Available via license: CC BY 4.0
Content may be subject to copyright.
Accepted manuscripts are peer-reviewed but have not been through the copyediting, formatting, or proofreading
process.
Copyright © 2018 Petroni et al.
This is an open-access article distributed under the terms of the Creative Commons Attribution 4.0 International license, which
permits unrestricted use, distribution and reproduction in any medium provided that the original work is properly attributed.
This Accepted Manuscript has not been copyedited and formatted. The final version may differ from this version.
Research Article: New Research | Cognition and Behavior
The Variability of Neural Responses to Naturalistic Videos Change with
Age and Sex
Agustin Petroni1, Samantha S. Cohen1,2, Lei Ai3, Nicolas Langer1,3,4, Simon Henin1, Tamara Vanderwal5,
Michael P. Milham3,6 and Lucas C. Parra1
1Department of Biomedical Engineering, City College of New York, New York, NY 10031, USA
2Department of Psychology, the Graduate Center of the City University of New York, New York, NY 10016, USA
3Center for the Developing Brain, Child Mind Institute, New York, NY 10022, USA
4Methods of Plasticity Research, Department of Psychology, University of Zurich, 8050, Switzerland
5Yale Child Study Center, New Haven, CT 06520, USA
6Nathan Kline Institute for Psychiatric Research, Orangeburg, NY 10962, USA
DOI: 10.1523/ENEURO.0244-17.2017
Received: 7 July 2017
Revised: 8 December 2017
Accepted: 14 December 2017
Published: 12 January 2018
Author Contributions: MM, LP, SC and NL Designed Research; NL Performed Research; AP, SC, NL, LA, and
LP Analyzed Data; AP, SC, LA, NL, SH, TV, MM, and LP Wrote the paper.
Funding: http://doi.org/10.13039/100000185DOD | Defense Advanced Research Projects Agency (DARPA)
W911NF-14-1-0157
Conflict of Interest: Authors report no conflict of interest.
This work was supported by the Defense Advanced Research Projects Agency (Contract W911NF-14-1-0157).
A.P. and S.S.C. are co-first authors.
Correspondence should be addressed to Lucas C. Parra, Department of Biomedical Engineering, City
College of New York, 160 Convent Ave., New York, NY 10031, USA. Phone: +1-212-650-8653; Email:
parra@ccny.cuny.edu
Cite as: eNeuro 2018; 10.1523/ENEURO.0244-17.2017
Alerts: Sign up at eneuro.org/alerts to receive customized email alerts when the fully formatted version of this
article is published.
1
1. Manuscript Title: The variability of neural responses to naturalistic videos change 1
with age and sex 2
2. Abbreviated Title (50 character maximum): Variability of neural responses to videos 3
3. List all Author Names and Affiliations in order as they would appear in the published 4
article 5
Agustin Petroni1*, Samantha S. Cohen1,2*, Lei Ai3, Nicolas Langer1,3,4, Simon Henin1, 6
Tamara Vanderwal5, Michael P. Milham3,6, Lucas C. Parra1+ 7
8
* A.P. and S.S.C. are co-first authors 9
10
1 Department of Biomedical Engineering, City College of New York, New York, NY 11
10031, USA. 12
2 Department of Psychology, The Graduate Center of the City University of New 13
York, New York, NY 10016, USA. 14
3 Center for the Developing Brain, Child Mind Institute, New York, NY 10022, USA. 15
4 Methods of Plasticity Research, Department of Psychology, University of Zurich, 16
8050, Switzerland. 17
5 Yale Child Study Center, New Haven, CT 06520, USA. 18
6 Nathan Kline Institute for Psychiatric Research, Orangeburg, NY 10962, USA. 19
4. Author Contributions: 20
MM, LP, SC and NL Designed Research; NL Performed Research; AP, SC, NL, LA, and LP 21
Analyzed Data; AP, SC, LA, NL, SH, TV, MM, and LP Wrote the paper. 22
5. Correspondence should be addressed to (include email address) 23
Lucas C. Parra, Department of Biomedical Engineering, City College of New York, 160 24
Convent Ave., New York, NY 10031, USA. Phone: +1-212-650-8653; Email: 25
parra@ccny.cuny.edu 26
6. Number of Figures: 9 27
7. Number of Tables: 0 28
8. Number of Multimedia: 0 29
9. Number of words for Abstract: 137 30
10. Number of words for Significance Statement: 58 31
11. Number of words for Introduction: 702 32
12. Number of words for Discussion: 1,817 33
13. Acknowledgements: We thank Stefan Haufe for suggesting code to compute ISC using 34
symmetrized between- and within-subject covariances. 35
14. Conflict of Interest: Authors report no conflict of interest 36
15. Funding sources: This work was supported by the Defense Advanced Research 37
Projects Agency (Contract W911NF-14-1-0157). 38
39
2
Abstract 40
Neural development is generally marked by an increase in the efficiency and diversity of neural 41
processes. In a large sample (N = 114) of human children and adults with ages ranging from 5 - 44 42
years, we investigated the neural responses to naturalistic video stimuli. Videos from both real-life 43
classroom settings and Hollywood feature films were used to probe different aspects of attention and 44
engagement. For all stimuli, older ages were marked by more variable neural responses. Variability 45
was assessed by the inter-subject correlation of evoked electroencephalographic (EEG) responses. 46
Young males also had less variable responses than young females. These results were replicated in 47
an independent cohort (N = 303). When interpreted in the context of neural maturation, we conclude 48
that neural function becomes more variable with maturity, at least during the passive viewing of real-49
world stimuli. 50
Significance Statement 51
Naturalistic videos were used to probe response variability with EEG in a large developmental cohort. 52
Our results are consistent with developmental theories positing that neural variability increases with 53
maturation, and that neural maturation typically occurs earlier in females. These results differ from 54
those observed with fMRI, where an increase in stereotyped responses with age is observed during 55
development. 56
Keywords 57
Inter-subject correlation, electroencephalography, naturalistic stimuli, evoked responses, development 58
Introduction 59
This study examines the relationship between the variability of neural responses and 60
development. Over the course of development, the accuracy and stability of behaviors generally 61
increase. This performance improvement is typically accompanied by a seemingly paradoxical 62
increase in the variability of neural responses both within and across subjects (Grady, 2012; Dinstein 63
et al., 2015). More variable electroencephalographic (EEG) responses across trials, characterized by 64
3
an increase in dimensionality and entropy, are associated with lower reaction time variability and 65
higher recognition accuracy (Mcintosh et al., 2008). Neural variability often presents as an increase in 66
the complexity of neural responses. This may be due in part to a developmental increase in the 67
repertoire of possible brain states (Vakorin et al., 2013) and this increase in complexity may underlie 68
the integration between distributed neural populations (Vakorin et al., 2011). EEG signal complexity 69
becomes elevated in late adolescence and is also elevated in females relative to males at this stage, 70
indicating that females may attain mature brain functioning prior to males (Anokhin et al., 2000). 71
Anatomical studies generally support the notion that females reach neural maturity prior to males 72
(Giedd et al., 1999; Lenroot and Giedd, 2006; Lenroot et al., 2007; Marsh et al., 2008). 73
Neural variability does not always accompany proficient behavior, however. Both theta band 74
coherence, a performance monitoring measure, and behavior are more variable across trials in 75
children (Papenberg et al., 2013). This suggests that neural variability does not always increase with 76
maturation. For adults, the variability in “functional connectivity” between different networks measured 77
with fMRI is elevated during rest and decreases during a cognitive task. The reverse is true for 78
children, whose brains become more variable during the task, and their performance is expectedly 79
lower than adults (Hutchison and Morton, 2015). 80
Recently, responses to naturalistic narrative stimuli have been used to examine how variability 81
in behavior and neural activity change with development. In these cases, variability is measured 82
across subjects rather than within individuals because it is expected that if an individual has a less 83
variable neural response across repeated renditions of stimulus, their neural response will also be 84
more similar to others who are responding to the same stimulus. Adults watch Sesame Street more 85
similarly to each other than infants do, as assessed by where their eyes fixate (Kirkorian et al., 2012). 86
Additionally, adults have more broadly similar neural responses to Sesame Street than children 87
