Download Humphreys (1939b) revisited: Is there a “verbal” PREE?

EUROPEAN JOURNAL OF BEHAVIOR ANALYSIS
2008, 9, xx - xx
NUMBER 2 (WINTER 2008)
1
Humphreys (1939b) revisited:
Is there a “verbal” PREE?
Frode Svartdal and Kjetil Heggelund
University of Tromsø, Tromsø, Norway
Humphreys (1939b) reported the PREE (partial reinforcement extinction effect) of a “verbal”
classically conditioned response. His interpretation in terms of expectation corresponds well to the
modern cognitive view of conditioning. To date, however, Humphreys (1939b) is the only study we are
aware of that attempts to demonstrate the involvement of a verbalized expectation in the PREE. We
performed two replications of Humphreys’ experiment, and offer an alternative interpretation to the
original conclusions. Specifically, we suggest that in Humphreys’ experiment, verbalized expectations
were not classically conditioned responses but rather responses established by confirmatory consequences and strong task demands. In a third experiment we demonstrate a PREE of the conditioned
verbal expectation response, but the effect was much weaker than that of the original experiment. Our
results suggest that the role of cognition in the generation of the PREE is weaker than assumed by
Humphreys and cognitive theory.
Keywords: Persistence, verbal response, reward, extinction, PREE
In 1939, Humphreys (1939a) published one
of the first papers on the partial reinforcement
extinction effect (PREE). In it, Humphreys
reported data from a classical eyelid conditioning procedure with students as participants. The
results demonstrated that partial, as compared
to continuous, reinforcement during acquisition causes increased resistance to subsequent
extinction. This effect, often coined the Humphreys’ paradox (e.g., Mazur, 2004), is now well
established as one of the basic findings of the
psychology of learning.
Later the same year, Humphreys (1939b)
published another paper that purportedly demonstrated the same effect, but now with verbalized expectations as the conditioned response.
Using a classical conditioning procedure1 with
24 trials, student participants wrote down, for
each trial, “yes” or “no” as to whether the US
would occur following the CS. The verbalized
anticipation of the US thus served as the CR.
Two reinforcement conditions, 100% or 50%,
were arranged. After acquisition, 12 extinction
trials with no reinforcement followed. The
results indicated that extinction performance
following the 50% reinforcement phase was
slower compared to performance following the
100% contingency. These results thus seemed to
conform well to the initial PREE findings, and
implied that behavioral (eyelid responses to the
CS) and cognitive (verbalized expectations of
the US) measures of the CR demonstrate a close
correspondence. Humphreys (1939, p. 298)
concluded that “extinction of verbal expectations of a stimulus duplicates rather closely the
extinction of conditioned eyelid reactions”, and
added: “This indicates that expectancy is, at least,
an important variable in conditioning” (ibid.).
1
Humphreys noted (1939b, p. 295) that his procedure is
“analogous” to a classical conditioning procedure, implying that
alternative descriptions are possible. However, the procedure fulfills the basic requirements for a classical conditioning procedure,
i.e., the presentation of two stimuli in a specified temporal order
independently of the behavior of the participant (e.g., Catania,
1992). We therefore use the term “classical conditioning” in the
description of the procedure throughout this paper.
We thank Reidar Kvadsheim, Oslo University College, for
helpful assistance in recruiting participants. Correspondence
concerning this paper should be addressed to F. Svartdal, Department of Psychology, University of Tromsø, 9037 Tromsø, Norway.
E-mail: [email protected]
1
2
Frode Svartdal and Kjetil Heggelund
Humphreys’ (1939b) report on the role of
expectancy in extinction was probably the first
experimental study that attempted to assess the
role of cognition in acquisition and extinction
of the conditioned response. His data were correlational: Because of the similarity between the
nonverbal and verbal PREE, the results were
taken to support the view that acquisition as
well as extinction involve explicit verbalization of contingencies. Interestingly, this view
corresponds well to a current and very popular
view of human conditioning that assumes
conditioning and extinction to be dependent
on explicit cognitive processes (e.g., Lovibond,
2004; Shanks & Lovibond, 2002; Shanks,
2005). Cognitive theory holds that learning
and extinction depend on consciously available
contingency knowledge, require attentional resources, and are influenced by language (Brewer,
1974; Lovibond, 2004; Shanks, 2005; Shanks
& Lovibond, 2002). Overall, this view is supported by substantial experimental evidence,
both within classical and operant conditioning
(see, e.g., Lovibond, 2004; Shanks, 2005, for
reviews). However, the cognitive interpretation
of extinction – i.e., that behavioral extinction
depends on changes in contingency beliefs
(Lovibond, 2004) – is largely theoretical and
has little empirical support.
