Vous êtes sur la page 1sur 21

When Is Schematic Knowledge Used in Source Monitoring?

Julia Spaniol and Ute J. Bayen


The University of North Carolina at Chapel Hill
Source monitoring involves judgments regarding the origin of information (M. K. Johnson, S. Hash-
troudi, & D. S. Lindsay, 1993). When participants cannot remember the source in a source-monitoring
task, they may guess according to their prior schematic knowledge (U. J. Bayen, G. V. Nakamura, S. E.
Dupuis, & C.-L. Yang, 2000). The present study aimed at specifying conditions under which schematic
knowledge is used in source monitoring. The authors examined the time course of schema-based guesses
with a response-signal technique (A. V. Reed, 1973), and multinomial models that separate memory and
guessing bias. Use of schematic knowledge was observed only when asymptotic oldnew recognition
was low. The time course of schematic-knowledge retrieval followed an exponential growth function.
Implications for theories of source monitoring are discussed.
At some time in the future you may remember reading this
article on source monitoring, but your episodic memory may fail
you when trying to remember in which journal you read it. You do
know that the Journal of Experimental Psychology: Learning,
Memory, and Cognition (JEP:LMC) is a journal that publishes this
kind of paper. Thus, the issues of JEP:LMC on your shelf may be
the first place to look for the article. When determining the source
of information we have encountered in the past, both memory for
the episode in which the information was acquired as well as prior
knowledge related to possible sources come into play. Our re-
search addresses the interplay of episodic memory and prior
knowledge in source monitoring.
Johnson, Hashtroudi, and Lindsay (1993) define source moni-
toring as the set of processes involved in making attributions
about the origins of memories, knowledge, and beliefs (p. 3).
Episodic-memory processes involved in source monitoring have
been studied extensively (for a recent discussion, see Mitchell &
Johnson, 2000). The use of prior knowledge in source monitoring,
however, is less well understood. In this study, we addressed two
questions. First, when does prior knowledge become available
during retrieval? To answer this question, we performed time-
course analyses by means of a response-signal technique (Reed,
1973). Second, how do episodic memory and prior knowledge
interact in source monitoring? We examined this issue by manip-
ulating episodic memory and observing the effect of this manip-
ulation on participants use of prior knowledge. In the next sec-
tions, we review a theoretical framework of source monitoring and
the evidence for the use of prior knowledge in source monitoring,
as well as relevant methodological considerations.
The Source-Monitoring Framework
Johnsons theoretical framework of source monitoring (Johnson,
1997; Johnson et al., 1993; Johnson & Raye, 1981, 2000) has
guided source-monitoring research and informed research in areas
as diverse as cognitive development (e.g., Markham, Howie, &
Hlavacek, 1999; Roberts & Blades, 1998) and cognitive aging
(e.g., Bayen, 1999; Bayen & Murnane, 1996; Trott, Friedman,
Ritter, Fabiani, & Snodgrass, 1999), social cognition (e.g., Klauer
& Wegener, 1998; Sherman & Bessenoff, 1999), eyewitness tes-
timony (for a review, see Lindsay, 1994), cognitive neuropsychol-
ogy (e.g., Dab, Claes, Morais, & Shallice, 1999; Mather, Johnson,
& De Leonardis, 1999), neuroimaging research (e.g., Nolde, John-
son, & DEsposito, 1998; Rugg, Fletcher, Chua, & Dolan, 1999),
schizophrenia (e.g., Franck et al., 2000; Keefe, Arnold, Bayen, &
Harvey, 1999), bilingualism (for a review, see Gerard & Scarbor-
ough, 1989), and consumer research (e.g., Law & Hawkins, 1997).
In the typical source-monitoring paradigm, participants are pre-
sented with items, each of which originates from one of several
sources (speakers, presentation modalities, backgrounds, etc.).
During a subsequent test phase, a mixed list of study items and
new items is presented. Participants are asked to indicate whether
a given test item was presented by Source A, Source B (etc.), or is
new.
According to Johnsons framework (Johnson, 1997; Johnson et
al., 1993; Johnson & Raye, 1981, 2000), source monitoring relies
on two kinds of information, namely, (a) qualities that are char-
acteristic of episodes, and (b) prior knowledge that is the basis for
judgment processes (Johnson et al., 1993). Characteristics of epi-
sodes can vary on multiple dimensions. For example, two sources
Julia Spaniol and Ute J. Bayen, Department of Psychology, The Uni-
versity of North Carolina at Chapel Hill.
This research was presented at the 42nd Annual Meeting of the Psy-
chonomic Society, Orlando, Florida, November 2001. Experiment 2 was
reported in Julia Spaniols masters thesis. The research reported in this
article was supported by National Institute on Aging Grants R01 AG17456
and AG17456-02S1 and by a Social Science Faculty Research Award from
the Institute for Research in the Social Sciences, The University of North
Carolina at Chapel Hill. We appreciate the assistance of Rhenelda H. Lewis
with the development of experimental materials, participant recruitment,
and data collection.
We thank Peter A. Ornstein, Rebekah E. Smith, and Patrick J. Curran for
helpful comments throughout the course of this research. We also thank
Marcia K. Johnson and two anonymous reviewers for their comments on an
earlier version of this article.
Correspondence concerning this article should be addressed to Ute J.
Bayen, Department of Psychology, The University of North Carolina at
Chapel Hill, Davie Hall, CB #3270, Chapel Hill, North Carolina 27599.
E-mail: ubayen@unc.edu
Journal of Experimental Psychology: Copyright 2002 by the American Psychological Association, Inc.
Learning, Memory, and Cognition
2002, Vol. 28, No. 4, 631651
0278-7393/02/$5.00 DOI: 10.1037//0278-7393.28.4.631
631
may differ with regard to perceptual characteristics (e.g., they may
look or sound different), spatial characteristics (e.g., two sources
may be in different corners of a room), temporal characteristics
(e.g., information may have been given today vs. yesterday),
affective characteristics (e.g., one source may evoke stronger emo-
tional reactions than another), or cognitive operations that took
place at the time of encoding. Source-monitoring decisions can
also involve judgment processes that make use of prior knowledge
about sources. Thus, according to this view, source monitoring is
not a purely episodic memory task. Instead, it relies on a combi-
nation of processes that draw on information from the episode and
from prior knowledge. As several authors have pointed out
(Bayen, Nakamura, Dupuis, & Yang, 2000; Johnson et al., 1993;
Mather et al., 1999), prior knowledge is useful in real-life source-
monitoring decisions, because there is often a relationship between
information and its source. For example, you do not have to
retrieve details from a specific episode from memory to judge that
the person who gave you health advice was probably your doctor,
not your hairdresser. However, relying on prior knowledge, source
schemas in particular, might lead to systematic source misattribu-
tions. For example, if your hairdresser did talk to you about your
health and you misattribute the information he gave you to your
doctor, then relying on this information may be costly.
Use of Prior Schematic Knowledge in Source Monitoring
Empirical evidence for the use of prior knowledge in source
monitoring comes from self-reports (Johnson, Foley, Suengas, &
Raye, 1988) and from laboratory experiments (Bayen et al., 2000;
Mather et al., 1999; Sherman & Bessenoff, 1999). The latter
studies systematically manipulated the expectancy of items for the
schema of the presenting source. A schema is a knowledge struc-
ture that organizes associated concepts (Alba & Hasher, 1983).
Bayen et al. (2000) investigated the kinds of cognitive processes
that underlie the use of schematic knowledge in source monitoring.
These authors reported experiments in which the sources were
either different scenes, or persons belonging to different profes-
sional groups. The degree to which the items were expected for the
source schema was manipulated. Half of the items were presented
by the source for which they were expected, and the other half
were presented by the source for which they were somewhat
unexpected. The source-identification results indicated that partic-
ipants used source schemas to make source judgments. However,
empirical measures of source identification cannot disentangle
source memory and source guessing (Murnane & Bayen, 1996).
By estimating multinomial-model-based parameters, Bayen et al.
(2000) were able to support a guessing hypothesis: The differences
in source-identification performance between experimental condi-
tions were due to differential biases when guessing the source,
rather than to improved memory for expected item-source combi-
nations. Participants tended to be biased toward guessing the
source for which the item was expected when they did not remem-
ber the actual source. Thus, the use of schematic knowledge in
source monitoring is reflected in source bias.
Heuristic Versus Systematic Processes in
Source Monitoring
In their source-monitoring framework, Johnson et al. (1993)
distinguish between systematic and heuristic processes when dis-
cussing the mechanisms underlying the use of episodic informa-
tion and prior knowledge in source monitoring. As noted by the
same authors, this distinction is highly similar to the one between
automatic and controlled processes. According to a common def-
inition of automaticity, processes that are characterized as auto-
matic are (a) fast, (b) largely under the control of stimuli rather
than intentions, and (c) require minimal attentional capacities. By
contrast, controlled processes are relatively slow, deliberate, and
attention demanding (e.g., Hasher & Zacks, 1979; Schneider &
Shiffrin, 1977). According to Johnson and Raye (2000), schemas
and other prior knowledge can be used heuristically or systemat-
ically in source decisions. However, no empirical research regard-
ing this issue has been reported to date. Next, we briefly review
research on retrieval of semantic information from memory and
evaluate it with regard to the automatic versus controlled
distinction.
Characteristics of Semantic Retrieval
The literature on memory models contains experimental and
simulation results suggesting that patterns of performance in epi-
sodic and semantic tasks can be explained with the same basic
mechanisms, and that the two share common retrieval dynamics
(e.g., Dosher & Rosedale, 1991; Hintzman, 1984; Shiffrin &
Steyvers, 1997). For example, Dosher and Rosedale (1991) exam-
ined the time course of recognition-memory judgments and that of
semantic-relatedness judgments. Their rationale was that if re-
sponses to the episodic and the semantic task differed in time
course, this would support the view that there are separate episodic
and semantic memory systems. The authors used a response-signal
method to compare the speed of the two types of judgments, and
to look for semantic intrusions in the recognition tasks and for
episodic intrusions in the relatedness task. The time course of the
two judgments was similar, and the authors reported a symmetric
pattern of intrusions at short retrieval intervals. According to
Dosher and Rosedale, these results do not provide support for a
separate-systems view. With regard to source monitoring, this
conclusion would lead one to expect a similar time course for
semantic retrieval as for episodic retrieval.
Other research suggests that semantic knowledge can be acti-
vated and used automatically. Automatic spreading-activation
mechanisms have been among the most popular explanations for at
least some aspects of semantic priming (e.g., Neely, 1991; Posner
& Snyder, 1975). Further, social-cognition research has estab-
lished that stereotypes can be activated automatically and influ-
ence behavior (Banaji & Greenwald, 1995; Banaji & Hardin, 1996;
Blair & Banaji, 1996; Chen & Bargh, 1997; Devine, 1989; Kunda
& Thagard, 1996).
Thus, the mathematical-modeling literature suggests that seman-
tic retrieval is not fundamentally different from episodic retrieval.
Furthermore, semantic-priming and social-cognition studies sug-
gest that semantic information in memory can be activated and
used automatically. However, because of the complex nature of the
attribution processes in source monitoring, it is conceivable that
the retrieval process is less straightforward in source monitoring
than it is in oldnew recognition or relatedness-judgment tasks.
Johnson and Raye (1981) reflect on this complexity when they
propose that episodic-memory processes, on the one hand, and
decision processes based on prior knowledge, on the other hand,
632
SPANIOL AND BAYEN
may be initiated simultaneously (p. 72). However, because de-
cision processes based on prior knowledge are assumed often to
involve more specific information, which may require additional
retrieval from memory, they are presumed normally to take longer
and for this reason might be considered a second stage (p. 72).
Multinomial Models of Source Monitoring
A basic requirement for studying the cognitive processes in
source monitoring is choosing a valid measurement model that
characterizes these processes (e.g., Bayen, Murnane, & Erdfelder,
1996; Johnson, Kounios, & Reeder, 1994; Batchelder & Riefer,
1990). Participant responses in the typical source-monitoring par-
adigm are categorical and, assuming there are two sources in the
study, can take on three values: Source A, Source B, or neither.
