Beruflich Dokumente
Kultur Dokumente
Discovering Cognitive
Architecture by
Selectively
Influencing
Mental Processes
Richard Schweickert
Purdue University, USA
Donald L. Fisher
University of Massachusetts Amherst, USA
Kyongje Sung
Johns Hopkins University
School of Medicine, USA
World Scientific
NEW JERSEY • LONDON • SINGAPORE • BEIJING • SHANGHAI • HONG KONG • TA I P E I • CHENNAI
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
Printed in Singapore.
R. S.
D. L. F.
K. S.
This page intentionally left blank
Preface
vii
viii Preface
Preface vii
Chapter 1: Introduction to Techniques 1
Chapter 2: Introduction to Process Schedules 7
Chapter 3: Selectively Influencing Processes in Task Networks 20
Chapter 4: Theoretical Basis for Properties of Means
and Interaction Contrasts 64
Chapter 5: Critical Path Models of Dual Tasks
and Locus of Slack Analysis 93
Chapter 6: Effects of Factors on Distribution Functions
and Consideration of Process Dependence 152
Chapter 7: Visual and Memory Search, Time Reproduction,
Perceptual Classification, and Face Perception 223
Chapter 8: Modeling with Order of Processing Diagrams 256
Chapter 9: Selective Influence with Accuracy, Rate,
and Physiological Measures 291
Chapter 10: Selective Influence of Interdependent Random
Variables 359
References 383
Author Index 403
Subject Index 409
ix
This page intentionally left blank
Chapter 1
Introduction to Techniques
1
2 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
presented and only had to repeat it. In a second condition (b) the
observer did not know which of five vowels was to be presented, and had
to repeat whichever was presented. Condition (b) requires two processes
in addition to those required in condition (a), namely, (1) the stimulus
vowel must be discriminated and (2) the response vowel must be chosen.
Thus, subtracting reaction time for (a) from reaction time for (b) gives
the time required for discrimination plus the time required for choice.
In a third condition (c) the observer did not know which of five
vowels was to be presented, but only had to respond to one, say, i, by
repeating it when it was presented. No choice was required for the
response, although discrimination of the stimulus vowel was required.
Thus, subtracting response time for (b) from response time for (c) gives
the time required for choice. Subtracting response time for (a) from
response time for (c) gives the time required for discrimination.
Donders reported that the mean reaction time for condition (a) was
201 msec, for (b) 284 msec, and for (c) 237 msec. Then the time for
discrimination is c − a = 36 msec, and the time for choice is b − c = 47
msec.
An important feature of this Subtractive Method of Donders is that
the experimenter can determine, based on the observer's responses,
whether the intended task was performed. Nonetheless, the experimenter
does not know how the task was performed. A criticism at the time,
based on introspection, was that changing from one condition to another
changes the nature of the processing, even of processing that precedes
the process allegedly inserted (Külpe, 1895, p. 414).
In principle, the results of Donders’ subtractive method can be
checked. To take a simple case, suppose the experimenter can insert and
remove two processes, with durations x and y. By producing three tasks
intended to have reaction times, respectively, of
x
y
x+y
Introduction to Techniques 3
the experimenter can check whether the duration of the third task is
indeed the sum of the durations of the first two tasks. This example is an
over simplification, but with three or more processes, if the experimenter
inserts and deletes processes in the right combinations, the number of
observed response times can be made large enough so that values of
unknown process durations can be solved for in more than one way,
allowing a check. Curiously, this does not seem to have been done at the
time.
Finally, suppose the response time for a level of Factor A combined with
a level of Factor B is the sum of the corresponding durations of process A
and process B.
It is easy to check that when both factors are at level 1, the reaction
time is τ11, when Factor A is at level 1 and Factor B is at level 2, the
reaction time is τ12, and so on. Each factor changes the duration of only
one process. Hence, the data can be represented by two processes in
series, with durations as above, and with each factor selectively
influencing a different process. (See Dzhafarov and Schweickert, 1995,
for a representation in which the reaction times and process durations are
random variables, rather than fixed constants.)
Clearly, it is arbitrary to use .5τ11 in the durations above, so this is not
the only way to represent the data with two processes in series. More
troubling is that a quite different process arrangement can also represent
the data. We have implicitly assumed that a process begins processing at
a starting point and stops processing at a finishing point, and if a second
process follows, its starting point is the finishing point of the first
process. McClelland (1979) and Townsend and Ashby (1983) showed
that factors can have additive effects on reaction time in a different kind
of model, where as soon as a process begins, it starts sending output to its
successor. McClelland’s model is called the cascade model, and Eriksen
and Schultz (1979) call such models continuous flow models. An
analysis of the cascade model by Roberts and Sternberg (1993) showed
6 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
that it failed to account for aspects of their data. But it often happens that
two process arrangements account for the known data equally well. In
the end, the choice between them can only be based on nonempirical
considerations such as simplicity, plausibility and taste.
If two factors do not have additive effects on reaction time, it is
possible that each factor prolongs a different process, so assumption (2)
above is satisfied, but the processes are not in series, so assumption (1) is
violated. Sternberg (1969) pointed out that if processes are in parallel,
the effect of prolonging two of them would be the maximum of the
effects of prolonging them separately.
In some situations, the processes are not all in series and they are not
all in parallel. Evidence comes from dual tasks, in which two stimuli are
presented, and a response is made to each (Telford, 1931). Consider the
case of two stimuli presented at the same time. When results are
compared with the corresponding single tasks in which each reaction is
made separately, the following outcomes are typical (e.g., Schvaneveldt,
1969), although not always found. In the dual task, the subject is not
carrying out all the single task processing for the first response, followed
by all the single task processing for the second response, because the
time required to do the dual task is less than the sum of the times
required to respond to each stimulus separately. On the other hand, in
the dual task the subject is not carrying out all the single task processing
for the first response simultaneously with all the single task processing
for the second response, because the time to do the dual task is longer
than the maximum of the times required to respond to each stimulus
separately. Something more general than pure serial or pure parallel
processing is needed.
Chapter 2
Bar charts are a natural way to represent the mental processes required
for a task; they are especially useful when intuition about process
durations is important. Bar charts are also called Gantt charts. Figure
2.1 gives an example for processes in a dual task. Stimulus s1 is
presented, followed after a stimulus onset asynchrony (SOA) by stimulus
s2. Responses r1 and r2 are made to stimuli s1 and s2 respectively.
There are three sequential processes for each stimulus, a perceptual
process, A, a central process, B, and a motor preparation process, C. (A
motor movement follows motor preparation, but because reaction time is
ordinarily measured at response onset, the motor movement that follows
is ordinarily not illustrated.) In this model, the perceptual processes for
the two stimuli, A1 and A2, are executed concurrently. However, the
response to the second stimulus is delayed because, in accord with
Welford’s (1952, 1967) single channel theory, the central processing, B2,
7
8 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
for the second stimulus cannot begin until the central processing, B1, for
the first stimulus is finished. The central processes B1 and B2 are
executed sequentially. The first use of this model that we are aware of is
by Davis (1957); it was popularized by Pashler and Johnston (1989). For
more discussion, see Pashler (1994).
predictions they all make. If one of the general predictions fails for an
experiment, there is no possible directed acyclic task network in which
the experimental factors selectively influence different processes.
Fig. 2.2. Directed acyclic network equivalent to Gantt chart for dual task.
The directed acyclic task network in Figure 2.2 is made of vertices joined
by arcs. Processing begins with the presentation of a stimulus at the
starting vertex of the network. A mental process is represented by an arc
10 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
directed from one vertex to another. The starting vertex of an arc, at the
tail, represents the starting point of the process. The ending vertex of the
arc, at the head, represents the finishing point of the process. Responses
are made at the ending vertices of the network.
Sometimes an arc does not represent a mental process, but merely
indicates that one process precedes another. For example, a stimulus
onset asynchrony is represented by an arc directed from the onset of one
stimulus, a vertex, to the onset of another stimulus, another vertex. This
SOA arc does not represent a mental process. As another example,
suppose a process stops using a certain resource at some point,
represented by vertex, and at another point, represented by another
vertex, a second process starts to use the resource. An arc from the first
vertex to the second vertex can be used to represent the fact that the
resource must be released by the first process before it can be used by the
second. If the resource is available the instant it is released, the duration
of the arc is 0. For convenience, we will often refer to arcs as processes,
even when there is no processing going on. An arc with duration 0
representing precedence is called a dummy process.
By starting at a vertex and moving along arcs in the direction of their
arrows until another vertex is reached, one traces a path. More precisely,
a path from a vertex u to a vertex z consists of the vertex u, followed by
an arc directed from u to a vertex v, followed by an arc directed from v to
a vertex w, and so on, with the last arc having ending vertex z. A single
vertex is considered a path. To indicate that one process immediately
precedes another, the head of the arc representing the first process is
incident with the tail of the arc representing the second. If one process
precedes another (not necessarily immediately), there is a path from the
head of the arc representing the first process to the tail of the arc
representing the second; the path will go along arcs in the direction
indicated by the arrows.
We say a vertex precedes another vertex if there is a path having at
least one arc from the former vertex to the latter vertex. A process
preceding a process, a vertex preceding a process, and so on are defined
similarly. A path that goes from a vertex u to the same vertex u, and that
has at least one arc, is called a cycle. An acyclic network has no cycles,
so a vertex or process does not precede itself. We assume precedence is
Introduction to Process Schedules 11
present trial, several targets are present on the screen. Suppose the
response “present” is made as soon as any process finishes with the
answer “target.” The processes working on the nontarget items can be
ignored, because they will not trigger a response. Then the response
“present” is made at an OR gate. At first it might seem that a single
network with both an AND gate and an OR gate is required. However,
the trials can be separated into target present trials and target absent
trials, with a different network for each type. The network for the
“present” response has an OR gate, the network for the “absent” response
has an AND gate. The task is represented by one OR network and one
AND network. More information on representing tasks with networks is
given in a later chapter.
The duration of an arc x is a nonnegative random variable, D(x). On
a particular trial, each arc is assumed to take on a particular value from
its probability distribution. The duration of an arc representing a process
is the duration of the process. The duration of an arc representing an
instantaneous action, such as a resource becoming available, is zero on
every trial.
The duration of a path is the sum of the durations of all the arcs on it.
A path can consist of a single vertex; in that case, the path duration is 0.
Since arc durations are random variables, the duration of a path is a
random variable also. To be specific, suppose a vertex u precedes a
vertex v on a particular path. The durations of the arcs on this path will
vary from trial to trial, so the duration of the path will vary also.
If there is more than one path from u to v, and we are interested in the
longest path from u to v, the path with the longest duration may not be
the same path on each trial. Despite this complication, we can speak of
the duration of the longest path from u to v; it is a random variable whose
value on a particular trial is the sum of the arc duration values on that
path which happens to be the longest for that trial. (On a given trial,
there may be several paths tied as longest or shortest from one vertex to
another; this turns out to not affect our conclusions.)
The time elapsing between the occurrence of vertex u and the
occurrence of vertex v is denoted D(u,v). If all vertices are AND gates,
D(u,v) is the duration of the longest path between vertices u and v. On a
particular trial, the longest path from the starting vertex of the network,
Introduction to Process Schedules 13
stands for Cognition, Perception, Movement, and also for Critical Path
Method. GOMS stands for Goals, Operators, Methods and Selection
Rules.) Estimates of the durations of component processes such as
typing and speaking were obtained from the videotapes and from the
human factors literature. For each task, a network was drawn for use of
the old workstation and another network was drawn for use of the new
workstation. With the networks and estimated durations, the time to
complete each task could be predicted for the new workstations.
Specifications indicated that several processes would be faster with
the new workstations. Surprisingly, predicted times to complete tasks
were longer. In the networks, some of the faster processes were not on
the critical path, so their shorter durations did not shorten completion
time. However, several other processes were inserted into the critical
path, thus increasing completion time. When the new workstations were
tested, completion times were indeed longer.
GOMS was developed by Card, Moran and Newell (1993), and CPM-
GOMS by John (1990). Cognitive task analysis is discussed in
Schweickert, Fisher and Proctor, 2003.
Processing Trees
Fig. 2.3. Processing trees for inclusion and exclusion conditions. Arcs are directed from
top to bottom.
pinclusion = R + (1 − R)F.
For a word in List 1, a subject in the exclusion group will not say the
word is old if it is recollected. However, a familiar word is more likely
to be from list 2 (presented recently), than from list 1 (presented earlier),
or new (not presented). So if a word is not recollected, but is judged
familiar, the subject will say the word is old. That is, for a word in List
1, the subject will say old (incorrectly) with probability
pexclusion = (1 − R)F.
The two equations above can be solved for the two unknowns, R and F.
To test the model, experimental factors expected to selectively
influence recollection or familiarity are manipulated. For example,
Jacoby (1991) proposed that a secondary task carried out during testing
would not change the familiarity of items, because familiarity was
established during study. However, the secondary task would harm
recollection, because recollection occurs during testing. Hence, the
secondary task is expected to decrease R leaving F invariant, as was
found. An example of a factor that does not selectively influence a
parameter is the presentation of words as anagrams instead of in the
usual way. This manipulation changes both R and F (Jacoby, 1991;
Jacoby, Toth & Yonelinas, 1993).
In a processing tree, at each vertex a process is executed. (In a task
network, processes were represented by arcs.) The first process to be
executed is represented by a special vertex, the root (at the top in our
illustrations). When a process is executed, it produces one of several
possible outcomes. These outcomes are represented by arcs leaving the
vertex representing the process. (Because the direction of all arcs is from
top to bottom, arrows can be omitted.) Such arcs are called the children
18 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
of the vertex. An arc is directed from its starting vertex to its ending
vertex. Each child of a vertex has a probability associated with it; this is
the probability the corresponding output is produced, given that the
process represented by the vertex is executed. The sum of the
probabilities associated with the children of a vertex is 1. When an
output is produced by a process, the arc corresponding to it is said to be
traversed, the ending vertex of the arc is said to be reached, and the
process represented by this vertex begins execution. This procedure
continues until a vertex with no children is reached. Such a vertex is
called a terminal vertex, and it produces a response. The responses fall
into mutually exclusive classes. Responses made at a particular terminal
vertex fall into one such class.
As in a task network, a path from a vertex u to a vertex z consists of
the vertex u, followed by an arc directed from u to another vertex v,
followed by an arc directed from v to another vertex w, and so on, with
the last arc having ending vertex z. A single vertex is considered a path.
A simple path is a path in which no vertex is repeated. We say a network
is connected if for any two vertices u and z there is a path from u to z, or
a path from z to u. A tree is network in which for every pair of vertices u
and z, there is exactly one simple path from u to z or exactly one simple
path from z to u, but not both. With our definition, a tree is connected.
Further, no vertex precedes itself on a path, so a tree is a directed acyclic
network. (With a task network, there may be more than one simple path
from one vertex to another, but this cannot happen in a tree.)
The probability of a path is the product of the probabilities associated
with the arcs on the path. The probability of a path consisting of a single
vertex is 1. Given that processing started at the root, the probability a
response is made at a particular terminal vertex is the product of the
probabilities on the path from the root to that terminal vertex (there is
exactly one such path). The probability a response in a particular class is
made is the sum of the probabilities that responses are made at terminal
vertices associated with that class. For short, we will sometimes say tree
to refer to a processing tree.
Because multinomial processing trees are so widely used, their
statistical analysis is well developed. See, for example, Batchelder and
Knapp (2004), Batchelder and Riefer (1986, 1990), Chechile and Meyer
Introduction to Process Schedules 19
Figure 3.1 illustrates a model for a dual task in which a subject produces
a time interval and, part way through the time interval, searches a screen
for a target (Schweickert, Fortin & Sung, 2007). Each trial had two
components. In the first component, a tone was presented. The subject
encoded its duration, to be used in the second component of the trial as
the duration goal of a time interval the subject would produce. When
ready for the second component, the subject pressed a button (noted as
event o1 in Figure 3.1). The button press blanked the screen and started
the time interval the subject was producing. After an interval (the
stimulus onset asynchrony, SOA), a display was presented (noted as
event o2 in Figure 3.1). The subject was to search through the display
20
Selectively Influencing Processes in Task Networks 21
and decide whether a target (a circle) was present among the distractors
(circles with a vertical line stem). The subject was to respond only after
he or she believed both that a time interval had elapsed whose duration
was the goal duration and also that a decision was made as to whether the
target was present in the display or not. The subject made a single
response by pressing a button. One button was pressed to indicate that
the target was present, another button was pressed to indicate that the
target was absent. The model is a simple AND network.
Fig. 3.1. Processes in a dual time production and visual search task. If the produced
interval is short, effects of prolonging SOA and visual search will be additive.
Three processes are illustrated, the produced time interval, the SOA,
and the visual search. (The visual search could be divided into
subprocesses, but the details are not relevant here.) The SOA and the
visual search are sequential. The time interval is concurrent with the
SOA and concurrent with the visual search.
For simplicity, consider trials on which the target is absent. On such
trials, the subject must search all the items in the display to correctly
decide the target is absent. Consider effects of manipulating three
factors. We can increase the time required for the search by increasing
the number of items in the display. We can increase the duration of the
SOA directly. Finally, we can increase the duration of the time interval
produced by the subject by giving the subject a longer duration goal.
In the initial condition illustrated in Figure 3.1, the goal duration has
elapsed (i.e., the produced time interval is over) before the search subtask
22 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 3.2. The combined effect of prolonging two concurrent processes will be less than
additive.
Fig. 3.3. If search is prolonged by 200, reaction time increases by 100. If the produced
interval is long, effects of prolonging SOA and visual search will be greater than additive.
With the initial condition in Figure 3.1, SOA and display size have
additive effects on reaction time. But with a different initial condition,
illustrated in Figure 3.3, these factors could interact. In Figure 3.3 the
produced interval is 700, longer than the sum of the SOA and the search
duration. With this initial condition, if the SOA duration is increased by
200, the increase in the reaction time is only 100. Likewise, if the search
duration is increased by 200, the reaction time increases by 100. Finally,
if the SOA duration is increased by 200 and the search duration by 200,
the reaction time increases by 300. The combined effect, 300, of
prolonging the SOA and the search is greater than the sum of their
separate effects (100 + 100). The factors of SOA and display size
interact.
In the task network, each of the three factors selectively influences a
24 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Slack
the longest path from the start of the network to r and (2) the duration of
the longest path that goes from the start of the network to r and that also
goes through arc A. Let o denote the starting vertex of the network, and
let A' and A" denote the starting and ending vertices of arc A. For two
vertices, say o and A', with the first preceding the second, let d(o, A')
denote the duration of the longest path between them. Let d(A) denote
the duration of process A. Then the total slack for A is
For more detail, see Schweickert (1978). With the formula, one can
see that on a particular trial a process is on the longest path from o to r,
that is, on the critical path, if and only if its total slack is 0.
Two related quantities are also used in the analysis of sequential
processes. Consider an AND network in which process A precedes
process B. The largest amount of time by which A can be prolonged
without delaying the start of B is the slack from A to B. Its value can be
found in the following way. Remove from the network all processes that
do not precede B', the starting vertex of B. (This includes removing
process B itself, but leaving vertex B'.) In the remaining network, B' can
be considered the terminal vertex. Then, by analogy with finding the
total slack for A, the slack from A to B is
negative. We will see that its value determines the form of the
interaction between factors selectively influencing processes A and B.
Here are examples of coupled slack values in later figures. In Figure 3.4,
s(A, r) = 225 and s(A, B) = 440, so k(A, B) = 225 440 = 215. In
Figure 3.7, the slack from B to C is 125, the same as the total slack for B.
Hence, k(B, C) = 0. However, if the duration of C were 100 (instead of
375 as in the figure), the total slack for B would be 175. The slack from
B to C would still be 125. Hence, the coupled slack for B and C would
be k(B, C) = s(B, r) s(B, C) = 175 125 = 50.
Typically in an experiment we do not know the durations of
individual processes on a trial, so we do not know the value of the slack
for any process. Values of this unobservable quantity are assumed to
have a probability distribution over all the possible trials. The random
variable taking on these values is denoted S(A, r); it is a function of the
random variables which are the process durations.
Selective influence
There are many ways that changing the level of an experimental factor
might selectively influence the duration of a process. For example, a
factor might make the duration of a process more variable, without
changing its mean. It is reasonable to assume that if changing the level
of a factor makes a process more difficult, it increases the mean duration
of the process. Unfortunately, this simple assumption does not lead to
many useful conclusions, so stronger assumptions are needed
(Townsend, 1990). Different assumptions about selective influence are
needed for different purposes. This chapter is concerned with expected
values of reaction times, so the assumptions need not be strong. For
many conclusions about expected values, dependencies between random
variables can be ignored; for example, the expected value of X + Y is the
sum of the expected values of X and Y, whether X and Y are correlated or
not.
Consider a factor selectively influencing a process A. Let a level of
the factor be denoted i, for i = 1, 2, .... If the brightness of a stimulus is
the experimental factor, then the levels 1 and 2 might indicate bright and
dim, respectively. Higher level numbers indicate greater process
Selectively Influencing Processes in Task Networks 27
0 if u < s(A, r)
u s(A, r) if u > s(A, r).
[x]+ = 0 if x ≤ 0
[x]+ = x if x > 0.
The result is that when the process A is prolonged, its mean response
time either increases or stays the same; i.e., it increases monotonically.
The reasoning is similar for other changes in the factor levels.
When there are two responses, it is customary to use the time from the
onset of the second stimulus to the onset of its response as the reaction
time to the second stimulus. Then for the model in Figure 2.1 it is not
hard to show that the response time to the second stimulus decreases
monotonically as the SOA increases. This may seem at first to contradict
the statement that increasing a factor level increases the mean response
time. However, with this way of measuring the response time to the
second stimulus, the location of the event used to start the clock (the
onset of stimulus 2) changes as the SOA changes. If instead the clock is
started at the onset of stimulus 1, the mean time at which the response to
the second stimulus is made increases as the SOA increases.
A note on OR networks
When the processes are in series, the factors have additive effects, so
the interaction contrasts are zero for every i and j.
The effects of selectively influencing two processes in a task network
depend on how the two processes are arranged. The major distinction is
between concurrent and sequential pairs of processes. (For a good
introduction, see Logan, 2002.) Sequential pairs are further
distinguished depending on whether they are in a structure called a
Wheatstone bridge. A Wheatstone bridge is illustrated in Figure 3.4.
Processes A and B are on opposite sides of the bridge. One place a
Wheatstone bridge arises is in a dual task, when subjects are instructed to
respond to the first stimulus before responding to the second stimulus.
This commonly given instruction, in effect, inserts a dummy process
between the two responses to establish their order. Figure 3.5 shows the
32 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
task network for the dual task model in Figure 2.2 drawn with the
additional constraint that response 1 precedes response 2. Figure 3.5 can
easily be redrawn as an AND network. It has the form of a Wheatstone
bridge. For analyzing times to make response 2, processes B1 and B2 are
on opposite sides of a Wheatstone bridge (as are other pairs such as A1
and C2). Other examples of models with a Wheatstone bridge are the
double bottleneck models of de Jong (1993), Ehrenstein, Schweickert,
Choi and Proctor (1997) and the stimulus-response compatibility model
of Kornblum, Hasbroucq, & Osman, (1990).
Fig. 3.4. Processes A and B are on opposite sides of an incomplete Wheatstone bridge.
Fig. 3.5. Instructing the subject to make response r1 before r2 creates a Wheatstone
bridge.
Selectively Influencing Processes in Task Networks 33
Fig. 3.6. Processes A and B are on opposite sides of a complete Wheatstone bridge.
Fig. 3.7. AND network used in simulations. Mean durations of processes not prolonged
are on arcs; mean durations of processes prolonged are in table headings.
36 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Table 3.1
Expected Values of Reaction Times
When Factors Influence Concurrent Processes A and B in Figure 3.7
All Process Durations Exponentially Distributed
μB
25 100 150 200 250
μA
Interaction Contrasts
μB
25 100 150 200 250
μA
300 - - - - -
450 - -4.2 -8.1 -12.2 -16.1
500 - -5.4 -10.4 -15.6 -20.8
550 - -6.4 -12.5 -18.8 -25.1
650 - -8.3 -16.1 -24.5 -32.9
700 - -9.1 -17.7 -27.1 -36.4
750 - -9.8 -19.2 -29.4 -39.7
The same patterns were found with each process duration sampled from
a truncated normal distribution, that is a distribution whose density
function is the normal distribution restricted to nonnegative values, and
renormalized so the area under it is one. The standard deviation of each
process duration was set to one fourth of its mean. The value of one-
fourth is representative of values typically found for response times
Selectively Influencing Processes in Task Networks 37
Table 3.2
Means of Simulated Reaction Times
When Factors Influence Concurrent Processes A and B in Figure 3.7
All Process Durations Have Truncated Normal Distributions
Interaction Contrasts
OR networks
Statistical considerations
Table 3.3
Expected Values of Reaction Times
When Factors Influence Sequential Processes B and C in Figure 3.7
All Process Durations Have Exponential Distributions
μB
25 100 150 200 250
μC
100 376.5 421.3 456.7 495.0 535.5
150 397.7 445.0 481.7 521.0 562.3
200 424.7 474.3 512.1 552.3 594.3
250 455.9 507.5 546.3 587.3 629.9
300 527.3 582.2 622.5 664.7 708.4
375 546.5 602.1 642.7 685.2 729.1
400 566.1 622.3 663.3 706.1 750.1
Interaction Contrasts
μB
25 100 150 200 250
μC
100 - - - - -
150 - 2.6 3.8 4.8 5.7
200 - 4.9 7.3 9.2 10.7
250 - 6.9 10.2 12.9 15.1
300 - 10.2 15.1 19.0 22.2
375 - 10.9 16.1 20.3 23.7
400 - 11.5 17.0 21.5 25.1
42 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Table 3.4
Means of Simulated Reaction Times
When Factors Influence Sequential Processes B and C in Figure 3.7
All Process Durations Have Truncated Normal Distributions
Interaction Contrasts
The task network illustrated in Figure 3.4 has an unusual feature. There
are three paths through the network, and only one of them contains both
A and B. If the arc from A to B has a short duration, then the path
containing them both will hardly ever be the critical path, so it will
appear as if A and B are not on a path together. In other words, although
A and B are in fact sequential, they might appear to be concurrent.
When factors selectively influence sequential processes on opposite
sides of an incomplete Wheatstone bridge (e.g., A and B in Figure 3.4),
the resulting patterns of mean response times can be similar to (or
identical to) the patterns observed when concurrent processes are
influenced. Fortunately, the patterns will be different provided a wide
range of levels of the factors are used, when large increments in process
durations overcome the relevant slacks. Once again, the patterns to be
expected are best illustrated by examples, which we will turn to after
explaining more about Wheatstone bridges.
The only way two factors selectively influencing sequential processes
A and B in a directed acyclic task network can produce a negative
interaction is for the network to contain a subnetwork in the shape of a
Wheatstone bridge, with A and B on opposite sides of the bridge
(Schweickert, 1978). To be more precise about what it means for one
graph to have the same shape as another, we need to explain what is
meant by two graphs to be homeomorphic. Consider a graph consisting
of two arcs in series, one from a vertex u to a vertex v, and another from
44 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
However, expected values of interaction contrasts need not all have the
same sign (Lin, 1999). An example is in Figure 3.9. Numbers on the
arrows are process durations. The duration of process F is 0 in the left
panel and 3 in the right; otherwise the networks are the same. Reaction
times for various durations of processes A and B for each network are in
Table 3.7. Corresponding interaction contrasts are in Table 3.8.
Interaction contrasts are all negative for the left hand network; further
they are monotonically decreasing with increasing values of durations of
A and B. Results are similar for the right hand network, except the
interaction contrasts are positive and monotonically increasing. So far, so
good. But suppose the left hand network occurs 1/10 of the time and the
right hand network 9/10 of the time. Resulting interaction contrasts are
in Table 3.9. They do not have the same sign, nor are they monotonic
with the durations of processes A and B.
Despite their potentially complicated behavior, processes on opposite
sides of a Wheatstone bridge (complete or incomplete) are on a path
together, and this fact can be used to distinguish them from concurrent
processes, as we now explain.
Table 3.5
Expected Values of Reaction Times: Factors Influence Sequential Processes
A and B On Opposite Sides of the Wheatstone Bridge in Figure 3.4
Baseline Process Durations and Prolongations Have Exponential Distributions
μB
25 100 150 200 250
μA
50 619 674 715 758 802
200 696 744 782 823 865
250 729 775 812 852 895
300 764 809 845 885 927
400 839 881 917 958 997
450 879 920 955 994 1035
500 920 960 995 1033 1074
Selectively Influencing Processes in Task Networks 47
μB
25 100 150 200 250
μA
50 - - - - -
200 - - 7 -10 -12 -14
250 - - 9 -12 -15 -18
300 - -10 -15 -18 -21
400 - -12 -18 -20 -25
450 - -13 -19 -24 -27
500 - -14 -21 -25 -29
Table 3.6
Means of Simulated Reaction Times: Factors Influence Sequential Processes
A and B On Opposite Sides of the Wheatstone Bridge in Figure 3.4
All Process Durations Have Truncated Normal Distributions
Fig. 3.9. When sequential processes A and B are prolonged, interaction contrasts are
monotonically decreasing for the left hand network and monotonically increasing for the
right hand network. They are neither increasing nor decreasing in a mixture of the two
networks.
