Emotional brain states carry over and enhance future memory formation

Abstract

Emotional arousal can produce lasting, vivid memories for emotional experiences, but little is known about whether emotion can prospectively enhance memory formation for temporally distant information. One mechanism that may support prospective memory enhancements is the carry-over of emotional brain states that influence subsequent neutral experiences. Here we found that neutral stimuli encountered by human subjects 9-33 min after exposure to emotionally arousing stimuli had greater levels of recollection during delayed memory testing compared to those studied before emotional and after neutral stimulus exposure. Moreover, multiple measures of emotion-related brain activity showed evidence of reinstatement during subsequent periods of neutral stimulus encoding. Both slow neural fluctuations (low-frequency connectivity) and transient, stimulus-evoked activity predictive of trial-by-trial memory formation present during emotional encoding were reinstated during subsequent neutral encoding. These results indicate that neural measures of an emotional experience can persist in time and bias how new, unrelated information is encoded and recollected.

Full text

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
nature neurOSCIenCe ADVANCE ONLINE PUBLICATION 1
a r t I C l e S
We form lasting, vivid and detailed memories for only a subset of
our experiences. Emotion is one key factor that influences the fate of
our memories1,2. Compared to their neutral counterparts, emotional
events and stimuli are more robustly remembered, with higher levels
of confidence, vividness and detail1–4. The presence of emotion not
only increases recollection of emotionally arousing experiences them-
selves but also has been shown to retroactively modulate recollection
of neutral information preceding emotional arousal5,6. Despite such
findings documenting the impact of emotion on memory for informa-
tion before and during manipulations of emotional arousal, little work
has examined whether emotion can prospectively enhance memory
for subsequently encountered information minutes later. Moreover, it
is unknown whether the persistence of emotional arousal is capable of
impacting future brain states and modifying the neural structures that
support memory formation for subsequently encountered stimuli.
Here we assessed whether exposure to emotionally arousing stimuli
prospectively enhances memory for subsequently encountered stimuli
by biasing future states of brain activity.
A robust body of literature indicates that emotional arousal during
an experience enhances the consolidation of memory for that particu-
lar event, resulting in more persistent, vivid and detailed emotional
memories over time1–4. Mechanistically, emotional arousal has been
linked with the release of norepinephrine and epinephrine, which,
in concert with the amygdala, are thought to modulate hippocampal
processes during both the encoding and subsequent consolidation of
emotional experiences1,7–9. Supporting this notion, activity within
and connectivity between the amygdala, hippocampus and medial
temporal lobe cortex reliably predicts successful emotional memory
formation and consolidation10–13. In addition to enhancing memory
for emotional events themselves, emotional learning and exposure
to emotional stimuli—or the induction of arousal pharmacologically
or with shock—can retroactively enhance long-term retention of
preceding neutral information5,6,14–18 (but see refs. 15,19).
But what about unrelated neutral information that follows an
emotional experience? Can extended periods of emotional arousal
bias future brain states and, in doing so, prospectively modify the
neural structures supporting memory formation for unrelated, neu-
tral information? Prior work has examined the influence of emotion
on memory for information encountered seconds after emotional
arousal6,15,19, and a recent study found that individual differences
in arousal induced by a block of emotional stimuli were related to
enhanced memory discrimination for similar visual images pre-
sented a few minutes later20. Previous studies have also examined
the impact of stress on memory formation for both emotional and
neutral stimuli. Memory for both emotional and neutral images is
enhanced when encoding blocks are interleaved with a stress manipu-
lation21. At the neural level, amygdala connectivity remains altered
after stress induction22–24, and stress induction can alter activation
in brain regions related to memory formation for both emotional
and neutral stimuli21. However, to our knowledge, no prior investiga-
tions have examined how extended periods of emotional arousal can
prospectively bias future brain states and thereby modify the manner
in which unrelated, neutral information is encoded into memory.
Here, we tested whether arousal and brain states associated with
an extended (~20 min) emotional experience can carry-over and
bias the encoding of neutral stimuli encountered approximately
9 to 33 min later and thereby modulate how those stimuli are encoded
into memory (Fig. 1).
1Helen Wills Neuroscience Institute, University of California, Berkeley, Berkeley, California, USA. 2Department of Basic Neurosciences, University of Geneva Campus
Biotech, Geneva, Switzerland. 3Department of Psychology, New York University, New York, New York, USA. 4Center for Neural Science, New York University, New York,
New York, USA. 5Nathan Kline Institute, Orangeburg, New York, USA. 6These authors equally contributed to this work. Correspondence should be
addressed to L.D. (lila.davachi@nyu.edu).
Received 21 June; accepted 23 November; published online 26 December 2016; doi:10.1038/nn.4468
Emotional brain states carry over and enhance future
memory formation
Arielle Tambini1,6, Ulrike Rimmele2,6, Elizabeth A Phelps3–5 & Lila Davachi3,4
Emotional arousal can produce lasting, vivid memories for emotional experiences, but little is known about whether emotion
can prospectively enhance memory formation for temporally distant information. One mechanism that may support prospective
memory enhancements is the carry-over of emotional brain states that influence subsequent neutral experiences. Here we found
that neutral stimuli encountered by human subjects 9–33 min after exposure to emotionally arousing stimuli had greater levels
of recollection during delayed memory testing compared to those studied before emotional and after neutral stimulus exposure.
Moreover, multiple measures of emotion-related brain activity showed evidence of reinstatement during subsequent periods of
neutral stimulus encoding. Both slow neural fluctuations (low-frequency connectivity) and transient, stimulus-evoked activity
predictive of trial-by-trial memory formation present during emotional encoding were reinstated during subsequent neutral
encoding. These results indicate that neural measures of an emotional experience can persist in time and bias how new, unrelated
information is encoded and recollected.

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
2 ADVANCE ONLINE PUBLICATION nature neurOSCIenCe
a r t I C l e S
To this end, two groups of subjects underwent blood-oxygen-level-
dependent (BOLD) functional MRI (fMRI) scanning during inciden-
tal encoding of extended (23 min) blocks of emotional and neutral
complex scenes (emotional and neutral encoding conditions; Fig. 1a),
returning 6 h later for a surprise memory test. One group of subjects
first encountered an extended block of emotional stimuli followed by
neutral stimuli (‘E-N encoding order’), while a different group of sub-
jects encountered neutral stimuli followed by emotional stimuli (‘N-E
encoding order’). A separate group of subjects incidentally encoded
blocks of neutral stimuli under the same procedures as the fMRI
subjects (but outside of the MRI scanner), and two other groups of
subjects incidentally encoded blocks of emotional and neutral stimuli
(E-N and N-E encoding orders) outside of the MRI scanner (Fig. 1a).
We reasoned that the prospective influence of emotion, or a carry-
over of an emotional state into subsequent neutral encoding, should
be present when subjects encoded a long block of emotional stimuli
first and neutral stimuli second (E-N encoding order) but not for the
opposite encoding order (N-E encoding order) or when only neutral
stimuli were encoded outside of the context of emotional stimuli
(N-N encoding order). We first assessed the persistence of emotional
arousal from emotional encoding into subsequent neutral encoding
blocks by measuring skin conductance levels (SCL) as a proxy for sym-
pathetic nervous system activation25. Behaviorally, we asked whether
the carry-over of an emotional state would enhance the subjective
recollection of subsequently encountered neutral stimuli (in the
E-N encoding order relative to the other encoding orders). Finally,
using fMRI, we examined whether emotion would prospectively bias
future encoding-related brain activity in at least two ways (Fig. 1b):
first, we asked whether BOLD activity patterns during emotional
and neutral encoding were more similar when neutral stimuli were
encoded after versus before emotional stimuli (E-N versus N-E encod-
ing order) and second, whether brain regions supporting emotional
memory (for example, amygdala and anterior hippocampus) were
more active and showed greater levels of connectivity when neutral
stimuli were encoded after versus before emotional stimuli (i.e., when
a carry-over of emotional arousal may have been present).
RESULTS
Skin conductance
If emotional arousal following an extended block of emotional encod-
ing persists into neutral encoding, then overall SCL, which is related to
sympathetic nervous system activation25, should track this persistence
and differ based on emotion and encoding order. To account for indi-
vidual differences in baseline skin conductance, SCL was computed
as the change relative to a baseline rest period obtained before the
first encoding block (referred to as ‘relative SCL’; Online Methods).
When assessing relative SCL in the fMRI study, we found a significant
interaction between emotion (emotional versus neutral encoding)
and encoding order (Fig. 2a; F1,42 = 8.72, P = 0.0051; permutation
test, P = 0.0036). Elevated SCL showed evidence of persistence from
long-lasting blocks of emotional into subsequent neutral encoding
(Fig. 2a). Specifically, SCL increased during emotional encoding
compared to the preceding baseline rest scan (t21 = 3.21, corrected
P = 0.0167; permutation test, corrected P = 0.004) and remained height-
ened during a block of subsequent neutral encoding (SCL relative
to baseline rest scan, t21 = 3.06, corrected P = 0.0239; permutation test,
corrected P = 0.0096). Thus SCL levels were equivalent between emo-
tional and neutral encoding when neutral stimuli were encoded tens
of minutes after emotional stimuli (E-N encoding order, t21 = 0.51,
P = 0.61; permutation test, P = 0.65), despite the temporal separation of
these encoding blocks. By contrast, SCL was significantly greater during
emotional versus neutral encoding when neutral stimuli were encoded
before emotional stimuli (N-E encoding order; Fig. 2a; t21 = 3.19,
P = 0.0044; permutation test, P = 0.0022).
We next directly tested whether relative SCL during neutral encod-
ing was enhanced when neutral stimuli were encountered after emo-
tional stimuli, compared to before emotional stimuli (N-E encoding
order) or after neutral stimuli (N-N encoding order). Relative SCL
during neutral encoding was marginally enhanced for E-N versus N-E
(t42 = 1.62, P = 0.056, one-tailed t-test; one-tailed permutation test,
P = 0.053) and N-N encoding orders (comparison with second block,
t43 = 1.32, P = 0.097, one-tailed t-test; one-tailed permutation test,
P = 0.098; comparison with first block, t43 = 1.50, P = 0.071, one-tailed
t-test; one-tailed permutation test, P = 0.073). However, relative SCL
during emotional encoding did not differ between encoding orders
(t42 = 0.82, P = 0.42; permutation test, P = 0.41). These results provide
evidence for enhanced SCL when neutral stimuli were encountered
after emotional stimuli, supporting the notion that a block of emo-
tional stimuli can induce arousal that persists tens of minutes later,
into and during an extended block of subsequent neutral encoding.
Time
~23 min
Emotional
encoding
Neutral
encoding
9 min 9 min~23 min
Neutral
encoding
Emotional
encoding
Rest
Rest
Rest
Rest E N
N E
Emotional
encoding
Neutral
encoding
Neutral
encoding
Emotional
encoding
Memory
Arousal
Neural pattern similarity
Emotion-related activity
Neu :
Neu
fMRI +
behavior
Behavior
Neutral
encoding
Neutral
encoding
Rest Rest N N
fMRI +
behavior
Neutral
encoding
Neutral
encoding
Neu
vs.
a
b
Figure 1 Experimental design and predictions. (a) Experimental design:
separate groups of subjects performed blocks of incidental emotional
and neutral encoding (E→N, subjects that performed emotional followed
by neutral encoding; N→E, subjects that performed neutral followed by
emotional encoding): two groups encoded stimuli during fMRI scanning
(E→N and N→E encoding orders), while two other groups of subjects
underwent E→N and N→E encoding blocks under behavioral testing
conditions (Supplementary Fig. 1) and one other group encoded neutral
stimuli outside of the context of emotional stimuli under behavioral
testing conditions (N→N). Encoding blocks were separated by rest periods
(cream boxes). (b) We predicted that arousal associated with emotional
encoding would persist and carry over, prospectively biasing brain activity
during subsequent neutral (neu) encoding and enhancing memory for
neutral stimuli for the E→N encoding order. No emotional bias or
carry-over of emotion should be present in the other encoding orders.
A prospective influence of emotion on subsequent neutral encoding
predicts that several factors listed should be differentially present during
neutral encoding for the E→N vs. N→E encoding order (greater memory,
arousal, emotion-related activity and neural pattern similarity between
emotional and neutral encoding).

