simple correlational analysis was carried

Theoretical analysis of an out in order to acquire a rough idea as to

what extent the models seemed to be
alphabetic confusion matrix* consistent in different conditions and,
among themselves, to be measuring or
reflecting the same attributes within the
J. T. TOWNSEND data. One limitation, of course, in the
Purdue University,West Lafayette, Indiana 47907 present type of study is that the models
purport to explain individual behavior.
A study was undertaken to acquire a confusion matrix of the entire upper-case English Hence, the work reported here must be
alphabet with a simple nonserifed font under tachistoscopic conditions. This was viewed as testing the ability of the models
accomplished with two experimental conditions. one with blank poststimulus field and to handle a large amount of averaged
one with noisy poststimulus field, for six Ss run 650 trials each. Three mathematical human confusion data but not as providing
models of recognition, two based on the concept of a finite number of sensory states and a critical test of the models' assumptions
one being the choice model, were compared in their ability to predict the confusion that are meant to apply at the level of the
matrix after their parameters were estimated from functions of the data. In order to individual. An experiment is in progress
ascertain the facility with which estimates of similarity among the letters could lead to a collecting long-term confusion data at the
psychological space containing the letters, 'l1ij, the similarity parameter of the choice individual level.
model was input to an ordinally based multidimensional scaling program. Finally, A second major type of analysis
correlation coefficients were computed among parameters of the models, the scaled included in the present study was the use
space, and a crude measure of physical similarity. Briefly, the results were: (I) the of an ordinal-based multidimensional
finite-state model that assumed stimulus similarity (the overlap activation model) and the scaling procedure applied to estimated
choice model predicted the confusion-matrix entries about equally well in terms of a similarity parameters in the hope of
sum-of-squared deviations criterion and better than the ali-or-none activation model, learning more about the characteristics of
which assumed only a perfect perception or random-guessing state following a stimulus psychological space for letter
presentation; (2) the parts of the confusion matrix that fit best varied with the particular identification, when all letters are in the
model, and this finding was related to the models; (3) the best scaling result in terms of a stimulus population. The results of this and
goodness-of-fit measure was obtained with the blank poststimulus field condition, with a the analysis with the substantive models
technique allowing different distances for tied similarity values, and with the Euclidean as are discussed, interrelated, and compared
opposed to the city-block metric; and (4) there was agreement among the models in terms to a simple. physical measurement of
of the way in which the models reflected sensory and response bias structure in the data, interletter similarity.
and in the way in which a single model measured these attributes across experimental Apart from the hypothesis-testing
conditions, as well as agreement among similarity ami distance measures with physical function of an experiment. there exist
Similarity. aspects that perhaps should be viewed as of
an information-gathering nature. In this
The study reported in these pages sought 26 by 26 confusion matrix, it was required context, latencies and confidence ratings
to pursue several related goals. The broad of the models that they be of sufficiently were collected and the summary results
goal was establishment and investigation of simple structure that parameter estimation commented on, although the models
a tachistoscopic confusion matrix obtained be feasible. One of the few extant models employed here were not developed to the
from human observers attempting to that possesses this property without extent necessary to make predictions
identify single members of the complete simplifying assumptions is the choice concerning these.
upper-case English alphabet. Despite the model (Luce, 1959, 1963a). However,
abundance of data concerning legibility of work with finite-state detection models THEORETICAL DEVELOPMENTS
alphanumeric characters (e.g., see Cernog & (e.g.. see Atkinson & Kinchla, 1965; Luce, It was of interest in the present study to
Rose, (967), some of the results are of 1963b) also suggested the possibility of compare two models selected from a class
marginal significance to the psychologist extension of this type of model to more of finite-state models with the choice
because (1) of the highly specific applied complex situations. Because of the very model (Luce, 1963a). and a "threshold"
nature of the study, (2) too few trials or explicit nature of assumptions of these finite-state model with models postulating
subjects were run. (3) only a few of the models about the sensory and bias sensory confusion. The class of finite-state
letters of the alphabet were used, or (4) no processes. they can be helpful in models from which the two representatives
attempt was made to use mathematical investigation of perceptual characteristics were selected was motivated by the simple
models to separate the response-bias by manipulation of these assumptions, To finite-state models used in some detection
characteristics from the stimulus or sensory this end. two generalizations of the experiments (Atkinson & Kinchla, 1965;
characteristics. This study was intended to activation model suitable for application in Kinchla, Townsend. Yellott, & Atkinson.
complement the literature by rectification the present experiment were developed. As 1966) and multisymbol recognition
of these limitations. will be seen, one of these is like the choice situations (Townsend, 1966). In order to
Within the above-stated broad goal was model in that it possesses structure for provide continuity with the intuitions and
the subsidiary aim of examining the ability description of psychological similarity, but structure developed earlier. we will refer 10
of three simple mathematical models to the other assumes only two psychological the finite class of models as the general
predict average behavior in a complete sc nsor y states relating to perfect activation model.
identification confusion study. Because of information or no information at all. It was
the large array of data represented by the of interest to learn if better predictions are General Activation Model
obtained with models that assume sensory The nature of the activation model in
confusion states as opposed to the our development supposes a mapping of
'The data were gathered and part of the
analysis of this study was accomplished at the two-state type of model just mentioned as the set of N possible stimulus events to a
Univerxity of Hawaii with the aid of a University well as to compare the choice model with (finite) set of n hypothetical internal
of Hawaii intramural research grant. the activation representative. In addition, a sensory states. On the basis of the

