1999, 71, 355–373
JOURNAL OF THE EXPERIMENTAL ANALYSIS OF BEHAVIOR
NUMBER
3 (MAY)
CHOICE, CONTINGENCY DISCRIMINATION, AND
FORAGING THEORY
W ILLIAM M. B AUM , J ED W. S CHWENDIMAN ,
K ENNETH E. B ELL
AND
UNIVERSIT Y OF NEW HAMPSHIRE
Four pigeons were trained on eight or nine pairs of independent concurrent variable-interval schedules. The range of reinforcement ratios included extreme ratios (up to 532 to 1). Large samples of
stable performance were gathered. Contrary to the findings of Davison and Jones (1995), the generalized matching law described choice more accurately than a contingency-discriminability model.
Taking small samples (5 to 10 sessions) and applying a more liberal stability criterion used by Davison
and Jones only increased the unsystematic variance in the data and in estimates of generalizedmatching-law sensitivity. Because changing to dependent scheduling and inserting a changeover delay
had no systematic effect, the deviations from generalized matching reported by Davison and Jones
probably arose from imperfectly discriminated stimuli. Analysis of visits revealed that visits to the
nonpreferred alternative were brief and approximately constant. When choice between the preferred
(rich) and nonpreferred (lean) alternatives, regardless of position, was analyzed according to the
generalized matching law, sensitivities approximated 1.0, with bias in favor of the lean alternative.
This bias, which arose from an excessive frequency of visits to the lean alternative, explains undermatching as the result of fitting one line to a choice relation that consists of two displaced lines,
both with a slope of 1.0. The pattern of deviation from the generalized matching line confirmed
this account. The findings suggest an alternative analysis of choice that focuses on probability of
visiting the lean alternative as the dependent variable. This probability was directly proportional to
ratio of reinforcement. Matching, undermatching, and overmatching may all be explained by a view
of concurrent performance based on foraging theory, in which responding occurs primarily at the
rich alternative and is occasionally interrupted by brief visits to the lean alternative.
Key words: choice, generalized matching law, contingency-discriminability model, foraging theory,
visit duration, key peck, pigeons
Choice has been studied in the experimental analysis of behavior with concurrent
schedules of reinforcement. The matching
law (Herrnstein, 1961) was first proposed to
describe performance on such schedules.
The matching relation equates the proportion of responses at an alternative with the
proportion of reinforcement obtained from
that alternative:
B1
r1
5
,
B1 1 B 2
r1 1 r 2
(1)
where B 1 and B 2 are the rates of responding
and r1 and r2 are the rates of reinforcement
obtained from Alternatives 1 and 2. This
equation was later put into the algebraically
Some of this research was completed in partial fulfillment of the requirements for Jed Schwendiman’s master’s degree in psychology. We thank Suzanne Mitchell,
William Stine, Randolph Grace, John Nevin, and the University of New Hampshire Behavioral Research Group for
helpful comments.
Address correspondence to William M. Baum at the
University of New Hampshire, Department of Psychology,
Durham, New Hampshire 03824-3567 (E-mail: wm.baum@
unh.edu).
equivalent ratio form (Baum & Rachlin,
1969):
B1
r
5 1.
B2
r2
(2)
As an attempt to explain systematic deviations
from the ‘‘strict’’ matching of Equation 2,
Baum (1974b) offered the generalized
matching law:
s
B1
r
5b 1 .
(3)
B2
r2
12
The parameter b describes bias, or preference
due to differences in variables other than r1
and r2. When b equals 1.0 there is no bias. If
b is greater or less than 1.0, there is a bias
toward Alternative 1 or 2. The parameter s
has been interpreted as sensitivity to the reinforcement ratio (Lobb & Davison, 1975).
When s falls short of 1.0, a result called undermatching, choice tends to be more indifferent than predicted by strict matching,
whereas when s exceeds 1.0, the result called
overmatching, choice is more extreme than
predicted by strict matching (Baum, 1979).
355
356
WILLIAM M. BAUM et al.
Performance under concurrent pairs of
variable-interval (VI) schedules has commonly been described by the generalized matching law in its logarithmic form (for reviews,
see Baum, 1979; Davison & McCarthy, 1988;
Wearden & Burgess, 1982), which is linear:
log
1B 2 5 s·log 1r 2 1 log b,
B1
r1
2
2
(4)
where log b is the intercept and sensitivity s
is the slope. Within subjects, sensitivity is assumed to be constant across ratios of reinforcement.
For concurrent VI schedules, the value of
s has often been found to deviate from 1.0
(Baum, 1979; Wearden & Burgess, 1982). In
proposing Equations 3 and 4, Baum (1974b)
speculated that several factors might affect
sensitivity, such as penalties for switching between alternatives and failure of discrimination between alternatives. Although undermatching occurs frequently, its origins
remain a puzzle. Penalties on switching, however, are well known to increase sensitivity,
even to overmatching (Baum, 1975, 1982;
Boelens & Kop, 1983).
Several mathematical models have been
proposed to explain matching. For example,
Shimp (1966) proposed momentary maximizing, Myerson and Miezin (1980) proposed
a kinetic model, Herrnstein and Vaughan
(1980) proposed melioration, Hinson and
Staddon (1983) proposed hill climbing, and
Rachlin (1978) proposed global optimization. Although all of the theories that have
been proposed explain matching, few explain
the frequently observed deviations—undermatching and overmatching.
One attempt to explain undermatching
was put forward by Davison and Jenkins
(1985). Elaborating on the possibility of failure of discrimination, they proposed that undermatching might be explained by misallocation of reinforcers between alternatives
(but see Wearden, 1983, for an alternative interpretation). They argued that failure of discrimination between the stimuli associated
with the different alternatives would result in
some of each alternative’s reinforcers being
misallocated to the other. An experiment by
Miller, Saunders, and Bourland (1980) lent
support to this idea by showing that in a single-key procedure (Findley, 1958) response
distribution depended on the disparity between the stimuli associated with the different
component schedules and that undermatching increased as stimulus disparity decreased.
Hoping to explain undermatching and replace the generalized matching law, Davison
and Jenkins (1985) proposed the contingency-discriminability model, which may be written
1
2
B1
r 2 pr1 1 pr2
5b 1
.
B2
r2 2 pr2 1 pr1
(5)
Equation 5 includes two free parameters, b
and p. The parameter b describes bias and is
identical to parameter b of Equation 3. The
parameter p replaces s in Equation 3. Whereas s can be thought of as sensitivity to the
ratio of reinforcement from the two sources,
p is interpreted as proportional confusion between the two sources of reinforcement (Davison & Jones, 1995). It incorporates the idea
that whenever the stimuli associated with
schedules of reinforcement are imperfectly
discriminable, some proportion of the reinforcers obtained from each schedule are associated with the alternative schedule and its
discriminative stimulus. With two concurrent
schedules, p can range from 0 to 0.5. When
b equals 1.0 and p is zero (indicating no confusion between alternatives) the contingencydiscriminability model simplifies to strict
matching (Equation 2). When p equals 0.5,
the alternatives are perfectly confused, and
the equation predicts no change in choice as
reinforcement changes. Because p is a proportion, values of p between 0 and 0.5 dictate
that, as the ratio of reinforcement becomes
more extreme, more and more reinforcers
obtained from the rich schedule are associated with the lean schedule. Hence, the curve
defined by Equation 5 appears S-shaped in
logarithmic coordinates, but for moderate
values of p, the curve is approximately linear
over the range of reinforcement ratios usually
studied (between 10 and 0.1). Thus, the predictions of the contingency-discriminability
model and the generalized matching law differ significantly only when the reinforcement
ratio becomes more extreme than usual (e.g.,
ratios greater than 10:1 or less than 1:10).