(Cantlon and Li, 2013). While the neural responses of adults to Sesame Street correlate more with 88
each other in many parietal and frontal regions, children correlate more strongly with each other in a 89
4
specific region in the superior temporal cortex (Cantlon and Li, 2013). Generally, from ages 18-88, as 90
humans age, responses to videos increase in variability (Campbell et al., 2015). Taken together, these 91
studies demonstrate that neural variability changes with age. The nature of this relationship depends 92
on multiple factors including the metric of neural variability, the developmental stage sampled, and the 93
brain region(s) of interest. 94
Here, EEG was recorded from subjects with ages ranging from 5 - 44 years as they were 95
presented with both naturalistic (Dmochowski et al., 2012, 2014) and conventional stimuli. To assess 96
neural variability, the level of similarity across subjects was assessed with the inter-subject correlation 97
(ISC) of responses evoked by the stimuli. ISC of the EEG is indicative of attention, engagement, and 98
memory in healthy adults (Dmochowski et al., 2014; Cohen and Parra, 2016; Ki et al., 2016; Cohen et 99
al., 2017) . We found that neural responses, indexed by ISC, become more variable with age. Among 100
children, females have more variable neural responses than males. This increase in variability is not 101
due to a decrease in evoked response magnitude, and was reproduced in two independent cohorts 102
consisting of 114 and 303 individuals. These results are consistent with theories positing that 103
development coincides with an increased repertoire of neural representations (Mcintosh et al., 2008), 104
and the sex differences are consistent with the idea that young males are less neurally mature than 105
young females (Giedd et al., 1999; Lenroot and Giedd, 2006; Lenroot et al., 2007; Marsh et al., 2008). 106
Importantly, this is the first EEG study to report a measure of across-subject neural similarity with clear 107
age and sex effects. 108
Methods 109
Subjects 110
In the main study ages ranged from 6 to 44 years old (N = 114, 14.2 +/- 8.0 years old, 46 females, see 111
Figure 1A for a full age and sex distribution) as part of the Child Mind Institute - Multimodal Resource 112
for Studying Information Processing in the Developing Brain (MIPDB; 113
http://fcon_1000.projects.nitrc.org/indi/cmi_eeg/; (Langer et al., 2017)). In the replication study ages 114
5
ranged from 5 to 21 years old (N = 303, 11.3 +/- 3.9 years old, 135 females, see figure 1B for a full 115
age and sex distribution). This data was obtained from the Child Mind Institute Healthy Brain Network 116
(CMI-HBN; http://fcon_1000.projects.nitrc.org/indi/cmi_healthy_brain_network/; (Alexander et al., 117
2017)). Both the main and replication study data come from publically available datasets. All 118
experiments were performed in accordance with relevant guidelines and regulations. The study was 119
reviewed and approved by the Chesapeake Institutional Review Board. All subjects presented with 120
normal or corrected to normal vision. 121
Stimuli 122
Engaging, naturalistic videos were the primary stimuli. Specific videos were selected because they 123
contained content relevant to social cognition, classroom anxiety, and attention. Three videos featured 124
either educational content or depicted classroom scenarios: Fun with Fractals (Fract, MIT), a cartoon 125
that explains fractals with examples (4m 34s), How to improve at Simple Arithmetic (Arith, E-How), in 126
which a math teacher in a typical educational setting explains addition and multiplication (1m 30s), 127
and Pre-Algebra Class (StudT, Pearson Education)), showing an interaction between two students 128
and a teacher (StudT, for student-teacher interaction) during math problem solving (1m 40s). Two 129
videos were clips from conventional cinema: Diary of a Wimpy Kid (Wimpy, Universal Pictures), a 130
movie about a preteen starting middle school (1m 57s), and Despicable Me (DesMe, Universal 131
Pictures), which contains infant and toddler characters and emphasizes social interactions (2m 51s). 132
While the main cohort contains data from all stimuli, the replication cohort only had three stimuli: 133
Wimpy, Fract, and DesMe. The variability of the neural responses to these videos was measured 134
across subjects using the inter-subject correlation (ISC) of evoked responses (see below). As a 135
control condition, a “Rest” condition, during which subjects sat with their eyes-closed for 4m 20s, was 136
also analyzed. This period establishes the baseline level of ISC, as no time-aligned stimulus 137
entraining neural activity across subjects was presented. Finally, “Flash”, a stimulus condition without 138
any narrative content was used. During this stimulus a black and white grating pattern that flashed at 139
6
25 Hz was presented for three minutes, thus synchronously stimulating neural activity across subjects 140
(see Steady State Visual Evoked Potentials (SSVEP) Methods section). This stimulus elicits steady 141
state evoked potentials (Vanegas et al., 2015) and was included to explore the extent to which ISC is 142
driven by low level evoked responses. 143
Procedure 144
While seated in a dimly lit room wearing an EEG net, subjects watched a series of short videos in a 145
pseudorandom order. Stimuli were presented on a 17-inch CRT monitor (SONY Trinitron Multiscan 146
G220, display dimensions 330×240 mm, resolution 800×600 pixels, vertical refresh rate of 100 Hz). 147
Note that some subjects did not experience all stimuli due to time limitations (Langer et al., 2017). 148
Additionally, as explained below, poor data quality for some recordings caused additional data loss. 149
For the replication study, only three conditions were used: Wimpy, Fract and DesMe. 150
EEG recordings and preprocessing 151
EEG recordings were performed with an EGI Clinical Geodesic 128 channel system (Electrical 152
Geodesic Inc, Eugene, OR). Of the 128 channels recorded, 105 constituted the EEG recording and 11 153
represented EOG channels used for eye movement artifact removal. The remaining channels, mainly 154
recording from the neck and face, were discarded. First, noisy channels were selected by visual 155
inspection and replaced with by zero valued samples, thus eliminating those channels’ contribution in 156
subsequent calculations of covariance matrices. Recordings, initially at 500 Hz, were then 157
downsampled to 125 Hz, high-pass filtered at 1 Hz, and notch filtered between 59 and 61 Hz with a 158
4th-order Butterworth filter. Eye artifacts were removed by linearly regressing the EOG channels from 159
the scalp EEG channels (Parra et al., 2005). Next, a robust Principal Components Analysis (PCA) 160
algorithm, the inexact Augmented Lagrange Multipliers Method (Lin et al., 2013), was used to remove 161
sparse outliers from the data following Ki et al (2016). Briefly, robust PCA recovers a low-rank matrix, 162
A, from a corrupted data matrix D = A + E, where some entries of the additive errors E may be 163
arbitrarily large. Finally, individual recordings for some stimuli were discarded on the basis of visual 164
7
inspection because they remained noisy after both automatic and manual noise removal. This was 165
necessary because these subjects exhibited profound movement artifacts and/or the saline used for 166
the recordings dried out. Despite these steps taken, the data overall appear to be of poorer quality 167
than that collected in an electrically shielded room with conductive gel (saline was used here). The 168
noise in the data may have led to the relatively low ISC values reported in the paper compared to 169
previous studies (Ki et al, 2016; Cohen et al, 2017). However, it is unlikely that the noise contributed to 170
our results as under baseline conditions (Rest), there was no difference in power between the cohorts 171
(see Results). All signal processing was performed offline using MATLAB software (MathWorks, 172
Natick, MA, USA). 173
Inter-Subject Correlation (ISC) 174
As variability is the inverse of similarity, we measured the similarity of evoked EEG responses across 175
subjects. This approach has been used extensively to study concerted, inter-subject changes in blood-176
oxygen level dependent (BOLD) signal in fMRI (Hasson et al., 2004, 2009; Kauppi et al., 2010), and 177
has been adapted to leverage the improved time resolution facilitated by EEG. To determine the 178
neural similarity across subjects responding to the same stimulus (or in the same condition, in the 179
case of Rest) the inter-subject correlation (ISC) of the EEG signal was computed, as described 180