Selectively Influencing Processes in Task Networks 49
Table 3.7
Reaction Times for Various Durations of A and B
In the Networks of Figure 3.9
Table 3.8
Interaction Contrasts for Various Durations of A and B
In the Networks of Figure 3.9
Table 3.9
Interaction Contrasts for Various Durations of A and B
In a Mixture of the Networks of Figure 3.9
be close to 0, even when the factors both have large effects, as Table 3.1
makes clear. Nonetheless, with concurrent processes, a prolongation of
one process can, in principle, always be made large enough to eclipse the
effect of prolonging the other. The next section gives details about long
prolongations.
Distinguishing sequential from concurrent processes is important for
understanding the task network as a whole. It turns out that knowing
which pairs of processes are sequential and which are concurrent gives
sufficient information to construct an underlying directed acyclic graph
with a procedure called the Transitive Orientation Algorithm, see
Golumbic (1980) and Schweickert (1983b). The algorithm puts
sequential processes in order, and all possible directed acyclic networks
consistent with the classification of processes as sequential or concurrent
can be produced. An important property of this algorithm is that if a
proposed classification of pairs of processes as sequential and concurrent
is not possible with a directed acyclic graph, the procedure stops, and
indicates that the proposed classification is inconsistent with a directed
acyclic graph.
skip to the next section (on Additive Factors) without loss of continuity.
The next chapter explains the assumptions underlying the results.
Concurrent processes
− 671.111 + 616.093. On the other hand, in Table 3.2, for the truncated
normal distribution, the numbers in column 3 have converged to their
lower bound of −19 = −547 + 528. Likewise, the numbers in row 3 have
converged to their lower bound of −28 = −556 + 528. Note that the row
limit is different from the column limit.
The lower tail of the distribution of the process when prolonged is the
key to whether the bound in the Appendix is the limit. The lower tail is
thick for the exponential distributions but thin for the truncated normal
distributions. When the duration of a process A is exponentially
distributed (or is the sum of a few exponentially distributed durations)
there is a nonnegligible probability of a duration too low to put A on the
critical path, even when the mean duration of A is large. In contrast, with
the truncated normal distribution, as the mean increases the area under
the lower tail becomes negligible.
Sequential processes
end of the last column are close also (−27.4 is close to −29.2).
Evidentially, the rows and columns have nearly converged to the same
limit. However, their values are far from the parameter value, E[K(A, B)]
= −130.0 (found with the Order-of-Processing diagram). By comparison,
the interaction contrasts in the lower right hand corner of Table 3.6 are
close together (− 202 is close to − 203, for the last row and the last
column). Further, the interaction contrast of − 203 in Table 3.6 is fairly
close to the parameter value, E[K(A, B)] = − 213. Convergence, or near
convergence, was only obtained in Table 3.6 when the processes were
prolonged to large durations. The need for longer prolongations in Table
3.6 than in Table 3.5 is due, to some extent, to the use of simulations in
the former and exact values in the latter.
To summarize, the interaction contrasts converged faster for process
durations with exponential distributions than for those with truncated
normal distributions. The latter required very large prolongations for
convergence. Part of the difference is probably due to the use of an
algorithm for the exponential distribution and simulations for the normal.
With the exponential distribution, interaction contrasts did not converge
to the bounds in the Appendix; nonetheless, with sequential processes,
the rows and columns converged to the same quantity. With truncated
normal distributions, the interaction contrasts converged slowly, but
when they did converge it was to the bounds in the Appendix. The
results demonstrate the feasibility of investigating limiting values of
interaction contrasts experimentally, but high levels of the factors
selectively influencing the processes must be available. The limiting
values may not be good estimators of the network parameters.
Superprocesses
Fig. 3.10. The subnetwork between o and c is a stage, as is that between c and r.
Processes E and F together with vertices o and c form a superprocess.
As before, for a vertex x that precedes a vertex y, let d(x, y) denote the
duration of the longest path between them in an AND network and of the
shortest path between them in an OR network. Let f[d(a, b),..., d(c, e)]
and g[d(u, v),..., d(w, z)] be functions of path durations. If for all possible
values of the path durations in the arguments of f and g, f[d(a, b),...,
d(c, e)] = g[d(u, v),..., d(w, z)] then we write
Appendix
Interaction contrasts for extreme values of the factors are discussed here
for AND networks, results for OR networks are analogous. When the
factor selectively influencing process A is at level i, Ui denotes the
increase in the duration of A from its baseline D(A). When the factor
selectively influencing process B is at level j, Vj denotes the analogous
quantity for the increase in duration of B. Let the expected value of the
response time, when the factor prolonging A is at level i and the factor
prolonging B is at level j, be denoted E[Tij]. Then
x1
.
xn
x1 y1
x = and y = ,
x n y n
64
Theoretical Basis for Properties of Means and Interaction Contrasts 65
Probability spaces
(i) Ω S.
(ii) If S S, then ~ S S. Here, ~ S, the complement of S, is the
set of elements of Ω that are not elements of S.
(iii) If {S1, S2,... } is a countable set of subsets of S, then S1∪ S2 ∪...
S.
A set of subsets of Ω satisfying (i), (ii) and (iii) is called a sigma algebra.
(Hays, 1994, considers any subset of the sample space to be an event, but
he is speaking of a finite or countably infinite sample space.)
A probability measure on S is a function P from the set of events into
the real numbers such that
(i) P(Ω) = 1;
(ii) If S S, then P(S) > 0;
(iii) If {S1, S2,... } is a finite or countably infinite set of pairwise
mutually exclusive events, then
For some experiments, such as rolling a red die and a green die
simultaneously, it is natural to consider a random vector that associates a
list of real numbers with every outcome of the experiment. For example,
X1 can be set to the number of dots that come up on the red die, and X2
can be set to the number of dots that come up on the green die. Then
X1
X=
X 2
X1
X=
X n
x1
x= .
x n
which both expectations exist. For random vectors the relation ≤st is
stronger than the assumption that the corresponding joint cumulative
distribution functions are ordered by ≤. That is, for random vectors X
and Y, if X ≤st Y, then FX(t) ≤ FY(t), for all t, but the converse is not true.
(As a technical detail, for certain random variables the relation ≤st is not
transitive, because the integrals required for certain expected values do
not exist. With little loss of generality, we assume for random variables
discussed here that ≤st is transitive.)
Recall that if random vector X and random vector Y have the same
joint cumulative distribution function, we write X ≈ Y. If X ≤st Y and Y
≤st X, it follows that X ≈ Y. Some authors write X =st Y to indicate that
X and Y have the same cumulative distribution function. (Another
technical detail is that X and Y can have the same joint cumulative
distribution function while having joint density functions that differ at a
countable number of isolated points. With little loss of generality, we
assume this situation does not arise for random variables here.)
The next result characterizes the usual stochastic ordering (see, e.g.,
Müller & Stoyan, 2002, Theorem 3.3.5; Shaked & Shanthikumar, 2007,
Theorem 6.B.1):
The following two statements are equivalent for two random vectors
X and Y.
(1) X <st Y.
(2) There exist random vectors X̂ and Ŷ on the same probability
space, with X ≈ X̂ and Y ≈ Ŷ , such that P( X̂ < Ŷ ) = 1.
Conditional expectation
For continuous random variables X and Y, the joint density is the function
f ( x, y) such that for all real x and y ,
x y
P( X x, Y y )
f ( x, y )dxdy .
x
P( X x) f
X ( x)dx .
E X [ X | y] xf ( x | y )dx .
EY [ E X [ X | y ]]
xf ( x | y ) dx
f ( y ) dy
xf ( x, y )dxdy
xf
x ( x )dx E[ X ].
A
A=
B
where functions f3,..., fp+2 do not depend on the levels i and j. Again,
generalization to a finite number of factors and of vector components is
straightforward.
Suppose random variables A and B are selectively influenced by
Factor Α and Factor Β, respectively. Note that for different pairs of
levels <i, j > and <i', j' > random variables Aij and Ai'j' may not be
defined on the same probability space. But in the definition of selective
influence the same random vector C on the same probability space is
employed for every pair of levels. Hence, through the assumption of
selective influence there are random variables f1i(C) and f1i'(C) defined
on the same probability space that have distributions equivalent to Aij and
Ai'j', respectively. Note also that the level j of Factor Β is not needed to
specify Aij, so we may write simply Ai. This manner of defining selective
influence of factors on random variables brings conceptual clarity and in
a more general form can be employed widely, not only to process
durations, but to perceptual images and mental tests. More about it is in
Chapter 10.
Let Dij be a random vector whose component random variables are the
durations of the processes required for a task when Factor A is at level i
and Factor B is at level j; that is, random vector D depends on Factors Α
and Β. A pair of factor levels can be written as a vector <i, j>. These
vectors are partially ordered; if i' ≤ i and j' ≤ j, we write <i', j'> ≤ <i, j>.
Suppose that whenever <i', j'> ≤ <i, j>
74 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
D i ' j ' st D ij .
ˆ , D̂
Then whenever <i', j'> ≤ <i, j> there exist random vectors D i' j' ij
ˆ , D
and Ẑ on the same probability space, with 0 <st Ẑ , D i ' j ' D i' j' ij
D i ' j ' st D ij .
For discussion of partial orders of factor levels (indices) for which the
usual stochastic order leads to realizable monotonicity, see Fill and
Machida (2001).
For our purposes factors must do several things. Factors must selectively
influence process durations in the sense of Dzhafarov and his colleagues,
increasing factor levels must order process durations with some type of
stochastic ordering, and finally, increasing the level of a factor must add
something to the process duration it selectively influences. The
following assumption leads to these.
Consider the random vector
Theoretical Basis for Properties of Means and Interaction Contrasts 75
Uˆ *
1
ˆ
U I *
ˆ
V1 *
C =
> 0.
V J *
Aˆ
base
Bˆ
base
Cˆ 1
Cˆ p
Now suppose for every level i of Factor Α and every level j Factor Β
Aˆ base Uˆ 1 * Uˆ i *
Aij
Bˆ base Vˆ1 * Vˆ j *
Bij
Aij = C1ij Cˆ1 .
Cpij Cˆ p
U ij
V ij
A
baseij
Dij = B .
baseij
C1ij
Cp ij
Uˆ1 * Uˆ i *
Vˆ1 * Vˆj *
Aˆbase
Dij D̂ ij = Bˆ
base . (4.1)
Cˆ1
Cˆ p
0
0
Aˆ
base
D 11 D̂11 = Bˆ .
base
ˆ
C1
Cˆ p
Uˆ 1 * Uˆ i *
0
ˆ
ˆ
D i1 D Abase
i1 = ˆ .
Bbase
Cˆ1
Cˆ p
0
ˆ ˆ
V1 * V j *
ˆ
Abase
D1 j D̂1 j = Bˆ .
base
Cˆ1
ˆ
Cp
Theoretical Basis for Properties of Means and Interaction Contrasts 79
Now, let
Uˆ i Uˆ 1 * Uˆ 2 * Uˆ i *
and
Vˆ j Vˆ1 * Vˆ2 * Vˆ j * .
Because all the random variables are nonnegative, if i' ≤ i and j' ≤ j,
ˆ D
P (D ˆ ) 1;
i' j' ij
D i ' j ' st D ij .
U i ' j ' U ij
st ,
Vi ' j ' Vij
ˆ
Abaseij Abase
ˆ
baseij Bbase
B
C1 Cˆ1 .
ij
ij Cp
Cp ˆ
Interaction contrasts
u
v
abase
b base ,
c1
cp
Uˆ ij
Vˆij
Aˆ baseij
.
Bˆ baseij
Cˆ 1ij
Cˆ p
ij
In this sample, the value for the duration of process A when Factor A
is at level i is abase + u. In the sample, a value for the duration of process
A when Factor A is at level 1 is abase. Similarly, for process B, in the
sample the value for the duration of process B when Factor B is at level j
is bbase + v. In the sample, a value for the duration of process B when
Factor B is at level 1 is bbase.
If both factors are at level 1, the total slack for process A is given by
the following equation,
Schweickert (1978). (Recall that A' and A" denote the starting and
ending vertices of process A, respectively.) All expressions in the
equation are for the sample values in the case when both factors are at
level 1. The duration of the longest path from the starting vertex, o, to
the terminal vertex, r, is d(o, r). The duration of the longest path with
process A on it, going from o to r, is d(o, A') + abase + d(A", r). If process
A is on the longest path from o to r, then the total slack for A is 0.
Otherwise, the total slack for A is the difference in the durations of these
two paths.
When both factors are at level 1, the reaction time is t11 = d(o, r). For
a real number x, let [x]+ = max{0, x}. When Factor A is at level i and
Factor B is at level 1, the reaction time is ti1 = t11 + [u − s(A, r)]+.
Prolonging process A by u may change the total slack for process B.
Similar to Equation (4.2), when Factor A is at level i and Factor B is at
level 1, the total slack for process B is
Here the subscript i1 indicates the values are for the case when Factor A
is at level i and Factor B is at level 1. The total slack for process B when
process A is prolonged by u depends on whether processes A and B are
concurrent or sequential.
Concurrent processes
Because s(A, r) and s(B, r) are functions of abase, bbase, c1,..., cp, we can
write h as a function h(u, v, abase, bbase, c1,..., cp).
Let us consider the effect on h of changing the prolongations of A and
B. If u = v = 0, then h has value 0. For fixed values of abase, bbase, c1,...,
and cp, function h is monotonically decreasing in u and v.
From Equation (4.1),
U 11 0
V 0 .
11
Because
84 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
0 U ij
0 st ,
Vij
(To avoid double subscripts in the expected value symbol, U11 is written
as U11, and so on.) On each side of the inequality the random variables
conditionalized on have the same joint distribution for all levels of the
factors, namely the joint distribution of Aˆbase , Bˆ base , Cˆ 1, , Cˆ p . By taking
expected values on both sides of the inequality over the random variables
conditionalized on, it follows that 0 > E(h(U, V, Abase, Bbase, C1,..., Cp).
Hence, for all i and j, the interaction contrast h is 0 or negative (see also
Schweickert, 1978).
Factors selectively influencing two concurrent processes by
increments will ordinarily have a negative interaction. Exactly additive
effects are impossible except under extraordinary conditions, such as a
factor having no effect (see Schweickert & Townsend, 1989; Townsend
& Ashby, 1983; Townsend and Schweickert, 1989).
Further, because for i' ≤ i and j' ≤ j,
U i ' j ' U ij
st ,
Vi ' j ' Vij
That is, for every given value of j, (AB)ij is bounded from below.
Likewise, for ever given value of i, (AB)ij is bounded from below.
Sequential processes
Then
(Schweickert, 1978).
When u = v = 0, h is 0. When u and v are each so large as to
overcome all the slacks, the interaction contrast is h = s11(A, r) − s11(A,
B'), a constant not dependent on u and v. This constant is the coupled
slack between A and B,
k(A, B) = s(A, r) − s(A, B') = d(o, r) − d(o, B') − d(A", r) + d(A", B').
U 11 0
V 0 .
11
Because
0 U ij
0 st ,
Vij
That is, the expected value of the interaction contrast is nonnegative for
all levels of the factors.
Further, because for i' ≤ i and j' ≤ j,
U i ' j ' U ij
st ,
Vi ' j ' Vij
OR networks
Table 4.1
Effects of Prolonging Sequential Processes A and B
in a Critical Path Network
u – s(A,r) 0 –k
u – s(A,r) 0 –k
v – s(B,r) 0 –k
v – s(B,r) 0 –k
u – s(A,r) + v – s(B,r) + k k 0
Note: Process A precedes B. Reaction time is t11 when neither is prolonged, ti1 when A is
prolonged by u, t1j when B is prolonged by v, tij when both are prolonged. Coupled slack
k(A, Bʹ) is denoted k.
between vertices are of interest, and d(p, q) denotes the duration of the
shortest path from vertex p to a vertex q which follows it. Suppose there
is a path from the ending vertex of a process A to a vertex p. The surplus
from A to p is the longest amount of time by which the duration of A can
be shortened without decreasing the duration of the shortest path from o
to p, that is, without A becoming an arc on the shortest path from o to p.
The surplus from A to p is
The equation for slack has exactly the same form in OR and AND
networks; the difference is that the symbol d(p, q) is interpreted as the
duration of the shortest path between p and q in an OR network and of
the longest path in an AND network. The coupled slack between process
A and a process B following it has the same form in OR and AND
networks,
When two stimuli are presented close together in time, and each requires
a response, the response time to the second is usually longer than if it
were presented alone (Telford, 1931). The delay is called the
psychological refractory period (PRP). The underlying mental
architecture has been probed by hundreds of experiments; we see fine
detail in places, elsewhere even the broad outlines are faint. At this point,
resolution is best at the stimulus end of the system, coarser at central
processing, and roughest at the response end, where processing seems
most complex. This chapter does not have space for everything we
know, it focuses on how we know it, in particular, on how structure is
revealed by selectively influencing processes. For more on attention see
Johnson and Proctor (2004).
93
94 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Central limitations
limitations. The central limitations when there are two stimuli are due
either to a single central channel that can process only one stimulus at a
time (Welford, 1952, 1967), or to capacity constraints which slow their
simultaneous processing (Broadbent, 1958). In Welford’s single channel
theory, if a second signal arrives while the central mechanism is busy,
that signal must wait, causing the refractory delay.
Welford (1952) proposed an easily tested version of a single channel
model with sequential central processing of the stimuli. He proposed
estimating the central processing time of the first stimulus by its entire
reaction time. Let RT1 denote the reaction time to stimulus 1 (implicitly
in this model, this time is the same for both the single and dual task
conditions). Let RT2 (single) denote the response time to stimulus 2
when presented alone, and RT2 (dual) denote the reaction time to
stimulus 2 presented in the dual task condition. Let I denote the
interstimulus interval (now called SOA). Then
see Figure 5.1. The resulting model explains many aspects of the data,
but is wrong in several details (e.g., Ollman, 1968. For discussion, see
Luce, 1986 and Schweickert & Boggs, 1984).
Fig. 5.1. Welford’s Single Channel Model. Time is on the horizontal axis. Physical
onsets of the first and second stimuli are denoted s1 and s2, physical onsets of responses
to them are at times r1 and r2, respectively.
96 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 5.2. Dual task model of Davis (1957), with current notation. Also called Response
Selection Bottleneck model and Single Central Bottleneck model. Stimulus 1 and
stimulus 2 are presented at s1 and s2. Sensory, central and motor preparation processes
for s1 are denoted A1, B1, and C1; analogous notation for s2. Central refractory times
(currently often called switching times) occur after central processing of Task 1 and of
Task 2, denoted SWa and SWb, respectively. Responses to s1 and s2 are at r1 and r2. For
a single trial, SWb is ignored, and SWa is denoted SW. If another dual task trial follows,
SWb must finish before new central processing starts.
Critical Path Models of Dual Tasks and Locus of Slack Analysis 97
Response limitations
Fig. 5.3. de Jong double bottleneck model with Refractory Interval between start of
response 1 and start of response 2. Other notation as in Figure 5.2. (Based on de Jong,
1993, Figure 1.)
Fig. 5.4. de Jong double bottleneck model as a critical path network. (Refractory
Interval RI is not the same as response 2, r2.) For a later discussion of Karlin &
Kestenbaum, assign durations A1 = 40, B1 = 150, C1 = 200, SW = 49, RI = 200, SOA =
90 or 1150, A2 = 110, B2 = 100 or 190, C2 = 200.
Table 5.1
Some Factors Influencing Processes in Dual Tasks
M. C. Smith
In Smith’s (1969) experiment, the first stimulus was a digit from 1 to 4
on either a red or green background. The subject identified the stimulus
by pressing a button; each finger indicated a different digit, fingers on
the left hand indicated the red background, those on the right indicated
the green. The second stimulus was a 1 or a 2 on a grey background; the
subject identified the digit vocally. There were two factors. The
interstimulus interval was varied as was the number of alternatives for
the first stimulus.
Let RT1 and RT2 be the reaction times to the first and second stimuli,
respectively. The data, from Figures 1 and 2 of Smith (1969), are
presented in Table 5.2. The response time to the second stimulus
increases with the number of alternatives for the first, as found by Karlin
and Kestenbaum (1968). Data again support the hypothesis that
inadequate expectancy of the second stimulus is not the only cause of the
refractory delay.
104 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
The first two rows in Table 5.2 for RT2 + ISI again provide a clear
demonstration of slack. When the interstimulus interval was increased
by 100 msec, from 50 to 150 msec, the increase in response time RT2
was much less than 100 msec. Specifically, the increase in RT2 was 15
msec when there are 2 alternatives, and only a few msec when there are 4
or 8 alternatives. Most of the 100 msec increase in the interstimulus
interval is expended in overcoming the slack.
Table 5.2
Reaction Times from Smith (1969)
Several names have been given to the model in Figure 5.2, including the
Single Channel, the Response Selection Bottleneck, the Standard
Bottleneck, and the Central Bottleneck model. As often happens, slight
variations of a model are given the same name, and exactly the same
model is given different names. We usually call the model in Figure 5.2
the Single Central Bottleneck model, to emphasize that it has one
bottleneck, whose function, whatever it may be, is central. Pashler and
Johnston (1989) tested the model through its predictions about
interactions. Their experiment will be described, followed by the
predictions.
In Experiment 1 Pashler and Johnston (1989) presented a high or low
pitched tone, followed after a Stimulus Onset Asynchrony (SOA) by a
letter. The subject pressed a button with the left hand to identify the
tone, and a button with the right hand to identify the letter. Intensity of
the second stimulus was varied, as was the SOA. Further, the second
stimulus was either the same as on the previous trial (a repetition) or not.
Response times RT1 to the first stimulus and RT2 to the second stimulus
in the double stimulation task were recorded. In separate blocks of trials,
response time RT2 to the second stimulus was recorded in the same
paradigm, but with no response required to the first stimulus.
Pashler and Johnston (1989) represented the model as a Gantt chart,
see Figure 2.1; the corresponding critical path network is in Figure 5.2.
For each stimulus the three processes, A, B, and C are for perceptual,
central, and response preparation, respectively. Process B is sometimes
called the bottleneck process. Stimulus onset asynchrony is denoted
SOA, and SW represents the switching of attention from the central
processing of Task 1 to the central processing of Task 2. The model is
essentially that of Davis (1957); process SW was called by him central
refractoriness.
One prediction is about interactions of factors with the special factor
of single vs. dual task responding. Pashler and Johnston (1989) followed
108 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Table 5.3
Reaction Time Means (Standard Deviations) in Milliseconds
Pashler and Johnston (1989) Experiment 1
results from the later and earlier phases of the experiment allows
comparison of conditions (1, 1), (1, 2), (2, 1) and (2, 2), where the
number of alternatives for s1 is listed first, followed by those for s2.
From Figure 4 of Karlin and Kestenbaum, it is clear that the
interstimulus interval had little or no effect on RT1, so the interstimulus
interval does not precede r1. It is also clear that at every interstimulus
interval, reaction time to s1 was faster with one alternative for s2. This
could indicate a violation of selective influence. At long interstimulus
intervals, the second stimulus is usually presented after the response to
the first stimulus is made, so the results could indicate that the number of
s2 alternatives has effects in two places, in processing of Task 1 and in
processing of Task 2, a violation of selective influence. But the results
could simply be due to subjects having more practice by the time they
started the later sessions with one alternative for s2.
Mean reaction times to the second stimulus are as expected if the
interstimulus interval and central Task 2 processing are sequential. First,
RT2 decreases monotonically as ISI increases. Second, if the ISI value is
added to RT2 (to obtain the time from s1 to r2), resulting times increase
monotonically with ISI. Finally, interaction contrasts are all nonnegative,
except for a few negligibly less than 0.
Clearly, there is slack for the interstimulus interval, for example,
when the ISI is increased by 100 msec, from 90 to 190 msec, RT2
increases by only 2 msec in the (1, 2) condition. There is also slack for
the number of s2 alternatives. When the ISI is 1150 msec, changing the
number of s2 alternatives from 1 to 2 increases RT2 by 90 msec, but
when the ISI is 90 msec the increase is only 39 msec. Such slack has
puzzled investigators; it requires a path from s1 presentation to r2 that
does not include process B2. One possibility is that B2 is concurrent
with B1, so that A1, B1, SW, C2 form a path from s1 to r2 (a
modification of Figure 5.2). Another possibility is a path provided by a
second bottleneck toward the response end of the system, as in de Jong’s
(1993) model (Figures 5.3, 5.4).
A numerical example may be helpful. Using the hypothetical process
durations in Figure 5.4, the effect of prolonging process B2 depends on
the SOA. Suppose SOA = 90. When B2 = 100, the critical path to r2 is
Critical Path Models of Dual Tasks and Locus of Slack Analysis 113
Stimulus 2 discriminability
Experiment 1 of Johnston and McCann (2006) was described in the
earlier section on Task 1 Central Processing and SOA. Task 2
discriminability and SOA were also investigated. Recall that Task 1 was
auditory or visual (in separate blocks). Regardless of Task 1 modality,
there was no effect of Task 2 discriminability on RT1. The effect on RT2
increased monotonically as SOA increased. That is, for RT2, Task 2
discriminability had a positive interaction with SOA. The positive
interaction indicates that Task 2 discriminability selectively influenced a
Task 2 process sequential with SOA; logically, this process, which they
call classification, follows the SOA. Details of the interaction lead
Johnston and McCann (2006) to say that ordinarily Task 2 classification
follows Task 1 central processing (i.e., is after the bottleneck), but for
some subjects on some trials it does not. Their statement is not
parsimonious, but not easily rejected because a mixture of networks that
produce additivity with other networks that produce a positive interaction
would yield a positive interaction.
In later experiments, Johnston and McCann (2006) presented a
rectangle as s2, to be judged as narrow or wide. Discrimnability was easy
or hard. The effect of discriminabilty was additive with SOA for RT2.
Interpretation is straightforward. They concluded that discriminability
selectively influenced a Task 2 process which followed both the SOA
and central processing of Task 1. They say this classification process is
the earliest known Task 2 process to follow Task 1 central processing.
button press with the left hand, one button for each tone. The stimulus
for the second task was B or D. Response 2 was a button press with the
right hand in one condition (hand-hand) and a pedal push with the left or
right foot in another condition (hand-foot). The number of response 2
alternatives was manipulated. In the one-alternative condition, for both
B and D the subject pressed the same button or pushed the same pedal.
In the two-alternative condition, each letter was assigned a different
button or pedal. The SOA was 25, 250 or 800 msec. Summary data are
in Figure 5.5 (de Jong’s, 1993, Fig. 4), and for RT2, in Table 5.4, as read
from that figure.
Table 5.4
Mean RT2 (msec) in de Jong’s (1993) Experiment 2
Let’s start with statistical details. For RT1, there was a significant
effect of the number of Task 2 alternatives. We will not pursue it
because it is small and does not affect our conclusions about RT2, but
perhaps the number of alternatives for Task 2 influenced processes in
Task 1 as well as in Task 2, a violation of selective influence.
For RT2, there were significant effects of SOA, number of Task 2
alternatives, and response modality. There was a significant positive
(underadditive) interaction between SOA and number of Task 2
alternatives and a significant positive interaction between number of
Task 2 alternatives and response modality. Numerically, there was a
positive (underadditive) interaction between SOA and response modality,
nearly significant (p < .06).
The three way interaction of SOA, number of Task 2 alternatives and
response modality was significant. In a separate ANOVA on only trials
with a foot response, SOA and number of Task 2 alternatives had
additive effects.
An important result for de Jong’s purpose is the significant positive
(underadditive) interaction between SOA and number of Task 2
alternatives; it supports the existence of a response bottleneck under his
experimental conditions. Compare the critical path networks in Figures
5.2 and 5.4. If as in Figure 5.2 there is a cut vertex between the SOA and
process B2 (selectively influenced by the number of Task 2 alternatives),
then the factors would always have additive effects. Instead, they have
additive effects only when Response 2 is made with the foot, when
presumably process C2 is long. Results are explained by the network in
Figure 5.4. When process C2 is short, process B2 has slack because of a
Critical Path Models of Dual Tasks and Locus of Slack Analysis 117
Table 5.5
Mean RT2 + SOA (msec) in de Jong’s (1993) Experiment 2
Also for each pair of factors, the interaction contrasts are predicted to
be positive and monotonically increase with factor levels. We check
them with Table 5.5. The lowest level of each factor, used as baseline
for the following calculations, is SOA 25, one Task 2 alternative, and
hand response. (Rounding is done after, rather than before, calculations.)