Davison and Jones (1995) set out to test the
contingency-discriminability model by studying choice with reinforcement ratios that var-
CHOICE AND FORAGING
ied over a wider range than usual: up to
about 100:1 and down to about 1:100. They
found that the generalized matching law
(Equation 4) accounted for about 98% of the
variance in choice and that the contingencydiscriminability model (Equation 5) accounted for about 99% of the variance. They argued that the difference was significant after
fitting the generalized matching law both
with and without the most extreme ratios of
reinforcement. When the extremes were
eliminated the slope (s in Equation 4) increased, indicating that choice fell short at
extreme reinforcement ratios, as the contingency-discriminability model predicted.
Their results, however, may have arisen
from specific aspects of the procedures that
they used. In particular, Davison and Jones
(1995) used dependent scheduling (Stubbs &
Pliskoff, 1969), in which the set-up of a reinforcer for one alternative causes the schedules for both alternatives to stop until the setup reinforcer is obtained, and a 3-s
changeover delay (COD), which blocked reinforcement after a switch of schedules until
at least 3 s had elapsed. Both of these would
tend to add responses to the lean alternative,
reducing relative responding at the rich alternative, particularly at the extremes.
More important, Davison and Jones (1995)
used unusually confusable stimuli. They arranged a single-key procedure (Findley,
1958), in which the two schedules were associated with two different brightness levels
on the main key. More typical procedures associate the schedules with highly distinctive
stimuli, such as color or location. The present
experiment made a more conservative test of
the generalized matching law by associating
the two schedules with two separate keys, a
more typical method of presenting concurrent schedules.
Finally, Davison and Jones (1995) used a
stability criterion that may have been too liberal and samples of stable performance that
may have been too small. When reinforcement ratios reach 100:1 or more, several sessions can pass without a reinforcer on the
lean alternative. Their stability criterion allowed conditions to change when only 10 or
fewer reinforcers had been obtained at the
lean alternative in total. They then used the
last five sessions as their sample, with the result that their samples from the extreme con-
357
ditions included fewer than five reinforcers
on the lean alternative. Small samples might
mean just that estimates of response and reinforcer ratios would be more variable at the
extremes; if so, the deviations observed by
Davison and Jones might simply have been
due to chance.
Davison and Jones (1995) claimed that,
whereas the generalized matching law offers
only a description, the contingency-discriminability model offers an explanation, in the
sense that it derives undermatching from assumptions about basic principles (e.g., confusion). Equation 5 has the limitation, however, that it only predicts undermatching. It
cannot explain overmatching because, for
that, p would have to take on negative values,
an outcome with no meaning.
Houston and McNamara (1981; Houston,
McNamara, & Sumida, 1987) proposed a
model that has the potential to predict the
full range of variation in sensitivity, from undermatching to overmatching. It derives from
foraging theory, in which foraging and choice
are seen as composed of successive visits to
alternative sites where resources may be obtained. Accordingly, concurrent performance
would consist of visits, now to one schedule,
now to the other. Houston and McNamara
derived an optimal performance that consists
of responding primarily at the rich alternative
interrupted by brief visits to the lean alternative. At low rates of switching—the optimum—this predicts overmatching. Higher
rates of switching, though suboptimal, would
lead to reductions in sensitivity, first to matching and then to undermatching.
Casting concurrent performance in terms
of rich and lean alternatives rather than in
terms of locations or stimuli, foraging theory
points to a different sort of analysis of choice.
Although bias in Equations 4 and 5 is usually
taken to indicate favoritism toward one position or stimulus, independent of which is rich
or lean, it is possible to think of another sort
of bias: favoritism toward the rich or lean alternative, independent of position or stimulus. A bias in favor of the lean alternative
would mean too frequent visiting of the lean
alternative and would appear as undermatching in a fit of the generalized matching law.
A bias in favor of the rich alternative would
appear as overmatching.
Depending on availability of resources, cost
358
WILLIAM M. BAUM et al.
of switching from one resource site to another, and the way visits are aggregated, a foraging model may predict the whole range of
observed outcomes when choice, measured
as relative aggregated visits, is considered as
a function of relative resources. Departing
from generalized matching, however, it suggests other analyses that focus on visits to the
lean alternative. Some research favors such a
possibility (Baum, 1982; Baum & Aparicio,
1999; Buckner, Green, & Myerson, 1993).
Particularly when choice and resource ratios
vary over a wide range, from near indifference to extremes, it seems possible to put a
model like this to the test.
The present experiment used a conventional set of procedures to study choice over
an extremely wide range. It had two aims: (a)
to compare the predictions of the generalized matching law and the contingency-discriminability model, and (b) to test the validity and implications of the Houston-McNamara
view of concurrent performance as composed
of visits to sources of reinforcement.
METHOD
Subjects
Four White Carneau pigeons, numbered
26, 27, 299, 973, participated. Pigeon 299 was
maintained at 85% of free-feeding weight
615 g and was experimentally naive at the
beginning of the experiment. The other 3
birds were maintained at 80% of free-feeding
weights (615 g), and all had previous experience with a variety of concurrent schedules.
After each daily session the birds were
weighed and, if necessar y, were fed an
amount of pigeon chow sufficient to maintain
their designated body weights. Water and grit
were constantly available in the home cages.
Apparatus
Four typical three-key operant conditioning chambers (35 cm deep by 35 cm wide by
35 cm high) were used. The circular keys (2.5
cm diameter) were aligned horizontally 26
cm above the floor and were transilluminated
with white light. Each chamber was equipped
with a houselight 7 cm above the center key,
and an aperture (6 cm by 5 cm) that allowed
access to a grain magazine was located 13 cm
below the center key. The magazine was illuminated when operated, at which time wheat
was available. A force of approximately 0.10
N was required to operate a key. Operation
resulted in an audible feedback click. Each
chamber was enclosed in a sound-attenuating
box and was fitted with a ventilating fan,
which helped to mask extraneous noises.
Event scheduling and data recording were
controlled by a microcomputer located in the
next room.
Procedure
Pigeon 299 was gradually deprived of food
before being trained to eat from the food
magazine. An autoshaping procedure was
then used to train the bird to peck keys transilluminated by the white light. When it was
pecking reliably, it was trained on a Fixed-Ratio 2 schedule for two 30-min sessions. Then
the bird was trained with a VI 30-s schedule
on the center key before concurrent VI 30-s
schedules on the left and right keys were introduced. All birds were exposed to concurrent VI 30 VI 30 for several days before Condition 1. All VI schedules were random-interval
schedules with a 0.5-s time base and differed
only in the probability of setting up reinforcement.
During the session both the left and right
keys were transilluminated with white lights
of approximately equal intensity. Pecks to the
keys were reinforced on independent concurrent VI schedules. The overall programmed reinforcement rate was held constant at 2 per minute. During reinforcement
the keylights were darkened and the food
magazine was raised and illuminated. Reinforcement duration was slowly adjusted separately for each bird during the first and second conditions from an initial duration of
2.00 s so that postsession feeding could be
kept at a minimum. Pigeons 26, 27, 299, and
973 were allowed to eat for 2.39, 2.00, 2.50,
and 2.07 s, respectively. Sessions commenced
in blackout and lasted until 80 reinforcers
had been obtained (usually about 45 min). If
a reinforcer had become available on one of
the schedules but was uncollected at the end
of the session, the next day’s session began
with a reinforcer available on that schedule.