previously (Dmochowski et al., 2012, 2014; Cohen and Parra, 2016; Ki et al., 2016). ISC assesses the 181
level of correlation in the EEG across time among a group of subjects as they respond to the same 182
stimulus. Larger ISC values imply more similarity in fast EEG responses across subjects (< 1s). This 183
indicates that the signals are more reliable due to decreased inter-subject variability. It has also been 184
found that subjects who pay more attention to the stimulus have higher ISC values (Ki et al., 2016). An 185
advantage of the technique is that the stimulus need only be presented once to each subject because 186
evoked responses are compared across individuals. As repeated trials are unnecessary, responses 187
are more similar to natural situations in which people experience uniquely presented novel stimuli. 188
Additionally, in contrast to event related potentials, the technique can be applied to continuous and 189
8
dynamic natural stimuli without the need for specific event markers (Ben-Yakov et al., 2012). As such, 190
the approach is data-driven both spatially and temporally. The approach is “data driven” spatially 191
because the data from the subjects determines the best combination of electrodes (which are spatially 192
distributed across the scalp and therefore may correspond with different anatomical regions) that 193
maximize the correlation across subjects. The approach is “data driven” temporally because the 194
ultimate correlation values are determined by the temporal fluctuations in the EEG signals. In contrast, 195
a more traditional approach to EEG data analysis would be to select electrodes that have previously 196
been shown to elicit a certain effect (or ERP) and measure event-locked responses from these 197
electrodes. We are not taking this approach. Rather, the electrodes that we chose and the time 198
periods that maximize correlation are determined directly by the data itself. 199
ISC utilizes correlated component analysis to identify linear combinations of EEG electrodes 200
that capture most of the correlation across subjects (Dmochowski et al., 2012). Correlated component 201
analysis is similar to principal component analysis (PCA) except that rather than maximizing variance 202
within one dataset, it selects projections,࢜ࣕࡾࡰ, where ܦ is the number of electrodes, that maximize 203
the correlation between multiple datasets. These projections can be thought of as virtual sensors (or 204
component sources) of activity that are optimized to capture most of the correlation between subjects. 205
They are the eigenvectors of ܴௐ
ିଵܴ. Where ܴௐ is the average within-subject covariance: 1
ேσܴ, 206
and ܴ is the average between subjects cross-covariance: 1
ேሺேି1ሻσσ ܴǡஷ, and ܴ ൌσሺ
ݔሺݐሻെ
௧
207
ݔҧሻሺݔሺݐሻെݔҧሻ் measures the cross-covariance of all electrodes in subject ݇ with all electrodes in 208
subject ݈. Vector ݔሺݐሻ is the scalp voltages at time ݐ in subject ݇ and ݔҧ is their mean value in time. 209
Following previous research, we use the three components, or eigenvalues ofܴௐ
ିଵܴ, that 210
represent the largest fraction of the correlation across subjects. These components can be optimized 211
for all subjects together, or for a subset of the entire cohort. The subsets used in this paper are 212
stimulus (Wimpy, DesMe, Fract, Arith, StudT, Flash, and Rest), age group (young vs old), sex (male 213
9
vs female), and sex and age group combined (young-male, young-female, old-male, and old-female). 214
ISC components are computed within subsets of the entire sample to examine potential differences in 215
the spatial distribution of activity across groups, although the spatial patterns are largely consistent 216
(Figure 7). 217
To calculate the ISC for individual subjects as they respond to the same condition as their 218
peers, the correlation between each individual’s EEG responses and the responses from all other 219
individuals is calculated (Cohen and Parra, 2016; Ki et al., 2016). The ISC values reported throughout 220
the paper are this measure of how well each individual correlates with the others. The projections, 221
࢜ࣕࡾࡰ, used to compute this subject-specific ISC value are either computed across all subjects or 222
within the subgroups listed above (divided by either stimulus, age, sex, or age and sex). The ISC for 223
each subject is therefore: 224
ܥ ൌ௩
ோ್ǡೖ௩
௩
ோೢǡೖ௩
, 225
where ܴǡ ൌ1
ሺேି1ሻσሺܴ ܴ
ሻ
ǡஷ , and ܴ௪ǡ ൌ1
ሺேି1ሻσሺܴ ܴ
ሻ
ǡஷ . ISC for subject ݇ is therefore 226
σܥ
3
ୀ1. A simplified template for the code to compute the correlated components and the ISC for 227
individual subjects is available at http://www.parralab.org/isc/ 228
Steady state visual evoked potentials (SSVEPs) 229
To determine the strength of low-level sensory evoked responses across individuals, we leveraged the 230
steady state visual evoked potential (SSVEP) paradigm (Flash) that was part of the data collection 231
effort (Langer et al., 2017). Stimulus and analysis followed established techniques (Vanegas et al., 232
2015). Briefly, the stimulus consisted of four circular ‘foreground’ stimuli (vertical grating, radius 2°) 233
that were flickered on-and-off at 25 Hz and embedded in a static ‘background’ grating, which is known 234
to generate reliable SSVEPs (Vanegas et al., 2015). This stimulus was presented in trials of 2.4 s 235
duration with inter-trial intervals of 1s which included a fixation cross presented for 0.5 s. The stimuli 236
were presented in several conditions that varied in their contrast and in the phase relationship 237
10
between the foreground and the background. A total of 128 trials were present (12 conditions total: 238
four foreground contrasts - 0% 30%, 60% and 100%, and three background conditions - parallel 239
phase, orthogonal phase, and no surround stimuli). Artifacts were rejected by removing trials for which 240
the power (or absolute value) of any electrode exceeded more than three standard deviations above 241
the mean. EOG activity was regressed out of the EEG, as described above. The initial 200ms of each 242
trial was removed to eliminate the onset of the visual evoked response. Data were Fourier 243
transformed for each trial, power in a 0.5 Hz bin surrounding the 25 Hz band was extracted, and then 244
averaged across all trials, regardless of condition (thus ignoring details of the foreground-background 245
interaction). Since the EEG activity measured with this paradigm is known to be dominated by primary 246
visual cortex (V1) responses, power was averaged over the five most relevant occipital electrodes 247
(O1-O5; (Vanegas et al., 2015)). 248
Dimensionality of EEG Responses 249
To gain a sense of the dimensionality of the EEG responses across subjects, the eigenvalue spectrum 250
was extracted from each subject’s covariance matrix (covariance between all electrodes measured 251
across time). These covariance matrices measure the correlation between electrodes for each 252
subject. The sum of the eigenvalues represent the overall power in the data. To assess the 253
dimensionality of the data, lines were fit to the loglog plot of the eigenvalue spectrum of each subject’s 254
covariance matrix. A shallower slope of the linear fit indicates that the there is appreciable power over 255
a larger number of dimensions. Two-way ANOVAs and subsequent post-hoc t-tests were employed to 256
compare power and the slopes of these linear fits for each age and sex group both across all stimuli 257
and within each stimulus. 258
Results 259
We sought to determine whether and how the variability of EEG differs across age and gender in 260
children and adults ranging from 6 – 44 years of age. To assess the variability in EEG signals across 261
subjects, the intersubject correlation (ISC) between individuals and their peers was assessed in 262
11
response to both naturalistic videos and artificial stimuli. ISC can be thought of as a measure of the 263
similarity of neural responses (Dmochowski et al., 2012). If subjects respond more similarly to their 264
peers, they will have a larger ISC value, which indicates that they have a less variable neural 265
response. 266
Intersubject correlation varies between stimuli 267
ISC is a stimulus-driven measure of attention (Ki et al., 2016) because neural responses are more 268
correlated across subjects when they naturally attend to a stimulus than when they are engaged in a 269
dual task. It is therefore expected to be indicative of varying levels of engagement (Cohen et al., 270
2017). A one-way ANOVA determined that ISC significantly depended on the stimulus (F(7) = 78.26, p 271
= 10-68; mean +/- STD ISC values: Wimpy: 0.053 +/- 0.036; DesMe: 0.035 +/- 0.023; Arith: 0.019 +/- 272
0.013; Fract: 0.026 +/- 0.016; StudT: 0.012 +/- 0.009; Flash: 0.030 +/- 0.019; Rest: 0.001 +/- 0.004), 273