40 Copyright 1971, PsychonomicJournals.Lnc.. Austill, Texas Perception & Psychophysics, 1971, Vol. 9 (lA)
particular internal state activated, S is (k ) Hence, s, through S26 represent "certain"
assumed to make a decision as to the peS) sensory states since P(Rilsj) = I; but in the
stimulus event, thus yielding a mapping of k (k l "uncertain" state, so. S is assumed to guess
the n internal states to the set of N possible L P(Sj) Si with probability Pi. which may depend
responses. For the complete identification j= 1 on learning and motivation variables.
experiment with N stimuli. the sensory (k) Multiplying these two matrices yields the
activation matrix would appear as S = (ajj) In this expression, P(Si) represents the theoretical confusion matrix C '" (Cij)
where i = I, 2,' ., Nand j = 0, I. 2, •• , n. a priori probability of stimulus Si and the where i = I, 2, ". 26, j = I, 2,.. •• 26. and
The second subscript refers to the superscript merely indicates that the Cjj = ajo ij + (J ~. aj)pj where
hypothetical sensory states so, s\ , ••• , Sn' stimulus is a member of the k-stimuli
subset of equal-density stimuli at the
Similarly. we may write the decision- or 0"= /1 wheni=j'l
bias-process matrix as D = (bi.j), where observed point. Under the expressed II 0 otherwise.
i=0.1.2,···.n and j=I.2... ·,N. circumstance this above quantity is both
Finally. we obtain the general theoretical the Bayesian probability (hence the term A few comments are in order concerning
confusion matrix by premultiplying D by "B ay esian probability matching") of related work on this model. For the
S: stimulus Sj given the particular observed two-signal case. when a, = a2, this model
point and the conditional probability of Sj reduces to that developed for two-interval
given the stimulus must be in the k-stimuli forced-choice detection experiments by
subset. We will not attempt to distinguish Atkinson and Kinchla (1965) and by
such theories from the finite-state theories Kinchla and associates (1966). Smith
here. (1968) has studied properties of confusion
where i = 1,2, ... , Nand k= 1,2. -, N, In order to obtain tractable special cases matrices generated by this mode, which he
where. in general, the only constraints are and to investigate the relative importance terms a "pure perceptibility" model, and
that of pairwise sensory confusions as opposed he has shown that it can be viewed as a
to "pure" guessing (where S is in a state of special case of the choice model; this result
n N complete uncertainty regarding the will be mentioned again after the brief
L ai' = 1 and that L b~ = I. presented stimulus), we develop the presentation of the choice model.
j=O I j= I
all-or-none and overlap activation models The all-or-none activation model
for application to letter-recognition requires estimation of 2N - I parameters,
The interpretation of the internal sensory experiments. 51 in the present experiment.
states we wish to make in this paper is in
terms of sensory confusions among the All-or-None Activation Model
various stimuli. For example, depending on The first specialization posits that S Overlap Activation Model
the actual set of stimuli, it might be either obtains such information from the The basic intuition for the overlap
possible that S is certain that the presented display as allows him to respond perfectly activation model is that in addition to a
stimulus is one of a particular subset of or he is thrown into an uncertain state unique activation state relating a letter in a
stimuli (the cardinality possibly being where he has no partial information and one-one fashion to the correct response.
equal to zero or to the number of possible must respond by guessing. Note that S there exist partial-information states such
stimuli) and hence, S must choose from could be in this uncertain state either that S is in a state of uncertainty with
this subset on the basis of learning or because he fails to detect anything at all or respect to two letters; thus. this model
motivational factors. because such characteristics of the letter as assumes that pairwise similarity can affect
It should be noted that one can devise he observes do not aid in identifying it. If interletter confusability. When in a state of
con tinuous-state models that possess we index the letters of the alphabet by the confusion, S is postulated to respond
properties that reasonably mimic numbers I through 26 and refer to S. as according to the relative magnitude of the
finite-state models. For example, if the stimulus i (i.e .. the it h letter of the two concerned response-bias probabilities.
sensory results of each stimulus alphabet) and R1 as response j, we may We thus express the sensory activation
presentation could be described by a write the appropriate sensory activation matrix as composed of entries for the pairs
uniform distribution, then there would matrix as S =' ai.,) where i = I, 2,' , 26, (S" tLk)' where i = 1,2, .' ,26,
exist, in general, areas of nonoverlap of j = 0, I, 2, ..• , 2/;' and where j = 1,2, • ,26, and k= i.i + I. ",26.
probability density; if an observation and where the entry for (S], tj.k) is given
occurred in such an area, a reasonable
strategy would be for S to respond with aii =
1 a, when j = 0,
at when j = i,
1 by

the associated response, with probability

one. On the other hand, if an observation
1 o otherwise,
~jk = I
~Jk when j = i, k = i, or both, I
were obtained at a point where, say. k where, as before, j refers to sensory state si' I 0 otherwise.
stimuli gave rise to equal a priori densities, These sensory states are related to the set
S might well guess on the basis of his of overt responses by the decision matrix
knowledge of the presentation D = (b i , ), where i = 0, I, "',26, Note that ti,i is the certain state for Sj, and
probabilities. For example, he might j = I, 2, ..• , 26, and tt,j is the uncertain state for Sj and Sj, i =1= j,
choose the signal with the highest a priori and can proceed from presentation of
or a posteriori probability (which either s, or s, (~ij = ~jj). The corresponding
maximizes average number correct), or he decision matrix is given by entries

I
might employ a Bayesian probability corresponding to the pairs (tj,b Rill) where
matching strategy. The second strategy Pj when i = 0 ) j.k are as before and m = I, 2, •• '. 26, The
would occur when S chooses, say, S. from bil '" I when i = j, entry for ('I,k, Rill) is equal to one when
the k alternatives with probability o otherwise. j = k = m, to gl/(gj + gk) when j = m =1= k, to
gk/(gj + gk) when k = m *- j, and to 0

Perception & Psychophysics, 1971, Vol. 9 (IA) 41

otherwise, where gi is the response bias response, as given by multiplying the scale constant for constant stimulus and
parameter for Rj. values of the bias and similarity sensitivity conditions.
The theoretical confusion matrix, given parameters, relative to the total strength Similarly, a mapping from the
as the result of the two foregoing processes relating the given stimulus to other stimuli. parameters of the choice model to the
is written C = (Ci,j) where i = 1, 2,"", 26, For the present study, this results in the overlap model or vice versa may be
j = I, 2," -, 26, theoretical confusion matrix C = (Cij), provided. Under some circumstances, this
where i=I,2,"',26, j=I,2,"',26, may imply nontestability of the two
and models against one another. However, the
models do not predict identical
1)jj ~j relationships in the data, and in certain
and
c··
1)
= 26 cases data may support one and falsify the
~ 1)jk ~k
other. For example, suppose that
k=l 1)ij = 1)kl = 1), for i *- j and k *- I, and that

Cjj = ~jj (gj; g)when i *- j. In addition, the similarity parameters are

{3i = (3j =~, for all i, j. The mapping from
the choice model parameters to the overlap
symmetric (1)ij = 1)jj), and it is assumed that model parameters is then given by
Now, it is clear that this model 1)ii= I for all i, thus fixing the unit. As
represents a deviation from the all-or-none indicated above, this results in ~jj = peRi I Sj) + P(Rj lSi)
model since not only is stimulus confusion (N(N + 1)/2] - I (350) free parameters
allowed, but a pure, no-information, to be estimated from the data. 21)
guessing state is precluded. It is also a For our purposes, the most important
one-step extension in the sense that we part of Smith's (1968) result relating the
could have postulated the existence of threshold to the choice model can be 1-(N-I)7)
k-wise confusion states. In general, this last shown readily by viewing the relationship ~jj = I + (N - 1)7)'
possibility generates an unmanageable that is assumed to obtain between the
number of parameters. However, if one is confusion matrix probabilities and the where N is again the number of
will ing to posit further theoretical similarity parameters: experimental stimuli. But, for
structure, then the number of parameters 1) > I/(N -- I), a reasonable possibility for
may be reduced. This approach was taken P(Rj I Sj)P(Rj lSi) confusable stimuli, ~ij < 0, which is
by Nakatani (1968) in the development of 1)jj = peRi I Sj)P(Rj I Sj) . contradictory to the overlap model.
a hybrid model that results in a set of Presently, we know of no corresponding
confusion states, where from two to all N theoretical results against the choice model
If we now substitute the theoretical
objects may be confused. Although when we map from the overlap to the
expressions for P(Rj I Sj) from the
Nakatani's work was independent of the choice model. Thus, it may be that the
threshold model, we obtain
present results, both have employed the
choice model is the more general of the
expedient of using the ratios of the
two despite the almost equal number of
relevant bias, for example, gi/(gj + gj), to parameters.
describe the S's guessing bias for a
particular confusion state.
Multidimensional Scaling Model
The overlap model is favored by (I - 0j)Pj (l-oj)Pi
Although the preceding substantive
properties relevant to the present aims:
OJ + (I -- 0j)Pj OJ + (1- OJ)Pj' models possess sufficient structure to
(I) It allows comparison of a two-way
describe interletter (sensory) confusability
confusion activation model with an
and this shows that for the threshold and one (choice model) posits a function
activation model that allows only N-way
model and the choice model to hold for a relating the similarity scale to distance in a
confusion or no confusion, and (2) it
given set of data, 1)ij must be factorable metric space, they cannot be used facilely
possesses just one more parameter,
into independent numbers associated with to test assumptions about the appropriate
(N(N + 1)]/2, than the choice model,
Sj and Sj. Hence, although it can be metric or the number of dimensions in a
which allows for a fair comparison between
developed from entirely distinct hypothesized underlying visual space. To
them for the large amount of data reported
conceptions, the threshold model can be this end, a program developed by Kruskal
herein. Hence, for the present experiment,
viewed as a special case, when one neglects (1964a, b) was employed to obtain
351 parameters were estimated for the
underlying assumptions giving rise to the information about possible dimensions of
overlap activation model and 350 for the
parameters, of the choice model. To see confusion, as well as to test applicability of
choice model.
that they are, nevertheless, not the same the Euclidean vs the city-block metric to
Choice Model theory, it suffices to note that for the an interletter confusion matrix. Kruskal's
The ch oice model requires less two-signal case, the threshold model method, an extension of Shepard's ordinal
discussion, being a straightforward generates linear and the choice model scaling procedure (Shepard, 1962a, b),
application of that model as presented by curvilinear isosensitivity curves. Or, to put permitted this to be accomplished with a
Luce (1963a). As does the overlap it another way, if the threshold model minimum of substantive assumptions.
activation model, this model postulates an provided good fits to several sets of data
influence of pairwise similarity on stimulus generated by varying motivational METHOD
confusions. Both the similarity parameters conditions, then only the Pi parameters Apparatus
and bias parameters are assumed to lie on would change; but by Smith's result, this A Gerbrands two-field, two-mode
ratio scales. Given a particular stimulus, the would perturb not only the bias parameter tachistoscope (Model T-2B-I) was used to
probability of any possible response is (3j, but also the similarity parameter 1)jj present the letters. A simple mechanical
assumed to be the strength of that which in terms of the theory should remain shutter was employed to increase the

