Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Journal Pre-proof
Exploring brain-behavior relationships in the N-back task
Bidhan Lamichhane, Andrew Westbrook, Michael W. Cole, Todd Braver
PII: S1053-8119(20)30170-1
DOI: https://doi.org/10.1016/j.neuroimage.2020.116683
Reference: YNIMG 116683
To appear in: NeuroImage
Received Date: 4 April 2019
Revised Date: 17 February 2020
Accepted Date: 24 February 2020
Please cite this article as: Lamichhane, B., Westbrook, A., Cole, M.W., Braver, T., Exploring
brain-behavior relationships in the N-back task, NeuroImage (2020), doi: https://doi.org/10.1016/
j.neuroimage.2020.116683.
This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.
© 2020 Published by Elsevier Inc.
Exploring brain-behavior relationships in the N-back task
Bidhan Lamichhane
1*
, Andrew Westbrook
2,3
, Michael W. Cole
4
& Todd Braver
1
1
Department of Psychological and Brain Sciences, Washington University in Saint Louis, 1
Brookings Drive, Saint Louis, MO, 63130, USA.
2
Donders Institute for Brain, Cognition and Behaviour, Radboud University, Kapittelweg 29,
6525 EN Nijmegen, The Netherlands.
3
Department of Cognitive, Linguistics, and Psychological Sciences, Brown University, 190
Thayer Street, Providence, RI, 02912, USA.
4
Center for Molecular and Behavioral Neuroscience, Rutgers University, Newark, New Jersey,
USA
* Corresponding author: Bidhan Lamichhane, bidhanlamichhane@wustl.edu
Credit Author Statement:
Bidhan Lamichhane: Data collection, data analysis, writing- original draft, writing-
reviewing.
Andrew Westbrook: Data collection, data analysis, writing- reviewing.
Michael W. Cole: writing- reviewing.
Todd Braver: Supervision, conceptualization, data analysis, draft preparation, writing-
reviewing.
1
Exploring brain-behavior relationships in the N-back task
Bidhan Lamichhane
1*
, Andrew Westbrook
2,3
, Michael W. Cole
4
& Todd Braver
1
1
Department of Psychological and Brain Sciences, Washington University in Saint Louis, 1
Brookings Drive, Saint Louis, MO, 63130, USA.
2
Donders Institute for Brain, Cognition and Behaviour, Radboud University, Kapittelweg 29,
6525 EN Nijmegen, The Netherlands.
3
Department of Cognitive, Linguistics, and Psychological Sciences, Brown University, 190
Thayer Street, Providence, RI, 02912, USA.
4
Center for Molecular and Behavioral Neuroscience, Rutgers University, Newark, New Jersey,
USA
* Corresponding author: Bidhan Lamichhane, bidhanlamichhane@wustl.edu
Keywords:
Working memory; N-back; frontal-parietal network; default mode network; salience network;
dorsolateral prefrontal cortex.
2
Abstract
Working memory (WM) function has traditionally been investigated in terms of
two dimensions: within-individual effects of WM load, and between-individual
differences in task performance. In human neuroimaging studies, the N-back task has
frequently been used to study both. A reliable finding is that activation in
frontoparietal
regions
exhibits an inverted-U pattern, such that activity tends to decrease at high load
levels. Yet it is not known whether such U-shaped patterns are a key individual
differences factor that can predict load-related changes in task performance. The current
study investigated this question by manipulating load levels across a much wider range
than explored previously (N = 1-6), and providing a more comprehensive examination of
brain-behavior relationships. In a sample of healthy young adults (n = 57), the analysis
focused on a distinct region of left lateral prefrontal cortex (LPFC) identified in prior
work to show a unique relationship with task performance and WM function. In this
region it was the linear slope of load-related activity, rather than the U-shaped pattern
that was positively associated with individual differences in target accuracy.
Comprehensive supplemental analyses revealed the brain-wide selectivity of this pattern.
Target accuracy was also independently predicted by the global resting-state connectivity
of this LPFC region. These effects were robust, as demonstrated by cross-validation
analyses and out-of-sample prediction, and also critically, were primarily driven by the
high-load conditions. Together, the results highlight the utility of high-load conditions for
investigating individual differences in WM function.
3
1. Introduction
Understanding the neural basis of working memory and executive control
(WM/EC) functions has been a major aim of cognitive neuroscience research. One of the
key drivers of such research efforts are the well-established findings that WM/EC
function is strongly dominated by individual differences, and moreover, that these
individual differences clearly contribute to real-world cognitive abilities (i.e.,
intelligence) and important life outcomes (e.g., computer programming skills, ability to
learn complex new task, SAT/GRE success, etc; (Ackerman, Beier, & Boyle, 2005;
Engle, Laughlin, Tuholski, & Conway, 1999; Kyllonen & Christal, 1990).
The N-back task has been one of the most commonly used experimental
paradigms for exploring the neural basis of WM/EC (Cohen et al., 1997; Gevins &
Cutillo, 1993). The N-back is well-established to robustly activate frontoparietal brain
regions that which general consensus hold to be critical for WM/EC function (Dosenbach
et al., 2006; Owen, McMillan, Laird, & Bullmore, 2005). An advantageous feature of the
N-back is that WM load can be varied in an incremental, parametric fashion by
increasing the value of N (Braver et al., 1997). This is a critical component of the
paradigm, since as N-back levels increase, task performance shows a reliable decrement,
while the subjective experience of cognitive effort and task difficulty also increases
(Ewing & Fairclough, 2010; Otto, Zijlstra, & Goebel, 2014; Westbrook, Kester, &
Braver, 2013; Westbrook, Lamichhane, & Braver, 2019) The concomitant increase in
task difficulty and drop in performance is useful psychometrically as it drives variability
4
among participants, and thus the potential to detect the neural correlates of individual
differences.
The consequences of N-back load manipulations on brain activity and the
relationship between activity and behavior remain unclear, however. For example, there
is uncertainty about which load levels are optimal for detecting brain activity patterns and
brain-behavior relationships. This uncertainty is, in part, because of non-linear inverted
U-shaped load functions, in which BOLD activity increases within frontoparietal brain
regions as N increases across lower load- levels (e.g., 0, 1,2) but then starts to decrease as
load increases to higher levels (N ≥ 3). A common observation is that inverted-U patterns
emerge when participants reach their capacity limits (Van Snellenberg et al., 2015). For
example, inverted-U functions are observed under very high cognitive demands, with
decreasing working memory capacity, with advancing cognitive age, and with cognitive
impairments (Callicott et al., 1999; Cappell, Gmeindl, & Reuter-Lorenz, 2010; Jaeggi et
al., 2007; Nyberg, Dahlin, Stigsdotter Neely, & Backman, 2009). One interpretation of
inverted-U patterns is that individuals “disengage” at these high-load levels (i.e.,
discontinue applying full cognitive effort), since the task may be too difficult to perform
adequately when load is at supra-capacity levels (Callicott et al., 1999; Jaeggi et al.,
2007; Van Snellenberg et al., 2015). Alternatively, inverted-U patterns may reflect
shifting strategies (e.g., a shift towards responding based on familiarity at high load
levels, rather than utilizing active maintenance (Juvina & Taatgen, 2007). A systematic
analysis of inverted-U patterns and their relationship with task performance across
individuals and across a wide range of load levels is needed for testing these hypotheses.
At the same time, uncertainty about the meaning of inverted-U patterns makes it unclear
5
whether very high levels of N-back load are even suitable for investigating brain-
behavior relationships. Consequently, very high levels of N-back loads (N > 3) have not
been investigated.
It also remains unclear which brain regions or neural markers are the best
predictors of individual differences in performance. For example, it is unclear whether
the brain-behavior relationships are more focal, i.e., related to activity patterns in
circumscribed brain regions, or more widespread and non-selective, and also captured
well by other neural measures, such as functional connectivity. The lateral prefrontal
cortex (LPFC) has traditionally been implicated as most likely to mediate performance in
WM/EC tasks, as this region is thought to play a critical role in active goal maintenance,
and cognitive control functions (Kane & Engle, 2002; Miller & Cohen, 2001).
Nevertheless, evidence pointing to a specific and unique role for focal LPFC regions in
mediating brain-behavior relationships in the N-back is limited. Although some studies
have found evidence implicating the LPFC, in many others it is just one relevant region
out of many that can predict individual variation in behavioral performance in the N-back
and related WM/EC tasks (Choo, Lee, Venkatraman, Sheu, & Chee, 2005; Harvey et al.,
2005).
Furthermore, other work has suggested that qualitatively distinct neural markers,
such as resting-state functional connectivity, may be equally or even more strongly
predictive of N-back task performance (Cole, Yarkoni, Repovs, Anticevic, & Braver,
2012). Specifically, Cole et al (2012) used a functional connectivity measure termed
global brain connectivity (GBC), to demonstrate strong relationships between
connectivity and multiple aspects of cognitive function, not only N-back performance but
6
also WM capacity and fluid intelligence. Interestingly, this work also provided a unique
perspective, in that the findings highlighted the contributions of a particular left PFC
region (center coordinates: -44, 14, 29) that was unique among brain regions in showing
robust brain-behavior relationships in both activation and GBC effects, across multiple
indices. Consequently, this work suggests that there may in fact be focal brain-behavior
effects found within distinct PFC regions, which may co-exist with (or even be stronger
than) the more widespread network effects that have been reported. However, until now
there have been no analyses directly comparing the predictive power of activity versus
connectivity effects. It also suggests the potential utility of the GBC metric for revealing
the functional importance of connectivity profiles within individual regions in a manner
that can be compared with activation profiles. However, a limitation of Cole et al (2012)
and other prior work was that such comparisons were not a direct focus of analysis, and
moreover, a restricted range of N-back load levels were examined.