indicating that the stimuli significantly varied in engagement level (Cohen et al., 2017). It is worth 274
noting that these ISC values are relatively low compared to previous research (Cohen and Parra, 275
2016; Ki et al., 2016). There are two factors that contribute to this discrepancy: the lower production 276
quality and therefore engagement level elicited by these stimuli and the relatively poor quality of the 277
EEG data (see Methods). Note also that ISC for EEG is generally lower than ISC of fMRI (e.g. 278
Lahnakoski et al., 2017) which has a slower time scale and higher signal-to-noise ratio, both factors 279
that can contribute to higher correlations (Haufe et al., 2017). As expected, ISC in the Rest condition 280
was not significantly different from zero (t-test, t(45) = 0.52, p = 0.4), confirming the notion that ISC 281
reflects stimulus-induced correlations (Dmochowski et al., 2012). A one-way ANOVA was therefore 282
performed on all stimuli excluding Rest, confirming that ISC strongly varies between stimuli (F(6) = 283
71.70, p = 10-55). Tukey post-hoc pairwise comparisons revealed that ISC was significantly stronger 284
when evoked by the qualitatively more engaging stimuli (Wimpy and DesMe), than it was for 285
educational videos (Arith, Fract, StudT; Tukey post-hoc pair-wise comparisons between each pair of 286
videos, Tukey’s HSD: p < 10-4). Among the more engaging videos from conventional cinema, Wimpy, 287
12
a movie trailer for the feature film “Diary of a Wimpy Kid”, evoked a higher level of neural similarity 288
than DesMe, a scene from the animated film “Despicable Me” (Tukey’s HSD: p = 10-7). Among the 289
relatively less-engaging educational videos, Fract elicited the highest level of ISC, which was 290
significantly higher than StudT (Tukey’s HSD: p = 10-6), but not Arith (Tukey’s HSD: p = 0.2). 291
Interestingly, Arith elicited a level of ISC similar to Flash (Tukey’s HSD: p = 0.5), and the level of ISC 292
elicited by Flash was significantly higher than StudT (Tukey’s HSD: p=10-7). 293
Intersubject correlation decreases with age 294
We hypothesized that neural similarity changes with age and therefore examined the correlation 295
between ISC and age. Here, ISC is computed in individuals by measuring the extent to which each 296
subject correlated with the other people in the same stimulus condition. For all of the stimuli excluding 297
Rest, there was a negative relationship between age and ISC (all r’s=-0.68 +/- 0.09, all p’s<10-10, FDR 298
corrected following Benjamini and Hochberg (1995), Figure 2). ISC did not vary with age during Rest 299
(r = -0.10, p = 0.5, N = 46). This was expected since Rest contained no stimulus to drive EEG signal 300
similarly across subjects. 301
These results indicate that ISC decreases with age. However, most of the subjects in the main 302
study were from the lower half of the age distribution (see Figure 1A). Since the components used to 303
measure ISC are optimized to capture the correlation across all subjects, the components may have 304
been biased by these younger subjects who constituted a majority of the sample. The cohort was 305
therefore divided into two age groups of equal size to eliminate this potential measurement bias. The 306
median split resulted in groups whose ages ranged from 6-14 (mean age 10.74 +/- 2.03) and 15-44 307
(mean age 23.65 +/- 8.04). The ISC was then recomputed from components extracted separately in 308
each group. A two-way ANOVA with factors of age and stimulus revealed that ISC was significantly 309
modulated by both stimulus (all excluding Rest, F(5, 393) = 63.64, p = 10-47) and age (F(1, 393) = 310
335.46, p = 10-53, Figure 3A). For all stimuli, ISC was much higher in the younger age group. 311
Intersubject correlation is elevated in males 312
13
Sex is an important factor that influences the developmental trajectory of the human brain (Giedd et 313
al., 1999; Lenroot and Giedd, 2006; Lenroot et al., 2007; Marsh et al., 2008). We therefore explored 314
the relationship between sex and ISC. A two-way ANOVA with factors of sex and stimulus (excluding 315
Rest) revealed main effects for both sex (F(1, 393)=53.11, p = 10-12) and stimulus (F(5, 393) = 30.12, 316
p = 10-26, Figure 3B). Tukey’s post hoc tests revealed that ISC was consistently higher in males for all 317
stimuli except for Flash where it was marginally significant (Flash: p = 0.06; Wimpy: p = 0.03; DesMe: 318
p = 10-6; Arith: p = 0.003; Fract: p = 10-4; StudT: p = 10-4). To examine whether the sex difference 319
depended on age, the data was separated into four groups with the same age division between 14 320
and 15 years as above (young-male, young-female, old-male, old-female). ISC was measured within 321
each group and averaged across all stimuli available for each subject to ensure sufficiently large 322
sample sizes (excluding control conditions - Flash and Rest, Figure 4). A two-way ANOVA with sex 323
and age as factors confirmed the age effect (F(1, 87) =98.85, p = 10-16), and the sex effect was 324
marginally significant (F(1, 87) = 3.83, p = 0.05). A direct comparison between the sexes in each age 325
group revealed that the sex effect was marginally significant among the young ages (t(53)=2.02, 326
p=0.05, 6-14 years), but not present for the old ages (t(33)=0.28, p=0.8, 15-44 years). 327
The effect of age on inter-subject correlation is not due to evoked response difference 328
The relationships between ISC, age, and sex may be partially driven by the reduction of evoked 329
response magnitude with age (Goodin et al., 1978; Tomé et al., 2015). Although correlation, which 330
ISC measures, is theoretically independent of magnitude, it is possible that a decrease in magnitude 331
corresponds with a decrease in the signal-to-noise ratio, which would result in a smaller ISC. The 332
magnitude of evoked responses was therefore assessed with the Flash stimulus which elicited steady-333
state visually evoked potentials (SSVEPs, see Methods). SSVEP magnitude weakly declines with age 334
(r=-0.22, p=0.02, N=109, Figure 5A) and a two-way ANOVA with age and sex as factors (the same 335
age/sex groups as Figure 4) found the age effect to be marginally significant (F(1,106)=4.00, p=0.05, 336
14
Figure 5B). There was no significant relationship between sex and SSVEP strength (F(1,106)=3.3, 337
p=0.08). 338
Since both SSVEP amplitude and ISC decrease with age, we reasoned that SSVEPs could be 339
used to factor out the effect of evoked response strength (Goodin et al., 1978; Tomé et al., 2015). 340
Indeed, ISC and SSVEP amplitude are correlated across subjects (r = 0.41, p = 0.0001, N = 84, 341
Figure 6A). To control for the effect of evoked response strength, each individual’s SSVEP amplitude 342
was linearly regressed against ISC, and the portion that could be explained by the SSVEP was 343
subtracted (ISC calculated within the same age/sex group as Figure 4). A two-way ANOVA with age 344
and sex as factors revealed that this residual ISC still significantly varies with age (F(1,81)=85.49, p = 345
10-14), but does not vary with sex (F(1,81)=0.08, p = 0.8, Figure 6B). Additionally, the sex effect is no 346
longer present in the younger group when SSVEP strength is controlled for (t(49)=0.11, p=0.9). The 347
lack of a sex effect may mean that the relationship between sex and neural variability is due in part to 348
evoked response magnitude, but the lack of an effect may also result from the reduced number of 349
subjects for which SSVEP magnitude was available: 84 vs 114. Regardless, neural variability, as 350
assessed by ISC, does increase with age, regardless of the strength of evoked responses. 351
Correlated component topographies similar across age and sex groups. 352
ISC was measured using components of the EEG that maximize correlations between subjects. These 353
components are linear combinations of electrodes and can be thought of as virtual sensors (See 354
Methods). To determine if the spatial distribution of the corresponding activity differed across groups, 355
the “forward model,” which represents how the components look on the surface of the scalp, was 356
computed for the largest three components which were used to compute ISC (Parra et al., 2005). 357
These component topographies were very similar across all age/sex groups for the strongest two 358
components - C1 and C2 (Figure 7, minimum cosine similarity was 0.97 for C1 and 0.78 for C2). The 359
third component (C3) was less similar across the groups (cosine similarity ranged from 0.89 to 0.31), 360
but it also constituted a much weaker portion of the ISC (C1 = 0.016 +/-0.009, C2 =0.008 +/-0.005, 361
15
and C3 =0.004 +/- 0.003, computed as in Figure 4 and averaged across all subjects and stimuli). 362
