Psychological Review

1«63, Vol. 70, No. 1, 61-79

A THRESHOLD THEORY FOR SIMPLE

DETECTION EXPERIMENTS 1
R. DUNCAN LUCE
University of Pennsylvania

The two-state "high" threshold model is generalized by assuming that

(with low probability) the threshold may be exceeded when there is
no stimulus. Existing Yes-No data (that rejected the high threshold
theory) are compatible with the resulting isosensitivity (ROC) curves,
namely, 2 line segments that intersect at the true threshold prob-
abilities. The corresponding 2-alternative forced-choice curve is a 45°
line through this intersection. A simple learning process is suggested
to predict S's location along these curves, asymptotic means are derived,
and comparisons are made with data. These asymptotic biases are
coupled with the von Bdk&y-Stevens neural quantum model to show
how the theoretical linear psychometric functions are distorted into
nonsymmetric, nonlinear response curves.

A classic postulate of psychophysics questioned by some detection workers,

is that some stimuli or differences e.g., Swets (1961), Swets, Tanner, &
between stimuli never manage to Birdsall (1961), and Tanner & Swets
affect the central decision making (1954a) who have argued that thresh-
centers; others, of course, do. In a olds, if one still wishes to call them
phrase, peripheral thresholds were that, are introduced only at the
assumed to exist. At least three types central decision level itself. What is
have been distinguished: absolute, important in this view is that the
difference, and detection. It is not, value of the "response threshold"—
however, clear that there is any real usually it is called something else,
difference among them. Absolute such as a decision criterion or cutoff—
thresholds seem to be the same as is not a fixed feature of the organism,
detection ones except that the only but rather it is a parameter under the
noise is internal, and many difference control of the experimental instruc-
threshold experiments differ from de- tions, information feedback, payoffs,
tection experiments only in the nature and other motivational factors. Two
of the background stimulus, e.g., a versions of such a threshold-free
pure tone or noise. decision theory have been developed
Recently the literal interpretation in detail. For signal detectability
of the threshold postulate has been theory see Birdsall (1955), Green
1
This research was supported in part by (1960), Licklider (1959), Peterson,
Grants NSF G-17637 and NSF G-8864 from Birdsall, and Fox (1954), Swets and
the National Science Foundation to the Birdsall (1956), Swets, Fanner, and
University of Pennsylvania. Birdsall (1955, 1961), Tanner (1955,
I wish to express my appreciation to R. R. 1956), Tanner and Birdsall (1958),
Bush, Eugene Galanter, Francis W. Irwin, W.
D. Larkin, Donald Norman, and Elizabeth F. Tanner and Norman (1954), and
Shipley for the many discussions we have had Tanner and Swets (1954a, 1954b).
of the ideas included in this paper. In addi- For the choice theory see Luce (1959),
tion, Elizabeth F. Shipley has graciously al- Restle (1961), Shepard (1957), and
lowed me to include portions of the data from
her thesis, which shortly will be reported in Shipley (1960, 1961). A number of
full elsewhere. experiments have been reported which
61
62 R. DUNCAN LUCE

agree with the main features of both way by a continuous, threshold-free

theories. In addition to those re- model. The only published attempt
ported in the above references, there that I know of is by Barlow (1961),
are studies by Clarke, Birdsall, and and his rationalization seems com-
Tanner (1959), Creelman (1959,1960), pletely ad hoc to me.
Egan, Schulman, and Greenberg Second, Shipley (1961) has obtained
(1959), Green (1958), Green, Birdsall, some simultaneous detection and rec-
and Tanner (1957), Shepard (1958), ognition data which indirectly sug-
Swets (1959), Swets, Shipley, McKey, gest that detection thresholds exist.
and Green (1959), and Veniar (1958a, On each trial either a 1,000-cps tone
1958b, 19S8c). Although the two in noise, a 500-cps tone in noise, or
theories differ conceptually, their pre- noise alone was presented, and the
dictions are so similar that it has been subject was required to decide whether
impossible as yet to decide between or not a tone was present and, in-
them. dependent of his detection response,
In the course of evaluating signal to attempt to recognize which it was.
detectability theory, a contrasting but (Controls were run in which no
equally explicit, sensory threshold recognition response was required and
model has been stated (Swets, 1961; in which recognition was only required
Swets et al., 1961; Tanner & Swets, when the subject said a tone was
1954a). It postulates that the thresh- present; there did not seem to be any
old is well above the noise level. interaction between the forced rec-
There is no doubt that this model is ognition responses and the detection
inadequate, and it has been concluded responses.) If we separate the two
that if thresholds exist they must be detection responses, then we can ask
so far down in the noise that the how well he recognizes when he says
notion of a threshold ". . . is not a he heard a tone as against when he
workable concept . . . [and] for said he did not hear one. If there
practical purposes, not measurable" really is a sensory threshold and if he
(Swets et al., 1961, p. 336). At least reports no tone present only when the
two sets of behavioral data do not threshold is not exceeded, then there
jibe easily with this view. should not be any differential recogni-
First, there are studies, beginning tion of the tones on the no-detect
with von B6kesy (1930) and Stevens, trials. This is what happens, as can
Morgan, and Volkmann (1941), of the be seen in Table 1, for both the Yes-
detection of energy increments of a No and forced-choice designs.
pure tone background. Some of the This paper has two main purposes.
results reported seem consistent only First, a simple threshold model is
with a quantal (threshold) model. described which appears to give as
Although a number of people are satisfactory an account of the response
dissatisfied with aspects of the experi- data as do the continuous detection
mental procedure and although the theories. Second, a way is suggested
psychometric function has not always to graft onto this sensory threshold
been found to be rectilinear as pre- model a decision process which pre-
dicted by some quantum theorists, the dicts in some detail the biasing effects
recurring n: (n — 1) relation between of information feedback, payoffs, and
the probability one and zero intercepts presentation probabilities. It is note-
of the psychometric function has not worthy that in conjunction with the
been accounted for in any satisfactory present threshold model the usually
 A THRESHOLD THEORY FOR SIMPLE DETECTION EXPERIMENTS 63
TABLE 1
PROBABILITY OF RECOGNITION CONDITIONAL ON THE DETECTION RESPONSE

