Artigo Stevens 1971

Psychological Review
1971, Vol. 78, No. 5, 426-450
ISSUES IN PSYCHOPHYSICAL MEASUREMENT

S. S. STEVENS 2
Harvard University
Two classes of ratio-scaling procedures are outlined—magnitude matching

and ratio matching—and their assets and liabilities are noted. Partition-
scaling procedures, which are supposedly designed for interval scaling, pro-
duce results that can be described by a power function with a virtual or
"as if" exponent. Since the virtual exponent is smaller than the actual
exponent of the continuum, the category scale is nonlinear. The virtual
exponent provides a convenient descriptor of several kinds of partition
operations. Other topics discussed include individual differences among
subjects' exponents, procedures of averaging, and the effects of stimulus
range on exponents. It is suggested that the power law asserts a nomothctic
imperative.
The task here is to review the matching comment on only a few of the features that
procedures used to determine the power func- characterize the principal methods. Con-
tions that govern the growth of sensation sideration is also given to some of the in-
magnitude and to consider some of the terval or partition methods, and to the dis-
sources of deviation and perturbation that tinction between the virtual exponent and
have raised questions concerning the nomo- the actual exponent. Other problems dis-
thetic quality of the psychophysical power cussed concern individual differences, aver-
law. aging, and range effects.
Since all procedures of measurement in-
volve matching operations, the interesting MAGNITUDE MATCHING
differences among different scales and differ- These procedures include all direct equa-
ent kinds of measurement can often be re- tions between two continua. Three principal
duced to a basic question : What was matched varieties of magnitude matching have been
to what, and how? In the domain of psy- distinguished.
chophysics, numerous scaling methods have
been invented, many of them useful for the Cross-Modality Matching
determination of ratio scales of apparent When the stimulus for a continuum can be
magnitude. The approaches of ratio scaling readily varied by means of a control of some
can be catalogued in different ways, but for kind, it becomes possible to match that con-
present purposes they fall into two general tinuum to any other continuum. Figure 1
classes : magnitude matching, which includes gives examples of matching functions pro-
the subclasses (a) cross-modality matching, duced when several different continua were
( b ) magnitude estimation, and (c) magni- matched to vibration on the fingertip.
tude production; and ratio matching which Ideally, the experiment comparing two
includes the subclasses (a) cross-modal ra- continua should provide for a balanced de-
tio matching, (6) ratio estimation, and (c) sign in which each continuum is matched in
ratio production. turn to the other continuum. A balanced
Since there are endless variations on psy- procedure may help to assess and correct the
chophysical procedures, it is possible here to regression effects that are always present in
1
Supported in part by Grant NS-02974 from the the matching operation (Stevens & Green-
National Institutes of Health (Laboratory of Psy- baum, 1966). Tn the typical experiment, the
chophysics Report ppr-366-133). observer tends to shorten the range of which-
2
Requests for reprints shoultl be addressed to S. ever variable he controls. Even within the
S. Stevens, Laboratory of Psychophysics, Harvard
University, 33 Kirkland Street, Cambridge, Massa- same modality the regression effect shows up
chusetts 02138. in matching functions. Thus, two somewhat
426
ISSUES IN PSYCHOPHYSICAL MEASUREMENT 427
Relative intensities of criterion stimulus

FIG. 1. Equal sensation functions obtained by cross-modality matches be-
tween various continua and a 60-Hz. vibration on the fingertip. (The vi-
bration amplitudes were set by the experimenter. The observer adjusted the
other stimulus to produce an apparent match—from Stevens, 1968a.)
different functions were obtained, depending B to a common continuum C. The two

on which auditory stimulus the subject ad- matching functions provide exponents whose
justed in matching a tone to a noise. ratio determines the exponent of the power
Because it is often difficult to give the sub- function relating A to B. Whether the re-
ject control of the stimulus, many cross- gression effects in the two matching func-
modality comparisons have not yet been tions A to C and B to C are exactly equal
made. It is difficult, for example, for the may not be known, of course, but the pro-
subject to vary the heaviness of lifted weights cedure may still cancel a major portion of the
in order to match heaviness to loudness. regression bias.
When a balanced design in the matching An instance of the second paradigm was
of two continua is impracticable, an evalua- provided by Moskowitz (1969), who asked
tion of the regression effect may sometimes observers to match both numbers and loud-
prove possible by way of a third continuum. ness to a wide variety of taste mixtures.
Two principal paradigms can be distin- From each of the 68 pairs of experiments
guished. that Moskowitz conducted we can derive an
1. Continuum A is adjusted to match each estimate of the power-function exponent re-
of two continua, B and C. The ratio of the lating number to loudness. The geometric
exponents of the matching function A to B mean of the 68 estimates was .67, which
and A to C determines the exponent of the agrees with the value of the exponent
function relating B to C. The derived func- adopted for loudness calculation (Stevens, in
tion would presumably be free of the regres- press). The standard deviation was .51
sion e,ffect, provided the regression oc- decilog, or about 12%. That degree of scat-
casioned by adjusting A remained constant ter may be regarded as an empirical guide to
when the criterion continuum was changed the amount of variability to be expected in
from B to C. The hoped-for constancy tests of transitivity among the exponents of
might be upset by such factors as disparities the power functions. When the regression
in difficulty or range. effect has been canceled, about two-thirds of
2. A second procedure for counterbalanc- the measured exponents may be expected to
ing the regression effect can be utilized lie within plus or minus half a decilog of the
whenever it is possible to match both A and predicted exponents (see Stevens, 1969).
428 S. S. STEVENS
The regression effect, of course, is only ject has learned to recognize a particular
one of the sources contributing to the sys- stimulus, little or no new information is ob-
tematic errors that affect the outcome of ex- tained from subsequent judgments of its
periments, but it is one of the most obstinate, repeated presentation. Furthermore, biases
and therefore perhaps the most important. due to range and spacing of stimuli seem to
And it may be composed of more than a have less effect when the subject is limited
single factor. to one judgment per stimulus.
Untrained, inexperienced college subjects
Magnitude Estimation seem to do as well at the matching tasks as
those who have had many years of practice.
This procedure is actually a form of cross- Hence, there is no need to "train" the sub-
modality matching in which numbers are jects. Indeed, since there is no right or
matched to stimuli. When first used, the wrong to the subjects' responses, it is not
procedure was called absolute judgment clear what would be meant by training.
(Stevens, 1953), later numerical estimation Under some circumstances, the nature of the
(Stevens, 1954), and still later magnitude task may profitably be clarified by allowing
estimation (Stevens, 1955a). That last the subjects to begin by matching numbers
name appears to have stuck. In this con- to an easier continuum, such as apparent
text, the number continuum can be regarded length of lines, or apparent size of circles.
as another perceptual modality. Magnitude Averaging can be done by computing geo-
estimation or "number matching" has be- metric means or medians. The log-log slope
come a popular method, mainly because of its (exponent) determined by the geometric
convenience. The subject brings the num- means is not affected by the fact that each
bers with him, so to speak, and the experi- observer uses a different unit of modulus.
menter needs only to provide the target When it is desired to adjust the judgments
stimuli to which the numbers are to be to a common modulus, a good method is to
matched. The nature of the task can be minimize the squares of the individual sub-
portrayed in terms of a typical set of written ject's intercept differences. The procedure
instructions to the subject. is: convert all scores to logs, compute grand
You will be presented with a series of stimuli in mean of logs, and adjust each log score for
irregular order. Your task is to tell how — they each observer by whatever additive constant
seem by assigning numbers to them. Call the first makes the observer's mean correspond to the
stimulus any number that seems to you appropriate. grand mean. That procedure of modulus
Then assign successive numbers in such a way that
they reflect your subjective impression. For ex- equalisation permits each of an observer's
ample, if a stimulus seems 20 times as •—, assign estimates to contribute to the correction fac-
a number 20 times as large as the first. If it tor to be applied to that observer's modulus.
seems one-fifth as —, assign a number one-fifth as
large, and so forth. Use fractions, whole numbers,
or decimals, but make each assignment proportional Magnitude Production
to the — as you perceive it.
Here the experimenter presents the num-
Experience has shown that it is usually bers one at a time in irregular order, and the
better not to designate a standard. The sub- subject adjusts the stimulus to produce an
ject then remains free to choose his own apparent match. The numbers themselves
modulus. If possible, stimuli should be pre- should normally approximate a geometrical
sented in a different irregular order to each progression. For example, in an extensive
subject, but the first stimulus is usually study of loudness and its inverse, softness,
chosen from among those in the middle re- the successive numbers presented were in the
gion, rather than from one end of the range. ratio 2 to 1 and ranged from 1.25 to 80
Between 10 and 20 stimuli may be presented (Stevens & Guirao, 1962). Sample results
at a session. A good schedule provides for are shown by the triangles in Figure 2.
one judgment, or at the most two judgments, Because of the regression effect, the power
per stimulus by each subject. After the sub- functions obtained by magnitude production
are typically steeper (have larger exponents) 100

than those obtained by magnitude estima-
tion. An unusually large regression angle is
illustrated in Figure 2. It is often assumed
that the unbiased function lies between the
two functions obtained by estimation and
production, and indeed it may. But where?
Are the error sources in the one procedure
exactly balanced by the error sources in the
other ? Although exact balance may be pos-
sible, it seems hardly likely that no asym-
metry exists. An average function may
nevertheless be desired, in which case it may
be well to compute the geometric mean of
the two exponents. The geometric mean is
invariant under an interchange of the two 30 40 SO 60 70 60 90 dB 100
coordinates (Indow & Stevens, 1966). Relative sound pressure
When the results of magnitude estimation FIG. 2. Magnitude estimation and magnitude
and production are combined in an appropri- production of loudness. (Each point is the geo-
ate way, the combined procedure offers ad- metric mean of two estimates or two productions
vantages over either procedure alone. by each of 10 observers.) (Reprinted with per-
mission from an article by J. C. Stevens and M.
A particular version of the combined pro- Guirao published in the Journal of the Acoustical
cedure designed to produce a balanced func- Society of America, 1964, Vol. 36. Copyrighted
tion has been spelled out by Hellman and by the Acoustical Society of America, 1964.)
Zwislocki (1968) and applied to loudness
functions. tions that ranged over about 10,000 to 1,
Another way of combining some of the which implies an exponent near .8. The ex-
features of production and estimation is to periment was repeated some months later.
permit the subject to set stimulus levels at The geometric mean of the 22 exponents for
his own pleasure and to report the apparent the 11 subjects was .7, which is fairly close
magnitude. The experimenter in effect gives to the value 2/3, which has been proposed as
up all control over the stimuli. A radical the standard value (Stevens, in press).
procedure of that kind was used by J. C.
Stevens and Guirao (1964) to show that in-
RATIO MATCHING
dividual subjects produce power functions
and that the power function is not, as some The earliest form of ratio matching ap-
writers had suggested, an artifact of averag- pears to have been devised by Merkel (1888)
ing. The key parts of the instructions were in order to determine what he called the
as follows: "doubled stimulus." It was a direct ratio-
Your task is to set the tone to different levels of scaling method, but its potentialities were
loudness and to assign numbers to each of the not effectively exploited. The method would
loudnesses. Make your numbers proportional to now be classed as one of the varieties of ratio
the loudness you hear. You may make as many production. It is convenient to distinguish
settings as you want. Try to cover a wide range
of loudness.
three subclasses of ratio matching, as follows:
Eleven subjects, chosen at random from Cross-Modal Ratio Matching