For the PREE specifically, Humphreys’
(1939b) experiment appears to be the only
study that has attempted to empirically document the involvement of explicit cognition in
extinction of a conditioned response. In fact,
we are aware of only three additional studies
that actually have investigated the role of beliefs, expectations and verbalizations related to
the PREE (Svartdal, 2000, 2003; Svartdal &
Silvera, 2008). Whereas Humphreys’ results
are consistent with a cognitive interpretation
of PREE, the three other studies provide no
evidence for the involvement of explicit cognitive processes in the PREE. For example,
in one set of experiments (Svartdal, 2003),
student participants were exposed to CRF vs.
PRF operant contingencies, then to extinction
contingencies. As expected, the conventional
PREE appeared in behavioral measures. We
also obtained information about explicit and
implicit verbalizations of the learning contingencies. Verbalizations closely matched acquisition contingencies. In contrast, beliefs about
extinction persistence did not reflect the PREE
and instead demonstrated an undifferentiated
and gradual decrease in persistence over extinction trial probes. In a later study (Svartdal &
Silvera, 2008, Exp. 2), we actually demonstrated
a behavior-cognition dissociation under extinction conditions. Following exposure to CRF vs.
PRF contingencies, participants had a small
number of extinction trials. The conventional
PREE appeared in the behavioral measures,
but no difference was observed in probes for
explicit or implicit contingency beliefs; instead,
these probes indicated a gradual but non-differentiated extinction response. In contrast to
Humphreys’ conclusion, our findings indicate
that contingency beliefs under extinction conditions are not related to the immediately preceding reinforcement history and therefore cannot
explain the behavioral PREE. Thus, whereas
probes are sensitive to extinction conditions in
the sense that verbalized probes indicate gradual
extinction, thus supporting cognitive theory in
a general sense, such probes are not sensitive to
the CRF vs. PRF manipulation.
Because Humphreys’ experiment may be
the only study to directly support the notion
that verbalization is closely correlated with the
PREE, it is of importance to verify its validity. To our knowledge, no replication of this
experiment has been performed. Furthermore,
the 1939b paper has been cited more than 200
times, and its interpretation of conditioning
and extinction is almost identical to a modern
cognitive view of conditioning and extinction.
The present paper presents data from replications and extensions of the Humphreys’ investigation, and argues that Humphreys may not
have demonstrated the PREE in his explicit
cognition (verbalization) study. We will argue
that the response he observed most probably
was shaped by consequences of responding
and strong task demands. If correct, this will
question the validity of Humphreys’ conclusion
and hence its support of a cognitive account of
the PREE.
We start by presenting a critical evaluation
Humphreys Revisited
of the Humphreys’ 1939b investigation (see
also Williams & Hartley, 1982). The procedure
was very simple. First, over 24 acquisition trials, participants guessed whether the US would
occur, given the CS. Two mutually excluding
response alternatives, “yes” or “no”, were available. In the 100% condition, “yes” responses
were consistently confirmed by the occurrence
of the US, and any “no” response would be disconfirmed by the occurrence of the US. Then,
under the subsequent extinction phase, “no”
answers would now be continuously confirmed
by the absence of the US, and any “yes” response
would be disconfirmed. Given such powerful
consequences of the emission of two response
alternatives under acquisition and extinction,
it is highly likely that the prediction task was
greatly facilitated. Accordingly, Humphreys
reported CRF “learning” curves that quickly
approached asymptotic level during acquisition trials, and also quickly diminished to zero
during extinction trials. Humphreys noted
that these results differed from typical animal
data; in fact, extinction was complete by the
third extinction trial (see Humphreys, 1939b,
Figure 2).
We hypothesize, therefore, that the verbal
“yes” response observed during acquisition
trials in the 100% condition was not a classically conditioned response; if anything, it
was most likely a response strongly controlled
by immediate consequences and hence could
more correctly be seen as an operant response.2
However, the task demands in the guessing
task generated a guessing pattern in the 100%
condition that may render this interpretation
doubtful as well. In fact, if the procedure is
seen as a simple discrimination task (“predict
if a stimulus will occur, given a signal”), most
participants will easily provide the correct response within few acquisition trials relying on
confirmatory feedback. Then, under extinction,
the correct response is easily identified applying
the same confirmatory strategy.
2
It should also be noted that Humphreys’ procedure,
although formally a classical conditioning procedure, did not
involve an unconditioned response (UR) that could elicit an
US. The lack of the usual automatic US-UR relation (this was
present in his 1939a experiment in the form of blink responses
to airpuffs) further indicates that the role of classical conditioning
in the 1939b experiment was marginal.