In the current study, we sought to remedy this gap in the literature, by capitalizing
on a novel experimental design in which a large sample of individuals (n = 57) each
performed the N-back task under a very wide range of load levels, from N = 1-6. This
design afforded a unique opportunity to test whether lower or higher load levels were
more predictive of individual differences in task performance. In addition, we directly
assessed the predictive power of a focal, a priori LPFC
region of interest highlighted in
both meta-analyses and our own prior work (Cole et al., 2012; Rottschy et al., 2012), to
capture individual differences in behavioral performance in terms of both its activity and
functional connectivity (i.e., GBC), and further to rigorously characterize the form of this
predictive relationship. Finally, we used newer cross-validation methods and a rigorous
7
out-of-sample test (involving data from the Human Connectome Project; HCP) to
establish predictive validity.
To preview our findings, we observed that this focal region of LPFC was a
reliable neural predictor of behavioral performance, and moreover that the predictive
utility was strongest when aggregating across all load levels, rather than just in selecting
lower levels. Moreover, in a direct comparison, high load levels were more predictive of
behavioral performance than low load levels. Nevertheless, additional independent
variance was explained by activation global resting state functional connectivity, such
that the strongest predictions of individual differences were found when both measures
(LPFC activity and GBC in LPFC) were aggregated in the model. Lastly, we observed
that this a priori defined LPFC region was one of the best predictors of behavioral
performance in a large out-of-sample dataset (HCP). Together, the results suggest the
importance of including a wide range of variability in N-back paradigms to target
individual differences, and highlight the unique contributions of a focal LPFC region for
predicting WM performance.
2. Materials & Methods
2.1. Participants
Fifty-eight participants were recruited from the Washington University in St.
Louis community. One was excluded (participant disclosed neurological problem after
data was collected) yielding a final sample of 57 participants (27 male and 30 female;
mean age = 24.28 ± 5.1 years) in N-back task. However, an additional 6 subjects out of
57 had technical issues with their resting-state fMRI data. Consequently, the remaining
8
51 participants were used for analyses of resting-state functional connectivity. All
included participants were healthy, right-handed, neurologically normal, not currently
taking any psychoactive medication, native English speakers, and with normal-to-
corrected vision including no color-blindness.
2.2. Experimental task and data collection procedure
All MRI data were collected in a 3 Tesla Siemens Trio scanner. After MR safety
screening and consent, participants were scanned in six BOLD fMRI runs while they
performed each of one level of the N-back task (N = 1:6; see Figure 1). To facilitate
individual difference analyses, all participants performed N-back conditions in the same
order of increasing N-back load levels (i.e., 1-back in first scan, 6-back in last scan). The
experimental session also included two additional resting-state scans and collection of a
T1-weighted anatomical scan (these are described further below).
All N-back fMRI BOLD scans consisted of 3 task blocks, each approximately 2
minutes in duration, that were preceded and followed by a 30 second rest fixation period
(marked by a central crosshair; 4 total per scan; total scan duration = 520 seconds). Task
blocks consisted of 64 stimuli (lower-case consonants), presented in the center of a screen
in large (32 point) Arial font. Each stimulus was presented for a maximum of 2 seconds
during which participants were instructed respond by button press as quickly as possible,
without sacrificing accuracy, to indicate whether the stimulus was a target (N-back
repeat) or not (using middle/index fingers). Upon button press, letters were replaced by a
fixation underscore ‘_’, until the next letter appeared, 2 seconds after the previous letter
was presented. There were 16 target items in each task block, and a variable number of
lures, depending on the task level (8 for the 1-back, 6 for the 2-back, 5 for the 3-back, and
9
3 for the 4-, 5-, and 6-back, each) where a lure is considered to be any stimulus repeated
within two positions of the target position (e.g., a 1, 2, 4 or 5-back repeat would be
considered a lure in the 3-back block). The number of lures decreased for higher load
levels to “flatten” performance functions – attenuating differences in performance from
lower to higher load levels.
At the end of each block, participants received brief feedback on their
performance in terms of percentage accuracy on target and non-target items. This was
presented for 5 seconds, followed by 25 seconds of resting fixation prior to the next block
(total 30s, in between blocks rest). The total duration of each scan was 520 seconds (260
scans). Each N-back level was presented in a unique color (black, red, blue, purple,
green, brown for the 1-back, 2-back, 3-back, 4-back, 5-back, 6-back respectively as
shown in Figure 1); task instructions referred to the condition by color (rather than
numerical load descriptor), to minimize potential demand characteristics regarding
difficulty. This last feature of the procedure was not relevant for the current study, and
was put in place solely for purposes of a subsequent study phase (in a separate
experimental session) that was not the focus of the current work.
Two resting-state scans were also conducted, one prior to beginning the N-back
task scans and after completing these scans; each was 530 seconds long. All fMRI BOLD
scans were acquired using the following parameters: 2000 msec TR, 27 msec TE (spin-
echo time), 90 degree flip angle, 4×4×4 mm voxels with a 256×256 field of view with 34
slices. Anatomical T1-weighted images were also collected with the following
parameters: 2400 msec TR, and 3080 msec TE (spin-echo time), 8 degree flip angle,
1×1×1 mm voxels, and 176 slices.
10
Figure 1: Schematic of single blocks of tasks for the 1-, 2-, and 6-back (“black”, “red”,
and “brown” task). Each color corresponded to a single load level and indicated the N-
back rules for a given run of stimuli.
2.3. Behavioral analysis
Behavioral performance was analyzed separately for target and non-target trials,
and by examining both accuracy and reaction time measures. Some studies of the N-back
have analyzed behavioral performance in terms of the signal detection measure d', since
this provides a relative measure of sensitivity to the target/non-target status of items
while controlling for response bias (Wickens, 2002). In contrast, for the current analysis
this measure may be less appropriate particularly for comparing load levels, since the
load levels varied in the proportion of non-target trials that were lures (i.e., lower lure
frequency at higher load levels). Thus, to be more conservative, all primary results are
described in terms of target or non-target accuracy. Nevertheless, we did conduct
supplementary analyses with d’ (reported in Supplemental Results), with most effects
largely unchanged (and in fact, some effects, with LPFC were even stronger).
11
2.4. Imaging analyses
All neuroimaging analysis was conducted using AFNI (https://afni.nimh.nih.gov)
software, with the following processing steps.
2.4.1. Task fMRI preprocessing
After the converting raw DICOM images to NIFTI format, data were temporally
aligned within each brain volume, and corrected for movement, yielding 6 estimated
motion parameters (three translation: x, y, z and three rotation: pitch, yaw, roll). As an
additional quality control step, data were also censored (scrubbed) for motion transients
using a frame-wise displacement threshold of 0.3 mm. Functional images were then
registered to the Montreal Neurological Institute (MNI) atlas space, which also involved
up-sampling from 4×4×4 mm to 3×3×3 mm voxels. Precise registration was verified
visually for every participant and cost functions were tailored to optimize registration for
each participant. Image intensities were scaled to have a mean value of 100, and a range
of 0-200. Finally, images were spatially smoothed with a Gaussian full-width half
maximum (FWHM) = 8 mm filter.
General linear models (GLMs) were fit using the 3dDeconvolve function in
AFNI, to analyze the relationship between task conditions on voxel-wise BOLD
activation levels. All GLMs incorporated the 6 estimated motion parameters and
polynomial functions (-polort 4) to capture low-frequency signal drifts as nuisance
covariates. N-back task activations were modeled by a block design of boxcar functions
spanning each 128-second stimulus run, convolved with a gamma hemodynamic
response function.
2.4.2. Resting state fMRI preprocessing
12
Resting-state fMRI data were pre-processed using AFNI’s standard resting state
pre-processing procedure (also see Jo et. al., 2010). Specifically, in addition to standard
steps taken for task BOLD data (i.e., spike-correction, temporal alignment, motion
correction, registration to MNI atlas space), more conservative censoring (scrubbing) of
motion transients was also performed (given that motion transients are known to have a
large impact of functional connectivity estimates, (Power et al., 2011) using a frame
displacement threshold of 0.2 mm. Likewise, the data were band-pass filtered (0.01-0.1
Hz), and nuisance signal from locally-averaged white matter (ANATICOR procedure
available in AFNI (Jo, Saad, Simmons, Milbury, & Cox, 2010)) and the 6 estimated
motion parameters were regressed out of the time-series prior to connectivity analyses.
