Resting-state networks of believers and non-believers: An EEG microstate study

Atheism and agnosticism are becoming increasingly popular, yet the neural processes underpinning individual differences in religious belief and non-belief remain poorly understood. In the current study, we examined differences between Believers and Non-Believers with regard to fundamental neural resting networks using EEG microstate analysis. Results demonstrated that Non-Believers show increased contribution from a resting-state network associated with deliberative or analytic processing (Microstate D), and Believers show increased contribution from a network associated with intuitive or automatic processing (Microstate C). Further, analysis of resting-state network communication suggested that Non-Believers may process visual information in a more deliberative or top-down manner, and Believers may process visual information in a more intuitive or bottom-up manner. These results support dual process explanations of individual differences in religious belief and add to the representation of non-belief as more than merely a lack of belief.

, and alterations of microstates after drug-intake (Schiller, Heinrichs, Beste & Stock, 2021;Schiller, Koenig, et al., 2019). Microstate analysis of electroencephalography (EEG) quantifies quasi-stable spatial-electrical distributions on the scalp (i.e., topographies). These quasi-stable topographies typically last between 40 and 120 ms before snapping into another topography, i.e., another microstate. Different microstates are thought to reflect activation in different, distributed neural networks. Resting-state microstates are calculated from passive EEG recording and may thus index individual differences in functioning in core neural networks. Four prototypical types of microstate networks (termed A, B, C, and D) account for approximately 80% of variance in nearly all individual resting EEG recordings (Koenig et al. 2002), and are reliable across individuals (Khanna et al., 2014;Liu, Xu, Zou, He, Zou & Gao, 2020).
Based on combined EEG and fMRI studies, it has been found that Microstate A is associated with brain activity in the temporal cortex and the left insula, indicating a role in phonological processing (Britz et al., 2010;Custo, Van De Ville, Wells, Tomescu, Brunet & Michel, 2017). Microstate B is associated with areas of the occipital cortex, indicating a role in visual processing (Britz et al., 2010;Custo et al., 2017). This view is supported by increased contribution of Microstate B in eyes-open EEG compared to eyes-closed EEG (Seitzman, Abell, Bartley, Erickson, Bolbecker & Hetrick, 2017). Though more difficult to characterize, Microstate C has been associated with default mode and task-negative networks as evidenced by brain activity in the anterior and posterior cingulate cortex, inferior frontal gyrus, insula, and precuneus (Britz et al., 2010;Custo et al., 2017). These results indicate that Microstate C is a task-negative network with a role in more intuitive or automatic (bottom-up) processing (Gronchi & Giovannelli, 2018;Jilka et al., 2014;Menon & Uddin, 2010;Vatansever et al., 2017). This view is supported by decreased activity of Microstate C during a serial subtraction task requiring top-down executive control (Seitzman et al., 2017), increased activity during initial bottom-up encoding of visual stimuli compared to later top-down retrieval (D'Croz- Baron, Bréchet, Baker & Karp, 2021), and decreased activity during fluid reasoning tasks involving spatial relationships and visualization (Zappasodi et al., 2019). Conversely, Microstate D is associated with brain activity in right-lateralized areas of the frontal and parietal cortex, namely the inferior parietal cortex, the right middle and superior frontal gyri, and the insula (Britz et al., 2010;Custo et al., 2017), indicating that Microstate D has a role in deliberative or analytic (top-down) processing (Damoiseaux et al., 2006;Mantini, Perrucci, Del Gratta, Romani & Corbetta, 2007). This view is supported by increased activity of Microstate D during a serial subtraction task requiring top-down executive control (Seitzman et al., 2017), increased activity during the preparation of a top-down directed decision process (D'Croz- Baron et al., 2021), and increased activity during a fluid reasoning task involving spatial relationships (Zappasodi et al., 2019). Microstate analysis of resting EEG is thus ideally suited to examine differences between Believers and Non-Believers with regard to fundamental neural networks related to deliberative (Microstate D) and intuitive reasoning (Microstate C).
In the current research, we examined how Believers and Non-Believers differ with regard to neural network activation at rest. In particular, we expected that Non-Believers would be characterized by increased contribution of Microstate D to the resting-state EEG, indicating increased activation of a neural network associated with deliberative or controlled (top-down) processing. Conversely, we expected that Believers would be characterized by increased contribution of Microstate C to the resting-state EEG, indicating increased activation in a neural network associated with more intuitive or automatic (bottomup) processing.