42 Perception & Psychophysics, 1971, Vol. 9 (IA)

number of independent fields to three. stimuJus duration was selected for each S the best estimated overall value of the zero
Stimuli were presented on 8 x II in. such that his overall probability correct entries. The rationale behind the
white index cards, using an IBM Executive was about 0.5, that is, at threshold. The convention was that if the "true"
Directrix typewriter, one letter per card, selected duration was then used for that S confusion probabilities were greater than
with the stimulus population consisting of the remainder of CI. Following CI, I day 0.0 I, then one or more confusions would
one deck made up of five English was taken to recalibrate Ss for CII. have occurred on the average in the 150
upper-ease alphabets in a quasirandom (i.e., However, performance with the trials represented in each row of the
shuffled) order. A prestimulus fixation poststimulus noise proved much more confusion matrix. On the other hand, if the
point was placed approximately 10 min variable; therefore, each S was recalibrated probability of confusion were less than
below the locus of the letter presented in every day during the five experimental 0.01 but greater than zero, one would
the stimulus field. A singleletter subtended sessions. Both conditions included 20 expect to observe cases where the entry for
an angle of about 30 min and the fixation warm-up trials at the beginning of each 150 trials turned out to be zero. Thus, in a
point an angle of about 6 min at S's eye. session. sense, one extra parameter was estimated
The possible visual fields were white with Each experimental condition consisted for each of the models. The calculated sum
fixation point, white with letter, and noise. of calibration followed by five of squared deviations varied between the
The noise field was produced by typing experimental sessions, each session obtained minimum and about 1.50 for the
alternately with "normal" and "expanded" occupying approximately I h during which overlap and choice models but varied only
spacing modes on the Executive Directrix 130 triaJs of five pseudorandomized a few I OOths for the threshold model. As it
typewriter, and successive rows of letters (shuffled) alphabets were presented. Each turned out, when estimates were also
were made to overlap. The characters row of the resulting confusion matrices obtained for "zero" confusion probabilities
typed were selected haphazardly from the thus contained J50 points for both between 0.01 and 1.00, in only one case
upper-case English alphabet. Use of the conditions. The amount of data per was the sum of squared deviations reduced
resulting noise field made perception of individual was insufficient to estimate and that by only 0.02; this occurred for
presented letters much more difficult. stability but the group-average confusion the choice model in CL
The luminance for CI (Without noise), matrices appeared quite constant from the It should be noted that in addition to
the first experiment, was at all times first half to the second half of the the analytical difficulties associated with
5.6 fl., and for CII (with noise)i the second conditions. All the models assume division by zero, the scaling properties of
experiment, the luminance was 5.6 fl, for steady-state behavior is generating the Luce's choice theory require stronger
the prestimulus and stimulus fields and for response proportions. assumptions when the set of choice
the noise field was 4.0 fl. The study required 13 consecutive days, alternatives consists of some perfectly and
including two weekends. A S was run at some imperfectly discriminable pairs
Procedure the same time each day. (Luce, 1959). The overlap activation model
The stimulus exposure sequence for CI Estimation and the choice model required estimation
was: white prestimulus field with fixation In all three models, estimation of the of 351 and 350 parameters, respectively,
point; stimulus field containing a randomly sensory and bias parameters was effected and the all-or-none activation model
selected letter; poststimulus field by setting each parameter equal to a required estimation of 51 parameters. The
containing fixation point. The sequence for function of confusion-matrix values number of degrees of freedom associated
CII was: white prestimulus field; stimulus associated with the parameter within a with the empirical confusion matrix was
field containing a randomly selected letter; particular theory. The exact formulae used 650.
noise field with jumbled letters; white are given in the appendix. Subjects
poststimulus field containing fixation The stimulus durations, chosen to yield Six females recruited from introductory
point. probability-correct values close to 0.50, psychology classes at the University of
The trials were S-paced, S initiating a were large enough to allow zero probability Hawaii were employed as Ss. Their vision
stimulus exposure following an alerting of confusion entries in the confusion was required to be 20/20 after correction.
buzzer sounded by E. The displayed letter matrix. This caused a problem in
followed the press of the start button by estimation since the functions used to RESULTS AND DISCUSSION
I sec. The only instruction to S with regard estimate the parameters often contained Tables I and 2 show the empirical
to his response speed was that a letter observed (P(Ri/Sj) terms in the confusion matrices for the two
response was to be made within 5 sec after denominator. In order to allow estimation experimental conditions. Table 3 shows
the stimulus exposure, after which he was of all relevant parameters, a convention predictions from the three models
to rate his sensory accuracy by giving a was uniformly adopted across the models. compared with the empirical values for the
confidence rating (CR) of I to 4. Ss were Since each row in the confusion matrix was letter "p"-a symbol possessing curved- as
instructed that a CR of I meant absolute based on 150 trials, it was decided to take well as straight-line features. The table
certainty, 2 relative certainty, 3 relative the best estimate of the "zero" confusion illustrates the superior ability of the
uncertainty, and 4 absolute uncertainty, probabilities (in terms of sum of squared overlap and choice models to reflect
i.e., guessing at random. Following S's deviations of predicted from observed similarity as well as the tendency to altered
response, E told him the letter that had values) lying between 0.000001 and 0.01. structure from CI to CIl.1 Although the
been presented. Each trial consumed about That is, the computer performed a series of predictions do seem to follow the general
15 sec, with approximately 5 sec iterations, and on each iteration exactly pattern of the confusions in the empirical
intervening between feedback and the next one value in the above range was tried for matrix. a numerical index of precision may
buzzer sound. Latencies were recorded by all the zero confusion entries. For each be helpful in interpretation. An idea of
E on each trial from a Hunter timer linked such value, the set of parameters was relative accuracy of the three models can
to a voice-operated relay and a microphone estimated for the model at hand and the be gained from Table 4, which indicates
positioned close to S's mouth. sum of squared deviations computed; that the sum of squared deviations of predicted
Two days of practice and calibration value associated with the smallest sum from empirical points over all 676 cells for
preceded 5 experimentaJ days in CJ. A of squared dcviations was chosen as each confusion matrix.