2.4.3. Region of interest (ROI) analyses
A region-of-interest (ROI) approach was used primarily for task activation and
connectivity analyses. The key ROI was a particular left prefrontal cortex (LPFC) region
that was selected based on prior findings demonstrating the strong involvement of this
region in WM function and individual differences in brain-behavior relationships (Cole,
Ito, & Braver, 2015; Cole et al., 2012; Rottschy et al., 2012). To define the ROI, we
created a spherical region (6mm radius) with center coordinates based on Cole et al, 2012
(MNI coordinates [-44, 14, 29]). Note that this is region overlaps with that identified by
Rottschy et al (2012) as part of the WM core network that they refer to as left inferior
frontal gyrus pars opercularis (-46, 10, 26); however, the extent of the sphere that we
used extends from the inferior frontal gyrus to the middle frontal gyrus. Nevertheless, to
reduce ambiguity, from here on we use the term LPFC to refer to this particular ROI.
13
To further assess the predictive utility of this ROI, supplemental analyses
compared its activation to a comprehensive, brain-wide set of focal ROIs, as well as
larger-scale brain networks. For these, we used the set of 264 (spherical, 6mm radius)
nodes centered on loci defined in Power et al (2011), as these were drawn from both task
fMRI meta-analyses and large-sample functional connectivity datasets, and have already
been pre-structured into brain network communities. After fitting GLMs, regression
weights were extracted and averaged across voxels for the LPFC ROI (and for
supplementary analysis for each of the nodes and networks defined in Power et al, 2011).
Between-subjects analyses of load-level and individual difference effects were conducted
using these averaged regressions weights. For GBC analyses, resting-state timeseries data
were averaged within the LPFC node and also for each node in the Power et al (2011) set,
such that pairwise correlations between all nodes could be calculated for each participant.
These correlation values were then included in similar brain-behavior analyses testing for
individual difference effects.
3. Results
3.1. N-back performance and load effects
As predicted, overall N-back accuracy (Table1 and Figure2), and target accuracy
decreased with increasing load (formally, repeated-measures ANOVA shows reliable
effects of task load (F
5,280
= 218.14, p < 0.001, Figure 2A). Similarly, as predicted, overall
non-target accuracy also decreased with increasing load (F
5, 280
= 24.93, p < 0.001, Figure
2B). Nevertheless, mean performance remained relatively high even at the highest load
conditions. Also, response times (RT) for both targets (F
5,280
= 42.75, p < 0.005, Figure
14
2C) and non-targets (F
5,280
= 32.53, p < 0.001, Figure 2D) showed the significant effect of
load and differ reliably (RTs slower for targets than non-targets, Table 1) at every load
levels. These accuracy and RT patterns imply that participants did not resort to guessing
or random responding, even at the highest load levels.
We next used linear mixed-effects models (equation 1) to test for linear effects of
load on accuracy and RT. Specifically, we tested whether accuracy and reaction time, as
independent variables (
ℎ
) could be predicted by load
i
(
)
, with loads nested
within participants
j
. Note that
and
are random intercepts, allowing the intercept
and slope to vary by subject, respectively.
ℎ
=
+
+
Eq (1)
=
+
=
+
For target and non-target accuracy, there were significant, negative linear effects
of load (target accuracy: slope = -10.25, t(56) = -23.2, p < 0.001; non-target accuracy:
slope = -1.62, t(56) = -7.3, p < 0.001). For target RT, but not for non-target RT, positive
linear effects were observed (target RT: slope = 0.03, t(56) = 7.34, p < 0.001; non-target
RT: slope = 0.004, t(56) = 1.63, p > 0.1). We further tested an extension of the mixed-
effects model that included a quadratic term
1
for target accuracy (slope = 1.18, t(56) =
5.46, p < 0.001) and both RT measures (target: slope = -0.017, t(56) = -8.8, p < 0.001;
non-target: slope = -0.018, t(56) = -9.4, p < .001). These quadratic patterns indicate that
1
In this case the model was extended to include an additional term:
Beha
=
0
+
1
+
2
2
+
, which was also allowed to vary by subject
=
+
15
target accuracy did not decline linearly at high load levels but rather declined
asymptotically, while RT slowed with load, but then sped up again when N-back load
was very high.
N-back task Target (mean) ± standard deviation Non-target (mean) ± standard deviation
Accuracy (%) Response time
(ms) Accuracy (%) Response time (ms)
1-back 87.7
±9.0 592
±82 95.7
±6.6 558
±88
2-back 77.3
±12.4 710
±11 92.4
±8.5 678
±128
3-back 58.1
±15.3 787
±123 90.5
±7.2 713
±145
4-back 52.1
±11.9 771
±139 90.2
±7.8 681
±132
5-back 43.1
±15.3 758
±142 88.9
±7.9 654
±119
6-back 37.7
±15.0 763
±147 86.4
±9.0 618
±121
Table 1: Behavioral results: performance (%) and response time (ms) (± standard deviation).
Figure 2. Bar plot of N-back performance. (A) Target: mean performance (accuracy, %)
by load level and (C) response time (RT). (B) Non-target: mean performance (accuracy,
%) by load level and (D) response time (RT). Error bars indicate standard error of the
mean.
16
3.2. Overview of N-back fMRI results
In order to investigate neural substrates of working memory, we used an a priori
approach, focusing on a particular prefrontal region of interest and, as a supplemental
analysis, on a comprehensive set of additional brain-wide nodes and networks. The focal
left LPFC region of interest was selected because it has been repeatedly shown to exhibit
strong connections with WM and behavioral performance in both meta-analyses and
specific studies, including those conducted in our lab (Cole et al., 2015; Cole et al., 2012;
Rottschy et al., 2012; Wager, Spicer, Insler, & Smith, 2014). In particular, not only was
this region included as part of the “core WM network” in Rottschy et al (2012), but also,
in Cole et al (2012, 2015) this region was unique in exhibiting robust brain-behavior
relationships in terms of both activity and connectivity patterns. Consequently, for this
LPFC ROI, we focused on both its activity, as well as its resting-state functional
connectivity.
After characterizing the load function observed in this ROI, we conducted a
comprehensive set of analyses characterizing brain-behavior relationships with it
(additional supplementary analyses compared this ROI to others across the brain as well
as to brain networks- reported in Supplemental Results). A road-map to these analyses is
as follows. First, we explored a variety of measurement and statistical approaches to
modeling the load function and reducing the dimensionality of activity and behavioral
performance variables: factor analysis, linear slope estimation, and linear mixed effects
modeling. Second, after characterizing brain-behavior relationships in terms of neural
activation, we tested whether similar relationships are present in regards to the functional
connectivity of the LPFC ROI, using the GBC metric. Additionally, we tested whether
17
GBC and neural activity measures each served as unique predictors of behavioral
performance, using a multiple regression approach. Moreover, to ensure the predictive
validity of this approach, these analyses were confirmed using cross-validation. Third, to
establish the generality of LPFC predictive power, we examined this in terms of out-of-
sample prediction, using both the N-back task and an additional out-of-scanner measure
of WM function from the HCP dataset. Lastly, to further understand the source of LPFC
brain-behavior relationships, we separately examined the relative predictive power of
high versus low load conditions.
3.3. WM Load function in LPFC
Prior studies have not investigated patterns of recruitment of these networks
beyond 3-back, and so the current study provides novel information on the role of these
regions at extremely high load levels. In particular, when plotting load-related activity
across participants in LPFC we observed a monotonic increase in activity across lower
level loads, when then shifted to a decreasing pattern beyond N = 3, thus exhibiting a
clear inverted-U profile (Figure 3A; although see below for evidence of a different profile
when subdividing participants according to performance). This visual pattern was
quantitatively confirmed by linear mixed effects modeling (as in equation 1 replacing
behavior (
ℎ
) by brain activity-BOLD
ij.
A non-significant linear term (slope =
0.002, t(56) = 1.29, p = 0.26) and a significant quadratic term (slope = -0.005 , t(56) = -
5.33, p < 0.001 was obtained. These findings confirm that in LPFC the effects of N-back
load appears to follow the inverted-U pattern, with a decrease in activity at high load
levels (see Supplemental Results for parallel findings at the brain network level).
18
Figure 3. Bar plot of N-back activity (beta parameter), mean over all participants by
load level, in (A) LPFC, (B) Anatomical location the left LPFC region from Cole et al.
(2012). Error bars indicate standard error of the mean.
3.4. Statistical modeling of activity-based brain-behavior relationships
Our primary focus in this study was to identify the relationship between load-
related activity and behavior to elucidate the source of individual differences in WM.
Rather than testing for multiple correlations across multiple load levels and BOLD
response profiles, we initially explored a measurement model perspective, employing
factor analysis to reduce dimensionality and test for correspondence between single
factors of performance and BOLD signal. To do so, we first validated that the BOLD and
accuracy measures could each be adequately captured by single factors (see
Supplemental Results).
Next, we tested for correlations between participants’ behavioral accuracy and
BOLD activity factor scores in an analysis, which is formally equivalent to a structural
equation modeling or latent variable approach. In the LPFC, we found a significant
positive correlation, such that higher BOLD activity factor scores were associated with
higher N-back task accuracy (r = 0.28, p = 0.034; Figure 4A). We also conducted parallel
19
analyses using the signal detection index d’ rather than target accuracy, and found similar
results (see Supplemental Results).
This hypothesis-driven confirmatory ROI analysis, was supplemented by an
additional exploratory follow-up analysis that also examined the remaining brain
networks and all individual Power nodes (see Supplemental Results for details).