Thus, for the most part, differences in ISC between age and sex groups were not due to differences in 363
the spatial distribution of neural activity across these groups. 364
Dimensionality of EEG Responses 365
To determine whether the differences in ISC across groups was due to diverse responses across 366
subjects or to more highly dimensional responses within subjects, the eigenvalue spectra of the EEG 367
covariance matrices were analyzed (Figure 8). The sum of these spectra represents the overall power 368
of the data. In general, the younger age group (using the same median split as above) had more 369
power than the older age group across all stimuli (F(1,438) = 452.13, p = 10-69). This suggests that 370
there was more overall power in the EEG of the young group. This power difference was only present 371
during the stimuli, not during rest (t(44) = 0.6, p = 0.5), suggesting that younger subjects have stronger 372
stimulus driven evoked responses. To assess the dimensionality of the EEG responses, a linear 373
model was fit to each subject’s eigenvalue spectrum (see Methods), and the slopes were compared 374
between the groups. A difference in dimensionality is reflected by a difference in this slope, with a 375
shallower slope indicating that there is a higher number of dimensions with appreciable signal. The 376
slopes of the linear fit did not differ across the age groups (across all stimuli: F(1,438) = 2.74, p = 0.1). 377
This suggests that the stimulus evoked responses are not inherently higher dimensional in the young. 378
For the sex comparison, males had higher overall power (across all stimuli: F(1,438) = 71.25, p = 10-
379
16), for all stimuli and not during rest (t(44) = 0.9, p = 0.3). Here females had a shallower slope than 380
males (across all stimuli: F(1,438) = 152.12, p = 10-30). This suggests a greater complexity of 381
responses within females. 382
Replication of results 383
To confirm these findings, the results were replicated in an independent cohort (N=303) with a 384
reduced stimulus set: Wimpy, Fract, and DesMe. Replicating the results above, ISC also decreased 385
with age in this cohort (Wimpy: r = -0.44, p = 10-14, N = 276; Fract: r = -0.37, p = 10-10, N = 270; 386
16
DesMe: r = -0.41, p = 10-12, n=281, Figure 9A). A two-way ANOVA with age and condition as factors 387
revealed that ISC is modulated by age (F(1,799) = 35.33, p = 10-9) and stimulus (F(2,799) = 272.903, 388
p = 10-91, Figure 9B). A two-way ANOVA with sex and stimulus as factors revealed that ISC was also 389
significantly modulated by sex (F(1,823) = 11.12, p = 0.0009, Figure 9C), and stimulus (F(2,823) = 390
430.95, p = 10-129). Finally, a two-way ANOVA which divided the data across age and sex groups, and 391
averaged ISC across stimuli, replicated the main effect of age (F(1,291) = 17.68, p = 10-5), and did not 392
find an effect of sex (F(1,291) = 2.59, p = 0.1, Figure 9D). Follow up analyses that examined a 393
potential sex difference in ISC in each age group revealed that the difference in ISC was present 394
among the young ages (t(224)=2.29, p=0.02, 5-14 years), but not the old ages (t(67)=0.59, p=0.6, 15-395
21 years). When the median was calculated according to the median of the replication distribution 396
(split at 10/11 years, see Figure 1B for age distribution), the above results were unchanged. In 397
summary, all results from the main experiment replicated in this independent cohort. 398
Discussion 399
The present work demonstrated age- and sex- related variability among individuals with respect to 400
their neural responses to complex naturalistic stimuli. Specifically, ISC was significantly correlated with 401
age for both naturalistic videos and artificial visual flashes. Younger subjects (6-14 years) exhibited 402
less variable neural responses than older subjects (15-44 years). A parallel finding revealed that 403
young males exhibited more similar responses to the stimuli than young females, a difference which 404
was only present in the younger cohort. These age and sex effects may result from neural 405
development, consistent with the notion that neural maturation occurs later in males than in females 406
(Lenroot et al., 2007; Marsh et al., 2008; Mous et al., 2017). A quantitative analysis of the spatial 407
distribution of the correlated activity revealed that the observed age and sex differences are largely 408
driven by the same neural components, lending more weight to the idea that the observed differences 409
in age and sex stem from a common developmental feature. Finally, a replication study with 303 410
participants yielded similar results. 411
17
A possible confound for the present results is that the neural correlations found across 412
subjects are due to correlations in overt behaviors such as eye movements. However, it is unlikely that 413
eye movements follow the same developmental trajectory as neural responses because eye 414
movement trajectories evoked by videos actually become more similar with age (Kirkorian et al., 415
2012). Thus, although the gaze patterns evoked by videos seem to converge with maturity, potentially 416
driving similar bottom-up neural processes, neural similarity as measured by ISC, decreases with age. 417
This indicates that patterns of neural activity may potentially increase in their diversity with age as top-418
down factors relating to the interpretation of naturalistic stimuli develop. Even in the condition where 419
subjects were instructed to maintain a fixed gaze position (Flash), ISC decreased with age. Future 420
studies with fine-grained eye-tracking during EEG could more definitively answer this question. 421
The observed ISC magnitude changes with age and sex may also be partially dependent on 422
evoked response magnitudes which typically decrease with age. While the amplitudes of auditory 423
event related potentials and their components decline with age (Goodin et al., 1978; Tomé et al., 424
2015), other components increase with age (Dinteren et al., 2014), or remain stable across 425
development (Kujawa et al., 2013). Although correlation, as measured by ISC, is in principle 426
insensitive to magnitude, it is possible that weaker stimulus evoked responses in adults may be 427
overpowered by non-stimulus related neural activity (i.e., “noise”) (Hammerer et al., 2013). In this 428
case, a smaller fraction of the signal would correlate across adults in comparison to children. To 429
control for the effect of age, the magnitude of steady state visual evoked potentials (SSVEPs) was 430
regressed from the ISC. The result indicates that SSVEP amplitude cannot explain the age effect, but 431
it may explain the sex effect, indicating that males have stronger evoked responses than females 432
(Figure 3B and 6B). However, it is worth noting that ISC and SSVEPs measure very different facets of 433
neural activity. SSVEPs, extracted from early visual processing areas in V1, likely represent low-level 434
visual processes. ISC, on the other hand, may be driven by higher level cortical areas since the spatial 435
distributions of the two dominant components (Figure 7) do not resemble low-level sensory evoked 436
18
responses. Parallel work indicates that the first component (C1), which captures the majority of the 437
correlated activity, is a supramodal component that is driven by both auditory and visual stimuli 438
(Cohen and Parra, 2016). 439
It is also possible that ISC decreases with age because adults process the world with more 440
diverse brain activity. In this view, adults have more highly variable stimulus-evoked responses and 441
their neural activity is therefore less similar across subjects. In this case, it would be likely that the 442
dimensionality of neural responses, a measure of their complexity, increases with age (Anokhin et al., 443
2000; Mcintosh et al., 2008; Vakorin et al., 2013). There was no clear trend, indicative of a difference 444
in the dimensionality between the young and old group. However, it does appear that females have 445
more diverse responses than males, a result that deserves further exploration and could possibly 446
underlie the reduction in ISC in this group. 447
The present results appear to be consistent with Campbell et al. (2015), who using fMRI also 448
found a decrease of ISC with age. However, while we study an age range dominated by development 449
and corresponding improvements in cognitive performance (between 6-44 and 5-23 years in each 450
cohort), Campbell et al. (2015) examined a range (18-88 years) that exhibited a deterioration in fluid 451
intelligence and reaction time. These measures correlated with a decrease in ISC. While Cantlon and 452
