Publication - The Cambridge Face Memory Test

Neuropsychologia 44 (2006) 576–585

The Cambridge Face Memory Test: Results for neurologically intact
individuals and an investigation of its validity using inverted
face stimuli and prosopagnosic participants
Brad Duchaine a,∗ , Ken Nakayama b
b

a Institute of Cognitive Neuroscience, University College London, 17 Queen Square, London WC1N 3AR, UK
Department of Psychology, Harvard University, Vision Science Laboratory, 33 Kirkland Street, 7th Floor, Cambridge, MA 02138, USA

Received 21 October 2004; accepted 1 July 2005
Available online 19 September 2005

Abstract
The two standardized tests of face recognition that are widely used suffer from serious shortcomings [Duchaine, B. & Weidenfeld, A. (2003).
An evaluation of two commonly used tests of unfamiliar face recognition. Neuropsychologia, 41, 713–720; Duchaine, B. & Nakayama,
K. (2004). Developmental prosopagnosia and the Benton Facial Recognition Test. Neurology, 62, 1219–1220]. Images in the Warrington
Recognition Memory for Faces test include substantial non-facial information, and the simultaneous presentation of faces in the Benton
Facial Recognition Test allows feature matching. Here, we present results from a new test, the Cambridge Face Memory Test, which builds
on the strengths of the previous tests. In the test, participants are introduced to six target faces, and then they are tested with forced choice
items consisting of three faces, one of which is a target. For each target face, three test items contain views identical to those studied in the
introduction, five present novel views, and four present novel views with noise. There are a total of 72 items, and 50 controls averaged 58. To
determine whether the test requires the special mechanisms used to recognize upright faces, we conducted two experiments. We predicted that
controls would perform much more poorly when the face images are inverted, and as predicted, inverted performance was much worse with a
mean of 42. Next we assessed whether eight prosopagnosics would perform poorly on the upright version. The prosopagnosic mean was 37,
and six prosopagnosics scored outside the normal range. In contrast, the Warrington test and the Benton test failed to classify a majority of
the prosopagnosics as impaired. These results indicate that the new test effectively assesses face recognition across a wide range of abilities.
© 2005 Elsevier Ltd. All rights reserved.
Keywords: Face recognition; Prosopagnosia; Neuropsychology

Face recognition is one of the most intensively studied
aspects of human cognition involving scientists from a wide
range of related fields. Because of this, it is important that
researchers have access to well-designed standardized tests
of face recognition. Such tests would provide a means to
compare the performance of participants in different laboratories. In addition, they would provide researchers with a
ready-made tool so they would not need to create a test and
develop norms. Lastly, neuropsychologists and neurologists
require additional tests that can contribute to classifying individuals who have face recognition impairments.
∗

Corresponding author. Tel.: +44 20 7679 1005; fax: +44 20 7916 8517.
E-mail address: [email protected] (B. Duchaine).

0028-3932/$ – see front matter © 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.neuropsychologia.2005.07.001

Currently, there are two commonly used standardized
tests of face recognition, the Benton Facial Recognition Test
(BFRT) (Benton et al., 1983) and the Recognition Memory Test for Faces (RMF) (Warrington, 1984). They are
widely used with normal participants and neuropsychological participants, but both suffer from serious shortcomings
that make them potentially misleading tests of face recognition (Duchaine & Weidenfeld, 2003; Duchaine & Nakayama,
2004). In the BFRT, participants are simultaneously presented with a target face and six test faces. Participants must
choose the three test faces that match the target face. Because
the target face and the test faces are presented simultaneously, participants can use a feature matching strategy. An
experiment with normal participants showed that a substan-