Given that each test item originates from either Source A or Source
B, or is new, participant responses can be summarized in a 3
(sources) 3 (responses) matrix (see Batchelder & Riefer, 1990;
Murnane & Bayen, 1996). As several authors have pointed out
(e.g., Batchelder & Riefer, 1990; Bayen et al., 1996; Johnson et al.,
1994), performance in source-monitoring tasks depends on item
recognition (i.e., discriminating between old and new items),
source memory (i.e., discriminating which source presented the
information), and various response biases. These processes cannot
be observed directly, because they jointly contribute to a partici-
pants response on each trial in a source-monitoring test. However,
the contributions of these components to overall performance can
be estimated from the observed data by using a multinomial
processing tree (MPT) model of source monitoring.
MPT models are a class of formal models for categorical data
that assume discrete cognitive states (e.g., item recognition and
source memory). Probabilities of attaining these states are esti-
mated from raw data through maximum-likelihood parameter es-
timation (Batchelder & Riefer, 1999; Hu & Batchelder, 1994;
Riefer & Batchelder, 1988). Parameter-estimation techniques,
goodness-of-fit measures, and strategies for power analysis are
readily available (Hu & Batchelder, 1994; Hu & Phillips, 1999;
Riefer & Batchelder, 1988). MPT models have been applied in a
variety of cognitive research domains (see Batchelder & Riefer,
1999, and Erdfelder, 2000, for recent reviews). MPT models of
source monitoring (Batchelder & Riefer, 1990; Bayen et al., 1996)
permit the separate estimation of item-recognition parameters,
source-memory parameters, and guessing-bias parameters in the
source-monitoring paradigm. The current study used the two-high
threshold (2HT) model of source monitoring (Bayen et al., 1996),
because it has been shown to be superior to alternative multinomial
models of source monitoring. It provides valid and independent
measures of both oldnew item recognition and source memory
(Bayen et al., 1996), and the 2HT assumption for item memory
yields an adequate fit to receiver operating characteristic (ROC)
data (Snodgrass & Corwin, 1988).
Figure 1 shows the four-parameter submodel of the 2HT multi-
nomial model of source monitoring. This is the submodel we used
in the current study; the full, eight-parameter version is shown in
Bayen et al. (1996, Figure 3). In our experimental paradigm, two
sources presented items that were expected for one source and
somewhat unexpected for the other. At test, participants indicated
whether items had been presented by the source for which they
were expected, by the source for which they were somewhat
unexpected, or were new. The first processing tree of the multi-
nomial model in Figure 1 represents a test trial on which the
participant is tested with a target item that was presented by the
source for which it is expected. With probability D, the participant
recognizes this item as old. The participant remembers, with
probability d, that this item was previously presented by the source
for which it is expected. With the complementary probability 1
d, the participant has no source memory for this item and has to
guess the source. With probability g, the participant guesses that it
was presented by the source for which it is expected, and with the
complementary probability 1 g, the participant guesses that it
was presented by the source for which it is somewhat unexpected.
If the participant fails to recognize the item as old (with probability
1 D), the participant guesses that it is a target item with proba-
bility b, or that it is a distractor item with probability 1 b. With
probability g, the participant guesses that the source of an unrec-
ognized item is the source for which the item is expected, and with
probability 1 g, he or she will guess that the source of an
unrecognized item is the source for which the item is somewhat
unexpected. The probability of a given response (expected,
unexpected, or new) to a Source E item is given by the sum of
Figure 1. Four-parameter version of the two-high threshold multinomial
model of source monitoring. E item that is expected for its source; U
item that is somewhat unexpected for its source; N new item; E the
response is source for which the item is expected; U the response is
source for which the item is somewhat unexpected; N the response
is new item; D probability of recognizing an old item or knowing that
a distractor item is new; d probability of remembering the source of an
old item; g probability of guessing that an item was presented by the
source for which it is expected; b probability of guessing that an
unrecognized item is old. Adapted from Source Discrimination, Item
Detection, and Multinomial Models of Source Monitoring, by U. J.
Bayen, K. Murnane, and E. Erdfelder, 1996, Journal of Experimental
Psychology: Learning, Memory, and Cognition, 22, p. 202. Copyright 1996
by the American Psychological Association.
633
USE OF SCHEMATIC KNOWLEDGE IN SOURCE MONITORING
the probabilities from all branches leading up to that response. The
second tree (items presented by the source for which they are
somewhat unexpected) can be interpreted in a similar way. If an
item is new (third tree), the participant either knows it is new with
probability D, or does not, with probability 1 D, in which case
its oldnew status and its source have to be guessed in a similar
way as was described for items presented by the source for which
they are expected.
In this study, we performed time-course analyses based on
multinomial-model parameters estimated for multiple response
times. In the next section, we describe the experimental technique
with which we manipulated response times to obtain these
multinomial-model parameter estimates.
Response-Signal Technique
As suggested by Bayen et al. (2000), the approach chosen here
for studying the nature of retrieval and use of schematic knowl-
edge in source monitoring is a response-signal method (Reed,
1973, 1976), sometimes referred to as speedaccuracy trade-off
procedure. Participants make speeded responses to test stimuli
following a signal to respond. The interval between stimulus onset
and the response signal is systematically varied. The dependent
variable of interestfor example, recognition memoryis plotted
against response time to chart a retrieval time-course function. As
first demonstrated by Reed (1973), the increase in recognition as a
function of response time can be described with an exponential
function shifted to the right, away from zero. Three parameters
characterize this function: an intercept parameter capturing the
point in time after which the dependent measure departs from zero
or chance level; a slope parameter describing the speed with which
the dependent measure increases as a function of response time;
and an asymptotic parameter indicating the level of the dependent
measure reached when response time is unlimited. The current
study examined the time-course functions of two dependent vari-
ables: oldnew recognition and source guessing bias. The validity
of Reeds (1973) exponential time-course function has been estab-
lished for model-based measures of oldnew recognition memory
(e.g., Benjamin & Bjork, 2000; Dosher & Rosedale, 1991; Hintz-
man & Caulton, 1997; Hintzman, Caulton, & Levitin, 1998; Hintz-
man & Curran, 1994, 1997; McElree, Dolan, & Jacoby, 1999;
Mulligan & Hirshman, 1995; Reed, 1973, 1976) and of source
memory (Kinjo, 1999; McElree et al., 1999). However, the time
course of schema-based source guessing bias has not been previ-
ously explored, and one of the goals of the present study was to
determine the shape of the time-course function for such bias.
Combining Multinomial Modeling and the
Response-Signal Technique
Johnson et al. (1994) used a response-signal procedure to ex-
amine the time course of memory processes underlying reality
monitoring, a task that requires participants to discriminate internal
from external sources (Johnson & Raye, 1981). Johnson et al.s
(1994) participants were asked to distinguish between items pre-
viously seen, items previously imagined, and new items. The
authors aggregated responses over participants and fitted a sub-
model of the one-high-threshold (1HT) model of source monitor-
ing (Batchelder & Riefer, 1990) to the aggregate response matrices
for each of four signal lags. The 1HT model differs from the 2HT
model previously described in that it assumes that new items
cannot be detected as new, that is, the decision about the oldnew
status of new items can only be made on the basis of guessing.
After comparing the point estimates of item recognition and source
memory, Johnson et al. (1994) concluded that item information is
available sooner than source information, and that source infor-
mation for imagined events is available sooner than source infor-
mation for perceived events.
McElree et al. (1999) fitted Johnson et al.s (1994) data to the
exponential time-course model and found the results inconclusive
with respect to the question about the relative speed of item
recognition and the two types of source memory. McElree et al.
attributed this in part to the size of the confidence intervals around
the multinomial-model parameter estimates and suggested using a
greater number of preasymptotic response lags. In the present
study, we followed McElree et al.s recommendations by using
seven response lags, and by choosing as high a number of obser-
vations per individual as possible to be able to perform time-course
analyses on individual-participant data. In addition, we fitted the
2HT model of source monitoring (Bayen et al., 1996, see above)
and not the 1HT model, which has been criticized for its incon-
sistency with the shape of empirical ROC curves (Kinchla, 1994).
Probability-Matching Theory
What determines the amount of source bias that people show in
source-monitoring tasks? We propose that source bias depends on
probability matching. This term refers to the phenomenon that
participants match their response biases to the perceived ratio of
different item types at test. Probability matching has been observed
in the context of bias in simple oldnew recognition tasks (e.g.,
Buchner, Erdfelder, & Vaterrodt-Plunnecke, 1995; Healy &
Kubovy, 1978; Ratcliff, Sheu, & Gronlund, 1992; Van Zandt,
2000). Erdfelder and Bredenkamp (1998) found that probability
matching is also involved in source bias. A similar finding was
obtained by Dodson and Johnson (1996). In the context of our
study, probability matching for source bias means that participants
adjust their source bias to the perceived ratio of target items that
were presented by the source for which they are expected to target
items that were presented by the source for which they are some-
what unexpected. That is, the probability of guessing the source for
which a target test item is expected is matched to the perceived
probability that a target test item was presented by the source for
which it is expected. In the current study, the actual ratio of items
that are expected for their source to items that are somewhat
unexpected for their source is 1:1. The better a participants
memory for the study list, the more likely this participant adjusts
his or her guessing to the 1:1 ratio, and accordingly splits guesses
evenly between expected and somewhat unexpected re-
sponses. This behavior is reflected in a guessing parameter g at or
near .5. However, a participant with poor memory for the study list
is less likely to realize the actual ratio of items that are expected for
their source to items that are somewhat unexpected for their
source. We hypothesize that such a participant defaults to a guess-
ing rule that attributes a high percentage of test items to the source
for which they are expected (resulting in a guessing parameter g
that exceeds .5). Such a guessing rule matches the real-world
probability of presentation of items by the source for which they
634
SPANIOL AND BAYEN
are expected. For example, in the real world, the probability that a
weatherman says rain is higher than the probability that a sports
reporter says rain. Thus, participants who cannot match experi-
mental probabilities because of poor memory for the presentation
episode will match probabilities according to their preexisting
schemas.
Prior research in our laboratory (Spaniol & Bayen, 2000) has
been consistent with this theory suggesting that memory and
source bias, although statistically independent when measured
with multinomial-model parameters, may be functionally depen-
dent. In our experiments, schema-based source bias was more
pronounced in older adults than in younger adults, unless younger
and older adults were equated on item memory and source mem-
ory. Thus, source guessing was more prone to influences of sche-
matic knowledge when memory for the study list was low (see also
Mather et al., 1999; Sherman & Bessenoff, 1999). By contrast,
source bias was relatively immune to influences of knowledge
when memory for the study list was high.
The Current Study
The goal of Experiment 1 was to determine the time course of
source bias in source monitoring. We manipulated source, expect-
ancy of items for their source, and response-signal lag to derive
individual time-course curves for multinomial parameter estimates
of item memory and source bias. Experiment 2 used the same basic
paradigm as Experiment 1, and, in addition, included a between-
subjects manipulation of memory. This manipulation allowed us to
test the hypothesis, derived from probability-matching theory, that
schema-dependent source bias is more likely when memory for the
study list is poor.
Experiment 1
The basic procedure, similar to that used by Bayen et al. (2000),
consisted of a source-monitoring test with two sources in which
the semantic relatedness of items and sources was systematically
varied. For example, the word playoffs was presented at study
either by a source for which it was expected (sports reporter) or by
a source for which it was somewhat unexpected (weatherman).
During the source-monitoring test, we presented participants with
target items and distractor items (each of which was also seman-
tically related to one of the two sources), and participants had to
indicate whether these test items had been presented by Source A,
Source B, or were new. We obtained estimates of item recognition,
source memory, oldnew bias, and source bias, as well as
goodness-of-fit indices by fitting the 2HT model of source moni-
toring (Bayen et al., 1996) to the observed response frequencies.
We combined this basic procedure with a response-signal method
in which participants were rewarded for responding within a
certain time limit after the response signals. Participants were told
to try to be accurate, but they were rewarded for timely responses
regardless of accuracy. We chose this payoff scheme to ensure that
participants would be motivated to respond as quickly as possible
after the response signals, and thus to minimize the number of
invalid trials.
We estimated separate sets of multinomial-model parameters for
each individuals performance at each signal lag after pooling the
data over eight studytest blocks. We then performed time-course
analyses for two multinomial-model parametersitem memory
and source bias. In line with the prior studies cited above, we
predicted an increase in item memory as a function of response
time, with the time-course curve following a shifted exponential
function. We also predicted that the time-course curve of source
bias would show a similar shifted exponential form as item mem-
ory. This prediction was derived from experiments discussed
above that showed similar retrieval functions for episodic and
semantic tasks (Dosher & Rosedale, 1991).