3
Performance in the 50% group of the Humphreys (1939b) experiment was more complex.
During acquisition, responding seemed to
random, i.e., the “yes” and “no” answers had
no relation to the occurrence of the US. Then,
under extinction conditions, participants first
demonstrated an increased probability of “yes”
responses, then they shifted to “no”. Thus,
increased extinction “persistence” seemed to
occur. As there was no way for participants to
reliably predict the occurrence of the US during
acquisition, “yes” and “no” guesses were confirmed and disconfirmed randomly. “Adaptation” to the PRF contingency would therefore
imply “yes” responses in about half the trials.
This is what Humphreys actually observed. In
consequence, the slower reduction of the “yes”
response under extinction was probably caused
by that prior chance relation between the CS
and the US (learned irrelevance). Further, the
PRF contingency implied the presentation of
the US on average every second or third trial. In
light of this, the burst of “yes” responses early in
extinction might reflect the expectation that the
US was due or even overdue (i.e., a gambler’s
fallacy; Tversky & Kahneman, 1974).
In summary, we argue that the verbal “yes”
and “no” responses observed in 100 and 50%
conditions the Humphrey’s 1939b experiment
probably had little to do with conditioning and
hence with the PREE. For the PREE to occur, a
history of conditioning must precede extinction
trials. This premise is not well satisfied in the
Humphreys’ experiment. In the 50% condition
there is little evidence for any specific conditioned response, and in the 100% condition
consequences and strong task demands greatly
facilitated adaptation. Of course, as the PREE
is observed in classical as well as in operant
conditioning (e.g., Mazur, 2004), it is not critical whether the verbal conditioning response is
respondent or operant. It is critical, however,
that the response involved in the demonstration
of the PREE is in fact a conditioned response. If
this premise is not satisfied in the Humphreys
1939b study, this will question the validity of
his study and hence its support of a cognitive
interpretation of the PREE. In the remainder
of this paper we present data to support and
Frode Svartdal and Kjetil Heggelund
4
extend this contention, and to demonstrate that
cognition is not strongly involved in generation
of the PREE.
Experiment 1
The first experiment was designed as a replication of Humphreys (1939b) original study
with a few minor modifications, and its purpose
was to replicate the original experiment and to
confirm our interpretation of the Humphreys’
results, i.e., that responding was strongly determined by confirmatory consequences. Whereas
Humphreys used a within-subjects design
with repeated measures, Experiment 1 applied
a between-groups design in which separate
groups of participants were exposed to the task
under CRF or PRF conditions. The rationale
for this modification was that PREE will occur
more reliably when comparing separate groups
of subjects (e.g., Amsel, 1992, 1994; Capaldi,
1994; Nevin, 1988). Also, in contrast to Humphreys’ original experiment we did not include
the chance series control condition.
Method
Participants. Participants were 27 students
of both sexes (mean age = 24.0 yrs.) from the
University of Tromsø.
Apparatus. We implemented Humphreys’
procedure in software via a web interface (Inquisit Web Edition; www.inquisit.com). The
experiment could then be performed on any
PC with a web browser.
Procedure. Participants first received general
instructions and were then were guided to an
introductory page explaining the procedure
of the experiment and requirements for its
completion. The experimental session started
with random selection of experimental condition, CRF (100% reinforcement) or PRF (50%
reinforcement). All participants then received
brief instructions on the computer screen. Each
trial started with the CS (a green square, 40 x 40
mm) being presented on the computer screen.
The participant was then, by a written message
presented on the screen, requested to guess
whether another stimulus, a red square (same
size and position as the CS), would follow. In
the CRF condition, the US always followed;
in the PRF condition, the US followed the
CS in 50% of all experimental trials. The PRF
schedule was identical for all participants in
that group, and implied a maximum run of 2
of extinction or US trials (the actual sequence
was us, us, ext, us, ext, ext, us, ext, ext, us, us, ext,
us, ext, ext, us, ext, ext, us, ext, us, us, ext, us).
Participants responded in terms of “yes” or “no”,
indicated by the keyboard keys “Y” and “N”.
These guesses (anticipation of the red square)
thus served as “conditioned responses”. The
acquisition phase consisted of 24 trials. Then,
12 extinction trials followed. In this phase, no
US was presented; the CS was presented, and
participants guessed whether or not the US
would occur.
Results and discussion
There were no effects of sex or age, so
these factors were disregarded in the analyses.