Subject

Design Presentation i 2 3

Yes No Yes No Yes No

Yes-No 500 cps 90 78 89 74 88 31

1,000 cps 8 76 20 75 12 42
noise 41 72 57 79 48 32

Correct Incorrect Correct Incorrect Correct Incorrect

Forced- 500 cps 88 43 87 63 87 43

choice 1,000 cps 14 47 25 63 13 41

Note.—These data from a simultaneous detection and recognition experiment are reported by Shipley (1961).
On each trial of the Yes-No experiment either a 500-cps tone in noise, a 1,000-cps tone, or noise alone was presented,
and the subject was required to respond whether or not a tone was presented and, independent of that response,
to recognize which tone it was. In the table, the conditional probabilities (decimal points are omitted) that a
presentation is recognized as the 500-cps tone are estimated separately for those trials when the subject said he
detected a tone (Yes columns) and for those when he said he did not (No columns). On each trial of the two-
alternative forced-choice experiment, either 500 cps or 1,000 cps appeared in exactly one of the two temporal
intervals. The subject was required to state which interval contained the tone and which tone it was. The condi-
tional probabilities that the presentation is recognized as the 500-cps tone are estimated separately for those trials
when the subject chose the interval containing the tone (Correct columns) and for those when he was Incorrect
(Incorrect columns). Note that the conditional probabilities depend on the presentation in the Yes and Correct
columns but are independent of the presentations in the No and Incorrect columns.

assumed expected-payoff model is Yes and No, although in practice he

completely unacceptable. Instead, a usually selects one of two buttons to
learning process is postulated. Thus, press. In many of the acoustic experi-
the present psychophysical theory is, ments, the background is white noise
in part, an asymptotic learning theory, and the stimulus a pure tone, but this
as seems sensible. is not necessary. For example, in the
von Be"kesy-Stevens quantal experi-
YES-NO DETECTION EXPERIMENTS ments, the background is a pure tone
of one energy and the background plus
It is generally agreed that one of stimulus is a tone of the same fre-
the simpler detection experiments is quency but different energy. None-
the Yes-No design. On each trial un- theless, I shall conventionally speak
ambiguous signals mark off a time of the background as noise.
interval during which either a back-
ground or the background plus a Let n denote a typical presentation
stimulus 3 is presented, and the subject of noise, s a typical presentation of
is required to indicate whether or not stimulus plus noise, and Y the Yes
he thinks the stimulus is there. Often and N the No responses. The basic
the possible responses are said to be data are the relative frequencies of a
8
Y response given s, j>(Y\s), and of Y
In the signal detectability literature, the given «, f>(Y\ri), which are assumed
physical event to be detected by the subject
has generally been called a "signal." As to arise from and therefore to estimate
long as one is working with tones in noise and the true conditional response prob-
the like, this does not seem inappropriate; abilities p(Y\s) and p(Y\n). With
however, nothing in this theory restricts one
to signals in this sense, so I have elected to or without "hats," it is clear that
use the more general term "stimulus." p(N\s) = 1 - p(Y\s) and p(N\n)
64 R. DUNCAN LUCE

= 1 — p(Y\n). Our problem is, first, that noise alone generates a detection
to explain how these two conditional observation, i.e., that it "passes" the
probabilities relate to one another threshold, and q(s), the true prob-
when we vary such experimental ability that stimulus plus noise gener-
parameters as the a priori probability ates a detection observation. We
P of presenting s (and so 1 — P of «), assume that q(s) > q(n).
the physical magnitudes of 5 and n, In the absence of data, one might
and the payoffs, The proposed an- have supposed that P(Y\s) = q(s)
swers, although far from complete, and P(Y\n) = q(ri), but this cannot
permit some experimental evaluation be because as a matter of fact the
of the model. Second, given estimates values of p(Y\s) and p(Y\ri) depend
of parameters from Yes-No data, we upon at least P, the instructions, and
must try to account for the data from the payoffs; and these differences are
other experimental designs involving much too large and systematic to be
the same s, n, and subject. ascribed to variability in the data.
We shall suppose that thresholds Evidently, then, the subject must
exist in the following sense. When convert some of the D observations
either the noise alone or the stimulus into N responses or some of the D
plus noise is presented, the organism observations into F responses, de-
enters one of two hypothetical states pending upon how he wishes to bias
denoted D and D. A "detection ob- the outcome. On the assumption that
servation" will be said to have oc- the D observations are all indistin-
curred when he goes into State D and guishable, or, at least, that the 5 and n
not to have occurred when he goes distributions of D observations are
(or stays) in D. These states are the same and that this is also true of
assumed to be internal to the subject the D observations, it is plausible that
and therefore cannot be directly ob- the bias involves responding "in-
served in terms of behavior. Whether correctly" to some random fraction
they can be studied by physiological of the observations. If so, we obtain
methods is an open question that we two different sets of equations depend-
need not discuss here. We do not upon which bias is introduced:
suppose that the same state neces- if p(Y\ri) < q(n), then
sarily results whenever a particular
stimulus is presented, but rather that p ( Y \ s ) = tq(s)
p(Y\n) = tq(n), [1]
the state entered is determined by a
random process that is characterized or if p(Y\ri) > q(n), then
by fixed probabilities for a given
subject, stimulus, noise, and experi- p(Y\s) = 2(5)+«[l -q(s)
ment. Just where the variability p(Y\n) = <z(») + «[l - < z ( [2]
enters in is not specified by the theory.
The underlying conditional prob- where 0 < t, u < 1.
ability model for these detection For the moment, we are not con-
observations (not responses) is cerned about the actual values of the
bias parameters t and «; rather we
Presentation Presen- assume that any particular value can
probability tation Observation
be made to arise, and we eliminate
D D
P these unknowns from Equations 1 and
1- P 2 to obtain the dependence of p(Y\s)
upon p(Y\n) with q(s) and q(n) as
In words, q(n) is the true probability parameters:
 A THRESHOLD THEORY FOR SIMPLE DETECTION EXPERIMENTS 65

p(Y\n), iip(Y\n) < q(n).

q(n}
p(Y\s) = [3]
q(s) - q(n)
iip(Y\n) >
1 - q_(n) '