Li (2013) studied a cohort that was more comparable to ours in age (4 to 25 years), they find that ISC 453
of fMRI was generally higher among adults (above age 18) than it was in children (below age 11). In 454
total, it appears that ISC as assessed by fMRI increases with development and declines in older age, 455
which potentially opposes our result with EEG. These differences may be due to important 456
methodological discrepancies between these studies and ours. To more definitively establish the 457
effect of age on ISC more work should be done using both fMRI and EEG. 458
The idea that maturity is marked by variability is not new (Campbell et al., 2015). It aligns with 459
theories from neural systems modeling and human studies (Mcintosh et al., 2008; Vakorin et al., 460
2011). In these models, moderate amounts of noise or variability facilitate efficient responses in 461
19
complex environments. Increased variability may be the reason for reduced evoked response 462
magnitudes since event related potentials are obtained by averaging across many events that are 463
inherently sensitive to signal noise. It is therefore possible that the increased variability of evoked 464
responses across trials with age results in reduced ERP magnitudes. 465
In the age range examined, neural development is a dynamic process. At the macro level, 466
longitudinal structural neuroimaging shows that cortical thinning occurs from childhood through early 467
adulthood, progressing in a caudal to rostral pattern (Gogtay et al., 2004; Giedd et al., 2015). At the 468
micro level, synaptic pruning and myelination, particularly in the frontal lobe, are ongoing during this 469
period (Rakic et al., 1994; Huttenlocher and Dabholkar, 1997; Cox et al., 2016). From a functional 470
perspective, studies of functional connectivity and task-based fMRI suggest that functional maturation 471
tends to follow a “diffuse to focal pattern” (Durston et al., 2006; Grill-Spector et al., 2008; Fair et al., 472
2009; Kelly et al., 2009), and may correspond to the extraordinary advances in behavior during 473
childhood (Xiao et al., 2016). Speculatively, the decreased ISC strength in older ages may reflect 474
greater inter-individual variability that results from the interplay of structural and functional 475
“streamlining” of neural architecture with distinct life experiences (e.g. cortical thinning, synaptic 476
pruning and diffuse-to-focal shifts in functional patterns). However, one limitation of the present study 477
is that it is cross-sectional rather than longitudinal, it is therefore difficult to make developmental 478
claims based on the age-based differences demonstrated here (Kraemer et al., 2000). 479
The age-related effect, may also be echoed by the sex difference in neural variability. 480
Longitudinal studies have demonstrated that females mature prior to males in a range of anatomical 481
measures (Lenroot et al., 2007; Lim et al., 2015). However, differences in developmental trajectories 482
between males and females may be complicated by the fact that the sexes ultimately differ in their 483
mature neuroanatomy (Marsh et al., 2008). Here, sex-related differences in neural variability were only 484
seen among younger subjects, suggesting that this is a development-related difference. Prepubescent 485
and early teenage years are marked by sex differences in behavioral maturity that may not be present 486
20
in later years (Mous et al., 2017). The difference in neural variability may also be due to pubertal stage 487
since it is known that females reach pubertal maturity 2-3 years prior to males (Sisk and Foster, 2004). 488
However, physiological pubertal stage was not measured here, and it is therefore not possible to 489
determine whether the sex differences were related to this factor. 490
Among the different stimuli used, the clips from conventional cinema (Wimpy and DesMe) 491
evoked a higher level of ISC than the educational videos (Arith, Fract, and StudT). The Hollywood 492
clips were rich with scene cuts and dynamic visual cues and are therefore expected to elicit strong 493
levels of ISC (Poulsen et al., 2017). However, previous research has also shown that engagement 494
with narrative stimuli modulates ISC, and it is therefore likely that these Hollywood clips are more 495
effective at engaging attention and thus elicit stronger ISC (Dmochowski et al., 2014; Ki et al., 2016; 496
Cohen et al., 2017). Although the ISC differences between age and sex may be due to each cohort’s 497
average level of attention, no independent measures of engagement or attention were collected. It is 498
therefore not possible to determine whether the present effects are driven by attention or differences 499
in low-level stimulus features. Most of the videos were aimed at younger audiences (i.e., Despicable 500
Me, Diary of a Wimpy Kid), and older subjects may have therefore been less interested in them. 501
However, this was not uniformly the case, for instance, the video about Fractals (Fract) may have 502
been equally interesting to both children and adults, while the Flash stimulus may be equally boring for 503
all ages. Thus, these two stimuli provided an important control for attentional effects on the age-504
related differences in ISC. Future work may benefit from looking at objective measures of engagement 505
(Cohen et al., 2017) in the different cohorts studied here. An understanding of such factors, and their 506
impact on behavior, may be of relevance to models of media-based addiction (e.g., internet addiction, 507
pornography addiction), as well as commercial neuroscience enterprises. Regardless, it is of note that 508
the age effect seen for the naturalistic videos was echoed in the SSVEP condition. Since this stimulus 509
should be equally (un)engaging for all ages, this favors an interpretation based on neural maturation 510
rather than attention. 511
21
Future work should recruit a larger sample of subjects above age 15 to determine whether the 512
age-related decline in ISC observed in later teenage years continues in adulthood, or might even 513
reverse in older age (Grady, 2012; Campbell et al., 2015). Future studies with clinical cohorts could 514
explore the potential link between ISC and behavioral markers of neural development. It is possible 515
that neural variability is not only a marker of maturity, but it is also a marker of neuropsychiatric 516
disorders (Dinstein et al., 2015). The methods used here provide a novel way of assessing such 517
markers under complex, naturalistic conditions. 518
Overall, the current results regarding intersubject correlation in children and adults are 519
interpreted in the context of neural maturation. Although males are delayed in the development of the 520
neural variability that appears to be a mark of maturity, the data presented here indicate that with 521
normal development they are no different than females as adults. Thus, with maturity, neural function 522
becomes more variable. 523
References 524
Alexander LM et al. (2017) The Healthy Brain Network Biobank: An open resource for transdiagnostic 525
research in pediatric mental health and learning disorders. bioRxiv. 526
Anokhin AP, Lutzenberger W, Nikolaev A, Birbaumer N (2000) Complexity of electrocortical dynamics 527
in children: Developmental aspects. Dev Psychobiol 36:9–22. 528
Ben-Yakov A, Honey CJ, Lerner Y, Hasson U (2012) Loss of reliable temporal structure in event-529
related averaging of naturalistic stimuli. Neuroimage 63:501–506 Available at: 530
http://linkinghub.elsevier.com/retrieve/pii/S1053811912007185. 531
Benjamini Y, Hochberg Y (1995) Controlling the false discovery rate: a practical and powerful 532
approach to multiple testing. J R Stat Soc Ser B 57:289–300. 533
Benninger C, Matthis P, Scheffner D (1984) EEG development of healthy boys and girls. Results of a 534
longitudinal study. Electroencephalogr Clin Neurophysiol 57:1–12. 535
Campbell KL et al. (2015) Idiosyncratic responding during movie-watching predicted by age 536
22
differences in attentional control. Neurobiol Aging 36:3045–3055. 537
Cantlon JF, Li R (2013) Neural Activity during Natural Viewing of Sesame Street Statistically Predicts 538
Test Scores in Early Childhood. PLoS Biol 11. 539
Clarke AR, Barry RJ, McCarthy R, Selikowitz M (2001) Age and sex effects in the EEG: Differences in 540
two subtypes of attention-deficit/hyperactivity disorder. Clin Neurophysiol 112:815–826. 541
Cohen SS, Henin S, Parra LC (2017) Engaging narratives evoke similar neural activity and lead to 542
similar time perception. Sci Rep 7:1–10. 543
Cohen SS, Parra LC (2016) Memorable audiovisual narratives synchronize sensory and supramodal 544
neural responses. eNeuro Available at: http://eneuro.sfn.org/cgi/doi/10.1523/ENEURO.0203-545
