Full text of DOI:10.3758/BF03195775

Memory & Cognition
2002, 30 (6), 893-907
Recognition memory for source and occurrence:
The importance of recollection
JOEL R. QUAMME
University of California, Davis, California
CHRISTINA FREDERICK
University of California, Berkeley, California
NEAL E. A. KROLL and ANDREW P. YONELINAS
University of California, Davis, California
and
IAN G. DOBBINS
Harvard University, Cambridge, Massachusetts
Previous recognition memory studies indicate that when both recollection and familiarity are ex-
pected to contribute to recognition performance (e.g., discriminating studied items from nonstudied
items) the dual-process and the unequal-variance signal detection models provide comparable ac-
counts of performance. When familiarityis not expected to be useful (e.g.,when items from two equally
familiar sources are discriminatedbetween), the dual-process model provides a significantlybetter ac-
count of performance. In the present study, source recognition was tested under conditions in which
familiaritycould have been used to perform a list-discriminationtask; participantswere requiredtodis-
criminatebetweenstrong studied items,weak studied items, and new items. The dual-process model pro-
vided a better account of performance than did the unequal-variance model. Moreover, the results in-
dicatedthat the unequal-varianceassumption in a single-processsignal detectionmodel was not a valid
substitution for recollection and that recollection was used to make recognition judgments even when
assessments of familiarity were useful.
Many classic and current models of recognition mem- distribution that exceeds the response criterion, and the
ory assume that old/new recognitiondecisions (i.e., “was false alarm rate is defined as the proportion of the new-
this item presented before?”) are based on an assessment item distribution that exceeds the response criterion (see
familiarity
of unidimensional memory strength, or
(e.g., Macmillan & Creelman, 1991, for an overview of signal
Banks, 1970; Donaldson, 1996; Green & Swets, 1966; detection theory).
Hirshman & Master, 1997; Inoue & Bellezza, 1998). Test
Dual-process models of recognition (e.g., Atkinson &
items are represented as falling along a familiarity contin- Juola, 1974; Jacoby, 1991; Jacoby & Dallas, 1981; Man-
uum so that studied (i.e., old) items are more familiar, on dler, 1980, 1991; Yonelinas, 1994) assume that a famil-
average, than nonstudied (i.e., new) items. Most models iarity judgment by itself is not sufficient to explain how
incorporate signal detection theory by assuming that recognition judgments are made. These models assume
recollection
variability in the familiarity values of new and old items that an additional process, the
of qualitative
results in partially overlapping normal distributions. Be- information,may also contributeto recognitionjudgments.
cause old and new distributionsoverlap, an individualmust In dual-process models, familiarity or strength has been
select a specific familiaritylevel (i.e., a response criterion) conceptualized as continuous (Atkinson & Juola, 1974;
in order to decide whether or not a test item was presented Yonelinas,1994), automatic (Jacoby, 1991;Jacoby & Dal-
previously. Items falling above the criterion are judged to las, 1981), and item specific (Mandler, 1980, 1991). Con-
be old, and items below the criterion are judged to be new. versely, recollectionis often referred to as a threshold pro-
The hit rate is defined as the proportion of the old-item cess (Rotello, Macmillan, & Van Tassel, 2000; Yonelinas,
1997, 1999a), controlled(Atkinson& Juola, 1974;Jacoby,
1991) and accompanied by the retrieval of contextual as-
sociations (Gardiner, 1988; Jacoby, 1991; Tulving, 1985;
This research was supported by National Institutes of Health Grant
MH59352-01awarded to A.P.Y. We thank Robert Greene, David Riefer,
Yonelinas, 1997, 1999a).
and two anonymousreviewers for insightfuland constructivecomments
The aim of the present paper is to assess the ability of
on earlier versions of this article. Correspondenceconcerning this article
single- and dual-process models of recognitionto account
for source recognition receiver-operating characteristics
should be addressed to J. R. Quamme, Department of Psychology, Univer-
sity of California, Davis, CA 95616 (e-mail: jquamme@ucdavis.edu).
893
Copyright 2002 Psychonomic Society, Inc.

894
QUAMME, FREDERICK, KROLL, YONELINAS, AND DOBBINS
(ROCs) in conditions under which both models predict 1996). However, according to this model, the asymmetry
that familiarity should be useful for making recognition is caused by a disproportionatenumberof highlyconfident
judgments. Participants were required to discriminate hits that represent recollected items. That is, familiarity is
studied target items from both studied lures that differed assumed to reflect an equal-variance signal detection
in strength and new items. This type of task requires the process that alone producesa symmetrical ROC. However,
simultaneous recognition of source and occurrence and some proportion of old items are expected to be recol-
provides unique constraints on how familiarity is used in lected, which has the effect of increasing the hit rate and
the two models to make recognition decisions. Specifi- producing an asymmetrical ROC, like those observed in
cally, we asked whether recollection contributesto perfor- standard recognition tasks.
mance when it is unnecessary (i.e., when familiarity could
The major difference between UVSD and DPSD mod-
be sufficient) and whether unequal variances of familiar- els is that the latter predicts an ROC with a nonzero inter-
ity distributionsare required if recollectionis assumed to cept equal to a proportion of old items that exceed a rec-
contribute.First, we will provide a brief overview of ROC ollection threshold. Unlike the signal detection process
curves in recognition. Then we will describe the use of that assumes Gaussian memory strength distributions,
familiarity-only and dual-process models to account for the thresholdrecollectionprocess assumes noncontinuous
recognizing both occurrence and source.
distributions for which there is some degree of nonover-
lap between new and old distributions (i.e., the strengths
of some old items exceed the highest possible strength of
any new items; see Murdock, 1974, for a more complete
The Use of Receiver-Operating
Characteristics in Model Comparison
A direct method for assessing the predictions of mod- treatment of threshold theory). A threshold process pre-
els and the outcomes of experiments is the examination dicts an ROC with a nonzero intercept that is linear within
of ROCs. An ROC is an empirical function that relates hit the unit square, rather then curved (Murdock,1974;Swets,
rates to false alarm rates in a decision task, such as 1986).1 If familiarity (a signal detection process) and
recognition. ROCs can be generated by having partici- recollection (a threshold process) both contribute to per-
pants make confidence ratings for each recognition judg- formance, as the dual-process model assumes, the ROCs
ment. The confidence continuumoften ranges from very should be asymmetrical and curved, but they should be
new
confident
judgments (i.e., rejections) to very confi- slightly more linear than those predicted by signal detec-
judgments(i.e., endorsements). Hit rates are plot- tion theory alone.
ted against false alarm rates across levels of confidence, Contrastingthe UVSD and the DPSD models by exam-
with the assumption that different levels of confidence re- ining performance in standard recognition test procedures
sult from more or less stringent criterion settings. has been difficult because, in standard recognition tests,
old
dent
For a signal detection model of recognition with equal the DPSD model assumes that both recollection and fa-
variances of old and new distributions,a curvilinearROC miliarity contribute to performance and predicts ROCs
is predicted that is symmetrical about the diagonal and that are almost identical to those predicted by the UVSD
always passes through the coordinates0,0 and 1,1. When model. In some recognitionstudies, the DPSD model pro-
z
plotted in -space, an equal-variance signal detection vides a small advantage over the UVSD model. In other
(EVSD) model predicts a linear function with a slope of studies, the pattern is reversed, and in many, the model fits
1.0. However, findings suggest that recognition memory are not significantlydifferent (see Yonelinas,1997,1999a;
ROCs are generally not symmetrical about the diagonal Yonelinas et al., 1996; for further discussion, see Glanzer,
(Ratcliff, McKoon, & Tindall, 1994; Ratcliff, Sheu, & Kim, Hilford, & Adams, 1999; Yonelinas, 1999b).
Gronlund, 1992). Specifically, in probability space, the
However, the two models do make different predic-
y
ROC function tends to be “pushed”toward the -axis, and tions under conditions in which the contribution of rec-
z
in -space, the slope is usually less than 1.0, suggesting ollection is relatively large and that of familiarity is rel-
that the variance of old items is greater than the variance of atively small. Under these conditions, the DPSD model
new items. Consequently, most unidimensional signal de- predicts that the ROC should not deviate much from lin-
tectiontheoriesof recognitionnow incorporatean unequal- earity. For example, relatively linear ROCs have been
variance assumption, and some global-matching models found for recognition tasks in which individualsdiscrimi-
explicitlypredict greater old-item variance than new-item nate between intact and recombined arbitrary associations
variance (e.g., Gillund & Shiffrin, 1984;Hintzman, 1986). between studied words (Kelly & Wixted, 2001; Rotello
Thus, an unequal-variancesignaldetection(UVSD) model et al., 2000; Yonelinas, 1997), among facial features
can overcome one of the major deficienciesof the EVSD, (Yonelinas,Kroll, Dobbins, & Soltani,1999), and between
in that it is compatible with the ROCs reported in most source or list contexts (Yonelinas, 1999a). These types of
experiments.
tasks require individuals to retrieve information about
Yonelinas(1994) developedthe dual-processsignal de- the studied association between two items or between an
tection (DPSD) model, which combines a familiarity as- item and its context. That relatively linear ROCs are ob-
sessment based on signal detection theory with a thresh- served for many of these tasks is evidence that a contin-
old recollection process. The DPSD model also predicts uous signal detection process does not contribute to the
ROCs that are curved and asymmetrical along the diago- recognition of arbitrary associations. One interpretation
nal (Yonelinas, Dobbins, Szymanski, Dhaliwal, & King, of these results is that the familiarity of an item–item or