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
nature neurOSCIenCe ADVANCE ONLINE PUBLICATION 3
a r t I C l e S
Memory performance
Next, we compared memory for neutral stimuli as a function of
encoding order, to assess whether a long block of emotionally arous-
ing stimuli is capable of prospectively enhancing memory for neutral
stimuli encountered 9 to 33 min later. Memory for stimuli encoun-
tered during the encoding sessions was assessed in a surprise memory
test 6 h later using a remember (R)–know (K) procedure26. We first
assessed whether emotional and neutral memor y differed based
on encoding order in the fMRI study (reported below), as well as
in separate groups of behavioral participants that did not undergo
fMRI scanning (Supplementary Fig. 1). As expected, overall memory
accuracy (proportion of total R and K hits minus false alarms) was
higher for emotional versus neutral stimuli (main effect of emotion:
F1,42 = 16.0, P = 0.00025). However, memory accuracy differed as
a function of encoding order (Fig. 2b and Supplementary Fig. 1;
encoding order by emotion interaction: F1,42 = 10.6, P = 0.0022). As
in the SCL data reported above, no reliable difference in emotional
versus neutral memory was found in the fMRI study when neutral
stimuli were encountered after emotional stimuli (E-N encoding
order; Fig. 2b; t21 = 0.66, P = 0.52), during which emotional arousal
showed evidence of persistence into subsequent neutral encoding (via
enhanced SCL; Fig. 2a). However, a robust enhancement in emotional
memory was found for the opposite N-E encoding order (Fig. 2b;
t21 = 4.37, P = 0.00027), in which no carry-over of emotional arousal
could be present. Our main question concerned whether exposure to
extended blocks of emotionally arousing stimuli could prospectively
enhance memory for neutral stimuli encountered tens of minutes
later. Memory for neutral stimuli was significantly greater when
they were encoded after versus before emotional stimuli and out-
side of the context of emotional stimuli (Fig. 2b and Supplementary
Fig. 1): corrected recognition rates of 62.7% were found in the
E-N encoding order compared to 53.5% in the N-E encoding order
(t42 = 2.06, P = 0.045) and 52.0% and 46.5% in the N-N encoding
order (first and second blocks of N-N encoding order; first block,
t43 = 2.41, P = 0.02; second block, t43 = 3.31, P = 0.002; permutation
test, P = 0.0022). In contrast to neutral stimuli, memory accuracy
for emotional stimuli did not differ between E-N and N-E encoding
orders (Fig. 2b; t42 = −0.78, P = 0.44). Note that the same pattern of
results was observed in separate behavioral groups, with enhanced
levels of memory for neutral stimuli when they were encountered
after emotional stimuli (E-N encoding order) relative to the other
encoding orders (Supplementary Fig. 1). These results demonstrate
that emotion can prospectively enhance memory for neutral stimuli
encountered tens of minutes later.
Given prior work indicating that emotion specifically enhances
the subjective sense of recollection (for example, see ref. 3) we asked
whether the memory benefit for neutral stimuli encountered after
emotional encoding blocks showed the same specificity. Comparing
emotional and neutral memory based on encoding order in the fMRI
study, a significant triple interaction was found between R versus
K responses, emotional versus neutral stimuli and encoding order
(F1,42 = 14.0, P = 0.0006). As predicted, R responses differed for
emotional versus neutral stimuli based on encoding order (emotion
by encoding order interaction: F1,42 = 20.7, P = 4.5 × 10−5; Fig. 2b
and Supplementary Fig. 1). Consistent with a carry-over or persist-
ence of emotional arousal into subsequent neutral encoding, levels
of emotional and neutral R responses were similar when neutral
stimuli were encountered after emotional stimuli in the fMRI study
(E-N encoding order; Fig. 2b; t21 = 0.49, P = 0.63). However, in
the opposite N-E encoding order, R responses were significantly
higher for emotional compared to neutral stimuli (Fig. 2b; t21 = 5.99,
P = 6 × 10−6). Crucially, similar to the combined memory measure
reported above, R responses were significantly greater for neutral
stimuli when they were encountered after versus before long blocks
of emotional stimuli (Fig. 2b and Supplementary Fig. 1; t42 = 2.16,
P = 0.037) and versus neutral stimuli in the N-N encoding order
(Fig. 2b and Supplementary Fig. 1; first block, t43 = 2.20, P = 0.03;
second block, t43 = 2.87, P = 0.006). This enhancement in neutral R
responses was found for both the fMRI study (reported here) and
the behavioral participants (Supplementary Fig. 1). However, in
contrast to neutral stimuli, R responses for emotional stimuli did
not reliably differ across E-N and N-E encoding orders (t21 = −1.38,
P = 0.17; Fig. 2b and Supplementary Fig. 1). In contrast to recol-
lection-based R responses, K responses did not significantly differ
for neutral versus emotional stimuli across encoding orders (Fig. 2b
and Supplementar y Fig. 1; emotional stimuli, t42 = 0.97, P = 0.34;
neutral stimuli, t42 = −0.48, P = 0.64; emotion by encoding order
interaction, P = 0.09).
Emo
R
~
Relative SCL (µS)
Time (blocks)
~
E N
Overall Memory (R+K)
b
0
0.2
0.4
0.6
0.8
Memory accuracy (hits − FAs)
Emotional
Neutral (E N)
Neutral (N E)
Neutral (N N)
−4
−2
0
2
4
6
8
N E N N
RR
Neu Neu
RRR
Emo Neu
R R R
Neu
E N N E N N
E N N E N N E N N E N N
a
KnowRemember
**
*
*
*
** **
*
**
*
**
**
**
Figure 2 Skin conductance levels and behavioral results. (a) Mean
galvanic SCLs are shown for each encoding order in the fMRI study
(n = 22 for each N→E and E→N encoding order) and for the N→N encoding
order for the behavioral study (n = 23), with each data point representing
data from each subject. Larger (filled) data points indicate means across
subjects. SCL is plotted relative to each subjects’ baseline SCL during the
first rest period. Each encoding block (emotional, emo; neutral, neu) has
three measurements, corresponding to the mean relative SCL during three
encoding scans (or 7.7 min epochs) for that block. Asterisks above data
points denote significant differences in SCL from the first baseline rest
scan (corrected for multiple comparisons in each encoding order).
(b) Memory was assessed for stimuli seen during encoding using a remember
(R)–know (K) procedure. Memory accuracy (hits minus false alarms) is
shown for all hits (combined R and K responses, left bars), R responses
(middle bars) and K responses (right bars) for the fMRI study (E→N and
N→E encoding orders; n = 22 for each encoding order) and the N→N
encoding order (behavioral study, n = 23). See Supplementary Figure 1
for similar results for E→N and N→E encoding orders in the behavioral
study. All error bars represent s.e.m. across subjects and individual dots
represent data from each subject. ~P = 0.053, *P < 0.05, **P < 0.005.