For SOA and Number of Task 2 Alternatives, interaction contrasts
are, for SOA = 250,
k(SOA, B2) = 80; k(SOA, C2) = 55; and k(B2, C2) = 69.
Then, if the processes are in the order SOA, B2, C2 we predict the
reaction time for prolonging all three processes from the effect of
prolonging each process individually and two of the three coupled slacks
(see the Appendix):
Critical Path Models of Dual Tasks and Locus of Slack Analysis 119
The left and right sides are equal (within rounding). The coupled slack
not used in the above equation, k(SOA, C2), corresponds to the first and
last process in the sequence of three. No other order fits the data so well,
so we conclude B2 is the process in the middle.
In addition to demonstrating once again that SOA precedes central
Task 2 processing, de Jong (1993) found that the factor of responding by
hand vs. by foot selectively influences response preparation for Task 2.
He showed that central processing and response preparation for Task 2
are sequential, and showed that in some circumstances there is a path
concurrent with them. Quantitative analysis shows central processing for
Task 2 is between the SOA and response preparation for Task 2.
Pashler (1984) presented a light bar for Task 1, and the response was a
button press with the left hand to identify whether the bar was high or
low in position. Task 2 was a visual search for a letter. There were two
Task 2 factors. The contrast of the display of letters to be searched
through for Task 2 was high or low, and the target letter was either
present or absent. The visual search was done either alone or in a dual
task, in different blocks. In the dual task, subjects were instructed to
respond for Task 1 before responding for Task 2. The SOA was fixed at
100 msec.
120 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Whether the visual search was done alone or as one of the dual tasks,
s2 contrast and target presence/absence had additive effects on search
task reaction time. The additivity supports the hypothesis that the two
factors selectively influenced two different processes.
Pashler (1984) compared single and dual task conditions to obtain
information about process arrangement. The effect of contrast was
greater in the single task condition than in the dual task condition. If we
assume the prolongation produced by decreasing contrast was the same
in the single and dual task conditions, then, because the effect was
smaller in the dual task condition, there was slack in the dual task
condition for the process selectively influenced by contrast. With the
notation of Figure 5.2 the slack is expected if s2 contrast selectively
influenced process A2. However, the effect of target presence/absence
was the same in the single and dual task conditions. If we again assume
the prolongation was the same in the single and dual task conditions,
there was no slack in the dual task condition for the process selectively
influenced by target presence/absence. With the notation of Figure 5.2,
the additivity is expected if target presence/absence selectively
influenced B2.
In another experiment on A2 and B2, Pashler and Johnston (1989)
found additive effects of s2 intensity and s2 repetition; this is explained if
these factors selectively influence A2 and B2; the work was discussed in
the section on SOA and B2.
Central processing for the two tasks of a dual task has been investigated
in the Psychological Refractory Period (PRP) paradigm and in Stroop-
like tasks, where conflict is induced by the stimuli.
PRP: Discriminability
We introduced Johnston and McCann’s (2006) experiment in the section
on Task 1 central processing and SOA. Discriminability of stimulus 1
was either easy or hard, as was discriminability of stimulus 2. There was
no effect of Task 2 discriminability on RT1. But Task 1 discriminability
and Task 2 discriminability had additive effects on RT2, indicating that
these two factors selectively influenced sequential processes.
groups. For the three original groups combined with the new group,
increased memory set size and target absence increased the reaction time
for each task, but subtraction difficulty only increased the reaction time
for the arithmetic task. Results can be explained with a simple model in
which the two tasks were carried out one after the other, the memory
search preceding the subtraction.
Stroop tasks
In the Stroop (1935) task, the name of a color is presented in colored
print, with the color of the print to be named. Typically, responses are
slower and less accurate when the name and print color conflict than
when they agree.
Response time includes time for all processing except set up prior to the
stimulus and tear down after the response. A more discriminating
analysis can be made by referring to a time mark within the task. Mental
processes are not directly observable, but some are accompanied by
Critical Path Models of Dual Tasks and Locus of Slack Analysis 127
was to identify a high or low pitched tone by pressing a button with the
left or right foot. The second subtask was to press a button with the left
or right index finger to identify the second stimulus as an X or an O. The
SOA values were 50, 200 and 500 msec. The SOA was constant
throughout a block of trials. Note that Task 1 and Task 2 use different
sense modalities and different responding limbs. A delay in Response 2
produced by performance of Task 1 cannot be due to conflicting
demands for the same organ or limb.
Figure 5.6 (adapted from Osman & Moore, 1993, Figure 1) illustrates
predictions about the effect of the SOA on the onset of the Task 2 LRP
(LRP2-onset). The top panel illustrates the hypothesis that preparation
of the second response begins in the processing prior to the delayed
central process B2. In that case, the interval from s1, presentation of the
first stimulus, to LRP2-onset would not change with SOA. But the
interval from LRP2-onset to r2, the second response, would decrease as
SOA increases.
The bottom panel illustrates the hypothesis that preparation of the
second response begins sometime during the delayed central process B2,
or (not illustrated) sometime during response process C2. In the bottom
panel case, the interval from s1 to LRP2-onset would decrease as SOA
increases. But the interval from LRP2-onset to r2 would not change as
SOA increases.
Osman and Moore (1993) also state that if the hypothesis considered
in the bottom panel is true, the interval from the completion of the central
processing for Task 1 to r1, the response for Task 1, is less than or equal
to the interval from LRP-onset to r1. Further, if processing is as
illustrated in either panel, and the LRP for the first task arises during the
central process B1, then neither the interval from s1 to LRP1-onset, nor
the interval from LRP1-onset to r1 would change as SOA changes.
We have space only for major results; see Osman and Moore (1993)
for more details. As usual in dual tasks, mean RT2 decreased as SOA
increased. Contrary to both models illustrated in Figure 5.6 there was a
significant effect of SOA on mean RT1; because it is small it does not
affect the main conclusions.
Critical Path Models of Dual Tasks and Locus of Slack Analysis 129
Fig. 5.6. Top. Model for Task 2 Lateralized Readiness Potential (curved line, LRP)
arising during Sensory Processing of stimulus 2 (A2). Time from s2 to LRP onset is
invariant with SOA. Bottom. Model for Task 2 Lateralized Readiness Potential arising
during Central Processing of stimulus 2 (B2). Time from LRP onset to r2 is invariant
with SOA. Notation as in Figure 5.2. Based on Osman & Moore (1993) Figure 1.
Fig. 5.7. Vincentized cumulative distribution functions (CDFs) of reaction times for the
first (RT1) and second (RT2) tasks. (CDFs are shown for each task at each stimulus onset
asynchrony [SOA], for each task on single-task blocks, and for the first task on catch
trials. [Note: The first-task CDFs for the long SOA and catch trials are so similar that
they are difficult to distinguish from a single CDF.] From Osman, A., & Moore, C. M.,
1993, The locus of dual-task interference: Psychological refractory effects on movement-
related brain potentials. Journal of Experimental Psychology: Human Perception and
Performance, 19, Fig. 6. Copyright 1993 American Psychological Association.
Reproduced with permission.
not important for the purposes of Osman and Moore (1993), the
cumulative distribution functions test an assumption important for
analysis of selective influence by increments on reaction time. The
cumulative distribution functions do not cross, except at the tails for
RT1. Except for these tails, reaction times for the SOA and the trial
types of single and catch support the “usual stochastic ordering”
assumption (also called stochastic dominance, see Chapter 4).
Critical Path Models of Dual Tasks and Locus of Slack Analysis 131
Fig. 5.8. Stimulus (S)-locked and response (R)-locked lateralized readiness potentials
(LRPs) for the first and second tasks. (LRPs are shown for each task at each stimulus
onset asynchrony, for each task on single-task blocks, and for the first-task on catch
trials. The top portion of each panel shows the LRPs, and the bottom portion shows the
effects on these LRPs of which response occurred in the other task. S1 = onset of first
stimulus; S2 = onset of second stimulus; RT1 = first reaction time; RT2 = second reaction
time.) From Osman, A., & Moore, C. M., 1993, The locus of dual-task interference:
Psychological refractory effects on movement-related brain potentials. Journal of
Experimental Psychology: Human Perception and Performance, 19, Fig. 7. Copyright
1993 American Psychological Association. Reproduced with permission.
The LRPs are in Figure 5.8 (Osman & Moore, 1993, Figure 7). A
question arises whenever brain potentials are averaged over trials at each
time t. The potentials are measured throughout the performance of the
task, but performance on some trials stops sooner than on others. All
trials are available for averaging at stimulus onset, but as time goes on
fewer and fewer trials remain available. Osman and Moore (1993)
followed a common procedure. When the early part of the potential is
132 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
between the R-locked LRPS (locked to R2). And in particular, SOA did
not have a statistically significant effect on the R-locked LRP-onsets. In
summary, effects are as expected if the LRP onset for the second
response occurs within the central processing of Task 2 (called B2 in
Figure 5.6).
between the end of B1 and the start of C1, process FTb is structurally
equivalent to process SW in the Single Central Bottleneck Model.
We discuss two predictions Jentzsch, et al. (2007) tested for trials
when r1 is completed before s2 is presented. First, a factor selectively
influencing A2 will have an underadditive interaction with SOA (i.e., a
positive interaction). Second, a factor selectively influencing B2 will
have additive effects with SOA. (See Chapter 3 for the basis of these
predictions.)
block of trials.
Trials on which RT1 was less than both SOAs (i.e., less than 400
msec) were analyzed. As one would expect, there was no effect of SOA
or s2 contrast on RT1. Also, as one would expect, there was an effect of
s2 contrast on RT2. Of importance for the authors’ purposes, when SOA
increased, RT2 decreased; that is, there was a residual PRP effect.
Moreover, the combined effect of SOA and s2 contrast was as predicted
from the Extended Response Selection Bottleneck model; that is, there
was an underadditive interaction (a positive interaction), as predicted if
the source of the residual PRP effect is a process (or more than one
process) concurrent with both SOA and A2 (see Figure 5.9). Results do
not distinguish between possibilities of FTb, or FTe, or both in Figure
5.9. We will speak of a single process FTx, keeping mind that there may
be more than one.
Clearly the required process FTx must precede r2. Further
information comes from LRP analysis. Suppose FTx must precede the
onset of the LRP for r2. Predictions are that (1) increasing the SOA will
decrease the interval between s2 and LRP-onset, and (2) prolonging A2
by changing the contrast of s2 will increase the interval between s2 and
LRP-onset. But neither manipulation will change the interval between
LRP-onset and r2. Results were as predicted. Note that the results have
the same form as found by Osman and Moore (1993) (although all their
trials were combined).
On the other hand, suppose as before FTx is concurrent with SOA and
A2, but FTx need not precede the LRP-onset. There are several ways this
could happen while still satisfying the constraint that FTx precedes r2.
Suppose, for example, that the LRP-onset is an event within process B2,
and FTx is concurrent with B2, but precedes C2. (In that case, FTx
would be the refractory interval in the model of de Jong, 1993, see
Figure 5.4). Because FTx is relatively long, such an arrangement
predicts that neither SOA nor s2 contrast will affect the interval between
s1 and LRP-onset. However, such an arrangement also predicts that
increasing SOA will decrease the interval between s2 and LRP-onset, as
will increasing s2 sensory processing difficulty by changing s2 contrast.
Results did not satisfy these predictions. The conclusion is that FTx
136 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
precedes the LRP-onset. Because the model makes the same predictions
for trials on which r1 follows s2, analyses were also done for all trials
combined. Those analyses are consistent with those we summarize here;
see Jentzsch, et al (2007) for details.
effect on RT2; that is, there will be no residual PRP effect. (2) The SOA
is short enough that RI is relevant, at least sometimes. In that case,
further increase in SOA can lead to a decrease in RT2 (a residual PRP
effect can occur); moreover, increasing SOA and prolonging B2 will
have underaddive effects (a positive interaction). With such a model, it
is not possible to observe both (1) a residual PRP effect and (2) additive
effects of SOA and a factor selectively influencing B2. With the
extended selection bottleneck model, on the contrary, both effects can
occur together.
Let us assume, as is generally agreed, that stimulus-response
compatibility for Task 2 (s2-r2 compatibility) selectively influences
process B2. Experiment 3 of Jentzsch, et al (2007) tested the effects of
SOA and s2-r2 compatibility. Most aspects of the design were the same
as in their Experiment 2. In particular, the SOAs were 400 and 600
msec. And at low s2-r2 compatibility, s2 was either an X or an O,
presented in high contrast. A new aspect of Experiment 3 was that at
high s2-r2 compatibility, s2 was an arrow pointing either left or right, in
the direction of the response to be made. The various s2 stimuli were
randomly presented in each block, as were the two SOAs. (There was no
electrophysiological recording.)
With all trials combined, SOA and s2-r2 compatibility had additive
effects on RT2, indicating that SOA and B2 are sequential processes with
no long path of processes concurrent with them. This is what McCann &
Johnston (1992) found. Of particular interest is what happened when
Response 1 preceded presentation of Stimulus 2.
For trials in which RT1 was less than the shorter SOA (400 mesc),
increasing SOA significantly decreased RT2, a residual PRP effect.
Further, decreasing s2-r2 compatibility significantly increased RT2.
However, these did not interact. (For these trials, RT1 was not affected
by SOA or by s2-r2 compatibility.)
In conclusion, the extended selection bottleneck model in Figure 5.9
is supported and models with a single bottleneck sufficiently late that B1
need not precede B2 are not supported.
138 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
of the first response, let r1 denote the onset of the first response, and then
add a further process following r1 to denote movement; call it MT1. The
end of this further process is the response offset, and its duration is the
motor time.
In the Single Central Bottleneck model, the last Task 1 process to
precede r2 is B1, the central processing of Stimulus 1. According to the
model, Response 1 preparation and movement occur too late to have an
effect on response time 2. However, increasing the movement time of
Response 1 increased RT2 in an experiment by Ulrich, Fernández,
Jentzsch, Rolke, Schröter and Leuthold (2006). They explain their
results by modifying the Single Central Bottleneck model so Response 1
offset precedes Task 2 response preparation; that is, MT1 precedes C2.
In the modified model, the movement following preparation for each
response is represented.
In more detail, Task 1 was identification of a tone as high or low
pitched. The response was made with the left hand by sliding a handle
along a smooth track on a box, forward for one tone and backward for
the other tone. The response ended when the handle reached the end of
the box; the starting position of the handle made the distance longer to
one end than to the other end. The stimulus for Task 2 was an X or an O;
the subject identified it by pressing a button with a finger of the right
hand. Response 2 movement time was not measured.
Results are in Figure 5.10 (Ulrich, et al., 2006, Figure 2). The middle
panel shows the movement time for response 1 (MT1); it is longer when
the response distance is longer, as intended with the apparatus. The time
of onset of Response 1, RT1, is actually shorter when the distance to be
moved is longer. Ulrich, et al (2006) explain that greater force is needed
to start a long movement than a short one (e.g., Schmidt, Zelaznik,
Hawkins, Frank & Quinn, 1979), and greater force produces a shorter
time for response onset criterion to be recorded (Ulrich & Wing, 1991).
The crucial finding is that Response 2 reaction time is longer when
the movement time for Response 1 is longer, contrary to the Single
Central Bottleneck model, but explained by the modification of Ulrich, et
al. (2006). Recall that with their modification, the movement for
Response 1, MT1, immediately precedes preparation for response 2, C2.
140 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 5.10. Task 1 reaction time (RT1; top panel), Task 1 movement time (MT1; middle
panel), and Task 2 reaction time (RT2; bottom panel) as a function of stimulus onset
asynchrony and movement distance of the Task 1 response. From Ulrich, R., et al., 2006,
Motor limitation in dual-task processing under ballistic movement conditions,
Psychological Science, 17, Fig. 2. Copyright 2006 Wiley. Reproduced with permission.
Response grouping
Responses in dual tasks are not tidy with respect to factors selectively
influencing processes because subjects sometimes delay the first
Critical Path Models of Dual Tasks and Locus of Slack Analysis 141
response and make the two responses close together. This is called
grouping. Grouping is comfortable for some reason. Control is more
efficient if two responses are produced together (Rinkenauer, Ulrich &
Wing, 2001). Control may be intermittent (Craik, 1948), so responses
may be made together to bundle signals. Nothing observable indicates
whether a subject grouped responses or simply made them close together
because the tasks happened to finish at about the same time. Grouping
makes interpreting response times difficult, particularly those to the first
stimulus.
Because there is little theory or data for guidance, Ulrich and Miller
(2008) used simulations to study a grouping model and some of its
variants. They assumed that when the subject did not group, he or she
used the Single Central Bottleneck model. When the subject grouped, he
or she used a model proposed by Borger (1963) as formulated by Pashler
and Johnston (1989). Borger’s model is in Figure 5.11 (based on Ulrich
& Miller, 2008, Figure 2). The duration of the interval between r1 and
r2 is a nonnegative random variable, D. For response time 2, with
Borger’s model it essentially makes no difference whether subjects group
or not. With the grouping schedule, response time 2 is
Fig. 5.11. The Borger (1963) model for response grouping. Motor preparation for both
responses begins after B2. Notation as in Figure 5.2, except duration of response 1 motor
preparation is C1'. Duration of response 2 motor preparation consists of the same amount
of time C1' plus a nonnegative random variable D. (Based on Ulrich and Miller, 2008,
Figure 2.)
RT1 = A1 + B1 + C1.
Response Time 1
In the simulations, as SOA increased RT1 decreased for some models
(including BorS1 and WA mentioned above) and increased for other
models (including WB mentioned above).
Remarks
Appendix
Fig. 5A.1. If the rate of concurrent processing is slower than that of sequential
processing, then the time at which the first process is finished is later with concurrent
processing.
The term k(A, C) does not appear, and its absence indicates that A and C
are at the extremes with B in the middle. Each coupled slack parameter
arises from slack between two processes. The three processes have only
two spaces between them, so one of the coupled slacks is irrelevant.
Derivation in Schweickert (1978).
Suppose process A is prolonged by ΔA, which is larger than both s(A, r1)
150 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
and s(A, r2). Let t1(ΔA, 0) denote the response time for r1 when A is
prolonged by ΔA, and denote other times similarly. Then (see Chapter 3)
Hence,
t1(0, ΔB) − t1(0, 0) + k1(A, B) = t2(0, ΔB) − t2(0, 0) + k2(A, B). (5A.3)
Both Equations (5A.2) and (5A.3) can be true; this happens for example
if A and B are in series. But if Equation (5A.2) is true and Equation
(5A.3) is false, then process A precedes process B.
If process A precedes process B, an equation similar to Equation
(5A.2) is useful because it does not require estimates of coupled slacks,
Analyzing comparability
Ashby and Townsend (1980) found a simple but important relation for
cumulative distribution functions for processes in series. It is easy to
explain the relation intuitively. With our usual notation, suppose a task is
performed by executing process A followed by process B. Suppose
Factor Α changes the duration of process A, leaving process B
unchanged, and Factor Β changes the duration of process B, leaving
process A unchanged. Suppose when Factor Α is at level i and Factor Β
is at level j the response time is Tij = Ai + Bj. Additive effects of the
factors on mean reaction times are predicted, that is E[T22] + E[T11] =
152
Effects of Factors on Distribution Functions 153
and
The two sums consist of the same terms in different orders. This
suggests that under suitable assumptions,
where “≈” means “has the same distribution as.” That is, for every time
t, if we let FT22+T11(t) and FT12+T21(t) denote the cumulative distribution
functions of T22 + T11 and T12 + T21, respectively, then for every time t
FT22+T11(t) = FT12+T21(t).
Roberts and Sternberg (1993) proposed a test based more directly on the
sampled response times. Their test was derived using the assumption
that for every combination (i, j) of levels of the factors Α and Β, the
durations Ai and Bj are independent. Independence is a sufficient, but not
necessary condition for their test of FT22+T11(t) = FT12+T21(t); see their
chapter for discussion. Independence of process durations is a strong
assumption, but there is evidence that it sometimes occurs. Sternberg
(1969) pointed out that factors selectively influencing processes in series
with independent process durations would have not only additive effects
on means, but on variances and cumulants at all levels. Additive effects
of factors on both means and variances were reported by Sternberg
(1969, p. 305); further reports are mentioned in the section on Process
Dependence below.
We first describe the test of Roberts and Sternberg (1993) and then
give their explanation of it. To start, every possible pair is formed with
the first member of the pair a response time from the set of response
times observed in combination (2,2) of factor levels and the second
member of the pair a response time from the set of response times
observed in combination (1,1). This is the Cartesian product of the two
response time sets. Then every pair of response times in this Cartesian
product is added. The result is a sample of values of T22 + T11. From this
sample an estimate of the distribution function FT22+T11(t) is formed; for
Effects of Factors on Distribution Functions 155
Fig. 6.1. Results of Summation Test. Panel A: detection data. Panels B and C:
identification data with 2 and 8 alternatives, respectively. Left of each panel: cdfs for
four factor level combinations. Right of each panel: cdfs for summations. Note: From
Roberts and Sternberg (1993), The meaning of additive reaction-time effects: Tests of
three alternatives. In Meyer, David E., and Sylvan Kornblum (Eds.), Attention and
Performance XIV: Synergies in experimental psychology, artificial intelligence, and
cognitive neuroscience, figure 26.2. Copyright 1993, Massachusetts Institute of
Technology, by permission of The MIT Press.
Effects of Factors on Distribution Functions 157
Subjects were very highly practiced, and data were combined over
different sessions. It is likely that the durations of both processes A and
B differ for different subjects; in other words, different subjects may
induce a covariance between process durations. Likewise, the different
stimulus numerals might induce a covariance. To avoid this, calculations
were done for combinations of levels of nuisance factors separately, then
averaged over combinations. For example, with eight alternatives in the
identification experiment, calculations were done for each subject and
numeral separately, then averaged over subjects and numerals. Before
averaging, observations were rescaled with a linear transformation
intended to increase the sensitivity of the test. See Roberts and Sternberg
(1993) for details.
Then
E[Tij ] tpf Ai (t )dt t (1 p) f Bj (t )dt pE[ Ai ] (1 p) E[ B j ].
0 0
1 1
[ F22 (t ) F11 (t )] [ F21 (t ) F12 (t )] .
2 2
Roberts and Sternberg (1993) call this the Mixture Test. Note that the
cumulative distribution function for each combination of levels can be
estimated separately; that is, there is no need for the Cartesian products
used in the Summation Test. Results are in Figure 6.2 for the same three
data sets as the Summation Test was applied to in Figure 6.1. Clearly,
Effects of Factors on Distribution Functions 159
the mixture test fails, so the Alternate Pathways Model can be rejected
for these data. See Roberts and Sternberg (1993) for more details about
carrying out the test.
The Alternate Pathways Model and the independent serial processes
(stages) model lead to contrary predictions. For example, they make
different predictions about variances. With independent serial processes,
Fig. 6.2. Results of Mixture Test. Panel A: detection data. Panels B and C:
identification data with 2 and 8 alternatives, respectively. Note: From Roberts and
Sternberg (1993), The meaning of additive reaction-time effects: Tests of three
alternatives. In Meyer, David E., and Sylvan Kornblum (Eds.), Attention and
Performance XIV: Synergies in experimental psychology, artificial intelligence, and
cognitive neuroscience, figure 26.4. Copyright 1993, Massachusetts Institute of
Technology, by permission of The MIT Press.
160 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
For the Alternate Pathways Model, with levels 1 and 2 of each factor,
the interaction contrast for the variances is
The factors will not have additive effects on the variance except trivially,
when one or the other has no effect on the mean.
Interestingly, the Alternate Pathways Model predicts the factors will
have additive effects on the squared reaction times (the second raw
moments). Reasoning as with expected values, additivity follows from
Data cannot simultaneously satisfy both the Summation Test and the
Mixture Test, except trivially, when, for example, there are no effects.
The data sets analyzed by Roberts and Sternberg (1993) passed the
Summation Test but failed the Mixture Test. To consider the opposite
situation, Roberts and Sternberg (1993, Note 20) generated simulated
data that pass the Mixture Test. Failure of the Summation Test is
predicted, but would the failure be noticeable?
To ensure realistic effect sizes, variances and so on, simulated data
were generated from the actual data sets discussed earlier, which passed
the Summation Test. Recall that for each data set there were two
experimental factors, each with two levels. In a particular data set, let Tij
denote the reaction time when the first factor is at level i and the second
at level j. Simulated data for the four factor level combinations were
generated as follows. The observed reaction times T11 and T22 were used
directly as simulated reaction times T*11 and T*22, respectively. Then the
observed reaction times T11 and T22 were pooled, and the pool randomly
divided into two halves. One half was used as T*12, the other half as
T*21. For each data set, the procedure was carried out once for each
Effects of Factors on Distribution Functions 161
1 1
F *12 (t ) F *21 (t ) F *11 (t ) F *22 (t ) ;
2 2
so,
Table 6.1
Statistical Tests for the Summation Test
Detection 42 0 39 29
Numeral Naming
2 Alternatives 66 0 69 37
8 Alternatives 61 0 91 57
Note: From Roberts & Sternberg (1993, Note 20) and S. Sternberg (personal
communication, June 21, 2011).
Fig. 6.3. The Summation Test fails for Simulated Data by Roberts and Sternberg that pass
the Mixture Test: Detection. S. Sternberg (personal communication, June 21, 2011).
Effects of Factors on Distribution Functions 163
Fig. 6.4. The Summation Test fails for Simulated Data by Sternberg and Roberts that pass
the Mixture Test: Identification, 2 Alternatives. S. Sternberg (personal communication,
June 21, 2011).
Fig. 6.5. The Summation Test fails for Simulated Data by Sternberg and Roberts that
pass the Mixture Test: Identification, 8 Alternatives. S. Sternberg (personal
communication, June 27, 2011).
164 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Statistical mimicking
investigator finding mean additivity might decide that the most plausible
explanation is that the factors selectively influence serial stages. It
would be natural for the investigator to use the Summation Test as a
further diagnostic. But with the simulation it led to erroneous
confirmation of serial stages.
There are warnings to the investigator in the simulated data. One
warning is large reaction time variances, signaling low power. A
measure of variability is the coefficient of variation, the standard
deviation divided by the mean. For reaction times, it is typically less
than .5 (some examples are in Luce, 1986). For the three data sets in
Figure 6.1 A, B and C, the coefficients of variation are .11, .14 and .08,
respectively, from the grand means and variances in Tables 26.2 and 26.3
of Roberts and Sternberg (1993). But for the simulated reaction times it
is unusually large, .86, when both factors are at their high levels. This
suggests an unusually large number of observations would be needed to
achieve good power. Indeed, an ANOVA on variances failed to detect
the interaction that is present for the Alternate Pathways Model. Another
warning to the investigator is that both the Summation Test and Mixture
Test were passed, theoretically not possible. The simulations
demonstrate that an investigator ignoring the warnings could be misled;
visually the agreement between the two sides of the Summation Test is
remarkably good for the Alternate Pathways Model simulation (Van
Zandt & Ratcliff, 1995, Figure 14).
To summarize, the distribution equality tests of Ashby and Townsend
(1980) and Roberts and Sternberg (1993) are valuable because they
provide radically different information from the usual analysis of mean
reaction times. Success is evidence that factors selectively influence
different serial processes, whose durations are stochastically
independent. Failure is evidence that one or the other assumption is
wrong. In particular, failure together with additive effects of the factors
on mean reaction times could be the result of factors selectively
influencing serial processes with dependent response times. Van Zandt
and Ratcliff (1995) demonstrate that, as with all tests, attention to both
incorrect acceptance and incorrect rejection is required.
166 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Tij = Ai ♦ Bj.
E[T] = S (t ) dt ,
0
(6.2)
For all levels i and j, for every t, the survivor function for the duration
of process A does not depend on j,
SAij(t) = SAi(t).
SBij(t) = SBj(t).
Likewise, the levels of Factor Β can be ordered j = 1, 2,... so that if j < j'
then for every time t
With two processes there are two cases, parallel and serial, both
considered by Townsend and Nozawa (1995). The following theorem
based on their work describes the survivor interaction contrast for two
parallel processes.
(d) The same conclusions follow for C(t), but with the signs reversed.
The probability that the task is not finished at time t is the probability
that A is not finished at time t and B is not finished at time t. With
marginal selective influence and stochastic independence, the survivor
function of the reaction time is
STij(t) = SAi(t)SBj(t).
distribution function is
t
FT (t ) P[T t ] FA (t x) f B ( x)dx FA (t x) f B ( x)dx. (6.4)
0 0
t t
FXi (t ) f
0
Xi (t ' )dt ' FXi* (t ) f
0
Xi* (t ' )dt ' (6.5)
For all t the first multiplier in the integrand is nonnegative and for all x in
the interval (0, τ) the second multiplier is positive. Hence for t in the
interval (0, τ) SIC(t) < 0.
174 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
SIC (t )dt 0.
0
Pij[A > a and B > b|C = c] = Pij[A > a|C = c]Pij[B > b|C = c].