No changeover delay was used. This was
done to avoid forcing extra responses to the
lean alternative and to avoid ambiguity about
the status of responses made during the
COD. This ambiguity arises because the COD
359
CHOICE AND FORAGING
Table 1
The order of conditions, scheduled reinforcement ratios, and sessions per condition for each
subject.
Pigeon 26
Pigeon 27
Order
Scheduled
ratio
Sessions
1
2
3
4
5
6
7
8
9
10
11
12
13
14
4:1
1:9
64:1
1:32
128:1
1:256
4:1
1:8
64:1
1:256 d
128:1 d
1:256 d,c
128:1 d,c
1:256 d,c
27
29
37
44
44
90
47
27
50
77
66
53
58
78
Scheduled
ratio
1:4
9:1
1:64
32:1
1:128
256:1
1:4
8:1
1:128 d
256:1 d
1:128 d,c
256:1 d,c
Pigeon 299
Pigeon 973
Sessions
Scheduled
ratio
Sessions
27
36
34
40
50
147
49
44
67
91
75
69
1:4
1:9
64:1
1:32
128:1
1:256
4:1
1:8
128:1 d
1:256 d
128:1 d,c
1:256 d,c
27
24
46
29
55
137
47
54
54
111
66
78
Scheduled
ratio
1:4
9:1
1:64
32:1
1:128
256:1
1:4
8:1
1:128 d
256:1 d
1:128 d,c
256:1 d,c
Sessions
27
24
42
36
65
144
52
54
79
62
75
69
Note. A d indicates dependent scheduling; a c indicates the use of a 3-s COD.
is discriminated: Responding during the
COD occurs at a high rate that differs little
between alternatives (Baum, 1974a). As a result, responding during the COD exhibits extreme undermatching. Because post-COD responding tends toward overmatching, the
common practice of combining the two as if
they were equivalent produces a result somewhere in between, often approximating
matching (Baum, 1974a). It is unclear, however, that they are equivalent. It may be argued that the COD resembles travel between
alternatives and should be omitted from calculations of choice (Baum, 1982). These considerations, combined with findings that
matching may be obtained without a COD
(e.g., Heyman, 1979; Shull & Pliskoff, 1967),
favored omission of the COD.
Table 1 shows the order of conditions,
scheduled ratio of reinforcement, and total
number of sessions in each condition for
each bird. Each condition remained in effect
for a pigeon until the logarithm of the behavior ratio appeared stable across sessions by
visual inspection and a minimum of 10 reinforcers had been obtained from the lean
schedule over the period of stable performance. In practice, this required a minimum
of 24 and a maximum of 147 sessions. With
one exception (Pigeon 299, Condition 2),
when conditions were changed, the positions
of the rich and lean schedules were always
reversed. Except for Pigeon 299, the first condition was replicated for every bird.
Following the initial set of eight or nine
conditions, some additional extreme conditions were conducted for each bird, two or
three with dependent scheduling and then
two or three with dependent scheduling and
a 3-s COD.
RESULTS
To compare the contingency-discriminability model with the generalized matching law,
three separate analyses of choice were performed for each bird. The first used the data
from the longest period of stability, as determined by visual inspection. The numbers of
responses, reinforcers, and changeovers were
summed across sessions. These numbers,
along with the obtained ratios of reinforcement for each condition and pigeon, appear
in the Appendix. The second analysis used
sums of responses and reinforcers from just
the last five sessions, as did the analysis done
by Davison and Jones (1995). In a few conditions, when the last five sessions included
no reinforcer on the lean side, up to 3 more
days were included to provide a calculable reinforcement ratio. Because all the conditions
in the present experiment were continued
beyond the number of sessions at which they
would have terminated according to the sta-
360
WILLIAM M. BAUM et al.
Fig. 1. Choice relations from data summed over the longest period of stability. Log2 response ratios for each
subject are plotted as a function of log2 obtained reinforcer ratios. The conditions with independent scheduling and
no COD (Conditions 1 to 9 for Pigeon 26 and Conditions 1 to 8 for the other 3) are indicated by triangles. The
conditions with dependent scheduling and no COD are indicated by squares. Circles indicate the conditions with
dependent scheduling and a COD. The solid straight lines are the best fits of the triangles to Equation 4 by least
squares regression. The equation and goodness of fit (r2) are shown in each graph. The dashed curves show the best
fits to Equation 5. The parameters of the fits are shown in Table 2.
bility criterion used by Davison and Jones, the
third analysis attempted to assess the possible
effects, had that stability criterion been applied to this experiment. Their criterion was
used to pick the five sessions that would have
been the last five sessions had that stability
criterion been used to terminate conditions
in this experiment. In a few conditions, up to
five earlier sessions’ data were included, as in
the second analysis, to provide a calculable
reinforcement ratio.
Figure 1 shows the logarithms (base 2) of
the response ratios plotted as a function of
the logarithms (base 2) of the obtained reinforcement ratios for each subject. Base 2
logarithms were used because, with them, a
log unit represents a doubling or halving of
ratio, and this allows distinctions among ratios to be seen more easily and in a more
intuitive way than with base 10 logarithms.
The only difference between base 2 and base
10 logarithms is multiplication by a constant;
that is, they differ only in the size of the unit.
Figure 1 shows the best fits of Equations 4
and 5 to the points for the conditions with
independent scheduling and no COD. Least
squares regression was used to find the best
fitting lines for the generalized matching law.
Solver in EXCELt was used to find the best
fitting parameters, also by the method of least
squares, of the contingency-discriminability
model. The points for conditions with dependent scheduling, with or without a COD, deviate in no systematic way from the other
points. Including them in the analysis produced no substantial changes in results.
Table 2 shows all of the parameters and
proportion of variance accounted for (r2) by
both the generalized matching law and the
contingency-discriminability model for each
subject. For the large sample, all 4 birds
showed undermatching, with sensitivity ranging from 0.80 to 0.86. For the contingencydiscriminability model, the estimates of p
were all low, ranging from 0.0024 to 0.0111.
The biases greater than 1.0 indicate some
361
CHOICE AND FORAGING
Table 2
Best fitting estimates of sensitivity (s) and bias (b) from the generalized matching law and
proportion of confusion (p) and bias (b) from the contingency-discriminability model with
maximum variance accounted for by each model for each of four types of analysis.
Generalized matching law
Contingency discriminability
s
b
r2
p
b
r2
26
Large sample
Last five sessions
Davison-Jones
0.81
0.79
0.67
0.79
0.91
1.30
.9695
.9857
.9676
.0086
.0106
.0272
0.82
1.01
1.31
.9431
.9624
.9309
27
Large sample
Last five sessions
Davison-Jones
0.86
1.00
0.83
1.60
1.86
2.30
.9695
.9441
.9465
.0024
2.0030
.0096
1.64
1.76
2.29
.9511
.9530
.9252
299
Large sample
Last five sessions
Davison-Jones
0.81
0.70
0.88
1.42
1.42
1.54
.9965
.9688
.9940
.0111
.0237
.0061
1.46
1.41
1.64
.9816
.9727
.9858
973
Large sample
Last five sessions
Davison-Jones
0.80
0.83
0.71
1.23
1.32
1.02
.9681
.9562
.9337
.0041
.0022
.0140
1.26
1.30
0.90
.9209
.9236
.9167
Pigeon
tendency to favor the left key. For all 4 pigeons, goodness of fit (r2) was higher for the
generalized matching law than for the contingency-discriminability model.