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
4 ADVANCE ONLINE PUBLICATION nature neurOSCIenCe
a r t I C l e S
The above results indicate that subjective recollection of neutral
stimuli was relatively greater when they were encountered 9 to 33 min
after blocks of emotional stimuli, mirroring enhanced SCL found
during neutral encoding. In contrast to differences in neutral memory
and SCL during neutral encoding, levels of subjective recollection
and SCL during emotional encoding did not consistently differ as a
function of encoding order. This pattern of results was observed in
two independent data sets (during fMRI scanning and in a separate
behavioral study). Together, these findings are consistent with the
notion that long-lasting blocks of emotional arousal can persist on an
extended timescale and prospectively enhance the later recollection
of neutral stimuli encountered at least 9 to 33 min later.
Low-frequency connectivity carry-over effects
Next, we sought to examine evidence for the hypothesis that exposure
to long blocks of emotional stimuli created a specific, arousal-related
brain state that could bias subsequent BOLD activity patterns during
neutral encoding. Given the extended time period of the encoding
sessions (~23 min) as well as the temporal gap between emotional and
neutral encoding (~9 min), the carry-over of such an emotional brain
state may have occurred at a relatively slow timescale (for example, on
the order of minutes). Thus, we reasoned that this brain state might be
manifest in low-frequency (LF) fluctuations in the BOLD signal, i.e.,
BOLD signal changes that were slower than trial-by-trial evoked activ-
ity. Such slow LF fluctuations have previously been shown to track
different behavioral states27–29. To test this hypothesis, we examined
whether patterns of LF connectivity that characterized emotional
encoding blocks carried over or showed evidence of reinstatement
during subsequent neutral encoding blocks (Fig. 3a).
Given the importance of the amygdala in mediating the influence
of emotion on memory and cognition30–32, we asked whether large-
scale patterns of LF amygdala connectivity representative of an emo-
tional brain state showed evidence of reinstatement during neutral
encoding blocks. Such reinstatement would result in greater simi-
larity between emotional and neutral encoding LF amygdala con-
nectivity patterns when neutral stimuli are encountered after versus
before emotional stimuli (E-N versus N-E encoding order). To this
end, we computed LF connectivity across cortical and subcortical
voxels using the amygdala as a seed region. Connectivity was com-
puted separately for each encoding scan in each subject, resulting
in distinct multivariate images or vectors representing LF amy-
gdala connectivity patterns during emotional and neutral encoding
(Fig. 3b). We then asked whether these global patterns of LF amy-
gdala connectivity were more similar (i.e., exhibited higher levels
of correlation) between emotional and neutral encoding blocks
for the E-N encoding order (in which a carry-over of emotional
arousal into subsequent neutral encoding was found) versus the N-E
encoding order (in which no carry-over of emotional arousal should
be present). LF amygdala connectivity patterns were significantly more
correlated (i.e., more similar) during emotional and neutral encod-
ing for the E-N versus the N-E encoding order (Fig. 3b; t38 = 2.42,
P = 0.02). This result indicates that large-scale, multivoxel patterns of
LF amygdala connectivity representative of emotional encoding can
carry over and manifest tens of minutes later when unrelated neutral
stimuli are later encountered.
At a more fine-grained level, we next examined specific interac-
tions between the amygdala and anterior hippocampus, which are
thought to underlie the memory benefit for emotional stimuli9,12,32.
We targeted the anterior hippocampus since it is expected to show
the strongest interactions with the amygdala, based on anatomy33–36
and prior reports of functional correlations between amygdala and
anterior hippocampal BOLD activity37 and since anterior hippoc-
ampal activation is consistently predictive of successful emotional
encoding10. As shown in Figure 4, levels of LF amygdala–anterior
hippocampal connectivity also showed evidence of being rein-
stated or carrying over from emotional to neutral encoding blocks.
First, connectivity during emotional versus neutral encoding differed
as a function of encoding order (emotion by encoding order interac-
tion, F1,38 = 4.57, P = 0.039). Mirroring our behavioral and SCL find-
ings, similar levels of LF amygdala–anterior hippocampal connectivity
were found during emotional and neutral encoding when neutral
stimuli were encountered tens of minutes after emotional stimuli (E-N
encoding order, t19 = −0.82, P = 0.42), but a significant enhancement
in LF connectivity was found during emotional versus neutral encod-
ing for the opposite encoding order (N-E encoding order, t19 = 2.12,
P = 0.048). Moreover, LF amygdala–anterior hippocampal connec-
tivity was significantly greater during neutral encoding blocks when
neutral stimuli were encountered after versus before emotional stimuli
(E-N versus N-E encoding orders; t38 = 2.38, P = 0.023). Yet no reli-
able difference in connectivity was found during emotional encoding
blocks between encoding orders (t38 = 0.97, P = 0.34). This pattern of
results suggests that LF amygdala–anterior hippocampal connectivity,
like overall levels of skin conductance, showed evidence of carrying
over from temporally extended blocks of emotional encoding into
subsequent neutral encoding.
Lastly, we found similar evidence for a carry-over effect in levels of LF
connectivity from emotional into neutral encoding within the ventral
a
Task frequency
Low-frequency fluctuations
Amygdala power
0.05 0.1 0.15 0.2 0.25
Frequency (Hz)
r = 0.73
Low-frequency
filtered BOLD
60 120 180 240
Time during encoding (s)
Amygdala
Anterior hippocampus
b
Emotional-neutral amygdala
correlation pattern similarity (z)
E N N E
0
0.2
0.4
0.6
0.8
1
Emo Neu
Neu Emo
r = 0.77
r = 0.46
Amygdala
seed
*
Figure 3 Low-frequency connectivity as a function of encoding order.
(a) Low frequency connectivity analysis approach. Left: plot shows the
power spectrum of amygdala BOLD encoding data for an example subject.
A peak in the power spectrum can be seen for the task frequency, 0.0714 Hz.
To examine LF connectivity, BOLD data were filtered below 0.06 Hz.
Right: plot shows LF filtered time courses of amygdala and anterior
hippocampus BOLD data from one subject during emotional encoding.
(b) Similarity of multivoxel LF amygdala connectivity patterns between
emotional and neutral encoding blocks as a function of encoding order.
Example data from one subject from the E→N encoding order (top left)
and one subject from the N→E encoding order (bottom left) are shown.
An example amygdala ROI (used as the seed region) is shown in the left
inset. Images show sagittal slices containing LF connectivity patterns
with the amygdala. Vectors next to each image depict the analysis
approach; the similarity of multivoxel connectivity patterns in each
subject was computed between emotional and neutral encoding blocks.
Right plot shows group data (n = 20 subjects in each encoding order)
for the similarity of multivoxel LF amygdala correlation patterns between
emotional and neutral encoding. Individual data points represent the
similarity (correlation) for each subject. Error bars, s.e.m. *P < 0.05

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
nature neurOSCIenCe ADVANCE ONLINE PUBLICATION 5
a r t I C l e S
anterior insula (vAI) network (Supplementary Fig. 2). We additionally
probed connectivity in this network as emotion-related activity has
consistently been shown in the vAI (refs. 38–41), and greater levels
of vAI network connectivity have been related to heightened arousal
ratings of the emotional stimuli used in this experiment38.
Taken together, these results provide evidence that emotional brain
states, measured by correlations in LF BOLD fluctuations present dur-
ing emotional encoding, can carry-over and become reinstated tens
of minutes later when participants encountered unrelated, neutral
information. Multiple signatures of emotion-related LF connectiv-
ity showed evidence of carrying over from emotional encoding into
subsequent neutral encoding and were present in both local networks
(for example, amygdala–anterior hippocampal connectivity) and in
global pattern similarity across the brain (in multivoxel amygdala
connectivity patterns).
Event-related subsequent memory carry-over effects
In addition to asking whether correlations in LF fluctuations in the
BOLD signal showed evidence of carry-over from blocks of emo-
tional to subsequent neutral encoding, we also examined whether
emotional encoding prospectively influenced transient, stimulus-
evoked or event-related BOLD activity during subsequent exposure
to neutral stimuli. Given our data suggesting that emotion prospec-
tively enhanced behavioral signatures of memory for subsequently
encountered neutral stimuli, we asked whether the neural structures
supporting memory formation for neutral stimuli were modulated
on a trial-by-trial basis by the prior induction of an emotional state.
Specifically, we compared stimulus-evoked activity related to subse-
quent recollection-based memory (i.e., greater activation for stimuli
later labeled as R versus K) during emotional and neutral encoding
blocks based on the order of encoding blocks, using both univariate
and multivariate approaches. As a function of encoding order, we
compared (i) the similarity of global, multivoxel patterns support-
ing later recollection between emotional and neutral encoding and
(ii) univariate activation predicting subsequent recollection between
emotional and neutral encoding.
First, we asked whether global, multivoxel patterns that charac-
terize recollection-based emotional memory formation were rein-
stated and similarly supported the later recollection of neutral stimuli
encountered after emotional stimuli. Such a reinstatement would
result in greater similarity between multivoxel patterns supporting
later recollection during emotional and neutral encoding when neu-
tral stimuli are encountered after versus before emotional stimuli.
To test this prediction, we measured activation patterns related to
successful recollection-based memory formation (differences in
activation estimates for trials later labeled as R versus K) separately
for emotional and neutral stimuli in each participant (Fig. 5a).
This resulted in multivoxel activity patterns across cortical and sub-
cortical voxels that were characteristic of recollection-based memory
formation, which we then compared between emotional and neutral
stimuli as a function of encoding order. Enhanced levels of similarity,
or correlation, were found between multivoxel patterns supporting
recollection-based memory for emotional and neutral stimuli when
neutral stimuli were encountered after versus before emotional stimuli
(in the E-N versus N-E encoding order; Fig. 5b; t40 = 3.00, P = 0.0046).
A nonparametric permutation test was performed to ensure that this
difference in multivoxel pattern similarity was not driven by differ-
ences in R versus K bin sizes between encoding orders (P = 0.00083;
Supplementary Fig. 3). This result demonstrates that spatially broad
activation patterns related to recollection-based memory formation
were more similar between emotional and neutral stimuli when neu-
tral stimuli were encountered after versus before emotional encoding
blocks. This result suggests that patterns of brain activity support-
ing successful recollection-based memory formation of neutral
Amygdala–anterior
hippocampus correlation (z)
Emo Neu EmoNeu
0
0.5
1
**
Amygdala
Anterior hippocampus
a b
NE
NE
Figure 4 Low-frequency amygdala–anterior hippocampal connectivity as
a function of encoding order. (a) Anatomically defined amygdala (red) and
anterior hippocampus (blue) ROIs are shown for an example subject.
(b) LF amygdala–anterior hippocampus connectivity as a function of
encoding order across n = 20 subjects per encoding order. LF amygdala–
anterior hippocampal connectivity is greater during neutral encoding for
the E→N vs. N→E encoding order (red vs. blue bars). Individual data
points represent connectivity for each subject. Error bars, s.e.m. *P < 0.05.
SRM emotional–neutral
pattern similarity (z)
Emo Neu
r = 0.22
Neu Emo
r = 0.004
Whole-brain Hippocampus
–0.08
0
0.08
0.16
0
0.1
0.2
0.3 ** *
a
b
N E
E N
N E N EE NE N
Figure 5 Similarity of multivoxel subsequent recollection-based memory
differences between emotional and neutral encoding as a function of
encoding order. (a) Example data are shown from one subject of the
E→N encoding order (top) and one subject of the N→E encoding order
(bottom). Sagittal slices depict subsequent recollection patterns
(R – K activity estimates) across cortical and subcortical voxels. (b) Group
data for the similarity of multivoxel encoding patterns (activity patterns
supporting subsequent recollection memory (SRM)) between emotional
and neutral encoding based on encoding order (n = 21 subjects per
encoding order). Greater similarity of patterns supporting subsequent
emotional and neutral recollection was found for the E→N vs. N→E
encoding orders for both global patterns across cortical and subcortical
voxels (whole-brain, left bar plot) and within the hippocampus (right bar
plot). Individual data points represent similarity (correlation) for each
subject. Error bars, s.e.m. *P < 0.05, **P < 0.005.