This equation describes a very simple and P was varied from 0.1 to 0.9 in
function, namely, a straight line steps of 0.2. In Figure 1, 1 have pre-
segment from (0, 0} to (q(n), q(s)) and sented the data and detectability
another from (q(n), q(s)) to (1, 1), the curves for one subject from each
two portions of which we shall speak experiment, choosing in each case the
of as the lower and upper limbs, re- subject that most favors signal de-
spectively. tectability theory. All of the data
In the signal detectability literature and the curves of the present thresh-
the function relating p( Y \ s) to p (Y \ n) old model are shown in Figure 2 for
has been called a receiver operating the visual experiment and in Figure 3
characteristic or, more briefly, an for the acoustic experiment. Through-
ROC curve, but it seems more ap- out the theoretical curves were fit by
propriate to call it an isosensitivity eye, because no optimal statistical
curve.4 In that theory, it is truly a procedure is known. (The theoretical
smooth curve. Examples of curves crosses in Figure 3 will be discussed
generated by detectability theory are later.)
shown in Figure 1 and of ones gener- In evaluating the acoustic data,
ated by Equation 3, in Figures 2 and 3. two facts are important. First, the
The high threshold model discussed p(Y\s) coordinate of each data point
by Swets (1961), Swets, Tanner, and is based upon a sample of 300 P, ob-
Birdsall (1961), and Tanner and servations and the p(Y\n) coordinate
Swets (1954a) is the special case of on 300 (1 — P) observations. Sec-
this one in which q(n) = 0, and so it ond, successive pairs of points reading
consists only of the upper limb, i.e., of around the curve, were generated
the line segment from (0, q(s)) on the under identical experimental condi-
ordinate to (1, 1). This is not a satis- tions. Thus, there can be little doubt
factory summary of the data, but the that there is variability beyond the
two line segments of Equation 3 do binomial associated with each ob-
about as well as any of the continuous servation point.
theories with the same number of free I would judge that the visual data
parameters, namely, two. For ex- slightly favor the threshold model and
ample Swets, Tanner, and Birdsall the acoustic data, the detectability
(1955, 1961) report data on visual model. Although different modalities
brightness for four subjects, where the may well involve different processes,
payoffs were varied and P was held neither set of data seems particularly
at 1/2. And Tanner, Swets, and conclusive. One feature of both sets,
Green (1956) report acoustic data on however, casts suspicion upon the
the detection of a 1,000-cps tone in
present threshold model. The model
white noise for two subjects, where the
(symmetric) payoffs were held fixed says that the isosensitivity curve has
4
a sharp corner which all too often
1 hope that the greater naturalness of this seems to float free of these data points.
term, as compared with equisensitivity, will
be adequate compensation for mixing Greek Of course, this is exactly what would
and Latin roots. happenjwere the true function a
66 R. DUNCAN LUCE

Visual Brightness Acoustic

Subject 4 Subject 2

FIG. 1. Yes-No detection data and the corresponding theoretical isosensitivity curves
derived from signal detectability theory. (The visual brightness data, reported by Swets
et al., 1955, 1961, were obtained under the same stimulating conditions with a presentation
probability of 0.5, but with different payoff matrices. The detection of a tone in noise data,
reported by Tanner et al., 1956, was obtained under the same stimulating conditions with a
fixed symmetric payoff matrix, but with different presentation probabilities.)

cornerless curve; hence, this threshold (D, D) occurs, never when (D, D)
model does not deserve serious con- occurs, some proportion v when {D, D)
sideration unless the lonely corners occurs, and another proportion w
are explained. A reason is suggested when (D, D) occurs. If we assume
later. that the observation probabilities in
the two intervals are independent,
TWO-ALTERNATIVE FORCED-CHOICE then the probability of, say, a (D, D}
EXPERIMENTS observation when (s, n) is presented is
2(s)[l ~ Q.(n)~] because q(s) is the
In the two-alternative forced-choice probability that the stimulus plus
design two time intervals are denned noise exceeds the threshold and
and the stimulus is, and is known to 1 — q(n) is the probability that noise
be, in exactly one. Thus, the two alone fails to exceed it. The other
presentations are the ordered pairs cases are similar, and they lead to
(5, n) and (n, s). The subject responds
by saying which interval, 1 or 2, he
believes to have contained the stimu- (w)] + vq(s)q(n)
lus. Assuming the above threshold
<Z (*)][! -«(»)] [4]
formulation, there are four possible
observations, (D, D), (D, D}, (D, D) /»(! !<»,
and (D, D), of which two, (D, D) and vq(n)q(s)
(D, D) give the subject no indication
of which response to make. It seems ], [5]
plausible, at least when the payoffs where 0 < v, w < 1.
are not too extreme, that the subject It follows by subtraction that
should apply biases only to these two
ambiguous cases. Thus, we assume p(\\(s,n)) = p(l\(n,s))
that he always responds 1 when , [6]
 A. THRESHOLD THEORY FOR SIMPLE DETECTION EXPERIMENTS 67

Subject 1 Subject 2

.0 0 .2 .4 .6 .8 1.0
p(Y|n)
Subject 3 Subject 4
1.0

.4

.2

.2 .4 .6 .8 1.0 0 .2 .4 .8 1.0
p(Y|n)

KIG. 2. Yes-\o visual brightness detection data from Swets, Tanner, and Birdsall (1955,
1961) and the corresponding theoretical isosensitivity curves derived from a threshold theory.
(Each coordinate of each point is based upon 200 observations.)

Subject 1 Subject 2

0 ,2 .4 .6 .8 1.0 .4 .6 .8 1.0

•"IG. 3. Yes-No acoustic (tone in noise) detection data from Tanner, Swets, and Green (1956)
and the corresponding theoretical isosensitivity curves derived from a threshold theory.
68 R. DUNCAN LUCE

so the isosensitivity curve is a biasing parameters so as to maximize

line segment with slope 1 running his expected payoff. Let the payoff
from <g(»)[l -<?(*)], g(s)[l -g(n)]> to structure be
<l-g( S )[l-g(»)],l-g(»)r.l-g( s )]>. Presentation Presen-
Thus, for example, if q(s) = 0.9 and probability tation Response
q(n) = 0.2, the segment runs from Y N
(0.02, 0.72) to (0.28, 0.98).
It is also easy to see from Equations
P
1 - P [ On
"21