Perception & Psychophysics, 1971, Vol. 9 (IA) 43

 Tahle 1
_ _ _ _ _ _ _ _E=I:.:.:nr:piric~.L.!-·onfusi().!!_Matrix: Condition I

Response
ABC D E F G H J K L M N 0 P Q R STU V W X Y Z

.58 .00 .00 .01 .00 .01 .00 .02 .02 .03 .10 .02 .00 .02 .01 .02 .00 .05 .00 .04 .00 .01 .00 .06 .00 .01

B .02 .26 .02 .05 .02 .00 .04 .07 .01 .02 .01 .01 .03 .08 .03 .03 .03 .18 .01 .01 .06 .01 .00 .01 .00 .00

C .01 .01 .50 .01 .05 .03 .03 .02 .01 .01 .03 .04 .00 .00 .07 .03 .01 .02 .01 .04 .01 .01 .01 .01 .01 .01

D .01 .04 .01 .46 .01 .00 .05 .04 .01 .01 .00 .00 .01 .01 .12 .02 .05 .05 .00 .02 .06 .01 .00 .01 .01 .00

E .01 .03 .03 .01 .36 .07 .00 .03 .03 .04 .06 .11 .00 .00 .03 .05 .00 .01 .00 .07 .01 .01 .00 .01 .03 .01

F .00 .01 .00 .01 .02 .33 .01 .03 .09 .05 .03 .07 .01 .01 .00 .03 .00 .01 .01 .18 .01 .01 .00 .01 .07 .01

G .01 .01 .08 .01 .01 .01 .34 .09 .01 .01 .02 .03 .00 .04 .11 .03 .03 .03 .01 .01 .07 .02 .00 .00 .01 .00

H .01 .01 .01 .OJ .00 .01 .01 .50 .01 .01 .01 .01 .03 .15 .03 .02 .00 .04 .00 .03 .04 .01 .03 .00 .01 .00

.00 .00 .00 .00 .01 .03 .01 .01 .57 .08 .01 .11 .00 .00 .01 .01 .00 .02 .00 .09 .00 .01 .00 .00 .03 .00

J .01 .00 .00 .OU .00 .01 .01 .02 .15 .48 .01 .03 .00 .00 .03 .01 .00 .01 .01 .08 .04 .01 .00 .02 .05 .01

K .03 .01 .01 .OJ .01 .02 .00 .06 ,1")5 .01 .50 .04 .01 .02 .04 .03 .00 .03 .00 .01 .01 .01 .01 .05 .03 .00

L .01 .00 .01 .01 .01 .01 .00 .03 .14 .03 .02 .60 .00 .01 .03 .01 .00 .nl .00 .07 .00 .00 OJ .00 .02 .00
M .00 .01 .00 .01 .00 .00 .01 .10 .00 .01 .01 .00 .1i2 .08 .05 .01 .01 .01 .01 .01 .01 .01 .05 .01 .00 .00

N .03 .01 .00 .01 .00 .00 .01 .06 .00 .00 .03 .00 .07 .54 .03 .01 .03 .04 .01 .01 .02 .02 .04 .01 .01 .01

o .01 .01 .06 .05 .00 .01 .11 .01 .oo .02 .01 .00 .01 .03 .51 .02 .10 .01 .00 .02 0') .00 .01 .00 .00 .01

P .01 .02 .02 .01 .01 .09 .01 .05 .03 .01 .01 .01 .01 .01 .06 .52 .00 .06 .00 .03 .01 .01 .00 .01 .01 .01

Q .01 .01 .01 .01 .00 .00 .11 .06 .00 .01 .01 .01 .00 .03 .28 .01 .36 .01 .00 .00 .05 .00 .01 .01 .00 .00

R .00 .04 .01 .00 .01 .01 .01 .09 .02 .01 .03 .03 .02 .05 .03 .05 .00 .49 .01 .02 .01 .01 .01 .03 .01 .00

S .01 .01 .03 .01 .02 .00 .02 .06 .03 .03 .06 .03 .00 .03 .04 .05 .00 .03 .43 .03 .01 .02 .00 .03 .01 .02

T .01 .00 .00 .01 .04 .05 .01 .01 .16 .04 .01 .05 .01 .01 .04 .04 .00 .00 .01 .42 .01 .OJ .01 .01 .05 .01

U .00 .00 .01 .01 .00 .00 .01 .11 .01 .00 .01 .02 .02 .05 .07 .01 .01 .01 .00 .02 .)5 .04 .04 .01 .00 .01

v .00 .00 .00 .01 .01 .00 .00 .03 .02 .02 .01 .01 .00 .01 .02 .02 .01 .01 .00 .02 .07 .60 .02 .01 .09 .01

W .01 .00 .00 .01 .00 .00 .00 .08 .00 .01 .02 .01 .05 .07 .05 .01 .01 .02 .00 .02 .05 .08 .45 .03 .01 .01

x .01 .00 .00 .00 .01 .00 .00 .03 .02 .01 .07 .01 .00 .04 .02 .01 .00 .00 .00 .05 .00 .03 .01 .55 .08 .04

Y .01 .00 .00 .00 .00 .01 .00 .05 .02 .03 .03 .03 .00 .03 .06 .01 .00 .00 .01 .06 .01 .08 .01 .03 .51 .01

Z .01 .01 .01 .01 .01 .00 .00 .03 .01 .06 .03 .00 .00 .00 .02 .00 .01 .00 .01 .03 .01 .01 .01 .05 .03 .66
-- _._------------- ------------

In order to acquire some intuition on

the size of these errors, we may compare
26 (
L
1)2 =
Pij - -26-
26
L Pij - I.
main diagonal. This table can be used to
discuss several points. First of all, the
the obtained values in Table 4 with a 1,)= I 1,)= I choice model and the overlap model are
prediction following from random about equal in their ability to fit the data
prediction (i.e., on the average, uniform or This index is largest when one value on for CI, although the overlap model appears
equiprobable confusions and probability each row is I and the rest are 0; it then a little more accurate in CII. Secondly, it
correct) on the part of a model. That is, all attains the value of 25. Table 4 indicates appears that the large increment in number
confusion entries are simply 1/26. If we let the sum of deviations squared, as well as of parameters (in fact, a factor of seven)
Pij refer to entries in the empirical this (equiprobable) index, for the two for the two activation models permitting
confusion matrix, then the sum of squared conditions. Also, it shows the sum of sensory confusion does not seem to
deviations under random prediction would squared deviations for the main diagonal purchase a corresponding increase in fit,
be only and the index figured only on the relative to the calculated index. In fact, it

44 Perception & Psychophysics, 1971, Vol. 9 (IA)

 Table 2
Empirical Confusion Matrix: Condition 2
-3'" Response
JAB C D E F G H I J K L M N 0 P Q R STU V W X Y Z
Vi
A .83 .01 .00 .01 .00 .01 .00 .01 .01 .01 .01 .02 .00 .00 .02 .00 .00 .00 .02 .01 .00 .03 .00 .00 .00 .01