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
6 ADVANCE ONLINE PUBLICATION nature neurOSCIenCe
a r t I C l e S
stimuli were prospectively biased by the presence of prior emo-
tional stimuli tens of minutes earlier, with activation patterns sup-
porting later neutral stimulus recollection being more similar to
emotional encoding patterns when neutral stimuli were preceded by
emotional stimuli.
Next, we examined whether similar effects were found within the
hippocampus, as emotion is thought to enhance recollection-based
memory formation via hippocampal mechanisms1,7–9. Hippocampal
activity patterns showed evidence of carry-over effects or reinstate-
ment of activity patterns representative of emotional encoding dur-
ing subsequent neutral encoding blocks, both in multivoxel patterns
predictive of later recollection (Fig. 5b; t40 = 2.02, P = 0.049) and
in the anterior versus posterior localization of hippocampal voxels
contributing to subsequent recollection36,37 (Supplementary Fig. 4
and Online Methods). These findings indicate that patterns of hip-
pocampal activity supporting recollection-based memory formation
of neutral stimuli were also biased by the presence of prior emotion,
specifically by enhancing the contribution of the anterior hippocam-
pus to memory formation.
Lastly, we asked whether individual brain regions supporting suc-
cessful recollection-based memory formation were more similar
between emotional and neutral stimuli when neutral stimuli were
encountered after emotional stimuli than when neutral stimuli were
encountered before emotional stimuli. To this end, we identified
regions that showed significant subsequent recollection effects (R > K
BOLD responses) for emotional stimuli as well as greater R > K activity
when neutral stimuli were encountered after versus before emotional
stimuli (E-N versus N-E encoding order). This carry-over recollection-
based memory effect was operationalized as brain regions that showed
a conjunction between six contrasts, depicted in Figure 6a (Online
Methods; note that this analysis also controls for differences in R and
K bin sizes based on encoding order). Six regions emerged from this
whole brain conjunction analysis (Fig. 6b; family-wise error (FWE)-
corrected, P < 0.05): the left amygdala, the right perirhinal cortex, the
right posterior inferior temporal gyrus and three regions in bilateral
inferior prefrontal cortex. Similar regions have previously been shown
to predict successful memory formation for emotional versus neutral
stimuli10, and a formal decoding analysis performed on the uncor-
rected conjunction map revealed the highest similarity with terms
“emotional stimuli,” “affect” and “salient” (using neurosynth.org42,
http://neurosynth.org/decode/?neurovault=JOYXPMRX-26012).
Here we found that regions associated with emotional processing and
memory formation were predictive of subsequent neutral memory
when neutral stimuli were encountered after, but not before, extended
blocks of emotional stimuli. The engagement of these additional brain
regions subserving neutral memory formation after emotional arousal
supports the idea that emotion- and encoding-related mechanisms were
reinstated and thus impacted memory formation 9 to 33 min later.
DISCUSSION
It is widely acknowledged that emotion can modulate what and
how we remember. Although prior work has shown that emotion
can retroactively influence memory for preceding neutral experi-
ences5,6,14,15, less is known about how emotional experiences can
linger and prospectively enhance memory formation for neutral infor-
mation encountered many minutes later. We found that exposure to
extended blocks of emotion-evoking stimuli induced a lasting emo-
tional state that enhanced participants’ later recollection of neutral
images encountered 9 to 33 min later, suggesting that the impact of
emotion carried over into and biased subsequent stimulus process-
ing and encoding. We directly queried the persistence or carry-over
of an emotional brain state into subsequent neutral encoding blocks
by analyzing global multivoxel patterns across the brain and activ-
ity in emotion- and arousal-related brain regions. First, we found
that LF fluctuations of the BOLD signal previously shown to track
behavioral or neural states27–29 carried over from emotional encod-
ing into subsequent neutral encoding blocks. These included global
multi voxel amygdala–whole-brain connectivity patterns, as well as
correlated activity in targeted circuits (amygdala–anterior hippoc-
ampus and ventral anterior insula network). Second, we found that
brain regions exhibiting successful encoding effects for neutral stimuli
were also modulated by preceding emotional experiences. Thus, both
large-scale patterns and overall levels of BOLD activity supporting
recollection-based memory formation were reinstated during neutral
encoding after emotional encoding blocks. Together, these findings
provide evidence that the induction of a relatively lasting emotional
state was associated with brain states that could later be reinstated,
biasing the way future neutral events were encoded and potentially
imbuing neutral experiences with emotional properties that enhance
their recollection.
Previous studies of emotion’s influence on memory have demon-
strated that the neurohormonal changes underlying physiological
arousal are critical for mediating emotion-enhanced memory7,8, spe-
cifically subjective recollection43, although these might not be the
only contributing factors. In the present study, we induced an emo-
tional state through the use of stimuli previously rated as subjectively
negative and arousing (compared to stimuli rated as neutral).
We then used SCL, a noninvasive assessment of autonomic arousal,
as our primary measure of emotion25. Notably, increased levels of
emotional arousal induced by the exposure to emotional stimuli
Left IT cortexInferior PFC
Left amygdala Right perirhinal
cortex
y = –6 x = 34
y = 22 y = –60
a b
R R
K K
* *
*
E N
N E
BOLD responseBOLD response
Figure 6 Subsequent recollection-based memory carry-over effects.
(a) Schematic of response patterns required for a subsequent recollection
carry-over effect, or reinstatement of emotional subsequent recollection
effects during later neutral encoding (six-way conjunction analysis).
Regions showed significant subsequent recollection effects or significantly
greater activity estimates for R vs. K trials during emotional encoding
(both encoding orders) and neutral encoding when neutral stimuli were
encountered after emotional stimuli (E→N encoding order, red bars).
Regions also showed significantly greater R – K differences between the
three conditions with arrows. Individual contrasts were thresholded at
*P < 0.05. (b) Brain regions showing emotional subsequent recollection
effects that carry over during later neutral encoding, isolated from the
conjunction analysis shown in a. Six regions emerged from this analysis
(shown in red): left amygdala (P = 0.013), right perirhinal cortex
(P = 0.004), left inferior temporal (IT) cortex (P = 0.0088) and three
inferior prefrontal cortex (PFC) regions (two shown; left P = 0.0006,
right P = 0). Results are FWE-corrected at P < 0.05, and the resulting
map was smoothed for visualization. Statistical map is displayed on the
group template anatomical image and can be found in MNI space on
neurovault.org (http://neurovault.org/collections/JOYXPMRX/).

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
nature neurOSCIenCe ADVANCE ONLINE PUBLICATION 7
a r t I C l e S
persisted in time and likely influenced the encoding, and perhaps
consolidation, of subsequently encountered neutral stimuli, resulting
in increased levels of subjective recollection when the neutral memory
was assessed 6 h later. This enhancement in recollection of neutral
stimuli was observed across two independent data sets. Crucially,
mirroring our behavioral findings, SCL was distinctly higher when
neutral stimuli were encountered 9 to 33 min after emotional stim-
uli versus before emotional stimuli as well as after neutral stimuli.
For these reasons, together with prior evidence linking emotion-
related neurohormonal changes with subjective recollection43, we
believe the subsequent influence of emotion on memory for future
neutral events may be mediated by noradrenergic activation triggered
by the emotional images.
It is important to note that our finding of enhanced memory for
neutral stimuli encoded after emotional stimuli is similar to, but dis-
tinct from, prior demonstrations that emotional arousal or adrenergic
agonists administered after the encoding of neutral stimuli retroac-
tively enhances later memory for neutral information5,16,18,44,45. Post-
learning arousal is proposed to heighten memory consolidation of
preceding neutral information5,16,45. Here we show that pre-encoding
emotional arousal that persists during exposure to neutral stimuli can
prospectively influence both their initial encoding and their recollec-
tion 6 h later. Notably, BOLD activity during neutral encoding tens of
minutes after emotional encoding in many respects resembled BOLD
activity when participants viewed and processed emotional stimuli.
Correspondingly, neutral stimuli encoded during heightened arousal
showed the same mnemonic profile as emotional stimuli (enhanced
recollection) relative to neutral stimuli encoded before and outside
the context of emotional arousal. Taken together, these findings paint
a picture of emotion as capable of biasing neutral memory via multiple
mechanisms: retroactively enhancing the consolidation of previously
encoded information, as well as prospectively biasing future brain
states and mechanisms underlying the encoding of new experiences
into memory.
Our findings that LF connectivity (amygdala–hippocampus and
ventral anterior insula network correlations), subsequent recollec-
tion effects in individual brain regions (amygdala, perirhinal cortex,
inferior prefrontal cortex, inferior temporal cortex) and patterns of
hippocampal recollection effects characteristic of emotional encoding
carried over into subsequent neutral encoding blocks are consistent
with prior work showing that emotion is related to BOLD activity
and connectivity in similar regions10,11,37, 38,41,46,47 (assessed using a
reverse analysis in neurosynth.org42). Notably, however, the present
data extend these findings in several ways. First, consistent with a
carry-over hypothesis, brain regions related to emotional processing
and memory formation were not only engaged when participants
viewed emotional stimuli but also during the subsequent encoding
of neutral stimuli. This result suggests that emotion can bias BOLD
activity over an extended time period (at least 9 to 33 min later).
Second, differences in LF connectivity during emotional and neutral
encoding in the N-E encoding order suggested that emotion mod-
ulated not only trial-by-trial evoked brain activity but also lower-
frequency background BOLD activity. This modulation of background
BOLD activity by emotion complements previous findings showing
that background BOLD activity reflects specific types of information
processing, such as distinct memory states29 and the processing of
distinct stimulus classes27,28.
Although the present study found evidence that an extended
emotional experience could bias future brain states and memory
encoding, we note that it is unclear which specific aspects of our
experimental design were critical for this effect to emerge. In our
design, encoding blocks lasted for 23 min; thus, it is unclear how
much time is necessary to induce a state of emotional arousal that will
persist and bias future behavior. Moreover, it is also unclear how long
emotional arousal may potentially last and whether this is related to
the duration of the initial arousal induction. It is also noteworthy that
participants performed the same task when they encountered emo-
tional and neutral stimuli in our study (rating of visual complexity), so
it is unknown whether this similarity in task context was necessary for
eliciting a carry-over of an emotional state into later neutral encod-
ing. Lastly, it is also unknown how expectations or explicit strategies
developed by participants may have facilitated the reinstatement of
emotional arousal or brain states during neutral encoding. Future
work is needed to understand how these different factors contributed
to the present findings.
The present results add to our understanding of the many ways
emotion can influence memory for unrelated neutral events. When
examining memories for events themselves, there is evidence that
more neutral details of an emotional event may be less well remem-
bered48,49, suggesting a trade-off in memory49,50. However, the
influence of emotion on memory can also extend over time, between
events. Not only can emotional arousal following a neutral event influ-
ence the storage of that event but, to the extent that arousal persists, it
can also influence memory for future neutral events that are tempo-
rally and semantically distinct. Our results suggest that this prospec-
tive memory enhancement may be due to a carry-over of the brain
states that underlie arousal and its influence on memory.
METHODS
Methods, including statements of data availability and any associated
accession codes and references, are available in the online version of
the paper.
Note: Any Supplementary Information and Source Data files are available in the
online version of the paper.
ACKNOWLEDGMENTS
We thank E. Bar-David for expert assistance with data collection for the fMRI
study and A. Patil, M. Kelemu, C. Brennan and D. Antypa for assistance with
behavioral data collection. This work was supported by Dart Neuroscience (L.D.);
NIMH grants MH074692 (L.D.), MH062104 (E.A.P.) and MH092055 (A.T.); and
by grants from the Swiss National Science Foundation (PZ00P1_137126), the
German Research Foundation (DFG RI 1894/2-1), and the European Community
Seventh Framework Programme (FP7/2007-2013) under grant agreement
334360 to U.R.
AUTHOR CONTRIBUTIONS
A.T., U.R., E.A.P. and L.D. designed the experiment and wrote the paper. A.T. and
U.R. collected and analyzed the data.
COMPETING FINANCIAL INTERESTS
The authors declare no competing financial interests.
Reprints and permissions information is available online at http://www.nature.com/
reprints/index.html.
1. LaBar, K.S. & Cabeza, R. Cognitive neuroscience of emotional memory. Nat. Rev.
Neurosci. 7, 54–64 (2006).
2. Phelps, E.A. & Sharot, T. How (and why) emotion enhances the subjective sense
of recollection. Curr. Dir. Psychol. Sci. 17, 147–152 (2008).
3. Ochsner, K.N. Are affective events richly recollected or simply familiar? The
experience and process of recognizing feelings past. J. Exp. Psychol. Gen. 129,
242–261 (2000).
4. Kensinger, E.A. & Corkin, S. Memory enhancement for emotional words: are
emotional words more vividly remembered than neutral words? Mem. Cognit. 31,
1169–1180 (2003).
5. Anderson, A.K., Wais, P.E. & Gabrieli, J.D.E. Emotion enhances remembrance of
neutral events past. Proc. Natl. Acad. Sci. USA 103, 1599–1604 (2006).
6. Knight, M. & Mather, M. Reconciling findings of emotion-induced memory
enhancement and impairment of preceding items. Emotion 9, 763–781 (2009).