B. Duchaine, K. Nakayama / Neuropsychologia 44 (2006) 576–585

tial proportion were able to score in the normal range on a
modified version of the BFRT in which the face was occluded
so that only the eyebrows and the hairline were presented
(Duchaine & Weidenfeld, 2003). Furthermore, a number of
prosopagnosics have been shown to score normally on the
BFRT (Duchaine & Nakayama, 2004; Newcombe, 1979;
Nunn, Postma, & Pearson, 2001) including some who in addition to their deficits with face memory tests also have deficits
with face perception tests (Duchaine, Yovel, Butterworth, &
Nakayama, in press). These converging results make it clear
that the possibility of feature matching makes the BFRT a
poor measure of face recognition ability. The RMF is limited by the nature of the images used in the test. During the
inspection phase of the RMF, participants are presented with
50 target images for 3 s each. Following this, participants are
presented with 50 forced-choice items consisting of a target
face and a distracter face. Images contain substantial non-face
information that can be used to discriminate between target
and distracter images. These include hair, clothing, posture,
emotional expressions, and image imperfections. When normal participants were presented with a modified version of
the RMF that occluded only the facial information, many participants were able to score in the normal range (Duchaine &
Weidenfeld, 2003). In addition, some prosopagnosics have
scored normally on the test (Duchaine, 2000; Nunn et al.,
2001), including one whose performance fell nearly to chance
when the non-facial information was occluded (Nunn et al.,
2001).
Because of the problems with these tests, scientists and
practitioners are left without an effective standardized test of
face recognition. To address this deficiency, we have created a
test of face memory that maximizes the strengths of the BFRT
and the RMF. Like the BFRT, our test will have sections with
different levels of difficulty and test items with novel views
of target faces. Like the RMF, our test will involve a memory
paradigm with multiple faces, which will make simultaneous
feature matching impossible. However, unlike in the BFRT
and the RMF, the face stimuli will be limited strictly to facial
information (e.g. no clothing, no hair line). Our test is also
akin to everyday face recognition in that participants will have
an opportunity to gradually acquire knowledge of target faces
from a wide range of views. They will see each target face 17
times throughout the entire test. Although they do not receive
feedback after seeing each test item, repetitive viewing should
provide the opportunity to develop better representations after
viewing images in test items.
Because the test will measure face memory, performance
on the test will depend on both perceptual mechanisms and
memory. As a result, the test will not provide a means to
measure the perceptual processes uncontaminated by memory processes, and our laboratory is currently developing a
test of face perception. However, face memory, not face perception, is the ability that determines our success in identity
recognition in everyday life, and so it is especially important
to measure it. When tests of face perception are developed,
the combination of tests of face perception and face memory

577

should provide a means to assess the contributions of perceptual processes and memory to variability in face recognition
ability.
We call our test the Cambridge Face Memory Test
(CFMT). The test will be available free of charge when
used for research purposes. In the following sections, we
will describe the results of testing with neurologically normal
participants. Following this, we discuss experiments aimed
at evaluating the validity of the test by testing neurologically
normal participants with inverted face stimuli and examining
whether individuals with face recognition impairments score
poorly with the CFMT.

1. Method
1.1. Stimuli
The faces are those of men in their 20s and early 30s, and
each individual was photographed in the same range of poses
and lighting conditions. Men’s faces were used, because men
and women perform equivalently with men’s faces whereas
women show an advantage with women’s faces (Lewin &
Herlitz, 2002; McKelvie, Standing, St. Jean, & Law, 1993).
All faces were cropped so that no hair was visible and facial
blemishes were removed. The men posed with neutral expressions.
Six individuals were chosen as target faces. We used six
targets, because it is a challenging yet manageable number
of faces for normal subjects to encode after brief exposures.
Twelve images of each target face were selected, and the same
poses and lighting conditions were used for each target face.
Test items consisted of a target face along with two distracter
faces with the same pose and lighting. Forty-six individuals
were used as distracter faces. Many of the distracter individuals were presented repeatedly, and this repetition meant
that participants could not simply make a familiar/unfamiliar
discrimination on test items with repeated distracters.
1.2. Procedure
The test consists of four stages (practice, introduction/same images, novel images, and novel images with
noise). Completing the test takes between 10–15 min.
1.2.1. Practice
The practice stage familiarizes participants with the procedure used in the introduction/same images stage by presenting cartoon faces in the same fashion that the target faces will
be presented. After instructing the participant to memorize
the following faces, three study images of Bart Simpson are
presented for three seconds each: a left 1/3 profile, a frontal
view, and a right 1/3 profile. Then a test item consisting of
one of the study views of Bart along with two other cartoon
faces is presented. Participants are instructed to press the key
corresponding to the number below the target face (1, 2, or

578

B. Duchaine, K. Nakayama / Neuropsychologia 44 (2006) 576–585

3). Two more test items follow, and each consists of one of
the study faces along with two distracter faces.

are never used as distracter faces. Feedback is not provided
during the test.