Interestingly, even when using a liberal statistical significance threshold (i.e., uncorrected
alpha level of 0.05), no other nodes or networks exhibited a positive correlation with
behavior, with the sole exception of the node that was located the closest anatomically to
our LPFC ROI.
Figure 4. Scatter plots. (A) Accuracy factor score and LPFC activity factor score (beta-
parameter). (B) Target accuracy factor score and LPFC load slope (linear fit of LPFC
activity on N-back load).
It is noteworthy that this analysis, which collapses load-related activity to a single
value, was associated with the target accuracy measure. However, it raises the question
whether the effects are specific to accuracy. Conversely, another parallel analysis which
substituted target RT as the behavioral index found no significant correlations with LPFC
(r = -0.13, p = 0.31). To more stringently confirm this specificity, we compared the
20
strength of the correlations between target accuracy and target RT, using Meng’s
correction for non-independent correlation coefficients (Meng, Rosenthal, & Rubin,
1992). Indeed, for this LPFC ROI, relationship with target accuracy was significantly
stronger than the relationship with target RT (LPFC: Z=2.57, p < 0.01). This finding is
particularly noteworthy, given that a supplementary analysis found that within-subject
(i.e. load-related) BOLD patterns related better to RT than accuracy measures (see
Supplemental Results). Thus, the results suggest that between-subjects co-variation in
brain activity and behavior are dissociable from within-subjects relationships (i.e.,
focused on load-related brain-behavior covariation).
The previous analyses above modeled load-related activity in terms of a factor-
score, which is conceptually similar to a cross-load weighted average. The second set of
analyses examined LPFC activity in terms of the linear-trend present in each individual’s
load-related data. Specifically, we estimated the linear slope of load-related LPFC
activity for each participant and tested whether this slope was correlated with their
behavioral accuracy factor score. For the LPFC, this correlation was significantly
positive, indicating that the participants showing a more positive linear load function also
had higher N-back task accuracy (r = 0.36, p = 0.006, Figure 4B).
To illustrate this effect more clearly, we subdivided participants into those that
had a positive linear slope in load-related activity (slope > 0; i.e., consistently increasing
with load) and those that had a linear slope that was not positive (slope
≤
0; i.e., flat or
decreasing activity with load). We then examined behavioral accuracy separately in these
two participant subgroups (Figure 5A). As can be seen, participants with a positive linear
slope (n = 38) showed consistently numerically higher accuracy than those without a
21
positive slope (n = 19), and the difference was significant at multiple individual load
levels (i.e., 1-back: p = 0.02; 2-back: p = 0.24; 3-back: p = 0.07; 4-back: p = 0.04; 5-
back: p = 0.04; 6-back: p = 0.66).
Figure 5. (A) Participants are divided into subgroups based on whether they exhibited
load-related increases in LPFC BOLD activity (positive linear slope; blue), or
flat/decreasing activity (non-positive linear slope; red). Bar plots display mean target
accuracy of the two participant subgroups for 1:6 back tasks, and demonstrate generally
better performance across all load levels in the (blue) subgroup showing load-related
increases. Error bars indicate standard error of the mean. (B) Plot of fitted bold activity
of N-back levels. Gray thin lines represent the linear slope of BOLD load effects in all
individual participants. Mean slope of the high accuracy participants (top third, n = 19;
yellow) and low accuracy participants (bottom third, n = 19; red)
Since the accuracy factor score includes components of both mean performance
(i.e., load independent) and changes in performance due to increasing load, we followed
up this analysis by testing for a more specific relationship between the LPFC linear load
slope in BOLD signal and the linear load slope in target accuracy. To do this most
robustly, we used linear mixed effects models, via a cross-level interaction. Here,
22
individual differences in behavioral performance were treated as an interaction term that
modulated the linear slope of load-related activity. In other words, the model predicted
BOLD activity in terms of linear effects of load, but with the slope of load-related
activation change modulated by the linear slope coefficient of task accuracy (target
trials). Consequently, similar to the previous analysis, this model tests whether more
accurate participants (who will exhibit a less negative / more positive linear slope in
target accuracy) also tend to show a more positively sloped linear load effect in BOLD
activation than less accurate participants. Additionally, the use of a linear mixed effects
model is more rigorous as it directly tests the cross-level interaction (individual
differences effect), while simultaneously controlling for random variation in both the
slope and intercept of load-related activity. Specifically, we tested whether BOLD signals
(
) could be predicted by linear, mean-centered load
(
)
i
, with load-levels
nested within participants
j
. Note that the linear load slope for accuracy (
)
estimated separately for each subject is a predictor of the load effect on BOLD signal at
the subject level of the model. Also,
gives the cross-level interaction of load and the
accuracy slope.
=
+
+
=
+
=
+
+
Indeed, the model reinforced the findings of the prior analysis, there was a
significant interaction between LPFC BOLD load slope and individual differences in
behavioral performance (indexed via the linear slope coefficient of target accuracy; t(56)
= 2.08, p =0.041). To demonstrate these effects, we visualize the predicted performance
23
across load levels and across participants, which also illustrates the complementary
nature of this analysis to the previous one (Figure 5B). Specifically, the linear slope
patterns clearly indicate the variability present across participants, and also that high
accuracy participants tended to have a positive linear slope (increasing BOLD signal with
increasing load), while the lower accuracy participants tended to exhibit a flat (or
decreasing) effect of load on BOLD signal.
Finally, it is worth noting that we also tested whether individual differences in the
strength of the U-shaped load related pattern in BOLD activity were related to behavioral
performance, using the quadratic term coefficient as the individual difference measure.
However, in none of these analyses was there a significant brain-behavior correlation
observed in the LPFC (and indeed no correlation was observed in any brain network
when tested in additional control analyses; see Supplemental Results).
3.5. Brain-behavior relationships with LPFC functional connectivity
In addition to the relationship between load-related activity and behavior, we were
also interested in investigating whether functional connectivity (FC), observed during the
resting state, was uniquely predictive of N-back task performance. We focused on a FC
measure which can be computed for specific, focal brain regions (as well as networks),
and which has been related to N-back performance in prior work: the global brain
connectivity (GBC) index. Specifically, in a prior study from our group (Cole et al.,
2012), we found that the GBC of this LPFC ROI was predictive of N-back task
performance, along with other relevant cognitive individual differences (working
memory capacity, fluid intelligence). Consequently, we tested whether this pattern would
24
replicate in a new dataset, and with a parametric manipulation of load.
We computed the GBC value in two steps and then correlated the value with our
accuracy factor score. First, the resting-state functional connectivity (rsFC) value
between LPFC and every other brain region (defined using the 264 Power node
parcellation) was computed and then Fisher r-to-z transformed. Next, these values were
averaged to create a single GBC score for each participant. As predicted, we found a
reliable correlation between this GBC score and the behavioral accuracy factor score and
this GBC value across participants (r(49) = 0.29, p = 0.039).
Another analysis compared the GBC of the LPFC relative to other brain regions
to determine whether this relative-GBC value was also predictive of behavioral
performance. To determine relative GBC, we first computed the GBC separately for not
only the LPFC, but also for each of the other 264 Power nodes in turn. Then, for each
participant we rank ordered these GBC values, to obtain the rank for LPFC relative to
other brain regions, which we call the relative-GBC value. We then correlated each
participant’s relative-GBC value against their behavioral performance, again using the
factor score measure. This correlation was also significant (r(49) = 0.30, p = 0.029).
This finding suggests that the higher the LPFC’s GBC relative to other brain regions
(irrespective of its overall value), the better was task performance.
Our results up to this point implicate distinct LPFC predictors of individual
differences in N-back target accuracy: activity (either through the factor score or linear
slope approach) and resting-state function connectivity (GBC). However, from the above
analyses, it is not clear whether these predictors explain overlapping or independent
variance in behavioral performance. If it is the latter, then the amount of variability in
25
behavioral performance that can be explained from these distinct neuroimaging measures
should increase when both are simultaneously included as predictors. In addition to
testing whether these regions explain overlapping of complementary variance, we also
conducted cross-validation analyses to more rigorously test the combined predictive
capabilities of these brain indices for WM behavioral performance.
To address these questions, we reformulated the analysis into a multiple
regression, using the accuracy factor score as the outcome variable, and LPFC activity
(linear slope coefficient) and functional connectivity (GBC) as simultaneous predictors.
The multiple regression indicated that, together, the predictors explained significant and
distinct components of behavioral variance, accounting for over 22% of individual
variation in N-back performance (i.e., overall model R
2
= 0.223) (LPFC activity slope:
beta = 62.4, t(47) = 2.86, p = 0.006; LPFC GBC: beta = 3.08, t(47) = 2.15, p = 0.036).
When compared individually, activity measure was somewhat stronger (R
2
= 0.148) than
the connectivity measure (R
2
= 0.09), yet this finding indicates that each of the two
measures accounted for a substantial portion of individual differences in WM
performance.
