Resting-State Functional Connectivity of the Amygdala in Autism: A Preregistered Large-Scale Study

Abstract

Objective:

Three leading neurobiological hypotheses about autism spectrum disorder (ASD) propose underconnectivity between brain regions, atypical function of the amygdala, and generally higher variability between individuals with ASD than between neurotypical individuals. Past work has often failed to generalize, because of small sample sizes, unquantified data quality, and analytic flexibility. This study addressed these limitations while testing the above three hypotheses, applied to amygdala functional connectivity.

Methods:

In a comprehensive preregistered study, the three hypotheses were tested in a subset (N=488 after exclusions; N=212 with ASD) of the Autism Brain Imaging Data Exchange data sets. The authors analyzed resting-state functional connectivity (FC) from functional MRI data from two anatomically defined amygdala subdivisions, in three hypotheses with respect to magnitude, pattern similarity, and variability, across different anatomical scales ranging from whole brain to specific regions and networks.

Results:

A Bayesian approach to hypothesis evaluation produced inconsistent evidence in ASD for atypical amygdala FC magnitude, strong evidence that the multivariate pattern of FC was typical, and no consistent evidence of increased interindividual variability in FC. The results strongly depended on analytic choices, including preprocessing pipeline for the neuroimaging data, anatomical specificity, and subject exclusions.

Conclusions:

A preregistered set of analyses found no reliable evidence for atypical functional connectivity of the amygdala in autism, contrary to leading hypotheses. Future studies should test an expanded set of hypotheses across multiple processing pipelines, collect deeper data per individual, and include a greater diversity of participants to ensure robust generalizability of findings on amygdala FC in ASD.

Despite enormous effort and interest, the neural basis of autism spectrum disorder (ASD) remains poorly understood. One long-standing hypothesis of autism has argued that functional connectivity (FC) between multiple brain regions is reduced, although both under- and over-connectivity have been reported (e.g., 1). A second long-standing hypothesis of autism as a disorder of social cognition has argued for atypical structure and/or function of the amygdala in autism (2). A third hypothesis is that autism features high variability within and across individuals (3), although this may be due to multiple subtypes of ASD as usually diagnosed.

These three leading neurobiological hypotheses have generally been tested piecemeal, often with unclear data quality (e.g., lack of signal-to-noise ratio [SNR] assessment in amygdala subregions and motion effects in resting-state functional MRI [rs-fMRI] data), resulting in inconsistent findings. For instance, some studies have reported amygdala hypoconnectivity (4, 5) and others hyperconnectivity (e.g., 6) or even just typical connectivity (7). Most studies have not used cytoarchitectonically distinct amygdala subdivisions, or have used approximate parcellations in a standard template brain space, limiting anatomical precision (e.g., 8–10). Amygdala SNR in blood-oxygen-level-dependent (BOLD) fMRI can suffer from severe signal dropout, but this is rarely explicitly assessed. Many studies have been based on small sample sizes (typically less than 50 per group), often from different data collection sites (11, 12). A final and related challenge is that rs-fMRI studies are particularly sensitive to analytic flexibility (13); the choice of processing pipeline is well known to have major effects on the reliability of group effects in ASD (14). These several challenges most likely interact, and they have precluded a clear answer to the main hypotheses about atypical brain function in ASD.

To test the three hypotheses more comprehensively, we placed a strong emphasis on the reliability and generalizability of our preregistered study. We investigated amygdala FC by assessing correlations between time series of pairs of brain regions (15) using rs-fMRI in a large available sample.

Different strategies for defining amygdala subregions have been introduced, some based on FC (16–18), others on anatomy (19, 20). We focused on two anatomically defined amygdala subdivisions—the basolateral amygdala and the cortico-centromedial amygdala, defined by the amygdala atlas in FreeSurfer (20). We tested group differences in FC magnitude (hypothesis 1), FC pattern similarity (hypothesis 2), and FC variability (hypothesis 3) between participants with ASD and neurotypical comparison participants. To further establish the robustness of our findings, all three hypotheses were tested across ranges of anatomical specificity of connectivity targets (including the whole brain, specific regions, and a previously implicated large-scale functional brain network (the default mode network [DMN] [21]), using two different fMRI denoising pipelines and three types of amygdala segmentation (the above-mentioned anatomical one, which was the focus of our study, as well as a functional one and the whole amygdala for comparison). Preregistered follow-up analyses further explored associations of amygdala FC with social functioning. The majority of analyses took a Bayesian approach to quantify support in favor of and against the null hypothesis, with careful control of data quality and covariates.