among students, staff, and secretaries, gave
the results shown in Figure 3. One subject J. C. Stevens set two different bright-
made as few as seven production estimations. nesses in front of the subject and asked him
Another made five times that many. Two to adjust one of two noises to make the ratio
subjects made settings that extended over of the noises match the apparent ratio of the
almost 100 decibels (db.) and made estima- brightnesses (see S. S. Stevens, 1961a, pp.
430 S. S. STEVENS
10'
Relative sound pressure (subdivision = 20 dB)
FIG. 3. Individual functions obtained when each of the 11 observers set the stimulus level and esti-
mated the loudness. (Each point represents a judgment. There was no averaging—from J. C. Stevens
& Guirao, 1964.)
17-18). The results confirmed the relative tion is the name commonly used for pro-
values of the exponents for loudness and cedures that require the subject to set a
brightness, showing, in fact, that the two stimulus to one-half (or some other fraction)
exponents are approximately equal. of the standard. (For a tabulation of many
of the numerous ratio productions that have
Ratio Estimation been made with acoustic stimuli, see Stevens,
1955a.)
Here the subject matches numerical ratios Ratio production has fallen into disuse
to apparent stimulus ratios. In the "com- mainly because magnitude matching seems to
plete" version of the procedure, stimuli are be a superior procedure. The biases in ratio
presented in all possible pairs and the ap- production are such that the method often
parent ratios are estimated (Ekman, 1958). fails to produce a clean power function.
Other versions use fewer stimulus pairings,
and some versions provide for reporting in INTERVAL MATCHING AND VIRTUAL
terms of fractions or percentages. For ex- EXPONENTS
ample, in an early experiment, Ham and
Parkinson (1932) presented a sound at one Although the judgment of intervals or dif-
level followed by a sound at a lower level, ferences, as required in various kinds of par-
and asked the subjects to estimate what per- titioning operations, may produce satisfactory
centage of the loudness remained. results on metathetic continua, systematic
biases afflict partitions carried out on pro-
thetic continua. Furthermore, the partition-
Ratio Production ing operations can produce at best an interval
In this once-popular scaling procedure, the scale, not a ratio scale. Nevertheless, one or
subject is required to find or produce the another form of interval matching has pro-
stimulus that seems to stand in a prescribed duced data that have played a role in the
relation to a standard stimulus. As we have establishment of the psychophysical power
seen, Merkel invented that kind of task with law (Stevens, 1953).
his method of doubled stimulus. Fractiona- In order to describe the procedures used
and the results obtained in the various kinds than differences, and without intending it, he
of partition operations, it is convenient to may actually produce ratio matches between
distinguish two exponents: a virtual or func- the pairs of stimuli on the two continua.
tional exponent and the actual exponent of Efforts to judge intervals often encounter
the continuum in question. The virtual ex- a dramatic hysteresis effect, which makes the
ponent is the one the observer appears to be judgment highly contingent on the order in
using when he makes his partition judgments. which the stimuli are presented (Stevens,
It is an "as if" exponent. The value of the 1957b).
virtual exponent a turns out to be lower
than that of the actual exponent /?. Since Interval Estimation
a < ft, the scales created by partitioning are
nonlinear relative to the corresponding mag- Here the subject may be asked to assign
nitude scales created by magnitude or ratio numbers to represent the sizes of apparent
matching. differences. For example, Dawson (1968)
Perhaps the best known example of a presented pairs of loudnesses and asked the
partition scale is the Munsell scale for the observers to make a magnitude estimation of
lightness of grays. That scale has been the apparent difference in each pair. He
determined and redetermined by several also presented pairs of visual areas. The
kinds of partition operations. A series of typical biases that emerge under partitioning
gray papers may also be scaled by magnitude procedures were apparent in the results,
estimation, as was shown by Stevens and especially in the judgments of loudness dif-
Galanter (1957). The virtual exponent of ferences. A constant loudness difference (in
the Munsell scale is approximately .33 and sones) is not judged to be constant; rather a
is decidedly lower than the actual exponent, given difference is judged smaller when it is
approximately 1.2, obtained by magnitude moved up the stimulus scale.
estimation. Another demonstration of the bias in in-
Let us now consider four varieties of interval judgments—the operation of a virtual
terval scaling procedures. exponent—is contained in the results of Beck
and Shaw (1967) who asked 28 subjects to
judge four loudness intervals, 5, 10, 15, and
Cross-Modal Interval Matching 20 sones in width, each located at four stim-
This procedure seemed to work well in ulus levels. The median estimations for
a 1953 experiment when subjects adjusted three of the interval sizes are shown in
markers along a line (position, a metathetic Figure 4 as a function of the sound pressure
continuum) in order to match the apparent level of the tone at the lower end of the in-
spacing of a series of loudnesses. Subjects terval. The curved lines show the general
also matched marker position to the apparent trend of the data.
spacing of the heaviness in a series of lifted If the world were so constructed that equal
weights. The same principle is involved, of prothetic intervals appeared equal to the per-
course, in numerous rating scales: the sub- ceiving subject, the lines in Figure 4 would
ject expresses his opinion by marking a posi- be straight and horizontal. The downward
tion on a line. Newhall (1950) used a some- trend of the data in Figure 4 illustrates the
what similar method, involving markers on typical result obtained in partition judgments
a two-dimensional grid, in order to deter- of whatever variety: equal intervals are not
mine spacings among the apparent light- judged equal at different locations on a pro-
nesses of gray papers. His results agreed thetic continuum. As an interval of a con-
with the Munsell scale. As a general stant size moves up the scale of the contin-
method, however, interval matching suffers uum, the constant interval is judged to be
from a basic ambiguity, especially when smaller and smaller.
metathetic position is not one of the continua The curves in Figure 4 were generated by
used. If two prothetic continua are involved, a partition model in which it was postulated
the subject may find it easier to match ratios that the observer's judgments are governed
432 S. S. STEVENS
100 which 29 observers made magnitude estima-
Interval tions of another set of loudness intervals—
intervals that were constructed to be constant
- 50
in size as determined by the lambda scale, a
scale that was constructed so as to agree with
30 a particular set of bisection data (Garner,
o>
T3
= 20
1954). Over the stimulus range of interest
here, the lambda scale has an effective ex-
ponent of approximately .3. In other words,
10 the "constant" intervals provided by the ex-
perimenters were generated by a function
aa
whose exponent coincided with that of the
virtual exponent. As we should expect,
ZO 30 40 50 60 70 80 90 therefore, the judged size of the intervals
Bottom end of interval - dB did not show a downward drift with increas-
ing stimulus level. In fact, when the judg-
FIG. 4. Showing how the judgment of an inter-
val of a constant size depends on the location of the
ments are plotted as in Figure 4, but with
interval. (Observers made magnitude estimations lambda interval rather than sone interval as
of sets of intervals 5, 10, 20 sones wide. In another the parameter, the data describe functions
set of experiments (triangles) the intervals were that are very nearly horizontal. Thus the
approximately 30 sones wide. The stimulus level principle is clear: When the generating func-
at the bottom end of the interval is shown by the
abscissa. The ordinate gives relative values only. tion used to set up the equal intervals has
As a constant interval moves upward in sound pres- the same exponent as the virtual operating
sure level, the perceived size of the interval de- function employed by the observers in their
creases. The family of three curves was generated partition judgments, then the intervals all
by assuming that instead of the actual exponent of appear equal.
the sone scale .6, the observers used a virtual or "as
if" exponent equal to .3. Triangles from Dawson, It is of interest next to consider the other
1968; other data from Beck & Shaw, 1967.) extreme and to ask what happens when the
equal intervals are generated by a function
by a power law that does not have the actual with an exponent that is lower than the vir-
exponent of the continuum, but rather has a tual exponent. Since a power function with
virtual or functional exponent equal to .3. a very low exponent resembles a logarithmic
The fit of the curves is only fair, for the data function (see Figure 5), we can examine the
do not provide enough information to dis- problem by setting up equal logarithmic or
tinguish between the family of functions gen- equal decibel intervals. Again the results
erated by the virtual exponent value .3 and turn out as expected. Decibel intervals ap-
the family given by some other nearby ex- pear to grow larger as their absolute level is
ponent. If the observer's virtual exponent raised. Thus a series of successive 10-db.
were .6, it would correspond to the actual intervals beginning at 40 db. produced the
exponent, and the lines in Figure 4 would following magnitude estimations: 1.51, 1.89,
then become straight and horizontal. As the 2.65, 4.14, and 9.10 (Dawson, 1968). Those
virtual exponent becomes smaller, the family values represent averages over four different
of curves tilts more steeply downward, and experiments. They show that the 10-db. in-
the distance between the curves decreases. terval 80-90 db. was judged to be about six
To a first approximation, then, it appears times larger than the 10-db. interval 40-50
that the observer judges loudness intervals db. If plotted in Figure 4, the judgments of
as if his power function had a virtual ex- 10-db. intervals would describe a curve that
ponent about half as large as the actual sweeps upward rather than downward. In
exponent of the continuum. That principle other words, observers' judgments demon-
was rather nicely confirmed in a second strate that the virtual exponent is decidedly
experiment by Beck and Shaw (1967) in greater than zero, and that a logarithmic
function does not accord with interval judg-