Figure 2 shows the results when the data
were summed over only the last five sessions
of each condition for each bird. This analysis
produced sensitivity estimates that varied over
a wider range: 0.70 to 1.00. Relative to the
sensitivity values found for the large samples,
sensitivity for Pigeons 26 and 973 remained
about the same, for Pigeon 27 it increased,
and for Pigeon 299 it decreased (see Table
2). Again the estimates of p in the contingency-discriminability model were small, and the
one for Pigeon 27 was actually negative, an
uninterpretable result. The results for goodness of fit were more variable than for the
large samples: For 2 birds, r2 was higher for
the generalized matching law, and for 2 others, r2 was higher for the contingency-discriminability model.
Figure 3 shows the results of applying the
stability criterion employed by Davison and
Jones (1995). After 11 sessions the relative response rates from sets of three sessions were
assessed for monotonic trend. When a monotonic increase or decrease had been absent
on at least five (not necessarily consecutive)
occasions, the data were summed over the
last five sessions at the point at which Davison
and Jones would have changed conditions.
The results from Davison and Jones’s stability
criterion produced estimates of sensitivity
that also were more variable, ranging from
0.67 to 0.88, than those found for the large
samples (see Table 2). In addition, the sensitivity values for Pigeons 26, 27, and 973 decreased relative to the other two analyses. For
those same 3 birds, the estimates of p were
higher than for the other two analyses. For
all 4 birds, goodness of fit (r2) was higher for
the generalized matching law.
For all three analyses, the fitted lines had
slopes and bias values within the range of
those typically found in concurrent VI VI performance (e.g., Baum, 1979; Taylor & Davison, 1983; Wearden & Burgess, 1982). Although both Equations 4 and 5 fitted the data
well, Table 2 shows that in 10 of the 12 analyses the generalized matching law accounted
for more variance than did the contingencydiscriminability model.
The residuals, found by subtracting the obtained log response ratios from the predicted
log response ratios of each model, were analyzed as a function of obtained log reinforcer
ratios for each bird. If the choice ratios ap-
362
WILLIAM M. BAUM et al.
Fig. 2. Choice relations from data summed over the last five sessions of each condition. Log2 response ratios for
each subject are plotted as a function of log2 obtained reinforcer ratios. The solid straight line in each graph is the
best fit to Equation 4 by least squares regression. The equation and goodness of fit (r2) are shown in each graph.
The dashed curves show the best fits to Equation 5. The parameters of the fits are shown in Table 2.
Fig. 3. Choice relations from data summed over five sessions based on Davison and Jones’s (1995) stability criterion. Log2 response ratios for each subject are plotted as a function of log2 obtained reinforcer ratios. The solid
straight line in each graph is the best fit to Equation 4 by least squares regression. The equation and goodness of fit
(r2) are shown in each graph. The dashed curves show the best fits to Equation 5. The parameters of the fits are
shown in Table 2.
CHOICE AND FORAGING
363
Fig. 4. Analysis of residuals. Upper left: generalized matching linear model fitted to choice ratios derived from
the contingency-discriminability model with p 5 .01. Lower left: residuals from the upper left. The curve, fitted by
least squares, is a third-order polynomial. Upper right: residuals of points on the line from the contingency-discriminability curve. Lower right: residuals of Pigeon 973’s large samples (Figure 1) from the contingency-discriminability
model. The third-order polynomial was fitted only to the triangles (independent VI schedules with no COD). Other
symbols as in Figure 1.
proximated a straight line, the contingencydiscriminability model would deviate from
them systematically, producing a predictable
pattern of residuals. A different pattern
would result if the linear generalized matching relation were fitted to choice ratios that
approximated the S shape required by the
contingency-discriminability model. Figure 4
illustrates these patterns. The upper left panel shows points derived from the contingencydiscriminability model with no bias and p
equal to 0.01. The line was fitted by least
squares to the points. Its slope (sensitivity)
equals 0.81. Although the fit might be considered good (r2 5 .99), the points deviate systematically from the line. The lower left panel
shows the residuals. At the extreme left, they
are positive, they become negative at less extreme reinforcement ratios favoring Alternative 2, they become positive again for nonextreme reinforcement ratios favoring Alternative
1, and then they become negative again for
extreme reinforcement ratios favoring Alter-
native 1. The curve shows a third-order polynomial that captured the pattern of the residuals almost perfectly (r2 5 1.0). Only cubic
and linear components were required; the
quadratic component and zero-order constant both equaled zero. The upper right panel of Figure 4 shows the result of the opposite
comparison—assuming the choice ratios to
conform to a line while fitting the contingency-discriminability model to them. The result
is exactly the opposite pattern, reflected in
coefficients equal to the ones at the lower left
but opposite in sign. These coefficients may
be used to assess the patterns of the actually
observed residuals.
If the situation is like that depicted in the
two left panels of Figure 4, then we expect
that the cubic coefficient will be negative and
the linear coefficient will be positive. If the
situation is the opposite—that the contingency-discriminability model is inappropriate because the data are roughly linear—then the
cubic coefficient will be positive and the lin-
364
WILLIAM M. BAUM et al.
Table 3
Cubic, quadratic, and linear coefficients of the least squares third-order polynomial fits to
residuals for each model as a function of log obtained reinforcer ratio.
Generalized matching law
Contingency discriminability
Cubic
Quadratic
Linear
r2
Cubic
Quadratic
26
Large sample
Last five sessions
Davison-Jones
0.005
0.002
0.003
20.010
20.003
0.018
20.21
20.09
20.14
.53
.19
.36
0.008
0.006
0.008
27
Large sample
Last five sessions
Davison-Jones
0.000
0.006
0.003
0.024
0.029
0.026
20.04
20.25
20.10
.56
.60
.42
299
Large sample
Last five sessions
Davison-Jones
0.003
20.001
0.005
0.010
0.013
20.001
20.11
0.06
20.18
973
Large sample
Last five sessions
Davison-Jones
0.006
0.006
0.000
20.008
0.013
20.006
20.31
20.32
20.02
Pigeon
ear coefficient will be negative. The lower
right panel of Figure 4 shows the contingency-discriminability residuals for Pigeon 973’s
large samples. The polynomial was fitted just
to the triangles (independent schedules with
no COD). The curve conforms to the pattern
shown at the upper right, indicating that the
data deviate systematically from the model in
just the way that would be expected if the
data were approximately linear. Unlike the results of Davison and Jones (1995), the residuals representing the most extreme reinforcement ratios revealed none of the tendency
toward less extreme choice required by the
contingency-discriminability model. The additional points for dependent schedules and
COD are in keeping with the rest, indicating
again that these procedural variations had no
systematic effect on choice.
Table 3 shows the results of polynomial fits
like that shown in Figure 4 for all the pigeons
and all the data sets. The contingency-discriminability residuals for all 4 birds in all
three analyses showed patterns like that in
the right graphs in Figure 4: positive cubic
coefficients and negative linear coefficients.
The quadratic coefficients were often different from zero, but varied unsystematically. Table 3 confirms for all birds and analyses the
conclusion drawn for Pigeon 973’s large sam-
Linear
r2
20.011
20.002
0.019
20.41
20.31
20.41
.74
.70
.71
0.002
0.003
0.006
0.026
0.024
0.027
20.18
20.23
20.27
.74
.52
.57
.78
.45
.84
0.006
0.004
0.007
0.010
0.015
20.001
20.29
20.20
20.31
.97
.35
.91
.93
.89
.03
0.008
0.007
0.004
20.006
0.013
20.006
20.52
20.51
20.30
.97
.94
.23
ples: The choice ratios deviated from the contingency-discriminability model in the pattern that indicates that they approximated
linearity throughout the range. The generalized-matching coefficients generally disconfirm any such pattern as that shown on the
left in Figure 4. Except for the zero cubic coefficient for Pigeon 27’s large samples, the
pattern actually resembles that for the contingency-discriminability residuals. Although the
choice ratios deviated systematically from the
linearity of generalized matching, the pattern
was opposite to that shown on the left in Figure 4. This deviation remains to be explained; we shall return to it below.