We could not make any predictions about the time course of the
source-memory parameter d. Estimating source-memory parame-
ters in the multinomial model requires that item memory be at
medium to high levels. When this is not the case, not enough
observations are available for the estimation of the source-memory
parameter d. As a result, source-memory estimates are very unre-
liable, and unrealistic values of d are often obtained (near 0.00 or
near 1.00). We expected asymptotic item memory to be far below
the ceiling, given the high degree of semantic relatedness of targets
and distractors. At short signal lags in particular, item memory was
expected to be very low. In fact, we chose the length of the shortest
signal lag (150 ms) such that item memory would be at floor level
so that we were able to determine at which point in time item
information becomes available. Given these circumstances, we
could not make any predictions about the source-memory param-
eter d.
The experiments explored questions about the nature of the
processes underlying the use of schematic knowledge in source
monitoring. As mentioned earlier, according to the source-
monitoring framework (Johnson & Raye, 2000), participants in
source-monitoring tasks may draw on schematic knowledge auto-
matically (or heuristically), or in a more controlled fashion
(systematically), but no empirical research exists on this issue.
Obtaining time-course functions for schema-based source bias was
expected to provide crucial information regarding this issue. The
parameter of interest is the intercept of the source-bias time-course
function. Greater speed is associated with a smaller intercept, and
automatic processes are thought to be relatively fast. Thus, the
more automatic a cognitive process, the smaller the intercept of its
time-course function. Although there is no straightforward cutoff
for deciding whether automaticity is present, comparing intercepts
can still be informative. Specifically, we were interested in the
location of the source-bias intercept relative to that of the item-
recognition intercept. For example, if the source-bias intercept was
smaller than or identical to the item-recognition intercept, this
would indicate that schema-based source bias is relatively auto-
matic. However, if the source-bias intercept was greater than the
item-recognition intercept, this would point to a stronger role for
controlled, systematic processing underlying schema-based source
bias.
Method
Participants
We recruited undergraduate students for a pool of potential participants
through flyers posted on The University of North Carolina at Chapel Hill
campus. We conducted brief telephone interviews with interested candi-
dates to screen out students who were not native speakers of English, as
well as individuals who had taken courses in cognitive psychology. Only
right-handed individuals were included in the pool to hold constant poten-
635
USE OF SCHEMATIC KNOWLEDGE IN SOURCE MONITORING
tial confounding effects of handedness on response selection and speed on
the memory test. No color-blind person was included in the pool. From the
pool, we randomly recruited participants for Experiment 1 and for
Experiment 2.
In Experiment 1, 18 undergraduate students (7 men, 11 women) partic-
ipated for payment. Mean age was 19.8 years (SD 1.19). Participants
received $36 and additional payment on the basis of their individual
performance. We excluded the data from one participant because we
learned during the course of the experiment that he was not a native
speaker of English.
Design
The design was a 2 2 7 factorial. Expectancy of items (expected
Source A vs. expected Source B), actual source (Source A vs. Source B),
and response-signal lag (150 ms, 300 ms, 450 ms, 600 ms, 800 ms, 1,000
ms, 2,500 ms) were manipulated within subjects.
Materials
Sources. All participants completed eight studytest blocks. In each of
these blocks, each study item was presented by one of two sources, A or B.
At study, each source was identified by first name and picture. At test, the
two sources were also identified by descriptors, such as weatherman and
sports reporter. We selected the descriptors such that they would activate
well-known schemas. Each was associated with a familiar, distinctive
semantic field, such as weather and sports. The two sources from each
pair shared a common theme. In the current example, that theme was
television news. The Appendix contains a complete listing of source
pairs for the eight blocks.
There were four all-men source pairs and four all-women source pairs
(see Appendix). All names were one-syllable, easy-to-pronounce four-
letter names that are common in the United States, according to the 1990
U.S. Census (mean frequency rank: 360.5, SD 530.30; U.S. Census
Bureau, Population Division, 2000). Within a source pair, names started
with different letters and did not rhyme.
For the pictures, we used color photographs of eight men and eight
women, none of whom were residents of North Carolina and who were
therefore unlikely to be known to our participants. The pictures originated
from public World Wide Web sites and private homepages. Homepage
owners were contacted and consented to the use of their photographs. The
pictures matched the source descriptors with respect to expected age and
appearance. For example, we used pictures of young female college stu-
dents with casual hairstyles and no makeup for the source pair humanities
major and science major, but pictures of middle-aged women with
professional hairstyles, makeup, and jewelry for the source pair doctor
and lawyer. We took care to choose equally distinctive pictures for
Source A and Source B in each block, because similarity between sources
has been shown to affect source memory (e.g., Bayen et al., 1996), and
equal levels of source memory and source bias were desired across blocks.
Therefore, pictures for Sources A and B were required to differ in at least
one salient feature such as hair color, hair length, glasses, facial hair, etc.,
but they could not differ in more than two of these features. All pictures
were brought into a standardized format (200 200 pixels; size on
screen: 3.4 3.4 in.). They showed either only the head or the head and
a small portion of the shoulders.
Items. Items were 1,008 English words from 16 semantic fields asso-
ciated with the 16 sources used in this experiment. There were 63 items
from each semantic field. Some items had been used in previous studies
(Bayen et al., 2000; Bayen, Spaniol, Nakamura, Yang, & Dupuis, 1999;
Spaniol & Bayen, 2000), some were taken from published norms (Battig &
Montague, 1969; Hunt & Hodge, 1971; Shapiro & Palermo, 1970), and
some were generated through Latent Semantic Analysis (Landauer &
Dumais, 1997). The remaining items originated from a variety of sources
such as Web sites, books, and thesaurus searches. The items were chosen
on the basis of four criteria: strong association with Source A and weak
association with Source B or vice versa, low semantic ambiguity, and low
semantic and orthographic similarity to other experimental items. Any item
was excluded that failed to meet these criteria according to the independent
judgments of each of the authors and an additional rater. Because of the
need for a large number of items per semantic category, it was impossible
to control formally for word frequency. However, pilot testing with college
students confirmed that all items were familiar words. Word length varied
between 3 and 13 letters (M 7.06, SD 2.13), the number of syllables
varied between 1 and 6 (M2.33, SD 0.29), and items usually consisted
of one word, rarely of two words (e.g., cold front). Mean word length did
not differ significantly between expected-Source A items and expected-
Source B items in any of the studytest blocks.
We conducted a pilot study with 19 undergraduate students at The
University of North Carolina at Chapel Hill. Its purpose was to ensure that
the studytest blocks produced comparable results with regard to memory
and bias, and that all items were familiar to college students. Each pilot
participant completed one session of four blocks randomly chosen from 10
blocks we had created. The source-monitoring tests in this pilot study were
carried out without performance-related payoffs and without response
signals. We excluded two blocks after pilot testing, because they did not
produce levels of source bias comparable with other blocks.
Procedure
All sessions were run individually. Each participant completed three
sessions, each of which lasted 90120 min. Between 20 and 30 hr passed
between sessions. Participants completed a practice task and two studytest
blocks in the first session, and three studytest blocks in the second and
third sessions, respectively. Material presentation and response collection
were directed by an IBM compatible personal computer. The order of the
eight studytest blocks was determined randomly for each participant.
At the beginning of the first session, participants signed a consent form,
and the experimenter gave them a brief verbal summary of the experimen-
tal protocol. Participants were encouraged to ask questions if they did not
understand the instructions. They were then seated at a computer and began
the practice task.
Practice task. The practice task, a modified version of a lexical deci-
sion task used by Johnson et al. (1994),
1
lasted approximately 45 min and
was administered once, at the beginning of the first session. Its purpose was
to train participants at making speeded three-choice responses under sim-
ilar conditions as in the source-monitoring task. Therefore, response lags,
response keys, and performance-related payoffs were the same as in the
studytest blocks (described below). Participants were informed that de-
pending on their performance, they could earn between $0 and $3.78 in this
part of the experiment. They were presented with a random series of 126
five-letter, two-syllable letter strings. After the list was presented once, it
was repeated twice, both times in a new random order. Half of the letter
strings were words such as river, and half were nonwords such as polef.
Half the words and nonwords appeared in all uppercase letters, the other
half appeared in all lowercase letters. Participants had to decide whether a
letter string was an uppercase word, or a lowercase word, or a nonword.
They were instructed to respond as quickly as possible after the response
signal, a 700-Hz tone. At the end of the practice task, participants were
informed how much money they had earned so far, and they took a 5-min
break.
Source-monitoring task. Participants were presented once with each
study item and were rewarded for speeded responding during the test phase
of each of the eight study-test blocks. There were no rewards for accuracy.
1
We thank Marcia Johnson for generously providing the computer
program for the lexical decision task.
636
SPANIOL AND BAYEN
All instructions for study and test phases were presented on the computer
screen. Participants had as much time to read the instructions as they
wished, and they pressed the space bar to proceed from one instruction
screen to the next. To ensure that participants understood the instructions,
the experimenter gave a brief verbal summary of the instructions before the
study phase and again before the test phase in each block. Participants were
encouraged to let the experimenter know if they had any questions.
Studytest blocks were separated by 5-min breaks. Although participants
were familiar with the procedure after the first studytest block, the
instructions could not be skipped in any of the following studytest blocks.
The study phase of each studytest block consisted of one presentation
of an 84-word study list. In the instructions, participants were informed that
they would see pictures of two people (e.g., Phil and Jack), accompanied
by words that these people say. They were instructed to read the words
carefully and remember them as best as they could, because they would
later be asked to remember the words and to remember who said which
word.
Of the 84 items presented during the study phase of each studytest
block, 42 were expected-Source A items, and 42 were expected-Source B
items. Of each group of items, 21 were presented by Source A and 21 by
Source B. Thus, one half of the study items were presented with the source
for which they were expected, and the other half were presented with the
source for which they were somewhat unexpected.
For purposes of counterbalancing, we randomly divided all expected-
Source A items and all expected-Source B items of each block into three
groups. The assignment of these item groups to sources was counterbal-
anced following the scheme in Bayen et al. (2000, Experiment 1).
Items were presented one at a time, centered at the bottom of the
computer screen in capital letters. Above the item appeared the picture of
the source in the middle of the screen. Below the picture, in a centered
position, appeared the name of the source in capital letters (e.g., PHIL),
followed by a colon to indicate that the word is spoken by the source. The
presentation order of the words was randomized by participant. Each word
appeared for 5 s and was immediately followed by the next word.
On completion of the study phase, participants were given an arithmetic
task to prevent effects of recency in the test phase. They were asked to
count backward by three silently from a number shown on the screen. After
10 s, they were prompted to write the number they had counted to on a
sheet provided by the experimenter.
Immediately after the arithmetic task, instructions for the source-
monitoring test appeared on the computer screen. Participants were given
a description of each source (e.g., It might be helpful for you to know that
PHIL is a WEATHERMAN and JACK is a SPORTS REPORTER). Next,
participants were informed that they would see words presented at the top
of the computer screen. Each time a word would come up on the screen, it
would be their task to decide whether it was a word that had been said by
Source A, by Source B, or by neither. They were instructed to indicate their
answers by hitting color-coded keys on the computer keyboard, and it was
explained which key corresponded to which answer. Participants were
informed that the color and position of the source names and descriptors
(e.g., Phil the weatherman) would match the color and position of the
corresponding response keys.
The instructions reminded participants that they would be making
speeded responses to a beep, as they had done in the practice task. They
were instructed to respond accurately, but to remember that the most
important aspect of this task was to respond as soon as possible after
hearing the beep. Next, participants were reminded of the payment rule.
They would win $0.01 for each response that was neither too fast nor too
slow. It was pointed out that, depending on the participants performance,
he or she could make between $0 and $1.26 on this studytest block.
Participants were asked to rest their left and right index fingers, as well as
the thumb of their dominant hand, on the answer keys. The experimenter
summarized the instructions and asked if the participant had any questions.
After the experimenter initialized the test phase, the test display was shown
for twenty seconds with a GET READY message in the same location in
which the test words would appear. The test display was similar to that
used in Bayen et al. (2000, Experiment 2). In the display, both source
pictures were presented side-by-side on the screen. Below the source
pictures appeared the descriptors along with the source names in 28-point
font capital letters, for example, PHIL the WEATHERMAN and JACK the
SPORTS REPORTER. For each studytest block, half of the participants
saw Source A on the left and Source B on the right, and half of the
participants saw Source A on the right and Source B on the left. The source
names and source descriptors were printed in green (source shown on
left-hand side) and yellow (source shown on right-hand side). Below them
appeared the third response option NEITHER in red letters, centered on the
screen.