METHODS

A preregistered analysis plan on existing data aggregated in the Autism Brain Imaging Data Exchange (ABIDE) I and II data sets (22, 23), including descriptions of methodological details and minor deviations, can be found at https://osf.io/cvqmb and in the online supplement. We provide a summary here.

Sample Information

The ABIDE data sets had been collected across multiple sites (22, 23) and approved by local institutional review boards. We included all available fMRI data sets published in Brain Imaging Data Structure (BIDS) format (24) as of May 2020 (http://fcon_1000.projects.nitrc.org/indi/s3/index.html), accessed using DataLad (https://github.com/datalad/datasets.datalad.org). We included participants with a diagnosis of ASD as well as participants who had no psychiatric illness (typically developed comparison subjects). Participants had to be between 16 and 50 years of age and have an estimated full-scale IQ (FSIQ) >70. We excluded participants with diagnoses of pervasive developmental disorder not otherwise specified as well as those with comorbid psychosis, epilepsy, or genetic developmental diseases other than ASD. The aim of these criteria was the selection of a more homogeneous sample with superior fMRI quality, in order to increase power to detect hypothesized effects. ASD classification was determined by clinical judgment and/or gold-standard diagnostic instruments (the Autism Diagnostic Observation Schedule [ADOS] [25]; the Autism Diagnostic Interview–Revised [26]). (For site-specific details, please refer to http://fcon_1000.projects.nitrc.org/indi/abide/.)

Data Quality Exclusion Criteria

We excluded participants who had in-scanner head motion with framewise displacement (FD) >0.2 mm at >40% of time points, had poor T₁ segmentation quality, or had low BOLD SNR in the amygdala (Figure 1).

FIGURE 1.

Participant inclusion and exclusion in a resting-state functional connectivity study of the amygdala in autism^a

^aABIDE=Autism Brain Imaging Data Exchange; ASD=autism spectrum disorder; FSIQ=full-scale IQ; PDD-NOS=pervasive developmental disorder not otherwise specified; QC=quality control.

MRI Data

Each individual had one structural MRI sequence available (T₁-weighted, e.g., three-dimensional magnetization-prepared rapid acquisition gradient-echo or a vendor-specific variant) and a resting-state fMRI sequence (echo planar imaging, T₂*-weighted; see references 22, 23 for details). One nuisance regressor for each ABIDE data collection site was added to account for all site-specific variance (e.g., scanner type or sequence) in all analyses.

MRI Data Processing

After preprocessing with fMRIPrep, version 20.2.6 (27), functional data were further processed to reduce non-BOLD-related signal (e.g., physiological noise) with rsDenoise (28; https://github.com/adolphslab/rsDenoise/releases/tag/ama_ release). The denoising strategy employed in the main analyses resembled that used in reference 29 (pipeline A). To test the effect of the denoising strategy on the results, we applied a second denoising pipeline (30) (pipeline B). See the preregistration for details on pilot analyses on independent data that justified our choices of denoising strategies.

Definition of amygdala subdivisions.

To increase anatomical specificity, and motivated by the known functional differences between amygdala subregions, amygdala nuclei were defined in each individual participant with FreeSurfer (20, 31) and aggregated into two previously described subdivisions (32, 33): the basolateral amygdala (BLA, including the lateral, basal, accessory basal, and paralaminar nuclei) and the cortico-centromedial amygdala (CCM, including the medial, central, cortical, and cortico-amygdaloid transition area) in subject (native) space. Note that left and right amygdala regions of interest (ROIs) were pooled into bilateral amygdala ROIs for all analyses. To explore generalizability over segmentation strategy, we additionally preregistered to use template-space functional subdivisions (MNI152, 2-mm resolution [16]; dorsal, medial, and ventrolateral), as well as the whole amygdala (not divided into subdivisions) as defined by FreeSurfer in subject space.

Cortical parcellations and subcortical segmentations.