Likewise, the levels of Factor Β can be ordered so if j* < j then for every
Effects of Factors on Distribution Functions 175
time t
The proof is similar to that of Theorem 6.1 here, see their Theorem 1, A-
4 for details.
For Theorem 6.2, the conclusions are true for processes A and B in
series, but with stochastic independence not required, and stochastic
dominance weakened to conditional stochastic dominance, stated with
notation of this case as:
P[T t ] FA (t x | B x ) f B ( x )dx.
0
When the expression for P[T < t] now in the proof of Theorem 6.2 is
replaced by the expression above, and conditional stochastic dominance
is assumed, the conclusions follow immediately.
Task networks
A glance at Figures 6.7 and 6.9 shows how survivor interaction contrasts
differ for concurrent and sequential processes. The following
summarizes results for nonzero interaction contrasts. Simulations
illustrating the results are described in detail in a later section and in the
Appendix.
For concurrent processes, the survivor interaction contrast is simple;
it is positive with OR gates and negative with AND gates. For the latter,
see Figures 6.6 and 6.7. For sequential processes in a serial-parallel
network with OR gates, the survivor interaction contrast is negative over
an interval close to time 0. It may or may not change signs as t increases.
For any time t greater than 0, the area from 0 to t bounded by the
survivor interaction contrast and the time-axis is negative.
For sequential processes in a serial-parallel network with AND gates,
the survivor interaction contrast is negative over a short interval after 0
and then changes sign at least once. For any time t greater than 0, the
area from t to ∞ bounded by the survivor interaction contrast and the
time-axis is positive. See Figures 6.8 and 6.9.
In a Wheatstone bridge, when a pair of sequential processes are
selectively influenced the results are the same as if they were in a serial-
parallel network, with one exception. For two sequential processes on
opposite sides of the bridge the survivor interaction contrast at small
values of t is positive with OR gates (Figures 6.10 and 6.11) and
negative with AND gates (Figures 6.12 and 6.13).
For sequential processes on opposite sides of the bridge, the survivor
interaction contrast need not change signs as prolongations become long.
In simulations with long prolongations with OR gates, it remained
positive (Figures 6.14 and 6.15). However, in simulations with AND
gates when prolongations are large, the survivor interaction contrast can
be made to change sign for large values of t (Figures 6.16 and 6.17).
These results do not require independent process durations. The figures
are from simulations with dependent process durations, described below.
The remainder of the chapter is organized as follows. Results for
survivor interaction contrasts are stated in more detail. A proof assuming
178 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
This section fills in details of the synopsis above and can be skipped
without loss of continuity. As before, the density function for a random
variable X is denoted by fX(t), the cumulative distribution function by
FX(t), and the survivor function by SX(t). A random variable X will be
indexed by a level k of a factor if the distribution of X depends on the
level of the factor, or potentially might depend on it. For example, if
Factor Α has level i and Factor Β has level j, then Tij denotes the reaction
time, a random variable, when Factor Α has level i and Factor Β has level
j. The cumulative distribution function of Tij is usually denoted GTij(t)
with corresponding density gTij(t). In an independent network, process
durations are mutually stochastically independent.
Each inequality for a survivor function holds for the corresponding
cumulative distribution function with the direction of the inequality
reversed. Likewise, each inequality stated for the survivor interaction
contrast, SIC(t), holds for the cumulative distribution interaction contrast,
C(t), with the direction of the inequality reversed. More details can be
found in Schweickert, Giorgini and Dzhafarov (2000), Schweickert and
Giorgini (1999) and Dzhafarov, Schweickert and Sung (2004).
Fig. 6.6. Survivor functions of simulated reaction times. Factors selectively influence
concurrent processes, B1 and A2 in the Response Selection Bottleneck Model. Functions
for (Low, High) and (Mid, High) are nearly superimposed, because reaction time is
governed by the maximum of the prolongations of B1 and A2. Parameters: Unique
component of B1 has α of 4 and 36 at Low and Mid levels, respectively. Unique
component of A2 has α of 4 and 96 at Low and High levels, respectively.
Fig. 6.7. Interaction contrasts for survivor functions of concurrent processes in Figure
6.6 are never positive.
180 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 6.8. Survivor functions of simulated reaction times. Factors selectively influence
sequential processes, A1 and B1 in the Response Selection Bottleneck Model.
Parameters: Unique component of A1 has α of 4 and 96 at Low and High levels,
respectively. Unique component of B1 has α of 4 and 36 at Low and Mid levels,
respectively.
Fig. 6.9. Interaction contrasts for survivor functions of sequential processes in Figure 6.8
are negative at low times and then change sign.
Effects of Factors on Distribution Functions 181
Fig. 6.10. Survivor functions of simulated reaction times. Process prolongations are
short. Factors selectively influence sequential processes B1 and B2 on opposite sides of
an OR Wheatstone bridge. Parameters: Unique component of B1 has α of 48 and 96 at
Mid and High levels, respectively. Unique component of B2 has α of 48 and 96 at Mid
and High levels, respectively.
Fig. 6.11. Interaction contrasts for survivor functions in Figure 6.10 are nonnegative.
182 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 6.12. Survivor functions of simulated reaction times. Process prolongations are
short. Factors selectively influence sequential processes B1 and B2 in an AND
Wheatstone bridge. Functions at (Mid, Low) and (Mid, Mid) level combinations are
superimposed because reaction times are determined by the maximum of the
prolongations of B1 and B2. Parameters: Unique component of B1 has α of 4 and 36 at
Low and Mid levels, respectively. Unique component of B2 has α of 4 and 12 at Low
and Mid levels, respectively.
Fig. 6.13. Interaction contrast for survivor functions in Figure 6.12 is negative at small
times. It is slightly positive at large times.
Effects of Factors on Distribution Functions 183
Fig. 6.14. Survivor functions of simulated reaction times. Process prolongations are long.
Factors selectively influence sequential processes B1 and B2 in an OR Wheatstone
bridge. Parameters: Unique component of B1 has α of 4 and 96 at Low and High levels,
respectively. Unique component of B2 has α of 4 and 96 at Low and High levels,
respectively.
Fig. 6.15. Interaction contrasts for survivor functions of Figure 6.14 are nonnegative. It
is not known whether such interaction contrasts can change sign.
184 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 6.16. Survivor functions of simulated reaction times. Process prolongations are long.
Factors selectively influence sequential processes B1 and B2 on opposite sides of an
AND Wheatstone bridge. Parameters: Unique component of B1 has α of 4 and 96 at Low
and High levels, respectively. Unique component of B2 has α of 4 and 96 at Low and
High levels, respectively.
Fig. 6.17. Interaction contrasts for survivor functions of Figure 6.16 are negative at small
times and then change sign.
Effects of Factors on Distribution Functions 185
SIC(u )du 0.
t
Effects of Factors on Distribution Functions 187
SIC (u )du 0.
0
SIC(u)du 0.
0
Then
The proof is analogous when all the gates are AND gates.
FTi (t ) Fi * ( x ) g ** (t x )dx ;
0
Then
S Ti * ( t ) S Ti ( t ) [1 FTi * ( t )] [1 FTi ( t )]
t
[ Fi* ( x ) Fi** ( x )]g ** ( t x ) dx 0.
0
Process Dependence
Capacity
where SA(t) and SB(t) are the survivor functions for the individual
durations of A and B, respectively. Suppose when A and B are executed
simultaneously, they are positively dependent; that is, as the duration of
A increases the duration of B tends to increase. Then the time to
complete either A or B when executed simultaneously tends to be shorter
than the time to complete A alone and the time to complete B alone. That
is,
On the other hand, when A and B are executed simultaneously if they are
negatively dependent,
SAorB(t) < SAalone(t)SBalone(t).
Effects of Factors on Distribution Functions 193
h(t) = f(t)/S(t).
t
H (t ) h(t ' )dt '.
t 0
t
f (t ' )
H (t )
t 0
S (t ' )
dt ' ln S (t ).
SAorB(t) = SA(t)SB(t),
so
H A or B (t )
Co (t ) .
H A (t ) H B (t )
K A (t ) K B (t )
Ca (t ) ,
K A and B (t )
t
f (t ' )
K X (t ) F (t ' ) dt ' .
0
Cross-talk
saying, e.g., left for H, right for S. The words for the second response
were deliberately chosen to be names of locations for the first response.
The two stimulus features were orthogonal so information about the
correct first response is irrelevant to the correct second response.
If the participant responded to a red H with his left hand to indicate
red and said left to indicate H the two responses were compatible. But if
the participant responded to a red S with his left hand to indicate red and
said right to indicate S the two responses were incompatible. By design,
the color and the letter identity were statistically independent.
Nonetheless, both responses were faster if the responses were
compatible. An explanation in terms of cross-talk is that while the first
response is being selected, the second response is also being selected.
Irrelevant information about the likely second response is transmitted to
the process selecting the first response, increasing or decreasing the
duration of the first response selection process.
Further evidence for cross-talk was found by Logan and Schulkind
(2000), who found a role for task set in producing it. In the dual task of
their Experiment 2 stimuli for both tasks were digits. For a magnitude
judgment task, the response was to indicate whether the digit was large
or small. For a parity judgment task, the response was to indicate
whether the digit was odd or even. In one condition, both tasks in the
dual task were magnitude judgment or both were parity judgment. In the
other condition, one task was magnitude judgment and one task was
parity judgment. They found an effect of the two stimuli being from the
same category (same magnitude or same parity) when the tasks were the
same, but not when they were different. In other words, cross-talk
occurred when the set for the tasks was the same, but not when it was
different. A similar result was reported later by Lien, Schweickert and
Proctor (2003). There are now many reports of cross-talk (e.g., Dutta,
Schweickert, Choi & Proctor, 1995; Logan & Delheimer, 2001; Logan
& Gordon, 2001; Schweickert, Fortin & Sung, 2007).
Two issues are raised by the presence of cross-talk in dual tasks. The
first is whether response selection for the two tasks goes on concurrently,
rather than sequentially. Hommel (1998) proposes such concurrency,
but says there is still a limitation in processing because selection of the
Effects of Factors on Distribution Functions 197
first response must finish before the selection of the second response can
finish. This proposal can easily be represented by modifying the Single
Central Bottleneck model (Figure 5.2). An arrow (SWa) now indicates
that B1 must finish before B2 starts. The arrow is simply moved so it
indicates that B1 must finish before B2 finishes. (The resulting model is
similar to the 1973 response interdiction model of Keele). This issue, the
change in architecture from sequential to concurrent processing, presents
no problem for selective influence. The second issue is whether cross-
talk ruins selective influence, and we will see below that it can.
Coactivation
Assuming two parallel processes, the left side of the inequality states
that the probability the system’s RT is less than t is the probability one of
the process durations, TA or TB, is less than t. In other words, this term
says the RT of a parallel self-terminating model is the duration of the
target process that finishes first when there are two targets available. This
probability is clearly bounded by the right side of the inequality because
(See Colonius, 1990; Colonius & Verberg, 1994; Townsend & Wenger,
2004, and Ulrich and Miller, 1997, for other related bounds.)
Violations of this inequality reject the race model. At any time t the
amount of violation is P[TAorB t] − P[TA t] − P[TB t]; see Colonius
and Diederich (2006) for an interpretation of this quantity. Violations
suggest a coactivation system, although other systems are logically
possible (Fific, Nosofsky & Townsend, 2008). Behaviorally, the
violation means that when there are two redundant targets, RT for
detecting either target tends to be faster than the minimum of two target
processes, each executed separately in the single target condition.
visual quality. Similarly, T12 = min{A1, B2} when the second stimulus is
of low visual quality only, and so on. Suppose when there is coactivation
it occurs in such a way that T11 can be written c min{A1, B1}, where c is a
constant, 0 < c < 1. Note that for all t > 0, t/c > t. Suppose coactivation
occurs in the same way for other conditions, e.g., P[T12 t] = P[c
min{A1, B2} t] = P[min{A1, B2} t/c].
Suppose process durations are stochastically independent. Suppose
stochastic dominance is produced so for all t, SA1(t) SA2(t) and SB1(t)
SB2(t). Then the survivor interaction contrast when there is coactivation
is nonnegative for all t, as it would be without coactivation.
E[T] = k/u.
Consider two factors, each of which changes the rate of one process.
Suppose one factor with levels 1 and i leads to rates uL1 and uLi, with uL1
> uLi. Suppose the other factor with levels 1 and j leads to rates uR1 and
uRj, with uR1 > uRj. Note that as the level of a factor goes up, the
corresponding rate goes down, so the corresponding time to reach
criterion goes up. Let T11 denote the reaction time when both factors are
at level 1; other reaction times are denoted similarly. Townsend and
Effects of Factors on Distribution Functions 201
Insertions
A = uA + cA + λcB.
B = uB + cB + λcA.
Fig. 6.18. The combined effect of prolonging two concurrent processes in an AND
network can be greater than the sum of the effects of prolonging them individually when
the process durations are correlated.
(Survivor functions in Figure 6.6 are for the time from the onset of the
first stimulus to the second response, RT2 + SOA.)
Figures 6.10 through 6.17 show results when the Response Selection
Bottleneck Model is modified so Response 1 must be made before
Response 2. With the modification it becomes a double bottleneck
model in the form of a Wheatstone bridge (Figure 3.5). Figures 6.10 and
6.12 are for short prolongations of processes B1 and B2 in, respectively,
an OR and an AND Wheatstone bridge; Figures 6.14 and 6.16 are for
long prolongations. With this model Response Time 2 is
(Survivor functions illustrated are for the time from the onset of the first
stimulus to the second response, RT2 + SOA.)
When the factors change the unique component of the duration of
each process, they behave as predicted for factors selectively influencing
processes, despite high correlations between the process durations and
changes in the correlations with factor levels. Correlations are in Tables
6.2 to 6.4 and, for the Wheatstone bridge, in Tables 6.A3, 6.A4, 6.A6 and
6.A7.
In the simulations, components of the process durations have gamma
distributions. Adding a common gamma distributed random variable to
unique gamma distributed random variables is one way of producing a
multivariate gamma distribution (Kotz, Balakrishnan & Johnson, 2000).
Details of the simulations are in the Appendix.
For the simulated reaction times, means and mean interaction
contrasts are reported in the Appendix. For sequential processes X and Y,
as factor levels increase the mean interaction contrast approaches the
coupled slack, k(X,Y), despite correlations between process durations.
This is as predicted in Chapters 3 and 4 for factors selectively
influencing processes.
Table 6.2
Response Bottleneck Model
Correlations Between Process Durations in Simulations
All Factors at Lowest Levels
A1 B1 C1 A2 B2 C2
A1 1.0000 0.3925 0.2846 0.5817 -0.0099 -0.0222
B1 0.3925 1.0000 0.2774 0.5816 -0.0244 -0.0039
C1 0.2846 0.2774 1.0000 0.4254 0.5938 0.6047
A2 0.5817 0.5816 0.4254 1.0000 0.2895 0.2823
B2 -0.0099 -0.0244 0.5938 0.2895 1.0000 0.3943
C2 -0.0222 -0.0039 0.6047 0.2823 0.3943 1.0000
Table 6.3
Response Bottleneck Model
Correlations Between Process Durations in Simulations
Factors Selectively Influencing B1 and A2 at Highest Levels
A1 B1 C1 A2 B2 C2
A1 1.0000 0.1684 0.3120 0.3014 -0.0106 0.0148
B1 0.1684 1.0000 0.1227 0.1013 -0.0084 0.0103
C1 0.3120 0.1227 1.0000 0.2294 0.5903 0.5959
A2 0.3014 0.1013 0.2294 1.0000 0.1708 0.1747
B2 -0.0106 -0.0084 0.5903 0.1708 1.0000 0.3979
C2 0.0148 0.0103 0.5959 0.1747 0.3979 1.0000
Table 6.4
Response Bottleneck Model
Correlations Between Process Durations in Simulations
Factors Selectively Influencing A1 and B1 at Highest Levels
A1 B1 C1 A2 B2 C2
A1 1.0000 0.0783 0.1371 0.2583 -0.0035 0.0418
B1 0.0783 1.0000 0.1425 0.2737 0.0335 0.0174
C1 0.1371 0.1425 1.0000 0.4437 0.5882 0.5984
A2 0.2583 0.2737 0.4437 1.0000 0.3040 0.3044
B2 -0.0035 0.0335 0.5882 0.3040 1.0000 0.3990
C2 0.0418 0.0174 0.5984 0.3044 0.3990 1.0000
208 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
The function f1,i(C,S1) has the same distribution as the random vector
Effects of Factors on Distribution Functions 209
X1<i,..., m>. When the definition holds, random vector X1<i,..., m> does not
depend on the level of any factor except level i of the first factor, so we
can write X1<i,..., m> simply as X1<i>. We can do the same for each random
vector Xk. According to the definition, each random vector Xk has the
same distribution as does a function of two arguments, a source C of
randomness common to all of X1,..., Xn and a source of randomness Sk
unique to Xk. The common source of randomness allows the random
vectors to be interdependent. The unique source of randomness for
random vector Xk allows its distribution to change when the level h of
the factor selectively influencing it changes, with no change in other
distributions.
Because we mainly use the definition for factors selectively
influencing random variables, for convenience we restate it in a form
specifically for them.
The notation means that the random vector on the left-hand side has
the same distribution as the random vector on the right-hand side.
Here, the distribution of function f1,i(C, S1) is the same as that of
random variable A<i,..., m>, which we can write simply as A<i> when the
definition holds.
Now suppose with Definition 6.1, C takes on the value c, where c is a
vector of real numbers, a value for every component of C. Given c,
210 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
< X1<i,..., m>,..., Xn<i,...,m> > < f1,i(c, S1),..., fn,m(c, Sn) >.
Then by Theorem 6.2, for any value t, with t τ, for any value c of C,
Then
Concluding Remarks
Appendix
And the duration of process C1, which is concurrent with A2, B2 and C2,
was
The SOA was 100. The duration of each common component was a
gamma random variable with α = 8 and β = 4. All unique components
when factors were at their lowest levels were gamma random variables
with α = 4 and β = 4. All the gamma distributed random variables were
independent. A sample of each was taken for each trial. In particular,
independent samples were taken for the duration of each common
component.
When the sequential processes A1 and B1 were selectively
influenced, the factor selectively influencing A1 changed α of the unique
component of A1. Likewise, the factor selectively influencing B1
changed α of the unique component of B1. Values of α are in the
margins of the tables with results. When concurrent processes B1 and A2
were selectively influenced, the situation was the same, except one factor
selectively influenced A2 instead of A1.
Parameters when each factor was at its lowest level were the same as for
the Response Selection Bottleneck Model, except that α for process C1
was 28, to make C1 long enough to be important. The processes on
opposite sides of the Wheatstone bridge were B1 and B2; values of α for
them at various levels of the factors selectively influencing them are in
the tables with results. For the Wheatstone bridge with OR gates, the
SOA was 25, α for A2 was 1 and α for C1 was 16.
218 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Table 6.A1
Means of Simulated Reaction Times
When Factors Selectively Influence
Concurrent Processes B1 and A2 of the
Response Bottleneck Model
α for B1
α for A2 4 12 36 48 96
4 404 436 533 579 772
12 405 436 532 580 772
36 448 456 531 580 771
48 497 497 537 580 772
96 687 688 687 688 774
Interaction Contrasts
α for B1
α for A2 4 12 36 48 96
4
12 -1 -2 0 -2
36 -25 -46 -44 -46
48 -32 -89 -92 -94
96 -31 -129 -175 -282
Table 6.A2
Means of Simulated Reaction Times
When Factors Selectively Influence
Sequential Processes A1 and B1 of the
Response Bottleneck Model
α for B1
α for A2 4 12 36 48 96
4 404 404 450 495 688
12 405 408 480 528 720
36 448 480 575 623 818
48 497 528 625 672 864
96 687 721 815 863 1055
Interaction Contrasts
From Simulations k(A1, B1) = 84
α for B1
α for A2 4 12 36 48 96
4
12 2 29 31 30
36 32 81 83 85
48 32 82 84 83
96 33 82 84 84
Effects of Factors on Distribution Functions 219
Table 6.A3
OR Wheatstone Bridge
Correlations Between Process Durations in Simulations
All Factors at Lowest Levels
A1 B1 C1 A2 B2 C2
A1 1.0000 0.3925 0.2473 0.6119 -0.0153 -0.0022
B1 0.3925 1.0000 0.2407 0.6164 -0.0272 -0.0143
C1 0.2473 0.2407 1.0000 0.3842 0.5129 0.5245
A2 0.6119 0.6164 0.3842 1.0000 0.2953 0.3037
B2 -0.0153 -0.0272 0.5129 0.2953 1.0000 0.3986
C2 -0.0022 -0.0143 0.5245 0.3037 0.3986 1.0000
Table 6.A4
OR Wheatstone Bridge
Correlations Between Process Durations in Simulations
Factors Selectively Influencing B1 and B2 at Highest Levels
A1 B1 C1 A2 B2 C2
A1 1.0000 0.1701 0.2530 0.6044 -0.0286 0.0006
B1 0.1701 1.0000 0.1169 0.2609 0.0075 0.0119
C1 0.2530 0.1169 1.0000 0.3989 0.2084 0.5216
A2 0.6044 0.2609 0.3989 1.0000 0.1251 0.3131
B2 -0.0286 0.0075 0.2084 0.1251 1.0000 0.1721
C2 -0.0006 0.0119 0.5216 0.3131 0.1721 1.0000
Table 6.A5
Means of Simulated Reaction Times
When Factors Selectively Influence
Sequential Processes B1 and B2
On Opposite Sided of an OR Wheatstone Bridge
α for B1
α for B2 4 8 12 36 48 96
4 309 323 333 352 352 352
8 315 328 339 366 368 368
12 317 330 344 382 384 384
36 316 333 349 438 464 481
48 317 334 349 444 485 527
96 317 333 348 445 493 672
220 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
α for B1
α for B2 4 8 12 36 48 96
4 312 309 304 227 179
8 301 299 295 225 179
12 287 285 283 225 179
36 190 190 192 184 162
48 144 145 145 143 137
96
Note. Interaction contrasts calculated with highest factor levels as baseline. In the upper
left corner they approach k(B1, B2).
Table 6.A6
AND Wheatstone Bridge
Correlations Between Process Durations in Simulations
All Factors at Lowest Levels
A1 B1 C1 A2 B2 C2
A1 1.0000 0.3925 0.2169 0.5798 -0.0178 -0.0183
B1 0.3925 1.0000 0.2156 0.5870 -0.0283 -0.0129
C1 0.2169 0.2156 1.0000 0.3283 0.4645 0.4784
A2 0.5798 0.5870 0.3283 1.0000 0.2841 0.2802
B2 -0.0178 -0.0283 0.4645 0.2841 1.0000 0.4034
C2 -0.0183 -0.0129 0.4784 0.2802 0.4034 1.0000
Table 6.A7
AND Wheatstone Bridge
Correlations Between Process Durations in Simulations
Factors Selectively Influencing B1 and B2 at Highest Levels
A1 B1 C1 A2 B2 C2
A1 1.0000 0.1777 0.2510 0.5940 -0.0134 -0.0122
B1 0.1777 1.0000 0.0972 0.2456 -0.0203 -0.0078
C1 0.2510 0.0972 1.0000 0.3591 0.1765 0.4722
A2 0.5940 0.2456 0.3591 1.0000 0.1200 0.3265
B2 -0.0134 -0.0203 0.1765 0.1200 1.0000 0.1723
C2 -0.0122 -0.0078 0.4722 0.3265 0.1723 1.0000
Effects of Factors on Distribution Functions 221
Table 6.A8
Means of Simulated Reaction Times
When Factors Selectively Influence
Sequential Processes B1 and B2
On Opposite Sided of an AND Wheatstone Bridge
α for B1
α for B2
4 12 36 48 96
4 419 441 533 579 772
12 439 452 533 581 773
36 528 529 578 624 816
48 577 577 627 673 864
96 767 769 816 864 1056
Interaction Contrasts
From Simulations, k(B1, B2) = -65
α for B1
α for B2
4 12 36 48 96
4
12 -10 -21 -19 -20
36 -20 -64 -63 -65
48 -21 -64 -64 -66
96 -20 -65 -63 -65
Recall from Chapter 4 that a probability space is a triple, < Ω', S, P>,
where Ω' is a set, S is a set of subsets of Ω' that form a -algebra, and P
is a probability measure on S. A measurable space is an ordered pair
<Ω, > , where Ω is a set and is a -algebra of subsets of Ω, (e.g.,
Royden, 1968).
Now let <Ω', S, P> be a probability space and let <Ω, > be a
measurable space. A measurable function from <Ω', S, P> into < Ω, >
is a function f from Ω' into Ω such that for any S ,
{x: f(x) S} S.
222 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Note
Visual Search
223
224 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
target from the color feature map, which in turn guides the visual
attention directly to that location (Wolfe, 1994; Wolfe, Cave, and
Franzel, 1989). Thus, the number of distractors does not affect RT
much, because features are extracted in parallel without any attentional
resource.
In contrast, in a conjunction search, a combination of multiple
features is needed distinguish a target from distractors (e.g., searching for
a red letter “T” among green “T”s and red “L”s). This is because the
target shares some of its features with the distractors and cannot be
discriminated from the distractors by just one feature. Typically in a
conjunction search, the mean RT tends to increase significantly as the
number of distractors increases (i.e., a significant set-size effect). Two-
stage models explain the set size-effect in this type of search by
assuming serial engagement of attentive processes in the second stage.
Since the objects in a conjunction search are defined by multiple features
and the attentional system is assumed to work on one spatial location at a
time to integrate different features at that location, it follows that the RT
increases as the number of distractors increases. Although there is some
evidence against serial processing in some conjunction search conditions
(Nakayama and Silverman; 1986; McLeod, Driver, and Crisp; 1988;
Pashler, 1987; Wolfe et al., 1989), most recent versions of two-stage
models explain these exceptions fairly well with additional mechanisms
and assumptions (see Treisman & Sato, 1990). The following section
reviews two studies that explicitly test the two different processing stages
of two-stage models by using the method of selective influence and the
mean and cumulative distribution function interaction contrasts of RTs.
four different conditions: both stimuli bright, only the left or right
stimulus bright, and neither of them bright. According to early studies
(e.g., Pashler & Badgio, 1985; Johnsen & Briggs, 1973), manipulation of
the visual quality of a stimulus should selectively influence the feature
extraction process for the stimulus. Since feature extraction processes are
assumed to be parallel according to two-stage models, the effect of visual
quality on overall RT would follow patterns expected from parallel
processing, if the two-stage structure is correct.
Thus, Egeth and Dagenbach looked for patterns of mean RTs
suggested by Schweickert (1978), Schweickert and Townsend (1989),
and Townsend and Schweickert (1989) in an extension of the additive
factor method (Sternberg, 1969). (Chapter 3 describes these patterns.)
Specifically, the time to process two stimuli in parallel should be the
maximum of the times to process each individually. Making the
simplifying assumption that times are constants, not random variables,
suppose the RT for processing a single high visual quality stimulus is t.
Then the RT for processing two high visual quality stimuli is max{t, t} =
t if the stimuli are processed in parallel, but t + t = 2t if processed
serially. When the visual quality of one of the two stimuli deteriorates,
it takes more time to extract the individual features, say ∆t. This is in the
pre-attentive stage; according to two-stage models the second stage is not
affected. The overall RT becomes max{t, t + ∆t} = t + ∆t if processing
is parallel and t + (t + ∆t) = 2t + ∆t if processing is serial. At this point,
RTs do not provide enough information to distinguish serial and parallel
processing, because the two processing systems give us the same change
in overall RT (∆t). The serial and parallel models generate different RT
patterns when we include the condition in which both stimuli are of low
visual quality. The overall RT becomes (t + ∆t) + (t + ∆t) for the serial
processing model and t + ∆t for the parallel processing model.
When two stimuli are of low visual quality, only serial processing
results in a different RT from when only one of the stimuli is of low
quality. Note that this diagnostic used in Egeth and Dagenbach (1991) is
valid only when the processing of both stimuli is guaranteed. If one of
the stimuli is a target and the search is serial, subjects may not process
the second stimulus (distractor) when the first stimulus processed has
Visual/Memory Search, Time Reproduction, Classification, and Face Perception 227
Table 7.1
Mean Response Times (ms) of Searching for
Upright Ts among Ls, or Ls among Ts, in the Target Absent Condition
(Modified from Table 5 in Egeth & Dagenbach, 1991)
One of the major targets of criticism for two-stage models is the serial
attentive processing assumption in the second stage. FIT explicitly
assumes that when attention is engaged, only one spatial location can be
attentively processed at a time to identify a target. As discussed before,
the set-size effect found in conjunction search tasks was initially thought
to be critical evidence for the serial processing assumption. Although it
is easy to draw this conclusion when one sees linearly increasing RTs as
the number of objects increases, it is now well known that linearity of RT
as a function of set size does not imply serial processing of stimuli
(Townsend, 1971, 1972; Townsend & Ashby, 1983). Linearity can be
explained by special parallel processing models, such as limited-capacity
parallel processing. These models assume that mental processes require a
limited capacity resource, such as attention. When multiple parallel
processes are executed, they share the resource. The consequence is that
as the number of parallel processes increases, the amount of resource
allocated to each process decreases. The set-size-effect is explained by
assuming that process durations are inversely related to the amount of
resource allocated to each process. For example, given a fixed
Visual/Memory Search, Time Reproduction, Classification, and Face Perception 229
Fig. 7.1. Two possible networks for the second stage of two-stage models (e.g., Treisman
& Gelade, 1980). A. Four processes are connected in series (e.g., FIT). B. Four processes
are connected in parallel. Here, p and r stand for starting and ending of the networks,
respectively.