RECOGNITION MEMORY FOR SOURCE
895
an item–context pairing is not increased beyond those of familiarity is a unidimensional signal detection process
the individualitems and contextsthemselves(i.e., the con- and that when familiarity is used in a recognitiontask, par-
stituentparts). That is, whereas item information becomes ticipants are able to control response criteria only, and not
more familiar with study, there are cases in which associa- the actual computationof familiarity (see Clark, 1999, for
tive and source information does not (see Kelly & Wixted, a discussion of this assumption). For a familiarity-only
2001, for a similar view).
signal detection model of this sort to account for list or
Although a number of experiments have shown that source discrimination, it must be assumed that the famil-
the UVSD model accountspoorly for the shapes of source iarity distributions of items from the two lists have dif-
ROCs, a linear component is not always sufficient to ac- ferent means, different variances, or both. Specifically, an
count for the shape of the ROC. This has been demon- item can be judged as having come from a target list if the
strated directly with source ROCs when items from two familiarities of target items are generally lower or higher
sources differ in strength (Yonelinas, 1999a, Experi- than those of items from the otherlist. Similar familiarity-
ment 4), when encoding procedures in the study phases only models of list discrimination have been proposed to
are relatively rich (e.g., Qin, Raye, Johnson, & Mitchell, account for performance in experiments in which the
2001), or when multidimensionalrating scales are used to strength of the items from two studied lists was expected
combineitem recognitionand source recognitionin a sin- to differ (Greene, 1999; Maddox & Estes, 1997).
gle task (e.g., Slotnick, Klein, Dodson, & Shimamura,
Of particular interest is whether recollection is used
2000). These findings suggest either that recollection is when items from different sources have different levels
not always well described as a threshold process or that of familiarity. It would seem that individuals should use
familiarity can sometimes be used to discriminate among recollection if distributionsoverlap greatly and if perfor-
items from different sources. Regardlessof whether either mance based on familiarity alone is expected to be low.
or both of these possibilities hold, any adequate model However, this would equally apply for item recognition.If
must account for both linear and nonlinear ROCs and recollectionis not used to discriminatestudied items from
must provide empirically verifiable predictions of recol- nonstudied items that differ in strength, why should it be
lection and familiarity estimates across experimental used to discriminatestudied targets from studied lures that
conditions.To accountfor bothlinearand nonlinearROCs, differ in strength?
A
single-process signal detection models must assume that
the threshold process does not contribute to performance
Suppose, for example, that two study conditions, and
B
serve to increment the familiarity of items to different
A
under most circumstances (i.e., when ROCs are curvilin- levels. Specifically, the familiarity of items after study
switch
B
A B
,
ear) but that participants occasionally
to a thresh- is higher, on average, than the familiarityof items. If the
old process when continuous strength information is un- participants are later shown a mixed list of
, and new
helpful. Whereas it is generally accepted that a threshold items and are instructed to endorse only those items from
process of some sort is required to account for linear one source, a familiarityassessment should be able to sup-
ROCs, it is less clear whether a threshold process such as port the discrimination of targets from lures.
recollection regularly contributes to performance when
ROCs are nonlinear (Glanzer et al., 1999). It has been items from List (i.e., strong targets) and to reject all
If, for example, the instructions are to endorse only
A
A
B
B
considered previously that although recollection may be items from List (i.e., weak lures) and all new items,
used in some recognition tasks, it might not be used on the decision can be made by simply accepting all items
recognition tests when it is not required or when it pro- with familiarities that fall above a criterion (see Fig-
vides information redundant with that provided by fa- ure 1A). In this situation, discriminating between items
miliarity (e.g., an inclusion task in the process dissocia- from two lists is similar to discriminating between old
tion procedure; Clark, 1999; Jacoby, 1991; Ratcliff, Van and new items, because a familiarity value below the cri-
Zandt, & McKoon, 1995). A critical issue is whether it is terion is partially diagnostic of both new items and items
reasonableto assume that recollectiondoes not contribute from the lure source. If confidence ratings are also ob-
to recognition when familiarity should be sufficient to tained,it can also be assumed that they are arranged along
discriminatebetween targets and lures. In other words, do the familiarity axis, as is shown in Figure 1A, with highly
individualsuse both recollectionand familiarity any time confident endorsementson the high-familiarityend (e.g.,
both processes provide useful mnemonic information “6”) and highly confidentrejectionson the low-familiarity
aboutthe study phase (as is assumed by dual-processmod- end (e.g., “1”).
els), or do they use recollection only when familiarity is
However, if participants are to endorse items from the
B
insufficient (as is assumed by a single-process switching less familiar source (i.e., weak targets), they must dis-
view)? To evaluate these positions, it is first necessary to criminate target items not only from new items with lower
define what conditions must be met for familiarity to familiarities, but also from old lures with higher familiar-
A
support source discrimination.
ities (i.e., strong lures). In this case, a unidimensional
signal detectionprocess with one response criterion is in-
sufficient to produce old/new discriminationand list dis-
crimination simultaneously.
Source Memory Discrimination
Based on Familiarity
Here, we consider a simple form of familiarity-only
One possiblesolutionis to assume that individualsadopt
recognition.For the present purposes, it was assumed that two sets of response criteria, as is shown in Figure 1B. In

896
QUAMME, FREDERICK, KROLL, YONELINAS, AND DOBBINS
Figure 1. Hypothetical familiarity curves for strong studied items, weak stud-
ied items, and new items. Panels show (A) a test situation in which items from
the stronger source are targets and participants use a single criterion to dis-
criminate strong targets from weaker studied and nonstudied lures and (B) a
test situation in which items from the weaker source are targets and partici-
pants use two criteria, a lower criterion that is used to discriminate studied tar-
gets from weaker nonstudied lures and an upper criterion that is used to dis-
criminate weak studied targets from strong studied lures.
this model, a low criterion is chosen, below which items terion. Thus, each confidence point on the ROC repre-
are rejected as new, and a high criterion is chosen, above sents the placement of two criteria along the familiarity
which items are selected as having come from the non- axis.
target source. An endorsement is given to items with fa-
A B
/
miliarities above an old/new criterion but below an
Source Memory Discrimination
source criterion. Similar two-criterion strength models
have been proposed to account for source discrimination
tasks, such as reality monitoring (Hoffman, 1997; Marsh
& Bower, 1993).2 The two-criterion model has the impli-
cation that highly familiar items can be rejected if partic-
ipants are given instructionsonly to endorse items from a
less familiar source. Confidence ratings would not be
arranged from low to high familiarity going left-to-right,
as in Figure 1A, but rather as increasing and then decreas-
ing, as in Figure 1B. Specifically, the most confident en-
dorsement (i.e., “6”) would be given not to all items
above a high level of familiarity, but to items that fall
within an intermediate familiarity range. Less confident
endorsementsare given to items both aboveand below the
Based on Recollection and Familiarity
The DPSD model can accountfor source discrimination,
whether or not familiarity contributes to performance, by
assuming that an item’s source can also be recollected.
A
Items can be judged as being from Source if that infor-
mation is recollected,even if the familiaritydistributionof
A
B
Source items is no higher or lower than that of Source
items. The DPSD equationsfor predictingresponse prob-
abilities for items from a target source, a lure source, and
new items will be given below. Consider first a discrimi-
nation task in which one must endorse only items from a
strong source. The probabilityof identifying a strong tar-
get item as a target (hit rate) is equal to the probability
R
that it is recollected, s, plus the probability that it is
A B
/ criteria are reached, and
range, until the old/new and
then more and more confidentrejectionsare given to items
A B
F
judged to be familiar, s , in the absence of recollection:
P
R
+
s
2 R F
falling below the old/new criterion and above the / cri-
(“target” strong target) =
(1
) . (1)
s s
|