Cortical data were parcellated into 400 parcels across both hemi-spheres (200 per hemisphere) and further summarized into seven intrinsic cortical networks, including the DMN (34, 35) (the parcellation is not homotopic and thus considered per hemisphere), in fsaverage6 space. Eight bilateral subcortical regions (thalamus, caudate, putamen, pallidum, hippocampus, nucleus accumbens, ventral diencephalon, brainstem) were defined using FreeSurfer (31, 36, 37) in subject space and by the Harvard-Oxford subcortical atlas in MNI152 template space.

Functional connectivity.

FC was quantified as the product-moment correlation of time series responses between two regions of the brain (38). We first extracted time-series data per voxel or vertex (depending on the parcellation scheme) for each subject and then averaged across an entire ROI. Amygdala ROI time series were then correlated with time series from 400 cortical and eight bilateral subcortical ROIs, yielding a functional connectivity vector (“amygdala connectome”) indicating the strength of temporal correlation between an amygdala ROI and each of the target regions. Correlation coefficients were Fisher r-to-z transformed to normalize their distribution. For a subset of analyses, we normalized FC with respect to each subject’s brain-wide FC to increase sensitivity and decrease the effect of variability in global FC strengths across individuals.

Metrics of amygdala FC.

To estimate overall strength of amygdala FC to the rest of the brain (hypothesis 1), we calculated the median (to account for potentially skewed distributions) magnitude of amygdala ROI correlations per subject and for each group (ASD and neurotypical).

To estimate the similarity of amygdala FC patterns across subjects (hypothesis 2), we calculated the average amygdala FC pattern per group (ASD or neurotypical) and then tested the similarity of the group-specific average patterns with Kendall’s tau correlation: the vector containing the amygdala’s connectivity to each of the other brain regions from one group was correlated with the analogous vector from the other group. The magnitude of correlation between the group-specific amygdala FC vectors indicates the evidence against hypothesis 2 (larger correlation indicates greater evidence for similarity, i.e., no difference between groups), or evidence for the null hypothesis (smaller or no correlation indicates greater evidence for dissimilarity, i.e., differences between groups).

Because of the short fMRI runs, there were insufficient data available to assess within-subject reliability of FC. Instead, to compare the between-subject variability of amygdala FC patterns in the two groups (hypothesis 3), we first calculated, for each subject, the average between-subject similarity in FC with all other subjects within a group, and then tested whether this distribution of subject-wise similarity differed between groups, using a permutation test.

Statistical Inferences

We used Bayesian inference testing in JASP, version 16.03 (https://jasp-stats.org/), to quantify support both for and against the null hypothesis. Our motivation for using Bayesian inference rather than the more common significance testing was to provide a more informative conclusion regarding the strength of support for our hypotheses than dichotomous rejection or not of the null hypothesis (39). In short, FC strength (hypothesis 1) was assessed with Bayesian analyses of covariance, FC pattern similarity (hypothesis 2) was assessed with Bayesian correlations, and FC pattern variability (hypothesis 3) was assessed with permutation testing. For interpreting the Bayes factor, a simplified heuristic is provided in the footnote to Table 1 (40, 41).

TABLE 1.

Bayes factors for model comparison between the null model and a model including group, testing for group differences in functional connectivity magnitude^a

		Whole Brain		Top 10%
Region of Interest	Pipeline	Raw	Normalized	Raw	Normalized
Basolateral amygdala	A	6.972	2.247	0.112	0.660
	B	0.095	0.103	0.153	0.501
Cortico-centromedial amygdala	A	0.156	0.156	0.078	0.213
	B	0.397	0.643	0.114	0.160
Medial amygdala	A	0.769	0.244	0.410	0.120
Dorsal amygdala		0.124	0.155	0.124	0.209
Ventrolateral amygdala		0.693	0.303	0.134	0.566
Whole amygdala		7.508	2.446	0.155	0.855