16.2016. 546
Cox SR, Ritchie SJ, Tucker-Drob EM, Liewald DC, Hagenaars SP, Davies G, Wardlaw JM, Gale CR, 547
Bastin ME, Deary I. J (2016) Age differences in brain white matter microstructure in UK Biobank 548
(N = 3,513). bioRxiv 51771:1–34 Available at: http://www.oasisdiscussions.ca/wp-549
content/uploads/2015/07/References-Reported-side-effects-of-intravenous-midazolam-sedation-550
when-used-in-pediatric-dentistry-a-review.pdf. 551
Dinstein I, Heeger DJ, Behrmann M (2015) Neural variability: Friend or foe? Trends Cogn Sci 19:322–552
328 Available at: http://dx.doi.org/10.1016/j.tics.2015.04.005. 553
Dinteren R, Arns M, Jongsma MLA, Kessels RPC (2014) P300 development across the lifespan: A 554
systematic review and meta-analysis. PLoS One 9. 555
Dmochowski JP, Bezdek MA, Abelson BP, Johnson JS, Schumacher EH, Parra LC (2014) Audience 556
preferences are predicted by temporal reliability of neural processing. Nat Commun 5:1–9. 557
Dmochowski JP, Sajda P, Dias J, Parra LC (2012) Correlated components of ongoing EEG point to 558
emotionally laden attention – a possible marker of engagement? Front Hum Neurosci 6:1–9. 559
Durston S, Davidson MC, Tottenham N, Galvan A, Spicer J, Fossella JA, Casey BJ (2006) A shift from 560
diffuse to focal cortical activity with development. Dev Sci 9:1–20. 561
23
Fair DA, Cohen AL, Power JD, Dosenbach NUF, Church JA, Miezin FM, Schlaggar BL, Petersen SE 562
(2009) Functional brain networks develop from a “local to distributed” organization. PLoS Comput 563
Biol 5:14–23. 564
Gasser T, Verleger R, Petra B, Sroka L (1988) Development of the E E G of school-age children and 565
adolescents. I. Analysis of band power. Electroencephalogr Clin Neurophysiol 69:91–99. 566
Giedd JN, Blumenthal J, Jeffries NO, Castellanos FX, Liu H, Zijdenbos A, Paus T, Evans AC, 567
Rapoport JL (1999) Brain development during childhood and adolescence: a longitudinal MRI 568
study. Nat Neurosci 2:861–863. 569
Giedd JN, Raznahan A, Alexander-Bloch A, Schmitt E, Gogtay N, Rapoport JL (2015) Child 570
Psychiatry Branch of the National Institute of Mental Health Longitudinal Structural Magnetic 571
Resonance Imaging Study of Human Brain Development. Neuropsychopharmacology 40:43–49 572
Available at: http://www.nature.com/doifinder/10.1038/npp.2014.236. 573
Gogtay N, Giedd JN, Lusk L, Hayashi KM, Greenstein D, Vaituzis a C, Nugent TF, Herman DH, 574
Clasen LS, Toga AW, Rapoport JL, Thompson PM (2004) Dynamic mapping of human cortical 575
development during childhood through early adulthood. Proc Natl Acad Sci U S A 101:8174–576
8179 Available at: http://www.ncbi.nlm.nih.gov/pubmed/15148381. 577
Goodin DS, Squires KC, Henderson BH, Starr A (1978) Age-related variations in evoked potentials to 578
auditory stimuli in normal human subjects. Electroencephalogr Clin Neurophysiol 44:447–458. 579
Grady C (2012) The cognitive neuroscience of ageing. 13. 580
Grill-Spector K, Golarai G, Gabrieli J (2008) Developmental neuroimaging of the human ventral visual 581
cortex. Trends Cogn Sci 12:152–162. 582
Hammerer D, Li SC, Volkle M, Muller V, Lindenberger U (2013) A lifespan comparison of the 583
reliability, test-retest stability, and signal-to-noise ratio of event-related potentials assessed during 584
performance monitoring. Psychophysiology 50:111–123. 585
Harmony T, Marosi E, A. D de L, Becker J, Fernández T (1990) Effect of sex, psychosocial 586
24
disadvantages and biological risk factors on EEG maturation. Electroencephalogr Clin 587
Neurophysiol 75:482–491. 588
Hasson U, Malach R, Heeger DJ (2009) Reliability of cortical activity during natural stimulation. Trends 589
Cogn Sci 14:40–48 Available at: http://linkinghub.elsevier.com/retrieve/pii/S1364661309002393. 590
Hasson U, Nir Y, Levy I, Fuhrmann G, Malach R (2004) Intersubject synchronization of cortical activity 591
during natural vision. Science 303:1634–1640. 592
Haufe, S., DeGuzman, P., Henin, S., Arcaro, M., Honey, C. J., Hasson, U., Parra, L. C. (2017) 593
Elucidating relations between fMRI, ECoG and EEG through a common natural stimulus. bioRxiv 594
207456. Available at: https://doi.org/10.1101/207456 595
Hutchison RM, Morton JB (2015) Tracking the Brain ’ s Functional Coupling Dynamics over 596
Development. 35:6849–6859. 597
Huttenlocher PR, Dabholkar AS (1997) Regional differences in synaptogenesis in human cerebral 598
cortex. J Comp Neurol 387:167–178 Available at: papers3://publication/uuid/B4EE9A33-90C1-599
4EE3-BB6A-804EF096B3D0. 600
Kauppi J-P, Jaaskelainen IP, Sams M, Tohka J (2010) Inter-subject correlation of brain hemodynamic 601
responses during watching a movie: localization in space and frequency. Front Neuroinform 4:1–602
10 Available at: http://journal.frontiersin.org/article/10.3389/fninf.2010.00005/abstract. 603
Kelly AMC, Di Martino A, Uddin LQ, Shehzad Z, Gee DG, Reiss PT, Margulies DS, Castellanos FX, 604
Milham MP (2009) Development of anterior cingulate functional connectivity from late childhood 605
to early adulthood. Cereb Cortex 19:640–657 Available at: 606
http://www.ncbi.nlm.nih.gov/pubmed/18653667. 607
Ki J, Kelly S, Parra LC (2016) Attention strongly modulates reliability of neural responses to 608
naturalistic narrative stimuli. J Neurosci. 609
Kirkorian HL, Anderson DR, Keen R (2012) Age Differences in Online Processing of Video : An Eye 610
Movement Study. 83:497–507. 611
25
Kraemer HC, Yesavage JA, Taylor JL, Kupfer D (2000) How can we learn about developmental 612
processes from cross-sectional studies, or can we? Am J Psychiatry 157:163–171. 613
Kujawa A, Klein DN, Proudfit GH (2013) Two-year stability of the late positive potential across middle 614
childhood and adolescence. Biol Psychol 94:290–296 Available at: 615
http://dx.doi.org/10.1016/j.biopsycho.2013.07.002. 616
Lahnakoski, J. M., Jääskeläinen, I. P., Sams, M., & Nummenmaa, L. (2017). Neural mechanisms for 617
integrating consecutive and interleaved natural events. Human Brain Mapping, 38(7), 3360-3376. 618
Langer N, Ho EJ, Alexander LM, Xu HY, Jozanovic RK, Henin S, Petroni A, Cohen S, Marcelle ET, 619
Parra LC, Milham MP, Kelly SP (2017) A resource for assessing information processing in the 620
developing brain using EEG and eye tracking. Sci data 4:170040 Available at: 621
http://www.ncbi.nlm.nih.gov/pubmed/28398357%0Ahttp://dx.doi.org/10.1038/sdata.2017.40. 622
Lenroot RK, Giedd JN (2006) Brain development in children and adolescents : Insights from 623
anatomical magnetic resonance imaging. 30:718–729. 624
Lenroot RK, Gogtay N, Greenstein DK, Wells EM, Wallace GL, Clasen LS, Blumenthal JD, Lerch J, 625
Alex P, Evans AC, Thompson PM, Giedd JN (2007) Sexual Dimorphism of Brain Developmental 626
Trajectories during Childhood and Adolescence. Neuroimage 36:1065–1073. 627
Lim S, Han CE, Uhlhaas PJ, Kaiser M (2015) Preferential detachment during human brain 628
development: Age- and sex-specific structural connectivity in diffusion tensor imaging (DTI) data. 629
Cereb Cortex 25:1477–1489. 630
Lin Z, Chen M, Ma Y (2013) The Augmented Lagrange Multiplier Method for Exact Recovery of 631
Corrupted Low-Rank Matrices. arXiv Available at: 632
http://arxiv.org/abs/1009.5055%0Ahttp://dx.doi.org/10.1016/j.jsb.2012.10.010. 633
Liu Z, Woltering S, Lewis MD (2014) Developmental change in EEG theta activity in the medial 634
prefrontal cortex during response control. Neuroimage 85:873–887 Available at: 635
http://dx.doi.org/10.1016/j.neuroimage.2013.08.054. 636
26
Marsh R, Gerber AJ, Peterson BS (2008) Neuroimaging Studies of Normal Brain Development and 637
Their Relevance for Understanding Childhood Neuropsychiatric Disorders. J Am Acad Child 638
Adolesc Psychiatry 47:1233–1251. 639
Mcintosh AR, Kovacevic N, Itier RJ (2008) Increased Brain Signal Variability Accompanies Lower 640
Behavioral Variability in Development. 4. 641
Mous SE et al. (2017) Applied Neuropsychology : Child The association of gender , age , and 642
intelligence with neuropsychological functioning in young typically developing children : The 643
Generation R study. Appl Neuropsychol Child 6:22–40 Available at: 644
http://dx.doi.org/10.1080/21622965.2015.1067214. 645
Papenberg G, Hammerer D, Muller V, Lindenberger U, Li SC (2013) Lower theta inter-trial phase 646
coherence during performance monitoring is related to higher reaction time variability: A lifespan 647
study. Neuroimage 83:912–920 Available at: http://dx.doi.org/10.1016/j.neuroimage.2013.07.032. 648
Parra LC, Spence CD, Gerson AD, Sajda P (2005) Recipes for the linear analysis of EEG. 649
Neuroimage 28:326–341 Available at: 650
http://linkinghub.elsevier.com/retrieve/pii/S1053811905003381. 651
Poulsen AT, Kamronn S, Dmochowski J, Parra LC, Hansen LK (2017) Measuring engagement in a 652