One hundred twenty-six test words appeared one at a time. The order of
test word presentations was randomized by participant. Eighty-four of the
126 test words were target items that had been presented during study. The
other 42 test words were new items, half of which were expected-Source
A and half of which were expected-Source B items. Each cell of the 3
(sources: A, B, new) 2 (expectancy: expected-Source A vs. expected
Source B) test design was represented by three items in each lag condition
of each studytest block. With this constraint, for each participant, items
were randomly assigned to signal lags.
The location and color of each response option on the screen corre-
sponded to the location and color of the response key on the computer
keyboard. To attribute a test word to Source A or Source B, participants
either pressed the D key (corresponding to the source shown on the
left-hand side of the screen), which was marked with a green sticker, or
they pressed the K key (corresponding to the source shown on the right-
hand side of the screen), marked with a yellow sticker. To respond neither,
participants pressed the space bar, which was marked with a red sticker.
Participants were instructed to make their responses immediately after the
response signal, a 700-Hz tone of 100-ms duration. All responses made less
than 75 ms after the onset of the response signal elicited a TOO FAST!
message on the computer screen. All responses made more than 300 ms
after the onset of the response signal elicited a TOO SLOW! message on the
screen. Each test stimulus remained on the screen until the participant
pressed one of the three response keys. The intertrial interval was 2.5 s. At
the end of the third session, participants indicated their age and gender on
a questionnaire, then received their payment and were debriefed as to the
purpose of the study.
Results
Following a brief presentation of the response-time data, our
description of the results will focus on the analysis of participants
response choices. In a first step, we estimated multinomial-model
parameters for individual participants raw response-choice data
matrices for each signal lag. In a second step, we submitted the
parameter estimates for item memory and source bias to time-
course analyses to examine how memory and bias changed as a
function of response time. In addition to analyses at the individual-
participant level, we report group averages to capture the typical
time courses of item memory and source bias. An alpha level of
.05 was used for all statistical tests.
Reaction Times
Postsignal reaction time (RT) was defined as the interval be-
tween response-signal onset and response. Following a practice in
the speedaccuracy trade-off literature (e.g., Johnson et al., 1994),
responses that were not made within a 75500-ms window after
onset of the response signal were excluded from further analyses.
The total percentage of excluded responses was 4%. The mean
637
USE OF SCHEMATIC KNOWLEDGE IN SOURCE MONITORING
postsignal RT on nonexcluded trials was 236.50 ms (SE 0.44
ms).
Multinomial Modeling
Before conducting multinomial modeling, we pooled the raw
data in the following way. First, each response was coded with
regard to the expectancy of the test item for the source that had
presented it (expected, somewhat unexpected, or new) and its
expectancy for the source to which the participant attributed it
(expected, somewhat unexpected, or new). Pooling the data over
expected-Source A items and expected-Source B items, we ob-
tained 3 (sources) 3 (responses) frequency matrices. We ob-
tained a separate matrix for each participants responses at each
signal lag. We fitted the 2HT multinomial model of source mon-
itoring (Bayen et al., 1996) to each participants 3 3 response-
frequency matrices, estimating a separate set of parameters based
on each matrix.
The full 2HT multinomial model with eight parameters is not
identifiable, because there are only six degrees of freedom in the
data. Identifiable submodels of the full model are obtained by
imposing equality constraints on parameters. We fitted Sub-
model 4, which is the most parsimonious of the available submod-
els of the 2HT multinomial model (Bayen et al., 1996, p. 202).
This model fit data from prior experiments with procedures very
similar to the ones in the current experiment, albeit without re-
sponse signals (Bayen et al., 2000). Model 4 makes three assump-
tions regarding equality of model parameters. The first assumption
is that item memory is equal for items presented by both sources
and for new items. The second assumption is that source memory
is equal for items presented by both sources. Several schema
theories predict better memory for unexpected than expected in-
formation when a schema is activated at encoding, a prediction that
has found support in empirical research (for a review, see Graesser
& Nakamura, 1982). In our paradigm, however, the schemas were
activated at retrieval. It is, therefore, reasonable to assume that
they did not influence memory for items or sources, a finding that
emerged in a prior study from our laboratory that used procedures
similar to the current one (Bayen et al., 2000). The third assump-
tion of Model 4 is that the probability of guessing for one or the
other source is equal for recognized and unrecognized items. This
assumption has been met in most research studies on source
monitoring involving multinomial models (e.g., Batchelder &
Riefer, 1990; Bayen & Murnane, 1996), including the study that
used procedures similar to the current one (Bayen et al., 2000). We
evaluated whether the assumptions outlined above were met by the
data from Experiment 1 by subjecting Submodel 4 to goodness-
of-fit tests. The power to reject the model with N 144 (144
responses per participant per signal lag) and 2 degrees of freedom
is 0.91, assuming a medium effect size of Cohens (1988) w .3,
as calculated with the GPower program (Erdfelder, Faul, & Buch-
ner, 1996). We obtained parameters through maximum-likelihood
parameter estimation for a total of 119 models (17 Participants 7
Signal Lags) and evaluated model fit with the log-likelihood ratio
statistic G
2
, which is asymptotically chi-square distributed. We
carried out these and all other parameter estimations and goodness-
of-fit tests reported in this article with the MBT program by Hu
(1999). For 113 of the 119 models, Model 4, which has two
degrees of freedom, provided good fit to the data (i.e., G
2
5.99),
indicating that the assumptions outlined above were met by the
data. In six cases, the model did not provide satisfactory fit to the
data, as indicated by a statistically significant G
2
. These cases were
not excluded, because when we test 119 models with an alpha level
of .05, we expect about six of these models to be rejected on the
basis of chance alone. Individual item-memory and source-bias
parameter estimates are shown in Figure 2 (for participants with
source bias) and Figure 3 (for participants without source bias).
2
Time-Course Modeling
For each signal lag, we added each participants median RT for
the responses to the duration of the signal lag to obtain a measure
of the average total response time. We plotted item-memory and
source-bias parameter estimates against response time for the
purpose of analyzing the time course of these cognitive processes.
We determined the relationship between response time and item
memory, and the relationship between response time and source
bias for each individual participant by fitting exponential or linear
models to each participants multinomial-model-based parameter
estimates.
The curve-fitting procedures were implemented using the SAS
REG program to model linear growth, and the SAS NLIN program
to model exponential growth. The exponential-growth model has
the following form (Reed, 1973):
yt 1 e
t
, (1)
for t , else 0.
Equation 1 describes the growth of the dependent measure y
over response time, with y representing item-memory parameter D
or source-bias parameter g. is an asymptotic parameter, charac-
terizing the level of y when response time is not limited. is a
rate-of-rise parameter characterizing the slope in the preasymptotic
section of the curve. is an intercept parameter that reflects the
point in time at which y departs from zero. In the case of item
memory, floor level is D 0, indicating absence of item memory.
In the case of source bias, chance level corresponds to g .5 (i.e.,
a probability of .5 to guess either of the two sources). Because
floor level for this parameter is .5, and not 0, the intercept param-
eter from Equation 1 is not informative with respect to source
bias. To obtain a meaningful intercept parameter for source bias,
we set g equal to .5 in the regression Equation 1. Solving for the
corresponding time t, we obtained the estimated point in time at
which g departs from its floor level.
It should be noted that in time-course analyses such as those
reported in this article, the standard errors but not the point
estimates of the time-course parameter estimates are biased and
should therefore not be interpreted. This is because the standard
regression assumption of independent and identically distributed
(IID) residuals is violated because of the repeated measures char-
acter of the underlying data (Bryk & Raudenbush, 1992).
We report R
2
as a goodness-of-fit measure for the time-course
functions. R
2
is usually adjusted for the number of free parameters,
because adding predictors to a regression equation tends to inflate
2
Interested readers may contact Julia Spaniol or Ute J. Bayen to request
a copy of all multinomial-model parameter estimates and results of
goodness-of-fit tests for individual participants for all signal lags.
638
SPANIOL AND BAYEN
estimates of R
2
(Judd & McClelland, 1989). The adjustment is
reflected in this formula:
adjusted R
2
1

i1
n
y
i
y
i

2
/n k

i1
n
y
i
y
2
/n 1
, (2)
where R
2
is the mean squared variability accounted for by the
model; y
i
is the observed value of the dependent measure (item-
memory parameter D or source-bias parameter g); n is the number
of data points (i.e., the number of lags in the response-signal
experiment); k is the number of free parameters; y
i
is the predicted
value of the dependent measure; and y is the average observed
dependent measure. It should be noted that the greater the
parameters-to-observations ratio, the more difficult it is for R
2
to
take on high values. This affected R
2
s in the present study because
the parameters-to-observations ratio was very high (three param-
eters estimated from seven observations).
Item Memory
Replicating previous work on the time course of memory re-
trieval (e.g., Dosher & Rosedale, 1991; Hintzman & Caulton,
1997; McElree et al., 1999), we fitted Reeds (1973) exponential
time-course model to item-memory estimates. We obtained model
solutions for 16 of the 17 participants. The participant whose D
parameters could not be modeled with the exponential time-course
model showed no evidence of item memory as indexed by D,
regardless of response time. During his postexperiment debriefing,
this participant admitted that he had not tried to be accurate. Table
1 lists estimates of intercept, slope, asymptote, and adjusted R
2
for
all other participants, as well as the arithmetic means of the
parameter estimates. These means describe the typical time
course. When inspecting Table 1, the reader will notice that there
is a great deal of interindividual variability in the parameter
estimates and in goodness-of-fit, as indexed by the adjusted R
2
s.
We will discuss the multiple sources of variability in this experi-
ment and their effect on the outcomes of our analyses in the
General Discussion section. The average time course of item
memory is shown in Figure 4.
Source Bias
Plotting the source-bias parameter g against response time
yielded a mixed picture (see Figures 2 and 3). For some partici-
pants, the plots indicated an asymptotically rising trend, starting
out around g .5, whereas for other participants they appeared to
be random scatter around a straight line at g .5. g is bounded
Figure 2. Item recognition (D) and source bias (g) as a function of response time for the 7 participants in
Experiment 1 who showed source bias.
639
USE OF SCHEMATIC KNOWLEDGE IN SOURCE MONITORING
by 0 and 1, with .5 indicating the absence of bias. To describe the
time-course curves for g formally, we fitted (a) Reeds (1973)
exponential time-course function, and (b) a simple linear function
to each participants source-bias trajectory. Exponential model
solutions could be obtained for the source-bias curves of 7 of
the 17 participants (shown in Figure 2). Estimates of intercept,
slope, asymptote, and adjusted R
2
for each of these participants, as
well as the means of the parameter estimates, are reported in
Table 2. The mean intercept was 410 ms (SE 30 ms), the mean
slope was 8.98 (SE 3.62), and the mean asymptote was 0.71
(SE 0.04). We also fitted a simple linear regression line through
the source-bias estimates for each individual. The mean intercept
for the 10 participants whose bias curves could not be described
with an exponential function (shown in Figure 3) was 0.51
(SE 0.02), and their mean standardized linear slope was 0.02
(SE 0.01), which is not significantly different from zero,
t(68) 0.90. By comparison, those participants whose bias curves
were consistent with an exponential function had a mean linear
intercept of 0.56 (SE 0.07) and a mean standardized slope
of 0.42 (SE 0.02), which is significantly greater than zero,
t(47) 3.20. That is, when described with a simple linear func-
tion, 10 of the participants showed no increase in source bias over
time, whereas the remaining 7 participants showed significant
increase. Figure 5 presents the average linear function for the 10
bias-free participants and the average exponential function for
the 7 participants with significant source bias.
Figure 3. Item recognition (D) and source bias (g) as a function of response time for the 10 participants in
Experiment 1 who did not show source bias.
640
SPANIOL AND BAYEN
The 7 participants R
2
s for the exponential model and the linear
model can only be compared informally, because the two models
are not nested. The mean adjusted R
2
for the exponential model
was .53; the mean R
2
for the linear model was only .15. This
suggests that the exponential model is superior to the simple linear
model at describing the time course of source bias in those 7
participants who showed significant source bias in this experiment.