^aThe table lists Bayes factors (BF₁₀) for model comparison between the null model and a model including group (autism and neurotypical), testing for group differences in functional connectivity (FC) magnitude (raw or normalized by subject-wise whole brain FC magnitude) between amygdala regions of interest (ROIs) (anatomical: basolateral, cortico-centromedial, and whole amygdala; functional: medial, dorsal, and ventrolateral amygdala) and the whole brain or the top 10% strongest connected parcels defined by the neurotypical group, for two different denoising pipelines (pipeline A [29] and pipeline B [30]). Pipeline B was applied only to the main analysis of the anatomically defined amygdala ROIs (basolateral and cortico-centromedial), to test for robustness of the main results to variation in neuroimaging data preprocessing. As a simple rubric, values >1 suggest evidence in favor of hypothesis 1 (difference between the ASD and neurotypical groups, alternative hypothesis), and values <1 suggest evidence in favor of the null hypothesis (no difference between the ASD and neurotypical groups). The following is a simplified scale for interpreting the Bayes factor: values <0.1 would be strong evidence, 0.1–0.33 moderate evidence, and 0.33–1 anecdotal evidence toward the null hypothesis; a value of 1 would be evidence toward neither the null hypothesis nor the alternative hypothesis; values of 1–3 would be anecdotal evidence, 3–10 moderate evidence, and >10 strong evidence toward the alternative hypothesis (40, 41).

Nuisance Variables

Nuisance variables of no interest for the purpose of our study (sex, age, mean FD, percentage of time points with FD >0.2 mm, ABIDE data collection site, FSIQ) were included in the models as outlined in each specific hypothesis and mean-centered before applied to a regression (hypothesis 2 and 3).

RESULTS

Sample Information

After exclusions (see Figure 1), 488 participants (212 with ASD [22 of them female] and 276 neurotypical comparison subjects [44 of them female]) were included in the analyses. The groups did not differ significantly in age (W=26,588.0, p=0.084, rank-biserial correlation=−0.091) or biological sex (p=0.075, χ²=3.175), but the ASD group showed a significantly lower FSIQ compared with the neurotypical group (W=21,152.5, p<0.001, rank-biserial correlation=−0.277) (Figure 2; see also Table S1 in the online supplement). The groups showed similar levels of head motion (mean FD: W=31,172.0, p=0.215, rank-biserial correlation=0.065; percentage FD >0.2 mm: W=30,839.5, p=0.305, rank-biserial correlation=0.054).

FIGURE 2.

Age, full-scale IQ, and in-scanner head motion of participants in a resting-state functional connectivity study of the amygdala in autism^a

^a Raincloud plots combine a cloud of points (each representing one subject) with a box plot (median, 25% and 75% quartiles, interquartile range, and maximum and minimum values) and a one-sided violin plot per group. Percent with FD >0.2 mm refers to percentage of time points with FD >0.2 mm in the resting-state fMRI run. ASD=autism spectrum disorder; FSIQ=full-scale IQ; FD=framewise displacement.

FC Results

No robust evidence for underconnectivity in ASD.

Hypothesis 1 tested whether the ASD group had a lower median magnitude of FC than the neurotypical group. First, a Bayesian analysis of covariance on the raw FC values found moderate evidence for a groupwise magnitude difference across the whole brain for the BLA (Bayes factor [BF₁₀]=6.972; ASD group: mean=0.137, SD=0.109, 95% CI=0.122, 0.151; neurotypical group: mean=0.166, SD=0.117, 95% CI=0.152, 0.180) (see Table 1 and Figure 3; see also Table S4 in the online supplement). In other words, data were almost seven times more likely to occur under the model including the effect of diagnosis compared to the model without. For the CCM, by contrast, we found moderate evidence for equivalent (no difference in) FC magnitude between the two groups (BF₁₀=0.156; ASD group: mean=0.111, SD=0.105, 95% CI=0.096, 0.125; neurotypical group: mean=0.118, SD=0.113, 95% CI=0.105, 0.131).

FIGURE 3.