Changing over may be analyzed either by
calculating frequency of switching, measured
by rate or probability, or by calculating duration of visit, measured by time per visit or
responses per visit. In the absence of visit-byvisit measures (e.g., Baum & Aparicio, 1999;
Buckner et al., 1993), duration of visit was estimated by dividing number of pecks at an
alternative by half the number of changeovers. Probability of switching was the reciprocal of this: half the number of changeovers
divided by the number of pecks. This analysis
was done only for the large samples.
Figure 5 shows pecks per visit (PPV) as a
function of preference (base 2 log of the ratio
CHOICE AND FORAGING
365
Fig. 5. Visit duration, measured as average pecks per visit (PPV), as a function of preference, measured as log
(base 2) ratio of responses, preferred (rich) alternative (P) to nonpreferred alternative (N). Triangles represent
conditions with independent schedules and no COD. Squares represent conditions with dependent scheduling and
no COD. Circles represent conditions with dependent scheduling and a COD. Filled symbols represent the preferred
alternative; open symbols represent the nonpreferred alternative. The broken line in each graph represents the locus
of equality. Note logarithmic (base 10) y axis.
of pecks at the preferred alternative to pecks
at the nonpreferred alternative). The absence of significant position bias allowed the
results to be represented without regard to
which side was preferred; when plotted separately, the results for the two sides were symmetrical. For the conditions with no COD, as
soon as preferences exceeded indifference
(i.e., a log value of zero)—even for preferences as small as 2:1—duration of visits to the
nonpreferred side became brief, close to one
peck, and showed no systematic variation with
preference. Also for those conditions, log duration of visits on the preferred side increased linearly with preference, following
the major diagonal. This indicates that pecks
per visit on the preferred side almost equaled
preference, an equality that would be perfect
if the pigeons always made just one peck in a
visit to the nonpreferred side. The points lie
slightly above the major diagonal, however,
suggesting a small systematic deviation from
equality. Adding the 3-s COD increased PPV
on both sides, producing longer visits to the
nonpreferred side but fewer of them, with
the result that the COD had little effect on
preference.
Figure 6 shows, for the conditions without
a COD, log (base 2) preference as a function
of log (base 2) reinforcement ratio, preferred
side to nonpreferred side, regardless of position. For all 4 birds, the slope of the regression line was close to 1.0. In other words,
when position (i.e., left-right) bias was ignored, the results approximated matching
with bias. The negative values of log b indicate that there was a roughly constant proportional bias in favor of the nonpreferred
(lean) alternative, regardless of whether it
was on the left or the right.
The approximate matching shown in Figure 6 contrasts with the undermatching
shown in Figure 1. The difference arises from
the standard analysis’s reliance on position as
the defining characteristic of the alternatives,
in contrast with the reliance in Figure 6 on
rich versus lean or preferred versus nonpreferred, regardless of position. The approxi-
366
WILLIAM M. BAUM et al.
Fig. 6. Preference, measured as log (base 2) response ratio, preferred alternative to nonpreferred alternative, as
a function of relative reinforcement, measured as log (base 2) reinforcement ratio, preferred alternative to nonpreferred alternative. The diamonds represent conditions without a COD. The circles represent conditions with a COD.
The solid lines (equations shown) were fitted to the diamonds by the method of least squares. The broken lines
represent perfect matching.
mate linearity of the relations in Figure 6 implies a systematic deviation from the standard
matching relation defined in terms of position. If choice is governed by the difference
between rich versus lean instead of left versus
right, and there is a constant proportional
bias toward the nonpreferred (lean) alterna-
Fig. 7. The two-line relation implied by Figure 6 illustrated for Pigeon 973’s choice relation (large samples;
Figure 1). The two line segments, joined at indifference,
have slopes equal to 1.0. The displacements (bias) from
the broken line, the locus of strict matching, were derived by the method of least squares. Bias favors the lean
alternative, whether on the left or the right.
tive, as shown in Figure 6, then that bias,
when expressed in terms of position, will go
in opposite directions, depending on whether the rich alternative is on the left or on the
right.
Figure 7 shows, by way of illustration, this
implication of Figure 6. The data of Pigeon
973 are replotted from Figure 1, except that
two lines of slope 1.0 were fitted to the ratios:
one for ratios less than 1.0 and one for ratios
greater than 1.0. A horizontal line at indifference connects the two lines. The broken line
shows the locus of strict matching. If the rich
alternative was on the right (points to the
lower left), the bias favored the left, and all
points lay above the matching line, whereas
if the rich alternative was on the left (points
to the upper right), the bias favored the right,
and all points fell below the matching line.
As Figure 1 shows, a line fitted to all the
points at once has a slope less than 1.0. Figure 7 shows, however, that choice deviated
systematically from such a single line. As long
as the relations in Figure 6 have slopes approximating 1.0, they imply the two-line pattern shown in Figure 7.
CHOICE AND FORAGING
Were a single line fitted to the points in
Figure 7, as in Figure 1, the two-line pattern
implies a specific pattern of deviation in the
residuals. The residuals would be negative at
the extreme left, then positive for points at
the left nearer the middle, then negative
again for points at the right near the middle,
and then positive again for points at the extreme right. Inspection of the triangles in the
graph for Pigeon 973 in Figure 1 verifies this
pattern. It happens to be qualitatively the
same pattern as shown in the upper right
graph of Figure 4. It explains the similarity of
coefficients for the two models shown in Table 3. If the data conform to the two-line pattern shown in Figure 7, then the generalizedmatching residuals will approximately follow
the pattern on the right in Figure 4, with the
result that the cubic coefficient will be positive and the linear coefficient will be negative.
Thus, the prevalence of this pattern in Table
3 further verifies the presence of the two-line
pattern in all the birds’ data and almost all
of the data sets.
DISCUSSION
The results support a view of concurrent
performance similar to that suggested by
Houston and McNamara (1981): responding
predominantly at the rich alternative occasionally interrupted by visits to the lean alternative that are brief and constant. Such a view
contradicts any model of concurrent performance that assumes, explicitly or implicitly,
that responding adjusts at both alternatives—
for example, melioration and the kinetic
model (Herrnstein & Vaughan, 1980; Myerson & Miezin, 1980). It may be reconciled
with various optimality models, such as momentary maximizing, hill climbing, and global optimality (Baum, 1981; Hinson & Staddon, 1983; Houston & McNamara, 1981;
Shimp, 1966). Even though the results go
against the usual version of the generalized
matching law, based on position or stimulus
of the alternatives, they provide no support
for the contingency-discriminability model.
The contingency-discriminability model may
account for more of the variance in the data
obtained under procedures in which the stimuli associated with VI schedules are difficult
to discriminate (e.g., Davison, 1996; Davison
& Jones, 1995). However, under more typical
367
procedures it accounts for less variance in
choice than does the generalized matching
law (Figure 1), it deviates systematically from
the choice relation (Figure 4 and Table 3),
and it gains no support from the rich-lean
analysis shown in Figure 6.