The results of Experiment 1 indicate that item memory began to
depart from baseline levels at around 390 ms on average, whereas
source bias, if present, emerged at around 410 ms. This result
suggests that the retrieval of prior knowledge (as reflected in
source bias) is somewhat slower than the retrieval of episodic
information (as reflected in item memory). However, because of
the small sample size, it is impossible to know whether this
difference is reliable, and Experiment 2 served in part to increase
the sample size for this comparison.
Participants differed substantially in how much emphasis they
placed on accuracy. Over the course of the three sessions and
during the debriefing, some participants expressed anxiety over
having missed a lot of the answers. These participants, presum-
ably, tried to be as accurate as possible. By contrast, others
indicated that they had paid little or no attention to accuracy,
because rewards were contingent on speed only. A comparison of
the asymptotic item-memory levels showed that unbiased partici-
pants outperformed biased participants in item memory, although
this difference was not significant, D .53, SE .04 versus D
.39, SE .09; t(14) 1.50. As elaborated above, we hypothesized
that there is an inverse relationship between memory and source
bias. In Experiment 2, we investigated this issue systematically by
manipulating memory performance and examining effects on the
time course of source bias.
Experiment 2
In addition to examining the time course of episodic memory
and schema-based source bias, Experiment 2 focused on the rela-
Table 1
Exponential Time-Course Functions for Item-Memory
Parameter D in Experiment 1
Participant (intercept) (slope) (asymptote)
Adjusted
R
2
1 0.26 (0.15) 2.52 (2.06) .35 (.08) .51
2 0.43 (0.06) 1.82 (0.67) .47 (.07) .87
3 0.39 (0.03) 1.94 (0.37) .73 (.05) .96
4 0.36 (0.07) 1.04 (0.30) .58 (.06) .95
5 0.26 (0.33) 2.44 (3.25) .26 (.08) .17
6 0.42 (0.10) 5.45 (8.92) .08 (.03) .50
7 0.44 (0.10) 1.02 (0.59) .29 (.14) .82
8 0.37 (0.07) 2.48 (1.21) .67 (.11) .78
9 0.51 (0.02) 9.31 (4.83) .45 (.05) .77
10 0.46 (0.01) 6.81 (1.75) .74 (.04) .93
11 0.40 (0.02) 4.62 (1.16) .40 (.02) .94
12 0.44 (0.05) 3.31 (1.59) .57 (.09) .80
13 0.41 (0.07) 1.97 (0.89) .56 (.10) .81
14 0.25 (0.17) 1.46 (0.94) .36 (.08) .70
16 0.45 (0.04) 6.19 (2.67) .70 (.06) .82
17
18 0.35 (0.26) 1.24 (1.47) .24 (.11) .35
M 0.39 (0.02) 3.35 (0.61) .47 (.05)
Note. Standard errors are shown in parentheses. Dashes indicate that the
exponential model did not describe the item-memory estimates for Partic-
ipant 17.
Figure 4. Item recognition (D) as a function of response time. The curve
shows the average time course of the D parameters of 16 participants in
Experiment 1.
Table 2
Exponential Time-Course Functions for Source-Bias
Parameter g in Experiment 1
Participant (intercept) (slope) (asymptote)
Adjusted
R
2
3 0.37 (0.09) 2.02 (1.05) .73 (.04) .76
5 0.49 (0.04) 17.94 (20.03) .62 (.03) .59
6 0.39 (0.02) 3.82 (0.68) .87 (.02) .97
7 0.39 (0.10) 4.79 (6.49) .67 (.06) .16
11 0.30 (0.41) 0.84 (1.49) .84 (.13) .28
16 0.49 (0.04) 6.90 (5.03) .67 (.03) .58
18 0.42 (0.05) 26.53 (124.10) .60 (.02) .50
M 0.41 (0.03) 4.79 (3.62) .67 (.04)
Note. The 7 participants whose source-bias estimates could be described
with the exponential time-course model are shown. Standard errors are
shown in parentheses. indicates the time (in seconds) after which source
bias departs from its chance level of .5.
Figure 5. Source bias (g) as a function of response time. The curve shows
the average time course of the g parameters of the 7 participants in
Experiment 1 who exhibited source bias. Also shown is the average time
course of the g parameters of all other participants in Experiment 1.
641
USE OF SCHEMATIC KNOWLEDGE IN SOURCE MONITORING
tive use of episodic information and schematic knowledge in
source monitoring. Johnson and Raye (1981) propose that Which
processes play the predominant role in a particular decision should
depend on such factors as the amount of time a person has,
availability of additional information in memory, and the cost of
mistakes (p. 72). In Experiment 2, we tested these suggestions by
experimentally manipulating available response time (through
response-signal methodology), the number of exposures to the
study materials, and accuracy payoffs. We expected an increase in
the number of exposures to the study materials to increase the
availability of episodic information in memory, thus increasing
reliance on episodic information and increasing asymptotic esti-
mates of episodic memory. With increasing availability of episodic
information, reliance on prior knowledge should decrease accord-
ing to our probability-matching theory, as previously described,
which predicts an inverse relationship between episodic memory
and schema-dependent source bias. With respect to the manipula-
tion of the cost of mistakes, we expected that introducing accuracy
payoffs would also favorably affect episodic-memory estimates.
The expected increase in episodic memory should again decrease
the tendency to rely on schematic knowledge. We expected that
schema-dependent source bias would appear reliably only when
episodic memory was poor.
Each participant was assigned to one of three conditions. Con-
dition 1 was identical to Experiment 1, with only one presentation
per item and no accuracy rewards. There were two presentations in
Conditions 2 and 3. In Condition 2, the same payoff rule was
enforced as in Experiment 1 (speed rewards regardless of accu-
racy). In Condition 3, responses that were made in the legal time
window were only rewarded if they were also accurate.
We chose this design to test the hypothesis of a negative
relationship between schema-based source bias and episodic mem-
ory. It should be noted that our probability-matching theory pre-
dicts a reverse relationship between source bias and source mem-
ory in particular. However, for the same reasons as in
Experiment 1, it was impossible to measure source memory reli-
ably in Experiment 2. We assumed that increasing the number of
exposures to study items and offering accuracy payoffs would
increase source memory and item memory alike, although only the
effects on item memory could be empirically verified.
Method
Participants
From the same pool of potential participants used for Experiment 1, we
randomly recruited 18 undergraduate students6 men and 12 women
from The University of North Carolina at Chapel Hill to participate in
Experiment 2 for payment. The minimum payment was the same as in
Experiment 1. Individuals who had participated in Experiment 1 were
excluded from participation in Experiment 2. The mean age of participants
in Experiment 2 was 20.1 years (SD 0.98). Two participants were
excluded and replaced because of failure to follow the experimenters
instructions.
Design
For Experiment 2, we extended the design of Experiment 1 to include a
between-subjects factor, dubbed experimental group, to manipulate
memory performance. Thus, the design was a 3 2 2 7 mixed
factorial with experimental group as a between-subjects variable (1: one
study presentation, no accuracy emphasis, 2: two study presentations, no
accuracy emphasis, 3: two study presentations, accuracy emphasis), and
expectancy of items for their sources (expected Source A vs. expected
Source B), actual source (Source A vs. Source B) and response-signal lag
(150 ms, 300 ms, 450 ms, 600 ms, 800 ms, 1,000 ms, 2,500 ms) as
within-subject variables. Participants were randomly assigned to the three
experimental groups. Again, three test versions were used. These were
counterbalanced within experimental group.
Materials and Procedure
The same materials were used as in Experiment 1. The procedure varied
depending on experimental group. The procedure in Group 1 was identical
to that in Experiment 1; those for Groups 2 and 3 differed. These differ-
ences will now be described.
Group 2. In this condition, each study word was presented twice
during the study phase of each studytest block. After all 84 words were
presented once in random order, the list was immediately presented again
in a different random order. Each word was presented by the same source
both times. As in Group 1, participants earned $0.01 for each answer that
was neither too fast nor too slow. They were also informed that depending
on their performance, they could make between $0 and $1.26 on each
studytest block. There was no reward for accuracy.
Group 3. As in Group 2, each study word was presented twice during
the study phase of each studytest block. Further, participants earned $0.03
for every answer that was correct and neither too fast nor too slow. This
was enforced in the practice task and in the studytest blocks. Responses
that occurred within the legal time window but were incorrect were not
rewarded. Participants were informed that depending on their performance,
they could make between $0 and $3.78 on each studytest block.
Results
In addition to the analyses reported for Experiment 1, we present
group analyses that tested for effects of the between-subjects
manipulation (number of study presentations/payoffs) on charac-
teristics of the time-course functions.
Reaction Times
As in Experiment 1, responses that did not occur within a
75500-ms window after response-signal onset were excluded
from all analyses. Again, we excluded 4% of responses on the
basis of this criterion. The average postsignal RT on nonexcluded
trials was 240.09 ms (SE 0.42 ms).
Multinomial Modeling
As in Experiment 1, we fitted Submodel 4 of the 2HT multino-
mial model of source monitoring (Bayen et al., 1996) to each
participants raw-data matrix at each signal lag. A total of 126
models were estimated (18 Participants 7 Signal Lags). In eight
cases, the model did not provide satisfactory fit to the data, as
indicated by a significant G
2
. These cases were not excluded
because their number only slightly exceeded the expected number
of models rejected due to chance (at .05, about 6 of 126
models are expected to be rejected on the basis of chance alone).
Individual multinomial-model parameter estimates for item mem-
ory and source bias in Experiment 2 are shown in Figure 6 (for
participants with source bias) and Figure 7 (for participants with-
out source bias).
642
SPANIOL AND BAYEN
Time-Course Analyses
Item memory. As in Experiment 1, we fitted Reeds (1973)
exponential time-course model to the item-memory estimates (pa-
rameter D) of the multinomial model. We obtained model solu-
tions for 17 of the 18 participants. The participant whose item-
memory parameters could not be modeled with the exponential
time-course model was in Group 1. She did not show any discern-
ible growth in D as a function of response time. Table 3 lists
estimates of intercept, slope, asymptote, and R
2
for all other
participants, as well as mean parameter estimates for each exper-
imental group. The average time course of item memory for each
experimental group is shown in Figure 8.
We tested the effectiveness of the between-group manipulations
in a one-way multivariate analysis of variance, with experimental
group as the independent variable and parameter estimates of
item-memory intercept, slope, and asymptote as dependent vari-
ables. There was no significant effect of experimental group on
intercept, F(2, 14) 0.55, MSE 0.02, and slope, F(2,
14) 0.97, MSE 2.69. The effect of experimental group on the
item-memory asymptote, however, was significant, F(2,
14) 11.31, MSE 0.03. We performed Tukey tests to determine
which pairs of means for the item-memory asymptotes were sig-
nificantly different. Group 1 (M.33, SE .10) was significantly
different from Group 2 (M .71, SE .08) and Group 3 (M
.78, SE .03), but the difference between Group 2 and Group 3
failed to reach statistical significance.
Source bias. As in Experiment 1, the source-bias estimates
indicating use of schematic knowledge for source guessing were
rather heterogeneous across participants. Again, some participants
plots showed an asymptotically rising trend, whereas for other
participants, there appeared to be random scatter around a straight
line at or near the baseline level of .5 (see Figures 6 and 7). To
describe the time course for the source-bias parameter g formally,
we again fitted (a) Reeds (1973) exponential time-course function
and (b) a simple linear function to each participants source-bias
trajectory. We obtained exponential model solutions for the bias
curves from 7 participants, 5 of whom were in Group 1, and 2 of
whom were in Group 2. The 2 participants showing bias in
Group 2 also had the lowest asymptotic item memory estimates in
their group. Estimates of intercept, slope, asymptote, and R
2
for
each of these participants, as well as mean estimates for the biased
participants in Groups 1 and 2, respectively, are reported in Ta-
ble 4. The mean intercept was 0.49 s (SE 0.06 s) in Group 1
and 0.45 s (SE 0.06 s) in Group 2. The average time-course
functions for participants with source bias in Group 1 and Group 2,
respectively (exponential time-course functions), as well as for
participants without source bias (linear time-course function), are
shown in Figure 9.
Again, we also fitted linear regression lines through the source-
bias estimates. An informal comparison showed that in the partic-
ipants whose source-bias estimates were consistent with an expo-
nential growth function, mean adjusted R
2
was higher for the
Figure 6. Item memory (D) and source bias (g) as a function of response time for the 7 participants in
Experiment 2 who showed source bias.