ments.
As a matter of fact, although Dawson did
not use intervals that were exactly constant
in sones, his data agree fairly well with func-
tions having the form of those generated by exponent
a virtual exponent of about .3. The tri- O.I
angles in Figure 4 show the magnitude esti-

mations of the largest sone intervals em-
ployed, ranging from 16 to 31 sones. The
magnitude estimations of those intervals were
multiplied by a factor in order to bring each
of them to the value it would have had if all
the intervals had been 30 sones. That multi-
plicative correction appears to do little or no 2 3 4 S
Log a r i t h m i c scale
violence to the data. The path depicted by
the triangles in Figure 4 shows how the ap- FIG. 5. Sample power functions plotted in semi-
parent size of a 30-sone interval grows logarithmic coordinates. (As the exponent of the
smaller as the stimulus level used to define power function decreases, the graph of the function
becomes straighter. It thereby approaches the form
the bottom end of the interval increases. of a logarithmic function, which is described by a
straight line in these coordinates.)
Interval Production
serve as an upper bound and the bisection-
The production of prescribed intervals was scale exponents normally lie below the bound.
the method invented by Plateau (1872). He In a bisection experiment in which <£a is
asked eight artists to paint a gray such that set midway between </>3 and </>i the virtual ex-
the intervals from gray to black and gray to ponent of the power function can be deter-
white would appear equal. That type of mined by the bisection equation (Stevens,
procedure is often called bisection. When 19SSa), which may be written
more than two equal intervals are involved,
it is called equisection (Garner, 1954).
The intervals to be produced need not be The exponent a may be found by iteration.
equal. For example, in one of my own ex- It seems to be an invariant rule that, as with
periments, observers produced louclness in- other partitioning procedures, the value of
tervals corresponding to markers spaced the virtual exponent determined from bisec-
unevenly along a line. The results from that tion has a lower value than the exponent
procedure of multisection were consistent determined by procedures that call for ratio
with the results obtained by equisection. judgments.
Under conditions designed to make the The family of curves in Figure 4 tells us
judgment maximally easy, the results of bi- that a true bisection on a prothetic con-
section experiments have sometimes been tinuum will not appear correct to the ob-
found to agree fairly closely with the psy- server. The lower half of the interval will
chophysical functions produced by magnitude appear larger than the upper half. Conse-
and ratio matching. In other words, al- quently, when the observer himself makes
though the inevitable systematic difference the bisection, he lowers the bisecting point.
was clearly evident, the size of the difference In some bisection experiments, the bisection
proved to be fairly small (see Stevens, 1955a point has fallen low enough to agree with the
for loudness, 1961b for brightness). To a exponent used to generate the curves in
limited extent, then, the bisection experi- Figure 4 (Garner, 1954). In other experi-
ments have confirmed the exponents of the ments, the virtual exponent has been found
magnitude scale obtained by ratio-scaling to lie closer to the actual exponent (Stevens,
procedures. The magnitude-scale exponents 1955a).
434 S. S. STEVEN'S
Sum Production Under optimal experimental conditions,

the virtual exponent a, determined by bisec-
The virtual exponent describes the con- tion, may approach the value of the actual
vexity in the partition function under still
exponent of the continuum as an upper
another procedure. The bisection problem can bound. Normally, as we have seen, the bi-
be turned around, so to speak, and instead of
section exponent lies below the continuum
asking the observer to produce segments, the
exponent. Nevertheless, the invariance of
experimenter can present the segments and the bisection exponent under multiplicative
ask the observer to produce the whole, or
variations in the stimulus levels has provided
the sum. If the virtual function is convex,
evidence that equal stimulus ratios produce
we should expect that the perceived whole equal sensation ratios—which is the invari-
would prove to be less than the sum of the
ance principle that underlies the power law.
perceived parts. When observers were
shown two or more line segments and asked
to produce a line that appeared equal to the CATEGORY SCALES AND THRESHOLDS
sum, the line produced was longer than the The category scale calls for special con-
sum of the separate lengths (Krueger, 1970). sideration because it is by far the most com-
In other words, a line equal to the sum of mon and yet perhaps the least satisfactory
the separate lengths would have appeared form of partition scale. Most of the many
shorter than the apparent sum. The experi- users of category scales throughout science,
mental results demonstrate that the exponent education, engineering, and commerce do not
1.0, which observers ordinarily use in judg- intend a partitioning operation. Their hope
ing apparent length, has been replaced by a is to grade or assess some variable, but they
virtual exponent having a lower value. proceed to prescribe—and to limit!—the sub-
jects' response scale. Commerce does much
RELATION TO POWER LAW of its buying and selling with the aid of the
Perhaps the most important outcome of grading of goods by subjective assessment on
the work on partition scales lies in the dem- simple category scales, a crude procedure
onstration, by means of interval methods, perhaps, but it serves its limited purpose.
that the psychophysical function is a power The business of scaling and measurement is
law. The argument rests on the fact that the not a primary aim in industry as it is in psy-
apparent bisection point remains approxi- chophysics. Nevertheless, category scaling
mately constant when all the stimuli are is occasionally used in psychophysical experi-
increased or decreased by the same factor. ments, despite the demonstrated inemcacy
For example, Plateau's eight artists each of the procedure. Stevens and Galanter
worked in a different atelier, under a differ- (1957) examined some 70 different category
ent level of illumination, and yet they all scales on a dozen different perceptual con-
produced bisecting grays that, when viewed tinua and found little to justify the use of
in the same setting, were "presques identi- category scaling in quantitative studies.
ques." When that happens, the sensory Their hope was that category scaling would
magnitude must follow either a logarithmic thereafter fall into disuse. They hoped in
law or a power law. If the bisection point vain.
falls at the geometric mean between the end The categories may be designated by a
points used to define the bisected interval, limited set of adjectives, such as large,
then the log function is indicated. If the bi- medium, and small. Or the categories may
section point falls above the geometric mean, be designated by a finite set of numbers,
as it seems to have done in Plateau's experi- such as 1 to 6. Those were the numbers
ment, then a power function is indicated. used for history's first recorded category
And the virtual exponent a increases as a scale, the scale of stellar magnitude, which
function of the distance of the bisecting point dates from about 150 B.C. and which in a
above the geometric mean. Some of those much revised form still serves the astronomer
basic relations are illustrated in Figure 6. (see Stevens, 1960).
On prothetic continua, the category scale

is invariably nonlinear relative to the magni-
tude scale. When plotted against the scale
obtained by magnitude estimation, for ex-
ample, the data from category judgments
produce a curve that is concave downward.
Whenever the subject is asked to categorize,
he is forced to divide the continuum into
parts or segments in order to make it con-
form to the limited, finite set of numbers or
adjectives that he is required to use. In
other words, he is obliged to attend to differ-
ences or distances. Under those circum-
stances, the subject is forced out of ratioing
and into partitioning.
What we learn from category experiments
is that the human being, despite his great O.I 0.8 03 0.4 OS 0.6 0.7 OB 0.9 1.0
versatility, has a limited capacity to effect B i s e c t i o n point: log scale
linear partitions on prothetic continua. He
FIG. 6. Schematic diagram showing a family of
may do quite well, to be sure, if the con- power functions with exponents ranging from .1 to
tinuum happens to be metathetic, but, since 1.2. (As shown by the circles, the position of the
most scaling problems involve prothetic con- bisection point moves upward along the stimulus
tinua, it seems that category and other forms scale (abscissa) as the exponent increases. Bisec-
of partition scaling ought generally to be tion at the geometric mean would correspond to a
logarithmic function, or a vanishingly small ex-
avoided for the purposes of scaling. If, for ponent.)
some reason, an unbiased interval scale is
needed, it can be obtained from a ratio scale, measurement of a threshold becomes a sta-
for the ratio scale contains the interval scale tistical process: On the basis of repeated
(Stevens, 1946). samples, we make a statistical decision re-
The reverse is not possible, however. The garding the location of the boundary. Still,
ratio scale cannot be recovered from the in- the underlying experimental operation for
terval scale when only interval information is determining any kind of threshold always in-
available. volves a procedure of matching either stimu-
Is there a use for category scaling? Al- lus to category or category to stimulus.
though essentially useless for ratio scaling,
category methods play an indispensable role
POWER-GROUP TRANSFORMATIONS
in threshold measurements, where the prob-
lem reduces to the determination of bound- As we have seen, partition scales can often
aries between classes. In fact, all threshold be usefully described by power functions.
determinations involve category procedures, Consequently, the nonlinearity of the parti-
because the problem is to sort stimuli into tion scale can be conveniently described by
classes, for example, those that are detectable the difference between the virtual exponent
and those that are not, those that are dis- of the partition scale and the actual exponent
turbing and those that are not, or those that of the magnitude ratio scale. In other words,
are acceptable and those that are not, and the operation of partitioning causes the ob-
so on. server to behave as though a power-group
Psychophysical thresholds are boundaries transformation had been performed, that is,
between classes. Although the boundary as though an exponent had been altered.
that we call a sensory threshold may be Whenever two power functions differ in ex-
sharp at a given point in time, in a living ponent, they are nonlinearly related. Since
organism the boundary behaves as though it the virtual exponent is always less than the
were jittering about. Consequently, the actual exponent, the curvature of the parti-
436 S. S. STEVENS
tion scale, relative to the magnitude scale, is exponent as that produced by the magnitude
always in the same direction. estimation of differences, as in Figure 4.
More formally, we may express the sub- In category scaling, the difference between
jective magnitude ^ as a function of the the virtual exponent a and the actual ex-
stimulus </> by i/» = k<$>&, where k depends on ponent ft seems also to depend on variability,
units, and ft is the actual exponent of the or on the noise load imposed by the task.
continuum. The actual exponent ft is the Some continua are easier to judge than
one we hope to determine more and more others. For example, the curvature of the
accurately as we learn to control regression category scale was shown to increase and the
effects and other biases. The partition scale virtual exponent to decrease as the con-
value P can be expressed by a similar equa- tinuum was changed from length of lines to
tion, but with an additive constant P0 to take largeness of squares to loudness of tones
care of the arbitrary reference: P + P0 = (Stevens & Guirao, 1963). The variabilities
k<l>tt, where a is the virtual exponent. (standard deviations) with which the ob-
The amount by which the value of a is less servers set values on those three continua
than the value of ft determines the curvature under the procedure of magnitude production
of the partition scale. In some kinds of were: length, 1.0; largeness, 2.1; and loud-
equisection experiments, the curvature is so ness, 4.0 decilogs.
slight that the value of a. has been pushed to It is especially important to note that
within about 10% of the value of ft. At the although partitioning produces a power-
other extreme, in some forms of category group transformation that lowers the effec-
scaling, the value of « has fallen to very low tive value of the exponent, the resulting
values (Marks, 1968). virtual exponent a is always positive. We
It is interesting to note that the category must, of course, exclude from that generali-
scale for stellar magnitude cannot be ex- zation the category scale obtained with highly
pressed in terms of a power function, because skewed distributions of stimuli, because those
the scale is more curved even than a log- abnormalities are essentially artificial and
arithmic function. Otherwise said, the mid- can be remedied by straightforward proce-
point or bisection point of the visual category dures of experimental iteration (see Pollack,
scale of stellar magnitude falls below the geo- 1965). The iterated pure category scale
metric mean of the stimulus scale. The stel- seems always to have a positive exponent.
lar category scale is a rather special case, The importance of a virtual exponent that is
however, because as a stimulus array, the positive and decidedly different from zero lies
distribution of the stars is prodigiously in the evidence it provides that the category
skewed. An attempted representation of the scale is not a logarithmic function of the
stellar judgments by a power function with a magnitude scale. A logarithmic category
negative exponent was given by Marks scale has often been assumed (e.g., Torger-
(1968), and a similar treatment for bisec- son, 1961), but under that assumption, the
tions falling below the geometric mean was virtual exponent would lie near zero.
given by Fagot (1963). Negative expon- The term power group was the name
ents, however, imply inverse or reciprocal proposed for the group of permissible trans-
functions and do not seem to be appropriate formations on what I called a logarithmic
to the present problem. interval scale (Stevens, 1957b). The power
A category production scale for loudness group provides that any scale value x may
gave the virtual exponent .3, which is about be replaced by x' where x' = axb. The
half the value of the actual exponent for transformation preserves the equality of
loudness. The procedure of category pro- ratios, but not of differences. For the visu-
duction serves to diminish the effects of ally minded, it may be helpful to note that a
stimulus spacing and thereby to approximate power-group transformation changes the
the pure form of the category scale. It is curvature, or, in log-log coordinates, the
interesting that the pure category scale slope of a function.
should have approximately the same virtual It is remarkable how widespread are the
instances of power-group transformations. power functions have been recorded by one