RECOGNITION MEMORY FOR SOURCE
897
The probability of a false alarm to an item from the weak target source are weak and those from the lure source are
source (i.e., the lure source) is equal to the probability strong (see Figure 1B). When targets have intermediate
F
that the lure is judged to be familiar, w, if the item is not strength relative to other items, endorsements (i.e., com-
R
recollected,
:
ing from the target distribution) are given to items with
w
C
familiarity values above a lower criterion, i , but below
P
2 R F
) .
w w
(“target” weak lure) =
(
1
(2)
|
C
an upper criterion, j. Thus, the probability of endorsing
an item under these circumstances is the area of the fa-
Importantly, Equation 2 assumes that recollection of a
studied item is used to reject the item, whereas in Equa-
tion 1, recollectionis used to accept the item. By contrast,
items judged to be familiar are accepted in both equa-
tions. Note also that the model assumes that false alarms
are always the result of incorrect familiarity assessments
and that there is no provision for false recollection (e.g.,
recollecting that a weak source item came from a strong
source). The probability of a false alarm to a new item is
C
C
and j. Equations 4, 5,
miliarity distribution between
and 6 are modified accordingly,iso that
P
=
-
(“target” weak target)
|
¢
d
Cj
é
ê
æ
ç
ö
¢
d -
Sw
C
C
æ
i ö
w
w
Rw
+
-
R w
*
F
- F
1
(
,
,
(7)
)
ç
è
÷
ø
÷
Sw
è
ø
ê
ë
P
=
(“target” strong lure)
|
¢
d
-
Cj
é
æ
ç
ö
¢
d
-
Ss
æ
i ö
s
F
also equal to the probability that it is judged familiar,
provided it is not recollected,
,
s
n
-
Rs * F
- F
1
(
( 8)
(9)
)
ê
ç
è
÷
ø
÷
R
:
Ss
è
ø
n
ê
ë
P
2 R F
) .
n n
(“target” new) =
(1
(3)
and
|
P
F 2C 2 F 2C
( ) ( ).
j
i
(“target” new) =
Although in most cases participants cannot truly recol-
lect that an item was not presented, there may be circum-
stances under which qualitative information may be used
to reject new items (see Rotelloet al., 2000, and Yonelinas,
1997, for discussions). For simplicity, n was assumed in
the present experiment to be zero, leaving the false alarm
|
The equations represent a hybrid of both UVSD and
DPSD models inthat theycontainall the parametersof both
models. The UVSD and DPSD models make equivalent
predictions if all the recollection terms are zero (
w = .00) and if all the familiarity distributions have the
R
R
=
s
R
F
rate for new items equal to
.
n
S
S
w = 1.00). In this case, the
same standard deviation (
=
model reduces to a standarsd EVSD model. Under DPSD
Familiarity is assumed to be well described by signal
detection theory, so the probability that an item of any
type will be judged familiar is equal to the area under-
neath thepart of its familiaritydistributionthatexceeds the
R
assumptions,the standarddeviationterms are equal, but
and w are free to assume any value between zero and ones.
R
R
Free estimation of for target items allows the ROC to in-
tercept the ordinate between 0,0 and 0,1. Free estimation
of for lures allows the top end of the ROC to intercept
the upper abscissa between 0,1 and 1,1. was originally
incorporated to account for the asymmetry commonly
observed in ROCs (see Yonelinas, 1994), but if both
parameters are equal, the ROC predicted by the DPSD
model is symmetrical. Under UVSD assumptions, recol-
lection terms are fixed at zero, but the standard deviations
P F . C
response criterion, ( i ). Familiaritydistributionsare
assumed to be normal, and the mean of the new-item dis-
tribution is scaled to zero with unit variance. Thus, the
criterion settings, as well as the means of target and lure
distributions, are equal to their respective distances from
zero in standard deviation units (i.e., -scores) of the new
distribution. The means of the strong and the weak dis-
tributions are expressed as s and w, respectively. Equa-
tions 1, 2, and 3 can be rewritten accordingly(i.e., as Equa-
tions 4, 5, and 6, respectively), so that the probability of
judging any type of item to come from the target distrib-
ution is expressed as [(
area of normal distribution that exceeds a response cri-
R
R
R
z
d
¢
d¢
S
S
w) are free to
of old-item familiarity distributions (
,
s
assume any value. Accordingly, the ROC will become
“pulled” toward the axis of the distributionwith the larger
standard deviation. However, the function will always
pass through 0,0 and 1,1, given that recollectionterms are
zero.
Proponentsof both models have argued that the unique
assumptions of the other are unnecessary. Statistically,
however, the DPSD and the UVSD models cannot be di-
rectly compared, because they are not nested models. For
example, when moving from the UVSD model to the
F
d
¢
C S
i )/ y], the cumulative
–
yy
C
k
terion,
(one of criteria used to construct the ROC,
2, . . . , k ), given the mean,
standard deviation, (1, s, or w), of the distribution:
i
C
C
C
d¢ d¢
(0, s, or w), and
y
d¢
,
1
S
S
S
y
P
=
(“target” strong target)
|
¢
d
-
Ss
C
æ
i ö
s
Rs
+
-
-
Rs * F
1
,
(4)
(
)
ç
è
÷
ø
S S
DPSD model, two parameters must be fixed (
,
w), and
s
R R
two parameters must be freed (
, w). However, because
s
¢
d
-
C
Sw
æ
i ö
w
P
=
* F
recollection and signal detectionparameters are indepen-
dent and serve different functions, there is no a priori rea-
son why dual-process and unequal-variance assumptions
cannot both be made by the same model. Indeed, an indi-
(“target” weak lure)
1
R w
, (5)
(6)
(
)
ç
è
÷
ø
|
P
= F -Ci
(“target” new)
.
(
)
|
The familiarity terms for these equations are somewhat rect way to compare two models statistically is to com-
different for predicting responses when items from the pare both to a less restricted hybrid model that containsall

898
QUAMME, FREDERICK, KROLL, YONELINAS, AND DOBBINS
the free parameters of the other two. The two models of counted for, the fit of the hybrid should not be an im-
interest can be derived from the hybrid model by fixing provement over that of the DPSD model. Conversely, if
values of critical parameters. For example, a hybrid of the unequal variances are a necessary feature of the familiar-
DPSD and the UVSD models would embody both dual- ity process, the fit of the hybrid should be an improve-
R
R
w) and ment over that of the DPSD model.
process assumptions (i.e., free parameters
,
s
S
unequal-variance assumptions (i.e., free parameters
,
s
S
w); the UVSD model is derived by fixing recollection
METHOD
parameters to .00, and the DPSD is derived by fixing dis-
tribution variances to 1.00. Then, by comparing the
UVSD and the DPSD models with the hybrid, it can be
determined to what degree the parameter restrictions
specific to each model are unacceptable and serve to
worsen the fit. Specifically, a comparison of the hybrid
Participants
Thirty-six undergraduates from the University of California,
Davis participated for course credit.
Materials
Two hundred forty facial sketches of famous people were se-
with the UVSD model tests whether there is an improve- lected from the Corel Draw Clip Art Portrait Dictionary. Faces were
chosen that were judged likely to be familiar and recognizable to
ment in fit by adding recollection if the variances are al-
most college students. The faces chosen generally represented the
ready allowed to be unequal. A comparison of the hybrid
areas of politics, business, movies and television, sports, and intel-
with the DPSD model tests whether there is an improve-
lectualism. Faces were divided into three sets of 80 each. One set
was presented in each of two study phases, and the faces in the re-
ment in fit by allowing variances to be unequal when rec-
ollection already contributes to performance. Note that
any hybrid model of this sort must contain more free pa-
rameters than either of the critical models and, as such, is
less parsimonious than the others. Thus, fitting this type
of model requires more than the usual number of ROC
points.
maining set were used as nonstudied lures in the recognition test.
Across participants, there were six counterbalanced assignments of
sets to conditions and four different presentation orders for each of
the assignments. Thus, across participants, each face appeared
equally often in all three phases. All the faces were presented to the
participants on a 15-in. color computer monitor.
In the present study, an attempt was made to deter-
mine the relative necessity of unequal-variance and dual-
process assumptions in simultaneously accounting for
recognition of source and occurrence when familiarity
should be sufficient to make both types of discrimina-
tions. Specifically, we examined the degree to which par-
ticipants would use recollection to discriminate target
items from both studied lures and nonstudieditems when
the familiarity of target items differed from both types of
lures. For all the participants, two study lists were pre-
sented with different encoding instructions so that items
from the two lists differed in strength. However, at test,
one group of participants was required to endorse only
items from the stronger source, whereas the other group
Procedure
All the participants received two study phases and a single test
phase. In the first study phase, the participants viewed 80 faces and
made rapid gender judgments about each of them (i.e., the weak
list). Gender judgments were expected to constitute a shallow en-
coding task and lead to relatively weak encoding, because a correct
response required only the processing of structural features. The z
and / keys were used to respond male and female, respectively. The
participants were told to take enough time to correctly classify each
face as male or female but to use no more than the time required to
respond correctly. They were also instructed to respond carefully,
because some male faces had long hair and some female faces had
short hair.
In the second study phase, the participants viewed a second set
of faces and judged the area in which they felt the person was most
famous (i.e., the strong list). The participants were instructed to
was required to endorse only items from the weaker classify each face according to which of seven categories that per-
son best represented. Response categories included business fig-
source. Thus, the participants in the two conditionswere
ures, leaders (political, military, spouse of leader), performers
expected to use different decision criteria, but the actual
(actor, singer, model), media personalities (TV host, newscaster),
memory parameters for strong and weak lists were not
sports figures (athlete, coach, sportscaster), and intellectuals (sci-
entist, author, artist). The participants placed each face in one of
expected to differ across conditions. DPSD and UVSD
models were each fit to data from the two conditions si-
multaneously, and critical model parameters were con-
strained to be identical across conditions. Fits of the two
models were examined to determine which one provided
the best overall account of performance, and both models
were compared with a hybrid model to determine the con-
sequences of removing recollection or unequal-variance
assumptions. If recollection is unnecessary for source
recognition when familiarity can be used, the fit of the
hybrid should not improve relative to that of the UVSD
model. Conversely, the fit of the hybrid should improve
relative to that of the UVSD model if recollection is used
for source recognitioneven when familiarity is sufficient.
If equal variances of familiarity distributions are suffi-
the seven categories by pressing a corresponding number on the
keyboard. They were instructed to classify each face only if they ac-
tually knew who the person was, and not to simply guess on the
basis of some other information (e.g., the person’s attire). The par-
ticipants were told that they did not need to know an individual’s
name but that they needed only to recognize that person as fitting
into one of the categories. If, however, they did not know the per-
son, they pressed the “8” key. A table with the assignment of cate-
gories to numbers was always presented along with a face. As much
time was allowed as was necessary for this decision phase. Follow-
ing each classification response, the participants were told whether
or not they were correct and were given a brief description on the
computer screen of why the person was well known. On the basis
of this information, they were again asked whether they knew the
correct classification. The participants were free to modify their
first responses at this point before continuing to the next face. Fame
cient for the familiarity process once recollection is ac- judgments and the subsequent elaborate feedback procedure were