The data are presented in the same manner as
Humphreys’ (1939b) investigation to facilitate
direct comparison. Figure 1, left panel, shows
the acquisition data. “Probability of response
YES” was computed as the probability of “yes”
answers over blocks of four trials and subjected
to repeated measures ANOVA. The CRF group
gradually established a functional response with
a high asymptotic level, whereas the PRF group
seemed to answer “yes” or “no” in a random
manner. An overall ANOVA demonstrated a
significant main effect of condition (CRF vs.
PRF), F(1, 25) = 120.64 p < .01, a marginal
effect of trials, F(5, 125) = 2.08, p = .07, and a
significant effect of the condition x trial interaction, F(5, 125) = 10.46, p < .001. The significant
interaction effect indicates that responding
developed differentially over trials (see Figure
1, left panel). Note that whereas learning was
rapid in the CRF group, no learning seemed
to occur in the PRF group. Performance in
this group was random, and if anything the
probability of “yes” answers diminished over
acquisition trials.
The extinction data were analyzed over
the 12 extinction trials. Figure 1, right panel,
shows that the occurrence of “yes” responses
diminished quite rapidly in the CRF condition,
Humphreys Revisited
5
Figure 1. Acquisition (left panel) and extinction (right panel) of the verbal anticipation of the
US in the CRF and PRF groups, Experiment 1. The acquisition data are means of 4 trials, the
extinction data are shown trial by trial. Error bars have been omitted to facilitate direct comparison
to Humphreys’ (1939b) original.
whereas the “yes” responses in the PRF condition first increased and then diminished over the
12 extinction trials. The ANOVA demonstrated
a significant main effect of condition, F(1, 25) =
9.12, p < .001, a significant effect of trials, F(11,
275) = 7.19, p < .001, and a significant effect of
their interaction, F(11, 275) = 4.79, p < .001.
Again, the significant interaction effect indicates
a differential development of responding over
trials between the two groups.
Figure 2 shows the data from Humphrey’s
1939b investigation. These (Figure 2, right
panel, Series I and II) appear to be very similar
to our data. In particular, the CRF data in both
experiments demonstrated a rapid reduction in
the probability of “yes” responses, indicating
complete extinction by the 4th or 5th extinction
trial. Performance in the PRF groups was also
similar in the two experiments, with much
more “persistent” responding compared to the
CRF groups. Two features of the PRF extinction curves must be emphasized. First, note
the peak after 3-5 extinction trials. This burst
of “yes” responses was studied by Williams and
Hartley (1982). They suggested an analysis in
terms of the run lengths of the US. Because
there was no specific learning in the PRF group,
we propose that the peak can be understood
in terms of an increased expectation that the
US will occur, given the no-occurrence of the
US during initial extinction trials (i.e., a gambler’s fallacy). Second, note the relatively steep
extinction curve in the PRF group once the
“yes” responses were discounted. Following 3-4
extinction trials, the power of the confirmatory
feedback for “no” answers seemed to produce an
“extinction” curve that resembles that observed
in the CRF condition.
On the face of it, these data are compatible with the PREE, but as argued it is highly
likely that performance under the arranged
classical conditioning procedure was facilitated
by instrumental contingencies. Such operantclassical interactions are well known from
6
Frode Svartdal and Kjetil Heggelund
Figure 2. The original data from Humphreys’ (1939b) experiment. The left panel shows acquisition
data, the right panel extinction data. The percentages refer to reported expectation among 78
participants as to whether a green light would be followed by a red light. Series I is the CRF
condition (100% reinforcement), Series II is the PRF condition (50% reinforcement). The chance
series is not relevant for the present paper.
animal research (e.g., Brown & Jenkins, 1966;
Williams & Williams, 1969). In the present
case we suggest that the strong confirmatory
100% acquisition contingency established a
response strategy that could be applied also
under extinction conditions, albeit in reversed
form, creating almost perfect adaptation to the
extinction contingency in 3-4 trials. This confirmatory strategy probably also was applied in
the PRF condition following the initial extinction peak. Also note that this peak produced
a number of disconfirmatory outcomes (“yes”
followed by the non-occurrence of the US),
further suggesting the involvement of implicit
instrumental contingencies.
If it is recognized that the conditioning
procedure used in Humphreys (1939b) and
our Experiment 1 involves powerful implicit
instrumental contingencies in the CRF conditions, it would be of great interest to replicate
Humphreys’ experiment without an explicit
response that can be shaped by consequences.
A way of eliminating or reducing the role of
an instrumental contingency in this context
would be simply to omit the emission of explicit
verbalized responses during acquisition trials.
In this way, there would at least be no explicit
response that could be subject to differential
shaping, and a more unbiased performance
during extinction trials must be expected.