Participants
Ethical approval for this study was obtained from our university's Human Research Ethics Board (Protocol 00084513). Resting EEG was recorded in a sample of 104 participants, who were recruited from the department's research participant pool. For the current study, we had N = 69 participants available from this sample (M age = 19.28; SD = 1.22; females = 40) who had responded to the item "What is your current religious affiliation?" in a previous mass testing administered as part of enrollment in the research participant pool. Participants who identified as religious (Christian, Jewish, Muslim, Buddhist, Hindu, and Other) were classified as Believers (n = 43; M age = 19.42; SD = 1.18; females = 27). Participants who identified as Atheist or Agnostic were classified as Non-Believers (n = 26; M age = 19.04; SD = 1.28; females = 13). All participants had normal or corrected-to-normal vision.
Believers and Non-Believers did not significantly differ with regard to gender (X 2 = 1.09, p = .297), markers of intelligence (self-reported Math Ability and Openness to Experience (Gosling et al., 2003); both p > .344), social status (p = .663) or markers of predisposition towards mood and relational disorders (trait anxiousness, as indexed with the Behavioral Inhibition System scale [Carver & White, 1994], and emotional instability, as indexed with the Neuroticism subscale [Gosling et al., 2003]; both p's > 0.843).

Procedure
As indicated above, this study took place in two parts. First, participants in the research participant pool completed the mass testing portion online (including the item on religious belief and a questionnaire on Math Ability using 6 items with 7-point response scales). Second, participants (N = 104; of which n = 69 completed the mass testing item on religious belief) were scheduled to an in-lab component. Participants conducted the experiment at a computer station in an electrically-and sound-shielded EEG-cabin. They first completed an informed written consent and were then fitted with a 64-channel EEG headset (Brain Products GmbH, Munich, Germany). Participants then answered demographic questions and several personality questionnaires as part of a larger research project on individual differences in the neuroscience of self-regulation [including a brief measure of the Big 5 with 2 items per subscale and 5-point response scales (Gosling et al., 2003), an item on social status ("Where would you place your primary caregiver(s)? Choose the number that best describes where you think your primary caregiver(s) stood during your childhood relative to other people in your community" with a 10-point response scale (1 = the bottom, 10 = the top), and the Behavioral Inhibition System scale (Carver & White, 1994); all data available upon request]. Participants then completed a resting-state task during which EEG was recorded. The task involved sitting passively for four minutes, alternating 1 min periods with eyes-open and eyes closed. In line with standard procedures (e.g., Damoiseaux et al., 2006;Mantini et al., 2007;Lee et al., 2013; for reviews, see Newson & Thiagarajan, 2019), only eyes-closed periods were used for analysis (two minutes in total). Eyes-closed EEG recording provides a more reliable measure of resting-state brain activity compared to eyes-open periods, as spontaneous processing of surrounding environmental (e.g., visual and emotional) stimuli don't affect the EEG (Barry, Clarke, Johnstone, Magee & Rushby, 2007). Then, again as part of a separate line of research, participants completed a number of cognitive and behavioral tasks not related to the current research (again, all data available upon request; see also Leota et al., 2021;Nash, Tran, Leota & Scott, 2020. Finally, participants completed manipulation and compliance checks. Participants were then debriefed, had the headset removed and hair washed, and thanked for their time.