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
8 ADVANCE ONLINE PUBLICATION nature neurOSCIenCe
a r t I C l e S
7. McGaugh, J.L. Memory--a century of consolidation. Science 287, 248–251 (2000).
8. McGaugh, J.L. & Roozendaal, B. Role of adrenal stress hormones in forming lasting
memories in the brain. Curr. Opin. Neurobiol. 12, 205–210 (2002).
9. Phelps, E.A. Human emotion and memory: interactions of the amygdala and
hippocampal complex. Curr. Opin. Neurobiol. 14, 198–202 (2004).
10. Murty, V.P., Ritchey, M., Adcock, R.A. & LaBar, K.S. fMRI studies of successful
emotional memory encoding: A quantitative meta-analysis. Neuropsychologia 48,
3459–3469 (2010).
11. Ritchey, M., Dolcos, F. & Cabeza, R. Role of amygdala connectivity in the persistence
of emotional memories over time: an event-related FMRI investigation. Cereb. Cortex
18, 2494–2504 (2008).
12. Richardson, M.P., Strange, B.A. & Dolan, R.J. Encoding of emotional memories
depends on amygdala and hippocampus and their interactions. Nat. Neurosci. 7,
278–285 (2004).
13. Popa, D., Duvarci, S., Popescu, A.T., Léna, C. & Paré, D. Coherent amygdalocortical
theta promotes fear memory consolidation during paradoxical sleep. Proc. Natl
Acad. Sci. USA 107, 6516–6519 (2010).
14. Dunsmoor, J.E., Murty, V.P., Davachi, L. & Phelps, E.A. Emotional learning selectively
and retroactively strengthens memories for related events. Nature 520, 345–348
(2015).
15. Hurlemann, R. et al. Noradrenergic modulation of emotion-induced forgetting and
remembering. J. Neurosci. 25, 6343–6349 (2005).
16. Schwarze, U., Bingel, U. & Sommer, T. Event-related nociceptive arousal enhances
memory consolidation for neutral scenes. J. Neurosci. 32, 1481–1487 (2012).
17. Packard, M.G. & Teather, L.A. Amygdala modulation of multiple memory systems:
hippocampus and caudate-putamen. Neurobiol. Learn. Mem. 69, 163–203 (1998).
18. Roozendaal, B., Castello, N.A., Vedana, G., Barsegyan, A. & McGaugh, J.L.
Noradrenergic activation of the basolateral amygdala modulates consolidation of
object recognition memory. Neurobiol. Learn. Mem. 90, 576–579 (2008).
19. Strange, B.A., Hurlemann, R. & Dolan, R.J. An emotion-induced retrograde amnesia
in humans is amygdala- and beta-adrenergic-dependent. Proc. Natl. Acad. Sci. USA
100, 13626–13631 (2003).
20. Segal, S.K., Stark, S.M., Kattan, D., Stark, C.E.L. & Yassa, M.A. Norepinephrine-
mediated emotional arousal facilitates subsequent pattern separation. Neurobiol.
Learn. Mem. 97, 465–469 (2012).
21. Henckens, M.J.A.G., Hermans, E.J., Pu, Z., Joëls, M. & Fernández, G. Stressed
memories: how acute stress affects memory formation in humans. J. Neurosci. 29,
10111–10119 (2009).
22. van Marle, H.J.F., Hermans, E.J., Qin, S. & Fernández, G. Enhanced resting-state
connectivity of amygdala in the immediate aftermath of acute psychological stress.
Neuroimage 53, 348–354 (2010).
23. Veer, I.M. et al. Beyond acute social stress: increased functional connectivity
between amygdala and cortical midline structures. Neuroimage 57, 1534–1541
(2011).
24. Eryilmaz, H., Van De Ville, D., Schwartz, S. & Vuilleumier, P. Impact of transient
emotions on functional connectivity during subsequent resting state: a wavelet
correlation approach. Neuroimage 54, 2481–2491 (2011).
25. Sequeira, H., Hot, P., Silvert, L. & Delplanque, S. Electrical autonomic correlates
of emotion. Int. J. Psychophysiol. 71, 50–56 (2009).
26. Rajaram, S. Remembering and knowing: two means of access to the personal past.
Mem. Cognit. 21, 89–102 (1993).
27. Al-Aidroos, N., Said, C.P. & Turk-Browne, N.B. Top-down attention switches coupling
between low-level and high-level areas of human visual cortex. Proc. Natl. Acad.
Sci. USA 109, 14675–14680 (2012).
28. Norman-Haignere, S.V., McCarthy, G., Chun, M.M. & Turk-Browne, N.B. Category-
selective background connectivity in ventral visual cortex. Cereb. Cortex 22,
391–402 (2012).
29. Duncan, K., Tompary, A. & Davachi, L. Associative encoding and retrieval are
predicted by functional connectivity in distinct hippocampal area CA1 pathways.
J. Neurosci. 34, 11188–11198 (2014).
30. Phelps, E.A. Emotion and cognition: insights from studies of the human amygdala.
Annu. Rev. Psychol. 57, 27–53 (2006).
31. Hermans, E.J. et al. How the amygdala affects emotional memory by altering brain
network properties. Neurobiol. Learn. Mem. 112, 2–16 (2014).
32. McGaugh, J.L. The amygdala modulates the consolidation of memories of emotionally
arousing experiences. Annu. Rev. Neurosci. 27, 1–28 (2004).
33. Pitkänen, A., Pikkarainen, M., Nurminen, N. & Ylinen, A. Reciprocal connections
between the amygdala and the hippocampal formation, perirhinal cortex,
and postrhinal cortex in rat. A review. Ann. NY Acad. Sci. 911, 369–391
(2000).
34. Saunders, R.C., Rosene, D.L. & Van Hoesen, G.W. Comparison of the efferents
of the amygdala and the hippocampal formation in the rhesus monkey: II.
Reciprocal and non-reciprocal connections. J. Comp. Neurol. 271, 185–207
(1988).
35. Stefanacci, L., Suzuki, W.A. & Amaral, D.G. Organization of connections between
the amygdaloid complex and the perirhinal and parahippocampal cortices in
macaque monkeys. J. Comp. Neurol. 375, 552–582 (1996).
36. Fanselow, M.S. & Dong, H.-W. Are the dorsal and ventral hippocampus functionally
distinct structures? Neuron 65, 7–19 (2010).
37. Dolcos, F., LaBar, K.S. & Cabeza, R. Interaction between the amygdala and the
medial temporal lobe memory system predicts better memory for emotional events.
Neuron 42, 855–863 (2004).
38. Touroutoglou, A., Hollenbeck, M., Dickerson, B.C. & Feldman Barrett, L. Dissociable
large-scale networks anchored in the right anterior insula subserve affective
experience and attention. Neuroimage 60, 1947–1958 (2012).
39. Uddin, L.Q., Kinnison, J., Pessoa, L. & Anderson, M.L. Beyond the tripartite
cognition-emotion-interoception model of the human insular cortex. J. Cogn.
Neurosci. 26, 16–27 (2014).
40. Kurth, F., Zilles, K., Fox, P.T., Laird, A.R. & Eickhoff, S.B. A link between the
systems: functional differentiation and integration within the human insula revealed
by meta-analysis. Brain Struct. Funct. 214, 519–534 (2010).
41. Chang, L.J., Yarkoni, T., Khaw, M.W. & Sanfey, A.G. Decoding the role of the insula
in human cognition: functional parcellation and large-scale reverse inference. Cereb.
Cortex 23, 739–749 (2013).
42. Yarkoni, T., Poldrack, R.A., Nichols, T.E., Van Essen, D.C. & Wager, T.D. Large-scale
automated synthesis of human functional neuroimaging data. Nat. Methods 8,
665–670 (2011).
43. Rimmele, U., Lackovic, S.F., Tobe, R.H., Leventhal, B.L. & Phelps, E.A. Beta-
adrenergic blockade at memory encoding, but not retrieval, decreases the subjective
sense of recollection. J. Cogn. Neurosci. 28, 895–907 (2016).
44. Nielson, K.A., Yee, D. & Erickson, K.I. Memory enhancement by a semantically
unrelated emotional arousal source induced after learning. Neurobiol. Learn. Mem.
84, 49–56 (2005).
45. Nielson, K.A. & Powless, M. Positive and negative sources of emotional arousal
enhance long-term word-list retention when induced as long as 30 min after
learning. Neurobiol. Learn. Mem. 88, 40–47 (2007).
46. Kensinger, E.A., Addis, D.R. & Atapattu, R.K. Amygdala activity at encoding
corresponds with memory vividness and with memory for select episodic details.
Neuropsychologia 49, 663–673 (2011).
47. Dougal, S., Phelps, E.A. & Davachi, L. The role of medial temporal lobe in item
recognition and source recollection of emotional stimuli. Cogn. Affect. Behav.
Neurosci. 7, 233–242 (2007).
48. Rimmele, U., Davachi, L., Petrov, R., Dougal, S. & Phelps, E.A. Emotion enhances
the subjective feeling of remembering, despite lower accuracy for contextual details.
Emotion 11, 553–562 (2011).
49. Kensinger, E.A., Garoff-Eaton, R.J. & Schacter, D.L. Effects of emotion on memory
specificity: memory trade-offs elicited by negative visually arousing stimuli. J. Mem.
Lang. 56, 575–591 (2007).
50. Mather, M. Emotional arousal and memory binding. Perspect. Psychol. Sci. 2,
33–52 (2007).