1.2.2. Introduction/same images
Participants are instructed that they will now begin the
test, and they are introduced to the first target face in the
same way that they were introduced to Bart Simpson during
the practice stage. Three study images are presented for three
seconds each. The images are a left 1/3 profile, a frontal view,
and a right 1/3 profile. Fig. 1 Panel A shows an individual
in the three views. Three test items are then presented and
participants are instructed to pick out the individual whom
they were just shown (see Fig. 1, panel B). Each test item
includes an item identical to a study item. Because the study
and test images are the same, the participants could respond
correctly by recognizing the image rather than face (Hay &
Young, 1982). There are six target faces, and this procedure
is repeated for the five remaining target faces. Target faces

1.2.3. Novel images
Immediately before this stage, participants are presented
with a single review image that has a frontal shot of each target
face. They are given 20 s to review this image. Following
the review image, participants are presented with 30 forcedchoice test items (6 target faces × 5 presentations) in a fixed,
random order. Each test item contains three faces, one of
which is a target face. Participants are instructed that each
test item will contain one of the six target faces and told
to respond with the key corresponding to the number under
the target face. All are novel images in which the lighting,
pose, or both vary (see Fig. 1, panel C). Appendix A displays
examples of the poses and lighting used for target items in
the novel images and novel images with noise sections, and
the lighting and poses used were the same for all six target

Fig. 1. Examples of stimuli similar to test stimuli. None of these items was used in the test. In the test, test faces are numbered 1, 2, and 3 from left to right, but
we omitted this to save space. Panel A shows study views of a target face. Study views are presented for three seconds each. Panel B displays a test item from
the introduction. Face 3 is the same image as the rightmost study view in Panel A. Panel C shows an item from the novel image section (face 1 is the target).
Panel D displays a test item from the novel images with noise section (face 3 is target).

B. Duchaine, K. Nakayama / Neuropsychologia 44 (2006) 576–585

579

faces. When participants are presented with test items in the
introduction, they know which target face will be the correct
answer. However, during this stage and the final stage, the
correct answer for an item can be any of the target faces, and
so the items are much more difficult.
1.2.4. Novel images with noise
Participants are presented with the review image again
for twenty seconds. Following this, 24 test items (6 target
faces × 4 presentations) are presented in a fixed, random
order. These items consist of novel images, and different levels of Gaussian noise were added to the face images (see
Fig. 1, panel D). Levels of noise for the faces in a test item
are identical. Noise was added to the faces for two reasons. At
the beginning of the final section, participants will have seen
each target image 13 times so the noise was added to keep
performance away from ceiling. Second, studies with normal participants indicate that noise forces increased reliance
on the special mechanisms that face recognition normally
depends on (McKone, Martini, & Nakayama, 2001).

2. Results
In this section, we discuss the results from our normal participants. Following this, we discuss two conditions aimed at
determining whether our test effectively measures face recognition. We do this by first administering the test when all faces
are presented inverted, and by giving the test in its upright
version to prosopagnosic individuals.

Fig. 2. Average cumulative performance for the 50 controls on the upright
version. Points display the average cumulative score for controls at each test
item. Error bars display the standard deviation for the cumulative scores.
Dashed lines divide the figure into the three different sections. Deviation
from perfect responding at the end of each section can be gauged by viewing
the distance between the intersection of the dashed lines and the cumulative
score for the final item in the section.