The lack of any systematic effect of dependent scheduling or adding a COD, shown in
Figures 1, 2, and 3, rules out these procedural
features as producing the difference in results between this experiment and that of
Davison and Jones (1995). Another important difference between their methods and
the present ones might have been the stability
criterion used and the number of sessions
completed per condition. Todorov, Castro,
Hanna, Bittencourt de Sa, and Barreto
(1983) reported effects of both number of
conditions and number of sessions per condition. Although they found within-subject
sensitivity to relative rates of reinforcement to
decrease with increasing number of conditions, neither the present experiment nor
that of Davison and Jones found any such effect of number of conditions. The finding by
Todorov et al. that running more sessions per
condition increased sensitivity, however, may
apply to the present comparison. The Davison-Jones stability criterion caused conditions
to change sooner in their experiment (after
a minimum of 18 and a maximum of 37 sessions). Had their criterion been applied in
this experiment, every condition would have
ended far sooner than it did (after 18 to 34
sessions). Thus the large numbers of sessions
completed per condition in the present study
may explain why the sensitivity estimates were
reduced for 3 of the 4 birds when the Davison-Jones stability criterion was applied (Table 2).
The small samples used by Davison and
Jones (1995) also resulted in fewer than five
reinforcers obtained on the lean alternative
in the most extreme ratio conditions. In some
conditions only one or two reinforcers were
obtained on the lean schedule. Such small
samples would increase the variability in estimates of relative responding and relative reinforcement. This inaccuracy is reflected in
the variability of the choice relations shown
in Table 2. For both analyses with small samples, the range of variation in sensitivity was
greater than for the large samples. This is
probably due mainly to the small numbers of
368
WILLIAM M. BAUM et al.
reinforcers in the small samples, because
when only one or two reinforcers occur on
the lean alternative, the reinforcement ratio
may be highly volatile. When rate of reinforcement at the lean alternative is low, accuracy demands that the sample size for measuring it must be proportionately large.
Although the liberal stability criterion and
the small samples may have reduced sensitivity and accuracy of estimates of choice and
relative reinforcement, it is unlikely that they
produced the systematic deviations from the
generalized matching law that Davison and
Jones (1995) observed. One cannot rule out
the possibility that the volatile small samples
deviated systematically by chance, but it
seems more likely that the difference arose
from their having used stimuli that were difficult to discriminate. The present experiment used two different locations, a difference known to produce good discriminations
in pigeons, whereas Davison and Jones used
two different brightness levels on the same
key. If this difference is the cause, the contingency-discriminability model would apply
only when the stimuli associated with the two
alternatives were poorly discriminated and
would have no bearing on typical procedures
with highly discriminable stimuli.
Both the generalized matching law and the
contingency-discriminability model may be
reconciled with the present data by reexpressing them in terms of rich versus lean alternatives instead of position or stimuli. The
matching law becomes
log
1B 2 5 log 1r 2 1 log c,
BP
rP
N
N
(6)
where P and N refer to preferred (rich) and
nonpreferred (lean) and c refers to bias favoring the rich or lean alternative. This approximates the lines fitted in Figure 6. It includes no sensitivity parameter, because the
slopes were close to 1.0. The contingency-discriminability model (Equation 5) would reduce to Equation 6 if the subscripts relating
to stimuli were replaced with P and N, and p
were zero. It might then apply to procedures
in which the alternatives’ stimuli are imperfectly discriminated, if it could be shown that
values of p greater than zero allowed correct
fits to the response ratios:
log
1B 2 5 log 1r
BP
N
2
rP 2 prP 1 prN
1 log c. (7)
N 2 prN 1 prP
This is constructive because it allows us to account for two types of undermatching: that
resulting from stimulus confusion and that associated with the generalized matching law. It
suggests, however, that the latter type may be
more apparent than real, because generalized-matching undermatching arises from fitting an inappropriate equation.
Although Figure 1 showed that fits of the
response ratios to generalized matching were
excellent when assessed by r2, goodness of fit
cannot be assessed only by r2. Even if the variance accounted for is high, the remaining
variance should be unsystematic. An equation
may be shown to be incorrect by showing that
the data deviate from it systematically. The residual patterns revealed in the coefficients in
Table 3 demonstrate that the response ratios
deviated systematically from the single line of
generalized matching and instead conformed
to a two-line pattern in which matching held
throughout, but with bias in favor of the nonpreferred (lean) alternative. Figure 7 illustrated this two-line pattern for response ratios
plotted in the usual way, according to position. Examination of Figure 6 also reveals
that Equation 6 fitted the data with no systematic deviation. One would conclude that typical undermatching arises from characterizing concurrent performance incorrectly,
because performance moves toward predominant responding at the rich alternative and
occasional brief visits to the lean alternative
(Figure 5).
Although such a view of concurrent performance allows Equation 6 to apply, it also
suggests that an alternative analysis, focused
on visits to the lean alternative, would be
more appropriate. If we assume that visits to
the nonpreferred alternative are brief and
constant on average, which Figure 5 showed
to be approximately true, then we may rewrite response ratios according to
BP
B
5 P ,
BN
N·D
(8)
where BP and BN are responses at the preferred and nonpreferred alternatives, N is the
number of visits to the nonpreferred alternative, and D is the number of responses
369
CHOICE AND FORAGING
made on a visit to the nonpreferred alternative. If D is close to 1.0, as in Figure 5, Equation 8 states that the response ratio will be
close to the average duration (number of responses) of a visit to the preferred alternative.
This result appears in Figure 5 in the close
approximation of the triangles and squares to
the major diagonal.
If we take the logarithm of Equation 8 and
rearrange, we obtain
log
1 N 2 5 log B
BP
BP
1 log D,
(9)
N
the equation that would describe the points
that parallel the major diagonal in Figure 5.
This equation was used to estimate log D and
D. For Pigeons 26, 27, 299, and 973, the estimates of log2 D equaled 0.437, 0.152, 0.538,
and 0.321. These correspond to estimates of
D of 1.35, 1.11, 1.45, and 1.25—averages of a
little more than one peck per visit to the nonpreferred lean alternative.
If we use Equation 8 to rewrite Equation 6,
we obtain
log
1N·D2 5 log 1r 2 1 log c.
BP
rP
(10)
N
Equation 10 approximates relations fitted to
the points in Figure 6. The negative values of
log c in Figure 6, which represent a value of
c less than 1.0, indicate a bias in favor of the
lean alternative. In contrast with the bias b of
the generalized matching relation (Equation
4), which is assumed to be independent of
reinforcement rates (r1 and r2), the bias c depends on reinforcement. It represents neither position bias nor stimulus bias, but rather favoritism toward the rich or lean
alternative, regardless of position or stimulus.
Equation 10 implies that the bias shown in
Figure 6 should be attributed to an excessive
tendency to visit the lean alternative. The value of c indicates the proportion of N that
should have been the number of visits to the
lean alternative for strict matching to obtain.
For example, the value of log2 c for Pigeon
973, which was about 21.0, would indicate
that the obtained number of visits should be
halved for strict matching; Pigeon 973 visited
the lean side about twice as often as it would
have according to perfect matching.
Thus, the apparent undermatching seen in
Figure 1 and the good approximation to the
two-line pattern seen in Figure 7 and implied
by the coefficients in Table 3 arose from an
excessive tendency to visit the lean alternative. Such an explanation implies that penalties on switching, such as COD and travel,
would increase apparent sensitivity (s) by decreasing the frequency of visiting the nonpreferred lean alternative, thereby increasing c.
If c increased to 1.0, strict matching would
result. If c increased beyond 1.0, as occurs
with travel requirements, the generalized
matching law would represent that bias—now
in favor of the rich alternative—as overmatching (Baum, 1982; Boelens & Kop, 1983). In
the present experiment, it appeared that the
COD had two effects: It decreased the frequency of switching and increased the number of pecks made in a visit to the lean side
(Figure 5). The consistency of the relative
measures for the conditions including a COD
with the rest (Figures 1 and 6) indicates that
the two effects tended to cancel each other
out.