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
nature neurOSCIenCe
doi:10.1038/nn.4468
ONLINE METHODS
Subjects. fMRI Study. Fifty right-handed native English speakers with normal or
corrected-to-normal vision participated in the study. Six subjects were excluded,
two due to a high rate of spike artifacts in the fMRI data and four due to very poor
memory performance (miss or false alarm rates > two s.d. above the mean of all
subjects). The average age of the remaining 44 subjects was 24.1 years (range:
19–34) and included 23 male and 21 female participants. Informed consent was
obtained from all participants in a manner approved by the institutional review
board at New York University. No a priori sample size calculations were per-
formed, but the sample size was chosen based on common practices in the field43.
Subjects were assigned to groups (E-N and N-E encoding orders) in alternating
fashion as the data was collected. Age and gender were approximately matched
across encoding orders (E-N: 11 female, 11 male, mean age = 24.9; N-E: 10 female,
12 male, mean age = 23.4).
Behavioral groups. Eighty-two participants with normal or corrected-to-normal
vision participated in the behavioral studies. The E-N and N-E encoding order
groups were collected simultaneously, with participants assigned to groups in
alternating fashion as the data was collected (n = 55 participants total). Most of
the participants were recruited at New York University (31/55, 16 in E-N encod-
ing order, 15 in N-E encoding order) while the rest of the participants were
recruited at the University of Geneva (24/55 participants, 12 in each encoding
order). The N-N encoding order was run as a separate study and consisted of 27
participants, all recruited at New York University. In total, 11 participants were
excluded: one who did not follow the instructions (N-N group) and ten due to
very poor memory performance (miss or false alarm rates > two s.d. above the
mean of all subjects; three N-N participants, four E-N participants and three N-E
participants). Of the remaining 71 participants, 46 were female (mean age: 24.5
years, range: 18–33), resulting in a total of 23 participants for the N-N encoding
order and 24 participants for each of the N-E and E-N encoding orders. Age and
gender were approximately matched across the N-N, E-N and N-E encoding
orders (N-N: 18 female, mean age = 24.96; E-N: 13 female, mean age = 24; N-E:
15 female, mean age = 24.5). Informed consent was obtained from all participants
in a manner approved by the institutional review boards at New York University
and the University of Geneva.
Procedure. fMRI study (E-N and N-E encoding orders). Before entering the scan-
ner, subjects were informed about the scanning procedures and performed a
brief practice session of the image complexity task to be performed during the
encoding sessions as well as a one-back task used for localizer scans. Prior to
scanning, electrodes were placed on the index and middle fingers of the subject’s
right hand to measure galvanic SCL.
During the scanning session, subjects performed incidental encoding blocks
(emotional and neutral encoding) interleaved with rest scans (Fig. 1a). The order
of the encoding blocks was counterbalanced across subjects, such that half of
the subjects performed emotional followed by neutral encoding (referred to as
the E-N encoding order) and the other half of the subjects performed neutral
before emotional encoding (referred to as the N-E encoding order). Additional
scans were collected after the last rest scan but were not used in the present
analyses (localizer scans in which subjects viewed scenes, objects and other stimu-
lus categories). A high-resolution anatomical scan was collected after all func-
tional scans. After the scanning session, subjects returned to the lab to perform
a surprise memory test for stimuli encountered during emotional and neutral
encoding, approximately 6 h after the last encoding session. Subjects were not
informed before the scanning session that their memory would be tested for
stimuli encountered during the encoding sessions.
By necessity, data collection and analysis were not performed blind to the
conditions of the experiments. However, subjective outcome assessments were
not performed.
Behavioral groups (N-N, N-E and E-N encoding orders). The same experimental
procedures used in the fMRI study were used in the behavioral groups, except
that all procedures took place outside of the MRI scanner and participants in the
N-N encoding order incidentally encoded two blocks of neutral stimuli instead
of blocks of emotional and neutral stimuli. Subjects were exposed to an auditory
stimulus of white noise during encoding and during rest periods in between
encoding blocks to resemble the constant noise during fMRI scanning. A surprise
memory test for neutral stimuli was administered with the same delay (6 h) as
the fMRI study.
Encoding sessions. All fMRI participants and E-N and N-E behavioral groups
incidentally encoded both emotional and neutral stimuli in separate blocks (emo-
tional and neutral encoding). In the N-N behavioral group, subjects incidentally
encoded two separate blocks of neutral stimuli. During encoding, subjects rated
the visual complexity of complex scenes. Each block consisted of three 7.7-min
scans or epochs and contained a total of 99 trials (33 trials in each scan). The
order of stimulus presentation was randomized for each subject. Each trial lasted
for 14 s and included a 1-s fixation cue, presentation of a complex scene for 2 s, a
response window of 1 s, followed by the performance of a baseline ‘arrows’ task for
10 s in between trials51. Subjects were instructed to view each image for as long as
it remained on the screen and examine the visual complexity of each image. The
response window contained the words “Complexity?” and “High Medium Low.”
Subjects pressed their ring, middle or index finger of their left hand to indicate
complexity ratings of high, medium or low, respectively. During the arrows task,
a series of arrows were presented pointing either left, up, or right on the screen.
Subjects were instructed to push their ring finger when the arrow pointed to the
left, their middle finger when the arrow pointed up and their index finger when
the arrow pointed to the right.
Memory test. Approximately 6 h after the end of the last encoding session,
all subjects completed two separate memory tests (one for each encoding
session) outside of the scanner. The order of the memory tests matched the order
of their presentation during encoding, to roughly equate the study–test inter-
vals across the two encoding sessions. Each trial consisted of the presentation
of a complex scene, and subjects performed a remember (R), know (K) or new
(N) judgment.
Subjects were trained on R, K and N decisions before the start of the memory
test26. After reading detailed instructions, participants explained the meaning
of R and K judgments in their own words. During the practice trials, subjects
indicated why they judged a scene as remembered or known out loud to the
experimenter. The recognition test was administered once the participants had
correctly understood the instructions, i.e., they judged a scene as remembered
when it brought back to mind a specific detail from the episodic context in which
the scene had been experienced, such as a sensory detail, a thought or a feeling.
Memory tests were self-paced and contained 198 trials each. Half of the stimuli
(99) in the memory test were novel distractor scenes and half were those pre-
sented during the encoding session. For the fMRI study and the E-N and N-E
behavioral groups, the stimulus sets presented at encoding and test were coun-
terbalanced across encoding orders: for half of the subjects stimulus set 1 was
used during the encoding session and stimulus set 2 was used as distractors in the
memory test; the other half of the subjects viewed stimulus set 2 during encoding
set 1 as distractors in the memory test. For the N-N behavioral group, four neutral
stimulus sets were created and counterbalanced across subjects with respect to
the first versus second stimulus set during encoding and with respect to serving
as encoding or distractor sets during memory testing. Stimulus sets for both the
fMRI and behavioral studies were matched for image complexity and the presence
of people, animals, inanimate objects and scenes.
Stimuli. The complex scenes used in all encoding sessions spanned 465 ×
620 pixels.
fMRI study and E-N and N-E behavioral groups. The complex scenes used in
the encoding sessions were drawn from the International Affective Picture Set
(IAPS)52 as well as a few in-house stimuli (Supplementary Table 1). Stimuli
were classified as emotional or neutral based on the normative ratings provided
for emotional arousal (emotional scenes: M = 5.65, s.d. = 0.82; neutral scenes:
M = 3.92, s.d. = 0.96) and valence (emotional scenes: M = 2.83, s.d. = 0.92; neutral
scenes: M = 5.94, s.d. = 1.00) assessed with the Self-Assessment Manikin (SAM)
scale (1 = calm, 9 = excited; 1 = unhappy, 9 = happy)52. Stimuli used for emo-
tional and neutral encoding as well as for the encoding versus distractor sets in
the memory test were matched for image complexity and the presence of people,
animals, inanimate objects and scenes.
N-N behavioral group. In addition to the 198 neutral stimuli used in the fMRI
study, another 198 neutral complex scenes were added to the stimulus set to
construct two blocks of stimuli for neutral encoding for the N-N encoding order.
Stimuli were taken from the IAPS52, the Necki Affective Picture System53 and our
own images. As in the fMRI study, stimuli were selected based on normative rat-
ings provided for emotional arousal (M = 3.98, s.d. = 0.79) and valence (M = 5.99,

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
nature neurOSCIenCe doi:10.1038/nn.4468
s.d. = 0.95) assessed with the SAM scale described above. Note that arousal and
valence ratings for neutral stimuli were well-matched across the E-N/N-E encod-
ing orders and the N-N behavioral group. The stimuli were classified into four sets
of 99 neutral photos, which were matched for image complexity, the presence of
people/animals, inanimate objects and scenes and which did not differ between
emotional arousal and valence (all P-values > 0.85).
Rest scans. Each rest scan lasted for 9 min. Subjects were instructed to close
their eyes and simply rest and think about anything that they wanted but to try to
remain awake54,55. We used only the fMRI data from the first rest scan in the cur-
rent analyses, to isolate regions in the ventral anterior insula (vAI) network38.
Behavioral analyses. Memory for emotional and neutral stimuli was assessed
by examining proportions of hits and false alarms for each encoding order (E-N,
N-E and N-N), collapsed across R and K responses (overall memory accuracy),
as well as separately for R and K responses. To analyze memory as a function of
emotion and encoding order (fMRI study and behavioral groups), mixed-effects
ANOVAs were performed on corrected memory accuracy rates (proportion of
hits minus false alarms) with a within-subjects factor of emotion (emotional and
neutral stimuli) and a between-subjects factor of encoding order (E-N and N-E
encoding orders). Separate ANOVAs were performed for all hits and false alarms
(combined R and K responses) and for R and K responses separately. Follow-up t-
tests were performed to investigate differences in memory between emotional and
neutral stimuli within encoding orders (paired t-tests) as well as between encod-
ing orders (unpaired t-tests). To assess whether neutral memory was enhanced
when they were encoded after emotional versus after neutral stimuli (E-N ver-
sus N-N encoding orders), we performed planned comparisons between cor-
rected neutral memory accuracy (combined R and K responses and R responses
alone) in the E-N encoding order versus both blocks in the N-N encoding order
(behavioral study) using an unpaired t-test. All memory data were determined
to be normally distributed (using a Lilliefors test; see “Statistics” below) with the
exception of overall memory accuracy (combined R and K responses) for the
second block of neutral stimuli in the N-N encoding order. We thus additionally
performed nonparametric permutation tests when statistically evaluating this
data (when comparing neutral memory in the N-N encoding order to the E-N
encoding order, for both the fMRI and behavioral studies). To do so, we com-
puted the true difference in memory across each of these comparisons. We then
derived a null distribution of the difference in memory across these conditions
by performing 10,000 null simulations in which the group labels (E-N or N-N
encoding order labels) were randomly shuffled and the difference in memory for
each null permutation of the data was computed. The true difference in memory
was compared to this null distribution to estimate a P-value for these comparisons
(see “Results” and Supplementary Fig. 1).
Skin conductance level analyses. In order to examine differential levels of arousal
based on encoding order, mean skin conductance levels (SCL) were analyzed,
which have been related to sympathetic nervous system activation (see ref. 25
for a review). Specifically, mean SCL was computed in each subject for each rest
and encoding scan (fMRI study) or each rest and encoding block (N-N behav-
ioral study). Relative SCL changes were then used as our dependent measure of
interest, which were computed for each encoding and rest block by subtracting
mean SCL measured during the pre-encoding baseline rest period from mean
SCL during each encoding and rest block (for both the fMRI and N-N behavioral
group). This procedure of computing SCL changes relative to a baseline period
controlled for individual differences in SCL and allowed us to compare SCL
across different groups of subjects (E-N encoding order versus other encoding
orders). To compare relative SCL during emotional versus neutral encoding as a
function of encoding order (in the fMRI study), a mixed-effects two-way ANOVA
was performed with a within-subjects factor of emotion (emotional and neutral
encoding blocks) and a between-subjects factor of encoding order (E-N/N-E).
Follow-up t-tests were used to examine elevated SCL relative to baseline rest
before encoding (t-test in relative SCL versus zero, corrected for multiple com-
parisons across encoding and rest blocks) and differences in relative SCL during
neutral encoding in the E-N versus N-E and N-N encoding orders (unpaired
t-tests). One-tailed t-tests were used to assess differences in relative SCL during
neutral encoding as a function of encoding order, as we specifically hypothesized
that the presence of prior emotional encoding would enhance SCL during neutral
encoding compared to the opposite encoding order (i.e., neutral encoding SCL
would be higher for the E-N versus N-E and N-N encoding orders). However,
relative SCL data distributions were found to be not normally distributed for sev-
eral blocks of interest. We thus also performed nonparametric permutation tests
for all statistical tests involving SCL. The same methods were applied as in the
permutation tests as described for the memory analyses (see “Behavioral analy-
ses”). Null simulations were performed in which the data labels were randomly
permuted and the comparison of interest was computed. The true difference of
interest was then compared to the null distribution generated for that particular
test. To assess the significance of the encoding order by emotion interaction
observed in SCL, we computed the effect of interest as the difference in emotional
versus neutral encoding SCL across encoding orders (E-N emotional encoding
SCL – E-N neutral encoding SCL – (N-E emotional encoding SCL – N-E neutral
encoding SCL)) and compared this true difference to a null distribution derived
from permuting E-N and N-E encoding order labels for each participant.
MRI data acquisition. Scanning was performed using a 3T Siemens Allegra MRI
system with a whole-head coil. Functional (BOLD) data were collected using
a gradient-echo planar pulse (EPI) sequence (repetition time (TR) = 2 s, echo
time = 30 ms; field of view = 192 mm; 31 slices oriented perpendicular to the long axis
of the hippocampus; 3 × 3 × 3 mm voxel size; 0.6 mm interslice gap; flip angle = 80°).
Functional scans contained 231 or 270 volumes or TRs for the encoding and rest
scans, respectively, after discarding the first four volumes to allow for T1 equili-
bration. High-resolution T1-weighted (magnetization-prepared rapid-acquisition
gradient echo) images were acquired after the last functional scan.
MRI preprocessing. The imaging data were preprocessed using SPM5 (Wellcome
Department of Cognitive Neuroscience, University College London, London,
UK). The BOLD data were first corrected for differences in slice timing acquisi-
tion followed by motion correction across all runs and the removal of low fre-
quency trends (< 0.009 Hz) from each scan. Each subject’s functional data were
then co-registered to their own T1-weighted anatomical image. For the definition
and analysis of anatomically defined regions of interest (ROIs; for example, amy-
gdala) as well as the analysis of whole-brain activity and connectivity patterns, the
functional data were analyzed in a subject-specific space and spatially smoothed
with a 6-mm full-width at half-maximum (FWHM) isotropic Gaussian kernel.
For group-level analyses, a group-level anatomical template was created using
advanced normalization tools (ANTs)56 based on the T1-weighted anatomical
images of all 44 subjects. Statistical maps from native subject spaces were then
transformed into this template space for group analyses.
Noise corrections were applied to the BOLD data from all encoding and rest
scans based on methods from Behzadi et al. (aCompCor)57. A nuisance ROI was
generated from white matter (WM) and cerebrospinal fluid (CSF) probabilistic
maps (see “Anatomical masks,” below). The WM and CSF maps were converted
to functional resolution and thresholded at probability values of 0.98 or 0.99 for
the WM maps and at 0.97, 0.98 or 0.99 for the CSF maps. After thresholding,
the WM mask was eroded by two voxels to avoid contaminating the WM signal
with GM signal57. We used an iterative procedure for defining the probability
threshold for each subject, such that the highest threshold was used with the
stipulation that the resulting CSF or eroded WM mask contained a minimum
of 10 voxels at functional resolution. On each iteration (for a given probability
threshold) we additionally excluded voxels from the nuisance ROI if their time-
course during any one encoding scan was modestly correlated (P < 0.2) with a
model of task activity57 (with task activity modeled as a 4-s boxcar corresponding
to fixation, stimulus presentation and the response window, convolved with the
canonical hemodynamic response function, HRF). The resulting voxels from
the thresholded WM and CSF maps were then combined to form one nuisance
ROI for each subject.
In order to extract dominant signals accounting for substantial variance in the
BOLD signal from the nuisance ROI, principal components analysis (PCA) was
performed on the BOLD data from the nuisance ROI. PCA was performed on the
z-scored BOLD data (note that LF trends were already removed from the data),
separately for each encoding scan and rest scan. We then performed simulations
to determine the number of principal components (PCs) to extract from the
nuisance ROI data for each scan. A null distribution of the expected eigenvalues
was generated separately for the encoding and rest data for each subject by per-
forming PCA on normally distributed data of equal rank to the encoding and rest