2.1. Performance of neurologically intact participants
with upright faces
Our participants were 50 college age individuals ranging in
age from 18–26 with a mean age of 20.2 (S.D. = 1.8). Twentynine of these participants were female while 21 were male.
They were paid for their participation.
Fig. 2 displays the average cumulative performance along
with the standard deviation (see Appendix C for means and
standard deviations). The figure is divided into the three sections of the test, and the intersections of the dashed lines indicate perfect performance. Because participants knew ahead
of time which target face would be present in each item in
the introduction, we expected them to perform very well
and in fact, they made few mistakes. However, they made
many more errors in the second section as is evidenced by
the decreasing slope in Fig. 2. For these items, participants
did not know which target face would be presented and all of
the images were novel views. Fig. 3 plots individual scores
at the end of each section, and this figure makes it clear that
there were a wide range of scores in the second section. In the
final section, participants were presented with novel images
degraded by noise. The slope of the line in Fig. 2 was even
flatter for this section, so it appears that the noise made these
items even more difficult.

Fig. 3. Individual cumulative scores for controls. We included the scores
of every other control so the figure is not overly cluttered. Scores at the
end of each section were computed. The largely similar relative position
of individual controls from section to section indicates that performance in
different sections depended on the same abilities.

580

B. Duchaine, K. Nakayama / Neuropsychologia 44 (2006) 576–585

The average total score out of 72 for the controls was 57.9
(S.D. = 7.9), which converts to 80.4% (S.D. = 11.0). Total
scores ranged from 43 to 72. The male participants averaged
56.5 (S.D. = 7.3) and the female participants averaged 58.9
(S.D. = 8.3). This difference was not significant.
2.1.1. Consistency of scores from section to section
To check if the different sections of the test relied on the
same abilities and representations, we looked at the participants’ consistency from section to section. We computed
correlation coefficients using each participant’s score for each
test section. Because participants performed so well in the
introduction/same images section, there was little variability,
and so the correlation coefficients were relatively low for the
same images-novel images comparison (r = 0.27, p = 0.06)
and for the same images-novel images with noise comparison (r = 0.35, p = 0.01).
In contrast, the scores for the novel images section and
the novel images with noise section were quite consistent,
and the correlation coefficient for this comparison was 0.74
(p = 0.001). In Fig. 3, the consistency of the participants is
apparent in that their rank at the end of the novel images with
noise is quite similar to their rank at the end of the novel
images section.
2.1.2. Item analysis
Next we conducted an item analysis to determine whether
the test contained items that did not effectively discriminate between good performers and poor performers. To do
this, we computed a correlation coefficient involving each
participant’s total score and their performance on each item
(correct or incorrect). Because performance was nearly perfect or perfect for the same image items in the introduction,
the correlations for these items were either not interesting or
we were unable to compute them. However, there was variability for all of the items in the other two sections except
for one so we were able to compute coefficients for 53 items.
The average correlation for the novel image items was 0.35
(S.D. = 0.13) and was 0.35 (S.D. = 0.13) for the novel image
with noise items. (Note that these equivalent values are not
a typo.) All but one of the correlations was positive (it was
−0.004) so 52 of 54 items contributed to the test’s sensitivity.
2.1.3. Analysis by face
To analyze the difficulty of the six target faces, we computed the percent correct for each target face. These percentages were 77, 69, 80, 81, 88, and 88. An inspection of the
two faces producing the two highest percentages leads us to
believe that it was because these faces were the most distinctive, but the order may have contributed as well.
2.2. Performance of neurologically intact participants
when the faces are inverted
The results in the previous section show that the test
produces a nice range of scores that are consistent from