Visual/Memory Search, Time Reproduction, Classification, and Face Perception 231
Fig. 7.2. Four different stimulus presentation conditions of Experiment 1 in Sung (2008).
Only the target-absent conditions are shown here. In these examples, the color of Ts are
green and the rest of the stimuli are red. The target is a red T, which is not shown in these
examples. Adapted from Sung, K. (2008). Serial and parallel attentive visual searches:
Evidence from cumulative distribution functions of response times. Journal of
Experimental Psychology: Human Perception and Performance, 34, Fig. 2. Copyright
2008 by American Psychological Association. Adapted with permission.
Table 7.2
Mean Response Times (ms) of Four Target-Absent Conditions
(Modified from Table 1 of Sung, 2008)
Target Absent
Mean RTs
Conditions (Figure 7.2)
A 553
B 630
C 624
D 687
Mean IC : 553-630-624+687 = − 19
The patterns of CDF IC further bolster the conclusion that the stimuli
of this particular conjunction search task can be processed in parallel,
rejecting the serial processing assumption of FIT. As explained in
Chapter 6, the CDF IC remains positive for all times t if two processes
selectively influenced by two factors are indeed connected in parallel in a
critical-path network (e.g., Figure 7.1B). Sung observed this pattern in
Visual/Memory Search, Time Reproduction, Classification, and Face Perception 233
his first experiment (i.e., finding a red T among green Ts, red Os, and
two special distracters). As shown in the bottom panel of Figure 7.3, the
CDF IC tended to remain positive for all times t. Also, the stochastic
dominance assumption for the test is clearly satisfied in all four cases
(top and middle panels in Figure 7.3).
Fig. 7.3. Top panel: CDFs from four target-absent conditions of Experiment 1 in Sung
(2008). Middle panel: stochastic dominance tests for four CDFS. Bottom panel: CDF IC
function calculated. A, B, C, and D represent corresponding target-absent conditions in
Figure 7.2. Adapted from Sung, K. (2008). Serial and parallel attentive visual searches:
Evidence from cumulative distribution functions of response times. Journal of
Experimental Psychology: Human Perception and Performance, 34, Fig. 4. Copyright
2008 by American Psychological Association. Adapted with permission.
234 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Memory Scanning
Fig. 7.4. Survivor ICs of five subjects performing memory search in Townsend and Fific
(2004). Thin solid lines in each plot represent the simple serial or parallel model that fit
best the observations represented by dotted dark lines. From Townsend, J. T., & Fific,
M. (2004). Parallel versus serial processing and individual differences in high-speed
search in human memory. Perception & Psychophysics, 66, Fig. 3(A). Copyright 2004
Springer. Reproduced with permission.
238 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
absence (“3”) of the target in the display and the end of time
reproduction. The response occurs at r, the end of the network.
The visual search stimuli were presented after the time reproduction
started to ensure that the visual search terminated at approximately the
same time as the end of the time reproduction. If the visual search
process goes on concurrently with the time reproduction but always
terminates before the end of time reproduction, the resulting RT
distributions would not reflect the effects of the experimental factors
selectively influencing the visual search process. This SOA before visual
search was the only substantial difference between the experiments in
Schweickert et al. (2007) and in Fortin et al. (1993). In Fortin et al.
(1993), the visual search started at the same time as the time
reproduction.
There are six different combinations of experimental factors available
for the CDF IC test in Schweickert, et al. (2007). Factors behaved as if
they selectively influenced the processes they were expected to in the
model of Figure 7.4; for example, manipulations of time interval and
SOA behaved as predicted for manipulations of concurrent processes.
There was one exception; one interaction contrast of the three tested for
display size and SOA was inexplicably not as predicted, details are in
Schweickert, et al. (2007). For the purpose of this chapter, we focus on
two important factors, display size (three levels) and the time interval to
be reproduced (three levels), to see if the factors selectively influence
Visual/Memory Search, Time Reproduction, Classification, and Face Perception 241
Table 7.3
Mean Response Times (ms) for 9 Different Conditions
of Display Size and Time Interval Factors
(From Table 1 of Schweickert, Fortin and Sung, 2007)
Mean ICs
IC1 = (5) – (4) – (3) + (1) = -46
IC2 = (6) – (4) – (3) + (1) = -50
IC3 = (8) – (7) – (2) + (1) = -53
IC4 = (9) – (7) – (3) + (1) = -84
IC5 = (6) – (5) – (3) + (2) = - 5
IC6 = (9) – (8) – (3) + (2) = -31
IC7 = (8) – (7) – (5) + (4) = - 7
IC8 = (9) – (7) – (6) + (4) = -33
IC9 = (9) – (8) – (6) + (5) = -26
Fig. 7.6. CDF ICs for time interval and display size factors. Adapted from Schweickert,
R., Fortin, C., & Sung, K., 2007, Concurrent visual search and time reproduction with
cross-talk. Journal of Mathematical Psychology, 51, Fig. 10. Copyright 2007 Elsevier.
Adapted with permission.
that the visual search did not start right away in Schweickert et al.
(2007), whereas it did in Fortin et al. (1993). But the mechanism that
would generate different RT patterns in these two experiments is not
known.
Perceptual Classification
Fig. 7.7. Two category structure and a sample stimulus, showing its two parts. To be
classified as category A, the color patch must have saturation value of s2 or s3 AND
middle line position of v2 or v3. H and L denote High (easy) and Low (hard) values.
Adapted from Fific, M., Nosofsky, R. M., & Townsend, J. T., 2008, Information-
processing architectures in multidimensional classification: A validation test of the
systems factorial technology. Journal of Experimental Psychology: Human Perception
and Performance, 34, Fig. 5. Copyright 2008 American Psychological Association.
Adapted with permission.
There were nine possible combinations of the two features values, but
Fific et al. examined only four conditions (i.e., s2v2, s2v3, s3v2, s3v3
combinations) to investigate the interaction contrast of the survivor
function of RTs. The reason for investigating these four conditions,
246 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 7.8. Mean RTs (left column), survivor functions (middle column), and SICs (right
column) of 7 subjects (rows) in Experiment 1. Adapted from Fific, M., Nosofsky, R. M.,
& Townsend, J. T., 2008, Information-processing architectures in multidimensional
classification: A validation test of the systems factorial technology. Journal of
Experimental Psychology: Human Perception and Performance, 34, Fig. 7. Copyright
2008 American Psychological Association. Adapted with permission.
248 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 7.9. Mean RTs (left column), survivor functions (middle column), and SICs (right
column) of 5 subjects (rows) in Experiment 2. The first 5 subjects of 8 are shown here.
Adapted from Fific, M., Nosofsky, R. M., & Townsend, J. T., 2008, Information-
processing architectures in multidimensional classification: A validation test of the
systems factorial technology. Journal of Experimental Psychology: Human Perception
and Performance, 34, Fig. 8. Copyright 2008 American Psychological Association.
Adapted with permission.
Visual/Memory Search, Time Reproduction, Classification, and Face Perception 251
processing times of the features and not the decision making process that
sums the outputs of these two processes, it means that it takes more time
to detect (not to classify) a stimulus with low saturation and brightness
values than one with high saturation and brightness values, which could
be the case as we saw in Egeth and Dagenbach’s (1991) visual search
experiments. However, it is also possible that the classification of (or
decision making about) the stimulus would also take longer if the
identified features are close to the classification boundary (e.g., [b2, s2]
of experiment 2) than if they are far from the boundary (e.g., [b3, s3])
(e.g., Ashby, Boynton, & Lee, 1994). This means that brightness and
saturation affect not only the different processes that occur before the
summation process of the coactive system, but also the decision making
process that occurs after the summation of the two processes. If decision
making is also changed, why is the pattern for a coactivation model
observed? A possible answer is discussed in Schweickert, Fortin, and
Sung (2007). When a factor influences two processes, not one, and its
effect is much greater in one process than in the other, the experimental
factor may still behave as if it selectively influences the first process
only. In terms of classification, this means that changing brightness and
saturation have major effects on separate feature processes, as well as
some minor effect on the decision making process. Although
Schweickert et al.’s (2004) observation is based on simulation, it is a
reasonable possibility. The issue is conceptual; better understanding of
coactive systems is needed, in particular of what it would mean for a
factor to selectively influence a process with coactivation present.
In summary, data of Fific et al. (2008) indicate that processing of two
separable feature dimensions is serial for most of their subjects, but
parallel for some; on the other hand processing of integral dimensions is
coactive. A study of processing of separable dimensions in the same
object would be informative (e.g., Fific, Little, & Nosofsky, 2010). A
clear specification of the models being tested may help us better
understand results of Experiment 2.
252 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Face Perception
Fig. 7.10. Four stimulus examples used in Ingvalson & Wenger (2005). Top-left: human
face without configural changes (or with centered eyes and mouth). Top-right: human
face with unambiguous configural changes from the top-left. Bottom-left: schematic face
stimulus without any configural changes. Bottom-right: schematic face stimulus with
unambiguous configural changes from the bottom-left. From Ingvalson, E. M., &
Wenger, M. J.,2005. A strong test of the dual-mode hypothesis. Perception &
Psychophysics, 67, Fig. 2. Copyright 2005 Springer. Reproduced with permission.
for other conditions such as inverted faces, and schematic faces. The
mean interaction contrasts were all positive for the face and face-like
stimuli, as predicted if the processes for featural and configural
information were concurrent and self-terminating (i.e., in an OR
network, see chapter 4). The survivor function interaction contrasts were
positive for all times t in all four subjects tested. This also is as predicted
if the processes were concurrent, and the response was made whenever
any of the concurrent processes detected a change (featural or configural)
between two face stimuli (i.e., the stopping rule of the parallel system is
self-terminating). Results support the dual-mode hypothesis, according
to which the processes are concurrent and self-terminating. With further
analyses, Ingvalson and Wenger (2005) also find support for unlimited
or super capacity, as predicted, but not for independence of rates of
featural and configural processes, contrary to prediction.
The results reported in Ingvalson and Wenger (2005) are very robust
in that the same parallel self-terminating system seems to work
Visual/Memory Search, Time Reproduction, Classification, and Face Perception 255
commonly for the inverted and the schematic face stimuli in all four
subjects tested. One thing to point out, however, is that the evidence
from the survivor function interaction contrast only supports concurrent
processing of featural and configural information but not the concurrent
processing of two features (colors of lips and irises) or two
configurations. That is, assuming there are four processes for each face,
two featural and two configural, it is possible that the two processes for
color of lips and color of irises (i.e., two features) can be sequential
(since they are spatially separated in a face) but concurrent with the two
processes for configural information. Alternatively, all four processes
could be concurrent. Since two features (colors of lips and eyes) always
co-varied in Ingvalson and Wenger’s (2005) experiment (as well as two
configuration manipulations), their results cannot help us differentiate
these two possible alternatives.
Concluding Remarks
AND Networks
256
Modeling with Order of Processing Diagrams 257
network. We begin with AND networks because they are the most
common task networks. We show how to use OP diagrams to compute
the moments of response time for them. We then show modifications
required for directed acyclic task networks with OR gates.
Order-of-Processing Diagrams
Assume that a process X with starting event X ' and ending event X" is
associated with each arc in a directed acyclic task network. A process
may be active, such as moving a limb. It may be more passive, such as a
communication delay. A process may indicate a precedence constraint.
The most common precedence constraint is that the ending of one
process must precede the start of another. Some other constraints can
easily be represented by arcs, for example, that the start of one process
must precede the start of another. Such precedence constraint arcs can
be considered processes of duration zero. We gain generality by
associating processes with these arcs as well as the arcs which represent
more standard processes (e.g., encoding, comparison and execution). If
an arc connects the start of process X to the start of process Y, then the
process corresponding to the arc may be identified as X ' Y ' ; other
processes are denoted similarly. Let Z be the set of all processes.
Assume that there are |Z| such processes.
States
Transitions
Fig. 8.1. Two processes, X and Y, in series represented in (a) a task network, and (b) an
OP diagram. The start and end vertices of processes X and Y in the task network are
labeled (Xʹ, Xʹʹ) and (Yʹ, Y′ʹ), respectively.
all processes which complete along any one path to that state. Thus, for
example, the completed set of state s3 will contain processes X and X"Y ' ,
i.e., K(s3) = {X, X"Y ' }.
f ( x k ) qk e ( q k x k ) , 0 x k
(8.1)
f ( x k ) 0, otherwise.
Set qij equal to the rate parameter of the process which completes
when a transition is made from state si to state sj. And set qii equal to the
sum of the rate parameters of the processes active in the current set C(si)
of state si. (Note that i and j are used to index states here; previously
they were used to index the levels of different factors.)
Then it can be shown that the mth response time moment E[T(1)m] can
be computed recursively by first computing in order E[T(|S|)], E[T(|S|
1)],..., E[T(1)], then computing in order E[T(|S|)2], E[T(|S|-1)2],...,
E[T(1)2], and so on, where E[T(i)m] is defined as follows (e.g., see
Howard, 1971, Equation 11.11.13, page 735; Kulkarni & Adlakha, 1986,
Equation 1, page 774):
m
E[T (i ) ]
mE[T (i )m1 ] jSi
qij [T ( j )m ]
, 1 i | S |,
qii (8.2)
0, i | S |
For example, suppose that there are only two processes in series, X
and Y, with respective rate parameters α and . Then, there are three
states in the OP diagram (assuming the transmission process, X"Y ' , is a
dummy process): X is current in state s1, Y is current in state s2, and no
processes are current in state s3. The first moment is easily computed:
E[T ( 3)] 0;
1 q23 E[T ( 3)] 1 0 1
E[T ( 2)] ;
q22
1
1 q12 E[T ( 2 )] 1 1
E[T (1)] .
q11
E[T (3) 2 ] 0;
2
02
2 E[T ( 2)] q23 E[T ( 3) ] 2
E[T ( 2) 2 ] 2;
q22
1 1 2
2 2
2 E[T (1)] q12 E[T ( 2) ]
2
E[T (1) 2 ]
q11
2 2 2
.
2
2
We can use the above methods when the duration of any one or more
processes is a general gamma random variable by representing each
262 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
|H |
E[ RT m ] E[ RT m | P h]Prob( P h )
h 1
This equation makes intuitively clear the general procedure one needs
to follow in order to obtain the expected completion time. This procedure
is described in detail in Goldstein and Fisher (1991). Here, we want only
to indicate the required computations when the durations are
independent.
To begin, let D represent the vector of process durations <X1,..., X|X|>.
Define fX(x1,..., x|x|) as the joint density function of the durations of the
processes in the task network. We assume that this density function is
known or can be estimated. Define Xhj as the process which completes
first in state shj, the jth state along path h. Define Thj as the duration of the
jth state along path h and set it equal to the duration of the first process to
complete in the jth state, i.e., Xhj, minus the duration of each of the
preceding states in which this process was current:
Th1 X h1
j 1 (8.3)
Thj X h1 T
i 1
ij hi ,
|H |
E[ RT ] g h f X ( x1 ,..., x| X | )dx1 ,..., dx|X | ,
h1 ( x1 ,..., x| X | )Rh
(8.4)
where gh = th1 + th2 + ... + thh', where h' is the index of the last state on
path h prior to the finish state, and where Rh is the region where each of
264 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
the state durations on path h is positive, i.e., Thj > 0, j = 1,..., h'. In order
to perform the integration, we need to rewrite the quantities thj and the
region Rh in terms of the Xi. First, the quantities thj can be rewritten in
terms of the Xi using Equation (8.3). For example, without loss of
generality, assume that the processes are numbered in such a way that
process Xi completes in the ith state along path h. Then, th1 = x1, th2 = x2 −
a12th1 = x2 − x1 (assuming that the second process is in the current set of
the first state), and so on. Second, note that given Equation (8.3), the
region Rh can be written as a system of inequalities where the duration of
each process must be greater than the duration of a linear function of the
durations other processes in X. For example, Th1 > 0 implies xh1 > 0; Th2
> 0 implies xh2 − a12Th1 = xh2 − a12xh1 > 0; and so on.
It may be instructive at this point to work through an example.
Consider a double stimulation task where some constant time (the
interstimulus interval) intervenes between the presentation of the first
and second stimulus. A very simple representation of processing in this
case as a directed acyclic task network is displayed in Figure 8.2a. The
corresponding OP diagram is in Figure 8.2b. In this figure, X1 represents
the processing of the first stimulus, X2 represents the processing of the
second stimulus, and X3 represents the “processing” or communication
delay of the interstimulus interval. The processing of the second
stimulus cannot be initiated until both the interstimulus interval has
elapsed (a logical requirement) and the first stimulus has completed
processing (the hypothesis). Assume that the durations X1 and X2 of,
respectively, processes X1 and X2 are independent uniform [300, 400]
random variables and the duration X3 of process X3 is a constant equal to
c, where 300 < c < 400.
Then, given the above assumptions we can write the joint density as
follows: fX(X1, X2, X3) = 1/10000 for 300 < X1, X2 < 400 and X3 = c; fX(X1,
X2, X3) = 0 otherwise. Let the top path in the OP diagram (Figure 8.2b)
be path 1. The states on this path all have a positive duration when X3 <
X1 < 400 and 300 < X2 < 400 (these assumptions define R1). Similarly,
consider the bottom path, path 2. The states on this path all have a
positive duration when X1 < X3 < 400 and 300 < X2 < 400 (these
assumptions define R2). From Equation (8.4) we compute the expected
Modeling with Order of Processing Diagrams 265
400 400
E[ RT ] (t
x2 300 x1 c
11 t12 t13 ) f x ( x1 , x 2 , x 3 )dx1dx 2
400 c
(t
x2 300 x1 300
21 t 22 t 23 ) f x ( x1 , x 2 , x 3 )dx1dx 2 .
Note that for path 1 we have from Equation (8.3), t11 + t12 + t13 = x3 +
(x1 − x3) + x2 = x1 + x2, and for path 2, t21 + t22 + t23 = x1 + (x3 − x1) + x2 =
c + x2. For example, substituting into the above equations, setting c =
300 and integrating, we find E[RT] = 700.
Fig. 8.2. Two concurrent processes, X1 and X3, represented in (a) a task network and (b)
an OP diagram.
OR Networks
Fig. 8.3. (a) Longest path task network with two concurrent processes, X1 and X2. (b)
The associated OP diagram. (c) A shortest path task network (the OR gate is indicated by
an open circle).
the shortest path task network in Figure 8.3c, neither T12 nor T13 appear,
respectively, as part of g1 or g2. Thus, we want to evaluate the sum of
multiple integrals below:
x1
E[ RT ] (t
x3 0 x1 0 x2 0
11 t14 ) f x ( x1 , x 2 ,x3 )dx 2 dx1dx3
x2
(t
x3 0 x2 0 x1 0
21 t 24 ) f x ( x1 , x 2 ,x 3 )dx1dx 2 dx 3 ,
where for h = 1, t11 = x2 and t14 = x3, and for h = 2, t21 = x1 and t24 = x3.
The above computations are quite simple when the durations of the
processes are independent exponentials (or sums of independent
exponentials). Details are in Fisher and Goldstein (1983). Computations
268 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Smith (1969)
1 1
E[ RT 1( j )] .
qa ( j ) qb
E[ RT 2( I , j )] I 1 e Iqc
qc
e e
Iqa ( j ) Iqc
qc q a ( j )
qc Iqa ( j ) 1 Iqa ( j ) Iqc
qa ( j )
e e Iq 1 e Iqc
c
qc q a ( j ) qc
e Iqc 1
.
q
a ( j ) qd
A grid search was used to find the best fitting parameters. (For no j
was denominator qc − qa(j) = 0, j = 2, 4, 8.) As a check, the partial
derivatives with respect to each parameter were produced with
MACSYMA, and parameter estimates minimizing the sum of the
squared error were found with the nonlinear regression program SAS
NLIN (SAS Institute, 1985). The parameter estimates from the grid
search were used as starting values in the nonlinear regression program.
The following reciprocals of the parameter estimates gave the best fit to
the observations in Table 5.2: 1/qa(2) = 25, 1/qa(4) = 125, 1/qa (8) = 200,
1/qb = 443, 1/qc = 278, 1/qd = 354. These are virtually identical to the
parameter estimates found by the grid search. With these estimates, the
model accounts for 99.4 percent of the variance.
Predicted values are quite close to observed values for RT2, see Table
8.1. For example, the differences between observed and predicted values
270 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 8.4. Processing in the double stimulation task run by Smith (1969) represented as: (a)
a PERT network and (b) an OP diagram. (Note that the response to the first stimulus is
emitted when process b completes and that the response to the second stimulus is emitted
when process d completes.)
Table 8.1
Observed (obs.) and Predicted (pred.) Reaction Times:
Smith (1969)
Stimulus 1 Alternatives
RT1
2 4 8
ISI obs. pred. obs. pred. obs. pred.
50 480 468 594 568 666 643
150 459 468 582 568 639 643
300 444 468 561 568 635 643
500 459 468 571 568 628 643
Stimulus 1 Alternatives
RT2+ISI
2 4 8
ISI obs. pred. obs. pred. obs. pred.
50 665 659 766 757 835 832
150 680 682 769 768 836 839
300 750 758 809 816 889 876
500 913 904 929 936 963 976
Three points of interest remain. First, we show how one can sometimes
represent two alternative networks as directed acyclic task networks,
namely, connectionist and queueing networks. Second, we show how to
incorporate resource constraints into response time modeling. Finally, we
talk briefly about cognitive behavior that cannot be represented as a
directed acyclic task network.
Connectionist networks
Table 8.2
Observed (obs.) and Predicted (pred.) Reaction Times:
Pashler and Johnston (1989)
McClelland, 1986), visual search (Phaf, Van der Heijden and Hudson,
1990) and categorization (Busemeyer & Myung, 1988; Gluck & Bower,
1988). Two criticisms of the use of such networks are considered here.
First, the network structure is usually presupposed, and it is not always
clear how to test that two processes in a network were arranged, say, in
parallel rather than in series. Second, only a few studies have used
connectionist networks to model reaction times in detail (e.g., Cohen,
Dunbar, & McClelland, 1990; Liu, Holmes & Cohen, 2008; Ratcliff, Van
Zandt & McKoon, 1999) and it is often not clear how to predict the time
it takes to perform a task represented as a connectionist network. We
address these two problems for a subset of connectionist networks, one
Modeling with Order of Processing Diagrams 275
and two layer networks without feedback of the activation, and where the
threshold gates are logically equivalent to AND or OR gates. Although
learning in connectionist networks is one reason they have received so
much attention, we are interested here in the network only after learning
has occurred because our concern is with the final architecture, not the
architecture as it evolves.
A simple one layer net without feedback is constructed as follows.
The net consists of a set of input and output nodes. (The set of input
nodes customarily does not count toward the number of layers.) Each
input node ni (i = 1,..., I) is connected to all output nodes nj (j = 1,..., J).
Associated with the connection or arc between input node ni and output
node nj is a weight wij. Activation is initiated at the input nodes by one
of the K input patterns pk (k = 1,..., K). An input pattern is a vector of I
components, the ith component being the input to node ni. The activation
aj at an output node nj is a weighted sum of the inputs a1,..., aI to nj:
I
aj w a
i 1
ij i
For example, a one layer net with three input nodes (n1, n2, n3) and
three output nodes (n4, n5, n6) is represented in Figure 8.5a. Each input
node is connected to every output node. If w14 = w24 = w34 = 1 and if a1 =
0, a2 = 1, and a3 = 0, then using the above equation the activation at node
n4 is a4 = 1. Assume that a response is made by output node nj when the
activation at the node exceeds some threshold dj.
Initially, the weights on the arcs are set randomly. The input patterns
are divided into subsets, a separate response being required for each
different subset. Assume for the sake of simplicity that there are J
different subsets of patterns, J output nodes and I = J input nodes. If the
input patterns satisfy certain assumptions, then the network can be
trained to make an appropriate response for each pattern in each subset.
Suppose that the network has been so trained. Then, if learning is
successful, for each input pattern pk in the set of trained patterns the
activation at exactly one of the output nodes, say node nj, will be above
276 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 8.5. (a) Representation of a one layer connectionist net with three input and three
AND output nodes; (b) the corresponding functional net when the output is n4; c) the task
network representation of the functional net; d) the task network representation of the
functional net if n4 were an OR node.
threshold; the activation at all other nodes will be below threshold. Thus,
the time to complete processing in the neural net will depend only on the
time it takes to increase the activation at node nj above threshold.
The threshold gates used in connectionist models are sometimes
logically equivalent to AND gates (Williams, 1986). For example,
consider the one layer net in Figure 8.5a. Assume that each activation ai
(i = 1, 2, 3) is either a 0 or a 1. And assume that node n4 has a threshold
of 3. Then n4 is equivalent to an AND gate. If an input pattern is
presented which increases above threshold just the activation at node n4,
then only the connections to node n4 are functional for this input pattern.
The resulting functional network is much simpler than the original one
and is formed from the connections in Figure 8.5b. Given the above
assumptions, this functional network can be interpreted as a simple
parallel AND network. To see this, let xij represent the process associated
Modeling with Order of Processing Diagrams 277
with the transmission of activation from input node ni to output node nj.
Then, node n4 will not respond until all of the processes x14, x24 and x34
have completed. As an AND network, this is graphed as in Figure 8.5c.
The distribution of the response time in this case requires computing the
maximum of the time to complete processes x14, x24 and x34. It is clear
that the results generalize to single layer nets with any number of input
and output nodes.
One or more of the nodes in the network under consideration could
just as easily have been an inclusive OR node. Such a node fires when
any one or more of its inputs are active. For example, if node n4 were an
inclusive OR node it would fire as soon as it received activation from
node n1, n2 or n3. This relation between an output node and its input
nodes can also be diagrammed as a directed acyclic network (Figure
8.5d). The duration of this process is simply the minimum of the time to
complete processes x14, x24 and x34. For more discussion of activation
functions realizing the Boolean AND and OR functions, see Williams
(1986).
A major problem with one layer nets such as the perceptron is that
they cannot be used to represent an exclusive OR gate. Such gates are
frequently required in order successfully to categorize a set of patterns.
However, a two layer net, unlike a one layer net, can successfully
categorize patterns which require an exclusive OR gate. Briefly, a two
layer net consists of a set of input nodes, hidden nodes (the middle layer)
and output nodes. An example is given Figure 8.6a. Nodes n1, n2, and n3
are the input nodes; nodes n4 and n5 are the hidden nodes; and nodes n6
and n7 are the output nodes. Each of the first two input nodes (n1 and n2)
has a connection to each of the two hidden nodes; each hidden node has a
connection to the two output nodes. In addition, the third input node (n3)
connects directly to the last output node (n7).
The activation ah at a hidden node nh (h = 1,..., H) is computed like
the activation at an output node in a single layer net:
I
ah w
i 1
ih ai
278 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 8.6. (a) Representation of a two layer exclusive-or connectionist network which
responds yes (n6) when exactly one of nodes n1 or n2 is activated and responds no (n7)
otherwise; (b) the corresponding functional network when exactly one feature is present;
(c) the PERT network representation of the functional network.
Queueing networks
Fig. 8.7. (a) A queueing network with two servers, s1 and s2, and room for three
customers in the buffer for server s1 and two customers in the buffer for server s2
(customer c2 is currently in the buffer at node n1 and customer c1 is on server s1); (b) the
PERT network representation of the processing of two customers through the queueing
network; (c) the corresponding OP diagram.
incorporated into the task network. In the second, less frequent form,
there is a limit on the number of concurrent processes. For example,
Fisher (1982, 1984) suggested that at most four comparison processes
can execute simultaneously during the standard consistent mapping
visual search task (i.e., letter targets in digit distractors); for recent
discussion see Cowan (2005). Task networks with just AND gates or
just OR gates cannot capture this type of limitation (though OP diagrams
can; see following section).
The most systematic and sustained work on capacity started with
work by Townsend (1971, 1974) addressing capacity issues in
distinguishing parallel and serial processing. The key component is a
measure of capacity (Townsend & Ashby, 1983; Townsend & Nozawa,
1995) that can be used in conjunction with techniques such as selectively
influencing processes to learn whether and how processes share capacity
when working together (Townsend & Wenger, 2004; Wenger &
Townsend, 2000).