B .02 .36 .01 .05 .01 .04 .03 .07 .02 .03 .01 .03 .01 .02 .10 .05 .00 .01 .02 .04 .01 .03 .01 .01 .01 .01

C .00 .00 .79 .01 .00 .01 .03 .02 .01 .01 .00 .01 .01 .00 .05 .01 .00 .01 .02 .00 .00 .01 .00 .01 .01 .00

D .01 .01 .01 .73 .01 .01 .00 .03 .00 .01 .00 .00 .00 .01 .11 .02 .03 .01 .01 .00 .01 .01 .00 .01 .00 .01

E .03 .01 .03 .00 .43 .09 .01 .01 .05 .01 .03 .09 .01 .00 .06 .01 .00 .01 .03 .03 .00 .01 .00 .02 .03 .01

F .05 .01 .02 .03 .05 .42 .01 .05 .04 .01 .01 .05 .01 .00 .05 .03 .01 .00 .01 .07 .01 .03 .00 .01 .04 .01

G .00 .00 .11 .01 .01 .01 .57 .04 .00 .01 .01 .00 .00 .01 .12 .02 .01 .01 .00 .00 .03 .03 .00 .00 .00 .01

H .05 .01 .00 .03 .03 .05 .03 .19 .09 .01 .01 .05 .03 .03 .10 .03 .01 .02 .02 .05 .05 .07 .00 .01 .01 .03

I .05 .01 .01 .02 .01 .03 .01 .07 .38 .03 .00 .12 .00 .01 .07 .01 .00 .01 .01 .05 .01 .05 .00 .00 .02 .01

J .03 .00 .00 .02 .01 .03 .01 .03 .08 .49 .01 .03 .00 .00 .06 .01 .01 .00 .02 .05 .01 .06 .01 .00 .01 .02

K .03 .00 .00 .00 .00 .03 .00 .05 .01 .01 .61 .05 .01 .00 .04 .01 .00 .01 .00 .02 .01 .01 .00 .06 .03 .02

L .03 .01 .03 .02 .02 .02 .01 .03 .04 .00 .02 .59 .02 .00 .07 .01 .00 .00 .01 .03 .01 .00 .01 .00 .02 .00

M .09 .01 .01 .01 .01 .01 .04 .06 .04 .03 .01 .03 .41 .05 .08 .02 .00 .01 .00 .00 .00 .05 .01 .01 .01 .03

N .• 04 .00 .01 .02 .01 .02 .01 .05 .05 .01 .01 .05 .01 .39 .09 .01 .01 .01 .01 .07 .01 .05 .02 .00 .03 .00

o .01 .00 .01 .02 .00 .01 .07 .01 .01 .00 .00 .01 .00 .00 .72 .02 .06 .00 .02 .01 .00 .01 .00 .01 .00 .01

P .00 .02 .01 .02 .01 .01 .00 .02 .03 .01 .01 .02 .00 .00 .05 .67 .01 .02 .03 .01 .01 .01 .00 .01 .02 .01

Q .02 .00 .03 .00 .00 .02 .02 .01 .01 .00 .00 .00 .00 .00 .32 .01 .52 .01 .03 .01 .00 .00 .00 .00 .00 .01

R .02 .01 .01 .03 .00 .01 .00 .08 .01 .00 .01 .03 .01 .00 .03 .13 .01 .52 .01 .03 .01 .02 .00 .00 .03 .01

S .03 .01 .01 .01 .00 .00 .03 .07 .01 .01 .00 .01 .00 .00 .01 .00 .00 .01 .78 .01 .00 .01 .00 .00 .00 .01

T .01 .00 .01 .01 .01 .OJ .01 .01 .08 .03 .01 .05 .00 .01 .OJ .01 .00 .01 .01 .58 .01 .03 .00 .00 .03 .02

U .03 .02 .01 .01 .01 .00 .01 .OJ .05 .02 .02 .01 .01 .00 .01 .OJ .00 .00 .04 .01 .53 .05 .00 .02 .01 .01

v .02.01.01.01.01.01.00.01.02.00.00.01.01.01.05.01.00. 00 .01 .01 .00 .69 .01 .02 .05 .01

W .05 .01 .01 .01 .00 .OJ .01 .05 .01 .02 .00 .02 .02 .03 .09 .04 .00 .01 .OJ .02 .OJ .37 .09 .03 .OJ .01

x .03 .01 .00 .00 .00 .01 .01 .06 .03 .00 .06 .04 .00 .00 .03 .01 .00 .00 .01 .OJ .00 .07 .01 .47 .09 .OJ

Y .01 .00 .01 .01 .01 .01 .00 .04 .03 .00 .02 .04 .01 .01 .07 .01 .00 .02 .03 .04 .00 .22 .00 .01 .36 .03

Z .01 .00 .00 .00 .00 .01 .00 .00 .01 .01 .00 .01 .01 .00 .01 .01 .00 .00 .01 .01 .00 .01 .00 .00 .03 .88

is possible to estimate how much the theoretically assigned parameters account the all-or-none with the off-diagonal cells,
choice and overlap model contribute with for approximately half of the error and the choice model distributed the error
their extra 199 parameters via their reduction (.44 for CI, .49 for CII) between the diagonal and off-diagonal cells
similarity structure as compared with free produced by this discounting of the 199 more than the other two models. One
assignment of parameters to those 199 cells largest deviations. As might also be reason the all-or-none activation model
in the threshold confusion matrix expected from Table 4, the off-diagonal performs as well as it does may therefore
associated with the largest deviations from entries were disproportionately represented be due to its potential for fitting the
the empirical matrix. That is, we simply in the lists of 199 largest squared diagonal values associated as they are with
insert the observed confusion probabilities deviations for the threshold model. Thus, greater possibility of variation.
in those specified cells, recompute the sum the all-or-none model may be more This last result suggests that an
of squared deviations, and contrast this efficient in this sense. Thirdly, it is act iva tion model that included the
number with the sum of squared deviations apparent that the overlap model found ali-or-none and the overlap might describe
for the other two models. We find that the greatest difficulty on the main diagonal, much of the structure of the data:

Perception & Psychophysics, 1971, Vol. 9 (IA) 45

 Table 3
. ._ .__ ._. ._ _.__ 1!~J>r_e~I.1I:Itive 1"heor~tic:11 Predictions
Condition I
Response
A BCD E F G R I J K 1 M N 0 P Q R s T u V W x y Z

EHP .01 .02 .02 .01 .01 .09 .01 .05 .03 .01 .01 .01 .01 .01 .06 .52 .00 .06 .00 .03 .01 .01 .00 .01 .01 .01

AON .02 .02 .02 .02 .02 .02 .02 .04 .02 .02 .02 .02 .02 .02 .02 .53 .02 .02 .02 .02 .02 .02 .02 .02 .02 .02
p
OlP .01 .02 .02 .01 .02 .04 .01 .04 .02 .01 .02 .01 .01 .01 .05 .56 .00 .05 .01 .04 .01 .02 .00 .01 .01 .00

eRe .01 .01 .02 .01 .02 .03 .01 .05 .02 .01 .02 .01 .01 .01 .05 .57 .00 .06 .00 .05 .01 .01 .00 .01 .01 .00
Condition 2
Response