section to section and that do not suffer from ceiling or
floor effects. However, these scores do not demonstrate
that the test actually assesses face recognition abilities. It
could simply activate general-purpose visual recognition
mechanisms.
To address this issue, we will first assess the effect of
inverting all of the faces in the test. Typically, inversion
decreases percent correct in face recognition experiments by
15–25% (Diamond & Carey, 1986; Scapinello & Yarmey,
1970; Yin, 1969) whereas inversion of many other objects
classes affects percent correct far less (Diamond & Carey,
1986; Scapinello & Yarmey, 1970; Yin, 1969). This disproportionate effect has been used to argue that upright faces
are processed in a manner that is qualitatively distinct from
the processing applied to other objects (Yin, 1969). Further work has shown that the specialized processing which
upright faces receive involves holistic or configural representation (Freire & Lee, 2000; McKone et al., 2001; Tanaka &
Farah, 1993; Tanaka & Sengco, 1997; Young, Hellawell, &
Hay, 1987) whereas most other types of objects, including
inverted faces, are represented more as a collection of parts
(Biederman, 1987). This distinction has also been supported
by a double dissociation between upright face processing and
inverted face processing (Farah, Wilson, Drain, & Tanaka,
1995; Moscovitch, Winocur, & Behrmann, 1997).
If the test relies on the mechanisms normally used for
upright face recognition, then we should find a large decrement in performance when the faces are inverted. However,
if we find the effect is not comparable to past face inversion
effects, this will indicate that performance did not depend
on the special mechanisms. We examined this prediction by
testing 20 new participants drawn from the same population
as the participants used for the upright version.
Fig. 4 plots the cumulative scores for participants in the
upright and inverted conditions. Even by the end of the
introduction/same images section, inverted scores are significantly worse than upright scores (t(68) = 5.0, p < 0.0001).
This difference suggests that normal face recognition mechanisms were contributing to upright performance even in this
very easy section. As Fig. 4 shows, the difference between
upright and inverted scores became much more pronounced
when novel images were presented, and this difference is
highly significant (t(68) = 7.9, p < 0.0001). For the section
with novel images with noise, the inverted average was
only 10% above chance (43%), and the difference between
upright and inverted was highly significant (t(68) = 6.0,
p < 0.0001).
The average inverted score for the entire test was 42.1
(S.D. = 4.7) or 58.4% correct (S.D. = 6.5). The upright mean
was 80.4% so inversion lowered performance by 22%, an
effect comparable to previous inversion effects. This difference was highly significant (t(68) = 8.4, p < 0.0001). The
inverted mean is two standard deviations below the upright
mean. There was little overlap between the scores in the two
conditions with inverted scores ranging from 33–50 whereas
the upright scores ranged from 43–71.

B. Duchaine, K. Nakayama / Neuropsychologia 44 (2006) 576–585

Fig. 4. Comparison of average upright cumulative scores to average inverted
cumulative scores. Error bars display one standard error above and below
the mean (upright n = 50; inverted n = 20). The figure is divided into the three
sections, and deviation from perfect responding at the end of each section
can be gauged by viewing the distance between the intersection of the dashed
lines and the cumulative score for the final item in the section.

2.3. Performance by individuals with face recognition
impairments with upright faces
The difference in performance for the upright and inverted
versions of the test indicates that the test activates the special
processes used to recognize upright faces. Next we address
this same issue by assessing the performance of individuals with face recognition impairments on the normal, upright
version of the test. Because the test appears to rely on the
special processes used with upright faces, we expect that the
individuals with impairments to these mechanisms will perform poorly on the test.
In addition, their scores will demonstrate whether the test
can contribute to assessments of individuals who may have
face recognition impairments. We will compare the prosopagnosics’ scores on the test to their scores on the BFRT (Benton
et al., 1983) and the RMF (Warrington, 1984). If scores on
the CFMT better classify the prosopagnosics than the BFRT
or the RMF, it will suggest that the CFMT could be a useful
measure for neurologists and neuropsychologists.
The eight participants in this section contacted our laboratory, because they complained of significant problems in
everyday face recognition. We will refer to these individuals
with labels indicating their sex (F or M) and their age at the
time of testing. Two out of this group suffered brain damage as young children (M26 and M41). The rest report no
head trauma and so appear to be congenital prosopagnosics.
Four of these individuals have been studied in other papers
on prosopagnosia (M26—Kosslyn et al., 1995; Hadjikhani
& de Gelder, 2002; M53—Duchaine, Dingle, Butterworth,

581

Fig. 5. Comparison of average upright cumulative scores for eight prosopagnosic participants and the control mean. The error bars for the controls
display one standard deviation above and below the control mean. Cumulative score after every six items is displayed.