RECOGNITION MEMORY FOR SOURCE
899
squared-error term that is computed only from variabil-
ity of the hit rate and assumes that false alarm rates are
perfectly known. That is, because the false alarm rate is
treated as a predictorvariable, the best-fitting ROC func-
tion minimizes only the squared deviations from the hit
rate. Both hit and false alarm rates contain error vari-
ance, and thus, we used a minimization function that re-
flected variability of both hit and false alarm rates. The
model fitting was performed using the solver function in
Table 1
Response Frequencies for Six Levels of Confidence
and Three Test Lists in Two Conditions
Response
Condition
Item
6
5
4
3
2
1
Total
Weak targets strong
91 126 119 149 206 749 1,440
252 274 196 208 206 304 1,440
93 144 253 261 640 1,440
Strong targets strong 613 298 168 127 118 116 1,440
weak
new
49
weak
new
117 127 131 243 365 457 1,440
47
61
98 266 360 608 1,440 Microsoft Excel 7.0 (see Dodson, Prinzmetal, & Shima-
mura, 1998; Yonelinas, 1994). The solver employs an ef-
ficient search algorithm that can be used to minimize the
value of a specified cell in an Excel spreadsheetby chang-
ing the values of other specified cells. The value in the
to-be-minimized cell can be defined as a function of the
values in to-be-changed cells. In this manner, the solver
expected to constitute a relatively deep encoding task that should
lead to relatively strong encoding, because the participants’ in-
volvement in the task was greater than in the first encoding phase.
Following the two study phases, all the participants were given a
recognition memory test. The test consisted of all 160 previously
studied faces, along with 80 new faces. Half of the participants were
instructed to endorse only the faces for which they had previously
made sex judgments (weak target condition) and to reject both non-
studied faces and faces for which they had made fame judgments
(i.e., both strong and new lures). The other half of the participants
were instructed to endorse only the faces for which they had made
fame judgments (strong target condition) and to reject both new
faces and faces for which they had made sex judgments (i.e., both
weak and new lures). Thus, the conditions differed according to
which list the participants were instructed to endorse. Each face was
rated according to how confident the participant was that it had ap-
peared in the target phase (sex or fame), with ratings ranging from
1 (certain that it did not appear in the target phase) to 6 (certain
that it did appear in the target phase). Responses of 1 or 6 indicated
that the participant had complete confidence, 2 or 5 indicated mod-
erate confidence, and 3 or 4 indicated low confidence. The partici-
pants were further instructed to spread out their responses so that
they used the entire range.
SSE
was used to minimize a total
term that was the sum
SSE
the
s of all the axes. The function to be minimized
was defined as
3
5
2
2
ˆ
SSE
P - P
ijk ijk
=
,
(10)
å åå
(
)
i
j
k
P
where the squared deviations between observed (
)
ijk
ˆ
P
and predicted ( ijk ) response probabilities are summed
i = j =
over
3 item types (strong, weak, and new items),
5 criterion settings (cumulated confidence levels), and
k
= 2 experimental conditions. Thus, the models were fit
to the observed data from the strong- and the weak-target
SSE
conditions simultaneously by minimizing an overall
term that was summed not only across confidence levels,
but also across the three types of items and the two con-
ditions. Because each participant contributed 5 response
probabilities for each type of item, there were 15 aver-
aged probabilities from each of the two conditions, for a
ANALYSIS
SSE
total of 30 data points. Thus, the overall
term was a
Each participantcontributed240 responses distributed
among 18 bins. Table 1 shows the total frequency count
summed across participantsfor each of the 18 bins in each
of the two conditions. Each frequency bin is associated
with one of the six confidence judgments made for faces
in one of the three lists (strong, weak, and new). Re-
sponse frequencies were converted to probabilities for
five different criterion settingsby cumulatingacross con-
fidence levels for each list. Performance at the most con-
servative criterion setting was estimated as the proportion
of faces receiving only the most confident endorsements
(i.e., only those faces receiving a 6). Probabilities for the
second most conservative setting were computed by
adding the proportion of faces receiving a response of 5
to the proportion of faces receiving a response of 6. Suc-
cessively less conservative settings were computed in this
manner, so that 15 endorsement probabilities were ob-
tained from each participantthat corresponded to five cri-
terion settings for each of the three types of faces.
sum of 30 squared errors between observed and model-
generated probabilities. This procedure may be best con-
ceptualizedas two simultaneousthree-dimensionalregres-
term reflects deviations of five
predicted points from five observed points on each of the
-, -, and -axes (target source, lure source, and new items,
respectively) in each condition.
The solver was used to find the best-fitting values of
SSE
sions for which the
x y
z
C
the signal detectioncriterion settings ( i ), as well as crit-
R
R
d¢ d¢ S S
, , , , w). Model-
w s w
ical model parameters (
,
generatedresponse probabsilitiesfor strong,sweak, and new
items for each level of confidence were computed using
the same criterion values. However, a different criterion
setting was estimated for each level of confidence within
a given condition, and different settings were estimated
for the two conditions. Thus, each criterion setting was
estimated to generate a predicted response probability
(i.e., a confidence point) on all three axes of one of the
two conditions. Note, however, that equality constraints
were imposed on the critical model parameters across the
two conditions.For example, regardless of whether test in-
Fitting the Models
ROCs are often assessed by examining the fits of re- structionsrequested the endorsement of strong fame items
gression models that predict hit rates from false alarm or weak sex items, it was assumed that all the weak sex
rates. This procedure is problematic in that it yields a items would have the same average familiarity across test