We focus here on the situation in which one has learned, through such
an analysis or in some other way, that there is a limit on the capacity
allocated to a subset of processes. Specifically, we want to show how
one can incorporate the limitation into the computation of the moments
of the response time. Full details are available in Goldstein and Fisher
(1992). Here, we assume the work Wx required to complete the
execution of a given process x varies from trial to trial. We assume that
the rate v(x, shi) at which work proceeds on a process x in the current set
of state shi is determined not only by the identity of the process and the
state, but also by the history of the system (i.e., the path) up until the
current state. We assume that no work is accomplished if a process is in
some set other than the current set so that v(x, shi) = 0 if x is not in C(shi).
Finally, we assume that the following relation obtains between the rate at
which work is performed in each of the states, the duration of the states,
and the total work on a given trial where some path, say h, is followed:
h
Wx v ( x , shi )Thi (8.5)
i 1
284 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
where h' is the index of the last state on path h prior to the finish state.
In order to compute the moments of the response time we assume the
joint density of the work requirements fw(w1,...,w|X|) is known or
estimated (this is similar to the assumption above that the joint density of
the process durations was known). The required computations follow
from Equation (8.6) in the form below:
|H |
E[ RT ] g h f w( w1 ,..., w| X | )dw1 ,..., dw| X | , (8.6)
h 1 ( w1 ,..., w| X | )Rh
100 50 2 w1
w2 w1 w3
E[ RT ]
w3 0 w1 0 w2 0
4
2
3
f w ( w2 , w1 , w3 )dw2 dw1dw3
100 100 100
w2 w1 w3
w3 0 w1 50 w2 0
4
2
f w ( w2 , w1 , w3 )dw2 dw1dw3
3
w2
100 100 2
w1 w2 w3
w3 0 w2 0 w1 0
2 4
f w ( w1 , w2 , w3 )dw1dw2 dw3 .
3
Completing the above, one obtains as the expected response time 52.17.
We have shown how one can represent AND networks, OR networks and
certain connectionist and queueing networks as directed acyclic task
networks, and then as OP diagrams. For each OP diagram, it was the
case that: a) one path was followed through the network on each trial; b)
the response time on a given trial was equal to the sum of the times spent
in each of the states along the relevant path; and c) only the joint density
function of the process durations or work requirements (not the state
durations) was known. Given this information, we then showed that one
could compute the moments of the response time. But to compute the
moments of the response time, one need not begin with a directed acyclic
task network. What general characteristics of a representation make it
possible to carry out the above program?
Goldstein and Fisher (1991) showed that the set of networks which
are OP representable differs from the set of directed acyclic task
networks represented previously in OP diagrams in several important
ways: a) there can be several start states; b) there can be several finish
states; c) there can be several processes which complete simultaneously
when a transition is made from one state to another (simultaneous
completion is allowed in a task network, but was not allowed previously
in the OP representation of it); d) there can be processes which are
interrupted during their execution; e) the transition out of a state can be
286 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Appendix
Fitting the Smith (1969) and Pashler & Johnston (1989) data
Smith
Recall that for the data of Pashler and Johnston (1989), there are seven
processes in the reduced model: A1B1, C1, SWa, A2(low), A2(high),
B2C2 and I. The duration of the interstimulus interval is known. Each
of the first six processes is assumed to have a gamma distribution. A
gamma distribution is described by two parameters, one for shape (which
we shall designate by α) and one for scale (). Then there are 12
parameters that we need to estimate.
Parameters
To begin, consider the processes associated with the response to the first
stimulus in the double stimulation task. There are two processes, A1B1
288 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
and C1, and thus four potential parameters. In order to reduce the
number of parameters, we assumed (arbitrarily) that the same shape
parameter, α, controlled the distribution of the durations of processes
A1B1 and C1. Thus, the duration A1B1 of process A1B1 has a gamma
distribution with parameters say, α11 and 11, and the duration C1 of
process C1 has a gamma distribution with parameters α12 and 12 where
α11 = α12. Now, note that if we require the mean and variance of our
prediction of the time RT1 to respond to the first stimulus in the double
stimulation task to equal respectively the mean and variance of the
observations, only one of the four parameters α11, α12, 11 and 12 remains
free. Specifically, recall that the predicted mean and variance are
obtained as follows:
If we set the above predictions equal to the sample mean, say M1, and
sample variance, say V1, we obtain:
M 1 (11 )( 11 ) (12 )( 12 )
V1 (11 )( 11 ) 2 (12 )( 12 ) 2
Once we set 11 to some value, say b11, then we can solve for 12 and α11
= α12. Specifically, we obtain:
and
M1
11 12 .
b11 12
Modeling with Order of Processing Diagrams 289
Simulation
The responses for the high intensity processes were expressed similarly.
A total of 1000 trials were used in the grid search through the
parameter space to evaluate the goodness of the fit of each combination
of parameters. Those values for the parameters were selected which
minimized the sum of the square of the difference between the 36
predictions and observations, i.e., the 18 means and 18 standard
deviations (two levels of intensity, three interstimulus intervals, three
different responses). The parameter values were: α11 = α12 = 9.5, 11 =
34.0, 12 = 28.3, α21(high) = α22(high) = 45.8, 21(high) = 1.0, 22(high) =
9.9, α21(low) = 8.4, α22(low) = 45.5, 21(low) = 11.9, 22(low) = 9.9, α3 =
1.0, 3 = 134.0.
Chapter 9
291
292 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Multiplicative Effects
The simplest way two factors can combine is through addition, next
simplest is through multiplication, to which we now turn.
Accuracy
Suppose the processes for performing a task are in series, output of one
being input to the next. The serial arrangement allows analysis of both
reaction time and accuracy. The reaction time will be the sum of the
durations of the individual processes, so two factors selectively
influencing two different processes will have additive effects on reaction
time (Sternberg, 1969). Now consider an individual process to be correct
if it produces the correct output for the input it was given. A simple
model is that the response is correct for the given stimulus if and only if
every process is correct, and that the probability all are correct is the
product of the probabilities that each is individually correct
(Schweickert, 1985). Then two factors that change the probability
different processes are correct will have multiplicative effects on the
probability of a correct response (sometimes called percent correct).
An assumption is made relating the duration of a process to its
probability of being correct. Suppose the probability a process A is
correct is a function πA(A) of its duration A (where we use the same
symbol for a process name and its duration). Then the probability of a
correct response for two processes A and B, with durations A and B,
respectively, is
Selective Influence with Accuracy, Rate, and Physiological Measures 295
π(correct) = πA(A)πB(B).
P(correct) = PA(A)PB(B).
E[P(correct)] = E[PA(A)PB(B)].
E[P(correct)] = E[PA(A)]E[PB(B)],
πijk = πi πjk.
1 1
5
k
ijk i
5
k
jk
.
Rates
rij = aibj.
The combined effect of changing the levels of both factors will be the
product of their separate effects; that is, the factors will have
multiplicative effects.
An example given by Roberts (1987) is an experiment by Clark
(1958). He manipulated reward schedule and food deprivation time for
rats that pressed a lever for food. Each group of rats had one of three
reward schedules, a Variable Interval of 1, 2 or 3 minutes. The groups
were tested after being fed, at delays of 1, 3, 5, 7, 10, 20, or 23 hours.
The effects of reward schedule and deprivation time on rate of lever
pressing were multiplicative. For an excellent review of many
experiments with multiplicative factors see Roberts (1987).
The basic model with a generator followed by a filter consists of two
sequential processes, but Roberts considered several arrangements of
processes. One, his Model 3, combines additive and multiplicative terms:
Processing Trees
Source monitoring
Prospective memory
Immediate memory
the item correctly. If the item is neither intact nor redinetgrated an error
is made. The probability of a correct response is
L + (1 − L − I)(1 − R),
where L is the probability the trace is lost. On the other hand, if some
phonemes are present in a partially complete trace, then the item
produced at recall will not be completely incorrect. By scoring items as
incorrect, partially correct and correct, Thorn, Gathercole, and Frankish
(2005) showed that long term knowledge has effects at more than one
locus in the process arrangement. For more details and current thinking
about the relation between long term and short term memory, see Thorn
and Page (2009).
We gain considerable understanding of what a model means about
mental processing when we find an experimental manipulation that
selectively influences a particular component of a model. There is a
concrete connection: This factor changes that parameter. But the further
question of what mental process the factor modifies is not necessarily
easy to answer.
Despite the complications, examining a simple tree structure is
worthwhile. In a later section on Tree Inference, we argue that if a
subject uses a complex tree to produce responses, but the experimenter
uses only a couple of factors that selectively influence processes, the
experimenter will find that a simple tree accounts for the results. So, let
us turn to some further examples of factors selectively influencing
parameters in the simple model of Eq. (9.1).
When a subject studies a list of items and then studies a second list,
memory for the second list is worse than if the first list had not been
studied, a phenomenon called proactive interference (or proactive
inhibition). The phenomenon is well established in long term memory
tasks. For a while after rapid forgetting was demonstrated in immediate
serial recall by Brown (1958) and by Peterson and Peterson (1959), it
seemed proactive interference did not occur in short term memory tasks.
Selective Influence with Accuracy, Rate, and Physiological Measures 305
Table 9.1
Keppel and Underwood (1962) Experiment 2
71 − 58 − 71 + 41 = −17.
Interaction contrasts in the table are negative. One can quickly see that
some other interaction contrasts are negative by noting that the
interaction contrasts themselves are monotonic in the rows and columns.
A check shows that all other interaction contrasts are negative, with
Selective Influence with Accuracy, Rate, and Physiological Measures 307
Table 9.2
Keppel and Underwood (1962) Experiment 2
Interaction Contrasts, Lowest Cell as Baseline
negligible exceptions.
It is worth fitting the model. Parameters were estimated to minimize
G2; estimation details are in the Appendix. Estimated values were I1 =
.723, I2 = .526, I3 = .407, and R1 = .959, R2 = .494, R3 = .296. With these
values the predicted probability of correct recall for Trial 1, Retention
Interval 18 seconds is, for example,
experiment by Tehan and Turcotte (1997), for which some data are
available3. For related work, see Tehan and Turcotte (2002). Tehan and
Turcotte (1997) manipulated the amount of proactive interference in
immediate serial recall by changing the similarity between the list to be
recalled and a preceding studied list. On trials to be analyzed, two four-
item lists were presented. The second list was always words. The first
list was words on half the trials (more interference) and letters on the
other half (less interference).
Method
Subjects were 40 introductory psychology students, who received course
credit for participating. Twenty subjects recalled by speaking and 20 by
writing. There were 10 one-list trials, not analyzed, and 40 two-list trials.
Instructions were to recall the most recent list, so when only one list was
presented subjects recalled that list and when two lists were presented
subjects recalled only the second list. On two-list trials subjects studied
the first list because they did not know whether or not they would need to
recall it.
Items were presented on a screen, one per second. The letter pool
was the English consonants, except “y”. For word lists, there was an
open pool and a closed pool. The open pool was 240 one syllable English
words, mostly concrete nouns. The closed pool was 16 words randomly
chosen from the open pool. Items were chosen randomly from the
designated pool, without replacement, to make the lists. Each subject was
presented with one set of trials using the open pool and one set using the
closed pool. The order of the sets was counterbalanced over subjects.
Data are available for the closed pool, but not for the open pool.
On half the two-list trials, a four digit number was displayed after the
second list. The subject was to decide if it was greater than or less than
5000. This task provided a delay prior to recall. Subjects who recalled
by speaking responded to the digit task by pointing up to indicate
“greater” and pointing down to indicate “less”. Those who recalled by
writing wrote + to indicate “greater” and − to indicate “less”. Subjects
were asked to recall the items in order.
Selective Influence with Accuracy, Rate, and Physiological Measures 309
Results
Frequency of recall over all subjects for the closed pool is in Table 9.3.
For statistical analysis, Tehan and Turcotte summed over serial positions.
A mixed ANOVA was conducted, with Interference, Delay and Word
Pool as within subject factors and Response Modality as a between
subject factor. Recall was significantly better with less interference, with
no delay, and with the closed word pool. There was a nonsignificant
trend for written recall to be better than oral recall. There were no
significant interactions.
Table 9.3
Frequencies of Recall from Experiment 2
of Tehan and Turcotte (1997), Closed Word Pool
Discussion
The experiment provides clear evidence of proactive interference in
immediate serial recall. Tehan and Turcotte found the effect on errors of
more interference was almost entirely due to more intrusions from the
prior list. More interference produced no significant differences between
omissions, transpositions or phonemic errors.
Tehan and Turcotte explained their results through reasoning of
Nairne and Kelley (1999), who say immediate serial recall requires two
discriminations, a discrimination of the current list from previous lists
and a discrimination among items in the current list. Between list
discrimination is more difficult when the prior and current lists are both
of words. Discrimination among items in the current list is more difficult
when items are phonologically similar (Tehan & Humphreys, 1995).
Tehan and Turcotte say that although phonological similarity was not
directly manipulated in this experiment, it may be larger in the closed
pool.
for a correct response, or (2) there is no such path. If there is such a path
the selectively influenced processes are ordered, if not they are
unordered. With four factors, there are six pairs of factors, and if we
consider the possibility that each pair is either ordered or unordered there
are 64 possibilities to contemplate. Moreover, if two factors are ordered
on a path, they can be ordered in two different ways. We need a quick
way to reduce the number of trees to consider. Qualitative tests are
invaluable for this.
If a factor changes parameters on children of a single vertex, we say
the factor selectively influences the vertex. If two factors selectively
influence two unordered vertices, the factors have additive effects. (This
will be discussed later in the section on Tree Inference) Let mij denote
the observed number of correct responses when Factor Α is at level i and
Factor Β is at level j. Choose any level i* of Factor Α and any level j* of
Factor Β as reference levels. We can consider an interaction contrast for
the number of correct responses in the same form as that previously
considered for mean reaction times, namely
If the two factors are additive, then all such interaction contrasts are 0.
In the experiment of Tehan and Tercotte (1997), despite the
nonsignificant ANOVA interactions, no two factors seem to have
additive effects. In particular, the interaction between Interference and
Response Modality is − 58 and the interaction between Delay and
Response Modality is − 42. (The reference levels are oral response, more
interference, and long delay.)
If two factors are not additive, but selectively influence different
vertices in a processing tree, then the two vertices are on a path from the
root to a terminal vertex, and so the vertices are ordered. Suppose the
first vertex is selectively influenced by Factor Α and the second by Factor
Β. Then the data must satisfy the following qualitative tests. These order
the levels of Factor Β (an order separate from the order of the vertices).
Let mij denote the observed number of correct responses when Factor Α
is at level i and Factor Β is at level j.
312 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Condition (a). Levels of Factor Β can be ordered so if j > j' then for
all i,
Fix levels i and i'. Then for all j and j' with j > j' all such interaction
contrasts have the same sign, negative, zero or positive.
Note that Conditions (a) and (b) do not require the levels of Factor Α
to be ordered. The qualitative Conditions (a) and (b) must be satisfied
for all combinations of levels of the factors other than Α and Β. (The
ordering of the levels of Factor Β and the sign of the interaction contrasts
may change from one combination to another.) With a plausible
additional assumption, Conditions (a) and (b) must also be satisfied when
frequencies are summed over all levels of the other factors; this usually
helps clarify the situation by increasing the sample size. For details, see
the later section on Tree Inference.
In the data of Tehan and Turcotte (1997), Condition (a) is easily seen
to be satisfied. With minor exceptions, the frequency of a correct
response decreases monotonically as Delay, Interference, and Serial
Position increase, and as Response Mode changes from written to oral.
Levels of each factor are ordered in the same way at all levels of the
other factors. For Condition (a) to be satisfied, only levels of the second
factor of every pair need be ordered; in fact, levels of all factors are
ordered.
Because Condition (a) is satisfied we turn to Condition (b). For a pair
of factors such as Delay and Interference, each with two levels, only one
interaction contrast is possible, so Condition (b) is automatically
satisfied. The only factor with more than two levels is Serial Position,
and we can test each other factor with it. Suppose in Condition (b)
Factor Β, the factor that selectively influences the vertex that comes
Selective Influence with Accuracy, Rate, and Physiological Measures 313
second on the path, has only two levels. Then Condition (b) is
automatically satisfied. Hence, we only need to test Condition (b) for
cases where Serial Position is in the role of Factor Β.
For Delay and Serial Position, using delayed-recall and serial position
4 as reference levels, the interaction contrasts for serial positions 1, 2 and
3 in that order are
(Frequencies are summed over the levels of the factors irrelevant to the
contrast.) These particular interaction contrasts have the same sign, so
Condition (b) is passed for them. All such interaction contrasts must
have the same sign. A quick test that some of the others have a negative
sign is whether the above interaction contrasts are monotonic with serial
position. They are not exactly monotonic, but close. Further tests reveal
no serious violation of Condition (b), so we continue to consider these
two factors for the model.
For Interference and Serial Position, using more-interference and
serial position 4 as reference levels, the interaction contrasts for serial
positions 1, 2, and 3 in order are
− 16, − 10, 2.
Two of the three are negative and the third is not far off. The interaction
contrasts are monotonic with serial position, an immediate test that
certain other interaction contrasts are negative. Further checks show no
serious violation of Condition (b) so we continue considering these two
factors for the model.
For Response Mode and Serial Position, using oral-response and
serial position 4 as baseline, the analogous interaction contrasts in order
are
2, 2, 32.
314 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
These all have the same positive sign. However, using as reference
levels oral-response and serial position 3, the interaction contrasts for
serial positions 1 and 2 are
− 30, − 30.
Fig. 9.1. A processing tree for data of Tehan and Turcotte (1997).
Any order of the vertices selectively influenced by the factors would lead
to the same equation. We first discuss model fit, and then consider
interpretations.
Observed frequencies and those predicted by the model in Figure 9.1
are in Table 9.4. Clearly, agreement is good, and there is no need to
modify the model. The model was fit for each response modality
separately. One way to use the two free scaling parameters is to fix two
of the processing tree parameters ahead of time to arbitrary values. That
was done; S1 (for less interference) was set to .6 for each response mode,
as was F1(for no delay). Other parameters were estimated to minimize
G2, see Appendix.
Parameters not fixed are higher for written responses than for oral,
with one exception. At the last serial position I4 is higher for oral than
written responses. This is a slight recency effect, consistent with the
finding that in immediate serial recall a recency effect is commonly found
316 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Table 9.4
Parameter Estimates and Model Fits for Experiment 2
of Tehan and Turcotte (1997)
with auditory presentation but not with visual presentation. This recency
effect is one reason that Serial Position and Response Modality behave in
a complex way.
Interpretation
A few ways to interpret processes in the tree are possible. It is one thing
to know a factor changes a process; it is another thing to know what that
process does. A difficulty in interpreting processes in this model is that
the order in which the factors have their effects is not established; the
model would fit just as well for any permutation of the three processes.
In the data of Keppel and Underwood (1962), we interpreted a longer
retention interval as increasing the probability an item is degraded
and proactive interference as decreasing the probability of redintegration.
Let us keep those interpretations, keeping in mind that other
interpretations are possible.
One way to think of serial positions is to note that because items must
be recalled in order, at the moment of recalling the item in position p, p −
1 items have already been recalled. Serial positions preceding an item
Selective Influence with Accuracy, Rate, and Physiological Measures 317
I = A + (1 − A)F.
1 − I = (1 − A)(1 − F).
Table 9.5
Data for Experiment 2 of Poirier, Schweickert & Oliver (2005)
For the serial position and list length, the upper right cells of the data
matrix are empty, so interaction contrasts can only be examined
piecemeal in the lower cells. For Serial Positions 1 through 4 and List
Lengths 5 through 7, interaction contrasts are in Table 9.6. Interaction
contrasts in the table were calculated as follows. Let mij be the observed
frequency of correct responses for serial position i, list length j, summed
over Word Length and Presentation Modality. (Summing over levels of
irrelevant factors requires that a plausible assumption be met, as
explained in the section on Tree Inference.) For the levels of the factors
in the table, the lowest frequency is for serial position 4, list length 7.
The interaction contrast for serial position i and list length j is
Interaction contrasts in the table are all negative, so the qualitative test
Condition (b) is passed for these levels of the factors. Further, the
interaction contrasts are monotonic with Serial Position and List Length,
indicating that certain other interaction contrasts have the same negative
sign. All possible such interaction contrasts need to be tested; with
negligible exceptions, all are negative. We consider Serial Position and
List Length as candidate factors in a model. If these factors selectively
influence two different vertices, the test is not informative about their
order.
Table 9.6
Interaction Contrasts for Serial Positions 1 through 4 and List Lengths 5 through 7
Summed over Word Length and Presentation Modality
Data from Experiment 2 of Poirier, Schweickert & Oliver (2005)
Serial Position
List Length 1 2 3 4 5 6 7
2
3
4
5 −259 − 210 − 133
6 − 89 − 67 − 44
7
Selective Influence with Accuracy, Rate, and Physiological Measures 323
In a moment we will fit a model for Serial Position and List Length,
keeping separate each combination of Word Length and Presentation
Modality. Before doing so, we need to check for each combination
whether interaction contrasts analogous to those just considered all have
the same sign. It turns out they all do, negative, with negligible
exceptions. The combination that fares the worst is long words read
silently. Interaction contrasts for this combination are displayed in Table
9.7, where it can be seen that violations are slight.
Table 9.7
Interaction Contrasts for Serial Positions 1 through 4 and List Lengths 5 through 7
Long words, read silently
Data from Experiment 2 of Poirier, Schweickert & Oliver (2005)
Serial Position
List Length 1 2 3 4 5 6 7
2
3
4
5 − 23 − 23 1
6 − 7 − 8 2
7
Qualitative tests for the other pairs of factors do not eliminate any
factors as candidates for selectively influencing vertices in a processing
tree. However, we will see in a moment that results of fitting a separate
model to each combination of Word Length and Presentation Modality
suggest it would not be fruitful to incorporate these two factors in a
processing tree model. Before turning to model fitting we briefly
continue with the qualitative tests, which are informative about some
factor pairs.
For Word Length and Serial Position, interaction contrasts were
calculated as follows. Let mi,w,j be the observed frequency of correct
responses at serial position i with word length w and list length j,
summed over presentation modalities. For each list length the last serial
position is not considered, so the lowest observed frequency is in the
penultimate position. Label this position with i = p. Then a contrast for
324 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
List length is irrelevant for this test. Every position does not occur in
every list length, so for a given position i the expression above was
averaged over the list lengths in which the position i occurred. (We are
interested in the sign of the interaction contrast, which would be the
same whether we add or average over list lengths. Averaging makes it
easier to compare the results for various serial positions.)
Results for serial positions 1 through 5 in that order are
Clearly, these do not have the same sign. This does not indicate that
these two factors do not selectively influence two different vertices. But
if the vertices are on a path that starts at the root, the vertex for Serial
Position cannot follow the vertex for Word Length.
Now consider Presentation Modality and List Length. For each
position, the cell with silent reading and list length 7 was the reference
for the interaction contrasts. Let mi,h,j be the observed frequency of
correct responses at serial position i, with presentation modality h (aloud
or silent), and list length j, summed over word length. For each serial
position the lowest observed frequency is at list length 7. Then a contrast
for serial position i and list length j is
Because Serial Position is irrelevant for this test, for a given list length j
the above contrasts were averaged over the serial positions occurring at
that list length, to form an interaction contrast for list length j. Resulting
interaction contrasts for list lengths 3, 4, 5, and 6, in that order, are
These do not have the same sign. Further, they are not monotonic, which
implies that other such interaction contrasts do not have the same sign.
We conclude that if Presentation Modality and List Length selectively
influence different vertices on a path that starts at the root, the vertex for
List Length is not the later one.
There are two more pairs of factors to consider, Word Length with
List Length, and Presentation Modality with Serial Position. It turns out
that there are only minor violations of Condition (b), i. e., Inequality
(9.3), for these pairs. If these factors selectively influence different
vertices on a path, Condition (b) is not informative about the order in
which the vertices occur.
length effects do not occur on the first few trials. It would be difficult to
produce a model in which these factors change parameters in a simple
way.
The generally good fit of the model supports the interpretation that
serial position selectively influences degradation, also supported by the
data of Tehan and Turcotte (1997). The generally good fit also supports
the interpretation that list length selectively influences redintegration.
From the parameter values, one can see that an increase in list length has
a more harmful effect on redintegration for long words than short words.
But as noted earlier, a model may be correct although interpretation of its
parameters is false. We now turn to experiments on that lead to a
reinterpretation of the parameter interpreted as redintegration probability.
Hulme, Maughan and Brown (1991) found better immediate serial recall
for words than nonwords. They explained this by saying subjects have a
representation of the features of a word in long term memory, and this
knowledge supports recall. No such representation is available for
nonwords. A nonword can be considered an extreme type of rare word,
suggesting that the long term representation of an infrequent word would
be less useful at recall than that of a frequent word. If so, redintegration
would be better for high than low frequency words, producing better
recall for high frequency words.
Lists of high frequency words and of low frequency words were
tested in immediate serial recall by Hulme, Roodenrys, Schweickert,
Brown, Martin and Stuart (1997). Recall was indeed better for high
frequency words. Typical serial position curves were found. Recall
decreased as serial position increased, but rose sharply at the last two
serial positions, a recency effect. Serial position and word frequency
interacted, with a smaller effect of word frequency at early serial
positions than in the middle. The combined effect of an increase in recall
due to serial position change and an increase due to word frequency was
smaller than the sum of their individual effects, a negative interaction.
Suppose as serial position increases, an item becomes more degraded,
Selective Influence with Accuracy, Rate, and Physiological Measures 327
pij = Ii + (1 − Ii)Rj.
The model accounted well for the data, except for the recency effect in
the last two serial positions; see Hulme et al. (1997) for details.
Further support for the redintegration interpretation comes from an
experiment by Hulme, Stuart, Brown and Morin (2003), in which lists of
alternating words and nonwords were tested in immediate serial recall.
As predicted, words were always recalled better than nonwords; even
when alternating in the same list.
But another experiment of Hulme et al. (2003) challenges the
interpretation of Word Frequency having its effect through
redintegration. In this experiment, lists of alternating high and low
frequency words were tested in immediate serial recall. Also tested were
pure lists of all high frequency and all low frequency words. Suppose
the frequency of a word produced its effect by making redintegration of
that particular word better. Then with a list of alternating high and low
frequency words, recall should alternate between high and low, in
correspondence with the frequency. Instead, recall for a word in
alternating lists did not depend on its frequency. In an alternating list,
recall for a word in a particular serial position, say position 4, was about
the same whether the word was high or low frequency. Recall for words
in alternating lists fell between recall for pure high frequency and pure
low frequency lists. Hulme et al. (2003) explained their results by noting
that inter-item associations are higher for high frequency than low
frequency words. The availability in memory of words in a list depends
on the associations between all of them, i.e., on their combined inter-item
associations. It follows that recall for words in alternating lists is
between that of high and low frequency words.
With this interpretation, R in Equation (9.1) is the probability of
retrieving a word through its associations with other list words. If so, it
328 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Two recent studies show a benefit of nocturnal sleep on recall when new
interfering associations were learned before testing of old ones
(Ellenbogen, Hulbert, Stickgold, Dinges & Thompson-Schill, 2006;
Ellenbogen, Hulbert, Jiang & Stickgold, 2009). A tree model allows
testing between two possibilities. One is that sleep decreases the
probability interference occurs. Another is that interference occurs with
the same probability with or without sleep, but given that interference
does occur, sleep decreases the probability it leads to a recall error.
In both experiments all subjects learned a list of A-B word pairs.
Some learned in the morning (the Wake group), some in the evening (the
Sleep group). All were tested 12 hours after learning, so those who
studied in the evening had a night of sleep before testing. Immediately
before testing, half the subjects learned a new list of A-C associations.
Learning new associations leads to poorer recall of old ones, attributed to
retroactive interference. Recall of B items was indeed worse when new
associations to C items were learned, whether subjects studied in the
morning or evening. But the decrement was less for subjects having a
night of sleep.
Selective Influence with Accuracy, Rate, and Physiological Measures 329
Table 9.8
Data of Ellenbogen et al. (2006)
Table 9.9
Data of Ellenbogen et al. (2009)
P + (1 − P)Q. (9.4)
Tree Inference6
We have seen examples of tasks in which processing trees fit data well.
But if the brain is complex, why are the trees simple? The answer is that
if a few factors are found that selectively influence processes in a tree,
the tree will be equivalent to a simple tree. An investigator often begins
with a processing tree based on intuition and tests it through goodness of
fit. But one can start with data from a factorial experiment and test
whether any processing tree exists that can account for the data, under
the assumption that the factors selectively influence processes in the tree.
If any such tree exists, the data impose a simple equivalent tree. The
approach is called Tree Inference (Schweickert & Chen, 2008;
Schweickert & Xi, 2011).