If the choice relation is determined by the
frequency of visits to the lean alternative,
then analysis should focus on rate or probability of visiting the lean alternative as the key
dependent variable. The present results point
in that direction, as do some theoretical treatments of concurrent performance (e.g.,
Houston & McNamara, 1981) and some experimentation (e.g., Baum, 1982). If we rearrange Equation 10, we obtain
log
1B 2 5 log 1r 2 2 log D 2 log c,
N
P
rN
(11)
P
which states that the probability of visiting the
nonpreferred (lean) alternative depends directly on the ratio of reinforcement (lean to
rich) with an intercept equal to 2(log D 1
log c) or bias equal to 1/Dc. If BP were measured as time spent responding, then N/BP
would represent rate of visiting the lean alternative instead of probability.
Figure 8 shows the application of Equation
11 to the large samples. The log (base 2) of
the probability of visiting the lean alternative—the reciprocal of visit duration (Figure
5)—is shown as a function of the log (base 2)
of the ratio of reinforcement, nonpreferred
(lean) to preferred (rich). The fits are all
good, with r2 ranging from .86 to .97. All the
slopes are substantially higher than for the
370
WILLIAM M. BAUM et al.
Fig. 8. Probability (log, base 2) of visiting the nonpreferred (lean) alternative as a function of relative reinforcement (log, base 2) at the nonpreferred alternative. Diamonds represent conditions with no COD. Circles represent
conditions with a COD. The lines (equations shown) were fitted to the diamonds by the method of least squares.
standard analysis shown in Figure 1, and all
are reasonable approximations to 1.0 (Baum,
1979). The slopes for the choice relations in
Figure 6 are close to the slopes for probability
of switching in Figure 8, as they should be. In
sum, the probability of visiting the lean alternative was approximately directly proportional to (i.e., matched with bias) the relative reinforcement for the lean alternative. The
points representing conditions with a COD
tend to fall below the others, indicating that
the COD decreased the probability of visiting
the lean alternative. Although too few conditions with a COD were conducted, if the
same matching relation held, the effect of the
COD would be on the intercept of the relation, presumably by increasing visit duration
D in Equation 11.
If we compare the intercepts in Figure 8
with those we derive from the estimates of D
and c based on Figures 5 and 6, there is a
rough correspondence. Pigeon 299, for
which log2 D and log2 c were approximately
equal but opposite in sign, shows an intercept
close to zero. Pigeon 973, for which the difference between log2 D and log2 c was greatest,
shows the largest intercept. The intercepts for
Pigeons 26 and 27, predicted from Equation
11 to be intermediate, are intermediate.
The standard analyses shown in Figure 1
suggested that choice relations that were uniform even out to reinforcement ratios over
200:1, but with undermatching. The analysis
comparing the preferred (rich) alternative to
the nonpreferred (lean) shown in Figures 5
and 6 restored an approximation to matching. At the same time, however, that analysis
suggested a focus on visits to the rich and
lean sides. Indeed, Figure 5 suggests that the
systematic variance in choice arises solely
from the systematic variance in visits at the
rich alternative. Duration of visits to the rich
alternative, however, is just the inverse of frequency of visiting the lean alternative. It thus
becomes possible to characterize concurrent
performance as responding on the rich alternative interrupted at some frequency, depending on the relative reinforcement there,
by brief visits to the lean alternative (cf.
Baum, 1982; Baum & Aparicio, 1999; Houston & McNamara, 1981; Houston et al.,
1987). When reinforcement approaches
equality at the two alternatives, responding
approaches indifference, and visits become
CHOICE AND FORAGING
brief at both alternatives (i.e., rate of switching approaches a maximum). Figure 8 shows
that such a characterization leads to an orderly and simple outcome, as simple as the
original matching law (Herrnstein, 1961):
Frequency of visiting the lean alternative is
directly proportional to the relative reinforcement at the lean alternative. An earlier paper
(Baum, 1982) argued that matching is a special case in the range of concurrent performances that can be generated by imposing
penalties on switching. This view of concurrent performance may both explain variation
in sensitivity and, at the same time, move us
to an understanding more general than the
matching law.
The findings in Figures 6 and 8 leave open
a number of questions. The contingency-discriminability model (Equation 7) may apply
to situations in which the concurrent alternatives are imperfectly discriminated. That
remains to be tested. If the alternatives are
highly discriminable, the present analysis suggests that concurrent performance consists of
responding on the rich alternative occasionally interrupted by brief visits to the lean alternative. It remains to be seen how general
these findings may be. For example, how
does including a COD over a broad range of
relative reinforcement affect the rich-lean relations shown in Figure 6 and the probabilities of visiting the lean alternative shown in
Figure 8? In approaching such questions, it
will be important to attend to the concerns
addressed in the present methods, that is, to
continue conditions until they are definitely
stable and to gather large samples of stable
performance. To draw quantitative conclusions, the data must be accurate and reliable.
REFERENCES
Baum, W. M. (1974a). Chained concurrent schedules:
Reinforcement as situation transition. Journal of the Experimental Analysis of Behavior, 22, 91–101.
Baum, W. M. (1974b). On two types of deviation from
the matching law: Bias and undermatching. Journal of
the Experimental Analysis of Behavior, 22, 231–242.
Baum, W. M. (1975). Time allocation in human vigilance. Journal of the Experimental Analysis of Behavior, 23,
45–53.
Baum, W. M. (1979). Matching, undermatching, and
overmatching in studies of choice. Journal of the Experimental Analysis of Behavior, 32, 269–281.
Baum, W. M. (1981). Optimization and the matching
371
law as accounts of instrumental behavior. Journal of the
Experimental Analysis of Behavior, 36, 387–403.
Baum, W. M. (1982). Choice, changeover, and travel.
Journal of the Experimental Analysis of Behavior, 38, 35–
49.
Baum, W. M., & Aparicio, C. F. (1999). Optimality and
concurrent variable-interval variable-ratio schedules.
Journal of the Experimental Analysis of Behavior, 71, 75–
89.
Baum, W. M., & Rachlin, H. C. (1969). Choice as time
allocation. Journal of the Experimental Analysis of Behavior, 12, 861–874.
Boelens, H., & Kop, P. F. M. (1983). Concurrent schedules: Spatial separation of response alternatives. Journal of the Experimental Analysis of Behavior, 40, 35–45.
Buckner, R. L., Green, L., & Myerson, J. (1993). Shortterm and long-term effects of reinforcers on choice.
Journal of the Experimental Analysis of Behavior, 59, 293–
307.
Davison, M. (1996). Stimulus effects on behavior allocation in three-alternative choice. Journal of the Experimental Analysis of Behavior, 66, 149–168.
Davison, M., & Jenkins, P. E. (1985). Stimulus discriminability, contingency discriminability, and schedule
performance. Animal Learning & Behavior, 13, 77–84.
Davison, M., & Jones, B. M. (1995). A quantitative analysis of extreme choice. Journal of the Experimental Analysis of Behavior, 64, 147–162.
Davison, M., & McCarthy, D. (1988). The matching law: A
research review. Hillsdale, NJ: Erlbaum.
Findley, J. D. (1958). Preference and switching under
concurrent scheduling. Journal of the Experimental Analysis of Behavior, 1, 123–144.
Herrnstein, R. J. (1961). Relative and absolute strength
of response as a function of frequency of reinforcement. Journal of the Experimental Analysis of Behavior, 4,
267–272.