Functional connectivity of anatomical amygdala subregions (BLA and CCM)^a

^aIn panel A, the strength of functional connectivity (FC) (z, magnitude; hypothesis 1) is displayed on 400 cortical parcels on an inflated cortex (fsaverage6 space), and eight subcortical regions per group in the bar representations. In panel B, the same plots are shown for the top 10% most strongly connected regions, as identified by the neurotypical group. In panel C, subject median FC magnitude is displayed in raincloud plots per amygdala subregion (BLA and CCM) for the whole brain (top), normalized FC with respect to each subject’s average FC across all pairs of brain regions (except those including the amygdala) (middle), and for the top 10% regions (bottom). In panel D, amygdala connectome pattern (hypothesis 2) is displayed as FC strength across seven large-scale functional networks and subcortical regions per group (ASD in red, neurotypical in black; shaded areas represent confidence intervals). ASD=autism spectrum disorder; BF₁₀=Bayes factor; BLA=basolateral amygdala; CCM=cortico-centromedial amygdala; Cont=control; DorsAttn=dorsal attention; LH=left hemisphere; NT=neurotypical; RH=right hemisphere; SalVentAttn=salience/ventral attention; SomMot=somatosensory motor; Subc=subcortical; ventral DC=ventral diencephalon; Vis=visual; z=strength of connectivity (Fisher r-to-z transformed).

Second, we normalized FC with respect to each subject’s average FC across all pairs of brain regions (except those including the amygdala) to increase specificity, given the variability in global FC strengths across individuals. Normalized FC showed anecdotal evidence for underconnectivity of the BLA in the ASD compared to the neurotypical group (BF₁₀=2.247; ASD group: mean=0.598, SD=0.385, 95% CI=0.546, 0.650; neurotypical group: mean=0.682, SD=0.393, 95% CI=0.635, 0.729 (see Table 1), and moderate evidence for equivalent (no difference) magnitude of CCM connectivity (BF₁₀=0.156; ASD group: mean=0.475, SD=0.417, 95% CI=0.419, 0.532; neurotypical group: mean=0.476, SD=0.431, 95% CI=0.425, 0.527).

Third, to further increase anatomical specificity, we tested FC between each amygdala ROI and a data-driven subset of regions by choosing the 10% of regions from the neurotypical group that had the strongest connectivity in that amygdala ROI connectome (i.e., between the given amygdala ROI and all the other 408 regions). We then took this new amygdala-ROI connectome, consisting of 41 FC values (10% of 408), as a new measure of the most specific amygdala connectivity to test for group differences. We found no evidence for differences in connectivity to the BLA and CCM between the two groups (all BF₁₀ values <0.7) (see Table 1). Thus, the overall results from testing hypothesis 1 showed an initially moderate effect in favor of underconnectivity in ASD (but only for the BLA) that became weak to absent with increased specificity of data.

No evidence for atypical amygdala connectivity patterns in ASD.

Hypothesis 2 quantified similarity between group-specific multivariate patterns of amygdala FC to the whole brain, to the top 10% most strongly connected regions, and to the DMN. For all three of these levels of anatomical specificity, and when controlling for nuisance variables, we found strong evidence for similarity between ASD and neurotypical amygdala connectomes, that is, against hypothesis 2 (BLA: whole brain, BF₁₀=1.46 e+131; top 10%, BF₁₀=1.58 e+22; DMN, BF₁₀=1.46 e+28; CCM: whole brain, BF₁₀=9.37 e+131; top 10%, BF₁₀=2.72 e+7; DMN, BF₁₀=1.38 e+30) (see Figure 3D and Table 2; see also Table S5 in the online supplement).

TABLE 2.

Bayes factors for pattern connectivity hypothesis testing for group similarity in functional connectivity pattern^a

		Pattern Similarity
Region of Interest	Pipeline	Whole Brain	Top 10%	DMN
Basolateral amygdala	A	1.46 e+131	1.58 e+22	1.46 e+28
	B	3.59 e+93	2.45 e+8	8.68 e+19
Cortico-centromedial amygdala	A	9.37 e+131	2.72 e+7	1.38 e+30
	B	1.03 e+84	5.22 e+9	7.51 e+15
Medial amygdala	A	1.31 e+111	3.84 e+6	1.72 e+21
Dorsal amygdala		4.14 e+98	78.88	1.58 e+24
Ventrolateral amygdala		6.75 e+103	2.13 e+8	5.76 e+19
Whole amygdala		2.88 e+130	4.12 e+12	2.18 e+28

^aThe table lists Bayes factors (BF₁₀) for pattern connectivity hypothesis testing for similarity between groups (autism and neurotypical) in functional connectivity (FC) between amygdala regions of interest (ROIs) (anatomical: basolateral, cortico-centromedial, and whole amygdala; functional: medial, dorsal, and ventrolateral amygdala) and the whole brain, the top 10% strongest connected parcels defined by the neurotypical group, and the default mode network (DMN) for two different denoising pipelines (pipeline A [29] and pipeline B [30]). Pipeline B was applied only to the main analysis of the anatomically defined amygdala ROIs (basolateral and cortico-centromedial amygdala), to test for robustness of the main results to variation in neuroimaging data preprocessing. Note that large positive BF₁₀ values (>10) shown here present strong evidence against hypothesis 2, with larger correlation indicating similarity of group-specific amygdala FC vectors (i.e., no difference between groups).