& Nakayama, 2004; Harris, Duchaine, & Nakayama, 2005;
M57—Duchaine, 2000; F46—Harris et al., 2005). To assess
whether these individuals did, in fact, have face recognition impairments, we tested them with two memory tests
with unfamiliar faces (Duchaine et al., 2003; Duchaine
& Nakayama, 2005) and a famous face test (Duchaine &
Nakayama, 2005). The z scores for the participants are presented in Appendix B along with their z scores on the CFMT,
and this table shows that their performance was clearly
impaired.
Fig. 5 shows the upright control average and the scores for
each prosopagnosic participant. Whereas the controls scored
nearly perfectly in the introduction/same images section,
many of the prosopagnosics made errors and the prosopagnosic average was significantly lower than the control average
(t(56) = 9.2, p < 0.0001). Like the inverted average, the prosopagnosic average in the section involving novel images plummeted relative to the upright control average (t(56) = 6.2,
p < 0.0001). By the end of the novel images section, all but
two of the prosopagnosics were more than two standard deviations below the mean. For the novel images with noise
added, the prosopagnosic average was just above chance
(34.9%), and this difference was quite significant (t(56) = 5.3,
p < 0.0001). The overall prosopagnosic mean was 36.5 (S.D.
= 9.7) or 50.7% (S.D. = 13.4), which is 2.7 standard deviations below the control mean (t(56) = 6.9, p < 0.0001). Scores
for the prosopagnosic participant ranged from 25 to 53.
2.3.1. Performance with different views
Fig. 6 displays percent correct on test items with front
views and those with side views in the novel views and novel

582

B. Duchaine, K. Nakayama / Neuropsychologia 44 (2006) 576–585

Fig. 6. Performance on items involving front views and side views from
the novel items and novel items with noise for upright controls, inverted
controls, and prosopagnosics.

views with noise sections. Among these items, there were 24
front views and 30 side views (18 right and 12 left). As is
apparent in the figure, percent correct was slightly higher for
the front views. The difference between the different views
was nearly identical in our three participant groups.
2.3.2. Performance on different face tests
Fig. 7 displays the scores for each prosopagnosic participant on the CFMT, BFRT, and RMF as standard deviations
from the normal control mean. Control means and standard
deviations for the BFRT and the RMF were obtained from
the manuals (Benton et al., 1983; Warrington, 1984). Neuropsychologists often classify scores two standard deviations
below the mean as impaired, and Fig. 7 shows that for the
CFMT scores for six of the eight prosopagnosic participants

were below this cut-off. F41’s score on the CFMT was 1.6
standard deviations below the control mean.
If we next consider the scores on the RMF, Fig. 7 shows
that scores on the CFMT and RMF for a number of prosopagnosics (M261 , F20, M53, and F41) were very similar. However, only three of the eight prosopagnosics scored more than
2 standard deviations below the mean. Especially problematic
are the scores of participants such as M41 and F46. They did
very poorly on the CFMT and other tests of face recognition,
yet scored normally on the RMF. Their normal performance
appeared to rely on non-facial information. M41 commented
that he was doing photograph recognition rather than face
recognition, and F46 remarked that she recognized the clothing and haircuts on many of the items. Other prosopagnosics
were also able to score well on the RMF. M53’s RMF score
and F29’s RMF score were near the control mean yet they
were clearly impaired on other face memory tests.
The BFRT suggests classifying scores of 40 and below as
impaired, so we have placed a dashed line in Fig. 7 at the standard deviation corresponding to a score of 40.5. The mean
score for the prosopagnosics was 42.4 (S.D. = 2.5), and six
of the eight prosopagnosics had scores classified as normal.
None had scores more than 2 standard deviations below the
control mean. The BFRT has three different types of items:
matching of identical front-views, matching of front-views
with three-quarter views, and matching front views under different lighting. All of the prosopagnosics scored perfectly on
items requiring matching of identical front-views. For matching different views, they average 19.4 out of 24 while their
mean was 17 out of 24 for matching under different lighting.
To compare how well each test discriminates between individuals with normal face recognition and those with impaired
face recognition, we have computed d for each test. d is a
bias-free measure of discrimination (Green and Swets, 1966).
The CFMT classified all 50 normal participants correctly
(hits) so its specificity is 100%. It correctly classified six of
the eight prosopagnosics (correct rejections) so its sensitivity
is 75%. Because d cannot be computed when there are zero
hits or false alarms, we changed the false alarm rate to one
out of 50. This produces an d for the CFMT of 2.7. Because
we did not run controls on the RMF and the BFRT, we will
charitably assume that, like the CFMT, no controls would
have been classified as impaired. On this assumption, the d
score for the BFRT is 1.4 and the d score for the RMF is 1.7.
3. Discussion
We created a new test of face memory in hopes that it can
supplement standardized tests of face recognition. The results
discussed in the previous section are very encouraging. Fig. 8
displays performance on the three sections of the test for the
three conditions. First consider the upright percent correct.
Because these scores are far off of the floor and the ceiling, the