P
I: .00 .02 .01 .02 .01 .01 .00 .02 .03 .01 .01 .02 .00 .00 .OS .67 .01 .02 .03 .01 .01 .01 .00 .01 .02 .01

.01 .01 .01 .01 .01 .01 .01 .02 .01 .01 .01 .01 .01 .01 .04 .68 .01 .01 .01 .01 .01 .01 .01 .01 .01 .01

OLP .00 .02 .01 .02 .01 .02 .01 .02 .02 .00 .01 .02 .00 .00 .05 .64 .01 .05 .02 .01 .01 .01 .01 .00 .01 .01

eRe .00 .01 .01 .02 .01 .02 .00 .03 .02 .00 .01 .02 .00 .00 .06 .70 .00 .03 .00 .01 .01 .02 .00 .00 .01 .01

EMP = empirical. AD}\' = all-ot-none activation model. DLP = overlap activation model, Cllt: = chotec model

Table 4 the probability of the certainty states vs Figures Ia and Ib show the curves for a
Sum of Squared Deviations of Theoretical the uniqueness of the probabilities of the fairly typical S relating, in the first part,
from Empirical Points sensory confusion states. Too, they would probability correct to latency and, in the
Model perhaps aid in the interpretation of the second, probability correct to confidence
Condi- All-or- Equi-
significance of the number of parameters rating. As the figures indicate, there is a
tion Choice None Overlap probable per se. For example, suppose the first continuous decrease in probability correct
subcase above fit no better than the as latency increases for both conditions.
Entire Confusion Matrix
CI .23 .49 .26 6.97
ali-or-none model and the second subcase Such a relationship between accuracy and
CII .36 .49 .23 8.14 fit as well as but not better than the latency has also been found to hold for
Main Diagonal
overlap model. This would provide an multisymbol processing tasks (Estes &
CI .07 .02 .18 5.98 indication that the number of parameters Wessel, 1967; Townsend, 1966), and
CII .14 .06 .18 7.84 might have more influence than the suggests that the phenomenon may be
particular structure of the models. Another connected with the same mechanisms in
possibility suggested by Smith (1968) both cases, perhaps with the processing
gj might be to fit probability mixtures of any characteristics of individual symbols. Other
P(R(S) = 0' -- two activation models or even an activation characteristics, shown in Fig. I a, are the
) I I) gj + gj
model and the choice model, i.e., longer mean latency and standard
aMI + (I - a)MII, where MI refers to deviations for CI than for CII. This is
+ (1- ~
k=1
Qik) gj' i oF j,
Model I and Mil to Model II and a is less
than 1 and greater than O. To the extent
probably connnected with the fact that
after recalibration on CII, the Ss' average
that the sum of errors squared is performance was better than on CI. To be
and independent of the value of a are the two sure, performance actually decreased in CII
N gi models explaining the same aspects of the on letters H, I, M, N, W, X, and Y,
~(R(S) =0·II + L ok--
koFi I gi + gk
data. Estimation techniques and computer indicating again a possible shift in
I I
programs are currently being developed to confusion structure, perhaps due to the
allow fits of these subcases to the data. presence of the poststimulus noise mask.
One interesting outcome was that ell Only one S did not exhibit this pattern; he
was not better fit by the ali-or-none model had a larger mean in CI but a smaller
than was CI. The absence of fading standard deviation.
afterimages undegraded by noise docs not Figure Ib shows the typical relationship
Two subcases of this model of interest seem to have diminished the sensory of probability correct to confidence rating.
could be extruded by (I) Jetting all 0jk be confusability of the stimuli. It appears that Although this result has been taken as
equal but allowing the Oji to be different the combination of the increase of stimulus indicative of the continuity of perceptual
(this case would have just one more duration necessary for adequate states (as, in fact, it may be), Krantz
parameter than the all-Of-none model), or performance in CII and the effects of the ( I 969) and Wickelgren ( 1968) have
(2) letting all the Ojj be equal but allowing mask did not radically affect the processing pointed out, in essence, that one must
the 0ij to be different. This case would characteristics of the Ss, at least with consider the S's processes that evaluate his
have just one more parameter than the reference to the fitting ability of the perceptual states and lead to the
choice models and the same as the overlap models tested here. It does appear that the distribution of confidence ratings found in
model. These cases should give some idea choice model predicts less successfully in the data.
of the relative importance of uniqueness of CII, although it is not clear why. In addition to acting as parameters of

46 Perception & Psychophysics, 1971, Vol. 9 (IA)

 1.0 letters having "tied" similarity scores to .35
0 52 CI have different distances in psychological 6----6 R' I SECONDARY

.9 space without penalty in terms of stress. .-- R' 2 SECONDARY

t::. 52 C2 The other, the secondary approach,
_ R" PRtMARY

t; .8 augments the stress value when tied pairs

.30 c>--<> R' 2 PRJMARY

l&J
a::
a::
.7 of letters fail to be associated with equal
distance. 25
0
u .6 As can be seen in Figs. 2a and 2b, in

...
>- .5
:J .4
general the Euclidean metric is superior to
the city-block metric, the primary en .20
en
approach is superior to the secondary w
iii approach, and CI is more easily fit into a a:
r-
~ .3
s multidimensional space than is CII. The en .15
exception to this is CII, r = I, which,
Q,. .2 except for spaces spanning three and IWO
.1 dimensions, is associated with lower stress .10
values than is CIl, r = 2. The reason for this
exception is not clear at present.
I 2 3 4 5 6 Each of the eight stress curves in fig. 2a
reaches Kruskal's (1964a, b) criterion of
RESPONSE LATENCY .10 in less than 10 dimensions, and rather
IN SEC nice "elbows" are present. Also, from
Klahr's (1969) recent results we can also 2 3 4 5 6 7 8 9 10

Fig. Ia, Probability correct as a function infer the "significance" of at least the DIMENSIONS
of latency for a typical S. primary curves. For example, in 50 sets of Fig. 2a. Kruskal measure of fit as a
randomly generated sets of )6 points, 5'lr, function of number of dimensions in the
o 52 CI of the cases were associated with stress space, Condition I.
fj. 52 C2 values of .170 or less for three dimensions
35
when fit with the Kruskal program. We

OX"--__
would expect this result in even fewer cases
1.0 with our greater number of points (26).
A
.30
.9 Nevel thclcss, there was some inconsistency
of spatial configurations for the two
t; .8 experimental conditions, for the two

-
25
l&J
~ .7
met rics, and even of the primary as opposed
to the secondary approach. Perhaps even
.......
l>--A R• I
R· 2
R'I
SECONDARY
SECONDARY
PRIMARY
oo .6 more unsettling was failure of a set of 20
<>--<> R' 2 PRIMARY

dimensions clearly related to our intuitions en

en
w
~ .5 concerning visual similarity to reveal itself. a:
f-
:J This result points up the difficulty inherent en
iii .4 In using scaling techniques as a detective
15