No robust evidence for greater between-subject variability of amygdala FC in ASD.

Hypothesis 3 tested for increased neural variability in amygdala FC in ASD by comparing the variability of amygdala connectome correlations between all pairings of subjects, averaged per subject, in each group (ASD and neurotypical). A permutation-based analysis (shuffling group membership of individual averaged Kendall’s tau correlation scores) showed that between-subject variability did not significantly differ between the ASD and neurotypical groups (one-sided permutation test, 10,000 iterations, Kendall’s tau: BLA, p=0.862; CCM, p=0.994; median group difference [ASD – neurotypical]: BLA, 0.006; CCM, 0.012). In an analysis that was not preregistered, we further explored the influence of similarity metric choice (Pearson’s correlation, Euclidean distance, cosine distance) and assessed variability within four data acquisition sites (those that had participant Ns >30) separately, with similar results. In sum, we found no consistently increased variability in the ASD group (see Figures S4 and S6 and Tables S6 and S19 in the online supplement).

Exploratory Analyses (Preregistered)

Selective associations with individual differences in social cognition in ASD.

We explored whether the magnitude results (hypothesis 1) were associated with individual differences in ADOS scores (combined social and communication scores; N=160 with valid data) and Social Responsiveness Scale (SRS) (42) scores (N=138 with valid data) in ASD. We found moderate evidence that higher SRS scores were associated with FC magnitude for the 10% strongest connected regions to the CCM when FC was normalized across the whole brain (BF₁₀=8.919; Kendall’s tau=−0.17), but not for non-normalized FC values, nor for magnitude across the whole brain (BF₁₀<0.21; see Table S14 in the online supplement). There was no such relation for the CCM and the ADOS (see Table S13 in the online supplement), and none for the BLA and the ADOS or SRS (all BF₁₀ values <0.2; see the online supplement for details and additional analyses, including for the neurotypical group; see also Tables S13–S18 and Figure S5 in the online supplement).

Results affected by choice of fMRI preprocessing pipeline, limiting robustness of findings.

We investigated whether results for hypothesis 1 would generalize to those obtained with a second denoising pipeline (pipeline B). The two pipelines differed in smoothing, temporal filtering, and motion time point censoring, while both included global signal regression (see the online supplement). We found that the inconsistent underconnectivity effect in ASD reported above was not robust to variations in denoising pipelines (see Table 1; see also Table S7 in the online supplement). We take these findings to show the sensitivity of the results to details of the fMRI processing. A post hoc exploratory analysis (not preregistered), which also included those participants who had previously been excluded for motion (total N=591), produced strongly enhanced group differences in the BLA (BF₁₀=470; see Table S20 in the online supplement). We take these additional findings to corroborate our above point: group magnitude differences in amygdala FC are unreliable, becoming weaker the more specific the data. This last result also suggests that spurious findings may result from a specific factor warranting future investigation: motion artifacts in the MRI data.

Analyses investigating pattern similarity (hypothesis 2) across the two tested pipelines (see Table S8 in the online supplement) indicated strong evidence against hypothesis 2—that is, we found similarity between amygdala FC patterns in the ASD and neurotypical groups regardless of preprocessing pipeline choice.

Similar directions of effects for functionally derived amygdala subdivisions.

To explore the robustness of our findings with respect to the choice of amygdala segmentation, we tested whether magnitude (hypothesis 1) and pattern similarity results (hypothesis 2) replicated when using template-space functionally defined amygdala subdivisions rather than the anatomically defined amygdala subdivisions used in all the analyses described above (16). For magnitude, we found no evidence for hypothesis 1; that is, there was no evidence for group FC differences when using these functional amygdala subregions (ventrolateral, medial, and dorsal), with respect to either the whole brain or the 10% strongest connected regions (all BF₁₀ values <1; see Table S9 in the online supplement).