643
USE OF SCHEMATIC KNOWLEDGE IN SOURCE MONITORING
exponential model (.57) than for the linear model (.39). This
suggests that the exponential model is superior to the simple linear
model at describing the time course of source bias in the 7
participants who showed source bias in this experiment.
To test formally whether the presence of source bias, that is, the
presence of an exponentially rising time-course curve for g, was a
function of the experimental group, we conducted a chi-square test
on the dichotomous variable source bias present versus not
present, with experimental group as the independent variable.
The test result was significant,
2
(2, N 18) 8.39, indicating
that the higher number of participants with source bias in Group 1
than in the other two conditions was unlikely to result from
random variation.
Comparing the Time Course of Item Memory and Source
Bias
To compare the time-course functions of item memory and
source bias, we pooled the raw data from 12 participants who
showed source bias. We did this separately for each signal lag. Of
Figure 7. Item memory (D) and source bias (g) as a function of response time for the 11 participants in
Experiment 2 who did not show source bias.
644
SPANIOL AND BAYEN
the 12 participants, 7 were in Experiment 1, and 5 were in the same
condition in Experiment 2 (Condition 1). Pooling across experi-
ments was legitimate, because all 12 participants had been ran-
domly sampled from the same pool and tested under identical
conditions. In a series of significance tests in which we set the
source-bias and item-recognition parameters to their baseline lev-
els and evaluated the resulting decrease in model fit, we found that
at the shortest lag (150 ms, median total response time 424 ms),
source bias was not significantly above baseline, g .50, G
2
(1)
.01; but at the next lag (300 ms, median total response time 544
ms), source bias was significantly above baseline, g .54,
G
2
(1) 5.56. However, item memory was already above baseline
at the shortest lag, D .05, G
2
(1) 5.56, and continued to be
above baseline at the next lag, D .09, G
2
(1) 20.13. These
results suggest that the onset of item memory occurred before the
onset of source bias.
Table 3
Exponential Time-Course Functions for Item-Memory
Parameter D in Experiment 2
Participant (intercept) (slope) (asymptote)
Adjusted
R
2
Group 1
1 0.41 (0.02) 1.90 (0.29) .19 (.01) .98
2 0.40 (0.09) 1.36 (0.67) .27 (.06) .81
3 0.41 (0.16) 1.75 (1.69) .22 (.08) .43
4 0.39 (0.06) 4.54 (2.50) .71 (.09) .73
5
6 0.43 (0.04) 4.19 (2.16) .26 (.04) .78
M 0.41 (0.01) 2.75 (0.67) .33 (.10)
Group 2
7 0.13 (1.56) 0.20 (1.57) .43 (2.44) .22
8 0.47 (0.03) 5.42 (1.46) .84 (.04) .94
9 0.42 (0.02) 3.84 (0.65) .87 (.04) .97
10 0.42 (0.02) 3.51 (0.81) .74 (.05) .95
11 0.47 (0.01) 6.57 (1.05) .87 (.03) .97
12 0.37 (0.12) 2.14 (1.35) .53 (.11) .66
M 0.42 (0.10) 3.61 (0.93) .71 (.08)
Group 3
13 0.41 (0.02) 3.67 (0.86) .87 (.06) .95
14 0.42 (0.03) 3.74 (1.18) .67 (.07) .90
15 0.38 (0.02) 3.94 (0.83) .83 (.04) .96
16 0.39 (0.06) 4.18 (2.42) .77 (.11) .70
17 0.41 (0.02) 5.59 (1.90) .74 (.06) .89
18 0.47 (0.01) 3.60 (0.44) .81 (.03) .99
M 0.41 (0.01) 4.12 (0.31) .78 (.03)
Note. Standard errors are shown in parentheses. Dashes indicate that the
exponential model did not describe the item-memory estimates for Partic-
ipant 5.
Figure 8. Item memory (D) as a function of response time. The curves
show the average time courses of the D parameters of 5 participants in
Group 1, 6 participants in Group 2, and 6 participants in Group 3 in
Experiment 2.
Table 4
Exponential Time-Course Functions for Source Bias in
Experiment 2
Participant (intercept) (slope) (asymptote)
Adjusted
R
2
Group 1
2 0.33 (0.33) 5.13 (17.75) .75 (.16) .33
3 0.47 (0.04) 2.37 (0.80) .92 (.06) .89
4 0.71 (0.06) 1.84 (0.70) .62 (.03) .86
5 0.48 (0.16) 1.88 (2.16) .64 (.07) .28
6 0.44 (0.24) 0.57 (0.94) .87 (.34) .57
M 0.49 (0.06) 2.34 (0.75) .76 (.06)
Group 2
7 0.39 (0.03) 2.92 (0.69) .91 (.03) .94
12 0.51 (0.05) 2.46 (1.11) .63 (.02) .80
M 0.45 (0.06) 2.69 (0.23) .77 (.14)
Note. The 7 participants whose source-bias estimates could be described
with the exponential time-course model are shown. Standard errors are
shown in parentheses. indicates the time (in seconds) after which source
bias departs from its chance level of .5.
Figure 9. Source bias (g) as a function of response time. The curves show
the average time courses of the g parameters of those 5 participants in
Group 1 and those 2 participants in Group 2 who showed source bias. Also
shown is the average time course of the g parameters of all other partici-
pants in Experiment 2.
645
USE OF SCHEMATIC KNOWLEDGE IN SOURCE MONITORING
General Discussion
In two experiments, we investigated a set of issues related to the
use of prior schematic knowledge in source monitoring. Prior
research has shown that preexperimental schematic knowledge
about sources influences biases in source guessing (Bayen et al.,
2000). Here, we addressed the following questions: What is the
time course of such source biashow much response time is
necessary for schema-based source bias to appear, and what is the
shape of its time-course function? We also asked whether drawing
on schemas for source guessing is the product of automatic or of
controlled processes. To answer this question, we compared the
time course of source bias with that of item recognition. And
finally, we were interested in the relationship between episodic
memory and schema-dependent source bias. According to our
probability-matching theory, we hypothesized that participants
base their ratio of guesses in favor of the source for which an item
is expected to guesses in favor of the source for which an item is
somewhat unexpected on the perceived ratio of test items that had
been presented by the source for which it is expected to test items
that had been presented by the source for which it is less expected.
When they do not remember enough of the study list to recognize
the actual ratio at test, participants default to the ratios represented
in their prior knowledge.
In our experimental paradigm, two sources presented words that
were expected for one source and somewhat unexpected for the
other. At test, participants indicated whether items had been pre-
sented by the source for which they were expected, by the source
for which they were somewhat unexpected, or were new. We
combined this basic paradigm with a response-signal method in
which participants made speeded responses after time intervals of
variable duration. In both experiments, we estimated individual
time-course functions for item memory and source bias from
parameter estimates based on an MPT model of source monitoring.
In Experiment 1, 7 out of 17 participants showed schema-based
source bias. In those individuals, source bias was characterized by
a shifted exponential time-course function with a slightly delayed
onset as compared with item memory. The participants with source
bias had slightly lower asymptotic item memory than did the
participants without source bias, although this difference was not
significant.
In Experiment 2, we randomly assigned participants to three
conditions that were expected to produce different levels of mem-
ory, the first condition being identical to Experiment 1. The
conditions differed with respect to the number of study presenta-
tions and with respect to the presence of accuracy payoffs at test.
The manipulation had a significant effect on asymptotic item
memory. Participants with source bias were most likely in the
experimental group with the lowest memory performance. This
finding was consistent with our probability-matching hypothesis.
Participants with good memory base their guessing on the ratio of
items presented by the source for which they are expected to items
presented by the source for which they are somewhat unexpected,
whereas participants with poor memory rely on preexperimental,
real-life ratios when guessing, which causes them to exhibit
schema-conform source bias. As in Experiment 1, the onset of
source bias occurred after the onset of item memory.
Item Memory
All but two participants in Experiments 1 and 2 showed shifted
exponential item-memory retrieval functions. On average, item
memory departed from zero at about 390 ms in Experiment 1 and
about 410 ms (Groups 1 and 3) and 340 ms (Group 2) in Exper-
iment 2. Intercepts of 400500 ms are typical in response-signal
studies of item recognition (e.g., Benjamin & Bjork, 2000; Dosher
& Rosedale, 1991; Hintzman & Caulton, 1997; Hintzman & Cur-
ran, 1994). In McElree et al.s (1999) reanalysis of source moni-
toring time-course data from Johnson et al.s (1994) study, the
estimated average intercept for item recognition was 88 ms. This
value is unusually low and may have resulted from some of the
methodological issues pointed out by McElree et al. Increasing the
number of presentations at study was beneficial for item memory
(higher item memory in Group 2 than in Group 1). Offering
payoffs for accurate responses (Group 3) led to an increase in item
memory, but this increase was not statistically significant.
The Time Course of Schema-Dependent Source Bias
Source bias, like item memory, was associated with a shifted
exponential time-course function, with an average onset of 410 ms
in Experiment 1, 490 ms in Group 1 of Experiment 2, and 450 ms
in Group 2 of Experiment 2. The finding that the time courses of
parameters tapping retrieval of episodic versus semantic informa-
tion share a common functional form (i.e., a shifted exponential
function) has not previously been demonstrated in the context of
source monitoring, but it concurs with previous findings in other
literatures. For example, Dosher and Rosedale (1991) used recog-
nition judgments and semantic-relatedness judgments to measure
episodic and semantic retrieval, respectively, and reported that the
time course for both types of retrieval followed a shifted exponen-
tial function. Unlike Dosher and Rosedale, however, we found that
the onset of episodic retrieval (i.e., item memory) preceded the
onset of semantic retrieval (i.e., source bias), and we discuss this
finding in the next section. Another interesting aspect of our
source-bias time-course findings is that participants who showed
source bias did not suppress it when they presumably had enough
time to consciously control their responses (e.g., at the longest
signal lag of 2,500 ms). This suggests two possibilities. Either
reliance on source-relevant schemas was a conscious strategy, or
participants were unaware of their schema-based bias.
Heuristic Versus Systematic Processes
Across the 7 participants in Experiment 1 and 5 participants in
Group 1 of Experiment 2 who showed schema-based source bias,
the mean source-bias intercept appeared greater than the mean
item-memory intercept (440 ms vs. 390 ms). An analysis in which
we pooled response frequencies across participants before estimat-
ing multinomial-model parameters suggested that item memory
indeed became available before source bias. In terms of the dis-
tinction between systematic and heuristic processes within the
source-monitoring framework (e.g., Johnson et al., 1993; Johnson
& Raye, 2000), this result suggests that influences of prior knowl-
edge in the form of schema-based source bias rely more strongly
on systematic processing than does episodic item memory, at least
in our paradigm. This finding does not concur with findings in the
646
SPANIOL AND BAYEN
social-cognition literature (e.g., Banaji & Greenwald, 1995; Banaji
& Hardin, 1996; Blair & Banaji, 1996; Chen & Bargh, 1997;
Devine, 1989; Kunda & Thagard, 1996) and the semantic-priming
(Neely, 1991; Posner & Snyder, 1975) and mathematical-modeling
literatures in cognitive psychology (e.g., Dosher & Rosedale,
1991; Hintzman, 1984; Shiffrin & Steyvers, 1997), all of which
suggest that semantic influences on memory performance can be
automatic, that is, fast and involuntary. A possible explanation for
these contradictory findings is that source monitoring is more
complex than other tasks that have been used to study semantic
retrieval (e.g., lexical decision tasks, relatedness judgments). In
source-monitoring studies, participants typically have several
types of information to consider, and have to choose between three
response alternatives. Biased guessing apparently does not begin
until some episodic information is already available.
Probability-Matching Theory
Across the two experiments, 14 participants exhibited schema-
conform source bias. The number of participants who showed
source bias varied significantly as a function of experimental
condition. Fifty percent of the participants who saw each study
item once only and were not rewarded for accuracy (Experiment 1
and the identical Condition 1 of Experiment 2) showed source bias.