4 and 5 that when v — w — q(n)q(s)/ then if the subject is on the lower limb
{q(n)q(s} + [1 - g(»)][l - g(s)]}, of the threshold model the expected
then p (11 (s, «)) = g (s) and £ (11 (n, s)) payoff is
= q(n). That is, the two-alterna-
tive forced-choice isosensitivity curve E(o) = Pp(Y\ s)0li
passes through the point whose co- + P[1 -p(Y\s)-]olt
ordinates are the true threshold + (1 - P)p(Y\n)on
possibilities. + (1 -P)[l -p(Y\n]on
These last two remarks, coupled
with the results about the Yes-No = t[Pq(s)(on - 012)
design, give a way to estimate the true + (1 - P)g(»)(o2l - OM)]
threshold probabilities. Suppose for + Pou + (1 - P)022.
the moment that the several response
probabilities are known. The point Because this equation is linear in t, the
(q(ri),q(s)} lies both on the 45° line maximum occurs either at t = 0 or
t = 1. A similar calculation for the
passing through (p(\. [ (n, s)) , p ( l \ (s, n)))
upper limb yields either u = 0 or
and on a line passing through ( p ( Y \ n ) ,
p(Y\s)} and either (0, 0) or (1, 1), de-u = 1. Thus, the expected payoff
pending upon which limb of the Yes- model places the subjects at one of
No model is involved. Thus, the three points: (0,0), (q(n), q(s)}, or
intersection of one of these two pairs (1, 1). This is clearly wrong (see
of lines is the point (q(n), q(s)). The Figures 2 and 3).
geometry is shown in Figure 4. Whether this prediction is wrong
So far as I know, no empirical because of the threshold model or be-
isosensitivity curves have been pub- cause of the expected payoff model is
lished for the two-alternative forced- less easy to decide. One thing about
choice experiment, so we cannot check the expected payoff model should be
noted: knowledge of the two subject-
our prediction that it is a straight line
with slope 1. This prediction differs determined conditional probabilities
considerably from the curve—which p ( Y \ s ) and p(Y\n) is needed to
is also symmetric about the diagonal calculate the values of the parameters.
from (0, 1) to (1, 0}—predicted by Certainly no one will claim that the
signal detectability theory. subject "knows" these, even un-
consciously, in a way that he can
ASYMPTOTIC LEARNING actually calculate expected values;
more likely, he arrives at his biases by
We turn next to the question of the a process of adjusting to his experience
values of the biasing parameters, t, u, —by learning. It is curious that no
v, and w. In the decision and choice one has yet evolved a learning theory
theory models for these experiments, which, asymptotically, predicts the
it has been customary to assume that maximization of expected values. An
the subject selects values for the alternative, and to my mind more
 A THRESHOLD THEORY FOR SIMPLE DETECTION EXPERIMENTS 69

„ i.o

34
"

p(Y|n)
-i-
. p(Y|n)
ptl|<n,s>)
1.0

qln)
LOWER LIMB UPPER LIMB
FIG. 4. The geometry relating the Yes-No isosensitivity curves, the two-alternative
forced-choice isosensitivity curves, and the true threshold probabilities.

reasonable, tack is to postulate di- Trial i the bias is ti. What is the bias
rectly a learning process, preferably ti+i on Trial i + 1 ? Of the events oc-
one that has already achieved some curring in Trial i, the only two that the
success in other areas, and to test its subject should rationally take into
asymptotic predictions against be- account in modifying the bias are his
havior. This we do.5 observation, D or D, and what he
Consider a subject who is operating later learned the presentation to have
on the lower limb of the Yes-No iso- been, s or n. If he is rational, he
sensitivity curve, and suppose that on certainly should not let the response
6
Conceptually, there is no special affinity
he made on Trial i or, for that matter,
between learning and thresholds, but in on any of the preceding trials influence
practice there are good reasons why it is his choice of bias. Because on the
easier to graft a learning mechanism on this lower limb the bias only tells him
threshold model than on the signal detect- how often to respond N to D observa-
ability model. In both there are three classes
of events that a subject might use to control tions, it seems clear that he should not
his learning: the hypothetical internal ob- change it when a D observation oc-
servations, his responses, and what he learns curs. When a D observation results
the presentation to have been. Both the first from an 5 presentation, the bias cer-
and third of these events form statistically tainly should not be lowered, which
stationary processes over trials, whereas the
response probabilities are changing. Thus, would only decrease the Y responses,
if the learning process is dependent upon the and it should not be increased when a
responses, the resulting stochastic learning D observation results from an n
model is mathematically quite complex and I presentation. That is, we expect
do not know how to analyze it. Although the
other two classes of events do not have this
particular complexity, the first can introduce
a different kind. The internal observations
that are assumed to occur in the detectability
model take on values in a continuum, and so
the learning model for this case must be
continuous, and such models are not yet The exact nature of the transition is
very well understood. The threshold model not obvious; however, a linear oper-
has the distinct advantage that there are only
a small number of observations states, which ator (Bush & Mosteller, 1955) is
results in a mathematically simple learning certainly one of the simplest possi-
process. bilities and one that has received
70 R. DUNCAN LUCE

considerable attention . So we assu me6 tity b depends upon P and upon the
two learning rate parameters 0 and 0',
which presumably in turn depend
upon the payoffs. We have no theory
for this dependence, so in general b
will have to be estimated from the
Then data or, as when we assume the learn-
ing rates to be equal for symmetric
payoffs, an assumption will have to
be made about 0 and 0'. Clearly, if
+ (l-P)p(D\n)(l-0% 0 4= 0 and 0' 4= 0, b ranges from 0
+ (l-P)p(D\n)ti when P = 1 to » when P — 0. At
some point when P is varied the
= ti{Pq(s)(l-6) subject presumably changes from
operating upon the upper limb to the
lower limb. (See Figure 3 where P
varies from 0.1 to 0.9.) We do not
have a general theory for when this
+Pq(*)0. change occurs, but it seems plausible
Taking expectations over 2; and that it should be somewhere in the
then the limit as i —> <» yields as the middle range of P values. If so, then
asymptotic expected bias MOO is bounded away from 0 to the
extent that q(s) is less than 1 and tx
, [8] is bounded away from 1 to the extent
+ &?(«) that q(n) is greater than 0. Or
where translated back to the isosensitivity
curve, the upper limb data points are
prevented from being near the corner
of the curve to the extent that q(s) is
less than 1 and the lower limb points
Similarly, we postulate the follow- are prevented from being near it to the
ing learning process for the upper extent that q(n) is greater than 0.
limb: An examination of Figures 2 and 3
suggests that the data are consistent
with this statement, which may ex-
plain why the corners seem to be
isolated. It also suggests that in-
formation feedback need not always
and a parallel calculation gives be beneficial in inducing subjects to
MOO = lim E («;) yield up the desired information, in
this case, the true threshold prob-
abilities, as is often assumed by
modern psychophysicists.
For the two-alternative forced-
According to Equation 9,rthe quan- choice model, we assume essentially
fl
Although I will state the operators in
the same learning process, namely
terms of the bias parameters, they could
equally well be stated in terms of the response (1-0X-+0, if (s, n) and (D, D)
probabilities because these probabilities are (1-0'X, if (n, s) and (D, D}
linear functions of the bias parameters. Vi, otherwise.
 A THRESHOLD THEORY FOR SIMPLE DETECTION EXPERIMENTS 71