Fig. 7. Comparison of performance on the CFMT, BFRT, and RMF for
the prosopagnosic participants. Scores are displayed as standard deviations
below the control mean.

1

M26’s score for the RMF was taken from Hadjikhani & de Gelder (2002).

B. Duchaine, K. Nakayama / Neuropsychologia 44 (2006) 576–585

Fig. 8. Percent correct for each test section for the three conditions. The error
bars show the standard deviation for each section for the upright condition.
The dashed line indicates chance performance.

test can assess a wide range of abilities. Each of the top five
possible scores (68–72) was achieved by only one participant
so the test challenges even individuals with very good face
recognition. Similarly, only five participants scored below 49
so it appears to discriminate in the low range of normal face
recognition abilities as well.
The test produced similar scores for men and women.
While the scores for the women were approximately two
points higher than the scores for the men, this difference was
not significant. A small sample of middle-aged participants
also suggests that the test can be used with older participants
as well. Nine middle-aged, college educated participants with
an average age of 46.6 (S.D. = 7.7) produced a mean slightly
higher (61.8) than our college age mean.
We predicted that inversion of the images would lead to
a drop in performance if the upright version activates the
special processes that contribute to upright face recognition.
Fig. 8 makes it clear that inversion affected performance in
every section. The difference between total scores for upright
and inverted was 22%, and this drop is comparable to or
greater than that seen in other recognition memory experiments comparing upright and inverted performance (Yin,
1969; Scapinello & Yarmey, 1970; Diamond & Carey, 1986).
We also investigated the validity of the test by testing
eight prosopagnosic individuals with the upright version.
Fig. 8 shows that the prosopagnosic mean on each section
was slightly lower than the inverted means, and their overall
mean was 2.7 standard deviations below the control mean.
Six of eight participants were also outside the range of control scores. All of the normal participants scored better than
2 standard deviations below the mean, and all but two of
the prosopagnosic participants were more than two standard
deviations below the mean.
It is particularly interesting to examine the scores of
prosopagnosic individuals who have shown normal performance with object recognition tests, because their normal or
at least relatively normal object recognition mechanisms may