For pattern similarity (hypothesis 2), we likewise found effects similar to those we found with anatomically defined amygdala ROIs: there was strong evidence against hypothesis 2. We found similar patterns of connectivity across the groups for the whole brain, for the 10% strongest connected regions, and for the DMN network (all BF₁₀ values >75; see Table S10 in the online supplement).

Similarity of effects in whole amygdala and BLA.

We also analyzed the amygdala as a whole, instead of segmentation into subregions. For magnitude (hypothesis 1), we observed similar underconnectivity results as compared to the BLA subregion: there was moderate evidence of underconnectivity between the whole amygdala and the whole brain (BF₁₀=7.5; see Table 1; see also Table S11 in the online supplement) and a weaker anecdotal effect when normalized across whole brain connectivity (BF₁₀=2.5). With regard to pattern similarity (hypothesis 2), we found strong evidence that patterns of connectivity from the whole amygdala to the rest of the brain, to the 10% strongest connected regions, and to the DMN network were similar between the ASD and neurotypical groups, replicating the subregion results (all BF₁₀ values >4×10¹²; see Table 2; see also Table S12 in the online supplement).

DISCUSSION

Autism is a pervasive developmental disorder whose behavioral phenotype undoubtedly arises from atypical brain function. The challenge for neuroscience is to elucidate that atypical neural functioning while exercising the utmost caution to avoid spurious positive findings. The results of this study demonstrate weak and inconsistent effects, or lack of evidence in support of prior hypotheses, based on a systematic analysis of a large fMRI data set of autistic participants using preregistered decisions related to quality control, anatomical targets, and denoising strategies.

Analyses testing our first hypothesis on amygdala FC magnitude provided moderate supporting evidence (Bayes factor; ~7) for underconnectivity between the BLA (but not the CCM) and the rest of the brain in ASD, initially supportive of the underconnectivity hypothesis. However, this effect became weaker when FC was normalized by the mean magnitude of FC across the whole brain, and it disappeared altogether when restricted to an analysis of the amygdala’s FC with only the 10% most strongly connected regions. The initial underconnectivity effect was not robust to a different denoising pipeline, and it was absent for the CCM subregion under all analyses. Analyses testing our second hypothesis showed that patterns of amygdala connectivity across different anatomical scales (whole brain, top 10%, and DMN) were robustly similar between the groups, regardless of processing pipelines and amygdala segmentations. Analyses testing our third hypothesis showed that between-subject variability was largely similar between the ASD and neurotypical groups (no robust evidence for increased heterogeneity of amygdala FC in ASD). There was thus inconsistent evidence for magnitude differences in amygdala FC (hypothesis 1), strong evidence against atypical patterns (hypothesis 2), and no consistent evidence for increased variability of FC in ASD (hypothesis 3).

Only a subset of participants (65% of the ASD group, 30% of the neurotypical group) had ADOS or SRS scores available, but one of these scores was indeed associated with neurobiological measures. Lower FC magnitude in the CCM subregion in ASD was associated with higher SRS scores, when restricted to the 10% top regions and when normalized across the whole brain (see Tables S15–S18 and Figure S5 in the online supplement). However, this finding was not robust against the second processing pipeline. In-depth behavioral assessments together with more stringent data quality control (e.g., larger samples with deeper high-quality data) could help future studies account for likely individual differences that contribute to poor statistical power to detect reliable group differences.

Limitations

The ABIDE data set is assembled across many different data collection sites, introducing considerable variability, even though the factor of data acquisition site was explicitly modeled in our analyses. Indeed, previous studies have argued that it is easier to classify acquisition sites from the neuroimaging data than to distinguish autism in the ABIDE data set (14). Both mechanisms for single-site, large-sample data collection and ways of generating fMRI data that are less susceptible to site-specific effects (such as movie fMRI, which is known to produce much more consistent brain activations than rs-fMRI [43]), could reduce site effects.