EEG Recording and Preprocessing
Continuous resting-state EEGs were recorded using a 64 Ag-AgCl channel ActiCHamp EEG system (Brain Products GmbH, Munich, Germany), positioned according to the 10/10 system and digitized at a sampling rate of 500 Hz (24 bit precision; bandwidth: 0.1-100 Hz). During the baseline recording, signals were referenced to TP9 electrode positioned over the left mastoid. All EEG data was preprocessed offline using Brain Vision Analyzer (Version 2.1.0.327; Brain Products GmbH, Munich). A notch filter (50 Hz) and a band-pass filter of 1.5-20 Hz were applied to the resting-state EEG data. Ocular artifacts were identified and removed using a semi-automatic independent component analysis for each participant. EEG channels that were defect or heavily affected by artifacts were interpolated using neighboring electrodes. Remaining artifacts were automatically identified (Min/max threshold: − 100 to +100 μV, maximum voltage step: 50 μV, lowest allowed voltage difference [maximum-minimum] in intervals of 100 ms: 0.5 μV) and removed from the signal. This automatic procedure was manually inspected and, if necessary, corrected, resulting in artifact free EEG signals for microstate analysis. Then, the signal was re-referenced to an average (whole-head) reference. Finally, artifact-free epochs with durations of at least two seconds were exported for microstate analysis (also see Khanna et al., 2014;Koenig, Lehmann, Merlo, Kochi, Hell & Koukkou, 1999).

Resting-State Microstates
Resting-state microstate analyses were conducted using a software plugin for EEGLAB (Delorme & Makeig, 2004) by Koenig (Koenig, 2017; for procedures, see Lehmann et al., 1987;Wackermann, Lehmann, Michel & Strik, 1993). To obtain individual electric potential field maps of the scalp, artifact-free EEG data from all available channels were extracted for each participant at time points of maximum global field power (GFP; Koenig et al., 2002). Extracting data at GFP-peaks only instead of using data from the whole epoch ensures an optimal signal-to-noise ratio (Koenig et al., 2002). In each participant, microstate maps from all GFP-peaks in the EEG recording were then submitted to an atomize-agglomerate hierarchical cluster analysis (AAHC) to identify the four most prevalent microstate-maps (Michel, Koenig, Brandeis, Gianotti & Wackermann, 2009;Murray et al., 2008). Third, maps of all subjects who completed resting-EEG measures (N = 104) were submitted to a second cluster analysis yielding four grand-mean microstate maps that we manually sorted to fit the standard order (Koenig et al., 2002). Note that we chose a fixed number of 4 clusters to maximize the comparability of our results with previous research (for an overview of studies using four microstate clusters, see Michel & Koenig, 2018). This approach allowed us to interpret the results with reference to previous findings on the functional significance of the four prototypical microstate types A, B, C and D. We included the full EEG sample in this step (N = 104) instead of including only Believers and Non-Believers (N = 69) in order to obtain more reliable grand-mean microstate maps for the back-fitting. Note that grand-means obtained from the whole sample and grand-means obtained from Believers and Non-Believers show no noticeable differences, and closely resemble the four canonical resting-state microstate types A, B, C and D known from the literature (see Table 1 and Table S1 in the supplementary material; for reviews, see Khanna, Pascual-Leone, Michel & Farzan, 2015;Michel & Koenig, 2018). Based on spatial correlations with the four grand-mean microstate maps obtained from our own data (N = 104), the four most prevalent microstate-maps of each individual participant were then assigned to the best-fitting grand-mean microstate type A, B, C or D. Next, individual potential field maps from GFP peaks were assigned to the best fitting of the four most prevalent microstate maps in each participant, resulting in a continuous temporal stream of microstates. Finally, we extracted average durations in milliseconds, average numbers of occurrences per second, and percentage contributions to the EEG signal (duration x occurrence) for each microstate type and each subject.
We also extracted transitions involving the microstates C and D (transitions from any other microstate type to these microstates and transitions from these microstates to any other microstate type). Transitions between microstates are calculated as the observed number of transitions relative to the expected number of transitions based on the occurrence of a network. For example, the transition from microstate B to D shows how much more often B switches to D per second (on average) than one would expect based on occurrences of microstate B. Thus, considering baseline occurrences of microstate B controls for the fact that switches from network B to any other network are more likely if B occurs very frequently (for details, see Lehmann et al., 2005). Note that transitions between microstate maps extracted from GFP-peaks (as was performed in our microstate analysis) might differ from transitions between microstate maps extracted from the whole EEG recording, as microstates between GFP-peaks are not considered (Michel & Koenig, 2018). This remains a matter of controversy in the current literature with further research needed (e.g., .

