y and z P w, then either x P w or z P y.
iii) If xP y and yP z, then either xPw or WP z.
17. At what steps in the proof of Theorem 3 is the finiteness of the set S used?
18. Show that procedure of the proof of Theorem 3 leads to m(z) = 0, m(y) = 1, m(x) =
2 for the relation of Example 12.
19. For Example 12, show that the function with scale values m(x) =
-17, m(y) = 23 m(x) = 10 also satisfies the definition of a measure.
21. Suppose that m and u are measures on the relational system < S,R > where S is
finite. Show that there is an order-preserving function f: M M where M = {m(xi)}
such that u(xi) = f(m(xi)) for all i.
22. How many different questions of the type "Do you prefer course i to course j?"
must you ask a student to construct a preference ordering for a set of70 courses?
23. Show that the relation of Example 9 is connected and transitive, but the system
<S, R > has no measure in the sense of Theorem 3.
25. a) Let be the set of all ordered pairs (a, b) of real numbers such that a = 0 or 1
and 0 b 1. Let R be the lexicographic order on S. Does the system < S, R> have a
measure?
b) Let T be the set of all ordered pairs (a, b) of real numbers such that a = 0
and b is either between 0 and 1 or between 2 and 3. Let R be the lexicographic order
on T. Does the system < T,R > have a measure?
V. AN ALTERNATIVE APPROACH
29. Prove that the conditions in the statement of Theorem 5 are necessary.
33. Use scale values of 2,3,5, 10 instead of 3,5,6,7 for the function m to determine a
measure w. Show that in the extension Rw you have (x2,x4) Rw (x4, x5) and (x4, x5) Rw
(x2, x4).
and
m(x y) = m(x) +m(y) for all x and y in S.
36. Show that the function m guaranteed by the theorem in Exercise 35 is unique up
to multiplication by a positive constant.
37. Show that the ordinary conceptions of length and weight satisfy the axioms of an
extensive measurement system.
SUGGESTED PROJECTS
3. It has been argued that in many situations, observed equality relations may not
be transitive. A person may judge rod x as long as rod y which in turn is judged as
long as rod z; yet x may be judged longer than z. Such judgments arise whenever the
differences between x and y and between y and z are too small to be noticed. The
combined difference, however, may be sufficiently large to make a difference between
x and z noticeable. The classic example is a sequence of cups of coffee each
containing one more grain of sugar than the previous cup. An observer could probably
detect no difference in sweetness between two adjacent cups. If "equally sweet" is a
transitive relation, then we would have to conclude that a cup with no sugar in it is
as sweet as one in which 10 teaspoons of sugar have been dissolved!
To handle such situations, R. Duncan Luce introduced the idea of a semi-order
as the type of relation to capture the notion of strict preference (See Exercise 15). If P
is a semi-order, then an indifference relation I can be defined by x I y if and only if
neither x P y nor y P x. Show that I is reflexive and symmetric, but not necessarily
transitive. Prove the following Representation Theorem: If P is a semi-order on a
finite set S, then there is a real-valued function f defined on S and a positive number
such that for all x and y in S,
f(x) > g(y) + if and only if x P y.
The constant may be interpreted as a single "just noticeable difference"
unit. Is this Representation Theorem true for infinite sets?
August 28, 2008 Chapter 7: Foundations of Measurement Theory Page 23
REFERENCES
Hlder, Otto, "Die axiome der quantititt und die lehre von mass," Berichte
ber Die Berhandlungen Der Knglich Schsischen Gesellschaft Der Wissenschaften Zu
Leipzig, Mathematisch-Physische Classe 53 (1901),1-64.
Suppes, Patrick, "A set of independent axioms for extensive quantities," Portugaliae
Maihematica 10 (1951), 163-172.
Cliff, Norman, Abstract Measurement Theory and the Revolution That Never
Happened, Psychological Science 3 (1992), 186 190.
Marley, A. A. J.. (ed.), Choice, decision and Measurement: Essays in the Honor of R.
Duncan Luce, (Mahwah, New Jersey: Erlbaum, 1997.).