Consider an experiment with two response classes, say, correct and
Selective Influence with Accuracy, Rate, and Physiological Measures 331
This case is illustrated in the left panel of Figure 9.2. One sees
immediately that if this model is true, Factors Α and Β have additive
effects on the probability of a correct response. This turns out to be
necessary and sufficient for a standard tree for unordered processes to
predict the probability of a correct response. In other words, if two factors
have additive effects on the probability of a correct response, parameter
values in the standard tree for unordered processes can be found that
predict the probability of a correct response (and thereby, of course, the
332 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Fig. 9.2. Left: standard tree for unordered processes. Right: standard tree for ordered
processes.
This case is illustrated in the right panel of Figure 9.2. The tree of
Gathercole, Frankish, Pickering and Peaker (1999) has this form for
incorrect responses. Note that parameter values in Equations (9.5) and
(9.6) are not unique, see Schweickert and Chen (2008).
Necessary and sufficient conditions for two factors to selectively
influence different processes in the standard tree for two ordered
processes are the following (Schweickert & Chen, 2008, Theorem 11).
Let pij be the probability of a correct response when Factor Α is at level i
and Factor Β is at level j. Matrix (pij) is produced by Factor Α and Factor
Β selectively influencing two vertices in the standard tree for ordered
processes, with the vertex indexed by i preceding the vertex indexed by j,
iff
Selective Influence with Accuracy, Rate, and Physiological Measures 333
1. The columns of (pij) can be numbered so j > j' implies for every i
that pij > pij'.
2. There exist levels i* and j* such that for every i there is a number
ri 0 with the property that for every j
For any pair of levels for which j > j', (pi*j − pi*j') > 0. If ri − ri' < 0, all
interaction contrasts of the form above will be negative, and for fixed j'
decreasing in j. If ri − ri' = 0, all will be 0. If ri − ri' > 0, all will be
positive, and for fixed j' increasing in j. This establishes Condition (b).
If the qualitative tests in Conditions (a) and (b) are satisfied, it is
worth estimating parameters and fitting the model. Parameter values are
not unique. Suppose the probability of a correct response at factor levels
i and j is given by Equation (9.6), for probabilities wi, xi, and yj, that is
pij = wi + xi yj.
An equation of the same form holds for other parameters, w*i, x*i, and
y*j, that is,
x*i = cxi,
w*i = wi − cdxi,
and
y*j = yj/c + d.
standard tree for ordered processes, with the vertex selectively influenced
by Factor Α preceding that selectively influenced by Factor Β. Suppose
each subject’s parameter values are different. The mixture of trees may
be equivalent to the standard tree for two ordered processes. The key
requirement is that for all trees in the mixture, the probabilities of
reaching the vertex selectively influenced by Factor Α are proportional,
the proportion not changing when the level i of Factor Α changes.
To see this, suppose for tree t when Factor Α is at level i and Factor Β
is at level j the probability of a correct response is
pij wi xi y j .
pij p
t
t ijt w x
t
t it
t
t it y jt w
t
t it xi y
t
t t jt .
xit are proportional. Then the mixture is equivalent to the standard tree
for ordered processes. If combining data over subjects or over levels of
irrelevant factors does not lead to one of the standard trees, one would
consider, of course, trees for individual subjects or combinations of
factor levels.
An important point is that if Factors Α and Β are not additive and
Conditions 1 and 2 are not satisfied, then the factors do not selectively
influence two different vertices in any processing tree. The standard tree
for unordered processes and the standard tree for two ordered processes
are ruled out, but so are all others. For more details, see Schweickert and
Chen (2008).
So far we have assumed a tree with three restrictions, (a) parameters are
probabilities, (b) there are two response classes, and (c) Factors Α and Β
each selectively influence a single vertex. It would be useful to
overcome these restrictions. Instead of probabilities, Roberts (1987)
used rates. Two response classes are not enough for tasks such as source
monitoring. Finally, in some tree models the same parameter appears in
several places; an example is the probability of guessing the correct
response.
Removing the first two restrictions is straightforward, but analysis is
complicated if factors change parameters at multiple vertices. Results
have been derived for trees in which factors are either additive or interact
multiplicatively (Schweickert & Xi, 2011). Beyond this, interactions in
the form of sums of products can sometimes be treated with the matrix
factorization model (Bamber & van Santen, 1978, 1980; Ollman, 1990)
described later.
Suppose a tree has a single root and each terminal vertex is labeled
with one of K classes of responses. Suppose the parameters on the arcs
are nonnegative numbers, not necessarily probabilities. Suppose when
the level i of Factor Α changes, parameters on some set of arcs change,
and when the level j of Factor Β changes parameters on a different set of
Selective Influence with Accuracy, Rate, and Physiological Measures 337
arcs change. Suppose these sets are mutually exclusive. There may be
some arcs whose parameters do not change when i or j change. If the
parameter on an arc changes when the level i of Factor Α changes, we
say the arc is indexed by i. An arc indexed by j is defined similarly.
When Factor Α is at level i and Factor Β is at level j, let pij.k denote the
value of the dependent variable for class k. The dependent variable is a
nonnegative number such as the probability of a response in class k or
the rate at which responses of class k are made. It may be larger than 1.
The situation is straightforward if no path from the root to a terminal
vertex contains both an arc indexed by i and an arc indexed by j. Factors
Α and Β will have additive effects, and the tree is equivalent to the
standard tree for unordered processes, enlarged to allow more than two
response classes. There is a parameter α, and for every response class k
there are parameters wi.k and zj.k such that
K K
with wi.k z 1 if the dependent variable is probability.
k 1 k 1 j .k
K K K
with wi.k y j .k z j .k 1 if the dependent variable is
k 1 k 1 k 1
suppose there is a vertex v such that a path from the root to v contains
arcs indexed by i and no arcs indexed by j, and one or more paths from v
to terminal vertices contain arcs indexed by j and no arcs indexed by i.
Suppose no other paths from the root to a terminal vertex contain both
arcs indexed by i and arcs indexed by j. Then the tree is equivalent to the
standard tree for multiplicatively interacting factors.
Under certain conditions, a tree will be equivalent to the standard tree
for multiplicatively interacting factors even if it has several paths from
the root to a terminal vertex containing arcs indexed by i and also arcs
indexed by j. Further, if every tree in a mixture of trees satisfies certain
conditions, the mixture of trees will be equivalent to the standard tree for
multiplicatively interacting factors. See Schweickert and Xi (2011) for
details.
The following are necessary and sufficient conditions for the standard
tree for multiplicatively interacting factors (Schweickert and Xi, 2011).
The key condition is that interaction contrasts can be written as a product
in which one multiplier depends only on i and the other multiplier
depends only on j. As before, interaction contrasts are calculated with
respect to reference levels. For Factor Β the reference level for one
response class may not be the same as that for another response class.
To emphasize this, when a level j of Factor Β is used as a reference level
for class k the level is written j(k). An expression such as pij*(k).k denotes
the value of the dependent variable for class k, when Factor Α is at level i
Selective Influence with Accuracy, Rate, and Physiological Measures 339
1. There exists a level i* of Factor Α and for every k' there exists a
level j*(k') of Factor Β such that for every i, j and k there exist ri,
0 < ri < 1, and sj.k, 0 < sj.k < 1, such that
2. For two levels i and i', ri ri' and for some k, for two levels j and
j', sj.k sj'.k.
K
3. For every j, s a , where a is a constant, 0 < a < 1.
k 1 j .k
Levels i and j.k can be ordered so the interaction contrasts fall into four
quadrants, positive when ri < ri' and sj.k < sj'.k, negative when ri < ri' and
340 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
sj.k > sj'.k, and so on. To be more specific, choose a level i of Factor Α. In
the matrix of interaction contrasts, (pij.k − pij'(k).k − pi'j.k + pi'j'(k).k), order the
columns so interaction contrasts in row i are monotonically increasing.
Then the interaction contrasts in all other rows should be monotonically
increasing or monotonically decreasing. Likewise, choose a level j of
Factor Β. For class k call this level j.k. Now in the matrix of interaction
contrasts order the rows so the interaction contrasts in column j.k are
monotonically increasing. Then the interaction contrasts in all other
columns should be monotonically increasing or decreasing. When an
ordering of levels i is found for some response class k it should be
possible to use that same ordering for all classes. The analogous
statement need not be true for ordering the levels j of Factor Β.
A plan for finding a tree with Tree Inference is the following. For a
given response class consider two factors. Are both ordered? If so, and
they are additive, fit the standard tree for unordered processes. If both
factors are ordered but they are not additive, are they multiplicative? If
so, consider a simple tree in which the probability of the response class is
pq. Is one factor ordered, but not the other? Check Conditions (a) and
(b), in Equations (9.2) and (9.3). If these are satisfied, try fitting the
standard tree for ordered processes. Is neither factor ordered? Choose a
reference level of Factor Α and a reference level of Factor Β. Calculate
interaction contrasts with respect to these levels. Do they form a 2 2
checkerboard, with positive interaction contrasts in upper right and lower
left, negative in lower right and upper left? If so, try fitting the standard
tree for multiplicatively interacting factors.
Contingency Matrices7
M
pgk p
h 1
gh phk .
A = (pgh).
B = (phk).
C = (pgk).
342 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Then
C = AB.
where
A1
A* and B* B 1 B 2 .
A 2
C C 21
C * * 11 .
C12 C 22
The rank of a matrix is the largest number of columns that are linearly
independent of all other columns. If X is a matrix with r rows and c
columns, then rank(X) < min{r, c}. For any level i of Factor Α, matrix
Ai has one row for each different stimulus and one column for each
mental state. Hence, rank(Ai) < min{S, M}. (The entries in each row
sum to 1, but this does not prevent the columns from being linearly
independent.) Likewise, for any j, matrix Bj has one row for each mental
344 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
state and one column for each different response. Hence, rank(Bj) <
min{M, R}.
Going further, matrix
A
A* 1
A 2
has 2S rows and M columns. Its rank is less than or equal to {2S, M}.
With I levels of Factor A the rank of this left factor is less than or equal
to min{IS, M}. Matrix
B* B1 B2
Physiological Measures
where C(t) is the total contribution from sources other than the two
groups of neurons.
Suppose manipulating one experimental factor changes activity of the
neurons producing signal A(t), changing that signal, but leaving B(t) and
C(t) invariant. Suppose manipulating another experimental factor
changes activity of the neurons producing signal B(t), changing that
signal but leaving A(t) and C(t) invariant. Finally, suppose manipulating
neither factor changes the coefficients b1 and b2. Then the combined
effect of manipulating the two factors will be the sum of their individual
effects. Because potentials are additive, the factors are additive. Factors
will be additive at every electrode location satisfying these assumptions;
each location will have its own coefficients b1 and b2, and its own
residual signal C(t). Note that if the factors both have effects at the same
time t the neurons that the factors are influencing are active at the same
time; that is, the factors are influencing simultaneous processes, not
sequential processes.
Testing factors for such additivity using voltage as the dependent
variable is called the Additive Amplitude Method. Additive effects of
two factors were found by Holcomb and Kounios (1992) in an
experiment on sentence processing. Subjects were presented with a
sentence, such as “No dogs are animals,” and responded whether it was
true or false. The following two factors were among those manipulated:
subject and predicate were unrelated and for a superset sentence, but
there was no interaction. It is noteworthy that the factors had additive
effects not only at the peak of the N400, but over the time interval
surrounding it. Holcomb and Kounios concluded that the factors affected
processes executing in parallel, one for relation and one for generality.
Looking for additive and interactive effects of factors on ERP was
also a goal of Gondan and Röder (2006), who investigated integration of
multi-sensory information. At a particular point on the scalp at time t let
the ERP amplitude be V(t) when a visual stimulus alone is presented, A(t)
when an auditory stimulus alone is presented, and AV(t) when the
auditory and the visual stimulus are presented together. Barth, Goldberg,
Brett and Di (1995) proposed that sensory integration is occurring at
times t at which
If AV(t) is the sum of A(t) and V(t) the above expression equals − C(t),
not 0.
Gondan and Röder (2006) introduced a new procedure. Typical
experiments on sensory integration can be considered 2 2 factorial
designs; one factor is presence or absence of the auditory stimulus, the
other is presence or absence of the visual stimulus. One condition is the
baseline, with no visual or auditory stimulation. To ignore the baseline is
to implicitly assume there is no common potential. Gondan and Röder
348 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
Here FTAV(t) is the cumulative distribution function for the reaction time
when tactile, auditory and visual stimuli are all presented; other notation
is similar. Violations of the race inequality showed evidence of sensory
integration (coactivation) between auditory and visual stimuli, but no
evidence of it between tactile and other stimuli.
Analysis of the ERP interaction contrast in Equation (9.9) showed
that prior to 84 msec, auditory and visual stimuli had additive effects on
ERP. Starting at 84 msec the interaction contrast was significantly
different from 0 at central electrode locations, evidence of sensory
integration. The time at which the interaction contrast differed from 0
and the shape of it depended on electrode location. In this experiment,
the expression proposed by Barth et al. (1995) and that proposed by
Gondan and Röder (2006) were similar for about the first 200 msec; in
this interval they lead to the same conclusions about whether sensory
integration is occurring. Later, after about 350 msec, the expressions
differ. That of Gondan and Röder (2006) is zero, interpreted as absence
of sensory integration, while that of Barth et al. (1995) is nonzero,
interpreted as an estimate of a potential common to all conditions.
Van Lankveld and Smulders (2008) found additive effects of two factors
on areas bounded by ERP curves. (Note that additive areas under the
ERP curves do not imply additive amplitudes of ERPs, or vice versa.)
Participants (all males) were asked to rate pictures in terms of felt
intensity (arousal) and experienced pleasure (valence). The main
question was whether erotic stimuli produce a specific ERP response. A
secondary question of the study is of primary interest here, whether
processing of arousal and of valence are processed independently
350 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
physiologically.
One special set of pictures was erotic (e.g., nude, heterosexual
behaviors); these were considered high arousal and positive valence.
Four other sets of pictures formed a 2 2 design of high and low arousal
crossed with positive and negative valence. Sports pictures (e.g., rafting
and parachute jumping) were considered high arousal and positive
valence. Pictures of traffic accidents, snakes, weapons, and horror were
considered high arousal and negative valence. Pictures depicting babies,
flowers, etc. were considered low arousal and positive valence
(low/positive). Finally, low arousal and negative valence pictures
(low/negative) showed crying people, garbage, cemeteries, etc. After
electrodes were attached, pictures were presented one by one. At offset
of each, the participant rated the picture on experienced pleasure and felt
intensity.
Van Lankveld and Smulders looked at the area under ERP curves in
two different time windows, which they called P300 (300-500ms) and
Positive Slow Wave (PSW; 500-700ms). The P300 ERP component is
thought to reflect the effect of perceptual saliency that attracts attention
(e.g., in the oddball paradigm) and PSW is thought to reflect the decision
or evaluation process; each component reflects other processes as well.
Figure 9.4 shows two critical dependent measures, the areas bounded
by ERP curves in intervals around P300 and PSW. Analyses showed
that the erotic pictures differed from the other high arousal positive
valence pictures, those of sports. Van Lankveld and Smulders interpret
this as indicating a specific ERP response to erotic stimuli.
In analyses of the nonerotic stimuli, in the P300 interval there was a
significant interaction between valence and arousal. However, in the
PSW interval, there were main effects of valence and arousal, but no
interaction. Van Lankveld and Smulders interpret the additivity as
evidence for independent processing of valence and arousal in this
interval.
Fig. 9.4. Mean areas under ERP curves in five different stimulus sets. Clear additivity is
shown when non-erotic (sports) images used as high arousal and positive valence
stimulus (bottom panel). From van Lankveld, J. J. D. M., & Smulders, F. T. Y., 2008,
The effect of visual sexual content on the event-related potential, Biological Psychology,
79, Fig. 4. Copyright 2008 Elsevier. Reproduced with permission.
brain area becomes active for information processing, blood around the
area has a ratio of oxyhemoglobins (hemoglobins with oxygen) and
deoxyhemoglobins (hemoglobins without oxygen) different from that of
brain areas not involved in the specific information processing. The
magnetic properties of hemoglobin depend on the level of oxygen in it.
The Magnetic Resonance Imaging (MRI) machine can detect the
different ratio of oxy- and deoxyhemoglobins, called the Blood-Oxygen-
Level-Dependence (BOLD) signal. Usually, a stronger BOLD signal
from an area indicates an increase in brain activity over some baseline;
neurons there fire more rapidly or more neurons there fire than before.
352 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
ones that participants had not seen before (new place). The long and
short-interval RS factors could be manipulated independently.
For example, two identical images in phase 2 (a no change condition
for the short-interval RS effect) could be an image seen in phase 1 (old
view condition for the long RS effect), an image that depicts a location
seen in phase 1 but from a different viewpoint (new view), or an image
not seen in phase 1 (new place).
Epstein et al. found that as the levels of long- and short-interval
repetition factors changed, the BOLD signal changed in the
parahippocampal region, which is partly responsible for memory
function, see Figure 9.5. Importantly, they found no interaction between
long and short-interval RS factors and concluded that two different RS
effects may be governed by two different brain functions that operate
independently (even though the functions are the results of neuronal
activities at the same physiological locations). Additivity of the two RS
factors can be clearly seen in Figure 9.5B.
Additional evidence for the conclusion of Epstein et al. (2008) was
that the long-interval RS effect was relatively invariant across the old
view and new view conditions and their activation changes were smaller
than those of the new place condition (see Figure 9.5B). This means that
the long-interval RS effect works as long as the images depict the same
location, regardless of the viewpoint. However, the short-interval RS
effect was much stronger in the no change condition compared to other
conditions (view change and place change). That is, the BOLD signal
change was smaller in the no change condition than in the place change
and view change conditions. This suggests that the short-interval RS
effect is mainly determined by the literal physical similarity of two
images presented successively and the long- interval RS effect by the
identity of the location that images show, even though in different
viewpoints.
Concluding Remarks
Fig. 9.5. A. Brain region (parahippocampal place area; PPA) that is selectively responsive
to scenes. The activated region is the result of region of interest analysis (ROI), which
identifies a region that is more responsive to experimental stimuli (scenes) than to others
(objects). B. fMRI response changes (%) in PPA for all nine conditions of RS factors.
The x-axis represents long-interval RS conditions. Different lines represent short-interval
RS conditions. (Adapted from Epstein, R. A., Parker, W. E., & Feiler, A. M., 2008,
Evidence for Dissociable Neural Mechanisms: Two Kinds of fMRI Repetition
Suppression? Journal of Neurophysiology, 99, Fig. 2. Copyright 2008 Society for
Neuroscience. Reproduced with permission.
Appendix
The model of Equation (1) was fit to the data using Excel solver.
Parameters were estimated to minimize G2. Let the observed frequency
in condition k be xk and the predicted frequency be mk. Then
G2 = − 2 Σ xk ln (mk/xk).
The sum is over both correct and incorrect frequencies in all conditions.
The goodness of fit statistic G2 is the log likelihood. For a given data set,
it is close in value to chi square, a commonly used alternative,
X2 = Σ(xk − mk)2/mk.
I' = 1 − (1 − I)/c
R' = 1 − c + cR.
In other words, the values of I and R are not unique; one is free to
choose a parameter c for change of scale. Parameter c must be chosen so
the transformed parameters I' and R' are between 0 and 1. If some
observed probability is 0 or 1 such a change of scale is impossible, but
this is not the case here. The scaling free parameter c adds 1 to the
356 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
For degrees of freedom for the experiment of Tehan and Turcotte, for
each response modality there are 32 observed frequencies. Because the
observed frequency of an error in a certain treatment is 200 minus the
frequency of a correct response, only 16 observed frequencies are
independent. The model has 8 parameters. But when calculating
degrees of freedom it is necessary to take into account that parameter
values are not unique; one set of valid parameters can be transformed
into another by multiplying by two scaling coefficients. In other words,
given x, y, and z that predict the probability of an error as a product, xyz,
transformed values x/a, y/b and abz for positive a and b will predict just
as well, provided a and b are within a range that leaves the transformed
values between 0 and 1. Another way to put it is that two parameters can
be fixed ahead of time (to values not too extreme for practical
computation). The number of parameters that need to be estimated is 6.
Hence the degrees of freedom are 16 − 6 = 10.
Fitting the data of Ellenbogen et al. (2006) and Ellenbogen et al. (2009)
Notes
359
360 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
on the same sample space and have a joint distribution. For each
treatment there is a random vector D = < A, B,..., Z>. Note we are
assuming no levels of factors outside Φ need be specified. We say the
distribution of the random vector <A, B,..., Z> depends only on factors in
the list Φ.
However, we do not assume the sample spaces for different
treatments are the same. In one treatment the outcome of the experiment
might be a button press and in another the outcome might be a spoken
word, forming different sample spaces. If in one treatment a stimulus is
bright and the duration of a perceptual process takes on a value, no value
is taken for what the duration of that process would be if the stimulus
were dim instead of bright. The random vectors for two different
treatments are not sampled together; that is, random vectors for different
treatments are not necessarily defined on the same sample space and do
not have a joint distribution. We say they are unrelated.
Selective Influence
values over which F(x) is constant, so F(x) may not have an inverse. We
define a function similar to an inverse as follows. For every p [0, 1]
let F − 1(p) = inf{x|F(x) = p}. Now let U denote a random variable
uniformly distributed between 0 and 1. The random variable F − 1(U) has
the same distribution as X, that is, F − 1(U) X. This is so because given
any x, P[F − 1(U) < x] = P[U < F(x)] = F(x). For one random variable to
be expressed as a function of another random variable the latter need not
have a uniform distribution, of course.
The definition of selective influence uses the notions of random
entities and measurable functions, which are defined in the Appendix of
Chapter 6. To avoid a possible source of confusion, note that symbol C
in the definition and symbols C1,..., Cp in Chapter 4 are unrelated.
For the special case where all random entities are random variables or
random vectors, the definition says the n random variables in the list
[A, B,..., Z ] are selectively influenced by the n Factors <Α, Β,..., Ζ>,
respectively, if there exists a random vector C (defined on some
probability space), and for every level k of every factor α in <Α, Β,..., Ζ>
there exists a real valued function fα,k such that for every treatment =
<i, j,..., m>
When the definition applies we sometimes write for short [A, B,..., Z]
↫ <Α, Β,..., Ζ>.
Remark 1
Remark 2
Remark 3
Remark 4
Marginal selectivity
Further,
A1 B 1
A1 1 − B 2
Hence
(e) B1 1 − B2.
A2 is stochastically independent of B1
A2 is not stochastically independent of B2.
Selective Influence of Interdependent Random Variables 369
Hα,k = fα,k(C).
On the other hand, suppose the random vector <H1,1,..., H1,I, H2,1,...,
H2,J,..., Hn,1,..., Hn,M> exists as described in the statement of the theorem.
Let C = <H1,1,..., H1,I, H2,1,..., H2,J,..., Hn,1,..., Hn,M>. For every level k of
factor α let
370 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
fα,k(C) = Hα,k.
Ai A (i ) B j B ( j)
and .
A (i ) B ( j)
n
(ij ) a k (i )bk ( j ) (10.2)
k 1
a 2
k 1
k 1 (10.3)
n
b
k 1
2
k 1.
Proof: Suppose Equations (10.2) and (10.3) hold. Let C1,..., Cn, SA
and SB be independent standard normal random variables. Let
372 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
n n
Ai 1 a (i ) S A ak (i )Ck f1i (C1 ,, Cn , S A )
2
k
k 1 k 1
n n
B j 1 bk2 ( j ) S B bk ( j )Ck f 2 j (C1 ,, Cn , S B ) .
k 1 k 1
n
n
V [ Ai ] 1 ak2 (i )V [ S A ] ak2 V [Ck ] 1.
k 1 k 1
Likewise, V [ B j ] 1.
Because Ai and Bj are standard, their correlation equals their
covariance,
Cosphericity
2. Correlations between Aij and Bij, ρ(i, j), i, j = 1,2, have the form of
Equation (10.2) with constraints (10.3).
Remark 5
Results above about cosphericity are stated for a 2 × 2 design, that is,
each factor has two levels. In a design with more factor levels, for the
factors to selectively influence random variables it is necessary that
cosphericity hold in every 2 × 2 sub-design. This is not sufficient,
however; see Kujala & Dzhafarov (2008, p. 142). Definition 10.1 might
apply for <i, j> {1, 2} × {1, 2} with a particular random entity C1 and
it might apply for <i, j> {2, 3} × {1, 2} with a different particular
random entity C2. It is not known whether one can always find a single
random entity C allowing Definition 10.1 to hold for the union of these
levels, <i, j> {1, 2, 3} × {1, 2}.
variables can be made. The Distance Test was developed for this
purpose by Kujala and Dzhafarov (2008); discussion here follows theirs.
The Distance Test is a consequence of the Minkowski inequality (e.g.,
Royden, 1968). For two random variables F and G defined on the same
probability space with a probability measure λ, the inequality has the
form, for any p > 1,
p
F (c ) G(c )
p p
p
F ( c ) G ( c ) d ( c ) p
d ( c ) p d ( c ) .
The integral is over all values of c. It follows that for any p > 1,
p p p
p
E [ F G ] p E [ F ] p E[ G ] .
p
p
E[ P Q ]
we can let
p
Dij p E[ Aij Bij ] .
max{D11, D12, D21, D22} < (D11 + D12 + D21 + D22)/2. (10.4)
p p
p
E[ P Q ] p E[ ( P R ) ( R Q ) ]
p p
p E[ P R ] p E[ R Q ].
p p p
p
E[ A11 B11 ] < p
E[ A11 A21 ] + p
E[ A21 B11 ]
and
p p p
p
E[ A11 A21 ] < p
E[ A11 B12 ] + p
E[ A21 B12 ] ,
so
p p p p
p
E[ A11 B11 ] p E[ A11 B12 ] p E[ A21 B12 ] p E[ A21 B11 ].
(10.5)
378 Discovering Cognitive Architecture by Selectively Influencing Mental Processes
p p
Dij p E[ Aij Bij ] = p
E[ Ai B j ] .
p p p p
p
E[ A1 B1 ] p E[ A1 B2 ] p E[ A2 B2 ] p E[ A2 B1 ] ,
so
Then
p
s ij p E[ hi ( Ai ) k j ( B j ) ] .
Concluding Remarks
Anderson, J. R., & Bower, G. H. (1974). Human associative memory. New York: Wiley.
Arbuckle, J., & Larimer, J. (1976). The Number of Two-Way Tables Satisfying Certain
Additivity Axioms. Journal of Mathematical Psychology, 13, 89-100.
Ashby, F. G. (2000). A stochastic version of general recognition theory. Journal of
Mathematical Psychology, 44, 310-329.
Ashby, F. G., Boynton, G., & Lee, W. W. (1994). Categorization response time with
multidimensional stimuli. Perception & Psychophysics, 55, 11-27.
Ashby, F. G., & Maddox, W. T. (1994). A response time theory of separability and
integrality in speeded classification. Journal of Mathematical Psychology, 38,
423–466.
Ashby, F. G., Prinzmetal, W., Ivry, R., & Maddox, W. T. (1996). A formal theory of
feature binding in object perception. Psychological Review, 103, 165-192.
Ashby, F. G., Tein, J.-Y., & Balakrishnan, J. D. (1993). Response time distributions in
memory scanning. Journal of Mathematical Psychology, 37, 526-555.
Ashby, F. G., & Townsend, J. T. (1980). Decomposing the reaction time distribution:
Pure insertion and selective influence revisited. Journal of Mathematical
Psychology, 21, 93-123.
Backus, B. T., & Sternberg, S. (1988, November). Attentional tradeoff across space early
in visual processing: New evidence. Paper presented at the meeting of the
Psychonomic Society, Chicago.
Bamber, D., & van Santen, J. P. H. (1978). A general method for the analysis of
conditional response frequencies. Unpublished manuscript, V. A. Hospital, St.
Cloud, MN.
Bamber, D., & van Santen, J. P. H. (1980, August). Testing discrete state models using
conditional probability matrices. Paper presented at the Mathematical
Psychology Meeting, Madison, WI.
Barth, D. S., Goldberg, N., Brett, B., & Di, S. (1995). The spatiotemporal organization of
auditory, visual and auditory-visual evoked-potentials in rat cortex. Brain
Research, 678, 177-190.
Bartle, R. G. (1964). The elements of real analysis. New York: Wiley.
Bartlett, J. C., & Searcy, J. (1993). Inversion and configuration of faces. Cognitive
Psychology, 25, 281-316.
Batchelder, W. H., & Riefer, D. M. (1986). The statistical analysis of a model for storage
and retrieval processes in human memory. British Journal of Mathematical and
Statistical Psychology, 39, 120-149.
383
384 References
Estes, W. K. (1991). On types of item coding and source of recall in short-term memory.
In W. E. Hockley & S. Lewandowsky (Eds.), Relating theory and data: Essays
on human memory in honor of Bennet B. Murdock (pp. 155-174). Hillsdale, NJ:
Erlbaum.