Figure 2 shows a dramatic example: The or more investigators in several sense modali-
change from estimation to production pro- ties : vision, hearing, taste, touch, kinesthesis,
duced a new power function with a larger and electrical pulses to the skin. The point
exponent, hence a different slope in log-log to be made here is that there appears to be a
coordinates, which means a different curva- tendency for the neurelectric exponent to de-
ture in linear coordinates. How is such a crease as the recording site becomes more
complex power transformation produced remote from the sense organ. That is to say,
with such apparent precision? The experi- following the power-law transduction per-
menter reverses the procedure, whereupon formed by the sense organ, there may ensue
the observers alter the curvature of their subsequent power-group transformations
response function in just such a way as to higher in the nervous system. But in the
preserve ratios. How is that possible? present state of knowledge, such principles
Many other circumstances produce power remain as vague as the evidence.
transformations. Among them we find the
effect of adaptation on visual brightness. THE PARTITION PARADOX
The exponent rises from .33 to .44 when the
The tendency of observers to make parti-
state of adaptation is changed from dark to tion judgments as though their effective or
about 100 db. above threshold (J. C. Stevens
virtual exponent is lower than the actual ex-
& Stevens, 1963). Even more drastic ponent of the continuum has impressed many
changes in the exponent—alterations by a people as paradoxical. Krantz (1970)
factor of three or more—may take place
wrote:
under inhibition, which includes visual con-
trast and auditory masking (Stevens, 1966). One of the long-standing puzzles is why the scales
Thus it appears that changes in the behavior obtained from category rating differ from those of
magnitude estimation [p. 40].
of sense organs under changing states of
adaptation, and under the inhibition created Why, in other words, does a constant inter-
by glare and masking, can be described by val on a prothetic continuum such as loud-
power-group transformations. Similar trans- ness not appear to remain constant when the
formations occur in the dramatic summation stimulus level is raised ?
of warmth when the irradiated area of the Perhaps no answer will satisfy everyone,
skin is increased. The exponent decreases but we can at least rule out the nature of the
by a factor of about 2 when the area is in- observer's task. We cannot blame partition-
creased by a factor of 10 (J. C. Stevens & ing as such, because partition judgments lead
Marks, 1971). Although I once regarded to linear results on metathetic continua. For
the power group as something of a curiosity example, a S-mel interval sounds the same
and useful only in the discussion of scaling size regardless of its location on the pitch
theory, the piling up of examples has led to continuum. But a 5-sone interval does not
the suggestion that the power group may sound the same size at different locations on
prove to be one of the most common trans- the loudness continuum (see Figure 4).
formations in the biological domain. Hence the problem must have to do with the
A review of research since the 1930s nature of the continuum. Among the fea-
(Stevens, 1970, 1971) turns up many ex- tures that distinguish the two kinds of con-
periments in which electrical recordings of tinua, there is one aspect of prime impor-
neurelectric effects in sensory receptors, tance. On a metathetic continuum, such as
nerve fibers, and neural complexes have ex- pitch, the error distribution, as measured by
hibited a power-law dependence on stimulus the size of the just noticeable difference
intensity. Thus, there exists rather direct (jnd) in subjective units (e.g., mels), re-
evidence that sensory systems are capable of mains the same all along the continuum. In
power-group transformations. Of course, other words, the absolute error is constant.
every placement of an electrode does not On a prothetic continuum it is the relative
produce a nice clean power function, but error that stays constant, so that the jnd for
438 S. S. STEVENS
loudness, measured in sones, grows very nate shows the stimulus change in decibels
large as loudness increases. that corresponds to a change of 1 sone. As
In order to illustrate how unlikely, indeed in the example above, at a level of 40 db. a
impossible, it would be for an observer to change of 1 sone corresponds to 10 db. At
make a correct judgment of a loudness differ- 100 db., tine required change is only .2 db.,
ence regardless of stimulus level, let us con- and it becomes still smaller at higher levels.
sider the following paradigm. We will as- The jnd (Weber fraction) is plotted in Fig-
sume that the loudness exponent is .6 and ure 7 as a horizontal line with an ordinate
that the interval from 40 to 50 db. corre- value of about .5 db. (see Miller, 1947). A
sponds to the subjective interval between a plot similar to Figure 7 can be constructed
loudness of 1 sone and a loudness of 2 sones, for any prothetic continuum. For visually
or a difference of 1 sone. That happens to presented lengths, for example, such a plot
be a clear and obvious difference. Now con- could be made to show how, in visual per-
sider the same 1-sone difference at 100 db. ception, a centimeter added to a centimeter
where the loudness is 64 sones. An increase makes a clearly perceived difference, whereas
from 100 to 100.2 db. makes a 1-sone differ- a centimeter added to a meter becomes only
ence, raising the loudness from 64 to 65 marginally detectable. Under successive
sones. But a change as small as .2 db. is so visual presentation, length behaves like loud-
small that it would be detected less than half ness, and a constant added difference be-
the time. It seems beyond possibility, there- comes lost to view as the stimulus increases.
fore, that a constant difference of 1 sone The smooth continuity of the curve in Fig-
could be judged to be the same size regard- ure 7 suggests that the underlying process
less of where it occurred on the continuum. that forces a given constant difference to be-
That illustrative example may help to sug- come less and less apparent as the level in-
gest why it is that every variety of partition- creases is a process that operates all up and
ing must give a distorted result on a pro- down the prothetic continuum. There are
thetic continuum. A constant difference no discontinuities, no sudden transitions.
transforms itself from obvious to undetect- When the observer attempts a comparison of
able as we go from weak to strong stimuli. differences at any place on the continuum,
Therein lies the essence of the prothetic it is as though his perception undergoes an
principle. asymmetrical distortion; a constant differ-
A graphic illustration of the underlying ence seems larger toward the lower than to-
principle is shown in Figure 7. The ordi- ward the higher part of the continuum.
From that basic asymmetry, it follows that
the operations of partitioning on prothetic
continua will fail to produce a linear, un-
biased interval scale. Where the asymmetry
does not exist, as on a metathetic continuum,
it becomes possible for partitioning to pro-
duce a linear interval scale.
INDIVIDUAL DIFFERENCES
Let us turn at this point to a problem that
70 SO was especially well formulated by Jones and
Sound pressure l e v e l Marcus (1961). Do different individuals
FIG. 7. Showing how the difference in decibels have different operating characteristics in
required to produce a 1-sone loudness difference their sensory systems ? If not, then why do
falls off with increasing sound pressure level. (At the slopes (exponents) differ when observ-
40 db. it requires an added 10 db. to add 1 sone. At ers match numbers to stimuli, as, for exam-
110 db. it requires an added .1 db. to add 1 sone. ple, in Figure 3 ?
The dashed line shows the approximate value of the
jnd, the increase that would be detected about half If we were to take some of the measured
the time.) values literally, it would be incumbent on us
to examine the implications. Consider in a range of exponents from about .3 to 1.5.