A calculation similar to that for the doubtedly has somewhat less binomial
Yes- No experiment yields variance. Moreover, the learning
process itself introduces added vari-
vv = !/(! + b). D2]
ance which more seriously affects our
Similarly, estimate of the Yes-No point than
f(l-0)w;+0, if (s, n)and(D, D) of the forced-choice one. Once q (n)
wi+1 = j (1 -6')wi, if (n, s) and (D, D) and q(s) are estimated, then the
I Wi, otherwise theoretical location of the data points
on the isosensitivity curves is de-
yields termined by Equations 1 and 8 or 2
w. = !/(!+ [13] and 11 for the Yes-No experiment
We note that vx = WK, as seems a and by Equations 4, 5, 12, and 13 for
priori reasonable, and that neither the forced-choice experiment, pro-
bias depends upon the underlying vided that we know b. If we assume
probabilities q(n) and q(s) as in the equal learning rates, then b = 1. The
Yes-No experiment. We also note comparison between data and theory
that if P = 1/2 and if the learning under that assumption is shown in
rates are equal, then the biases Figure 5; it is surprisingly good, but
are symmetric in the sense that unfortunately it does not permit us to
vx = Woo = 1/2. decide which limb is being used.
There is some suggestion that it may
EMPIRICAL TESTS be the lower one, for in four of the
six cases that and only that inter-
To test the model, we have four sets section is in the unit square, but this is
of data, all collected on W. P. Tanner's far from conclusive.
equipment in the Psychophysical Lab- Swets (1959) reported similar data
oratory, Electronic Defense Group, on three subjects for several different
University of Michigan. Shipley signal to noise ratios. The plots are
(1961) ran each of three subjects in, similar to those for Shipley's data;
among other conditions, the Yes-No however, the predictions do not seem
and two-alternative forced-choice de- to be quite so accurate. In part this
signs with P = 0.5 and with sym- is due to the smaller sample sizes used
metric payoffs. Each condition was by Swets.
run twice with different stimuli, pure
tones of 500 and 1000 cps. Each pres- Next, we have the acoustic data
entation, s or n in the Yes-No and from Tanner, Swets, and Green (1956)
(s, n) or (n, s} in the forced-choice which were presented in Figure 3.
Again, because the payoffs were sym-
design, occurred 800 times. Using the
metric, we assume equal learning
estimation scheme of Figure 4, values
parameters, so b is determined by P.
for q(n) and q(s) were obtained for
both limbs. If either or both inter- The predicted values, assuming that
sections lay outside the unit square, the P — 0.1 and 0.3 points are on the
lower limb and that the rest are on the
I selected the intersection of the 45°
line through the forced-choice data upper one are shown as crosses in
point and the edge of the unit square Figure 3; recall that successive pairs of
as the final estimate. This incorrectly data points were collected under iden-
attributes all of the error variance to tical experimental conditions. The
the Yes-No data point; however, predictions seem satisfactory for Sub-
because of the location of the two ject 1, but less so for Subject 2.
points, the forced-choice point un- Because no study has yet been made
72 R. DUNCAN LUCE

500 cps tone 1000 cps tone

Subject I

Subject 2

Subject 3

0 .1
p ( Y | n ) and

FIG. 5. Yes-No and two-alternative forced-choice acoustic (tone in noise) data reported
by Shipley (1961). (The theoretical curves are from the threshold model and the predicted
values—crosses—are asymptotic values derived from a linear learning process.)

of the learning process itself, I do not visual data in Figure 2 are less easy
know how adequate the assumption to make because the isosensitivity
0 = 8' is, but to the extent that it is curve was generated by varying the
wrong errors are introduced into our payoffs, not P. The only information
predictions. that we have about the payoffs used
Comparable predictions for the are the numbers
 A THRESHOLD THEORY FOR SIMPLE DETECTION EXPERIMENTS 73

I first tried using K — 1 to predict the

P responses. For three of the subjects
this seemed satisfactory, but by trial
which are the relevant criterion quan- and error I found that K = 0.5 is a
tities if one assumes that the subjects much better choice for Subject 1. The
maximize expected payoffs and that results are shown in Table 2. Note
they are described by the signal the rather sharp break in both the
detectability model. If we assume observed and predicted values of
that the learning rate parameter asso- pa(Y\n) as one moves from the lower
ciated with s presentations, 6, is to the upper limb, as indicated by
proportional to the difference of the the bold face vertical bars in the table,
two 5 payoffs, o u — 012, and that 6' is even though the changes in /3 are
proportional to 022 — 021, then small in that region.

k-ALTERNATIVE FORCED-CHOICE
EXPERIMENTS
Thus, in addition to q(n) and q(s), The two-alternative forced-choice
there is the free parameter K to be design can be readily generalized to
chosen when fitting data. The values one having k intervals, exactly one of
for q(n) and q(s) we get from Figure 2. which contains the stimulus. It is not
Because it is reasonable that the two easy to work out the response prob-
constants of proportionality relating abilities for any model, including this
0 and 6' to payoffs might be the same, one, except under the assumption that