The ABIDE sample is already nonrepresentative by excluding more severely affected individuals and infants because of MRI incompatibility, and the motion exclusion criteria amplified this bias (see Tables S2 and S3 and Figure S3 in the online supplement). Generalizations from this work are thus limited to a small section of the autism spectrum. Relatedly, age-specific nonlinear changes in amygdala FC have been found beyond early childhood (44) and could obscure findings in our analysis. Our focus was on autistic adolescents and adults, rather than children and infants, for several reasons: we would expect FC (and other brain metrics) to change as a function of development, making it more difficult to discover stable effects at the group level. For instance, signal quality would be expected to be considerably worse as a result of increased in-scanner motion in younger participants. Of note, our primary interest in this study was not in autism as a developmental disorder, but as a lifelong pervasive disorder that also has a clear atypical phenotype in adulthood. Future investigations should target developmental trajectories of amygdala FC, ideally longitudinally in the same participants.

The choice of rs-fMRI-based parcellation and targeted subsets of regions and networks (top 10%, DMN) may also limit the generalizability of our findings. Individually derived networks (45) may be most sensitive for discovering group differences, yet they require more data per individual (46) than available in the limited ABIDE data sets, which had only 4–12 minutes total.

Our approach of defining amygdala subregions individually in native space increased spatial specificity over template space approaches. The spatial reliability of the resulting subregions may not be ideal (47), especially for smaller subregions. We partially addressed this by combining individual subregions into complexes.

Future Directions

Future studies should formally test the specific influence of nuisance variables and denoising pipelines, since these lead to different conclusions. We (48) and others (14, 49) have previously noted the importance of motion correction in connectivity studies of autism, and different pipelines do this in different ways and with different degrees of completeness. When we conducted a post hoc exploratory analysis (see Table S20 in the online supplement) including participants who had been excluded for in-scanner motion, the results emphasized BLA underconnectivity. Is the underconnectivity effect thus driven merely by increased motion in autistic participants? Although we did not find any motion differences between groups on coarse metrics such as framewise displacement, we consider it likely that more subtle aspects of motion do indeed distinguish ASD and may influence any FC results (e.g., stereotyped rocking movements, motion reduction strategies, and complex interactions between IQ, motion, and group that may not be captured adequately simply by using nuisance regression; see reference 50 for a review). Future studies could specifically investigate potentially related FC differences in the somatomotor network (see Figure 3D).

We urge the accumulation of deeper and longitudinal neuroimaging data to assess potential neurobiological markers of atypical FC in autism. Deeper data per subject would allow for individual whole brain parcellation, more rigorous assessment of FC reliability across multiple runs, and identification of most variable FC connections across individuals. We are also in agreement with others that simply increasing sample size is not necessary, and may not be ideal, to address the present limitations (51). The ABIDE data set is heterogeneous in terms of the participants (varying ages, diagnoses by different clinicians) and the measurements taken (different sites, scanners, and scanning parameters). This could be improved by having a more uniform scanning protocol (as the Human Connectome Project did) and diagnosis verification by a single group. Needless to say, such large-scale coordination would require a large-scale funding effort. A longitudinal approach should also be combined with including children to address the developmental aspect of amygdala FC and complement the pervasive aspects in adulthood presented here. We further echo recent calls for using movie fMRI (43), which can produce data that are highly consistent across different sites and samples (52). If a uniform set of video stimuli were adopted by the research community, a much higher-quality data set could eventually be accumulated, complementing the ABIDE data sets. Because movies are tolerated for longer scanning sessions, movie fMRI could potentially address the need for more clinically representative imaging samples, further extending representation across the spectrum. There are, however, practical limitations in collecting deep data in individuals with more severe levels of impairment. For instance, including autistic individuals who are minimally verbal, have more prominent motor symptoms, or are otherwise more cognitively impaired (e.g., with an FSIQ <80) may not be ethical for comfort and compliance reasons with the longer fMRI scans that we are calling for here. We would further urge a uniform set of more in-depth behavioral assessments of all participants, including the SRS, the ADOS (for autistic participants), and standard neuropsychological and cognitive assessments (as, for example, in the Healthy Brain Network study [53], covering the developmental brain from ages 5 to 21 years).

CONCLUSIONS

It is important to note that we do not conclude that amygdala FC is generally typical in autism. Instead, we conclude that the evidence for atypical FC of the amygdala in autism is weak at best, and unreliable. Ultimately, our suggested ideal of representative, densely assessed, in-depth data can only be achieved by the consistent accumulation across multiple high-quality studies, an investment worth prioritizing if we are to better understand and delineate the neurobiological substrates of autism.