Experiment 2
Experiment 2 was a modified replication
of the Humphreys experiment. The effect
of implicit instrumental contingencies was
reduced by instructing participants not to
respond during acquisition trials; responding
only occurred in extinction trials. Participants
had 24 acquisition trials in which the CS was
followed by the US in all (CRF) or 50% of all
trials (PRF). In contrast to Experiment 1 of this
paper, participants now observed the CS-US
presentations but did not respond. They were
instructed to pay attention to the two stimuli
and detect any pattern that might govern their
presentation. Then, in 12 subsequent extinction trials, participants indicated, following the
presentation of the CS, whether or not the US
would occur. As in the previous experiment, the
US was never presented during extinction trials.
We decided to maintain an extinction session
Humphreys Revisited
of similar length to the original Humphreys
experiment.
By omitting the instructed response during
acquisition trials, this procedure is less prone
to the influence of implicit operant contingencies. It is of course possible that implicit
responses and expectations during acquisition
trials might be affected by the US presentations,
but we reason that the effect of such implicit
contingencies should be much less compared
to the explicit shaping that probably occurred
in Experiment 1.
As the experiment involved two phases, an
observation phase (acquisition trials), and then
a response phase (extinction trials), instructions
about this difference could be given at the outset
of the experimental session or in two parts. To
prevent an interruption of the flow of the experiment, we decided to provide all instructions
at the outset of the experiment. However, the
tradeoff with this solution is a possible loss of
initial extinction responses. We considered this
as less of a problem than the interruption with
new instructions would represent.
Method
Participants and apparatus. A total of 40
students of both sexes (mean age = 22.8 yrs.)
at the University of Tromsø participated. The
apparatus was identical to that of Experiment 1
except that the procedure for acquisition trials
now did not involve responding.
Results and discussion
There were no effects of sex or age, so these
factors were disregarded in the analyses. An
overall ANOVA indicated a nonsignificant effect of Condition, F(1, 38) = 0.01, a significant
effect of Trials, F(11, 418) = 4.51, p < .001
and a significant effect of the Condition x Trials interaction, F(11, 418) = 3.72, p < .001.
Figure 3 shows these data. It is apparent that
the effect of extinction conditions in reducing
the probability of “yes” responses was weaker
compared to the results of Experiment 1. This
was seen both in the CRF and PRF conditions.
The significant Condition x Trials interaction
indicates a conventional PREE, but extinction
performance was not significantly different
7
between the groups when the two first extinction trials were excluded from the analysis in
a planned comparison, F(1, 38) = 1.14, p =
.29. The present data thus demonstrate a weak
effect of the CRF vs. PRF conditions in extinction performance; in fact, the main difference
between the groups could be traced to different asymptotic levels when entering extinction
conditions.
Figure 3. Extinction of verbal anticipation of an
US in the CRF and PRF groups, Experiment 2.
E1-E12 refers to individual extinction trials.
It could be argued that the change in procedure from explicit responding to no responding
under acquisition would imply less attention to
the task, and that such reduced overall attention
could explain weaker extinction effects. Note,
however, that initial extinction performance
was quite similar to that of Experiment 1, and
the probability of “yes” response in the two first
extinction trials exceeded .80. When compared
to the results of Experiment 1, these results
seem to indicate a more realistic extinction
performance in that much of the effect of the
implicit instrumental contingencies has been
removed. However, we must assume that even
the procedure of Experiment 2 might involve
implicit instrumental contingencies and thus
exaggerate effects in the direction of a conventional PREE. In reality, therefore, Experiment
2 provides very weak evidence for the existence
of a PREE when a verbalized expectation is used
as the CR.
Frode Svartdal and Kjetil Heggelund
8
Experiment 3
The purpose of Experiment 3 was to repeat
Experiment 2 with one potentially important
modification. Humphreys (1939b), as well as
our Experiments 1 and 2, applied dichotomous
response alternatives, “yes” or “no”. It is possible
that the implicit feedback contingencies could
act in a very direct way on such responding, creating an exaggerated PREE in the Humphreys
(1939b) experiment and in our Experiment 1.
The weakened PREE seen in our Experiment
2 might then more correctly reflect the “net”
PREE when most of the effect of implicit instrumental contingencies has been removed. It
is possible, however, that the weak PREE might
also be attributed to the operationalization of
the response, i.e., that “yes” and “no” answers
do not reflect the more nuanced expectations
that may be generated during acquisition. If
conditioning and extinction depend on beliefs
(Lovibond, 2004), such beliefs must change
gradually to match gradually increasing and
decreasing response rates. Maybe, then, graded
response alternatives could render a more sensitive cognitive measure. We therefore applied
an expectation response that asked participants
to indicate the likelihood for US to occur in
five separate steps, from “very likely” to “very
unlikely”.