~ .3
device unless the scaled dimensions either
arc very elementary and obvious or unless 10
Q,. .2 they turn out to be equivalent to some
previously hypothesized dimensions From
.1 examination of letters having relatively 05
large projections on various dimensions one
2 3 4 could infer the Importance of direction of
lines and angles, roundness, and vertical
CONFIDENCE RATING linearity. hut, again. these did not appear
in the same form for the different DIMENSIONS
Fig. I b. Probability correct as a function conditions and approaches. and the letters
of confidence rating for a typical S. associated with onc of these attributes Fig. 2b. Kruskal measure of fit as a
were not always the same or even function of number of dimensions in the
similarity in the choice model, T/i i as completely consistent with the attribute. space, Condition II.
cst imarcd by thc formula has been The difference in the ability of CI and
suggested as possessing several properties CII to be fit into a multidimensional space Two of the possible dimensions
appropriate for a measure of similarity may part wily follow from the recalibrat ion mentioned above, roundness and vertical
(Shepard. 1958); for this reason. the procedure carried out in ell. To take an linearity, correspond to two of the
resulting values were used as input data for extreme example, if a S's performance dimensions reported by Kunnapas (1966).
the multidimensional scaling analysis. The varied because of daily fluctuauons In Nevertheless, despite this finding and
present application of Kruskal's ( 1964h) at tention. equal accuracy. as given by the despite the hopeful outcome of Brown and
program included two fit iechruqucs as rccahbration. may not have implied the Andrews's (1968) comparison of the scaled
well as fitting the Euclidean and city-block same underlying visual space. The space of a discrimination and judgment
mctrics. One technique, referred to as the confusion matrix for CII would then be a task, it should not be surprising if results
primary approach, allows two pairs of composite of several underlying spaces. should differ in the Kunnapas and the

Perception & Psychophysics, 1971, Vol. 9 ( I A) 47

 Table 5 reflecting the same underlying structure in stimulus similarity did not, and a model
Correlation Coefficients for the Bias Parameters the data, correlation coefficients were reflects these properties. To be sure, the
for the Three Models and the TwoConditions· obtained for analogous parameters in the differences in the scaling fits indicate that,
OlP CHC AON three models, and these are shown in at the least, the psychological spaces may
Model CII CI Tables 5 and 6 for bias and similarity not have been identical for the two
CII CI elI
parameters, respectively. In Tables 5 and 6, conditions. When examining the correlations
AON CI .48 AON =ali-or-none activation model, OLP = with the supposed metrics, we may note,
CI .53 .88 .53 overlap activation model, and CHC = as a reference, the correlation of about .75
OlP
CII .89 .83 choice model. The two conditions are given obtained by correlating l'lij with -In(71ij)
CI .23 .70 by CI and CII. Table 6 includes (de). The overlap parameters correlated
CHC CII .60 correlations with and among two possible slightly higher with -In(71ij) than they do
,. See text for explanation of headings. metrics and a crude physical measure of with 71ij. Also, a favorable relationship is
interletter similarity. The metrics are ones found between the metric -In(71ij) and the
suggested by Luce (1963a) to be associated metric from the multidimensional scaling
present type of studies; one demands with the choice model (del and the other is program. It is interesting to note that if the
conscious evaluation of perfectly that obtained from application of the relationship between 71ij in CI and CII were
perceptible stimuli and the other scaling program (d.), The measure of by way of a similarity transformation
immediate identification of severely physical overlap or similarity (PHOVLP) (y = ax, "a" constant) as choice theory
degraded images produced by was found by taking blown-up versions of demands if the similarity structure is the
tachistoscopic exposure. the font used in the experiment and same in CI and CII, then the correlation
The finding that a Euclidean metric is superimposing them in pairs on a 10 by 10 coefficient for -In(71ij) for the two
generally most appropriate for the present grid (the letters were eigh t units high and conditions would also be large. The fact
data may at first seem bizarre. Torgerson about five units wide, on the average). that it is not (r =.63) tells us that the
(1958) and Shepard (1964), for example, Then, the relative intersection was relationship must have been affine
have suggested that certain types of visual computed for each pair, the intersection (y =ax + b, "a," "b" constan ts) rather
dimensions may be conducive to the divided by the union of cells covered in the than strictly simply a stretch or shrinkage
appearance of a city-block metric, such matrix. relationship. This probably followed from
dimensions being "perceptually distinct" Bearing in mind that we usually fewer estimated similarities of zero in CI
and "compelling." Letters, such as ''0'' interpret the correlation coefficient r or r 2 than in CII. Of course, conventional
and "I," for instance, may seem to be as referring to a linear relationship, the application of the theory would assume no
made up of perceptually distinct types of obtained values are surprisingly large. zero similarities.
elements. There are several possibilities Within the bias table, the correlations Concluding the correlation analysis is
that might be related to this finding. Some within a condition are higher than across that of the physical overlap measure, a
of the structure that appears to be conditions; this may indicate an actual measurement that does not suffer from the
independent may actually be related. For change of biases by Ss from CI and Cll. built-in dependencies that characterize the
instance, the extent of linearity of, say, an Also, the highest correlations are obtained other numbers in the correlation tables.
ellipse is inversely related to its degree of between the overlap activation model and The results are again substantially in the
roundness, and perceptual blur due to the choice model, suggesting that despite expected direction with the highest
tachistoscopic conditions might produce their differing heritage, these two models correlations found with the choice
several gradations of linearity, thus reflected bias processes in a similar way, similarity parameters and with the scaling
producing perceptual dimensions better and different from the ali-or-none model, metric.
suited to a Euclidean metric. Also, of by virtue of their ability to depict stimulus The approach assumed in this paper has
course, since letters are made up of several similarity. been of a twofold nature. On the one hand,
kinds of elements, this would tend to Among the similarity parameter mathematical models were employed as
promote perceptual continuity among the correlations, the most salient result was the investigatory structures to reveal and
letters themselves. A third possibility is remarkable r of .97 of the choice model in explicate various aspects of the data; the
that restricted (in time) observations such the two conditions, as contrasted with the substantive recognition models and the
as those in the present study could lower lowest correlation of biases for the choice scaling technique were complementary in
the probability of sampling information model across conditions, which, of course, this aim. On the other hand, there was a
from several dimensions simultaneously we would expect if bias changed and theory-testing facet of the study which
and thus improve the fit of a Euclidean
metric relative to the city-block metric Table 6
(e.g., see the discussion by Hyman & Well, Correlation Coefficients for Similarity Parameters, Distance
Measures, and a Physical Similarity Measure.
1967). It should be mentioned that the
lack of clear-cut results relating to OlP CHC de d
- - s- -
particular dimensions of similarity and Model CII Cl CII CI CII CI ell PHOVlP
perhaps even the support of a Euclidean
Cl .58 .43 -.67 -.61 .48
metric may be due to an artifact caused by OlP ell .64 -.67 -.54 .53
scaling group data, especially if Ss employ
different dimensions in their identification CI .97 -.75 -.70 .70
CHC CII -.76 -.65 .65
process. An experiment is in progress that
includes the gathering of long-term data for CI .63 .86 ~-.69
de CII .75 -.58
individual Ss and the application of the
present analyses to them. CI .57 -.67
ds CII -.54
In an effort to obtain a rough idea of the
extent to which the various models were ,. See text for further explanation.