One third of the participants who saw each study item twice and
were not rewarded for accuracy (Group 2 of Experiment 2) showed
source bias. None of the participants who saw each study item
twice and were rewarded for accuracy (Group 3 of Experiment 2)
showed source bias. That is, source bias was present most often in
the condition with the lowest asymptotic item memory, namely in
Experiment 1 and in Condition 1 of Experiment 2. Thus, across
participants, the data supported our prediction of an inverse rela-
tionship between memory and source bias and was hence compat-
ible with our probability-matching theory of source bias. Accord-
ing to this theory, participants adjust their guessing to the ratio of
targets presented by the source for which they are expected to
targets presented by the source for which they are somewhat
unexpected, unless their memory for the study list is not good
enough to recognize this ratio. In this case, they rely on the ratios
in their preexperimental experiences, where items that fit the
schema for a source are most likely presented by that source.
Unfortunately, our sample was too small to allow for statistical
analyses of the relation between asymptotic source bias and as-
ymptotic item memory. Only a crude dichotomysource bias
present versus not presentcould be taken into account. However,
visual inspection of the scattergram in Figure 10, in which asymp-
totic source bias is plotted against asymptotic item memory, ap-
pears to indicate a negative relationship across participants, which
is consistent with our probability-matching theory.
Further, it should be noted that our probability-matching theory
predicts a reverse relationship between source bias and source
memory in particular. For reasons explained in the introduction, a
paradigm in which item memory is expected to be low (which is
necessarily the case with a response-signal paradigm with highly
similar targets and distractors) does not allow for the estimation of
reliable source-memory parameters. Therefore, although we
deemed it reasonable to assume that increasing the number of
exposures to study items and offering accuracy payoffs at test
increase source memory in addition to item memory, a more
comprehensive test of probability-matching theory would require
experimental paradigms that do allow for the estimation of reliable
source-memory parameters. Preliminary studies along these lines
(Spaniol & Bayen, 2000) suggest that source bias most likely
appears when source memory is very low. More stringent tests of
probability-matching theory would also require the manipulation
of the probability of presentation of item-source combinations in
the experiment and according to preexisting schemas.
It is noteworthy that the majority of participants in our study did
not show any source bias. Activation of schemas, with subsequent
use of these schemas for source guessing, involves a complex set
of cognitive processes that is prone to interindividual differences.
As described above, some of these differences can be accounted
for by differences in memory, but we believe that interindividual
differences in metacognitive strategies are another source of vari-
ability. It is reasonable to assume that over the course of the eight
studytest blocks, some participants realized that the expectancy
and the source of test items were never correlated. In some cases,
participants even verbalized their thoughts about this to the exper-
imenter well before the last studytest block. Because of their
insight into the study design, some participants may have chosen
a random-guessing strategy when they did not remember the
source, rather than matching their guesses to preexperimental
probabilities.
Sources of Variability
In both experiments, the R
2
s (see Tables 14) indicate that for
some participants the time-course models explain only a relatively
small proportion of the variance in memory and bias. There are
several reasons for this. As previously mentioned, the
observations-to-parameters ratios were low in this study. There
were 7 observations available to estimate three time-course param-
eters. According to a common rule of thumb, at least 15 observa-
tions are recommended per parameter (Pedhazur, 1997, p. 207).
However, obtaining more observations per parameter would have
required a greater number of signal lags, and hence longer test
lists. As explained in the Method section of Experiment 1, this was
Figure 10. Asymptotic item memory plotted against source bias across
participants in Experiments 1 and 2. For those participants whose source-
bias time course could be modeled with an exponential time-course model,
the estimated asymptote is shown. For the other participants, the mean
source-bias estimates across lags are shown. The regression line is also
shown.
647
USE OF SCHEMATIC KNOWLEDGE IN SOURCE MONITORING
not feasible because of constraints on the construction of the
experimental materials.
Considering the multiple sources of variability in this study, it is
remarkable that we were even able to detect any regularities in the
data. We expectedand obtainedconsiderable interindividual
differences in multinomial-model parameters. Such differences in
memory and bias likely resulted from differences between indi-
viduals in ability, motivation, strategies based on metamemory
assumptions, and compliance. When group multinomial models
are estimated based on raw data that have been pooled across
participants (a common practice in multinomial modeling), heter-
ogeneity among individuals violates the assumption of mutually
independent and identically distributed observations (IID as-
sumption) that underlies MPT models (e.g., Batchelder & Riefer,
1999). Simulation studies have shown that these models are robust
against small amounts of heterogeneity, but robustness does not
hold if individual differences are substantial (Riefer & Batchelder,
1991; Riefer, Hu, & Batchelder, 1994; Riefer & Rouder, 1992).
This was one of the reasons why we chose to estimate separate
multinomial models for each participant, but we did so at the
expense of parameter reliability. That is, we estimated individual-
participant parameters on the basis of relatively few observations.
The number of observations was limited because source pairs and
items had to meet a number of criteria as outlined in the Method
section of Experiment 1. Consequently, we obtained parameter
estimates with relatively large confidence intervals. Some of the
variability in our multinomial-model-based parameter estimates is
therefore due to lack of precision.
Other sources of variability for individual estimates were het-
erogeneity of the experimental materials, variability due to practice
effects and other effects of repeated testing, and trial-to-trial fluc-
tuations in participants cognitive states. These sources of vari-
ability will be explained in turn.
The IID assumption is not only violated in the presence of
interindividual differences, but also in the presence of heteroge-
neous experimental materials. Given that we needed large numbers
of items that fit certain criteria, it was impossible to completely
eliminate all heterogeneity among items. Differences existed in
word length, word frequency, concreteness, interitem similarity,
semantic relatedness of items and sources, and similarity of
sources within and across source pairs.
Finally, the IID assumption is also violated if participants
behavior changes over time, for example, as a function of practice.
It is likely that participants became more practiced and adjusted
their strategies over the eight studytest blocks. In particular, they
may have realized over blocks that in the experiment the ratio of
items presented by Source A and items presented by Source B
among expected-Source A items and expected-Source B items,
respectively, did not correspond to the ratios in the real world.
Such a realization may decrease schema-based bias over blocks.
Such sequential effects were likely a source of IID violations.
In sum, we did our best to minimize violations of the IID
assumption of the multinomial model, but some of these violations
were inevitable and affected the validity and precision of the
multinomial-model based parameter estimates. These, in turn,
served as input for the time-course models. At the time-course
modeling level, there were estimation errors due to the relatively
small number of multinomial-model parameter estimates that each
time-course model was based on, and due to the lack of reliability
of these estimates.
Implications for the Source-Monitoring Framework: How
Prior Knowledge and Episodic Memory Interact
How do we use our knowledge when we make attributions about
specific memories? This question is of central importance in
memory research. Research on schemas (e.g., Alba & Hasher,
1983; Bartlett, 1932; Bransford & Johnson, 1972) has been
groundbreaking in illuminating the interplay of prior knowledge
and episodic memory. In the current study, we examined the
mechanisms involved in the use of prior schematic knowledge and
episodic memory in source monitoring. Source monitoring is a task
that has received a great deal of attention during the past 2 decades,
and research on this topic has integrated diverse theoretical liter-
atures. Examples include dual-process theories of recognition,
cognitive aging, memory for context, encoding specificity, math-
ematical modeling, and the neuropsychology of memory (for a
recent review, see Mitchell & Johnson, 2000). Given the popular-
ity of the source-monitoring paradigm in current memory research
and the ubiquity of source-monitoring demands in real life, it is
crucial that we advance in developing a comprehensive theory of
the cognitive processes involved in source monitoring. Explaining
how prior knowledge influences source monitoring and how it
interacts with episodic memories is an important step in this
direction. Our findings suggest that prior schematic knowledge
affects source monitoring relatively late in processing, after some
information from the episode is already available. Our data also
support the notion that prior knowledge influences guessing only
when episodic memory is poor. Our interpretation of this finding
is that participants who at test remember little about the study list
cannot adjust their guessing to the actual ratio of items presented
by the source for which they are expected to items presented by the
source for which they are somewhat unexpected. Thus, they rely
on the ratios represented in their prior knowledge, that is, they
attribute test items to sources that represent a matching schema.
Several questions remain unanswered. First, is the reliance on
schemas for source guessing better characterized as a conscious
strategy or as unconscious behavior? In other words, are partici-
pants aware of their biased responding? This question should be of
particular interest to researchers studying stereotyping, eyewitness
testimony, and memory development over the lifespan. Second, we
found substantial interindividual differences in the use of sche-
matic knowledge in source monitoring. This is an intriguing find-
ing, because it suggests that some people are prone to schema-
based biases, whereas others are not. We speculated that
participants differed in their metacognitive strategies, but this issue
remains to be addressed in future research. Third, a comprehensive
test of probability-matching theory requires further research on the
relationship between source memory and source bias. Finally, we
have described memory and guessing processes involved in source
monitoring, but much research remains to be done to determine
basic mechanisms underlying these processes, and to describe the
specific interaction of information from various sources (e.g., the
episode, prior knowledge) during the course of remembering.
648
SPANIOL AND BAYEN
References
Alba, J. W., & Hasher, L. (1983). Is memory schematic? Psychological
Bulletin, 93, 203231.
Banaji, M. R., & Greenwald, A. G. (1995). Implicit stereotyping in false
fame judgments. Journal of Personality and Social Psychology, 68,
181198.
Banaji, M. R., & Hardin, C. D. (1996). Automatic stereotyping. Psycho-
logical Science, 7, 136141.
Bartlett, F. C. (1932). Remembering: A study in experimental and social
psychology. New York: Macmillan.
Batchelder, W. H., & Riefer, D. M. (1990). Multinomial processing models
of source monitoring. Psychological Review, 97, 548564.
Batchelder, W. H., & Riefer, D. M. (1999). Theoretical and empirical
review of multinomial process tree modeling. Psychonomic Bulletin &
Review, 6, 5786.
Battig, W. F., & Montague, W. E. (1969). Category norms of verbal items
in 56 categories: A replication and extension of the Connecticut category
norms. Journal of Experimental Psychology, 80, 146.
Bayen, U. J. (1999). Aging and source monitoring of characters in literary
texts. Aging, Neuropsychology, and Cognition, 6, 187200.
Bayen, U. J., & Murnane, K. (1996). Aging and the use of perceptual and
temporal information in source memory tasks. Psychology and Ag-
ing, 11, 293303.
Bayen, U. J., Murnane, K., & Erdfelder, E. (1996). Source discrimination,
item detection, and multinomial models of source monitoring. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 22, 197
215.
Bayen, U. J., Nakamura, G. V., Dupuis, S. E., & Yang, C.-L. (2000). The
use of schematic knowledge about sources in source monitoring. Mem-
ory & Cognition, 28, 480500.
Bayen, U. J., Spaniol, J., Nakamura, G. V., Yang, C.-L., & Dupuis, S. E.
(1999, November). Schematic knowledge about sources is used in
source monitoring. Poster presented at the annual meeting of the Psy-
chonomic Society, Los Angeles.
Benjamin, A. S., & Bjork, R. A. (2000). On the relationship between
recognition speed and accuracy for words rehearsed via rote versus
elaborative retrieval. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 3, 638648.
Blair, I. V., & Banaji, M. R. (1996). Automatic and controlled processes in
stereotype priming. Journal of Personality and Social Psychology, 70,
11421163.
Bransford, J. D., & Johnson, M. K. (1972). Contextual prerequisites for
understanding: Some investigations of comprehension and recall. Jour-
nal of Verbal Learning and Verbal Behavior, 11, 717726.
Bryk, A. S., & Raudenbush, S. W. (1992). Hierarchical linear models:
Applications and data analysis methods. Thousand Oaks, CA: Sage.
Buchner, A., Erdfelder, E., & Vaterrodt-Plunnecke, B. (1995). Toward
unbiased measurement of conscious and unconscious memory processes
within the process dissociation framework. Journal of Experimental
Psychology: General, 124, 137160.
Chen, M., & Bargh, J. A. (1997). Nonconscious behavioral confirmation
processes: The self-fulfilling consequences of automatic stereotype ac-
tivation. Journal of Experimental Social Psychology, 33, 541560.
Cohen, S. (1988). Statistical power analysis for the behavioral sciences
(2nd ed.). Hillsdale, NJ: Erlbaum.
Dab, S., Claes, T., Morais, J., & Shallice, T. (1999). Confabulation with a
selective descriptor process impairment. Cognitive Neuropsychol-
ogy, 16, 215242.
Devine, P. G. (1989). Stereotypes and prejudice: Their automatic and
controlled components. Journal of Personality and Social Psychol-
ogy, 56, 518.
Dodson, C. S., & Johnson, M. K. (1996). Some problems with the process-
dissociation approach to memory. Journal of Experimental Psychology:
General, 125, 181194.