900
QUAMME, FREDERICK, KROLL, YONELINAS, AND DOBBINS
instructions and that the stronger fame items would have are used, rather than deviations from a grand mean. That
the same average familiarity across test instructions. is, each model fit is assessed by the degree to which it ex-
x
y
Equalityconstraintswere made on the critical parameters plains the sum of six separate variance estimates ( -, -,
z
of both models.
and -axes in the two conditions),rather than the total vari-
The DPSD and UVSD models both contained 19 free ability among the 30 points. This procedure was deemed
parameters. Of these, 17 were shared by the two models. appropriate given the multidimensional nature of the data
d
¢
Two parameters were values of for strong and weak (i.e., a separate variance measure was obtainedfor each di-
items. The other 15 free parameters shared by all the mod- mension). Using deviationsfrom a grand mean gave larger
R2
els representedthe criterionsettingsfor thetwo conditions.
Criterion settings were assumed to be under participant isons were nearly identical for the two procedures.
control and were allowed to vary across the two condi- As was discussed earlier, a direct comparison of
estimates, but significance tests for model compar-
R2
tions. Whereas only1 setting was assumed to be necessary values for the DPSD and the UVSD models could not be
for each confidence point in the strong-target condition statistically assessed, because the models are not nested.
(see Equations4–6), 2 settings were assumed to be neces- Thus, nested comparisons were made between each of
R
sary for each confidence point in the weak-target condi- these models and the hybrid model that contained both
S
tion (see Equations7–9). Thus, 5 settings were estimated and parameters. A more complete description of this
from the strong-target condition, 1 for each level of con- analysis is given later. Also, because the observed data
fidence, and 10 settings were estimated from the weak- points represented proportions of response frequencies
target condition,2 for each level of confidence. To reiter- summed across participants, 95% confidence intervals
ate, faces given fame judgments were expected to lead to about the observed data were examined as well. For each
higher strength than were faces given sex judgments, and model, the number of predicted points was counted that
when instructed to endorse only those faces that had been fell outside a 95% confidence interval about the observed
R2
in the fame list (i.e., strong targets), the participantsneeded points. Regardless of the , any model fit should be re-
only to accept the most familiar items as having been stud- jectedif there is a significantlygreater numberof predicted
ied. Thus, the actualdecisionprocess and criterionsettings pointsfalling outside the confidenceintervalsthan would
in this conditionwere considered to be similar to those of be expected given a true 95% success rate. A .05 alpha
a simple old–new discrimination. However, if the partici- level was adopted for all significance tests.
pants were instructed only to endorse items that had re-
ceived a sex judgment(i.e., weak targets), one upper crite-
rion and one lower criterion were assumed to be necessary
RESULTS
for each level of confidence,because the participantswere
The average ROCs are presented in Figures 2 and 3,
required to accept only items of intermediate familiarity.
along with the model-generated data points. Hit rates ap-
For the DPSD model, two additional parameters were
y
x
pear on the -axes, and false alarm rates appear on the -
axes. Three plots are shown for each condition, one for
each possible pairwise relationship between item types.
For the weak-target condition,weak sex targets are plotted
against strong fame lures (panel A) and against new items
(panel B). Strong fame lures are plotted against new items
in panel C. For the strong-target condition, strong fame
targets are plotted against weak sex lures (panel D) and
against new items (panel E). Weak sex lures are plotted
against new items in panel F. Ninety-five percent confi-
dence intervals are given around each observed point.
used to represent the contributionof recollectionto strong
R
R
R
and weak items ( s and w, respectively). Both values
were constrainedto assume values between zero and one.
For the UVSD model, two additional parameters were
freed for the standard deviations of strong and weak fa-
S
S
miliarity distributions ( s and w, respectively). Equality
R
R
constraints across conditions were imposed on s and
in the DPSD model and on s and w in the UVSD modewl.
S
S
S
Both parameters were constrained to assume a value of
1.00 or greater.3 For the hybrid model, with which the
R R
S
S
s, and
w
DPSD and the UVSD were compared,
,
,
s
were all estimated under the same constrainwts.
R2
Note that because the of these fits reflect variability of
all three item types, confidence intervals were computed
for all item types and are shown for variables on both
axes in each ROC. Thus, the intervals in Figures 2 and 3
each represent one 2-dimensional side (i.e., a rectangle)
of a confidence “cube.” Model-generated points, rather
than continuous functions, are shown in the figures to il-
lustrate their locations within each confidence rectangle.
Note that because the fitting occurred for three dimensions
(sex, fame, and new items), it is not enough that the model-
The fits of the models were evaluatedby examiningthe
value for the fitted solution. The
4
R2
R2
was computed
SSE
using the overall
term and the sum of the six total
SST
sums of squared deviations( ik) for the three item types
in the two conditions. This SST term was defined as
3
5
2
2
SSE
P - P
ijk ik
=
,
(11)
(
)
ååå
i
j
k
P
where squared deviations of observed points ( ijk) from generated function pass through the two-dimensionalcon-
an item/condition mean ( ik) were summed across = 3 fidence rectangle, as is depicted in the individual ROCs.
–
P
i
j
k
item types, = 5 confidence points, and = 2 experimen- Rather, the model-generated points must lie within the
tal conditions. Note that in this equation, deviations of three-dimensionalconfidencecube to provide an adequate
each point from its corresponding item/condition mean fit. This is because each model-generated point repre-

RECOGNITION MEMORY FOR SOURCE
901
Figure 2. Observed and model-generated data points for the unequal-varaince signal detection model displayed as points
on an ROC. Error bars about each point represent 95% confidence intervals. Arrows indicate ROC points where model
predictions fall outside the confidence intervals for at least one of the two axes. Panels A, B, and C show ROCs from the
weak-target/strong-lure condition for (A) list discrimination with weak sex targets plotted against strong fame lures, (B) item
discrimination with weak sex targets plotted against new lures, and (C) strong fame lures plotted against new lures. Panels
D, E, and F show ROCs from the strong-target/weak-lure condition for (D) list discrimination with fame targets plotted
against weak sex, (E) item discrimination with strong fame targets plotted against new lures, and (F) weak sex lures plotted
against new lures.
sents the closest part of the function to the observed point (sex targets vs. new lures), and panel E shows the target/
in three dimensions. A proper representationof such a fit nonstudied ROC for the strong-target condition (fame
would require a three-dimensional scatterplot with a targets vs. new lures). Finally, ROCs comparing false
model-generated function passing through confidence alarms to studied lures versus nonstudiedlures are shown
cubes around each point. A function that passed through in panels C and F. Panel C shows the ROC for the weak-
a rectangle in two dimensions (e.g., strong targets and target condition (fame lures vs. new lures), and panel F
weak lures) might still pass outside the confidence inter- shows the ROC for the strong-target condition (sex lures
val about a value on the third dimension (e.g., new lures) vs. new items).
and would miss the cube.
The UVSD model accounted for 97.6% of the total
observed variance in the average ROCs and for between
96.5% and 98.4% of the variance of any of the three lists
The Unequal-Variance Signal Detection Model
Data points generated by the UVSD model are shown for either of the two conditions. The best-fitting UVSD
in Figure 2, along with observed points and confidence parameter estimates are shown in Table 2. Under the as-
intervals. The ROCs shown in panels A and D represent sumptions of the UVSD model, making fame judgments
source discrimination (i.e., target faces plotted against about faces during study led to greater discrimination of
d
¢
studied lures). Panel A shows the source ROC for the par- studied items from new lures ( s = 2.70) than did making
d
¢
ticipants who received instructions to endorse only items sex judgments( w = 0.74). The ability to discriminatebe-
from the sex list (i.e., the weak-target condition), and tween items in the weak list and items in the strong list ac-
panel D shows the source ROC for the participants who cording to the UVSD model is represented by a substantial
d
¢
ds¢2d¢
w = 1.96). Also, the standard deviation of
fame list (i.e., the strong-target condition). The ROCs the strong familiarity distribution ( = 2.07) showed an
received instructions to endorse only the items from the
as well (
S
s
S
shown in panels B and E represent the ability to discrimi- increase relative to that of the weak distribution (
=
w
nate studied targets from nonstudied lures. Panel B shows 1.00). These estimates are substantively consistent with
the target/nonstudied ROC for the weak-target condition the general finding that the ratio of nonstudied to studied

902
QUAMME, FREDERICK, KROLL, YONELINAS, AND DOBBINS
Figure 3. Observed and model-generated data points for the dual-process signal detection model displayed as points on
an ROC. Error bars about each point represent 95% confidence intervals. Arrows indicate ROC points where model pre-
dictions fall outside the confidence intervals for at least one of the two axes. Panels A, B, and C show ROCs from the weak-
target/strong-lure condition for (A) list discrimination with weak sex targets plotted against strong fame lures, (B) item dis-
crimination with weak sex targets plotted against new lures, and (C) strong fame lures plotted against new lures. Panels D,
E, and F show ROCs from the strong-target/weak-lure condition for (D) list discrimination with strong fame targets plot-
ted against weak sex lures, (E) item discrimination with strong fame targets plotted against new lures, and (F) weak sex lures
plotted against new lures.
standard deviationstends to decrease with increasing ac- those shown in Figure 2, but with best-fitting approxi-
d
¢
S S
d
¢
S S
= 2.70, / =
n s
curacy ( = 0.74,
/
= 1.00 vs.
mations generated by the DPSD model rather than by the
UVSD model. The DPSD model accounted for 99.0% of
w
n
w
s
0.48; see Glanzer et al., 1999).
Figure 2 also shows model-generated points that rep- the total observedvarianceand between 98.3% and 99.7%
resent the best-fit approximations to the observed data of the variance of any of the three item types for either
points. Arrows indicate ROC data points where the model
predictions fall outside the confidence intervals for at
least one of the two plotted axes. Six of the 30 predicted
values fall outside a 95% confidence interval of their re-
Table 2
Least-Squares Parameter Estimates
and Fit Indices of DPSD and UVSD Models
Fitted Parameter Values
Fit Indices
R2
df %CI
UVSD .00* .00* 2.70 0.74 2.07 1.00 .976 11 80
p
spective observed values (binomial = .0033). In the
¢
ds
¢
weak-target condition,the first three pointsfall below the
intervals for the weak sex targets, and the first two fall
below the intervals for the new lures (see panels A, B,
and C). In the strong condition,the first point falls below
the intervalfor new lures (see panels E and F). Thus, 80%
(24 of 30) of the UVSD model-generated data points fall
within the 95% confidence intervalsaround the observed
points.
Model
Rs
Rw
dw
Ss
Sw
DPSD .30 .16
Hybrid .00 .14
1.58 0.87 1.00* 1.00* .990 11 93
2.17 0.85 1.40 1.00 .993 97
9
Note—DPSD, dual-process signal detection model; UVSD, unequal-
variance signal detection model; Hybrid, hybrid dual-process unequal-
variance signal detection hybrid model; Rs, probability of recollecting
an item from the strong list; Rw, probabilityof recollecting an item from
the weak list; ds, signal detection discriminability of strong items from
new items; dw, signal detection discriminability of weak items from
¢
¢
new items; Ss, standard deviation of strong-item familiarity; Sw, stan-
dard deviation of weak-item familiarity. Range restrictions were im-
posed on Rs and Rw (0.00 $ R $ 1.00) and on Ss and Sw (S $ 1.00).
%CI, percentage of predicted data points that fall within a 95% confi-
Dual-Process Signal Detection Model
Data points generated by the DPSD model are shown
in Figure 3, along with observed points and confidence
intervals. The panels in Figure 3 reflect the same plots as dence interval. *Fixed parameter value.