Figure 3 the functions for observers EG The distributions of the ratio settings tended
(exponent .4) and PK (exponent 1.1). to be roughly log normal, however, and there
They both happen to be "normal" observers, appeared to be no reason not to average the
having similar audiometric thresholds. If, data, which is equivalent to averaging in-
as seems reasonable, the two observers ex- dividual exponents.
perience the same loudness when a tone is Consider this question: What would hap-
near threshold, what about a tone 100 db. pen if an experimenter were to report that
above threshold? The two exponents sug- each of 22 observers in a group made an
gest that the loudness at 100 db. would be identical setting when asked to produce a
about 3,000 times greater for PK than EG. two-to-one loudness ratio ? It is an interest-
Yet both observers react very similarly to ing question whether the scientific commu-
acoustic stimuli. For example, both observ- nity would regard such a result as a new
ers report that 100 db. sounds very loud; marvel, or whether scientists would simply
they both call 120 db. slightly painful; and reject it as an implausible outcome. We ex-
they both jump when it is first turned on. pect variability and we usually find it. The
Both observers set the intensity control of substantive question concerns whether and
the radio to roughly the same level for com- how to average the data.
fortable listening. Both carry on conversa- Although the operating characteristics of
tions in about the same level of voice. All normal sensory systems may have the same
those bits of evidence make it hard to credit or closely similar exponents, there are cir-
a difference in exponent that entails a 3,000- cumstances in which an abnormal exponent
fold difference in loudness at 100 db. can be demonstrated and in which the ab-
Another consideration is this. The value normality calls for careful measurement.
of an exponent reflects the curvature in the The otologist, contemplating middle-ear sur-
operating characteristic of the sensory sys- gery, needs to know whether so-called re-
tem. As a function of sound pressure, the cruitment, with its attendant large exponent,
exponent .4 suggests that the loudness func- characterizes the lower part of the patient's
tion for EG is a decelerating function, loudness function. If the exponent is ab-
sharply concave downward. The exponent normally large, middle-ear surgery is con-
1.1 suggests that the function for PK is con- traindicated. But if the exponent is normal,
cave upward—an accelerating function of the otologist may feel free to try to correct
sound pressure. In order to achieve such a middle-ear difficulty. (For a model de-
an unlikely difference in curvature, nature picting the manner in which Meniere's dis-
would have had to produce two radically dif- ease may produce power-group transforma-
ferent kinds of operating characteristics in tions on the loudness function, see Stevens &
the two auditory systems. Guirao, 1967.)
An alternative hypothesis is that all hu- Except for clinical cases, it seems fair to
man auditory systems operate on much the say that seldom, if ever, have investigators
same design, with very nearly the same ex- undertaken the serious effort that is needed
ponents, but that there are wide individual to establish the existence of individual dif-
differences in what observers take to be the ferences in sensory power functions. Mostly,
numerical value of a loudness ratio. We we are shown the variability that happens to
encounter that kind of individual difference be found in a particular matching response,
especially clearly when an observer is asked usually magnitude estimation. The obvious
to adjust one sound to make it appear half next step, usually not undertaken, would be
as loud, or twice as loud, as another sound. to try magnitude production. I have found
In an experiment with 22 observers, the that many individual differences vanish as
scatter of eight distributions of settings cov- soon as the matching task is inverted. For
ered a median range of 14 db. (For plots example, by magnitude estimation, the low-
of the distributions, see Stevens, 19S7a.) est loudness exponent in a group was .4, but
That range of settings would correspond to by magnitude production, that particular ob-
440 S. S. STEVENS
server produced the exponent .9; the geo- by being averaged out. An analogous strat-
metric mean of the two exponents is .6 (see egy of error cancellation by averaging finds
graphs in Stevens & Greenbaum, 1966). usefulness in psychophysics and in all the
Five observers in those same experiments rest of science.
(Stevens & Guirao, 1962) made both magni- In the long run, since scientists tend to be-
tude estimations and magnitude productions lieve only those results that they can repro-
of the loudness of noise. The ratio of the duce, there appears to be no better option
largest to the smallest exponent for estima- than to await the outcome of replications. It
tion was 1.42; for production, it was 1.27. is probably fair to say that statistical tests of
When the geometric means of the expon- significance, as they are so often miscalled,
ents for estimation and production were ex- have never convinced a scientist of anything.
amined, the range ratio fell to 1.14. In other By contrast, a tabulation of 178 determina-
words, the individual differences were less tions of the loudness exponent, based on 25
pronounced when both estimation and pro- years of accumulated results from several
duction were used in a balanced design. different laboratories, produced a median re-
The balancing of estimation by production sult .6, which became the exponent recom-
may be good for a start, but if we are seri- mended by the International Standards Or-
ously interested in the power function for a ganization (see Stevens, 1955a). Since
particular individual, we will not stop with then, a further accumulation of experimental
magnitude estimation and production. We determinations has begun to fix the second
will want to know the results from a bal- decimal place, and it now appears that the
anced array of additional cross-modality value 2/3 may be more representative (Stev-
matching tasks, the more the better. ens, in press). The value 2/3 happened to
correspond to the modal value of the 1955
VARIABILITY AND AVERAGING distribution, but at that time, the median
seemed a better choice than the mode.
Criticism of the power law has sometimes How to average data presents serious and
centered on one or another aspect of varia- interesting questions. The median is per-
bility. It would be good, of course, if varia- haps the single most unbiased measure of
bility could be reduced, so that the psycho- location, and it has often been used in psy-
physical functions could be determined with chophysics. A popular rule is: when in
higher precision. But empirical functions doubt, use the median. On the other hand,
always suffer from variability, and the cen- a more efficient average is often wanted, and
tral question is not so much whether a mea- the choice of an efficient measure can usually
surement is variable, or whether subjects be made to rest on the form of the distribu-
disagree, but whether averaging is appro- tion. Thus, two different averages have
priate. If the data can be appropriately proved appropriate in psychophysical scal-
averaged, it does not matter how widely the ing, each under a particular circumstance.
variability may range, provided the number In an equisection experiment, 45 subjects
of independent measurements can be in- divided a 40-db. segment of the loudness con-
creased. In principle, the standard error can tinuum into four equal-appearing intervals.
then be brought down to any desired level. Because the decibel measures of the subjects'
In electrophysiology, for example, mira- settings gave skewed distributions, it did not
cles of averaging are performed routinely by seem proper to average the decibel values,
computers programmed to dig a particular which would have been equivalent to com-
waveform out of the myriad variations in the puting geometric means. When the decibel
ongoing neural activity of the brain. A re- measures were converted into sones (a linear
peated click delivered to the ear can then be loudness value), the settings showed the de-
seen as an evoked potential at the scalp. sired symmetries. It was concluded that
The response to the click emerges remark- averaging should be done by computing the
ably clear and unencumbered by the noise of arithmetic means of the loudness values, and
the brain, for the noise has been suppressed an iteration procedure for determining those
values was outlined (Stevens, 1955b). The Stimulus order 5 8 2 6