Because of the nonsignificant PREE seen in
Experiment 2, we also could suspect that this
null result could be attributable to relatively
low statistical power. We therefore increased
the number of participants for Experiment 3
substantially. This increased n was made possible
because the experimental procedure could be
administered as a web application. Web-based
implementations of experiments may be associated with a high drop-out rates, and they
also lack the stringent control that laboratory
experiments provide (e.g., Birnbaum, 2000).
However, numerous web studies have consistently replicated classic experimental results
(e.g., Eichstaedt, 2002; Hewson & Charlton,
2005), and we reasoned that the relatively
simple procedure of the present experiments
would be well suited for administration over
the Web.
Finally, for this experiment we included
three conditions. In addition to the CRF and
PRF conditions, we included a group receiving 60% reinforcement during acquisition.
The rationale for this was that because the
PRF condition did not indicate any specific
conditioned response in Experiments 1 and 2,
maybe some effect would be observed under
conditions with a slightly increased rate of
reinforcement.
Method
Participants and apparatus. A total of 395
students at the Oslo University College participated. The apparatus was identical to that
of Experiment 2.
Procedure. As in Experiment 2, the acquisition trials did not involve responding. Further,
instead of giving dichotomous responses as to
whether the US would occur, participants now
answered by a number on the keyboard in five
grades of certainty, from “very unlikely” (1) to
“very unlikely” (5). Participants were randomly
allocated to the conditions CRF (100% reinforcement), PRF (50%), and PRF60 (60%).
Data scoring and analysis. Because the acquisition phase of the experiment did not require
any action, transition to the extinction phase
was not observed immediately by all participants. Thus, only 8 participants responded in
the first extinction trial. We therefore decided
to include extinction trials 2-12 in the statistical
analyses. Further, because of a relatively large
number of missing data over trials, we decided
that participants that responded in less than
8 of the extinction trials were excluded from
the analysis. Missing data in a given trial was
replaced by the mean of that variable. Finally,
participants that did not demonstrate any variability in their responses during extinction trials
were excluded from the analysis. This left 225
participants, which indicates a dropout rate not
far from 40%. This is a bit higher than the 30%
rate common for web-based studies (Birnbaum,
2000), but it must be remembered that we used
relatively strict inclusion criteria.
Results and discussion
The overall ANOVA of the extinction data
Humphreys Revisited
indicated a nonsignificant effect of Condition,
F(2, 222) = 2.27, p > .10, a significant effect
of Trials, F(10, 2220) = 20.92, p < .001 and
a significant effect of the Condition x Trials,
F(20, 2220) = 2,62, p < .001. Figure 4 shows
these data. Inspection of Figure 4 indicates that
no difference seemed to appear between the
two PRF conditions. Further, the significant
interaction effect indicates that the CRF and
the PRF conditions developed differently over
extinction trials, which is also apparent in the
figure. As this interaction involved a crossing of
curves early in extinction, the proper analysis
of predicted CRF vs. PRF group difference is
to compare late extinction trials. Thus, a contrast analysis demonstrated that the PRF and
PRF60 groups differed significantly from the
CRF group over the final 8 extinction trials,
F(1, 222) = 10.71, p = .001.
Figure 4. Extinction performance over
extinction trials 2-12 in the CRF, PRF and
PRF60 groups, Experiment 3.
The relatively rapid decrease of the anticipation response in the CRF group compared to the
slower decrease in the PRF groups demonstrates
a conventional PREE. Also note that the anticipation response in the CRF group demonstrates
a level well above the minimum 1 response
option at the end of the extinction trials. This
indicates an extinction rate quite similar to
Humphreys’ original eyelid data (1939b), and
very different from his verbal conditioning extinction data (1939b). Finally, note that even
though the difference between the CRF and
9
PRF groups was significant, the magnitude of
the PREE observed here is not striking.
General discussion
The present paper criticized the study of
Humphreys (1939b), and concluded that the
data originally presented by Humphreys probably did not demonstrate classical conditioning
of a verbalized anticipation response. Further,
Experiment 1 of this report replicated Humphreys’ procedure. The results were similar to
the original data. Analysis of the results in both
experiments supports the view that the “PREE”
observed in these experiments had little to do
with PREE; instead, the verbal anticipation
responses were most probably shaped by immediate consequences (i.e., the presentation of the
US following the verbalized expectations) and
thus involved strong task demands that greatly
facilitated the prediction task. To support this
conclusion, Experiments 2 and 3 tested Humphreys’ hypothesis by means of a procedure
with reduced possibility of the US acting as
a reinforcer to shape an explicit anticipation
response. In both experiments the PREE was
observed, but in a much weaker form compared
to Experiment 1.