RECOGNITION MEMORY FOR SOURCE
903
of the two conditions. The best-fitting DPSD parameter
The hybrid model accounted for 99.3% of the total
estimates are shown in Table 2. Fame judgments about variance across both conditions.The fit was significantly
F
faces during study led to both a higher probability of rec- better than that of the UVSD model [ (2,9) = 10.98,
R
MS
p
ollection ( s = .30), and a greater familiarity discrimina-
e = 0.000847, = .0039], indicating that even with a
d
¢
tion of studied items from new items ( s = 1.58) than did provision for unequal variances, the fit was improved
R
d¢
making sex judgments ( w = .16; w = 0.87). As was pre- when recollection was added to the model. By contrast,
dicted, the participants used familiarity to discriminate the fit was not significantly better than that of the DPSD
d
s¢2 d
¢
F
MS p
e = 0.000847, = .23], indicat-
between strong and weak items (
w = 0.71). These model [ (2,9) = 1.69,
estimates are substantively consistent with the general ing that no significant improvement in fit was gained
finding that levels-of-processing manipulations affect over that of the dual-process model by allowing unequal
both recollection and familiarity (see Yonelinas, 2001). variances for familiarity distributions. Thus, a compari-
To determine whether it was reasonable to assume that son of both models with the hybrid reveals that the equal-
Rs
R
were each equivalent across conditions, the variance restriction of the DPSD was reasonable and did
w
and
equality constraints were removed for these parameters, not compromise thefit of thatmodel. Conversely, thezero-
and the model was refit. Two degrees of freedom were recollection assumption of the UVSD was unreasonable
R
R
lost by allowing s and w to vary across conditions.The and did compromise the fit of the model.
modified model explained 99.3% of the variance and did
The best-fittingparameters of the hybrid model are pre-
not represent an improvement over the original model sented in Table 2. Taken at face value, the parameters sug-
F
MS
p
R
e = 0.000834, = .234]. As compared gest that recollection occurred only for weak items ( =
s
[ (2,9) = 1.72,
with the strong-target condition,
R
R
was larger for the .00, w = .14) and that an increase in variance occurred
s
R
S
S
weak-target condition (.47 vs. .19), and w was slightly only for strong items ( s = 1.40, w = 1.00). This finding
larger for the weak-target condition (.19 vs. .15). Thus, is substantively inconsistent with a body of literature on
for the present experiment, there may have been a ten- the effect of level of processing on estimates of recollec-
dency for the participants to use recollection more in the tion and familiarity (Yonelinas, 2001). On one hand, this
weak-target condition,but the fit of the model was not sig- appears to indicate that the improvement over the UVSD
R
nificantlycompromisedby assumingthat recollectionwas model was exclusively the result of adding w. However,
used equivalently in the two conditions. because there was no improvement of the hybrid over the
The model-generated points in Figure 3 represent the DPSD model, parameter differences between the hybrid
R
R
,
s
closest approximationof the DPSD model to the observed and theDPSD models(.16 vs. .14 for w, .00 vs. .30 for
S
data points. Of 30 estimated data points, 2 predicted val- 1.40 vs. 1.00 for s ) should be interpreted with caution.
ues fall outside a 95% confidence interval of their re- Although a hybrid account of recognition ROCs cer-
p
spective observed value (binomial = .45). In the weak- tainly cannot be rejected on the basis of the present ex-
target condition,the last point falls below the interval for periment, it did not appear to represent a substantive im-
weak targets (see panels D and F), and the first point falls provement over either the UVSD or the DPSD models and
below the interval for the new lures (see panels E and F). was not a statistical improvement overthe DPSD model. It
Thus, 93% (28 of 30) of the model-generated data points was, however, a statistical improvement over the UVSD
fall within 95% confidence intervals about the observed model.
points.
A comparison of the two models suggests that the
DPSD model provided a better account for the ROC data
than did the UVSD model. For the DPSD model, only 2
DISCUSSION
In the present study, unequal-varianceand dual-process
out of 30 pointsfell outside the 95% confidenceintervals, signal detection models were fit to recognition ROCs in
whereas for the UVSD model, 6 out of 30 points fell out- conditions in which familiarity should have been useful
side the confidence intervals.
for recognition of both source and occurrence. The par-
ticipants discriminated either strong targets from weak
studied lures and new lures or weak targets from strong
studied lures and new lures. A simple familiarity process
Dual-Process Unequal-Variance
Signal Detection Hybrid
A hybrid model was specified that contained all the was modeled as a UVSD model and did not provide an
free parameters of DPSD and UVSD models. The hybrid adequate fit to the ROCs of the two conditions simultane-
model is simply a dual-process model without the equal- ously. A DPSD model in which familiarity was augmented
variance assumptionfor the familiarity distributions.Thus, by recollectiondid provide an adequate fit of ROCs of the
the hybrid model contains two more free parameters than two conditions simultaneously. When both unequal vari-
do the other two models. Comparing this model with the ances and recollectionwere assumed for the same model,
other two models allows two assumptions to be tested. there was no improvement over the dual-process model,
Comparison with the UVSD model tests the assumption but there was an improvement over the unequal-variance
that givenunequal variances, recollectionis not necessary model. The results suggest that recollection contributed
to account for performance. Comparison with the DPSD to recognition in the present study even though above-
model tests the assumption that given the contribution of chance performance could have been achieved by using
recollection,familiarity distributionshave equal variance. only familiarity and that equal variances of familiarity

904
QUAMME, FREDERICK, KROLL, YONELINAS, AND DOBBINS
distributions were sufficient for the familiarity process miliarity. The first assumption distinguishes the present
once recollection was accounted for. The results further model of familiarity from multidimensional signal de-
suggestthat peopledo not limit their recognitiondecisions tection (MSD) models that propose that items, when stud-
to familiarity assessments when familiarity information ied, can become more familiar along more than one di-
is useful but also make use of recollectionwhen it is avail- mension (see Macmillan & Creelman, 1991). The second
able to them.
assumption about familiarity in the present model distin-
The results are inconsistent with the view that recol- guishes it from some global-matching models of recog-
lection contributesto performance only when familiarity nition, in which context parameters can be used to weight
information is relatively ambiguous. Although it appears familiarity distributions according to test instructions
that a threshold process such as recollection is necessary (e.g., the matrix model, Humphreys, Bain, & Pike, 1989;
to account for linear ROCs in associative and source SAM, Gillund& Shiffrin, 1984;see Clark, 1999, and Rat-
recognitiontasks, it has been argued that such findingsare cliff et al., 1995). Both alternatives are discussed below.
unusual and that the contributionof a threshold process is
not a general characteristic of either item recognition Multidimensional Signal Detection Theory
(Glanzer et al., 1999) or source recognition (Qin et al.,
The source recognition tasks presented here can be
2001; Slotnick et al., 2000). Familiarity-onlysignal detec- viewed in light of an MSD model of source recognition.
tion accounts of recognition ROCs assume unequal vari- For example, the present data could be accounted for by
ances of familiarity distributions in order to account for assuming that the two source lists are represented by spe-
the asymmetry of the ROCs but must also assume that in- cific strength axes and that studying an item serves to in-
switching
dividuals employ a
ear ROCs. That is, a familiarity-only view assumes that such, alltest itemswould have both a
fame list
strategy to account for lin- crease its strength along both axes to some degree. As
sex list
strengthvalue
strength value. Studying an item in the
whenever possible, participants make a familiarity as- and a
sessment only on the basis of a continuous signal detec- fame list would increase its fame strength to a greater ex-
tion process. However, when familiarity cannot be used tent than its sex strength, and vice versa for items in the
to make recognitiondecisions (e.g., when discriminating sex list. It is assumed that to make a recognition deci-
between items of two equally familiar sources), partici- sion, participantsrotate a single decision axis to an angle
pants switch to a threshold-based recollection process. so that the strength distributions,when projected onto the
The present data contradictthis type of switching view, in axis, are discriminable (i.e., the distance between target
that both processescontributedto performance in the same and lure distributions is maximized). Kinchla (1994) and
task.
Batchelder, Riefer, and Hu (1994) have summarized the
Note thatotherprocess-switchingstrategiesmay be con- advantages and disadvantages of MSD modeling, using
sistent with thepresent data.For example, the dual-process MSD theory for source-monitoring data, and Banks and
account presented here can be viewed as a trial-by-trial colleagues (Banks, 2000; Banks, Chen, & Prull, 1999)
switching model whereby participants either recollect an have applied MSD to the analyses of source recognition
item on a given trial or, failing that, resort to a familiarity and to inclusion and exclusion performance in a process
assessment. Thus, recollectedinformation may contribute dissociation paradigm.
to a decision on one trial, whereas a decision on another
trial may be based on a familiarity assessment.
Although the proposals are interesting, MSD models
run into the same problem as unidimensionalsignal detec-
In the present paper, we asked whether recollectionwas tion models, in that they cannot account for linear ROCs.
used, given that it was possible for the participants to As such, it must be assumed that linear ROCs occur when
make their decisionssolely on the basis of familiarity. We participantsabandona signal detectionprocess in favor of
considered this to be the reasonable approach, given that a threshold process for performance. Linear ROCs could
most models of recognition assume that familiarity is an be produced by introducingnonnormal distributions, but
important componentof recognition performance. How- it is unclear what effect this modification would have on
ever, one could easily have asked the opposite question: other aspects of the model (e.g., rotation of the decision
Given that people use a recollection process, is there ev- axis).
idence that they also use a familiarity process? If it is as-
The multidimensionalmemory representationassumed
sumed that recollection is a threshold process, nonlinear in MSD theory may, in some cases, be consistentwith the
ROCs can be taken as evidence for a familiarity process. memory representation assumed by the dual-process
Figures 2 and 3 show that the functions are clearly curvi- model. For example, consider a two-dimensional mem-
linear, which suggests that familiarity was used and that ory space in which the dimensions represent the strengths
a simple threshold process would not be sufficient.
of two sources (e.g., fame strength vs. new, sex strength
In the present study, a simple form of familiarity-only vs. new). If the space is rotated 45º, the dimensions will
recognition was tested in the context of a complex recog- more likely resemble item strength (old vs. new) and
nition task. The following assumptions were made about source strength (fame vs. sex). The latter solutionis more
familiarity: (1) Familiarity is a unidimensional signal de- compatiblewith multinomialand dual-process models of
tection process, and (2) participantsare able to control re- source recognition than is the former, but both will pro-
sponse criteria only, and not the actual computationof fa- vide equivalent statistical fits to the data, provided that