Dosher, B. A., & Rosedale, G. (1991). Judgments of semantic and episodic
relatedness: Common time-course and failure of segregation. Journal of
Memory and Language, 30, 125160.
Erdfelder, E. (2000). Multinomiale Modelle in der kognitiven Psychologie
[Multinomial models in cognitive psychology]. Bonn, Germany: Habil-
itation, University of Bonn.
Erdfelder, E., & Bredenkamp, J. (1998). Recognition of script-typical
versus script-atypical information: Effects of cognitive elaboration.
Memory & Cognition, 26, 922938.
Erdfelder, E., Faul, F., & Buchner, A. (1996). GPower: A general power
analysis program. Behavior Research Methods, Instruments, and Com-
puters, 28, 111.
Franck, N., Rouby, P., Daprati, E., Dalery, J., Marie-Cardine, M., &
Georgieff, N. (2000). Confusion between silent and overt reading in
schizophrenia. Schizophrenia Research, 41, 357364.
Gerard, L. D., & Scarborough, D. L. (1989). Language-specific lexical
access of homographs by bilinguals. Journal of Experimental Psychol-
ogy: Learning, Memory, and Cognition, 15, 305315.
Graesser, A. C., & Nakamura, G. V. (1982). The impact of a schema on
comprehension and memory. In G. H. Bower (Ed.), The psychology of
learning and motivation (Vol. 16, pp. 59109). New York: Academic
Press.
Hasher, L., & Zacks, R. T. (1979). Automatic and effortful processes in
memory. Journal of Experimental Psychology: General, 108, 356388.
Healy, A. F., & Kubovy, M. (1978). The effects of payoffs and prior
probabilities on indices of performance and cutoff location in recogni-
tion memory. Memory & Cognition, 6, 544553.
Hintzman, D. L. (1984). MINERVA 2: A simulation model of human
memory. Behavior Research Methods, Instruments, and Computers, 16,
96101.
Hintzman, D. L., & Caulton, D. A. (1997). Recognition memory and
modality judgments: A comparison of retrieval dynamics. Journal of
Memory and Language, 37, 123.
Hintzman, D. L., Caulton, D. A., & Levitin, D. J. (1998). Retrieval
dynamics in recognition and list discrimination: Further evidence of
separate processes of familiarity and recall. Memory & Cognition, 26,
449462.
Hintzman, D. L., & Curran, T. (1994). Retrieval dynamics of recognition
and frequency judgments: Evidence for separate processes of familiarity
and recall. Journal of Memory and Language, 33, 118.
Hintzman, D. L., & Curran, T. (1997). Comparing retrieval dynamics in
recognition memory and lexical decision. Journal of Experimental Psy-
chology: General, 126, 228247.
Hu, X. (1999). Multinomial processing tree models: An implementation.
Behavior Research Methods, Instruments, and Computers, 31, 689695.
Hu, X., & Batchelder, W. H. (1994). The statistical analysis of general
processing tree models with the EM algorithm. Psychometrika, 59,
2147.
Hu, X., & Phillips, G. A. (1999). GPT.EXE: A powerful tool for the
visualization and analysis of general processing tree models. Behavior
Research Methods, Instruments, and Computers, 31, 220234.
Hunt, K. P., & Hodge, M. H. (1971). Category-item frequency and
category-name meaningfulness (m): Taxonomic norms for 84 catego-
ries. Psychonomic Monograph Supplements, 6, 97119.
Johnson, M. K. (1997). Identifying the origin of mental experience. In
M. S. Myslobodsky (Ed.), The mythomanias: The nature of deception
and self-deception (pp. 133180). Mahwah, NJ: Erlbaum.
Johnson, M. K., Foley, M. A., Suengas, A. G., & Raye, C. L. (1988).
Phenomenal characteristics of memories for perceived and imagined
autobiographical events. Journal of Experimental Psychology: General,
117, 371376.
Johnson, M. K., Hashtroudi, S., & Lindsay, D. S. (1993). Source monitor-
ing. Psychological Bulletin, 114, 328.
Johnson, M. K., Kounios, J., & Reeder, J. A. (1994). Time-course studies
649
USE OF SCHEMATIC KNOWLEDGE IN SOURCE MONITORING
of reality monitoring and recognition. Journal of Experimental Psychol-
ogy: Learning, Memory, and Cognition, 20, 14091419.
Johnson, M. K., & Raye, C. L. (1981). Reality monitoring. Psychological
Review, 88, 6785.
Johnson, M. K., & Raye, C. L. (2000). Cognitive and brain mechanisms of
false memories and beliefs. In D. L. Schacter & E. Scarry (Eds.),
Memory, brain, and belief (pp. 3586). Cambridge, MA: Harvard Uni-
versity Press.
Judd, C. M., & McClelland, G. H. (1989). Data analysis: A model-
comparison approach. San Diego, CA: Harcourt Brace.
Keefe, R. S. E., Arnold, M. C., Bayen, U. J., & Harvey, P. D. (1999).
Source monitoring deficits in patients with schizophrenia; a multinomial
modelling analysis. Psychological Medicine, 29, 903914.
Kinchla, R. A. (1994). Comments on Batchelder and Riefers multinomial
model for source monitoring. Psychological Review, 101, 166171.
Kinjo, H. (1999). Recognition memory vs. source memory: A comparison
of their time-course in a speedaccuracy trade-off paradigm. Unpub-
lished doctoral dissertation, Columbia University, New York.
Klauer, K. C., & Wegener, I. (1998). Unraveling social categorization in
the Who said what? paradigm. Journal of Personality and Social
Psychology, 75, 11551178.
Kunda, Z., & Thagard, P. (1996). Forming impressions from stereotypes,
traits, and behaviors: A parallel-constraint-satisfaction theory. Psycho-
logical Review, 103, 284308.
Landauer, T. K. & Dumais, S. T. (1997). A solution to Platos problem:
The latent semantic analysis theory of acquisition, induction, and rep-
resentation of knowledge. Psychological Review, 104, 211240.
Law, S., & Hawkins, S. A. (1997). Advertising repetition and consumer
beliefs: The role of source memory. In W. D. Wells (Eds.), Measuring
advertising effectiveness. Advertising and consumer psychology (pp.
6775). Mahwah, NJ: Erlbaum.
Lindsay, D. S. (1994). Memory, source monitoring, and eyewitness testi-
mony. In D. F. Ross, R. J. Don, & M. P. Toglia (Eds.), Adult eyewitness
testimony: Current trends and developments (pp. 2755). New York:
Cambridge University Press.
Markham, R., Howie, P., & Hlavacek, S. (1999). Reality monitoring in
auditory and visual modalities: Developmental trends and effects of
cross-modal imagery. Journal of Experimental Child Psychology, 72,
5170.
Mather, M., Johnson, M. K., & De Leonardis, D. M. (1999). Stereotype
reliance in source monitoring: Age differences and neuropsychological
test correlates. Cognitive Neuropsychology, 16, 437458.
McElree, B., Dolan, P. O., & Jacoby, L. L. (1999). Isolating the contribu-
tions of familiarity and source information to item recognition: A time
course analysis. Journal of Experimental Psychology: Learning, Mem-
ory, and Cognition, 25, 563582.
Mitchell, K. J., & Johnson, M. K. (2000). Source monitoring: attributing
mental experiences. In E. Tulving & F. I. M. Craik (Eds.), The Oxford
handbook of memory (pp. 197214). New York: Oxford University
Press.
Mulligan, N., & Hirshman, E. (1995). Speedaccuracy trade-offs and the
dual process model of recognition memory. Journal of Memory and
Language, 34, 118.
Murnane, K., & Bayen, U. J. (1996). An evaluation of empirical measures
of source identification. Memory & Cognition, 24, 417428.
Neely, J. H. (1991). Semantic priming effects in visual word recognition:
A selective review of current findings and theories. In D. Besner &
G. W. Humphreys (Eds.), Basic processes in reading: Visual word
recognition (pp. 264336). Hillsdale, NJ: Erlbaum.
Nolde, S. F., Johnson, M. K., & DEsposito, M. (1998). Left prefrontal
activation during episodic remembering: An event-related fMRI study.
Neuroreport, 9, 35093514.
Pedhazur, E. J. (1997). Multiple regression in behavioral research: Expla-
nation and prediction. Fort Worth, TX: Harcourt Brace.
Posner, M. I., & Snyder, C. R. R. (1975). Attention and cognitive control.
In R. L. Solso (Ed.), Information processing and cognition: The Loyola
Symposium (pp. 5585). Hillsdale, NJ: Erlbaum.
Ratcliff, R., Sheu, C. F., & Gronlund, S. D. (1992). Testing global memory
models using ROC curves. Psychological Review, 99, 518535.
Reed, A. V. (1973, August). Speedaccuracy trade-off in recognition
memory. Science, 181, 574576.
Reed, A. V. (1976). List length and the time course of recognition in
immediate memory. Memory & Cognition, 4, 1630.
Riefer, D. M., & Batchelder, W. H. (1988). Multinomial modeling and the
measurement of cognitive processes. Psychological Review, 95, 318
339.
Riefer, D. M., & Batchelder, W. H. (1991). Statistical inference for
multinomial processing tree models. In J.-P. Doignon & J.-C. Falmagne
(Eds.), Mathematical psychology: Current developments (pp. 313336).
New York: Springer.
Riefer, D. M., Hu, X., & Batchelder, W. H. (1994). Response strategies in
source monitoring. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 20, 680693.
Riefer, D. M., & Rouder, J. N. (1992). A multinomial modeling analysis of
the mnemonic benefits of bizarre imagery. Memory & Cognition, 20,
601611.
Roberts, K. P., & Blades, M. (1998). The effects of interacting in repeated
events on childrens eyewitness memory and source monitoring. Applied
Cognitive Psychology, 12, 489503.
Rugg, M. D., Fletcher, P. C., Chua, P. M. L., & Dolan, R. J. (1999). The
role of the prefrontal cortex in recognition memory and memory for
source: An fMRI study. Neuroimage, 10, 520529.
Schneider, W., & Shiffrin, R. M. (1977). Controlled and automatic human
information processing: Detection, search, and attention. Psychological
Review, 84, 166.
Shapiro, S. I., & Palermo, D. S. (1970). Conceptual organization and class
membership: normative data for representatives of 100 categories. Psy-
chonomic Monograph Supplements, 3, 107127.
Sherman, J. W., & Bessenoff, G. R. (1999). Stereotypes as source-
monitoring cues: on the interaction between episodic and semantic
memory. Psychological Science, 10, 106110.
Shiffrin, R. M., & Steyvers, M. (1997). A model for recognition memory:
REMRetrieving effectively from memory. Psychonomic Bulletin &
Review, 4, 146166.
Snodgrass, J. G., & Corwin, J. (1988). Pragmatics of measuring recognition
memory: Applications to dementia and amnesia. Journal of Experimen-
tal Psychology: General, 117, 3450.
Spaniol, J., & Bayen, U. J. (2000, March). Who said what? Aging and
source memory. Paper presented at the 21st Annual Meeting of the
Southern Gerontological Association, Raleigh, NC.
Trott, C. T., Friedman, D., Ritter, W., Fabiani, M., & Snodgrass, J. G.
(1999). Episodic priming and memory for temporal source: Event-
related potentials reveal age-related differences in prefrontal function-
ing. Psychology and Aging, 14, 390413.
U.S. Census Bureau, Population Division (2000). Frequently occurring
first names and surnames from the 1990 census. Retrieved July 3, 2000,
from http://www.census.gov/genealogy/www/freqnames.html
Van Zandt, T. (2000). ROC curves and confidence judgments in recogni-
tion memory. Journal of Experimental Psychology: Learning, Memory,
and Cognition, 26, 582600.
650
SPANIOL AND BAYEN
Received July 23, 2001
Revision received January 22, 2002
Accepted January 22, 2002
Appendix
Source Pairs
Source A Source B
Name Descriptor Name Descriptor
Phil Weatherman Jack Sports reporter
Beth Science major Gail Humanities major
Jill Doctor Anne Lawyer
Barb Music critic Lynn Art critic
Neil Chef Seth Bartender
Tess Married Kate Divorced
Mark Mountains lover Chad Beach lover
Sean Foreign traveler Mike U.S. traveler
651
USE OF SCHEMATIC KNOWLEDGE IN SOURCE MONITORING