classroom : Synchronised neural recordings during a video presentation. Sci Rep 7:1–9 Available 653
at: http://dx.doi.org/10.1038/srep43916. 654
Rakic P, Bourgeois JP, Goldman-Rakic PS (1994) Synaptic development of the cerebral cortex: 655
implications for learning, memory, and mental illness. Prog Brain Res 102:227–243. 656
Sisk CL, Foster DL (2004) The neural basis of puberty and adolescence. Nat Neurosci 7:1040–1047 657
Available at: http://www.nature.com/doifinder/10.1038/nn1326. 658
Tomé D, Barbosa F, Nowak K, Marques-Teixeira J (2015) The development of the N1 and N2 659
components in auditory oddball paradigms: a systematic review with narrative analysis and 660
suggested normative values. J Neural Transm 122:375–391. 661
27
Uhlhaas PJ, Roux F, Singer W, Haenschel C, Sireteanu R, Rodriguez E (2009) The development of 662
neural synchrony reflects late maturation and restructuring of functional networks in humans. 663
Proc Natl Acad Sci 106:9866–9871 Available at: 664
http://www.pnas.org/cgi/doi/10.1073/pnas.0900390106. 665
Vakorin VA, Lippe S, McIntosh AR (2011) Variability of Brain Signals Processed Locally Transforms 666
into Higher Connectivity with Brain Development. J Neurosci 31:6405–6413 Available at: 667
http://www.jneurosci.org/cgi/doi/10.1523/JNEUROSCI.3153-10.2011. 668
Vakorin VA, McIntosh AR, Mišić B, Krakovska O, Poulsen C, Martinu K, Paus T (2013) Exploring Age-669
Related Changes in Dynamical Non-Stationarity in Electroencephalographic Signals during Early 670
Adolescence. PLoS One 8. 671
Vanegas MI, Blangero A, Kelly SP (2015) Electrophysiological indices of surround suppression in 672
humans. J Neurophysiol 113:1100–1109 Available at: 673
http://jn.physiology.org/lookup/doi/10.1152/jn.00774.2014. 674
Xiao Y, Friederici AD, Margulies DS, Brauer J (2016) Longitudinal changes in resting-state fMRI from 675
age 5 to age 6years covary with language development. Neuroimage 128:116–124 Available at: 676
http://dx.doi.org/10.1016/j.neuroimage.2015.12.008. 677
Figure Legends 678
Figure 1: Age and sex distributions for the main study (A) and replication study (B). A: Subjects in the 679
main study (N=114) had data for all of the stimuli. B: Subjects in the replication study (N=303) had 680
only three stimuli (Wimpy, Fract, DesMe) and contribute to the results in Figure 9. 681
Figure 2: Neural similarity, measured as the intersubject correlation (ISC) of neural activity, decreased 682
with age. Correlation values ranged from r=-0.58 to r=-0.78, indicating a consistent relationship 683
between maturity and neural variability. ISC was computed for each individual by correlating neural 684
responses from individual subjects with the neural responses from all other subjects for that stimulus 685
(regardless of age and sex). 686
28
Figure 3: ISC, a measure of neural similarity, was consistently higher among younger ages and males. 687
A: Across all stimuli, ISC was higher for younger subjects (6-14 years, light green) than older subjects 688
(15-44 years, dark green). B: Across all stimuli, ISC was higher for males (blue) than females (red). 689
For both A and B, ISC was computed separately within each age and sex group. Black lines indicate 690
the median. 691
Figure 4. Sex differences in the young do not exist in the old. Young males were more neurally similar 692
to each other than young females. This sex difference is absent in the older group. Here, ISC was 693
computed within each sex and age group separately and averaged across all stimuli except for Flash 694
and Rest. Black lines indicate the median. 695
Figure 5: Steady state visual evoked potential (SSVEP) magnitude depended on age, but not on sex. 696
A: SSVEP strength was weakly correlated with age across subjects, but it was no different between 697
males and females. B: SSVEP strength was no different between males and females. Black lines 698
indicate the median. 699
Figure 6: Relationship between ISC magnitude and SSVEP strength. A: SSVEP strength, a measure 700
of the magnitude of evoked responses, correlated with ISC strength, calculated using all stimuli except 701
for Flash and Rest. B: Comparison of ISC strength after SSVEP magnitude was regressed out (ISC - 702
SSVEP) between males and females in the young age and old age groups. While there was a 703
significant difference between the age groups, a difference between the sex groups was not present. 704
Black line indicates the median. 705
Figure 7: Spatial distributions corresponding to the three strongest components of intersubject 706
correlation (ISC: C1 - C3). Red and blue colors indicate positive and negative correlation of the 707
voltages on the scalp surface with the component activity. These maps are unitless due to an arbitrary 708
scale on the projection vectors. Here, the projections have been computed separately for the 709
combination of the two sex and age groups. As the scalp topographies were relatively consistent 710
29
across the groups, the differences in ISC across these groups was not due to differences in the spatial 711
topography of correlation within the group. 712
Figure 8: Eigenvalue spectra of the average covariances for each demographic group. Eigenvalues 713
measure the power of the signal in principal components of the EEG (correlated across time for each 714
stimulus). Each curve is the average eigenvalue spectrum for each group averaged across all stimuli 715
and subjects. A: Young subjects have more power than Old subjects in all dimensions. This is 716
represented by the upward shift in their average eigenvalue spectrum. B: The eigenvalue spectrum of 717
Females has a shallower slope than that for Males indicating that they have a more diverse set of 718
neural responses. 719
Figure 9: The results from the main study replicated in an independent cohort (N=303). A: ISC 720
decreased with age in the replication cohort. ISC was computed for each individual by correlating 721
responses from individual subjects to those from all other subjects (regardless of age and sex) for that 722
stimulus. Correlation values ranged from r=-0.37 to r=-0.44. Note that for every stimulus a different 723
number of subjects was available. B: Across all stimuli, ISC was higher for younger subjects (6-14 724
years, light green) than it was for older subjects (15-44 years, dark green) in the replication cohort. For 725
consistency, the split between the ages was consistent between this study and the main study C: 726
Across all stimuli, ISC was higher for males (blue) than females (red) in the replication cohort. For 727
both B and C, ISC was computed separately within each age and sex group. Black lines indicate the 728
median. D: Sex differences in the young disappeared with age in the replication cohort. Young males 729
were more neurally similar to each other than young females, and this sex difference was absent in 730
the older group. Here, ISC was computed within each sex and age group separately and averaged 731
across all stimuli used in the replication cohort. Black lines indicate the median. 732
5 1015202530354045
2
6
10
14
Number of subjects
Age
male
female
5 1015202530354045
10
20
30
40
Number of subjects
Age
A
B
Intersubject Correlation (ISC)
Age [years]
10 20 30 40
0
0.01
0.02
0.03
0.04
0.05
0.06 r=-0.76
p=7x10-13
N=65
10 20 30 40
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035 r=-0.68
p=2x10-10
N=68
10 20 30 40
0
0.01
0.02
10 20 30 40
0
0.02
0.04
0.06
0.08
0.1 r=-0.76
p=1x10-11
N=58
10 20 30 40
0
0.01
0.02
0.03
0.04
r=-0.77
p=4x10-13
N=66
0.005
0.015
0.025
0.03 r=-0.66
p=7x10-10
N=69
10 20 30 40
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07 r=-0.71
p=4x10-12
N=74
Wimpy DesMe Arith
Fract StudT Flash
ISC
Young
Old
ISC
0.14
0
0.02
0.04
0.06
0.08
0.1
0.12
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14 Male
Female
Wimpy
DesMe
Arith
Fract
StudT
Flash
Rest
Wimpy
DesMe
Arith
Fract
StudT
Flash
Rest
AB
ISC
0.09
0.01
0.03
0.05
0.07
Male FemaleMaleFemale
Young Old
[mean age = 10.74] [mean age = 23.65]
10 20 30 40
Age
0
1
2
3
4
5
SSVEP strength
Male
Female
Male Female Male Female
0
1
2
3
4
5
SSVEP strength
Young Old
BA
1234
SSVEP
0.02
0.04
0.06
0.08
ISC
Male Female Male Female
Young
-0.02
0
0.02
0.04
ISC - SSVEP
BA
Old
young-
males
old-
females
old-
males
young-
females
C1 C2 C3
10 010 2
log(Eigenvalue rank)
10 2
10 4
10 6
log(Eigenvalues)
Young
Old
10 010 2
log(Eigenvalue rank)
10 2
10 4
10 6
log(Eigenvalues)
Male
Female
BA
Wimp
y
DesMe
Fract
ISC
Wimp
y
DesMe
Fract
Male
Female
0.14
0.1
0.06
0.02
Young
Old
0.14
0.1
0.06
0.02
ISC
ISC
Age [years]
Male
Young
FemaleMaleFemale
Old
ISC
0.1
0.06
0.02
A
BCD
Wimpy
r=-0.44
p=1x10-14
N=276
5101520
0.02
0.06
0.1
0.14
Fract
r=-0.37
p=3x10-10
N=270
5101520
0.01
0.02
0.03
0.04
DesMe
r=-0.41
p=2x10-12
N=281
5101520
0.02
0.04
0.06
0.08