48 Perception & Psychophysics, 1971, Vol. 9 (IA)

 REFERENCES
especially manifested itself in the Substitution of the theoretical expressions for
ATKINSON, R. C, & KINCHLA. R. A. A
evaluation and comparison of the the confusion matrix entries, Pij' will suffice to learning model for forced-choice detection
recognition models. Although this show that the above arc the appropriate experiments, British Journal of Mathematical
approach proved fruitful in delineating Iormulac. & Statistical Psychology, 1965, 18, 183-206.
some of the characteristics of alphabetic BROWN. D. R., & ANDREWS, M. H. Visual form
discrimination: Multidimensional analyses.
confusion and suggesting the type of model Per ce p t ion & Psychophysics, 1%8. 3,
necessary to deal with visual confusions, Overlap Activation Model 401-406_
the spatial configurations resulting from In this model. structure existed to measure or BUSH, R. R. Estimation and evaluation, In R. D_
reflect pairwise stimulus similarity. The Lucc, R. R. Bush, and E. Galantcr (Fds.),
the scaling analysis were less than
parameters assigned this function we called ~iJ' Han d book of mathematical psychology.
satisfactory, in terms of describing a designating the overlap or similarity of stimulus i Vol. I. New York: Wiley, 1963_ 1'1'.429-469
psychologically intuitive set of dimensions. and stimulus .1. To denote the probability that CLRNOG. U. Y., & ROSE, I. C. (l-ds.).
To some extent, this may be a function of "perfect" sensory information was obtained. we l.egibility of alphanumeric characters anti
the sheer complexity of the English other symbols. II. A reference handbook.
used the symbol ~ii' As this model included no
Washington. D.C: National Bureau of
alphabet; it may prove valuable to build up pure guessing state. only perfect information and Standards Misrellaneous 262-2.1967.
a psychophysics of simple visual elements, pairwise confusion states, we employed the ESTFS, W. K., & WFSSEl., D. L. Reaction time
such as lines at various orientations, and assumption that when in a confusion state, say in relation to display size and correctness of
then begin to construct more complex. between s, and Sj. the S gave response Ri with re sp on sc in forced-choice visual signal
alphabets with these. In this way the laws probability gi!(gi + gil, where gi is the response detection. Perception & Psychophysics, 1966,
strength or response bias (like Pi in the 1.369-373_
relating to the emergent properties of the HYMAN, R., & WELL, A, Judgments of
all-or-none activation model); this implies the
more complex stimuli may become known. similarity and spatial models. Percep tu-n &
con-rant ratio rule on the guessing biasev, Psychophysics, 1967.2,233·248.
APPENDIX The expressions which yielded the sensory KINCHl.A. R. A.. TOWNSEND, J. r.,
Although there exist maximum likelihood parameters for tim model were: YELlOn . .I. I., & ATKINSON, R. C.
estimation procedures for the choice model and Influence of correlated visual cues on auditory
the ali-or-none model, for convenience and ~'i ~ Pij + Pii· ti.] z: 1,2.•••. N. i i= II signal detection. Perception & Psychophysics,
1966, I. 67-73.
homogeneity of estimation technique, all
N Kl.AHR, U. A Monte Carlo investigation of the
parameters were estimated by setting each
.L Pji. statistical significance of Kruskul's nonmctric
parameter equal to a function of the data I i.c., scaling procedure. Psychometrika, 1969, 34.
j= \ .
the modified method of moments (aush, 1963) I' #01 319-330.
the particular function, of course, being related KRANTZ. D. H. Threshold theories of signal
to the particular model and the specific detection. Psychological Review, 1969, 76,
and the guessing biases were obtained from the
parameter. For simplicity_we will here denote 308,324.
formula: KRliSKAL, .I. B. Multidimensional scaling by
1>(RjlSi) as l'ij'
optimizing goodness-of-fit to a nonmctric
, I hypo thcxix, Psychome trika. 1964a. 29. )·28.
Alt-er-None Activation Model gi = TN:- (i = I. 2..... NI.
KRl:SKAI . .I. B. Nonrne tric multidimensional
It will be recalled that there exist no similarity y ...1,[ scaling: A numerical method. Psychomctrik a,
parameters as such in this model, only stimulus ~ I'
i=] ji 1964b, 29. 115·130.
parameters reflecting the likelihood of obtaininu Kl·JNNAPAS. T. Visual perception of capital
sensory information, which is related in a letters. Scandinavian Journal of Psychology.
one-one fashion to the (correct) response (0,1. In 1966,7.189-196.
Choice Model
addition, if the S docs not find himself in this LlICF, R. D. Individual choice bchovior. New
The choice model and the relevant formulae York: Wiley. 1959.
certain state, he must guess from all 2h
arc well known; hence. we shall simply list them LLCE, R. D. Detection and recognition. In R. U.
alternatives and he picks alternative "i" \\ ith
and remark that this model possesses the same l.ucc, R. R. Bush, and L. Galarnc r (Lds.),
probability Pi. This model possesses a total 015 I
number of similanty and gue . . ,ing parameter' lL's., Ho n d b o ok o] mathematical psvrhotogv.
parameters for this experiment as compared with
one as the overlap model for a tOlal of 3511 Ihe Vol. I N",,, York: Wiley, 1963a. 1'1'.103·189.
650 degrees of freedom in thc data rill' sunilaritv parume tcr functions arc: r.uct.. R. D. A threshold model tor simple
estimates for the sensory parameters were: detection experiments. Psychological Review.

0i" {Ph <I

11ii - _[Pil ii
1'.Il.1'."It
] ' "' \i..J·~1.2.·".Nl
1963b, 70, 61· 79.
NAKATANI. I.. II. A confusion-choke stimulus
recognition model applied !O word

l
.
recognition. Technical Report No. 31. 1968,
and the bia\ naramctcr- are e.. unrated hy: l I u m a n Communication Laboratory.
University of California. Los Angeles.
SHLI'ARD, R. N. Stimulus and response
generalization: Deduction of the gL'lll'rali/Jtion
quotient from a trace model. P\H:hofog"ical

*i£
I z; I Review. 1958. 65. 242·256.
+ SHU'ARD, R. N. The analysis of proximities:
I t may be obxcrvcd that instead of \clcding a
Multidimensional scaling with an unknown
single letter to Lise in estimating all the bia-, distance function: I. Psychomernka. 1962a,
ratios. in order to possibly give more stability to 27, ] 25·140.
the estimates. tor each bias parameter. each letter SHJ-:PARD. R. N. The analysis of proximuics:
ti = 1.2, .... Nl wa-, used as the denominator and the average Multidimensional scaling with an unknown
taken. In terms of the model. this simply results distance function: II. Psychomctrika, 1962b.
where. as before, N is the number of surnuli. and in an overall multiplication of the {31 bv the 27,219·246.
responses m an experiment. TIll' estimates for the con-rant
SHI:PARD, R. N. Attention and the metric
structure of the stimulus. Journal of
guessing bia-, parameter" were given hy
Mathematical Psychology, 1964. I. 54-87.
SMITH, K. Models of confusion, Paper delivered
• N 2 to Psychonornic Sccic ty, St. Louis, 1968.
p' ~ ----~---- (i ~ 1.2.' ". N)
1 N N P'k TORCJ-:RSON, W. S. Mel/JOdI of scaling. New
~ ~--2+N 2 York: Wiley. 1958.

*, #o,.j
]=1 k> I Pji which. of course. cancc}, out in the CXpre"\10J1" TOWNSFND. J. T. Choice behavior in a
for Pit. cu c d-r ccogniuon task. Technical Report

Perception & Psychophysics, 1971, Vol. 9 (I A) 49

 No. 103, 1966, Institute for Mathematical probabilities. Psychological Bulletin, 1968.69. may be obtained from the American
Studies in the Social Sciences, Stanford 126-131. Documentation Institute.
University.
WICKElGREN, W. A. Testing two-state theories NOTE
with operating characteristics and a posteriori 1. The complete set of theoretical predictions (Accepted for publication April 24. J 970.)

50 Perception & Psychophysics, 1971, Vol. 9 (IA)