Herrnstein, R. J., & Vaughan, W. (1980). Melioration
and behavioral allocation. In J. E. R. Staddon (Ed.),
Limits to action: The allocation of individual behavior (pp.
143–176). New York: Academic Press.
Heyman, G. M. (1979). A Markov model description of
changeover probabilities on concurrent variable-interval schedules. Journal of the Experimental Analysis of Behavior, 31, 41–51.
Hinson, J. M., & Staddon, J. E. R. (1983). Matching,
maximizing, and hill-climbing. Journal of the Experimental Analysis of Behavior, 40, 321–331.
Houston, A. I., & McNamara, J. (1981). How to maximize reward rate on two variable-interval paradigms.
Journal of the Experimental Analysis of Behavior, 35, 367–
396.
Houston, A. I., McNamara, J. M., & Sumida, B. H.
(1987). Relative allocation on concurrent schedules
can depend on schedule parameters when behavioral
parameters are constant. Journal of the Experimental
Analysis of Behavior, 47, 127–131.
Lobb, B., & Davison, M. C. (1975). Preference in concurrent interval schedules: A systematic replication.
Journal of the Experimental Analysis of Behavior, 24, 191–
197.
Miller, J. T., Saunders, S. S., & Bourland, G. (1980). The
role of stimulus disparity in concurrently available reinforcement schedules. Animal Learning & Behavior, 8,
635–641.
Myerson, J., & Miezin, F. M. (1980). The kinetics of
372
WILLIAM M. BAUM et al.
choice: An operant systems analysis. Psychological Review, 87, 160–174.
Rachlin, H. (1978). A molar theory of reinforcement
schedules. Journal of the Experimental Analysis of Behavior, 30, 345–360.
Shimp, C. P. (1966). Probabilistically reinforced choice
behavior in pigeons. Journal of the Experimental Analysis
of Behavior, 9, 443–455.
Shull, R. L., & Pliskoff, S. S. (1967). Changeover delay
and concurrent schedules: Some effects on relative
performance measures. Journal of the Experimental
Analysis of Behavior, 10, 517–527.
Stubbs, D. A., & Pliskoff, S. S. (1969). Concurrent responding with fixed and relative rate of reinforcement. Journal of the Experimental Analysis of Behavior, 12,
887–895.
Taylor, R., & Davison, M. (1983). Sensitivity to reinforcement in concurrent arithmetic and exponential
schedules. Journal of the Experimental Analysis of Behavior, 39, 191–198.
Todorov, J. C., Castro, J. M., Hanna, E. S., Bittencourt de
Sa, M. C. N., & Barreto, M. de Q. (1983). Choice,
experience, and the generalized matching law. Journal
of the Experimental Analysis of Behavior, 40, 99–111.
Wearden, J. H. (1983). Undermatching and overmatching as deviations from the matching law. Journal of the
Experimental Analysis of Behavior, 40, 332–340.
Wearden, J. H., & Burgess, I. S. (1982). Matching since
Baum (1979). Journal of the Experimental Analysis of Behavior, 38, 339–348.
Received May 25, 1998
Final acceptance December 3, 1998
373
CHOICE AND FORAGING
APPENDIX
Data for each subject summed over the large samples. Numbers in parentheses give the number of sessions in the sample.
Reinforcement ratio
Pigeon Condition
26
27
299
973
1
2
3
4
5
6
7
8
9
10
11
12
13
14
1
2
3
4
5
6
7
8
9
10
11
12
1
2
3
4
5
6
7
8
9
10
11
12
1
2
3
4
5
6
7
8
9
10
11
12
(9)
(11)
(12)
(9)
(25)
(56)
(27)
(12)
(17)
(57)
(26)
(38)
(49)
(64)
(7)
(9)
(23)
(12)
(18)
(86)
(9)
(16)
(31)
(67)
(38)
(59)
(8)
(13)
(15)
(15)
(38)
(93)
(20)
(19)
(29)
(56)
(43)
(50)
(9)
(11)
(30)
(24)
(23)
(102)
(16)
(32)
(32)
(45)
(34)
(46)
Responses
Obtained
Left
Right
Left
Right
CO
4:1
1:9
64:1
1:32
128:1
1:256
4:1
1:8
64:1 d
1:256 d
128:1 d
1:256 d,
128:1 d, c
1:256 d,
1:4
9:1
1:64
32:1
1:128
256:1
1:4
8:1
1:128 d
256:1 d
1:128 d,
256:1 d, c
1:4
1:9
64:1
1:32
128:1
1:256
4:1
1:8
128:1 d
1:256 d
128:1 d, c
1:256 d,
1:4
9:1
1:64
32:1
1:128
256:1
1:4
8:1
1:128 d
256:1 d
1:128 d,
256:1 d, c
3.8:1
1:6.5
66.1:1
1:26.7
199:1
1:194
4.2:1
1:7.6
51.3:1
1:242
121:1
1:215
111:1
1:204
1:4.7
8.6:1
1:75.7
32.1:1
1:110
490:1
1:3.9
7.8:1
1:154
243:1
1:168
277:1
1:3.1
1:7.5
41.9:1
1:32.3
107.6:1
1:247
3.5:1
1:7.3
100:1
1:298
118:1
1:285
1:3.3
9.1:1
1:67.6
27.7:1
1:122
532:1
1:4.2
7.1:1
1:134
239:1
1:169
282:1
13,145
10,500
30,246
1,625
61,488
1,095
46,328
5,681
47,751
2,680
79,139
3,728
153,459
5,230
4,050
20,736
4,124
35,660
1,047
218,538
5,852
28,727
972
190,044
1,329
159,014
7,815
8,413
57,818
2,799
141,766
4,287
51,931
13,197
112,017
2,388
172,863
1,300
7,051
14,217
3,190
29,241
766
138,469
7,121
28,572
889
61,808
1,513
77,265
6,900
23,069
2,680
30,770
723
169,336
16,628
34,054
1,711
193,472
597
148,070
3,505
222,625
13,731
5,266
47,242
780
37,548
334
15,671
3,942
76,659
732
88,574
1,615
12,338
28,756
2,097
31,521
1,827
271,671
18,911
45,096
2,048
203,128
2,289
158,549
10,847
2,646
52,281
3,345
38,959
325
12,042
9,272
48,641
1,907
50,724
557
569
118
2,049
26
1,990
23
1,743
112
1,334
18
2,063
14
3,885
25
97
645
24
931
13
6,866
147
1,135
16
5,338
18
4,703
155
122
1,172
36
3,012
30
1,246
184
2,297
15
3,411
14
168
793
35
1,853
15
7,985
247
2,244
19
3,585
16
3,667
151
762
31
694
10
4,457
417
848
26
4,363
17
3,011
35
5,095
460
75
1,816
29
1,427
14
573
145
2,464
22
3,022
17
485
918
28
1,164
28
7,410
354
1,336
23
4,465
29
3,986
552
87
2,365
67
1,825
15
1,033
316
2,541
15
2,704
13
10,900
10,254
3,756
2,714
1,125
1,275
28,599
10,847
2,660
3,188
1,099
1,068
968
1,092
7,086
7,344
7,712
1,488
1,974
640
9,729
6,991
1,900
1,438
420
706
10,888
12,349
3,278
3,041
2,975
6,640
26,121
15,044
3,150
2,770
640
290
—
3,965
5,076
4,860
1,369
614
10,314
13,320
1,361
3,536
286
458
c
c
c
c
c
Note. A d indicates dependent scheduling; a c indicates the use of a 3-s COD.
View publication stats
Reinforcers
Programmed