TABLE 2
ASYMPTOTIC LEARNING MODEL PREDICTIONS OF VISUAL DATA

Sub- s
K
ject
8 8 6 4 2.5 2 1.5 i .75 .75 .50 .25 .16

p(Y\s) observed 64 64 82 79 — 63 85 77 89 84 88 91 96
1 05 80 5 predicted 64 64 67 72 — 76 84 86 87 87 89 93 95
p(Y\n) observed 2 3 2 6— 5 20 32 19 29 50 65 78
predicted 4 4 4 4— 5 26 33 39 39 48 65 74
p(Y\s) observed 48 59 73 75 71 80 90 84 87 89 95 93 95
2 05 85 1 predicted 58 58 63 69 74 86 86 87 88 88 89 91 93
p(Y\n) observed 0 0 2 3 4 11 19 23 18 25 25 41 66
predicted 3 3 4 4 4 12 14 18 22 22 28 42 52
f>(Y\s) observed 45 56 64 73 65 78 90 84 87 88 85 94
3 03 83 1 predicted 64 64 — 70 76 77 85 86 86 86 87 90 92
p(Y\n) observed 0 2 — 2 8 1 15 18 14 24 12 33 59
predicted 2 2 — 3 3 3 13 18 21 21 28 43 54
p(Y\s) observed 43 43 64 64 56 71 76 77 84 86 83 87 93
4 07 74 1 predicted 42 42 47 53 60 77 78 79 81 81 83 84 91
p(Y\n) observed 0 4 6 12 5 17 14 18 34 34 35 49 74
predicted 4 4 4 5 6 18 22 27 32 32 40 42 66
Note.—-Response proportions predicted by asymptotic learning model for visual detection data reported by
Swets, Tanner, and Birdsall (1955, 1961). The vertical bold face line indicates the transition from lower to upper
limb. Each observed proportion is estimated from 200 observations. Decimal points have been systematically
omitted.
74 R. DUNCAN LUCE

is correct one-feth of the time when

all intervals produce D observations.
These are the only ways a correct
response can occur, so

+i

Pk(O

-e(»)]}. [14]
For*> 2,pt(Q = (1 + A)/2, where
A ~ q(s) — q(n), whereas for k > 2,
pk(C) depends upon both q(n) and
q(s) and not just upon their difference.
Thus, in contrast to other theories,
FIG. 6. Maximum and minimum curves pk(C) is not uniquely determined by
of the proportion of correct responses in pz(C). To get an idea of the freedom
/fe-alternative forced-choice designs. (The involved, assume A is fixed, then the
data points for the detection of a tone in limiting possibilities are when q (n) = 0
noise are from Swets, 1959.)
and q(s) — A, in which case
the asymptotic response biases are p,(Q = [A(k- 1) + !]/*, [15]
equal. In the two-alternative case, and when q (n) = 1 — A and q (s) = 1 ,
this means setting vx = 1/2, which is in which case
what happens in the learning model
if the learning rates are equal and
P = 1/2. In general, it means that in
any ambiguous situation the several Typical examples of these bounds are
possibilities are used equally often. shown in Figure 6. The data points
The effect of this symmetry assump- are from Swets (1959) ; clearly they
tion is to make the probability of a fall within the bounds.
correct response independent of the Swets (1959) ran three other sub-
stimulus presentation. Correct re- jects in the Yes- No and the two- and
responses can occur in the following four-alternative forced-choice designs.
ways: A response is always correct If we estimate the values of q(s) and
when the stimulus produces a D q(ri) from the Yes- No and two-
observation and the_ k — 1 noise pre- alternative forced-choice data using
sentation produce D observations; it the method of Figure 4, then we can
is correct one-half the time when 5 predict what should be observed in the
and exactly one n produce D observa- four-alternative forced-choice experi-
tions, which can happen m
. / ment. 7 Because it is not always clear
( from the Yes-No data which limb was
ways; it is correct one-third of the 7
I wish to thank J. A. Swets for providing
time when s and exactly two n's pro- me with the raw data to make these calcu-
duce D observations; etc.; and it lations.
 A THRESHOLD THEORY FOR SIMPLE DETECTION EXPERIMENTS 75
TABLE 3
FOUR-ALTERATIVE FORCED-CHOICE EXPERIMENT

Estimated parameters P>(C)

Signal to
Subject noise ratio Upper limb Lower limb
in db Upper Lower
limb limb Observed
«(») «w «(») aW
1 9.4 13 77 25 89 67 62 62
14.5 0 87 13 100 90 82 82
16.6 0 89 11 100 92 85 88
2 9.4 19 72 33 86 58 53 52
11.7 17 74 28 86 62 57 63
14.5 7 78 29 100 75 64 75
16.6 2 83 19 100 84 75 79
3 9.4 0 69 25 94 76 65 68
11.7 7 80 27 100 77 66 73
14.5 0 83 17 100 88 80 85
16.6 0 92 8 100 94 88 90
Note.—Estimates of «(«) and q(s) from Swets (1959) Yes-No and two-alternative forced-choice data, and the
predicted and observed values of #4(0. Each subject made 500 observations at each signal level in each experi-
mental condition. Decimal points have been omitted on all the probabilities.

used, the calculations are reported for of the stimulus magnitude is usually
both limbs in Table 3. These predic- called a psychometric function. For
tions suggest that Subject 1 was example, in the von Bekesy-Stevens
operating on the lower limb; that quantal theory, the theoretical func-
Subject 2 was on the upper limb for tion is 0 for all stimulus increments
at least the three most intense stimuli; less than one amount, 1 for all incre-
and that the picture is not clear for ments larger than another, and a
Subject 3. It is certainly the case straight line between these two points.
that one of the two predictions is As no distinction has been made in the
always near the observed value. quantal literature between what we
Tanner, Swets, and Green (1956) are calling q(s) and px(Y\s), it is not
report four-alternative forced-choice perfectly clear which function is
data for the same subjects whose Yes- meant. There is no question that
No data are shown in Figure 3. Esti- in testing the theory, estimates of
mating q(n) =0.11 and q(s) = 0.68 px(Y\s) have been plotted against
from the Yes-No data for Subject 1, increment size, but an examination of
we predict p 4 (O = 0.63; 0.60 was ob- the theory itself suggests that we
served. For Subject 2, q(n) = 0.28, should interpret it as referring to q (s),
q(s) = 0.74, and we predict pt(C) Assuming that the above learning
= 0.51; 0.56 was observed. In both model for biasing is correct, the
cases, p*(C) was estimated from 297 relation between p»(Y\s) and q(s) for
observations. the lower limb bias is obtained from
Equations 1 and 8; it is
DISTORTION OF THE PSYCHOMETRIC
FUNCTION b \(V\*\
Koo = l"X
-* \1 ^} — f {flM
i V- 1 / = / \ ^ i 'i / \
q(s) + bq(n)
A plot of the Yes-No detection
probability versus a physical measure Because bq(n) > 0, px(Y\s) < q(s)
76 R. DUNCAN LUCE