Although this type of multiple regression analysis is informative, current literature
has pointed to the limitations of standard regression approaches in demonstrating
predictive validity due to over fitting with respect to a given dataset (Yarkoni & Westfall,
2017). Consequently, we next adopted a cross-validation approach popularized in the
machine learning literature – the leave-one-subject-out method – to provide further
validation of these results. Specifically, we tested the correlation between the predicted
and actual behavioral performance values on the left-out data. This correlation remained
26
significant, though as expected was of lower magnitude, with an adjusted R
2
= 0.15 (p =
0.006).
This result supports the predictive utility of the two neural indices of LPFC
function, the linear slope of load-related activity and GBC, for predicting N-back
performance in out-of-sample data. Interestingly, cross-validation analyses also
demonstrated that LPFC activity (linear slope) was a significant predictor in isolation
since it remained significant in leave-one-out cross-validation tests with it as the only
predictor variable (r=0.29, p=0.036). However, the same was not true for LPFC
functional connectivity (GBC), as it was no longer significant when included as the only
predictor variable (r = 0.17, p = 0.22).
This conclusion was further supported by a permutation test implemented by
computing the correlation between actual and predicted accuracy (in 1000 iterations)
after randomly shuffling accuracy values (in a linear model where LPFC activity slope
and LPFC GBC were simultaneous predictors of accuracy). As expected, the mean
correlation (mean correlation from 1000 iterations after Fisher r-to-z transformations)
between predicted and true accuracy was near zero across permutations (r = 0.0066). In
1000 iterations only 8 of the permutations was the correlation as high as in our original
dataset (r = 0.36). Thus, the cross-validation test confirms a highly significant correlation
between predicated and actual accuracy (p = 0.006)
2
.
3.6.Testing generalization of brain-behavioral performance prediction
2
Note that we also conducted a parallel set of analyses that replaced the linear
slope parameter with the factor score as the index of LPFC activity. These analyses
provided very similar conclusions to the ones above and are reported in Supplemental
results.
27
The prior analyses suggested the predictive validity of LPFC neural measures for
predicting individual differences in WM function. To further test for the generalizability
of these predictions, we examined whether this unique left LPFC ROI, which was not
well-identified in by any existing parcellation scheme (Cole et al., 2015; Cole et al.,
2012), exhibited predictive power related to N-back behavioral performance in another,
much larger dataset, albeit one that did not examine parametric manipulations of working
memory load. To do this, we made use of the publicly available HCP dataset, which
includes N-back data from the 2-back and 0-back condition. To simplify the analysis, we
included data from the 500-release set, since this was the last set to provide results from a
volume-based GLM analysis (i.e., the largest release that used an analysis approach
compatible with the use of volume-based, voxelwise ROIs). Furthermore, from this
release we used only unrelated participants (n = 198), to avoid potential confounds in
using twins and other related individuals. To provide the strongest test of generalization,
we tested whether 2-back activation in our LPFC ROI predicted an out-of-scanner
measures of WM and executive control function that were collected in that study: List
Sorting (i.e., a standard WM measure included as part of the NIH Toolbox; (Barch et al.,
2013; Gershon et al., 2013) and Penn Matrices (Bilker et al., 2012); a measure of fluid
cognition or fluid intelligence/gF). The advantage of using these out-of-scanner
behavioral measures is that any observed associations cannot be attributed to the
contemporaneous collection of brain activity and behavioral performance measures,
which serves as a potential confound when using N-back accuracy as the behavioral
measure. Instead, a correlation with a separate, out-of-scanner measure like List Sort or
PMAT performance would indicate that N-back-related LPFC activity reflects a more
28
stable trait-related index of WM/EC function. Indeed, the robust correlation between
LPFC 2-back activity and both List Sort performance (r=0.26, p < 0.001) and PMAT
(r=0.22, p < 0.0015) supports the hypothesis that LPFC recruitment is trait-like, and thus
generalizes across tasks. Moreover, when comparing the magnitude of this brain-behavior
correlation (with List Sort) relative to all the other (264) nodes in the Power parcellation,
we found that it was one of the strongest – indeed only 4 other nodes showed slightly
stronger correlations (highest r = 0.285) and three of these were also in the FPN. This
finding demonstrates clearly that this LPFC ROI can be expected to robustly reflect
brain-behavior relationships in other N-back datasets and with other behavioral measures
of WM/EC function.
For completeness, we also note that activity in this LPFC ROI was also reliably
correlated with in-scanner 2-back performance as well in this HCP dataset (r=0.23,
p=0.001). The magnitude of this correlation was comparable to that observed in our own
dataset when restricting to the 2-back activity level (r=0.23).
3.7. The relative predictive power of high vs. low load data in the N-back
One of the most unique and potentially counter-intuitive aspects of our study
design and analysis is that we examined N-back activity and performance at levels
beyond those standardly tested in either behavioral or brain imaging studies of the N-
back. Indeed, to our knowledge, this study is the first to examine six parametric levels of
N-back fMRI and performance data. The likely reason for the uniqueness of our design is
that conventional intuitions regarding the N-back are that the high load levels are too
difficult for participants to perform well, and thus likely would be less sensitive to
29
individual differences in brain activity and behavioral performance, due to floor effects.
We tested this assumption directly via analyses that separated the data into low
load (N = 1—3) and high load (N = 4—6) subsets, since it is the high load conditions that
are most unique to our study. We then conducted analyses that tested brain-behavior
relationships in various ways in these two subsets. First, we replicated the multiple
regression analysis described above in which we retained the behavioral accuracy factor
score that included all 6 load levels, but then split the LPFC activity predictor in two,
with one indicating the linear slope effect in only the low load conditions (1,2,3) and the
other indicating linear slope in only the high load conditions (4,5,6). Thus, there were a
total of 3 predictor variables (LPFC GBC, LPFC 123-slope, LPFC 456-slope). In a
multiple regression we found that the total explained variance was similar at 22%, but
that only the GBC and 456-slope predictors were statistically significant (LPFC GBC:
beta = 3.26, t(46) = 2.25, p = 0.03; 123-slope: beta = 8.06, t(46) = 0.95, p = 0.34; 456-
slope: beta = 33.58, t(46) = 1.97, p = 0.055).
Second, we conducted separate regressions with just the low load predictors
(along with LPFC GBC) predicting accuracy in just the low load conditions (by creating
a factor score summary over just N = 1—3) and just the high load predictors (+LPFC
GBC) predicting accuracy in just the high load conditions (again with N = 4—6
behavioral factor score summary). In this analysis, low load brain measures explained
only 5% of the variance in low load performance and neither of the predictors were
significant (LPFC GBC: beta = 0.08, t(48) < 1; 123-slope: beta = 13.23, t(48) = 1.64, p =
0.10). In contrast, high load brain data explained 16% of variance in high load
performance, and both predictors were significant (LPFC GBC: beta = 3.09, t(48)= 2.15,
30
p = 0.037; 456-slope: beta = 33.17, t(48)= 2.17, p = 0.035)
3
.
Together these results suggest that, counter to standard intuitions, brain-behavior
relationships are stronger in very high load conditions relative to low load conditions.
Thus, including very high load conditions increased our sensitivity to detect these brain-
behavior relationships. To quantify this sensitivity, we directly assessed the proportion of
total variability explained by individual differences (i.e., between-subject variability)
using the intraclass correlation coefficient (ICC), which provides a measure of both how
reliable are individual differences, and the relative proportion of variance that is due to
load effects (i.e., within-subject variability vs. individual differences). The ICC statistic is
typically used in test-retest reliability analyses, but for our purposes, each load-level
condition was treated as a “retest” event. All analyses were conducted using the ‘psych’
package in R (Revelle, 2018), and report the ICC(3,k) metric, which is the most
conservative.
When examining all load conditions (N = 1—6), the ICC estimate for target
accuracy was 0.83 (95% CI: 0.74—0.89); of that, the proportion of variance due to load
was 0.68 and to individual differences 0.14 (0.18 residual). For the LPFC BOLD data,
again with all load conditions, the ICC estimate was 0.93 (95% CI: 0.90—0.95); with
proportion of variance due to load 0.05 and to individual differences 0.65 (0.29 residual).
This indicates high reliability of both measures, but with varying sensitivity to individual
differences.
Next, we compared variance explained when separately examining low (N = 1—
3
Again we ran a parallel set of analyses that replaced the linear slope parameter
with the factor score as the index of LPFC activity. Again these additional analyses
provided the same conclusions as drawn above, attesting to their robustness. These
analyses are also reported in Supplemental results.
31
3) and high (N = 4—6) conditions. For both the behavioral and BOLD data the effects
were striking, with increased ICCs and proportion of variance due to individual
differences in the high load conditions (target accuracy: ICC = 0.85, 0.76—0.90 95% CI,
proportion of variance due to individual difference = 0.51 and to load = 0.20; BOLD:
ICC = 0.91, 0.86—0.95 95% CI, proportion of variance due to individual differences =
0.77 and to load = .01) compared to low load (target accuracy: ICC = 0.62, 0.41—0.76
95% CI, proportion of variance due to individual difference = 0.14 and to load = 0.59;
BOLD: ICC = 0.82, 0.73—0.89 95% CI, proportion of variance due to individual
difference = 0.53 and to load = 0.13). Taken together, these data support the idea, that
counter to standard intuitions, the high load condition is actually more sensitive for the
detection of individual differences in both behavioral performance and brain activity.