arithmetic mean is also an appropriate aver- SPL in 4B 90 95 100 109
age for other kinds of partition scales, such
as the category scale.
In experiments involving magnitude esti-
mation and other forms of cross-modality
matching, it is the geometric mean, not the
arithmetic mean, that appears to be the ap-
propriate average. Against a linear scale,
the response distributions are skewed, as in-
deed they must be when error is proportional
to magnitude. When error grows in pro-
portion to magnitude, so that the relative
error stays constant, a logarithmic trans- 0 02 0.4 06 0,8 1 0 12 1.4 1.6 1.8 2.0 2.2
formation tends to undo the skewness. The Log magnitude estimation
applicable principle states that when error is FIG. 8. Cumulative frequency distributions of
relative, the error distribution is log normal. magnitude estimations. (Each of 70 subjects, mak-
The log-normal model for magnitude esti- ing their first judgments ever, assigned whatever
mation was tested by J. C. Stevens who number seemed appropriate to eight levels of white
plotted the results obtained when 70 naive noise presented in the order shown. The data of
three subjects who used negative numbers or zeros
observers estimated the loudness of a white were not tabulated. In these coordinates, a straight
noise presented by a loudspeaker in a class- line signifies a log-normal distribution. Data from
room (J. C. Stevens & Tulving, 1957). In J. C. Stevens & Tulving, 1957.)
the first of two experiments, the observer
chose his own modulus by assigning to the naive listeners. Since it was a classroom
first stimulus (85 db.) whatever number experiment, the order of the stimuli could
seemed appropriate. In the second experi- not be made different for the different listen-
ment, the stimuli were presented in pairs, a ers. Consequently, the order of the stimuli
standard at 85 db. called 10 followed by a is reflected in the slopes of the cumulative
variable. When the cumulative frequencies frequency lines, or, in other words, in the
for this second experiment were plotted on standard deviations. The first stimulus has
probability paper against the logarithm of the smallest standard deviation (1.0 deci-
the magnitude estimations, the result was a log), the second stimulus the next smallest
family of straight lines. In other words, the (1.4 decilog), and so forth. If the stimulus
distributions were log normal. order had been made different for each ob-
Some five years later J. C. Stevens and server, the lines in Figure 8 would be more
Miguelina Guirao applied the modulus equal- nearly parallel. But a tendency has been ob-
ization procedure to the data of the first ex- served in numerous experiments for the vari-
periment, the one in which each of the 70 ability to increase at the low end of the scale,
subjects had chosen his own modulus. (The and to a lesser extent, at the high end. That
deviation of each subject's scores from the same tendency is apparent in Figure 8.
average function was minimized by the pro- Hence, it can be seen that the effect due to
cedure for modulus equalization described stimulus order cuts across the normal tend-
above.) The cumulative frequencies of the ency for the variability to be slightly lower in
judgments subjected to modulus equalization the middle range.
are shown in Figure 8. Again the straight The lines in Figure 8 make it clear that
lines demonstrate that the distributions are the geometric mean is an appropriate aver-
approximately log normal. age for the data. In this instance, the geo-
Other features of the results in Figure 8 metric means determine a power function
are also of interest, especially in view of the with the exponent .55.
fact that those were the first magnitude esti- A similar treatment was applied to the
mations ever made by that large group of data from the original "no standard" experi-
442 S. S. STEVENS
ment carried out in 1954. The elimination ard deviations ranged from 1.7 to 5.4 deci-
of the standard was an important procedural logs, with a median value of 2.2 decilogs.
change suggested by Geraldine Stevens. A The variability in the results of cross-
group of 32 observers each made two judg- modality matching for a group of observers
ments of each level of a 1000-hertz (Hz.) can be divided into three main components,
tone. (The details of the procedure are given namely, the variability attributable to differ-
in Stevens, 1956.) The level of the first tone ences from observer to observer in the effec-
was varied from one observer to another and tive modulus (intercept), the effective ex-
was assigned whatever number the observer ponent (slope), and the residual scatter in
thought appropriate. The cumulative fre- each observer's matches. It is sometimes
quencies were nicely log normal, both before useful to partial out those sources of varia-
and after modulus equalization. The effect bility (see Stevens & Stevens, 1960).
of the modulus equalization was to reduce
the standard deviations by a large factor, as RANGE EFFECTS
shown in Table 1.
It is significant to note that the standard Much has been written about the effects of
deviations were larger when the stimulus to stimulus range on the exponents of the sen-
be judged was 1000 Hz. (Table 1) than sory power functions. The range (loga-
when it was a white noise (Figure 8). In rithmic spread) of the stimuli used in cross-
numerous experiments, it has been found modality matches may affect the exponent,
that a noise is easier to judge than a tone. but the experimenter can design tactics to
Still more difficult for most observers are offset the biasing effects of range, if he so
magnitude estimations of apparent pitch. chooses. Analogous options face the experi-
Each of 20 subjects made two judgments of menter with regard to other distorting fac-
12 frequencies between 100 and 7500 Hz., tors, and the same principle applies to mea-
with no designated standard (Stevens & surements in physics as well as psychophys-
Galanter, 1957). With such a small number ics. A scientific measurement of serious
of scores, the cumulative frequency plots consequence can never be based on one
showed much scatter, but the overall pic- experiment, because multiple experiments
ture was log normal. Corrected by modulus are required to detect and correct the sys-
tematic errors. Multiple experiments are
equalization, pitch judgments yielded stand-
the rule in physics; they ought also to be
ard deviations that were not very different the rule in psychophysics.
from the corrected values for the loudness There is a negative correlation between
judgments in Table 1. For pitch, the stand- the measured exponents and the ranges of
the stimuli that were used in some of the
TABLE 1 experiments by which the exponents were
STANDARD DEVIATIONS IN DECILOGS OF MAGNITUDE
determined. For example, a range as great
ESTIMATIONS DETERMINED GRAPHICALLY as 90 db. has been used for loudness, which
FROM CUMULATIVE DISTRIBUTIONS has a low exponent, compared with a range
of about 10 db. for electric current through
Stimulus SD SD the fingers, which has a high exponent. The
unconnected corrected
negative correlation between range and ex-
110 5.6 3.0 ponent has led Poulton (1968) to say
100 5.4 2.3
90 4.7 1.8 that in designing the experiments to measure the
80 4.7 1.7 exponent, the experimenters did not adequately
70 4.6 1.7 compensate for the effects of the different physical
60 5.0 1.9 ranges . . . [p. 5].
50 5.5 3.0
40 5.8 3.2 To be sure, the experimenter could choose a
10-db. range for the study of loudness; but a
Note.—Thirty-two subjects made two judgments of each
stimulus (1000 Hz.), When corrected by modulus equalization,
90-db. range of electric current through the
the variability fell by a factor of approximately 2. fingers would prove insupportable. It seems
that stimulus ranges are to a very large exponent. Some of the lowest measured ex-
tent selected by experimenters because na- ponents apply to odor. Benzaldehyde (syn-
ture's exponents are what they are, not the thetic almond) gave the exponent .2 (Stev-
other way around. ens, 1957b). That was probably the first ol-
In suggesting that the experimenter should factory exponent ever determined. A similar
compensate for the effects of the different low value has since been found by Berglund,
physical ranges, Poulton directed attention Berglund, Engen, and Ekman (1971). The
to the wrong side of the equation. It is not stimulus range for benzaldehyde is relatively
the physical ranges that need compensation; short, at least as compared to loudness or
rather, the experimenter should try to en- brightness. From the point of view of the
sure that the subjective ranges are as com- observer, the subjective range of the odor
parable as possible. Stimulus measures have seems extremely short compared to the
much arbitrariness about them: measures of enormous subjective ranges that can be pro-
sound pressure give one loudness exponent; duced in loudness or brightness.
measures of sound power give an exponent
EFFECTS OF REPETITION
that is half as large. For apparent size, the
measured diameter of circles gives one ex- Although range effects may be present in
ponent, the measured area of circles gives any given experiment, the degree to which
another, and so on. they affect the outcome may sometimes
In comparing the exponents of different be altered by repeated presentation of the
continua, the experimenter would like to be stimuli.
able to select stimuli—regardless of how they An experiment showing how repeated pre-
happen to be measured—so that they would sentations of a very short range of luminous
produce a constant subjective range. If he targets can cause the measured brightness
could do that, then the correlation would exponent to increase on successive presenta-
uniquely fix the relative values of all the ex- tions was carried out in 1960 by A. W. F.
ponents. Of course, if the experimenter Huggins (reported in Stevens & Stevens,
knew in advance how to choose the stimulus 1960). The results are shown in Figure 9.
ranges that would produce the perfect corre- The stimuli, a series of Munsell grays rang-
lation, he would not need to run the experi- ing from black to white, were viewed under
ment. In effect, then, much of the extensive so-called reduction conditions, which made
work with cross-modality comparisons can them appear as luminous targets, not as sur-
be regarded as an effort by trial and error faces. The stimulus range covered only 16
to determine what stimulus ranges would be db., because that is all there is between a
needed to provide a constant subjective range black paper and a white one. The stimuli
on all continua and thereby make it possible were also presented under two levels of il-
to produce a perfect negative correlation be- lumination, which extended the range and
tween logarithmic range and exponent. which gave the filled points in Figure 9.
The correlation between stimulus range The exponent for the extended range is .35.
and exponent has been reported as high as The magnitude estimations for the shorter
— .94 by Teghtsoonian (1971), who proposed range on the first presentation follow very
closely the lowest five points on the extended
that a single scale of sensory magnitude serves a range, but later presentations give succes-
wide variety of perceptual continua, and that vari- sively steeper slopes (higher exponents).
ation in power law exponents is primarily due to
variation in dynamic ranges [of stimuli] [p. 71]. Although the limiting of the experimental
procedure to a single presentation of each
That is an interesting hypothesis, even stimulus may attenuate some of the effects
though a test of it would require that we of range, residual effects on the exponent
learn how to determine dynamic range, may still remain. Under some circum-
which may prove to be an elusive variable. stances, the residual effects can be balanced
It is important to note that a short range out in the experimental design, as is shown
does not necessarily produce a large ex- below.
444 S. S. STEVENS
100 observers would necessarily respond with

adjusted ranges that average larger than the
criterion range. That value of the adjusted
50 range would then determine an exponent
greater than 1.0. On the other hand, if the
30 criterion range was set at a very large value,
the observers would tend to match it with
.2 20 shorter ranges, determining thereby an ex-
o ponent smaller than 1.0. Between those two
extremes, the exponent would decrease
* ,0 monotonically as the range increased.
Since the two continua A and B can be
interchanged in their roles of criterion and
adjusted variable, in a balanced experiment
two paths would be traced by the exponent
as the criterion range was varied from very
small to very large. Those two paths would
cross, and the crossing point would presuma-
bly determine the unbiased exponent—un-
biased by the effect of range, but not neces-
55 60 65 70 75 80 85 90 sarily free of other possible biases.
Luminance in decibels A concrete example of the two paths fol-
lowed by the exponent when range is varied
FIG. 9. Showing how Munsell grays, extending in a fairly well-balanced experiment can be
from black to white, are judged when they are pre-
sented entirely alone with no other visible lumi-
constructed from the data of Stevens and
nance in the field of view. (The luminance then Poulton (1956). The loudness function for
covers a range of about 16 db., which seems sub- a 1000-Hz. tone was studied by allowing
jectively rather short. On the three successive pre- groups of 8 to 11 unpracticed observers to
sentations of the set of stimuli, the exponent be- make only a single judgment, either a ratio
came larger (unfilled symbols). The filled symbols
show the results for an extended luminance range. estimation or a ratio production. A stand-
The exponent is .35, which is much lower than the ard stimulus was sounded first and called
value 1.2 obtained when the Munsell grays are 100, and the observer expressed the apparent
placed one by one on a table in front of the ob- ratio by assigning a number to a second
server.) stimulus at a lower level. The exponents
corresponding to the median estimations of
COUNTERBALANCING FOR RANGE each group are shown by the triangles in
Figure 10. Each triangle represents a dif-
The general problem of the interaction of ferent group of listeners.
range and exponent may be thought of in For the ratio productions, the observers
terms of the matching of two continua, A adjusted a sone potentiometer to produce a
and B, each of which may serve in turn as prescribed fractional loudness relative to a
the adjusted and as the criterion continuum. standard. The exponent corresponding to
What does the observer tend to do to the the ratio productions (sone average) for
adjusted stimulus as a function of the range each group of subjects is shown by a circle
(logarithmic difference) between two cri- in Figure 10. Again each circle represents
terion stimuli set by the experimenter ? For a different group of listeners. The data
purposes of discussion let us assume that the show an approximate symmetry and suggest
true exponent determined by the slope A/B that for those naive observers making their
in log-log coordinates is 1.0. first loudness judgments, the exponent lies
If the experimenter were to set the cri- between .6 and .7. The range effect emerges
terion range to a very small value, say, to a as a dramatic but orderly variable, and since
small fraction of a decilog, on the average the it shows an approximate symmetry, it can in
Ratio to be produced changed the upper scale and the location of