RECOGNITION MEMORY FOR SOURCE
905
the item distributions do not change positions relative to It is, of course, quite plausiblethat familiarity is sensitive
one another. In fact, the dual-process model of Yonelinas to the specific goals and intentions of the participants.
(1994) is equivalent to a multidimensional model in However, it may be more useful to first rule out the role
which item distributions are Gaussian along a familiarity of recollection in these effects, given previous research
dimension and rectilinear along a recollection dimen- showing that familiarity behaves more like an automatic
sion. A decision is first made on the recollection dimen- process than like a controlled process (e.g., Jacoby, 1991;
sion by determining whether an item falls above a region Jacoby, Woloshyn, & Kelley, 1989; Rajaram, 1993). Fu-
of overlap between old and new items (i.e., whether it ex- ture studies will be useful in testing this notion.
ceeds a threshold). If not, the item’s strength is indepen-
dently assessed by an evaluationof its familiarity relative cific may turn out to be true, the global-matchingmodels
to a response criterion on the familiarity dimension. are not consistent with recognition memory ROC data.
Althoughthe assumptionthat familiarity is contextspe-
In summary, MSD theory represents an innovativeand First, some global models predict that recognition ROCs
powerful alternativeto the present dual-processmodel, in will be symmetrical, whereas others predict that they
that performance on a variety of source recognition tasks will always become less symmetrical as performance in-
can be explained by the rotation of a single decision axis creases, and both of these predictionshave been falsified
through a multidimensional memory space. However, the (see Ratcliff et al., 1992; Yonelinas, 1994). Second, they
MSD models do not account for linear ROCs. Moreover, produce Gaussian memory strength distributions, and
without some a priori consideration of the nature of the thus they are inconsistentwith studies demonstratinglin-
memory dimensions, it is difficult to determine just how ear ROCs (e.g., Rotello et al., 2000; Yonelinas, 1997,
substantivelydifferent an MSD model solutionis from that 1999a). Whether it is possible to modify these models to
of a dual-process or multinomial model solution. Much bring them in line with existing ROC data is not yet clear.
work needs to be done to determine whether MSD models
of recognition are both practical and psychologically
meaningful.
REFERENCES
Ashby, F. G. (1992). Multidimensional models of categorization. In
F. G. Ashby (Ed.), Multidimensional models of perception and cog-
nition (pp. 449-483). Hillsdale, NJ: Erlbaum.
Global-Matching Models of Recognition
Atkinson, R. C., & Juola, J. F. (1974). Search and decision processes
in recognition memory. In R. C. Atkinson, R. D. Luce, D. H. Krantz,
& P. Suppes (Eds.), Contemporary developments in mathematical
psychology: I. Learning, memory and thinking (pp. 243-293). San
Francisco: Freeman.
The second assumption about familiarity in the pres-
ent model distinguishes it from global-matching models
of recognition (e.g., SAM; Gillund & Shiffrin, 1984). In
global-matching models of memory, the familiarity or
strength is a single value that is computed as the match Banks, W. P. (1970). Signal detection theory and human memory. Psy-
chological Bulletin, 74, 81-99.
Banks, W. P. (2000). Recognition and source memory as multivariate
decision processes. Psychological Science, 11, 267-273.
Banks, W. P., Chen, Y., & Prull,M. W. (1999). Memory and awareness:
between the test item and all the items in memory. Both
SAM and the matrix model (Humphreys et al., 1989) can
account for certain types of source recognition (e.g., ex-
clusionin the process dissociationprocedure) by assuming
that participantscan “focus” retrieval on certain items and,
thus, that the actual computation of a familiarity value is
under partial control of the individual (Clark, 1999; Rat-
cliff et al., 1995). The use of context in the computation
of familiarity is necessary in a single-process model be-
cause context can be retrieved and used to make recog-
nition decisions (Humphreys et al., 1989; Ratcliff et al.,
1995). This is accomplished in global-matching models
Is memory information stratified into conscious and unconscious
components? In B. H. Challis & B. M. Velichkovsky (Eds.), Stratifi-
cation in cognition and consciousness (pp. 129-172). Amsterdam:
Benjamins.
Batchelder,W. H., Riefer,D. M., & Hu, X. (1994). Measuring mem-
ory factors in source monitoring: Reply to Kinchla. Psychological
Review, 101, 172-176.
Clark, S. E. (1999). Recalling to recognize and recognizing to recall.
In C. Izawa (Ed.), On human memory: Evolution, progress, and re-
flections on the 30th anniversary of the Atkinson–Shiffrin model
(pp. 215-243). Mahwah, NJ: Erlbaum.
Dodson, C. S., Prinzmetal, W., & Shimamura, A. P. (1998). Using
Excel to estimate parameters from observed data: An example from
source memory data. Behavior Research Methods, Instruments, &
Computers, 30, 517-526.
weight
by the addition of context parameters to
items ac-
cording to the retrieval instructions.If the test instructions
require endorsing one set of studied items and rejecting
another, items from the two sets can be given different
context weights by the individual in an attempt to dis-
Donaldson, W. (1996). The role of decision processes in remember-
ing and knowing. Memory & Cognition, 24, 523-533.
criminate between them. By contrast, dual-process the- Gardiner,J. M. (1988). Functional aspects of recollective experience.
Memory & Cognition, 16, 309-313.
Gillund, G., & Shiffrin, R. M. (1984). A retrieval model for both
recognition and recall. Psychological Review, 91, 1-67.
Glanzer, M., Kim, K., Hilford, A., & Adams, J. K. (1999). Slope of
ories generally assume that the intentionaluse of context
affects recognition decisions primarily through the rec-
ollection component and that the contribution of context
to the computationof familiarity cannot be controlledon
the basis of the test instructions alone. This is not to say
that context cannot affect familiarity in a dual-process
model. Rather, the assumption of dual-process models is
that any intentionalfocusing of retrieval on items of a spe-
cific context is reflected in recollection,not in familiarity.
the receiver-operating characteristic in recognition memory. Journal
of Experimental Psychology: Learning, Memory, & Cognition, 25,
500-513.
Green, D. M., & Swets, J. A. (1966). Signal detection theory and
psychophysics. New York: Wiley.
Greene, R. L. (1999). The role of familiarity in recognition. Psycho-
nomic Bulletin & Review, 6, 309-312.