To our knowledge, the data from Experiments 2 and 3 are the first to demonstrate that
predictions of the US in classical conditioning, a “verbalized” CR, are different following
CRF vs. PRF reinforcement and conform to
a conventional PREE. However, in spite of
the prediction task being relatively simple, the
group differences over extinction trials were
quite small, and PRF persistence was marked
even in the final extinction trials. This pattern
of results is similar to the eyelid conditioning
data of Humphreys (1939a), but different
from his verbal conditioning data (1939b). We
therefore conclude that Humphreys was correct
in assuming that PREE is demonstrable in the
verbal anticipation response, but that he in fact
did not demonstrate this effect.
In a related set of experiments on predictions
of persistence in operant conditioning, Svartdal
(2003) and Svartdal and Silvera (2008) have
shown that predictions of extinction persistence
10
Frode Svartdal and Kjetil Heggelund
in between-groups manipulations of CRF and
PRF do not conform to a conventional PREE.
Even when independent groups of participants demonstrated a behavioral PREE under
CRF vs. PRF conditions (Svartdal & Silvera,
2008, Experiment 2), corresponding verbalized persistence predictions were insensitive
to the immediately preceding reinforcement
history. These results from operant conditioning contrast the results of Experiments 2 and 3
of this report. We suggest that two differences
between the procedures may explain this apparent conflict. First, in the operant conditioning
procedure, predictions of persistence responses
address the person’s own behavior, whereas in
classical conditioning persistence predictions
address external and often salient events. This
difference relates not only to the difference of
focus during prediction measurement (own
behavior vs. external stimuli), but more generally also to differences in the complexity of
the immediately preceding reinforcement history involving these factors. Thus, the classical
conditioning situation is defined in terms of
repetitions of response-independent CS-US
presentations (in the present case 24 acquisition trials). In contrast, the operant situation
is defined in terms of a response-reinforcer
relation in which the participant’s own behavior during the acquisition phase is part of the
contingencies. Second, although the procedure
of Experiment 3 did not instruct responding
during acquisition trials, it is possible and even
likely that participants established implicit
anticipatory responses that were confirmed or
disconfirmed by consequences. These anticipatory responses were relatively easily translated
to explicit responses during extinction trials. In
contrast, the operant extinction experiments
discussed had no obvious relationship between
the conditioned operant response (key pressing
according to complex rules) and the prediction
measure (answers using a questionnaire). In
sum, these two procedural differences may at
least in part explain the dissimilarity between
these two set of results.
We believe that the extinction performance
seen in Experiments 2 and 3 of this report is
not strongly controlled by explicit cognition.
This follows from the acquisition procedure,
which did not require an explicit verbalized
response during each acquisition trial. The resulting extinction curves, and particularly the
modest differential change as a consequence of
CRF vs. PRF contingencies, speak for control
mechanisms that are not well represented cognitively. If correct, this conclusion in fact corresponds quite well to the previous studies of
operant extinction performance mentioned; the
difference between them would then be one of
degree rather than of principle. The conclusion
that cognition is not strongly involved in the
production of the behavioral PREE implies that
other non-cognitive mechanisms are involved.
Thus, we find it more likely that associationistic
mechanisms (e.g., Amsel, 1994; Capaldi, 1994)
and/or very simple heuristics (e.g., Wong, 1977)
may be responsible for the behavioral PREE.
For example, a simple “win stay, lose shift”
strategy may produce rapid behavioral changes
but not be reflected in explicit contingency beliefs. Clearly, as the present experiments provide
support for the notion that explicit cognition
is not strongly involved in the PREE, further
research must examine the actual mechanisms
that are involved.
Finally, the experimental procedure of
Experiment 3 of this report must be commented. In contrast to a traditional laboratory
experiment, Experiment 3 was implemented
in a web-based system. Such procedures may
be associated with a high drop-out rates, and
they generally lack the stringent control over
participants that a laboratory experiment provides (e.g., Birnbaum, 2000). In Experiment
3, a relatively large drop-out rate was observed.
In addition, we applied strict exclusion criteria.
However, if this affected our conclusions, it
did so by removing noise. Thus, corresponding
analyses with more liberal inclusion criteria indicated no effect of the CRF vs. PRF manipulation. In general, use of the Internet as a research
tool in the social and behavioral science is
becoming well documented (Birnbaum, 2000;
Hewson, Yule, Laurent & Vogel, 2003; Mann
& Stewart, 2000), and numerous web studies have replicated classic results (Eichstaedt,
2002; Hewson & Charlton, 2005). The latter
Humphreys Revisited
conclusion is in accordance with the results of
the last experiment of the present report in that
the results match Humphreys’ (1939a) original
eyelid conditioning results.