4. Discussion
Our study fills an important gap in our understanding of brain-behavior
relationships in working memory tasks, and in the N-back in particular. Although the N-
back is one of the most widely used paradigms to study working memory and executive
control, there is still a poor understanding of how brain activity varies by load and how
load-related activity patterns relate to task performance. Our study is unique in that we
examined brain-behavior relationships in an N-back study design that used a very wide
range of load levels, spanning from N = 1-6. To our knowledge, no previous studies of
the N-back have systematically examined brain activity and behavioral performance at
very high load levels (N > 3). The key question of interest was how LPFC activity varied
32
with load and in comparison to other regions, and whether LPFC contributed to
behavioral performance in a systematic way across load levels.
A systematic analysis of brain activity and performance across a large sample,
and a wide range of load levels revealed multiple novel observations. First, we clearly
replicated the inverted U-shaped pattern that has been a prominent feature of prior studies
(Callicott et al., 1999; Jaeggi et al., 2007; Van Snellenberg et al., 2015), leading to the
greatest activation at middle load levels (i.e., 2/3-back). This feature was not only
prominent in LPFC activity, but observed brain-wide, as it was present in many other
networks related to WM/EC as described in Supplemental Results. Second, we found
that within a focal left LPFC ROI, load-related activity reliably predicted N-back
behavioral performance. Moreover, this brain-behavior relationship was selective: it was
observed with the linear, rather than the inverted-U component of load-related activation,
it was not found with RT measures, and it was unique relative to other brain regions and
even whole-networks, as revealed in supplemental analyses (Supplemental Results).
Third, we found that global connectivity with this focal LPFC region was an independent
predictor of N-back task performance (i.e., even when considering LPFC activation).
Fourth, rigorous cross-validation analyses demonstrated the predictive utility of load
functions in the LPFC, which also generalized to a large out-of-sample dataset (HCP).
Finally, and potentially one of the most counter-intuitive aspects of our findings, the
highest load levels (4-6) of the N-back, rather than standard lower-load levels (N
≤
3),
were the most sensitive for detecting brain-behavior relationships, as they exhibited the
greatest individual variability in both performance and brain activity. Thus, together the
results highlight the utility of our approach in testing for brain-behavior relationships in
33
WM tasks such as the N-back when sampling across a very wide-range of load levels.
Nevertheless, the results do point to a number of puzzling and unresolved issues in terms
of the neural mechanisms of WM function, that we discuss next.
4.1. Load-related activity functions: Meaning of the inverted-U pattern
The current findings replicate and extend a now consistent pattern of inverted-U
load functions observed in neuroimaging studies of WM (Jaeggi et al., 2007; Jansma,
2004; Van Snellenberg et al., 2015). The inverted-U shape has puzzled investigators, and
several hypotheses have been proposed. The most prominent is that the inflection point,
in which activation levels start decreasing as load continues to increase, may reflect the
point in which WM capacity is exceeded (Callicott et al., 1999; Haier, Siegel, Tang,
Abel, & Buchsbaum, 1992; Neubauer, Grabner, Fink, & Neuper, 2005). This account is
bolstered by findings of a close correspondence between the load level in which the
inflection point occurs and independent measures of WM capacity (Vogel, McCollough,
& Machizawa, 2005). These findings have not only been observed in N-back paradigms,
but in delayed match-to-sample paradigms (such as the Sternberg item recognition task)
in which the inverted-U function has been linked to the active maintenance of
information in working memory (Cappell et al., 2010; Karlsgodt et al., 2007; Karlsgodt et
al., 2009).
Other accounts have postulated that inverted-U functions reflect task
disengagement, or a shift in processing strategy, which may occur somewhat
independently of available capacity (i.e., due to other considerations, such as cognitive
effort avoidance (Jaeggi et al., 2007; Jonides & Nee, 2006; Vogel et al., 2005). It is also
consistent with neuro-computational models of WM, which suggest that the balance
34
between recurrent connectivity and strong lateral inhibition leads to capacity constraints
that create sub-linear relationships between load and average activation (Chatham et al.,
2011; Edin et al., 2009; Rolls, Dempere-Marco, & Deco, 2013; Wei, Wang, & Wang,
2012).
Our results contribute to this literature in several ways. First, our design increased
sensitivity to the inverted-U pattern due to the wider range of load levels employed, at
least relative to standard N-back paradigms (Van Snellenberg et al., 2015). Thus, we
confirm that in the LPFC (and indeed in other brain networks as well; see Supplemental
Results), there is a definite non-monotonic pattern with activation levels being strongest
at N = 3, and decreasing from that at higher levels (e.g., Figure 3A). Interestingly,
however, we found that it was the linear rather than non-linear load effect that best
predicted individual differences in behavioral performance. In particular, highest
performing participants showed a definite linear increase in activity, whereas lower
performing participants tended to show decreasing activity. This finding suggests that
although there may be an overlying pattern of decreasing activity in all participants at
high load levels, it is the strength of the linear load-pattern (i.e., the tendency to
monotonically increase, or at least not decrease, activity with increasing load) that most
strongly discriminates high and low performers. Thus, the results suggest the continued
utility of linear load modeling, to test for individual variation in WM function, even given
the presence of non-monotonic patterns.
Nevertheless, a limitation of our study is that we did not have an independent
measure of WM capacity, which ideally would be assessed out-of-scanner, with well-
established psychometric measures (e.g., standard span or change detection tasks
35
(Conway et al., 2005; Kyllingsbaek & Bundesen, 2009; Luck & Vogel, 2013).
Consequently, a direction for future research would be to determine whether and how
WM capacity limits relate to the distinctions between high and low performers we
observed in terms of N-back load patterns in the current study, and moreover, whether
capacity indices can be used to predict where the inflection point in inverted-U patterns is
located or whether it exists (cf. van Snellenberg et al, 2015).
Data of this type (i.e., independent measures of WM capacity) would also be
informative with regard to our finding that between-subjects variability in BOLD signal
was not related to reaction times. Conversely, supplemental analyses confirmed that
within-subject (rather than between-subject) load-related inverted-U patterns seemed to
better track with reaction time (see Supplemental Results). It is possible that the
inflection point within the inverted-U BOLD activity load functions might predict not
only between-subjects variability in N-back accuracy, but also the inverted-U and
inflection point in load-related RT patterns (i.e., and related to WM capacity measures).
Such findings would support the idea that inverted-U patterns reflect something more
about how target/non-target response decisions are reached, rather than about the quality
of information storage per se. Although it was beyond the scope of current study to
directly test for such effects (which would require more trials at each load condition, and
more explicit manipulation of decision-related factors), this is an issue that could be
addressed in future work, particularly through the use of evidence accumulation decision-
making models: e.g., drift diffusion, linear ballistic accumulator (Ratcliff, Smith, Brown,
& McKoon, 2016).
4.2. Focal region vs. network-level contributions to WM performance
36
Although our focus was brain-behavior relationships within a focal brain region
(LPFC), we also investigated whether parallel relationships were observed at the network
level. In supplemental analyses we also tested cognitive brain networks including the
broader frontoparietal network, as well as the dorsal attention, cingulo-opercular, and
default mode networks. Paralleling our findings with the LPFC, in none of these
networks was the inverted U-shaped or linear pattern associated with performance. It is
worth noting that some analyses did reveal brain-behavior relationships in the DMN (see
Supplemental Results), but these relationships were observed at the average level of task-
related deactivation, rather than in terms of task-related activity that was specifically
load-related.
It is possible that these null findings with regard to “task-positive” networks is
actually a false negative, and that significant brain-behavior effects might have emerged
with larger sample sizes. In fact, other work using very large N-back datasets have
pointed to network-level prediction of WM performance (Bolt et al., 2018; Egli et al.,
2018). Nevertheless, it is also possible that the lack of load-related findings present at the
brain network level within the current dataset reflect a true pattern that is related to the
use of a wider-range of load levels than has previously been studied in the N-back. For
example, if inverted-U patterns reflect capacity limitations, the inverted-U patterns
observed at lower loads (N =1—3) would be primarily driven by low-capacity
individuals. These patterns would also thus obscure any linear effects that only emerge
for higher-capacity individuals across the wider-range of loads (N = 1—6). Thus the
wider-range may be necessary to capture and distinguish between linear and quadratic
effects, and relate them to individual differences in performance. Moreover, it is possible
37
that linear versus quadratic components may emerge to varying degrees across load levels
in different parts of a given brain network. Hence, when considering the full load range
(N = 1—6), subtler linear effects might be most sensitively be detected in focal regions
(e.g. the LPFC), rather than in entire networks.
Our results do confirm the functional importance of this focal left LPFC region
for WM task performance. Indeed, the results are consistent with the findings of many
meta-analyses, which have pointed to the reliable engagement of this particular region in
WM paradigms. For example, Rottschy et al. (2012) highlight this region as a key
component of what they refer to as the “core” WM network. Furthermore, in our own
prior work we found that this region was uniquely selective in predicting N-back task
performance both in terms of within-subject and between-subject indicators (Cole et al.,
2012). Although the current results do not highlight exactly how and why this region
contributes to WM in such a unique way, they do point to the need for further targeted
investigations of this region, to better reveal the mechanisms by which it contributes to
WM function
4
.