on the lower limb, and its maximum The following hypothesis is cur-
value, 1/[1 + &<?(«)], occurs when rently under investigation and it ap-
q(s) = 1. Similarly, Equations 2 and pears to have some merit. In neural
11 yield the result for the upper limb: quantum theory (Stevens etal., 1941),
those stimulus increments that cause
p*(Y\s) = < z ( s ) + M » [ l -g(5)] zero and one quantum changes are
assumed not to be detected, whereas
those that cause changes of two or
more quanta are. Let us_suppose that
In this case p(Y\s) > 9(5), and its this defines our states D and D, re-
minimum value, q(n) + [1 — <?(»)]/ spectively. In addition to this as-
(1 + &), occurs when q(s) = q(n). If sumption, let us postulate that the
we suppose that q(s) is a rectilinear subject also uses the change in the
function having p = 1 and p = 0 number of quanta excited to decide
intercepts in 2:1 ratio and that which bias to use. Specificially, let
g(w) = 0.05, then we get plots like us suppose that there is an integer
those shown in Figure 7, where & is a h > 0, such that if the stimulus pro-
parameter. duces a change of fewer than h quanta,
Once again, we are not sure when he imposes a lower limb bias, and that
the subject switches from lower to if it produces a change of h or more,
upper limb biasing. It is clear that he imposes an upper limb bias. We
such a switch must occur, for when he do not know what determines the
is on the lower limb px(Y\s) can never choice of h, but presumably it de-
reach 1, no matter how intense the pends in part upon instructions, pres-
stimulus is. entation probabilities, and payoffs.
In any event, we can see what sorts of
psychometric functions result from
different choices for h.
For h — 0, the subject always uses
an upper limb bias; these are the
upper functions shown in Figure 7.
For all other values of h, there are re-
gions of stimulation where either h — 1
or h new quanta are excited by the
stimulus, and so the data will be a
weighted average of the response
curves resulting from lower and upper
limb biases. The probability that h
additional quanta are excited depends
upon the probability that the energy
residue of the background plus the
I 2
QUANTAL UNITS
stimulus exceed h quantal units of
energy. In quantum theory it is
FIG. 7. Theoretical upper and lower limb usually assumed that the distribution
psychometric functions when q(s) is assumed
to be rectilinear with a 2:1 ratio of intercepts, of residues is (approximately) uni-
q(n) = 0.05 and b = 1/10, 1, and 10. (The form, in which case we simply have
upper limb curves are above, and the lower the following rule: Let 5 denote the
limb ones are below the q(s) curve, which is
indistinguishable from the lower limb value of the stimulus in quantal units,
b = 1/10 curve.) ii s < h — 1, then there is a lower
 A THRESHOLD THEORY FOR SIMPLE DETECTION EXPERIMENTS 77

h-i h=3
1.0 1.0

.8 .8

- .6
</>
.6
>
a. .4 .4

.2 .2

0 0

h=4

b=IO

.2

Q U A N T A L UNITS
FIG. 8. Theoretical px(Y\s) psychometric functions for different values of h when q(s)
is assumed to be rectilinear with a 2:1 ratio of intercepts, g(«) = 0.05, and b = 1/10, 1, and 10.
(For h = 3 and 4, the b = 1/10 curve is indistinguishable from the q(s) curve.)

limb bias; if h — 1 < s < h, then features of the model are noteworthy,
there is a lower limb bias with prob- of which two are really problems.
ability h — s and an upper limb one First, the biasing effects on the re-
with probability I — (h — s); and for sponse behavior that result from pay-
s > h, there is an upper limb bias. offs and presentation probabilities
Using this rule, typical functions are were treated as the asymptotic conse-
shown for h = 1, 2, 3, and 4 in Figure quences of a linear learning process,
8. For h larger than 4, the functions not as the usually assumed maximiza-
are just like those for h — 4, except tion of expected payoffs which, when
that the right hand plateau extends coupled with this threshold model,
over h — 3 quanta units. yields totally incorrect results. Sec-
ond, the dependence of the asymptotic
CONCLUSION response probabilities on the prob-
The central conclusion of this paper abilities of stimulus presentations is
is that there is at least one sensory explicit, but the dependence upon the
threshold model for simple detection payoffs is given only in terms of
experiments which is not clearly wrong learning rate parameters and so is
as judged by existing data. Four implicit. A theory relating the learn-
78 R, DUNCAN LUCE