3/4 1/2 1/4 the circles accordingly. The resulting cross-
i.o T-
over then fell at .65—a value smaller than
0.9
my assumed exponent. Thus the two as-
08 sumptions, one too high and one too low,
•»• 07
had succeeded in bracketing the exponent
c between .64 and .65, but closer to .65. With
v 06
c fallible empirical data, the exact value of the
o 05 exponent cannot be taken too literally, but it
o.
* 04 is nevertheless interesting that the crossover
111
03
value accords approximately with the con-
sensus of other measurements.
02
"0 10 20 30 40 dB 50 The search for other instances in which a
Ratio to be estimated pure range effect could be studied has been
FIG. 10. The range effect in a partially balanced only partially successful. Experiments with
experiment. (Each point represents a separate the required balanced design seem to be rare,
group of subjects who estimated a loudness ratio but we can compare two separate experi-
(triangles) or produced a loudness ratio (circles). ments that add up to a partly balanced de-
When the range effect produces symmetrical func-
tions, as is approximately true here, the exponent sign. In Figure 11, the triangles show the
may be uniquely determined. It is the value that exponents corresponding to the median ratio
makes the exponent corresponding to the crossover estimations made by groups of about 30 sub-
point equal to the exponent implied by the relation
of the scales at the top and bottom of the figure. jects (Poulton, 1969). Each group judged
The exponent so determined is .65. Data from one noise level set at 5, 20, or 35 db. above a
Stevens & Poulton, 1956.) standard noise (600 to 1200 Hz.) at 65 db.
The circles in Figure 11 represent the ex-
principle lead to a unique exponent de- ponents determined from the ratios produced
termined by the crossing point in Figure 10.
The actual crossing point depends, of
course, on the relation between the two scales Ratio to be produced
(top and bottom in Figure 10) against which 2 4 10
the two sets of data are plotted. But the re-
lation between the values on the two scales
(both logarithmic) also expresses an ex-
ponent. The problem, then, is to adjust the
relation between the two scales (ratios to be
produced and ratios to be estimated) so that
the exponent determined by the scale relation
coincides with the exponent determined by
the crossing point. The adjustment can be
carried out by iteration.
In plotting Figure 10, I first assumed that
the exponent was .60. Accordingly, the 0.4
0 10 20 30 40 dB 50
ratio 1/2 at the top of the graph was set
Ratio to be estimated
directly above 10 db. at the bottom of the
graph. The other ratios were then set ac- FIG. 11. Range effect shown in two noncompara-
cording to a logarithmic spacing. When the ble experiments. (Separate groups of subjects esti-
data were plotted, the crossover was found mated loudness ratios (triangles) of a band of
noise, 600-1200 Hz. (Poulton, 1969). A single
to correspond to .64 on the ordinate—a value group of subjects produced fractional and multiple
that was larger than my assumed exponent. ratios (circles) of a 40-tone complex (Geiger &
I next assumed a larger exponent, .67, and Firestone, 1933).)
446 S. S. STEVENS
by 31 listeners who made multiple and frac- THE NOMOTHETIC IMPERATIVE
tional settings of noise consisting of a 40-
tone complex (Geiger & Firestone, 1933). The scientist's contest with nature has
Both the multiple and the fractional produc- prospered to the degree that simplicities and
tions were averaged to determine the ex- uniformities have been detected amid the
ponents. The same group of listeners heard complexities that afflict observation and ex-
all the stimuli, and the order of the ratio periment. The simple invariances have often
productions was 1/2, 1/4, 1/10, . . . 2, 4, proved hard to find, however, because no ex-
and 10. periment can be performed without its "con-
In order to plot the circles in Figure 11, text," and no measurement can be made
the exponent was assumed to be .67, and without error. In psychophysics, each ex-
the ratio 2 on the upper scale was set di- perimenter records results that disagree to
rectly above 9 db. on the lower scale. The some extent with those of his colleagues,
crossover point then fell at .77. If the ex- and a penumbra of uncertainty surrounds
ponent is assumed to be .8, and if the scale even our best determinations. Consequently,
values and the circles are moved accordingly, there is room for many hypotheses and for
we find that the crossover point becomes many views regarding the structure of the
about .78. Thus the two assumed exponents psychophysical domain.
have bracketed the crossover exponents. First there are those whose working hy-
Consequently, the exponents that would be pothesis states that there exist laws to be
determined by the data in Figure 11, if they discovered. Heeding the nomothetic im-
were homogeneous and suitable for such a perative, those investigators refuse to be put
purpose, would lie between .77 and .78. off by the apparent chaos in the organism's
That value is on the high side, even though reactions to stimuli, and they try to order,
the measured exponents for noise are often classify, and systematize the behavioral facts.
larger than those for tones. The mean ex- Some people seem willing to gamble that the
ponent for white noise in an earlier compila- sense organs operate in beautifully simple
tion was approximately .72 (Stevens, 19SSa). ways and that what we take to be complexity
The groups of subjects who gave the re- lies more in our inept descriptions than in
sults that determined the triangles in Figure nature's actual comportment. As regards
11 subsequently judged all the other stimuli. the particular question of reactions to stimu-
The pooled results determine a power func- lus intensity, the nomothetic outlook assumes
tion with an exponent of about .73, provided that there exists an orderly input-output
no attempt is made to force the best fitting function, a simple law of some sort—how-
line to pass through the standard stimulus. ever difficult it might be to pin it down in its
Although experimenters have sometimes as- exact form. Perhaps it was that same nomo-
sumed that the function must pass through thetic outlook that drove Kepler to search
the standard, there is in fact no such re- for simple invariant laws in the baffling paths
quirement. The observer does not respond of the wandering planets, which were thought
to any stimulus, whether standard or vari- in early times to crisscross the skies impelled
able, with zero error. by their own volition.
The foregoing examples of the sources of The nomothetic viewpoint is consonant
error and distortion that may plague a psy- with an objective, operational approach, but
chophysical measurement are more illustra- it does not necessarily entail a particular
tive than exhaustive. As has often been philosophy of reality. In the fitting of sche-
said, there are many ways to perform a bad matics to empirical fact—what I have called
experiment. Not even the expert in statis- the schemapiric endeavor—there is no neces-
tical design can tell exactly how to perform sity that we take a stand regarding any ulti-
a good one. In principle, however, we can mate concern. We can simply note, for ex-
study each experimental ailment and utilize ample, that when the intensity of the visual
the rules of its behavior in order better to stimulus increases, the observer's response
diagnose the nomothetic substrate. changes in an orderly fashion. If he is in-
strutted to squeeze a hand dynamometer to purpose of implementing a schematic struc-