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
nature neurOSCIenCe
doi:10.1038/nn.4468
nuisance ROI data57. PCs from the true nuisance ROI data were then selected for
a given encoding or rest scan if their associated eigenvalues exceeded the 99%
confidence interval of the eigenvalues from the simulations. However, if different
numbers of PCs were chosen across any of the encoding and rest scans for each
subject, then the minimum number of PCs selected across all scans was used
for all scans. This procedure ensured that removal of any nuisance signals was
consistent across all encoding and rest scans, to avoid differentially biasing the
data in any particular scan. After selecting the number of PCs for the encoding
and rest data for the nuisance ROI, the associated temporal projection of each
selected PC was removed from the whole-brain BOLD data, in conjunction with
the six motion parameters estimated during processing, using linear regression
in a voxelwise fashion. TRs surrounding time periods of sudden motion (see
below for definition) were not included in the estimation of beta coefficients in
this regression58,59.
Additional measures were taken to address the influence of motion on BOLD
data and connectivity estimates, which was not completely addressed by the nui-
sance regression approach taken above60,61. All encoding and rest BOLD data
were temporally censored or ‘scrubbed’, such that TRs surrounding time peri-
ods of sudden motion were removed from the analyzed data61. To accomplish
this, we calculated frame-wise displacement (FD) of head motion based on the
motion parameters estimated during preprocessing61, and TRs including and
surrounding (1 TR before, 2 TRs after) FD values greater than 0.5 mm were
flagged to be scrubbed.
The scrubbing procedure was first performed at multiple levels: first, in the
estimation of beta coefficients when regressing nuisance signals from the BOLD
data58,59. For the encoding data analyzed via the general linear model (GLM;
analyses shown in Figs. 5 and 6 and Supplementary Figs. 2 and 3), the flagged
time points were removed from the BOLD data and the design matrices. For LF
connectivity analyses, the flagged data points were removed after the time-series
were low-pass filtered but before the estimation of correlation values. After this
scrubbing procedure was performed, a substantial number of time points were
missing for a small subset of subjects. Subjects were excluded from LF connec-
tivity analyses if less than 5 min of data were remaining in any given scan60. For
the encoding data, this included n = 2 subjects for emotional encoding and n = 1
subject for neutral encoding; for the data from the first rest scan, this included n
= 1 subject. Since a major goal was to compare connectivity between emotional
and neutral encoding as a function of encoding order (see below), we excluded
n = 4 total subjects from LF connectivity analyses (2 from each encoding order)
to obtain equal bin sizes across E-N and N-E encoding orders (one additional
subject from the E-N encoding order that had the most censored time points
was excluded).
Anatomical ROI definition. FSL’s FIRST segmentation was used to define the
amygdala and hippocampus, based on each subject’s high-resolution T1-weighted
scan. All ROIs were manually inspected and edited when appropriate to ensure
proper definitions according to Pruessner et al.62. The hippocampus was then
divided into anterior, middle and posterior portions by splitting the coronal sec-
tions into thirds along the anterior–posterior axis.
Anatomical masks. SPM5 was used to create white matter (WM), cerebrospinal
fluid (CSF) and gray matter (GM) probabilistic maps based on each subject’s
T1-weighted anatomical image. To create a mask of voxels in the cerebral cortex,
each subject’s anatomical image was parcellated using FreeSurfer, and all voxels
labeled in any structure in the cerebral cortex were combined to create a mask
of the cortex63. To create a mask of voxels in subcortical structures, FSL’s FIRST
segmentation was used and all voxels in any subcortical structure were combined
to form one subcortical mask.
Ventral anterior insula network analyses. The BOLD data from the first rest
scan was used to define regions in the ventral anterior insula (vAI) network38. To
accomplish this, the rest data were low-pass filtered (< 0.1 Hz) after the removal of
nuisance signals. MNI coordinates from a previously reported study were used to
define a seed region in the right vAI (ref. 38). To extract the time-course of the vAI
from these coordinates, we co-registered our group template brain to the MNI
template brain, computed the location of the MNI coordinates in our group tem-
plate space and then extracted the time course of the vAI in native subject space
using the inverse normalization procedure in ANTs. The correlation between this
vAI seed and all voxels in the brain were computed in native space, after discard-
ing time points flagged from the scrubbing procedure (see “MRI preprocessing”).
The resulting correlation values were Fisher z-transformed, transformed into
our group template space and averaged across subjects. This average correlation
map was thresholded at z > 0.35 to define four regions in the vAI network (bilat-
eral vAI, medial prefrontal cortex and posterior cingulate cortex, as in ref. 38;
Supplementary Fig. 2). To compute connectivity within this network during
encoding blocks, the group level ROIs comprising the vAI network were inverse-
normalized into native subject space, and the average of the pairwise correlations
between these ROIs was computed for each encoding scan.
Low-frequenc y encoding connectivity analyses. To examine low-frequency
(LF) connectivity during encoding blocks, the BOLD data were low-pass
filtered (< 0.06 Hz) below the task frequency of 0.0714 Hz (Fig. 3a). Before
computing connectivity, time points flagged from the scrubbing procedure
(see “MRI preprocessing”) were removed. Since each encoding block was
comprised of three scans, the scrubbed data were first z-scored within each
scan and then the data was concatenated across scans for each encoding block.
Pearson correlation coefficients were then computed between the concatenated
BOLD time-courses in each encoding block and correlation coefficients were
Fisher z-transformed.
First, we compared the global similarity (correlation) of multivoxel LF amy-
gdala connectivity patterns between emotional and neutral encoding blocks as
a function of encoding order (Fig. 3b). Multivoxel LF connectivity (correla-
tion) patterns with the amygdala were computed using the bilateral amygdala
LF filtered time-course. The Fisher z-transformed correlation value (measur-
ing connectivity with the amygdala) was then extracted for all voxels that were
labeled as gray matter (value of 0.3 or greater using the probabilistic GM mask
created in SPM5) and fell within a mask of cortical or subcortical structures (see
“Anatomical masks”). Voxels that fell in the amygdala ROI were excluded from
this analysis, although the exclusion of these voxels did not modify the results.
To measure the similarity of multivoxel LF amygdala connectivity patterns
between emotional and neutral encoding blocks, the Pearson correlation was
computed between the LF amygdala connectivity patterns during emotional and
neutral encoding across these cortical and subcortical voxels. These correlation
values (representing similarity of amygdala connectivity patterns between emo-
tional and neutral encoding blocks) were Fisher z-transformed, and the difference
between multivoxel pattern similarities for E-N versus N-E encoding orders was
assessed using an unpaired t-test (Fig. 3b).
To compare levels of LF connectivity between ROIs as a function of encod-
ing order (Fig. 4 and Supplementary Fig. 2), two-way mixed-effects ANOVAs
were performed on the z-transformed correlation values in each encoding block
using a within-subjects factor of emotion (emotional/neutral) and a between-
subjects factor of encoding order (E-N/N-E). Additional follow-up t-tests were
performed to assess differences based on emotion within encoding order as well
as differences between neutral encoding connectivity as a function of encoding
order. Paired t-tests were used to evaluate differences within an encoding order
as a function of emotion (between emotional and neutral encoding blocks), and
unpaired t-tests were used to evaluate differences across encoding orders between
condition types (i.e., neutral encoding connectivity as a function of encoding
order).
Subsequent memory analyses. GLMs were used to measure BOLD activation
associated with different levels of subsequent memory, specifically activity that
tracked future levels of recollection-based memory. To do so, separate regressors
were included for: (i) subsequent R, K and missed (M) trials (trials labeled ‘new’ at
test), modeled as 2-s boxcars; (ii) stimulus-preceding fixation periods (modeled
as 1-s boxcars); and (iii) response periods (modeled as 1-s boxcars), such that the
arrows task served as an implicit baseline. Each regressor was convolved with the
canonical HRF. GLMs were estimated for emotional and neutral encoding blocks
in native subject specific space (Fig. 5). For group level analyses, the resulting
contrasts (R – K trials) were normalized into our group template space (Fig. 6).
Two subjects were excluded from all subsequent recollection analyses: one
subject in the E-N encoding order had less than two K responses for emotional
stimuli and another subject from the N-E encoding order with the fewest number
of R or K responses whose exclusion served to match the number of subjects
from both encoding orders.