References
Amsel, A. (1992). Frustration theory: An analysis
of dispositional learning and memory. Cambridge: Cambridge University Press.
Amsel, A. (1994). Précis on Frustration theory:
An analysis of dispositional learning and
memory. Psychonomic Bulletin & Review,
1, 280-296.
Birnbaum, M. H. (Ed.) (2000). Psychological experiments on the Internet. San Diego:
Academic Press.
Brewer, K. S. (1974). There is no convincing
evidence for operant or classical conditioning in adult humans. In W. B. Weimer & D.
S. Palermo (Eds.), Cognition and the symbolic
process. Hillsdale, NJ: Erlbaum.
Brown, P. L., & Jenkins, H. M. (1968). Autoshaping of the pigeon’s key peck. Journal of
the Experimental Analysis of Behavior, 11,
1-8.
Capaldi, E. J. (1994). The sequential view:
From rapidly fading stimulus traces to the
organization of memory and the abstract
concept of number. Psychonomic Bulletin &
Review, 1, 156-181.
Catania, A. C. (1992). Learning. 3rd Ed. Englewood Cliffs, NJ: Prentice Hall.
Eichstaedt J. (2002). Measuring differences in
preactivation on the Internet: the content
category superiority effect. Journal of Experimental Psychology, 49, 283-91.
Hewson, C. M. Yule, P., Laurent, D., & Vogel.
C. M. (2003). Internet research methods: A
practical guide for the social and behavioral
sciences. London: Sage.
Hewson, C. & Charlton J. P. (2005). Measuring
health beliefs on the Internet: A comparison
of paper and Internet administration of the
Multidimensional Health Locus of Control
Scale. Behavioral Research Methods. 37(4),
691-702.
Humphreys, L. G. (1939a). The effect of random alternation of reinforcement on the
11
acquisition and extinction of conditioned
eyelid reactions. Journal of Experimental
Psychology, 1939, 25, 141-158.
Humphreys, L. G. (1939b). Acquisition and
extinction of verbal expectations in a situation analogous to conditioning. Journal of
Experimental Psychology, 25, 294-301.
Jenkins, W. O. & Stanley, J. C. (1950). Partial
reinforcement: A review and critique. Psychological Bulletin, 47, 3, 193-234.
Lewis, D. J. (1960). Partial reinforcement: A
selective review of the literature since 1950.
Psychological Bulletin, 57, 1, 1-28.
Lovibond, P. F. (2004). Cognitive processes
in extinction. Learning and Memory, 11,
495-500.
Mann, C., & Stewart, F. (2000). Internet communication and qualitative research: A handbook for research online. London: Sage.
Mazur, J.E. (2004). Learning and behavior.
4th Ed. Upper Saddle River, NJ: Prentice
Hall.
Nevin, J. A. (1988). Behavioral momentum and
the partial reinforcement effect. Psychological
Bulletin, 103(1), 44-56.
Shanks, D. R. (2005). Implicit learning. In K.
Lamberts and R. Goldstone, Handbook of
Cognition (pp. 202-220). London: Sage.
Shanks D. R. & Lovibond, P. F. (2002). Autonomic and eyeblink conditioning are closely
related to contingency awareness. Journal of
Experimental Psychology: Animal Behavior
Processes, 28, 38-42.
Svartdal, F. (2000). Persistence during extinction: Are judgments of persistence affected
by contingency information? Scandinavian
Journal of Psychology, 41, 315-328.
Svartdal, F. (2003). Extinction after partial
reinforcement: Predicted vs. judged persistence. Scandinavian Journal of Psychology,
44, 55-64.
Svartdal, F., & Silvera, D. H. (2008). Cognitive
measures of PREE: Predicted persistence
after continuous vs. partial reinforcement.
MS under review.
Tversky, A., & Kahneman, D. (1974). Judgment
under uncertainty: Heuristics and biases. Science, 185, 1124–1131.
Williams, C. D., & Hartley, R. E. (1982).
12
Frode Svartdal and Kjetil Heggelund
Humphreys’ acquisition and extinction of
verbal expectations revisited: Evidence for a
partial reinforcement hump effect. Psychological Reports, 51, 1023-1034.
Williams, D. R., & Williams, H. (1969).
Automaintenance in the pigeon: Sustained
pecking despite contingent non-reinforcement. Journal of the Experimental Analysis
of Behavior, 12, 511-520.
Wong, P. T. P. (1977). Partial reinforcement and
the learning of a strategy in the rat. American
Journal of Psychology, 90, 663-674.