match the apparent intensity of what he sees, ture which may be related by operational
then the force of his squeeze increases as a rules to an empirical structure (see Stevens,
power function of the physical luminance (J. 1968b).
C. Stevens, Mack, & Stevens, 1960). That As an experimenter, I feel free to use
is the first-order fact. But what shall we say terms like sensation and subjective, because
about it? We have obviously measured they can be defined operationally in the con-
something, but have we measured sensation ? text of psychophysical research. For some
As soon as that word appears, a discordant writers, however, the use of such words puts
chatter sets in, because many authors con- the user in the camp of Subjectivism, as op-
tend that sensation is not a thing that can posed to Behaviorism (e.g., see Baird, 1970).
be measured. I have long thought of myself as a behavior-
If I thought it would help, I should hap- ist, but it has not seemed defensible to assert
pily give up the word sensation in favor of that verbal taboos provide a tool for ad-
some other term, such as behavioral response vancing science, as some behaviorists have
or apparent effect, but the philosopher would seemed to believe. A viable injunction is
shortly discover that my new term is only a this: Use any words you care to, but let it
euphemism for what I regard as a straight- be plain what operations lie behind your ver-
forward construct, a construct that can per- bal forms. Otherwise said, construct the
haps best be communicated to other people schema as you will, but tell us precisely how
by my labeling it sensation. the empirics articulate with the schematic
A prime source of confusion in this seman- terms in order to produce the schemapiric
tic issue rests with the fact that each of us substance.
experiences sensations, and to each of us our Such protestations, I realize, will avail
sensations seem personal and private and in- nothing with those who want to maintain a
accessible to measurement. As scientists, we philosophical distinction between certain
should try to ignore that fact. To help free traditional views and to preserve a dualism
us of narcissistic introspection, there is the that to an operationist is devoid of meaning.
example of Plateau, the blind physicist, who Since I do not believe in the usefulness of
gave us our first interval scale for the light- the distinction, I am happy to read that
ness of grays. He himself did not need to It is impossible to locate Stevens' view precisely,
see the grays that the eight artists painted, since he constantly shifts back and forth between a
for he could record the reports of other ob- behaviorist and some other way of treating the
servers. Plateau's achievement drives home question . . . [Savage, 1970, p. 390].
a crucial point: the blind could develop the More damning than that, however, the whole
psychophysics of vision; the deaf could de- psychophysical enterprise is said by Savage
velop psychoacoustics. to be wrongly conceived, so that
Sensation then is a construct, a name However we reconstrue Stevens' law, it cannot be
given to a constellation of behaviors. The construed as one relating sensations to stimuli,
justification for saying that sensation is sub- since the former are incapable of measurement [p.
jective is that a human subject exhibits the 541].
behaviors. The heating coil in an electric More than half a thousand pages, it should
stove also exhibits interesting behaviors, but be said, are devoted by Savage to the thesis
quite arbitrarily we refrain from using the that we cannot measure what psychophysi-
term subjective for those behaviors or for cists seem so delightedly to be measuring.
the constructs built on them. The simple, Philosophical problems must not detain us,
operational dichotomy into subjective (mean- however, for philosophical issues are eternal
ing people) and objective (meaning not peo- and do not go away. More tractable, per-
ple) seems eminently convenient, but words haps, are some of the specific issues that
and slogans sometimes cause clash and con- strike directly at the nomothetic imperative.
flict. In the schemapiric view of science, A cluster of questions has concerned the
words and symbols serve only the neutral problem of picking the best sensory scale.
448 S. S. STEVENS
When three different scales have been created on our expectations regarding the outcome
on the same continuum by three separate of experiments, so that a discordant result
sets of operations—jnd's, category estima- becomes suspect until verified by adequate
tion, and magnitude matching—on what replication. In other words, the sensory
grounds do we presume to ascribe to one of power law takes precedence over the results
those scales a superior position? Helson of any particular experiment, just as the
(1964) seemed to deny that we can make power law governing gravitational attraction
any such judgment, for he said that usually remains unquestioned despite a par-
ticular experimenter's inability to confirm it
no one scale, however carefully established, can be
considered better than other scales obtained under by dropping objects in a laboratory. Both
different conditions of judging [p. 179]. kinds of power laws make possible many
kinds of predictions; yet they both can be
Other authors have also lamented the ab- shown to fail to some degree in particular
sence of criteria for determining the validity contextual circumstances.
of a choice of one kind of scale over another. The other view holds that the nomothetic
Validity is indeed the issue here. Which imperative has no compelling jurisdiction in
scale best measures what it is that we want psychophysics, because departures from the
to measure? Since it is that kind of ques- power law are too numerous to be ignored.
tion, the answer becomes a matter of opinion It is claimed that the ways in which the
—a value judgment. It appears that all human observer responds to stimulus in-
problems related to validity must seek their tensity depend on prior learning, adaptation,
ultimate solution in the pragmatic domain. range of stimuli, nature of the matching task,
If a scale does the job we want done, we and so on; and that until those many factors
usually accept it. and contextual influences can be discovered,
But opinions differ about problems and explored, and understood, it is premature to
solutions. How else can we understand the speak of a psychophysical law.
decision of an experimenter to limit the ob- The two views sketched above may prove
server's responses to a finite set of numbers, more extreme than the attitude of any par-
such as 1 to 7 or 1 to 20? That curious ticular scientist, but they illustrate two poles
maneuver of constraining the observer's re- of opinion. Perhaps the least nomothetic
sponses is a tactic that seems somewhat view yet expressed was that of Poulton
mulish to those who have allowed the ob- (1968) who concluded his review of "the
server a full range of responses and have new psychophysics" by saying, "The mechan-
witnessed the greater usefulness of the re- ism of response learning and response bias
sulting ratio scales. It may be true that in must be included in any adequate descrip-
the long run, superior procedures tend to tion. To this reviewer," he added, "they
replace inferior procedures, but in almost present the more interesting and challenging
two decades of practice, magnitude estima- problems." In my own view, the problems
tion has not displaced category estimation— of response bias rate no better than a nui-
nor does it seem likely to do so any time sance, an interesting nuisance, perhaps, as
soon. some of the foregoing sections have tried to
Even among those who have given up show, but nevertheless a diversion from the
category scaling in favor of ratio or magni- basic business of sorting out the fundamental
tude matching, there remains a matter of principles. Fortunately for science, how-
taste and opinion that divides the practition- ever, its practitioners are motivated by a
ers. One view holds that the construct we diversity of values and interests. It would
call sensory magnitude follows simple laws, stifle the enterprise if we all tried to crowd
and in particular that under proper circum- in on one single problem.
stances, the sensation magnitude experienced
by the typical (median) observer grows as a REFERENCES
power function of the stimulus magnitude. BAIRD, J. C. Psychophysical analysis of visual
The power law is thought to set constraints space. Oxford: Pergamon Press, 1970.
BECK, J., & SHAW, W. A. Ratio-estimations of Bulletin de I'Academic Royale de Belgique, 1872,
loudness-intervals. American Journal of Psy- 33(Ser. 2), 376-388.
chology, 1967, 80, 59-65. POLLACK, I. Iterative techniques for unbiased rat-
BEROLUND, B., BERGLUND, U., ENGEN, T., & EK- ing scales. Quarterly Journal of Experimental
MAN, G. Individual psychophysical functions for Psychology, 1965, 17, 139-148.
twenty-eight odorants. Perception and Psycho- POULTON, E. C. The new psychophysics: Six
physics, 1971, 9, 379-384. models for magnitude estimation. Psychological
DAWSON, W. E. An experimental analysis of Bulletin, 1968, 69, 1-19.
judgments of sensory difference. Unpublished POULTON, E. C. Choice of first variable for single
doctoral dissertation, Harvard University, 1968. and repeated multiple estimates of loudness.
EKMAN, G. Two generalized ratio scaling methods. Journal of Experimental Psychology, 1969, 80,
Journal of Psychology, 1958, 45, 287-29S. 249-253.
FAGOT, R. A. On the psychophysical law and esti- SAVAGE, C. W. The measurement of sensation.
mation procedures in psychophysical scaling. Berkeley: University of California Press, 1970.
Psychometrika, 1963, 28, 145-160. STEVENS, J. C., & GUIRAO, M. Individual loudness
GARNER, W. R. A technique and a scale for loud- functions. Journal of the Acoustical Society of
ness measurement Journal of the Acoustical So- America, 1964, 36, 2210-2213.
ciety of America, 1954, 26, 73-88. STEVENS, J. C., MACK, J. D., & STEVENS, S. S.
GEIGER, P. H., & FIRESTONE, F. A. The estimation Growth of sensation on seven continua as mea-
of fractional loudness. Journal of the Acoustical sured by force of handgrip. Journal of Experi-
Society of America, 1933, 5, 25-30. mental Psychology, 1960, 59, 60-67.
HAM, L. B., & PARKINSON, J. S. Loudness and STEVENS, J. C., & MARKS, L. E. Spatial summa-
intensity relations. Journal of the Acoustical tion and the dynamics of warmth sensation.
Society of America, 1932, 3, 511-534. Perception and Psychophysics, 1971, 9, 391-398.
HELLMAN, R. P., & ZWISLOCKI, J. J. Loudness de- STEVENS, J. C., & STEVENS, S. S. Brightness func-
terminations at low sound frequencies. Journal tion: Effects of adaptation. Journal of the Op-
of the Acoustical Society of America, 1968, 43, tical Society of America, 1963, 53, 375-385.
60-64. STEVENS, J. C., & TULVING, E. Estimations of loud-
HELSON, H. Adaptation-level theory: An experi- ness by a group of untrained observers. Ameri-
mental and systematic approach to behavior. New can Journal of Psychology, 1957, 70, 600-605.
York: Harper & Row, 1964. STEVENS, S. S. On the theory of scales of mea-
INDOW, T., & STEVENS, S. S. Scaling of saturation surement. Science, 1946, 103, 677-680.
and hue. Perception and Psycho physics, 1966, 1, STEVENS, S. S. On the brightness of lights and
253-272. the loudness of sounds. Science, 1953, 118, 576.
JONES, F. N., & MARCUS, M. J. The subject effects (Abstract)
in judgments of subjective magnitude. Journal of STEVENS, S. S. Biological transducers. Conven-
Experimental Psychology, 1961, 61, 40-44. tion Record, Institiite of Radio Engineers, 1954,
KRANTZ, D. H. A theory of magnitude estimation Part 9, 27-33.
and cross-modality matching. (MMPP 70-6) STEVENS, S. S. The measurement of loudness.
Ann Arbor: Michigan Mathematical Psychology Journal of the Acoustical Society of America,
Program, July 1970. 1955, 27, 815-820. (a)
KRUEGER, L. E. Apparent combined length of two- STEVENS, S. S. On the averaging of data. Science,
line and four-line sets. Perception and Psycho- 1955, 121, 113-116. (b)
physics, 1970, 8, 210-214. STEVENS, S. S. The direct estimation of sensory
MARKS, L. E. Stimulus-range, number of cate- magnitudes—loudness. American Journal of
gories, and form of the category-scale. Ameri- Psychology. 1956, 69, 1-25.
can Journal of Psychology, 1968, 81, 467-479. STEVENS, S. S. Concerning the form of the loud-
MERKEL, J. Die Abhangigkeit zwischen Reiz und ness function. Journal of the Acoustical Society
Empfindung. Philosophische Studien, 1888, 4, of America, 1957, 29, 603-606. (a)
541-594. STEVENS, S. S. On the psychophysical law. Psy-
MILLER, G. A. Sensitivity to changes in the in- chological Review, 1957, 64, 153-181 (b)
tensity of white noise and its relation to masking STEVENS, S. S. On the new psychophysics. Scan-
and loudness. Journal of the Acoustical Society dinavian Journal of Psychology, 1960, 1, 27-35.
of America, 1947, 19, 609-619. STEVENS, S. S. The psychophysics of sensory func-
MOSKOWITZ, H. R. Scales of intensity for single tion. In W. A. Rosenblith (Ed.), Sensory com-
and compound tastes. Unpublished doctoral dis- munication. Cambridge, Mass.: M.I.T. Press,
sertation, Harvard University, 1969. 1961. (a)
NEWHALL, S. M. A method of evaluating the spac- STEVENS, S. S. To honor Fechner and repeal his
ing of visual scales. American Journal of Psy- law. Science, 1961, 133, 80-86. (b)
chology, 1950, 63, 221-228. STEVENS, S. S. Power-group transformations
PLATEAU, J. A. F. Sur le mesure des sensations under glare, masking, and recruitment. Journal
physiques, et sur la loi qui lie 1'intensite de ces of the Acoustical Society of America, 1966, 39,
sensations a 1'intensite de la cause excitante. 725-735.
450 S. S. STEVENS
STEVENS, S. S. Tactile vibration: Change of ex- STEVENS, S. S., & GUIRAO, M. Loudness, recipro-
ponent with frequency. Perception and Psycho- cality, and partition scales. Journal of the
physics, 1968, 3, 223-228. (a) Acoustical Society of America, 1962, 34, 1466-
STEVENS, S. S. Measurement, statistics, and the 1471.
schemapiric view. Science, 1968, 161, 849-856. STEVENS, S. S., & GUIRAO, M. Subjective scaling
(b) of length and area and the matching of length to
STEVENS, S. S. On predicting exponents for loudness and brightness. Journal of Experi-
cross-modality matches. Perception and Psycho- mental Psychology, 1963, 66, 177-186.
physics, 1969, 6, 251-2S6. STEVENS, S. S., & GUIRAO, M. Loudness functions
STEVENS, S. S. Neural events and the psycho- under inhibition. Perception and Psychophysics,
physical law. Science, 1970, 170, 1043-1050. 1967, 2, 459-465.
STEVENS, S. S. Sensory power functions and STEVENS, S. S., & POULTON, E. C. The estimation
neural events. In W. R. Loewenstein (Ed.), of loudness by unpracticed observers. Journal of
Handbook of sensory physiology. Vol. 1. New Experimental Psychology, 1956, 51, 71-78.
York: Springer-Verlag, 1971. STEVENS, S. S., & STEVENS, J. C. The dynamics of
visual brightness. Cambridge: Harvard Univer-
STEVENS, S. S. Perceived level of noise by Mark
VII and dB(E). Journal of the Acoustical sity, Laboratory of Psychophysics, 1960.
TEGHTSOONIAN, R. On the exponents in Stevens'
Society of America, in press.
law and the constant in Ekman's law. Psycho-
STEVENS, S. S., & GALANTER, E. H. Ratio scales logical Review, 1971, 78, 71-80.
and category scales for a dozen perceptual con- TORGERSON, W. S. Distances and ratios in psycho-
tinua. Journal of Experimental Psychology, physical scaling. Acta Psychologica, 1961, 19,
1957 54, 377-411. 201-205.
STEVENS, S. S., & GREENBAUM, H. B. Regression
effect in psychophysical judgment. Perception (Received January 8, 1971; revision
and Psychophysics, 1966, 1, 439-446. received May 6, 1971)

Artigo Stevens 1971

Uploaded by

Copyright:

Available Formats

Artigo Stevens 1971

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Artigo Stevens 1971

Uploaded by

Copyright:

Available Formats

Psychological Review

1971, Vol. 78, No. 5, 426-450

ISSUES IN PSYCHOPHYSICAL MEASUREMENT

Two classes of ratio-scaling procedures are outlined—magnitude matching

Relative intensities of criterion stimulus

different functions were obtained, depending B to a common continuum C. The two

are typically steeper (have larger exponents) 100

Eleven subjects, chosen at random from Cross-Modal Ratio Matching

function does not accord with interval judg-

angles in Figure 4 show the magnitude esti-

Sum Production Under optimal experimental conditions,

On prothetic continua, the category scale

instances of power-group transformations. power functions have been recorded by one

to examine the implications. Consider in a range of exponents from about .3 to 1.5.

values was outlined (Stevens, 1955b). The Stimulus order 5 8 2 6

100 observers would necessarily respond with

Ratio to be produced changed the upper scale and the location of

strutted to squeeze a hand dynamometer to purpose of implementing a schematic struc-

You might also like