Feller, W. (1971). An introduction to probability theory and its applications. Vol. II (2nd
Ed.). NY: John Wiley & Sons.
Fific, M., Little, D. R., & Nosofsky, R. M. (2010). Logical-rule models of classification
response times: A synthesis of mental-architecture, random-walk, and decision-
bound approaches. Psychological Review, 117, 309-348.
Fific, M., Nosofsky, R. M., Townsend, J. T. (2008). Information-processing architectures
in multidimensional classification: A validation test of the systems factorial
technology. Journal of Experimental Psychology: Human Perception and
Performance, 34, 356-375.
Fific, M., & Townsend, J. T. (2003). Properties of visual search task on two items
revealed by the systems factorial methodology. Paper presented at the meeting
of the Society for Mathematical Psychology, Ogden, Utah.
Fill, J. A. & Machida, M. (2001). Stochastic monotonicity and realizable monotonicity.
The Annals of Probability, 29, 938-978.
Fisher, D. L. (1982). Limited channel models of automatic detection: Capacity and
scanning in visual search. Psychological Review, 89, 662-692.
Fisher, D. L. (1984). Central capacity limits in consistent mapping visual search tasks:
Four channels or more? Cognitive Psychology, 16, 449-484.
Fisher, D. L. (1985). Network models of reaction time: The generalized OP diagram. In
G. d’Ydewalle (Ed.), Cognition, Information Processing and Motivation
(Volume 3). Amsterdam: North-Holland Press, 229-254.
Fisher, D. L., Duffy, S. A., Young, C., & Pollatsek, A. (1988). Understanding the central
processing limit in consistent-mapping visual search tasks. Journal of
Experimental Psychology: Human Perception and Performance, 14, 253-266.
Fisher, D. L., & Glaser, R. A. (1996). Molar and latent models of cognitive slowing:
Implications for aging, dementia, depression, development, and intelligence.
Psychonomic Bulletin & Review, 4, 458-480.
Fisher, D. L., & Goldstein, W. M. (1983). Stochastic PERT networks as models of
cognition: Derivation of the mean, variance, and distribution of reaction time
using order-of-processing (OP) diagrams. Journal of Mathematical Psychology,
27, 121-151.
Fortin, C., Bedard, M. C., & Champagne, J. (2005). Timing during interruptions in
timing. Journal of Experimental Psychology: Human Perception and
Performance, 31, 276-288.
Fortin, C., Rousseau, R., Bourque, P., & Kirouac, E. (1993). Time estimation and
concurrent nontemporal processing: Specific interference from short-term-
memory demands. Perception & Psychophysics, 53, 536-548.
Garner, W. R. (1974). The processing of information and structure. NY: Wiley.
References 389
Gathercole, S. E., Frankish, C. R., Pickering, S. J., & Peaker, S. (1999). Phonotactic
influences on short-term memory. Journal of Experimental Psychology:
Learning, Memory and Cognition, 25, 84-95.
Gluck, M. A., & Bower, G. H. (1988). Evaluating an adaptive network model of human
learning. Journal of Memory and Language, 27, 166-195.
Goldstein, W. M., & Fisher, D. L. (1991). Stochastic networks as models of cognition:
Derivation of response time distributions using the order-of-processing method.
Journal of Mathematical Psychology, 35, 214-241.
Goldstein, W. M., & Fisher, D. L. (1992). Stochastic networks as models of cognition:
Deriving predictions for resource-constrained mental processing. Journal of
Mathematical Psychology, 36, 129-145.
Golumbic, M. C. (1980). Algorithmic graph theory and perfect graphs. NY: Academic
Press.
Gondan, M., & Röder, B. (2006). A new method for detecting interactions between the
senses in event-related potentials. Brain Research, 1073, 389-397.
Gray, W. D., John, B. E., & Atwood, M. E. (1993). Project Ernestine: Validating GOMS
for predicting and explaining real-world task performance. Human Computer
Interaction, 8(3), 237-309.
Greenwald, A. G. (1972). Doing two things at once: Time-sharing as a function of
ideomotor compatibility. Journal of Experiment Psychology, 94, 52-57.
Harris, J. R., Shaw, M. W. and Bates, M. (1979). Visual search in multicharacter arrays
with and without gaps. Perception and Psychophysics, 26, 69-84.
Haxby, J. V., Hoffman, E. A., Gobbini, M. I. (2000). The distributed human neural
system for face perception. Trends in Cognitive Sciences, 4, 223-233.
Hay, J. F., & Jacoby, L. L. (1996). Separating habit and recollection: Memory slips,
process dissociations, and probability matching. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 22, 1323-1335.
Hays, W. L. (1994). Statistics (5 th Ed.). Fort Worth: Harcourt, Brace College Publishers.
Heathcote, A., Brown, S., Wagenmakers, E. J., & Eidels, A. (2010). Distribution-free
tests of stochastic dominance for small samples. Journal of Mathematical
Psychology, 54, 454-463.
Henson, R. N., Rylands, A., Ross, E., Vuilleumeir, P., & Rugg, M. D. (2004). The effect
of repetition lag on electrophysiological and hemodynamic correlates of visual
object priming. Neuroimage, 21, 1674-1689.
Herman, L. M., & Kantowitz, B. H. (1970). The psychological refractory period: Only
half of the double stimulation story? Psychological Bulletin, 73, 74-88.
Hick, W. E., & Welford, A. T. (1956). Central inhibition: Some refractory observations.
Comment. Quarterly Journal of Experimental Psychology, 8, 39-41.
Hoffman, J. E. (1978). Search through a sequentially presented visual display. Perception
and Psychophysics, 23, 1-11.
Hoffman, J. E. (1979). A two-stage model of visual search. Perception and
Psychophysics, 25, 319-327.
390 References
Johnson, A., & Proctor, R. W. (2004). Attention: Theory and practice. Thousand Oaks,
Sage.
Johnston, J. C., McCann, R. S., & Remington, R. W. (1995). Chronometric evidence for
two types of attention. Psychological Science, 6, 365-369.
Johnston, J. C., & McCann, R. S. (2006). On the locus of dual-task interference: Is there a
bottleneck at the stimulus classification stage? Quarterly Journal of
Experimental Psychology, 59, 694-719.
Jolicoeur, P., & Dell’Acqua, R. (2000). Selective influence of second target exposure
duration and Task(1) load effects in the attentional blink phenomenon.
Psychonomic Bulletin & Review, 7, 472-479.
Kaerkes, R., & Mohring, R. H. (1978). Voresungen über Ordnungen und Netzplantheorie
(Lectures on orders and a theory of networks). Aachen, Germany: Technischen
Universität Aachen.
Karlin, L. & Kestenbaum, R. (1968). Effects of number of alternatives on the
psychological refractory period. Quarterly Journal of Experimental
Psychology, 20, 167-178.
Keele, S. W. (1973). Attention and human performance. Pacific Palisades, CA:
Goodyear.
Kelley, C. M., & Jacoby, L. L. (2000). Recollection and familiarity: Process-dissociation.
In E. Tulving & F. I. M. Craik (Eds.) The Oxford handbook of memory (pp.
215-228).
Kelley, J. E. & Walker, M. R. (1959). Critical path planning and scheduling. Proceedings
of the Eastern Joint Computer Conference (pp. 160-173). Boston, MA.
Keppel, G., & Underwood, B. J. (1962). Proactive inhibition in short-term retention of
single items. Journal of Verbal Learning and Verbal Behavior, 1, 153-161.
Kinchla, R. A. (1994). Comments on Batchelder and Riefer multinomial model for source
monitoring. Psychological Review, 101, 166-171.
Kirk, R. E. (1982). Experimental design: Procedures for the behavioral sciences (2nd
ed.). Montery, CA: Brooks/Cole.
Klauer, K. C. & Wegner, I. (1998). Unraveling social categorization in the “Who said
what” paradigm. Journal of Personality and Social Psychology, 75, 1155-1178.
Knapp, B. R. & Batchelder, W. H. (2004). Representing parametric order constraints in
multi-trial applications of multinomial processing tree models. Journal of
Mathematical Psycholology, 48, 215-229.
Kohfeld, D. L., Santee, J. L., & Wallace, N. D. (1981). Loudness and reaction time: II
Identification of detection components at different intensities and frequencies.
Perception & Psychophysics, 29, 550-562.
Kornblum, S., Hasbroucq, T., & Osman, A. (1990). Dimensional overlap: A cognitive
bias for stimulus-response compatibility--A model and a taxonomy.
Psychological Review, 97, 253-270.
Kotz, S., Balakrishnan, N., & Johnson, N. (2000). Continuous multivariate distributions:
Volume 1: Models and applications. NY: Wiley.
392 References
Kounios J., & Holcomb P. (1992). Structure and process in semantic memory: Evidence
from brain related potentials and reaction time. Journal of Experimental
Psychology.: General, 121, 459-479.
Kujala, J. V., & Dzhafarov, E. N. (2008). Testing for selectivity in the dependence of
random variables on external factors. Journal of Mathematical Psychology, 52,
128-144.
Kujala, J. V., & Dzhafarov, E. N. (2010). Erratum to “Testing for selectivity in the
dependence of random variables on external factors” [J. Math. Psych. 52 (2008)
128-144]. Journal of Mathematical Psychology, 54, 400.
Kulkarni, V. G. and Adlakha, V. G. (1986). Markov and Markov-regenerative PERT
networks. Operations Research, 34, 769-781.
Külpe, O. (1895). Outlines of psychology: Based upon the results of experimental
investigation. (E. B. Titchner, Trans.) London: Swan Sonnenschein & Co.
(Original 1893).
Kutas, M., & Hillyard, S. A. (1980). Reading senseless sentences: Brain potentials reflect
semantic incongruity. Science, 207, 203-205.
Lee, D. D., & Seung, H. S. (2001). Algorithms for nonnegative matrix factorization.
Advances in Neural Information Processing Systems 13: Proceedings of the
2000 Conference (pp. 556-562). Cambridge: MIT Press.
Li, X., Schweickert, R., & Gandour, J. (2000). The phonological similarity effect in
immediate recall: Positions of shared phonemes. Memory & Cognition, 28,
1116-1125.
Lien, M.-C., & Proctor, R. W. (2002). Stimulus-response compatibility and psychological
refractory period effects: Implications for response selection. Psychonomic
Bulletin & Review, 9, 212-238.
Lien, M.-C., Schweickert, R., & Proctor, R. W. (2003). Task switching and response
correspondence in the psychological refractory period paradigm. Journal of
Experimental Psychology: Human Perception and Performance, 29, 692-712.
Lin, A. (1999). Time to collision with two dimensional motion: Effects of horizontal and
vertical velocity. Unpublished doctoral dissertation, Purdue University, West
Lafayette, IN.
Liu, Y. (1996). Queueing network modeling of elementary mental processes.
Psychological Review, 103, 116-136.
Liu, Y. (2008). Queuing network modeling of the Psychological Refractory Period
(PRP). Psychological Review, 115, 913-954.
Liu, Y., Feyen, R., & Tsimhoni, O. (2006). Queueing Network-Model Human Processor
(QN-MHP): A computational architecture for multitask performance in human-
machine systems. ACM Transactions on Computer-Human Interaction, 13, 37-
70.
Liu, Y. S., Holmes, P., & Cohen, J. D. (2008). A neural network model of the Eriksen
task: Reduction, analysis, and data fitting. Neural Computation, 20, 345-373.
References 393
McLeod, P., Driver, J., & Crisp, J. (1988). Visual search for conjunctions of movement
and form in parallel. Nature, 332, 154-155.
Meyer, D. E., & Kieras, D. E. (1997a). A computational theory of executive cognitive
processes and multiple-task performance. Part 1. Basic mechanisms.
Psychological Review, 104, 3-65.
Meyer, D. E., & Kieras, D. E. (1997b). A computational theory of executive cognitive
processes and multiple-task performance. Part 2. Accounts of psychological
refractory period phenomena. Psychological Review, 104, 749-791.
Micro Analysis and Design (1985). Micro SAINT [Computer program]. Concord, MA:
MGA, Inc.
Miller, J. O. (1982). Divided attention: Evidence for coactivation with redundant signals.
Cognitive Psychology, 14, 247–279.
Miller, J. (1982). Discrete versus continuous models of human information processing: In
search of partial output. Journal of Experimental Psychology:Human
Perception and Performance, 8, 273-296.
Miller, J. (1988). Discrete and continuous models of human information processing:
Theoretical distinctions and empirical results. Acta Psychologica, 67, 191-257.
Miller, J. O. (1993). A queue-series model for reaction time, with discrete-stage and
continuous flow models as special cases. Psychological Review, 100, 702-715.
Miller, J., Ulrich, R., & Rolke, B. (2009). On the optimality of serial and parallel
processing in the psychological refractory period paradigm: Effects of the
distribution of stimulus onset asynchronies. Cognitive Psychology, 58, 273-310.
Mochs, J. (1988). Decomposing event-related potentials: A new topographic components
model. Biological Psychology, 26, 119-215.
Molenaar, P. C. M., & van der Molen, M. W. (1986). Steps to a formal analysis of the
cognitive-energetic model of stress and human performance. Acta
Psychologica, 62, 237-261.
Müller, A., & Stoyan, D. (2002). Comparison methods for stochastic models and risks.
NY: Wiley.
Nairne, J. S., & Kelley, M. R. (1999). Reversing the phonological similarity effect.
Memory and Cognition, 27, 45-53.
Nairne, J. S., Neath, I., & Serra, M. (1997). Proactive interference plays a role in the
word-length effect. Psychonomic Bulletin & Review, 4, 541-545.
Nakayama, K., & Silverman, G. H. (1986). Serial and parallel processing of visual
feature conjunctions. Nature, 320, 264-265.
Navon, D. & Miller, J. (1987). Role of outcome conflict in dual task interference.
Journal of Experimental Psychology: Human Perception and Performance, 13,
435-448.
Nosofsky, R. M., & Palmeri, T. J. (1997). An exemplar-based random walk model of
speeded classification. Psychological Review, 104, 266-300.
Ollman, R. T. (1968). Central refractoriness in simple reaction time: The deferred
processing model. Journal of Mathematical Psychology, 5, 49-60.
References 395
Ollman, R. (1990). The matrix product model and the method of additive factors.
Unpublished manuscript.
Oriet C., & Jolicoeur, P. (2003). Absence of perceptual processing during reconfiguration
of task set. Journal of Experimental Psychology: Human Perception and
Performance, 29, 1036-1049.
Osman, A., & Moore, C. M. (1993). The locus of dual-task interference: Psychological
refractory effects on movement-related brain potentials. Journal of
Experimental Psychology: Human Perception and Performance, 19, 1292-
1312.
Pashler, H. (1984). Processing stages in overlapping tasks: Evidence for a central
bottleneck. Journal of Experimental Psychology: Human Perception and
Performance, 10, 358-377.
Pashler, H. (1987). Detecting conjunctions of color and form: Reassessing the serial
search hypothesis. Perception & Psychophysics, 41, 191-201.
Pashler, H. (1994). Dual-task interference in simple tasks: Data and theory.
Psychological Bulletin, 116, 220-244.
Pashler, H., & Badgio, P. (1985). Visual attention and stimulus identification. Journal of
Experimental Psychology: Human Perception and Performance, 11, 105-121.
Pashler, H., & Johnston, J. C. (1989). Chronometric evidence for central postponement in
temporally overlapping tasks. Quarterly Journal of Experimental Psychology,
41A, 19-45.
Peterson, L. R., & Peterson, M. J. (1959). Short-term retention of individual verbal items.
Journal of Experimental Psychology, 58, 193-198.
Phaf, R. H., Van der Heijden, A. H. C. and Hudson, P. T. W. (1990). SLAM: A
connectionist model for attention in visual selection tasks. Cognitive
Psychology, 22, 273-341.
Poirier, M., & Saint-Aubin, J. (1995). Memory for related and unrelated words: Further
evidence on the Influence of semantic factors in immediate serial recall.
Quarterly Journal of Experimental Psychology, 48A, 384-404.
Poirier, M., Schweickert, R., & Oliver, J. (2005). Silent reading rate and memory span.
Memory, 13,380-387.
Prinzmetal, W., Ivry, R. B., Beck, D., & Shimizu, N. (2002). A measurement theory of
illusory conjunctions. Journal of Experimental Psychology: Human Perception
and Performance, 28, 251-269.
Raab, D. (1962). Statistical facilitation of simple reaction times. Transactions of the New
York Academy of Sciences, 24, 574-590.
Ratcliff, R. (1978). A theory of memory retrieval. Psychological Review, 88, 59-108.
Ratcliff, R., Van Zandt, T., & McKoon, G. (1999). Connectionist and diffusion models of
reaction time. Psychological Review, 106, 261-300.
Reynolds, D. (1964). Effects of double stimulation: Temporary inhibition of response.
Psychological Bulletin, 62, 333-347.
396 References
Riefer, D. M., & Batchelder, W. H. (1988). Multinomial modeling and the measurement
of cognitive processes. Psychological Review, 95, 318-339.
Riefer, D. M., & Batchelder, W. H. (1995). A multimomial modeling analysis of the
recognition-failure paradigm. Memory & Cognition, 23, 611-630.
Riefer, D. M., Hu, X. G., & Batchelder, W. H. (1994). Response strategies in source
monitoring. Journal of Experimental Psychology: Learning, Memory and
Cognition, 20, 680-693.
Rinkenauer, G., Ulrich, R., & Wing, A. M. (2001). Brief bimanual force pulses:
Correlations between the hands in force and time. Journal of Experimental
Psychology: Human Perception & Performance, 27, 1485-1497.
Roberts, S. (1987). Evidence for distinct serial processes in animals: The multiplicative-
factors method. Animal Learning & Behavior, 15, 135-173.
Roberts, S., & Sternberg, S. (1993). The meaning of additive reaction-time effects: Tests
of three alternatives. In S. Kornblum & D. E. Meyer (Eds.), Attention and
performance XIV: Synergies in experimental psychology, artificial intelligence,
and cognitive neuroscience--A silver jubilee 9 (pp. 611-654). Cambridge, MA:
MIT Press.
Roodenrys, S. (2009). Explaining phonological neighborhood effects in short-term
memory. In A. Thorn& M. Page (Eds.) (2009). Interactions between short-term
and long-term memory in the verbal domain (pp. 177-198). Hove: Psychology
Press.
Rouse, W. B. (1980). Systems engineering models of human-machine interaction. New
York: North Holland.
Royden, H. L. (1968). Real analysis (2nd ed.). London: Macmillan.
Rumelhart, D. E. (1970). A multicomponent theory of the perception of briefly exposed
visual displays. Journal of Mathematical Psychology, 7, 191-218.
Rumelhart, D. E., & McClelland, J. L. (Eds.) (1986). Parallel distributed processing
(Vol. 1). Cambridge, MA: MIT Press.
Ruthruff, E., Miller, J., & Lachmann, T. (1995). Does mental rotation require central
mechanisms? Journal of Experimental Psychology: Human Perception and
Performance, 21, 552-570.
Sanders, A. F. (1990). Issues and trends in the debate on discrete versus continuous
processing of information. Acta Psychologica, 74, 123-167.
SAS Institute, Inc. (1985). User’s guide: Statistics. Cary, NC: SAS Institute, Inc.
Schmidt, R. A., Zelaznik, H., Hawkins, B., Frank, J. S., & Quinn, J. T. (1979). Motor-
output variability: A theory for the accuracy of rapid motor acts. Psychological
Review, 86, 415-451.
Schneider, W., & Shiffrin, R. M. (1977). Controlled and automatic human information
processing: I. Detection, search, and attention. Psychological Review, 84, 1-66.
Schuberth, R. E., Spoehr, K. T., & Lane, D. M. (1981). Effects of stimulus and contextual
information on the lexical decision process. Memory & Cognition, 9, 68-77.
References 397
Thorn, A. & Page, M. (Eds.) (2009). Interactions between short-term and long-term
memory in the verbal domain. Hove: Psychology Press.
Thornton, T. L., & Gilden, D. L. (2007). Parallel and serial processes in visual search.
Psychological Review, 114, 71-103.
Tombu, M., & Jolicoeur, P. (2003). A central capacity sharing model of dual-task
performance. Journal of Experimental Psychology: Human Perception &
Performance, 29, 3-18.
Townsend, J. T. (1971). A note on the identification of parallel and serial processes.
Perception & Psychophysics, 10, 161-163.
Townsend, J. T. (1972). Some results concerning the identifiability of parallel and serial
processes. British Journal of Mathematical and Statistical Psychology, 25, 168-
197.
Townsend, J. T. (1974). Issues and models concerning the processing of a finite number
of inputs. In B. H. Kantowitz (Ed.), Human information processing: Tutorials
in performance and cognition. Hillsdale, NJ: Erlbaum, pp. 133-168.
Townsend, J. T. (1984). Uncovering mental processes with factorial experiments. Journal
of Mathematical Psychology, 28, 363-400.
Townsend, J. T. (1990). Truth and consequences of ordinal differences in statistical
distributions: Toward a theory of hierarchical inference. Psychological Bulletin,
108, 551-567.
Townsend, J. T., & Ashby, F. G. (1983). Stochastic modeling of elementary
psychological processes. Cambridge: Cambridge University Press.
Townsend, J. T., & Fific, M. (2004). Parallel versus serial processing and individual
differences in high-speed search in human memory. Perception &
Psychophysics, 66, 953-962.
Townsend, J. T., & Nozawa, G. (1995). Spatio-temporal properties of elementary
perception: An investigation of parallel, serial, and coactive theories. Journal of
Mathematical Psychology, 39, 321-359.
Townsend, J. T., & Schweickert, R. (1989). Toward the trichotomy method of reaction
times: Laying the foundation of stochastic mental networks. Journal of
Mathematical Psychology, 33, 309-327.
Townsend, J. T., & Thomas, R. D. (1994). Stochastic dependencies in parallel and serial
models: Effects on systems factorial interactions. Journal of Mathematical
Psychology, 38, 1-34.
Townsend, J. T., & Wenger, M. J. (2004). A theory of interactive parallel processing:
New capacity measures and predictions for a response time inequality series.
Psychological Review, 111, 1003-1035.
Treisman, A. M., & Gelade, G. (1980). A feature-integration theory of attention.
Cognitive Psychology, 12, 97-136.
Treisman, A., & Sato, S. (1990). Conjunction search revisited. Journal of Experimental
Psychology: Human Perception and Performance, 16, 459-478.
Tsao, D. (2006). A dedicated system for processing faces, Science, 314(5796). pp. 72-73.
References 401
Ulrich, R., Fernández, S. R., Jentzsch, I., Rolke, B., Schröter, H., & Leuthold, H. (2006).
Motor limitation in dual-task processing under ballistic movement conditions.
Psychological Science, 17, 788-793.
Ulirch, R., & Miller, J. (1997). Tests of race models for reaction time in experiments with
asynchronous redundant signals. Journal of Mathematical Psychology, 41, 367-
381.
Ulrich, R. & Miller, J. (2008). Response grouping in the psychological refractory period
(PRP) paradigm: Models and contamination effects. Cognitive Psychology, 57,
75-121.
Ulrich, R. & Wing, A. M. (1991). A recruitment theory of force-time relations in the
production of brief force pulses: The parallel force unit model. Psychological
Review, 98, 268-294.
Valls, V., Laguna, M., Lino, P., Pérez, A., & Quintanilla, S. (1998). Project scheduling
with stochastic activity interruptions. In J. Weglarez (Ed.), Recent advances in
project scheduling (pp. 333-354). Boston: Kluwer Academic Publishers.
van Lankveld, J. J. D. M., & Smulders, F. T. Y. (2008). The effect of visual sexual
content on the event-related potential. Biological Psychology, 79, 200-208.
Van Selst, M. & Jolicoeur, P. (1994). Can mental rotation occur before the dual-task
bottleneck? Journal of Experimental Psychology: Human Perception and
Performance, 20, 905-921.
Van Selst, M., & Jolicoeur, P. (1997). Decision and response in dual-task interference.
Cognitive Psychology, 33, 266-307.
Van Zandt, T. (2002). Analysis of response time distributions. In H. Pashler & J. T.
Wixted (Eds.), Stevens’ handbook of experimental psychology: Vol. 4.
Methodology in experimental psychology (3rd ed., pp. 461-516). NY: Wiley.
Van Zandt, T., & Ratcliff, R. (1995). Statistical mimicking of reaction-time data: Single
process models, parameter variability, and mixtures. Psychonomic Bulletin &
Review, 2, 20-54.
Van Zandt, T., & Townsend, J. T. (1993). Self-terminating versus exhaustive processes in
rapid visual and memory search: An evaluative review. Perception &
Psychophysics, 53, 563–580.
Vorberg, D. & Schwarz, W. (1988). Network models of reaction times. Paper presented at
the XXIV International Congress of Psychology, Sidney, Australia.
Voss, A., Rothermund, K., & Voss, J. (2004). Interpreting the parameters of the diffusion
model: An empirical validation. Memory & Cognition, 32, 1206-1220.
Ward, R., & McClelland, J. L. (1989). Conjunctive search for one and two identical
targets. Journal of Experimental Psychology: Human Perception and
Performance, 15, 664-672.
Waugh, N. C., & Norman, D. A. (1965). Primary memory. Psychological Review, 72, 89-
104.
Welford, A. T. (1952). The ‘psychological refractory period’ and the timing of high-speed
performance--a review and a theory. British Journal of Psychology, 43, 2-19.
402 References
403
404 Author Index
Simson, R., 348 Ulrich, R., 97, 106, 111, 133, 135, 137,
Slotnick, S. D., 19, 300 138, 139, 140, 141, 142, 143, 144,
Smith, M. C., 101, 103, 105, 268, 269, 145, 146, 191, 198, 271
270, 271, 287 Underwood, B. J., 305, 307, 310, 316,
Smith, R. E., 302 355
Smulders, F. T. Y., 349, 350 Valls, V., 202
Spoehr, K. T., 190, 296 Van der Heijden, A. H. C., 274
Stahl, C., 19 van der Molen, M. W., 33
Steffens, M. C., 301 Van Lankveld, J. J. D. M., 349, 350
Sternberg, S., 3, 4, 5, 6, 8, 20, 27, 34, van Santen, J. P. H., 336, 341, 358
58, 153, 154, 155, 156, 157, 158, Van Selst, M., 8, 101, 102, 111, 113
159, 160, 164, 165, 166, 167, 185, Van Zandt, T., 20, 164, 165, 213, 235,
189, 190, 213, 214, 222, 226, 227, 243, 274
234, 238, 243, 259, 291, 293, 294, Vaughan Jr., H. G., 348
359 Vorberg, D., 33, 195, 198
Stickgold, R., 328 Voss, A., 292
Stoyan, D., 27, 68, 69 Voss, J., 292
Stroop, J. R., 123, 124 Vuilleumeir, P., 352
Stuart, G., 317, 326, 327 Wagenmakers, E. J., 170
Stuart, G. P., 328 Walker, M. R., 11
Sung, K., 20, 176, 178, 187, 191, 196, Wallace, N. D., 34, 259
212, 214, 215, 229, 230, 231, 232, Wang, W.-C., 301
234, 239, 251 Wang, Z., 28, 31, 38, 40, 42, 45, 52, 54,
Teder-Sälejärvi, W. A., 347 62, 92
Tehan, G., 308, 309, 310, 311, 312, Waugh, N. C., 303
326, 356, 357 Wegener, I., 15
Tein, J.-Y., 235 Welford, A. T., 7, 8, 20, 94, 95, 97,
Telford, C. W., 6, 93 127, 134, 146, 191
Thomas, R. D., 190, 292, 293, 360 Wenger, M. J., 191, 194, 195, 198, 244,
Thompson-Schill, S. L., 328 252, 253, 254, 255, 283
Thorn, A., 304 West, A. N., 239
Thorn, A. S. C., 304, 317 Wickens, C. D., 282
Thornton, T. L., 229 Wiggs, C. L., 352
Tombu, M., 97, 191 Williams, R. J., 276, 277
Toth, J. P., 17, 300, 301 Wing, A. M., 141
Townsend, J. T., 5, 8, 9, 26, 27, 28, 34, Wolfe, J. M., 224, 225, 228, 229
37, 42, 44, 45, 50, 51, 60, 68, 75, Wu, C. X., 9, 94, 280
84, 152, 153, 165, 167, 168, 169, Wundt, W., 1
170, 172, 174, 190, 191, 194, 195, Xi, Z., 330, 336, 337, 338, 358
197, 198, 200, 201, 213, 214, 226, Yonelinas, A. P., 17, 301
228, 229, 235, 236, 238, 243, 244, Young, C., 280
246, 249, 256, 259, 283, 359, 360, Yu, J. C., 302
366, 367 Zelaznik, H., 140
Treisman, A., 224, 225, 229, 230, 282
Tsao, D., 252
Tsimhoni, O., 94
Turcotte, J., 308, 309, 310, 311, 312,
326, 356
This page intentionally left blank
Subject Index
409
410 Subject Index