Statistical Analyses
In order to examine the links between Believers and Non-Believers in resting-state neural network functioning, we first calculated the contribution (% of the total resting-state) of microstates C and D during resting-state (eyes-closed) as the dependent variables. Contribution is based on the values of microstate duration (average length of each microstate in milliseconds) and occurrences (total number of instances Table 1 Grand-mean microstates maps. N = 104. Grand-mean maps of resting-state microstates A-D. Note that the maps closely resemble the canonical microstate types A, B, C and D known from the literature (Koenig et al., 2002). of each microstate/sec) in the resting-state.
We conducted a one-way MANCOVA with belief group as the independent variable (Non-Believers vs Believers) and contribution of microstates C and D as the dependent variables, and the average microstate duration and occurrence across all four microstates entered as covariates to ensure that any differences between groups are associated with the specific microstate and not a general microstate stability. We then followed up with separate one-way ANCOVAs for each dependent variable, with the same covariates. To determine if duration or occurrences are driving any contribution differences between groups, we next conducted separate MANCOVAs with duration and occurrences of microstates C and D as the dependent variables. We then again followed up with separate one-way ANCOVAs for each dependent variable. Again, to control for a general microstate stability, we entered mean duration in the duration-specific analyses, and total occurrences in occurrencespecific analyses, as covariates. Finally, we used a MANOVA and follow-up t-tests to test for group differences between Believers and Non-Believers in microstate transitions involving the microstates C and D.

Resting-State Microstates
On average, there were 106.34 s of artifact-free resting-state EEG data available for microstate analyses (SD = 7.33;). AAHC-analyses resulted in an average variance-explanation of the EEGsignal of 74.89% (SD = 4.37;). This value is closely in line with previous research and demonstrates that four microstate clusters fit our EEG data well (see Table 1 for grand-mean microstate maps). On average, participants showed a contribution of microstate C of 27.26% (SD = 5.69; range: 12.09-37.92) and a contribution of microstate D of 25.00% (SD = 4.84;). These microstate characteristics as well as transitions involving microstates C and D were normally distributed (all KS-tests p > .844). For descriptive statistics of demographic data and microstate characteristics for the whole sample used in the study (N = 69), Believers (n = 43), Non-Believers (n = 26) and Non-Responders (N = 35) see Table S2 in the supplementary material.

Resting-State Microstates in Believers and Non-Believers
In our initial MANCOVA test, there was a statistically significant difference in microstate contribution to resting-state activation based on religious belief, F(2, 64) = 3.85, p = .026; Wilk's Λ = 0.893, partial η2 = 0.107 (for details on all main analyses of the study, see Table S3 in the supplementary material). Controlling for age, gender, self-reported Math Ability, Openness to Experience, social status, Neuroticism and trait anxiousness (Behavioral Inhibition) as separate or joint covariates resulted in highly comparable results (group remained significant as a predictor in all of these analyses (p = .021 to.028), and with similar effect-sizes (partial η2 = 0.107 to 0.116; see Table S4 in the supplementary material for these analyses as well as all subsequent analyses including these covariates; all reported effects of this study remained stable in these analyses).
Next, we examined if the contribution effects are due to either duration or occurrences differences, or both. In the MANCOVA test examining duration, there was a no significant difference in microstate duration based on religious belief, F(2, 66) = 1.49, p = .232; Wilk's Λ = 0.956, partial η2 = 0.044, and no significant differences in follow-up ANCOVA analyses (both p's > 0.171; partial η2 < 0.044). However, In the MANCOVA test examining occurrences, there was a significant difference in microstate occurrences based on religious belief, F(2, 66) Finally, we explored individual differences in transitions between networks in the resting-state, focusing on transitions involving Microstates C and D (i.e., ten separate transition variables). Results showed that, in a one way MANOVA test, there was again a statistically significant difference in transitions to and from Microstate C and D based on religious belief, F(10, 58) = 2.07, p < .042; Wilk's Λ = 0.737, partial η2 = 0.263. This multivariate effect appeared to be entirely driven by