ing parameters to payoffs is needed, BIRDSALL, T. G. The theory of signal de-

but in the meantime sequential data tectability. In H. Quastler (Ed.), In-
formation theory in psychology. Glencoe:
should be collected and the parameters Free Press, 19SS. Pp. 391-402.
estimated directly as in ordinary BUSH, R. R., & MOSTELLER, F. Stochastic
applications of stochastic learning models for learning. New York: Wiley,
theory. Third, the threshold analysis 1955.
CLARKE, F. R., BIRDSALL, T. G., & TANNER,
of the Yes-No detection experiment W. P., JR. Two types of R. O. C. curves
led to an isosensitivity curve that and definitions of parameters. J. Acousl.
consists of two distinct line segments, Soc. Amer., 1959, 31, 629-630.
but no criterion was developed about CREELMAN, C. D. Detection of signals of un-
which limb is in use during a par- certain frequency. Technical Memorandum
No. 71, 1959, University of Michigan,
ticular experimental run. This leads Electronic Defense Group.
to an ambiguity in the estimates of CREELMAN, C. D. Detection of complex
the true threshold probabilities, which signals as a function of signal bandwidth
continually proved to be a bothersome and duration. Technical Report No. 99,
problem in analyzing data and eval- 1960, University of Michigan, Electronic
Defense Group.
uating the model. The data in EGAN, J. P., SCHULMAN, A. I., & GREENBERG,
Figure 3 and in Table 2 hint at the G. Z. Operating characteristics determined
possibility that subjects may shift by binary decisions arid by ratings. J.
between the limbs when Acoust. Soc. Amer., 1959, 31, 768-773.
GREEN, D. M. Detection of multiple com-
ponent signals in noise, /. Acoust. Soc.
Amer., 1958, 30, 904-911.
On-o GREEN, D. M. Psychoacoustics and detec-
tion theory. J. Acoust. Soc. Amer., 1960,
is in the neighborhood of 1.5 or 2, 32, 1189-1203.
suggesting that it may be better to GREEN, D. M., BIRDSALL, T. G., & TANNER,
choose presentation probabilities and W. P., JR. Signal detection as a function
payoffs so that ft lies well outside this of signal intensity and duration. /. Acoust.
transition region. In this way, we can Soc. Amer., 1957, 29, 523-531.
LICKLIDER, J. C. R. Three auditory theories.
be reasonably sure whether the sub- In S. Koch (Ed.), Psychology: A study of a
ject is using a lower or upper limb bias. science. Vol. 1. New York: McGraw-Hill,
For example, if the payoff matrix is 1959. Pp. 41-144.
symmetric, one might use Ps in the LUCE, R. D. Individual choice behavior. New
York: Wiley, 1959.
neighborhood of 0.2 or 0.8, which PETERSON, W. W., BIRDSALL, T. G., & Fox,
places /3 in the neighborhood of 4 or J, W. C. The theory of signal detectab:lity.
respectively. Finally, the theoretical Trans. IRE Profess. Group Inform. Theory,
effects of this sort of biasing on the 1954, 4, 171-212.
psychometric function were shown to RESTLE, F. Psychology of judgment and
choice. New York: Wiley, 1961.
be quite striking (see Figure 8). If
SHEPARD, R. N. Stimulus and response
this, or some such, model is correct, generalization: A stochastic model relating
then fairly subtle analyses of response generalization to distance in a psychological
data are necessary in order to test space. Psychometrika, 1957, 22, 325-345.
adequately any theory of the psycho- SHEPARD, R. N. Stimulus and response
metric function, such as the neural generalization: Tests of a model relating
generalization to distance in psychological
quantum theory. space. /. exp. Psychol, 1958, 55, 509-523.
SHIPLEY, ELIZABETH F. A model for detec-
REFERENCES tion and recognition with signal uncer-
BARLOW, H. B. Comment on neural quanta. tainty. Psychometrika, 1960, 25, 273-289.
In W. A. Rosenblith (Ed.), Sensory com- SHIPLEY, ELIZABETH F. Detection and rec-
munication. New York: Wiley, 1961. ognition with uncertainty: An experiment
Pp. 786-790. and a revised model. Unpublished doc-
 A THRESHOLD THEORY FOR SIMPLE DETECTION EXPERIMENTS 79
toral dissertation, University of Pennsyl- measures. J. Acoust. Soc. Amer., 1958, 30,
vania, 1961. 922-928.
STEVENS, S. S., MORGAN, C. T., & VOLK- TANNER, W. P., JR., & NORMAN, R. Z. The
MANN, J. Theory of the neural quantum human use of information: II. Signal
in the discrimination of loudness and pitch. detection for the case of an unknown signal
Amer. J. Psychol., 1941, 54, 315-335. parameter. Trans. IRE Profess. Group
SWETS, J. A. Indices of signal detectability Inform. Theory, 1954, 4, 222-226.
obtained with various psychophysical pro- TANNER, W. P., JR., & SWETS, J. A. A
cedures. /. Acoust. Soc. Amer., 1959, 31, decision making theory of visual detection.
511-513. Psychol. Rev., 1954, 61, 401-409. (a)
SWETS, J. A. Is there a sensory threshold? TANNER, W. P., JR., & SWETS, J. A. The
Science, 1961, 134, 168-177. human use of information: I. Signal
SWETS, J. A., & BIRDSALL, T. G. The human detection for the case of the signal known
use of information: III. Decision-making exactly. Trans. IRE Profess. Group
in signal detection and recognition situa- Inform. Theory. 1954, 4, 213-221. (b)
tions involving multiple alternatives. IRE TANNER, W. P., JR., SWETS, J. A., & GREEN,
Trans. Inform. Theory, 1956, 2, 138-165. D. M. Some general properties of the hear-
SWETS, J. A., SHIPLEY, ELIZABETH F., ing mechanism. Technical Report No. 30,
McKEY, MOLLY J., & GREEN, D. M. 1956, University of Michigan, Electronic
Multiple observations of signals in noise. Defense Group.
J. Acoust. Soc. Amer., 1959, 31, 514-521. VENIAR, FLORENCE A. Effect of auditory cue
SWETS, J. A., TANNER, W. P., JR., & BIRD- on discrimination of auditory stimuli. /.
SALL, T. G. The evidence for a decision Acoust Soc. Amer., 1958, 30, 1079-1081.
making theory of visual detection. Tech- (a)
nical Report No. 40, 1955, University of VENIAR, FLORENCE A. Signal detection as a
Michigan, Electronic Defense Group. function of frequency ensemble. I. /.
SWETS, J. A., TANNER, W. P., JR., & BIRD- Acoust. Soc. Amer., 1958,30, 1020-1024. (b)
BALL, T. G. Decision processes in percep- VENIAR, FLORENCE A. Signal detection as a
tion. Psychol. Rev., 1961, 68, 301-340. function of frequency ensemble. II. /.
TANNER, W. P., JR. On the design of psycho- Acoust. Soc. Amer., 1958, 30,1075-1078. (c)
physical experiments. In H. Quastler VON. BEKESY, G. Uber das Fechner'sche
(Ed.), Information theory in psychology. Gesetz und seine Bedeutung fiir die
Glencoe: Free Press, 1955. Pp. 403-414. Theorie der akustischen Beobachtungs-
TANNER, W. P., JR. Theory of recognition. fehler und die Theorie des Horens. Ann.
J. Acoust. Soc. Amer., 1956, 28, 882-888. Phys., 1930, 7, 329-359.
TANNER, W. P., JR., & BIRDSALL, T. G.
Definition of d' and y as psychophysical (Received May 15, 1962)