Nevertheless, an important implication of the current results is that they clearly
underscore the potential importance of conducting region-focused analyses in addition to
network-based ones. Although network-focused analyses are useful for dimensionality
4
While revising this manuscript for publication, we took an initial step towards
better understanding of the anatomic and functional specialization of this left LPFC ROI,
by taking advantage of a new anatomic parcellation scheme (Ji et al., 2019), which
subdivided brain regions into not only standard WM/EC brain networks, but also newly
defines a left-lateralized language network, which also involves the LPFC. Overlapping
our ROI onto this parcellation revealed that our ROI primarily overlapped with the FPN,
with only a small fraction overlapping the language network (see Supplemental). This
finding confirmed our intuition that although the region may reflect a unique functional
region, it seems to belong within the FPN proper.
38
reduction, our results suggest the potential limitations of such approaches, as they may
obscure focal and unique contributions to functionality. As a concrete example of this
point, the recent study of Egli et. al., (2018) analyzed their large N-back dataset using
independent component analyses (ICA) as a data-driven dimensionality reduction
approach, which they argued revealed the presence of two unique networks: a parietally-
centered network related more to WM load effects and a frontally-centered network was
more involved with general sustained attention. However, close inspection of their own
data also points to the importance of the same left mid-lateral PFC region we focus on
here. Nevertheless, because of their network-focus, Egli et. al., (2018) do not highlight
this region in their results, which otherwise would cause its potentially unique
contribution to behavioral performance to be overlooked.
4.3. Global connectivity vs. activity within LPFC
Although the first-generation of WM neuroimaging studies focused exclusively
on relating BOLD response magnitude to load manipulations, the field has clearly shifted
to focus on functional connectivity as an important predictor of WM performance. In
Cole et. al., (2012), we highlighted the GBC metric as a potentially powerful summary
measure of functional connectivity that could be associated with focal brain regions.
Moreover, in that study we demonstrated that GBC within the left LPFC showed a strong
degree of individual variation, which critically appeared to have strong functional
consequences, in predicting not only N-back accuracy, but also broader measures related
to WM function (i.e., working memory capacity and fluid intelligence). The current study
replicated this pattern in a new dataset, and moreover replicated the finding that LPFC
39
GBC and LPFC activity served as independent predictors of a behavioral measure of WM
function (N-back accuracy).
The finding of independent sources of individual variation in both the GBC and
activity of LPFC begs the question of what each of these two metrics reflect, and how
they relate. More broadly, the information content of mean activation versus that of
functional connectivity is of growing general interest, moreover there have been concerns
raised about the growing divide in studies focused on functional connectivity
(particularly resting-state) and those focused on task-related activation. Recent attempts
have been made to integrate connectivity and activity-based analyses, of which a notable
example is activity flow mapping (Cole, Ito, Bassett, & Schultz, 2016). Our results,
however, are consistent with the idea that the activity and connectivity patterns associated
with LPFC contribute unique variance to behavioral performance. In particular, the
factors associated with individual variation in LPFC activity and with LPFC GBC seem
to be functionally independent, and so potentially reflect distinct causal mechanisms.
Moreover, the results also replicate other prior results suggesting the importance of
resting-state functional connectivity patterns (Sala-Llonch et al., 2012), particularly
involving the DLFPC, as an important and unique dimension of individual difference
with clear implications for WM function.
4.4. The importance of high-load WM conditions
Potentially the most surprising contribution of this study was the finding that
high-load (i.e., N > 3), rather than low-load working memory conditions were the most
sensitive for identifying individual differences. A common assumption that high-load
conditions in the N-back exceed most participants’ WM capacity predicts that
40
performance would be at floor for N
≥
3, and that variability in BOLD activity patterns
would merely reflect noise. On the contrary, we found that high-load conditions most
strongly differentiated individuals, behaviorally and in terms of activity patterns. Given
that individual differences analyses rely on between-subjects variability, higher load
levels may thus be critical for detecting brain-behavior relationships.
Indeed, it may have been precisely the utilization of high-load manipulations
which provided the necessary discriminating power to identify individuals who were able
to maintain high levels of performance and LPFC activity. Those with shallower
decreases in performance showed more positive linear slopes in activity patterns. It is
possible that under such high-load conditions, preservation of performance, and the brain
activity metrics are more closely reflecting processes related to cognitive control factors,
rather than simple active maintenance, such as sustaining cognitive goals, resisting
tendencies to distraction and mind-wandering, or potential for affective reactivity to
internal negative performance feedback signals. In fact, to speculate, it may be that the
high-load LPFC metrics may reflect control processes more closely than simple active
maintenance, and these may be the most critical dimensions of individual differences in
cognitively demanding tasks, such as the N-back. If so, the findings would be consistent
with the view that the N-back should be more strongly construed as a probe of cognitive
control functions, than pure working memory
per se
, and this might be particularly true at
high-load levels, which are most control-demanding. A key implication of the current
results is that if investigators are most interested in individual differences, high-load
rather low-load N-back conditions should be emphasized. Notably, this recommendation
is essentially opposite to common intuitions and predominant practice governing N-back
41
studies since the beginning. Of course, one caveat is that we can only make this
recommendation for studies involving healthy young adults, since that is the population
we studied here. Future work will need to determine whether high-load N-back
conditions are also equally efficacious and sensitive when examining other populations of
interest.
Acknowledgments
We would like to thank Sarah Adams in the Cognitive Control & Psychopathology Lab at
Washington University in Saint Louis for helping us in data collection.
Declaration of Conflicting Interests
The authors declared that they had no conflicts of interest with respect to their
authorship or the publication of this article.
Funding
This work was supported by National Institutes of Health grants: R37 MH066078, R01
AG043461, and R21 AG058206 to T.S.B, and R01 AG055556 and R01 MH109520 to
M.W.C. This work was also supported by grant 2011246 from the USA-Israel Bi-national
Science Foundation to T.S.B. and M.W.C. The content of this article is solely the
responsibility of the authors and does not necessarily represent the official views of the
funding agencies.
References
Ackerman, P. L., Beier, M. E., & Boyle, M. O. (2005). Working memory and intelligence: the
same or different constructs? Psychol Bull, 131(1), 30-60.
Barch, D. M., Burgess, G. C., Harms, M. P., Petersen, S. E., Schlaggar, B. L., Corbetta, M., et al.
(2013). Function in the human connectome: task-fMRI and individual differences in
behavior. Neuroimage, 80, 169-189.
42
Bilker, W. B., Hansen, J. A., Brensinger, C. M., Richard, J., Gur, R. E., & Gur, R. C. (2012).
Development of abbreviated nine-item forms of the Raven's standard progressive
matrices test. Assessment, 19(3), 354-369.
Bolt, T., Prince, E. B., Nomi, J. S., Messinger, D., Llabre, M. M., & Uddin, L. Q. (2018).
Combining region- and network-level brain-behavior relationships in a structural
equation model. Neuroimage, 165, 158-169.
Braver, T. S., Cohen, J. D., Nystrom, E. N., Jondies, J., Smith, E. E., & Noll, D. C. (1997). A
Parametric Study of Prefrontal Cortex Involvement in Human Working Memory.
Neuroimage, 5, 49-62.
Callicott, J. H., Mattay, V. S., Bertolino, A., Coppola, R., Frank, J. A., Goldberg, T. E., et al.
(1999). Physiological Characteristics of Capacity Constraints in Working Memory as
Revealed by Functional MRI. Cerebral Cortex, 9, 20-26.
Cappell, K. A., Gmeindl, L., & Reuter-Lorenz, P. A. (2010). Age differences in prefontal
recruitment during verbal working memory maintenance depend on memory load.
Cortex, 46(4), 462-473.
Chatham, C. H., Herd, S. A., Brant, A. M., Hazy, T. E., Miyake, A., O'Reilly, R., et al. (2011).
From an executive network to executive control: a computational model of the n-back
task. J Cogn Neurosci, 23(11), 3598-3619.
Choo, W. C., Lee, W. W., Venkatraman, V., Sheu, F. S., & Chee, M. W. (2005). Dissociation of
cortical regions modulated by both working memory load and sleep deprivation and by
sleep deprivation alone. Neuroimage, 25(2), 579-587.
Cohen, J. D., Perlstein, W. M., Braver, T. S., Nystrom, E. N., Noll, D. C., Jonides, J., et al.
(1997). Temporal dynamics of brain activation during a working memory task. Nature,
386, 604-607.
Cole, M. W., Ito, T., Bassett, D. S., & Schultz, D. H. (2016). Activity flow over resting-state
networks shapes cognitive task activations. Nat Neurosci, 19(12), 1718-1726.
Cole, M. W., Ito, T., & Braver, T. S. (2015). Lateral Prefrontal Cortex Contributes to Fluid
Intelligence Through Multinetwork Connectivity. Brain Connect, 5(8), 497-504.
Cole, M. W., Yarkoni, T., Repovs, G., Anticevic, A., & Braver, T. S. (2012). Global connectivity
of prefrontal cortex predicts cognitive control and intelligence. J Neurosci, 32(26), 8988-
8999.
Conway, A. R. A., Kane, M. J., Bunting, M. F., Hambrick, D. Z., Wilhelm, O., & Engle, R. W.