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
nature neurOSCIenCe doi:10.1038/nn.4468
To assess multivoxel patterns of subsequent recollection effects across the brain
and whether subsequent recollection patterns were more similar between emo-
tional and neutral encoding for the E-N versus N-E encoding order, subsequent
recollection effects were computed as the difference in activity estimates (beta
coefficients) for subsequent R – K trials. This R – K effect was then extracted
for all voxels that were labeled as gray matter (value of 0.3 or greater in a proba-
bilistic GM mask) and fell within a mask of cortical or subcortical structures
(see “Anatomical masks”). The Pearson correlation was then computed for each
subject between the emotional and neutral subsequent recollection patterns
across these voxels. Resulting correlation values were Fisher z-transformed and
the difference between the similarities for the E-N versus N-E encoding order
was assessed using an unpaired t-test (Fig. 5).
To ensure that differences in multivoxel subsequent recollection patterns
between emotional and neutral encoding based on encoding order were not
potentially driven by differences in R and K bin sizes across encoding orders
(as an interaction between R versus K responses, emotion and encoding order
was observed), a nonparametric permutation test was used that controls for dif-
ferent bin sizes across comparisons. Simulations were performed (n = 1,000) to
generate a null distribution of the expected effect size of interest (difference in
similarity between emotional and neutral R – K patterns between the E-N and
N-E encoding orders) that would arise by chance based on the properties of the
data while shuffling the conditions of interest (activity related to true subsequent
recollection differences). For each simulation, a new design matrix was created for
each subject by randomly shuffling the subsequent memory labels. The GLM was
then recomputed, the resulting subsequent recollection effects (R – K activity esti-
mates) were recorded and the similarity between the null multivoxel emotional
and neutral subsequent recollection patterns was computed for each simulation
in each subject. The true difference was then compared to the null distribution
of differences to assess statistical significance (Supplementary Fig. 3).
To examine the presence of univariate carry-over on subsequent recollection
effects (Fig. 6), i.e., subsequent recollection effects that were present during
emotional encoding and persisted into subsequent neutral encoding for the E-N
encoding order, we performed a series of conjunctive tests. Based on the R versus
K activity estimates, a carry-over effect on subsequent recollection was operation-
alized as a conjunction of the following effects (depicted in Fig. 6a):
• Emotional encoding R > K for E-N encoding order;
• Emotional encoding R > K for N-E encoding order;
• Neutral encoding R > K for E-N encoding order;
• Neutral encoding R – K for E-N encoding order > neutral encoding R – K
for N-E encoding order;
• Emotional encoding R – K for E-N encoding order > neutral encoding
R – K for N-E encoding order;
• Emotional encoding R – K for N-E encoding order > neutral encoding
R – K for N-E encoding order.
In other words, regions emerging from this contrast had to demonstrate sig-
nificant emotional subsequent recollection effects for both encoding orders, a
significant subsequent recollection effect during neutral encoding for the E-N
encoding order, with the stipulation that all of these subsequent recollection
effects had to be greater than the subsequent recollection effect when this emo-
tional carry-over effect should not be present (during neutral encoding when
it occurred before emotional encoding). All individual tests were thresholded
at P < 0.05 using a one-tailed t-test, as the directionality of all differences was
uniquely specified, and the conjunction of the above contrasts was performed.
Null simulations described above were performed to determine the voxel extent
that resulted in a whole-brain family-wise error (FWE) rate of P < 0.05 for this
conjunction analysis64. These null simulations were performed in normalized
template space rather than native subject space (used for null simulations for
subsequent recollection pattern analyses). For each null simulation, a new design
matrix was created by shuffling the subsequent memory labels of each subject,
GLMs were recomputed for each encoding order, the above contrasts were per-
formed and the cluster sizes resulting from the above conjunction analysis were
recorded. This approach generated a null distribution of cluster sizes associated
with a specific contrast (in this case a conjunction analysis). The cluster size
(73 voxels) was chosen as the smallest cluster for which the probability of observ-
ing this cluster size was less than 0.05 across 1,000 simulations.
We also assessed whether emotion-related activity predictive of subsequent
recollection was localized within the anterior versus posterior hippocampus,
based on prior notions that the amygdala should predominantly influence the
anterior hippocampus during emotional memory formation33,36,37. We then
examined whether such an anterior versus posterior bias was present dur-
ing neutral encoding when neutral stimuli were encountered after extended
blocks of emotional stimuli (E-N encoding order). To do so, we simply meas-
ured the anterior versus posterior position all voxels showing subsequent
recollection-based encoding effects (all voxels showing R > K activity esti-
mates) in each subject on a scale of −1 (fully posterior) to +1 (fully anterior).
This resulted in an anterior versus posterior bias score for all subjects, sepa-
rately for emotional and neutral encoding blocks. This measure was found to
not be normally distributed (using the Lilliefors test) for emotional encoding
in the N-E encoding order. We thus computed both parametric and nonpara-
metric permutation tests when assessing differences in the anterior–poste-
rior hippocampal subsequent recollection memory bias score as a function
of encoding order. Permutation tests were performed in the same manner as
described for memor y performance and SCL (see “Behavioral analyses” and
“Skin conductance level analyses”).
Statistics. To examine the impact of encoding order on SCL, memory accuracy
and neural measures, the data were first analyzed using mixed-effects ANOVAs
including within-subjects factors of emotion and between-subjects factors of
encoding order. Follow-up t-tests were used to examine differences between
emotional and neutral encoding within each encoding order (paired t-tests) and
between encoding orders (unpaired t-tests). All of the t-tests performed were two-
tailed, with the exception of the SCL data (as described in “Skin conductance level
analyses” and noted in the “Skin conductance” subsection of “Results”), as well
as when examining carry-over subsequent recollection effects at the whole-brain
level, since the direction of the subsequent memory effects were uniquely speci-
fied (Fig. 6; described in “Subsequent memory analyses”). Multiple comparison
corrections were performed for the SCL data (comparing relative SCL across
multiple scans versus zero in each encoding order in Figure 2a; described in “Skin
conductance level analyses”) as well as for the whole-brain analysis in Figure 6
(described in “Subsequent memory analyses”). The normality of data distribu-
tions was tested using the Lilliefors test. All data were found to be normally
distributed (did not reach significance to reject the null hypothesis that the data
were derived from a normal distribution) with the exception of mean SCL during
several blocks, overall memory accuracy (R + K hits) for the second block of the
N-N encoding behavioral group and the anterior–posterior bias scores of hip-
pocampal subsequent recollection memory for one of the four encoding blocks
(emotional encoding in N-E encoding order). Thus, tests of statistical significance
involving these data were performed using both parametric and nonparametric
permutation tests (both statistical tests were reported and resulted in the same
interpretation for all data). Note that nonparametric permutation tests were also
used for subsequent memory analyses (see “Subsequent memory analyses”).
Data and code availability. The data that support the findings of this study
are available on reasonable request from the corresponding author (L.D.).
The data are not publicly available because they contain information that could
compromise research participant privacy/consent. The statistical map associ-
ated with the whole brain conjunction analysis in Figure 6 is available (in MNI
space rather than group template space) on neurovault.org (http://neurovault.
org/collections/JOYXPMRX/). Standard software packages (SPM5 and FSL)
were used for processing the MRI data in addition to custom Matlab scripts.
Custom-written code is available upon reasonable request to the corresponding
author (L.D.).
A Supplementary Methods Checklist is available.
51. Stark, C.E.L. & Squire, L.R. When zero is not zero: the problem of ambiguous
baseline conditions in fMRI. Proc. Natl. Acad. Sci. USA 98, 12760–12766
(2001).
52. Lang, P.J., Bradley, M.M. & Cuthbert, B.N. International Affective Picture System
(IAPS): Instruction Manual and Affective Ratings (NIMH Center for the Study of
Emotion and Attention, 1999).
53. Marchewka, A., Zurawski, Ł., Jednoróg, K. & Grabowska, A. The Nencki Affective
Picture System (NAPS): introduction to a novel, standardized, wide-range, high-
quality, realistic picture database. Behav. Res. Methods 46, 596–610 (2014).

© 2016 Nature America, Inc., part of Springer Nature. All rights reserved.
nature neurOSCIenCe
doi:10.1038/nn.4468
54. Greicius, M.D., Krasnow, B., Reiss, A.L. & Menon, V. Functional connectivity in the
resting brain: a network analysis of the default mode hypothesis. Proc. Natl. Acad.
Sci. USA 100, 253–258 (2003).
55. Damoiseaux, J.S. et al. Consistent resting-state networks across healthy subjects.
Proc. Natl. Acad. Sci. USA 103, 13848–13853 (2006).
56. Avants, B.B., Epstein, C.L., Grossman, M. & Gee, J.C. Symmetric diffeomorphic
image registration with cross-correlation: evaluating automated labeling of elderly
and neurodegenerative brain. Med. Image Anal. 12, 26–41 (2008).
57. Behzadi, Y., Restom, K., Liau, J. & Liu, T.T. A component based noise correction
method (CompCor) for BOLD and perfusion based fMRI. Neuroimage 37, 90–101
(2007).
58. Power, J.D., Barnes, K.A., Snyder, A.Z., Schlaggar, B.L. & Petersen, S.E. Steps
toward optimizing motion artifact removal in functional connectivity MRI; a reply
to Carp. Neuroimage 76, 439–441 (2013).
59. Power, J.D. et al. Methods to detect, characterize, and remove motion artifact in
resting state fMRI. Neuroimage 84, 320–341 (2014).
60. Van Dijk, K.R.A., Sabuncu, M.R. & Buckner, R.L. The influence of head motion on
intrinsic functional connectivity MRI. Neuroimage 59, 431–438 (2012).
61. Power, J.D., Barnes, K.A., Snyder, A.Z., Schlaggar, B.L. & Petersen, S.E. Spurious
but systematic correlations in functional connectivity MRI networks arise from
subject motion. Neuroimage 59, 2142–2154 (2012).
62. Pruessner, J.C. et al. Volumetry of hippocampus and amygdala with high-resolution
MRI and three-dimensional analysis software: minimizing the discrepancies between
laboratories. Cereb. Cortex 10, 433–442 (2000).
63. Fischl, B. et al. Whole brain segmentation: automated labeling of neuroanatomical
structures in the human brain. Neuron 33, 341–355 (2002).
64. Maris, E. & Oostenveld, R. Nonparametric statistical testing of EEG- and MEG-data.
J. Neurosci. Methods 164, 177–190 (2007).