906
QUAMME, FREDERICK, KROLL, YONELINAS, AND DOBBINS
Hintzman, D. L. (1986). “Schema abstraction” in a multiple-trace
memory model. Psychological Review, 93, 411-428.
Hirshman, E., & Henzler, A. (1998). The role of decision processes
in conscious recollection. Psychological Science, 9, 61-65.
Hirshman, E., & Master, S. (1997). Modeling the conscious corre-
lates of recognition memory: Reflections on the remember–know
paradigm. Memory & Cognition, 25, 345-351.
to-reject in recognition: Evidence from ROC curves. Journal of
Memory & Language, 43, 67-88.
Slotnick, S. D., Klein, S. A., Dodson, C. S., & Shimamura, A. P.
(2000). An analysis of signal detection and threshold models of source
memory. Journal of Experimental Psychology: Learning, Memory, &
Cognition, 28, 1499-1517.
Swets, J. A. (1986). Indices of discrimination or diagnostic accuracy:
Their ROCs and implied models. Psychological Bulletin, 99, 100-
117.
Hoffman, H. G. (1997). Role of memory strength in reality monitoring
decisions: Evidence from source attributionbiases. Journal of Exper-
imental Psychology: Learning, Memory, & Cognition, 23, 371-383. Tulving, E. (1985). Memory and consciousness. Canadian Psychol-
Humphreys, M. S., Bain, J. D., & Pike, R. (1989). Different ways to
cue a coherent memory system: A theory for episodic, semantic, and
procedural tasks. Psychological Review, 96, 208-233.
Inoue, C., & Bellezza, F. S. (1998). The detection model of recogni-
tion using know and remember judgments. Memory & Cognition, 26,
299-308.
ogy, 26, 1-12.
Yonelinas, A. P. (1994). Receiver-operating characteristics in recogni-
tion memory: Evidence for a dual-process model. Journal of Experi-
mental Psychology: Learning, Memory, & Cognition, 20, 1341-1354.
Yonelinas, A. P. (1997). Recognition memory ROCs for item and asso-
ciative information: The contribution of recollection and familiarity.
Memory & Cognition, 25, 747-763.
Jacoby,L. L. (1991).A process dissociation framework: Separating au-
tomatic from intentionaluses of memory. Journal of Memory & Lan- Yonelinas, A. P. (1999a). The contribution of recollection and famil-
guage, 30, 513-541.
iarity to recognition and source-memory judgments: A formal dual-
process model and an analysis of receiver operating characteristics.
Journalof Experimental Psychology:Learning,Memory, & Cognition,
25, 1415-1434.
Jacoby, L. L., & Dallas, M. (1981). On the relationship between au-
tobiographical memory and perceptual learning. Journal of Experi-
mental Psychology: General, 110, 306-340.
Jacoby, L. L., Woloshyn, V., & Kelley, C. (1989). Becoming famous
without being recognized: Unconscious influences of memory pro-
duced by dividing attention. Journal of Experimental Psychology:
General, 118, 115-125.
Yonelinas, A. P. (1999b). Recognition memory ROCs and the dual-
process signal-detection model: Comment on Glanzer, Kim, Hilford,
and Adams (1999). Journal of Experimental Psychology: Learning,
Memory, & Cognition, 25, 514-521.
Kelly, R., & Wixted, J. T. (2001). On the nature of associative infor- Yonelinas, A. P. (2001). Consciousness, control, and confidence: The
mation in recognitionmemory. Journal of Experimental Psychology:
Learning, Memory, & Cognition, 27, 701-722.
3 Cs of recognition memory. Journal of Experimental Psychology:
General, 130, 361-379.
Kinchla, R. A. (1994). Comments on Batchelder and Riefer’s multi- Yonelinas, A. P., Dobbins,I., Szymanski, M. D., Dhaliwal, H. S., &
nomial model for source monitoring. Psychological Review, 101,
166-171.
King, L. (1996). Signal-detection, threshold, and dual-process mod-
els of recognition memory: ROCs and conscious recollection. Con-
sciousness & Cognition, 5, 418-441.
Yonelinas, A. P., Kroll,N. E. A., Dobbins, I. G., & Soltani,M. (1999).
Recognitionmemory for faces: When familiarity supportsassociative
recognition judgments. Psychonomic Bulletin & Review, 6, 654-661.
Macmillan, N. A., & Creelman, C. D. (1991). Detection theory: A
user’s guide. New York: Cambridge University Press.
Maddox, W. T., & Estes, W. K. (1997). Direct and indirect stimulus-
frequency effects in recognition. Journal of Experimental Psychol-
ogy: Learning, Memory, & Cognition, 23, 539-559.
Mandler, G. (1980). Recognizing: The judgment of previous occur-
rence. Psychological Review, 87, 252-271.
NOTES
Mandler, G. (1991). Your face looks familiar but I can’t remember
1. Technically, threshold models predict linear ROCs for confidence
your name: A review of dual process theory. In W. E. Hockley (Ed.), ratings only when all the recollected items are given the highest confi-
Relating theory and data: Essays on human memory in honorof Ben-
net B. Murdock (pp. 207-225). Hillsdale, NJ: Erlbaum.
Marsh, R. L., & Bower, G. H. (1993). Eliciting cryptomnesia: Un-
conscious plagiarism in a puzzle task. Journal of Experimental Psy-
chology: Learning, Memory, & Cognition, 19, 673-688.
dence rating. Thus, the dual-process model assumes that all recollected
items will lead to a highly confident endorsement or rejection. The
high-confidencerestriction on recollected items is supported by empir-
ical data showing that recollection-based discrimination is reflected al-
most exclusively in high confidence ratings (Yonelinas, 1994, 2001).
Murdock, B. B. (1974). Human memory: Theory and data. Potomac, To the extent that this assumption is violated, the ROC would be more
MD: Erlbaum.
curvilinear. Consequently,the contributionof familiarity would be over-
estimated, and that of recollection would be underestimated.
2. The model also bears a superficial resemblance to two-criterion
signal detection models of remembering and knowing in recognition
Ogilvie, J. C., & Creelman, C. D. (1968). Maximum likelihood esti-
mation of receiver-operating characteristic curve parameters. Journal
of Mathematical Psychology, 5, 377-391.
Qin, J., Raye,C. L., Johnson, M. K., & Mitchell, K. J. (2001).Source (Donaldson, 1996; Hirshman & Henzler, 1998; Hirshman & Master,
ROCs are (typically) curvilinear: Comment on Yonelinas (1999). 1997; Inoue & Bellezza, 1998). However, the remember/know models
Journal of Experimental Psychology: Learning, Memory, & Cogni- assume that participants use an upper remember/know criterion to di-
tion, 27, 1110-1115.
vide a single distribution of studied items into remember and know re-
gions. In the present model, the upper criterion is used to discriminate
between two separate distributions of studied items.
3. This constraint was included because most signal detection mod-
els of recognition assume that the variance of studied-item distribution
does not normally decrease relative to that of the new-item distribution
(i.e., the z-ROC does not have a slope greater than 1). For completeness,
we fit a UVSD model with no range restrictions on Ss and Sw. The fit-
Rajaram, S. (1993). Remembering and knowing:Two means of access
to the personal past. Memory & Cognition, 21, 89-102.
Ratcliff, R., McKoon, G., & Tindall, M. (1994). Empirical general-
ity of data from recognition memory receiver-operating characteris-
tic functions and implications for the global memory models. Jour-
nal of Experimental Psychology: Learning, Memory, & Cognition,
20, 763-785.
¢
Ratcliff, R., Sheu, C.-F., & Gronlund, S. D. (1992). Testing global ted weak-item distributionhad a dw of 0.69 and an Sw of 0.75. This sug-
memory models using ROC curves. Psychological Review, 99, 518- gests that making sex judgments about faces during study increased the
535.
mean of the distributionbut decreased the standard deviation. The overall
Ratcliff, R., Van Zandt, T., & McKoon, G. (1995). Process dissoci- R2 was .979,which didnotrepresent an appreciabledifference fromtheR2
ation, single-process theories, and recognition memory. Journal of of the UVSD model presented here with S range restrictions (R2 = .976).
Experimental Psychology: General, 124, 352-374.
Rotello,C. M., Macmillan,N. A., & Van Tassel, G. (2000). Recall-
4. The ROC fitting procedure can also be conducted on raw response
frequencies rather than cumulated frequencies. Maximum likelihood

RECOGNITION MEMORY FOR SOURCE
907
estimation (Ashby, 1992;Ogilvie & Creelman, 1968) is sometimes pre-
ferred over least squares estimation because variability in both hits and
false alarms can be accounted for. However, this problem was overcome
in the present study by minimizing squared error terms for all item
types. Both procedures, as well as the chi-square minimization tech-
nique employed by Slotnick et al. (2000), were performed on the pres-
ent data, and all three yielded similar results. The maximum likelihood
However, the conclusions based on model comparisons using all the
procedures were identical. Consequently, only least squares fits are re-
ported for models (R2), and F ratios are presented for comparisons
among models.
¢
fits led to slightly lower R estimates and slightly higher d estimates than
did the least squares or the chi-square minimization of cumulated data.
(Manuscript received September 10, 2001;
revision accepted for publication May 17, 2002.)