583

provide an alternative means to achieve a good score on the
test. One of the participants (F46) has performed normally
on a number of tests of object discrimination (Duchaine &
Nakayama, 2005)2 while another (M53) has performed normally on every non-face test on which we have tested him
(Duchaine et al., in press). Despite their good abilities with
many categories of objects, F46 scored 2.9 standard deviations below the normal mean while M53 was 2.4 below. This
suggests that the test forced reliance on the special processes
which are impaired in these individuals.
Two of the prosopagnosic participants (F41 and M57),
however, had scores within 2 standard deviations of the mean
and within the normal range. F41’s score of 45 is 1.6 standard deviations below the mean while M57’s score of 53 was
only slightly below the control mean. However, they had the
best scores among the prosopagnosics on the two tests of
face memory that we used to classify prosopagnosics (see
Appendix B for the scores and Duchaine & Nakayama, 2005
for details on the tests). Given that we tested a number of
prosopagnosics, a score like M54’s, which places her in the
bottom 5% of normal participants, is not particularly surprising. However, M57’s score gives no hint of his impairment.
The experimenter checked M57’s score immediately after the
test, and after seeing how good it was, asked M57 how he had
done so well. M57 responded that he intentionally attempted
to “lust” after the faces rather than simply memorize them.
He has been in many faces tests and he was interested in how
this would affect his performance, because he believes that he
processes faces differently when he is attracted to them. His
score suggests his encoding strategy may have worked, and
other experiments indicate that more attractive faces are better remembered than unattractive faces (Cross, Cross, & Daly,
1971; Shepherd & Ellis, 1973). Given all of the different types
of information that can be extracted from faces (emotionality, masculinity–femininity, attractiveness, etc.), our test, like
all current tests with faces, presents information that mechanisms other than those used for face recognition can operate
on. Because of these alternative routes to recognition, potential prosopagnosics should always be tested with a range of
tests.
Our comparison of the CFMT, BFRT, and RMF showed
that the CFMT classified 75% of the prosopagnosics correctly while only 25% were correctly classified by the BFRT
and only 38% were correctly classified by the RMF. Because
the CFMT and the RMF both test face memory, the disparity between these classifications is very problematic for the
RMF. The BFRT is sometimes used as a test of face recognition, and the normal performance by the prosopagnosics
demonstrates that it does not effectively classify individuals with face recognition impairments. However, the BFRT
despite its name was designed as a test of face perception, and
so normal performance on it by prosopagnosics along with
deficits for face memory performance could be explained as
a dissociation between intact face perception and impaired
2

Participant F46 was called F2 in Duchaine & Nakayama (2005).

B. Duchaine, K. Nakayama / Neuropsychologia 44 (2006) 576–585

584

face memory. This may account for some of the scores in the
normal range, but some of the prosopagnosics tested appear
to have impaired face perception. For example, M53 scored
45 on the BFRT, yet he shows no face inversion effect and
is impaired on a range of face processing tasks (emotion,
gender, attractiveness)(Duchaine et al., in press). Past results
showing that normal participants can score normally when
the majority of the face is occluded also indicate that normal
scores do not demonstrate normal face perception (Duchaine
& Weidenfeld, 2003). Thus, our results suggest that normal
scores on the BFRT and the RMF should be interpreted cautiously.
In summary, these results indicate that the Cambridge Face
Memory Test is a valid measure of face recognition ability
that is sensitive to a wide range of abilities. The test is available free of charge for research purposes. Because it will be
freely available, we hope to rapidly generate norms for different demographic groups.

Appendix C. Summary of scores for each condition
and section

Upright
Introduction
Novel images
Novel images with noise
Total
Inverted
Introduction
Novel images
Novel images with noise
Total
Prosopagnosics
Introduction
Novel images
Novel images with noise
Total

Mean

S.D.

Range

17.82
23.74
16.36

0.44
4.31
4.02

16–18
17–30
7–24

57.92

7.91

43–71

16.15
15.50
10.40

2.28
2.76
2.96

8–18
11–19
4–15

42.05

4.71

33–50

15.31
14.15
8.77

2.21
3.93
3.11

12–18
8–21
3–14

38.23

7.52

25–53

Appendix A

Appendix B. z scores for prosopagnosics on four tests
of face memory

M26
M44
F20
F46
M53
F29
F41
M57

CFMT

Famous faces

Old–new 1

Old–new 2

−4.2
−3.9
−3.8
−2.9
−2.4
−2.3
−1.6
−0.6

−11.6
−6.2
−5.7
−5.3
−7.3
−4.7
−2.2
−8.5

File Typeapplication/pdfFile Titledoi:10.1016/j.neuropsychologia.2005.07.001File Modified2017-12-14File Created2006-01-18