(2005). Working memory span tasks: A methodological review and user’s guide.
Psychonomic Bulletin & Review, 12(5), 769-786.
Dosenbach, N. U., Visscher, K. M., Palmer, E. D., Miezin, F. M., Wenger, K. K., Kang, H. C., et
al. (2006). A core system for the implementation of task sets. Neuron, 50(5), 799-812.
Edin, F., Klingberg, T., Johansson, P., McNab, F., Tegner, J., & Compte, A. (2009). Mechanism
for top-down control of working memory capacity. Proc Natl Acad Sci U S A, 106(16),
6802-6807.
Egli, T., Coynel, D., Spalek, K., Fastenrath, M., Freytag, V., Heck, A., et al. (2018). Identification
of Two Distinct Working Memory-Related Brain Networks in Healthy Young Adults.
eNeuro, 5(1).
Engle, R. W., Laughlin, J. E., Tuholski, S. W., & Conway, A. R. A. (1999). Working Memory,
Short-Term Memory, and General Fluid Intelligence: A Latent-Variable Approach.
Journal of Experimental Psychology: General, 128(3), 309-331.
Ewing, K., & Fairclough, S. (2010). The impact of working memory load on psychophysiological
measures of mental effort and motivational disposition. In: de Waard D, Axelsson A,
Berglund M, Peters B, Weickert C, editors. Human Factors: A system view of human,
technology and organisation. . Maastricht: Shaker Publishing.
43
Gershon, R. C., Wagster, M. V., Hendrie, H. C., Fox, N. A., Cook, K. F., & Nowinski, C. J.
(2013). NIH Toolbox for Assessment of Neurological and Behavioral Function.
American Academy of Neurology, 80, S2-S6.
Gevins, A., & Cutillo, B. (1993). Spatiotemporal dynamics of component processes in human
working memory. Electroencephalography and clinical Neurophysiology, 87, 128-143.
Haier, R. J., Siegel, B., Tang, C., Abel, L., & Buchsbaum, M. S. (1992). Intelligence and Changes
in Regional Cerebral Glucose Metabolic Rate Following Learning. Intelligence, 16, 415-
426.
Harvey, P. O., Fossati, P., Pochon, J. B., Levy, R., Lebastard, G., Lehericy, S., et al. (2005).
Cognitive control and brain resources in major depression: an fMRI study using the n-
back task. Neuroimage, 26(3), 860-869.
Jaeggi, S. M., Buschkuehl, M., Etienne, A., Ozdoba, C., Perrig, A. J., & Nirkko, A. C. (2007). On
how high performers keep cool brains in situations of cognitive overload. Cognitive,
Affective, & Behavioral Neuroscience, 7(2), 75-89.
Jansma, J. (2004). Working memory capacity in schizophrenia: a parametric fMRI study.
Schizophrenia Research, 68(2-3), 159-171.
Ji, J. L., Spronk, M., Kulkarni, K., Repovs, G., Anticevic, A., & Cole, M. W. (2019). Mapping
the human brain's cortical-subcortical functional network organization. Neuroimage, 185,
35-57.
Jo, H. J., Saad, Z. S., Simmons, W. K., Milbury, L. A., & Cox, R. W. (2010). Mapping sources of
correlation in resting state FMRI, with artifact detection and removal. Neuroimage, 52(2),
571-582.
Jonides, J., & Nee, D. E. (2006). Brain mechanisms of proactive interference in working memory.
Neuroscience, 139(1), 181-193.
Juvina, I., & Taatgen, N. A. (2007). Modeling control strategies in the N-back task. In
Proceedings of the 8th International Conference on Cognitive Modeling, New York. NY:
Psychology Press. pp 73-78
.
Kane, M. J., & Engle, R. W. (2002). The role of prefrontal cortex in working-memory capacity,
executive attention, and general fluid intelligence: An individual-differences perspective.
Psychonomic Bulletin & Review, 9(4), 637-671.
Karlsgodt, K. H., Glahn, D. C., van Erp, T. G., Therman, S., Huttunen, M., Manninen, M., et al.
(2007). The relationship between performance and fMRI signal during working memory
in patients with schizophrenia, unaffected co-twins, and control subjects. Schizophr Res,
89(1-3), 191-197.
Karlsgodt, K. H., Sanz, J., van Erp, T. G., Bearden, C. E., Nuechterlein, K. H., & Cannon, T. D.
(2009). Re-evaluating dorsolateral prefrontal cortex activation during working memory in
schizophrenia. Schizophr Res, 108(1-3), 143-150.
Kyllingsbaek, S., & Bundesen, C. (2009). Changing change detection: improving the reliability of
measures of visual short-term memory capacity. Psychon Bull Rev, 16(6), 1000-1010.
Kyllonen, P. C., & Christal, R. E. (1990). Reasoning Ability is ( Little More Than) working-
Memory Capacity ?! Intelligence, 14, 389-433.
Luck, S. J., & Vogel, E. K. (2013). Visual working memory capacity: from psychophysics and
neurobiology to individual differences. Trends Cogn Sci, 17(8), 391-400.
Meng, X., Rosenthal, R., & Rubin, D. B. (1992). Comparing Correlated Correlation Coefficients.
Psychological Bulletin, 111(1), 172-175.
Miller, E. K., & Cohen, J. D. (2001). AN INTEGRATIVE THEORY OF PREFRONTAL
CORTEX FUNCTION. Annu. Rev. Neurosci., 24, 167-202.
Neubauer, A. C., Grabner, R. H., Fink, A., & Neuper, C. (2005). Intelligence and neural
efficiency: further evidence of the influence of task content and sex on the brain-IQ
relationship. Brain Res Cogn Brain Res, 25(1), 217-225.
44
Nyberg, L., Dahlin, E., Stigsdotter Neely, A., & Backman, L. (2009). Neural correlates of
variable working memory load across adult age and skill: dissociative patterns within the
fronto-parietal network. Scand J Psychol, 50(1), 41-46.
Otto, T., Zijlstra, F. R., & Goebel, R. (2014). Neural correlates of mental effort evaluation--
involvement of structures related to self-awareness. Soc Cogn Affect Neurosci, 9(3), 307-
315.
Owen, A. M., McMillan, K. M., Laird, A. R., & Bullmore, E. (2005). N-back working memory
paradigm: a meta-analysis of normative functional neuroimaging studies. Hum Brain
Mapp, 25(1), 46-59.
Power, J. D., Cohen, A. L., Nelson, S. M., Wig, G. S., Barnes, K. A., Church, J. A., et al. (2011).
Functional network organization of the human brain. Neuron, 72(4), 665-678.
Ratcliff, R., Smith, P. L., Brown, S. D., & McKoon, G. (2016). Diffusion Decision Model:
Current Issues and History. Trends Cogn Sci, 20(4), 260-281.
Revelle, W. (2018). psych: Procedures for Personality and Psychological Research, Northwestern
University, Evanston, Illinois, USA, https://CRAN.R-project.org/package=psych Version
= 1.8.12.
Rolls, E. T., Dempere-Marco, L., & Deco, G. (2013). Holding Multiple Items in Short Term
Memory: A Neural Mechanism. PLoS One, 8(4), 1-13.
Rottschy, C., Langner, R., Dogan, I., Reetz, K., Laird, A. R., Schulz, J. B., et al. (2012).
Modelling neural correlates of working memory: a coordinate-based meta-analysis.
Neuroimage, 60(1), 830-846.
Sala-Llonch, R., Pena-Gomez, C., Arenaza-Urquijo, E. M., Vidal-Pineiro, D., Bargallo, N.,
Junque, C., et al. (2012). Brain connectivity during resting state and subsequent working
memory task predicts behavioural performance. Cortex, 48(9), 1187-1196.
Van Snellenberg, J. X., Slifstein, M., Read, C., Weber, J., Thompson, J. L., Wager, T. D., et al.
(2015). Dynamic shifts in brain network activation during supracapacity working
memory task performance. Hum Brain Mapp, 36(4), 1245-1264.
Vogel, E. K., McCollough, A. W., & Machizawa, M. G. (2005). Neural measures reveal
individual differences in controlling access to working memory. Nature, 438(7067), 500-
503.
Wager, T. D., Spicer, J., Insler, R., & Smith, E. E. (2014). The neural bases of distracter-resistant
working memory. Cogn Affect Behav Neurosci, 14(1), 90-105.
Wei, Z., Wang, X. J., & Wang, D. H. (2012). From distributed resources to limited slots in
multiple-item working memory: a spiking network model with normalization. J Neurosci,
32(33), 11228-11240.
Westbrook, A., Kester, D., & Braver, T. S. (2013). What is the subjective cost of cognitive effort?
Load, trait, and aging effects revealed by economic preference. PLoS One, 8(7), e68210.
Westbrook, A., Lamichhane, B., & Braver, T. (2019). The Subjective Value of Cognitive Effort is
Encoded by a Domain-General Valuation Network. J Neurosci.
Wickens, T. D. (2002). Elementary signal detection theory. New York, NY, US: Oxford
University Press.
Yarkoni, T., & Westfall, J. (2017). Choosing Prediction Over Explanation in Psychology:
Lessons From Machine Learning. Current Directions in Psychological Science, 21(6),
391-397.