Discussion
Initial research suggests that religious non-belief may reflect a different kind of belief system, rather than a lack of one (Farias, Newheiser, Kahane & de Toledo, 2013). However, little research has directly examined Non-Believers. Here, we applied a neural trait approach to better understand the neural origins of individual differences in religious belief. Neural traits are objective, stable, brain-based measures capable of revealing sources of heterogeneity in affective, cognitive, and behavioral processes Nash et al., 2015;Nash & Knoch, 2016;Schiller et al., 2014Schiller et al., , 2020Schiller, Gianotti, et al., 2019). Broadly, our results support dual process explanations of individual differences in religious belief. Research has shown that religious belief is related to more intuitive and heuristic reasoning and non-belief is related to more deliberative and analytic reasoning (Gervais & Norenzayan, 2012;Pennycook et al., 2012Pennycook et al., , 2016Shenhav et al., 2012). Consistent with this, we demonstrate that Non-Believers are characterized by neural resting networks associated with deliberative reasoning, whereas Believers are characterized by neural resting networks associated with intuitive reasoning.
Specifically, microstate analyses revealed that Non-Believers demonstrated increased contribution from Microstate D and decreased contribution from Microstate C to the resting-state EEG compared to Believers. Further, these microstate contribution effects were primarily driven by the number of occurrences of the respective microstate. Because Microstate D is associated with deliberative or analytic (topdown) processing, and Microstate C is associated with more intuitive or automatic (bottom-up) processing (Britz et al., 2010;Custo et al., 2017;Seitzman et al., 2017), these findings support dual process models of religious belief, which suggest that Non-Believers are characterized by increased deliberative and analytic reasoning, whereas Believers are characterized by increased intuitive or automatic reasoning (Gervais & Norenzayan, 2012;Pennycook et al., 2012Pennycook et al., , 2016Shenhav et al., 2012). Analysis of microstate transitions suggests that the contribution findings may be related to individual differences in resting-state network communication. Non-Believers showed more transitions from Microstate B to Microstate D, whereas Believers showed more transitions from Microstate B to Microstate C. Given that Microstate B is associated with visuo-spatial processing (Britz et al., 2010;Custo et al., 2017;Koenig et al., 2002;Seitzman et al., 2017), these results suggest that Non-Believers may process visual information in a more deliberative, top-down manner, and Believers may process visual information in a more intuitive, bottom-up manner.
Several previous microstate studies investigated bottom-up and topdown processing at rest. For example, it was found that believers in paranormal activity showed increased bottom-up visual processing (i.e. increased contribution of Microstate B) compared to skeptics (Schlegel et al., 2012), supporting the general idea of increased bottom-up processing in believers. However, note that microstate maps of believers and skeptics in paranormal activity systematically differed from each other, indicating that the two groups might differ with regard to microstate topography rather than their temporal dynamics. As microstate maps of Believers and Non-Believers were extremely similar in our research (see Table S1 in the supplementary material), an advantage of our study is that group differences can be clearly deduced on the temporal dynamics of microstate maps (i.e., contributions of Microstates C and D and transitions from Microstate B to D and from B to C). In another study, intranasal application of the hormone and neuropeptide oxytocin led to increased contribution of Microstate D and decreased contribution of microstate C compared to controls, a similar pattern of microstate dynamics as identified in our group of Non-Believers. As oxytocin is well-known for its anxiolytic effects (e.g., Yoshida et al., 2009; for a review, see Meyer-Lindenberg, Domes, Kirsch & Heinrichs, 2011), and religious belief has been found to show complex associations with anxiety (e.g., Laurin et al., 2008; for a review, see Kay et al., 2010), we speculate that temporal dynamics of resting networks related to top-down and bottom-up processing at rest might be related to individual levels of anxiety. Future research could therefore aim to investigate differences between people high and low in anxiety with regard to Microstate C and D, or the impact of manipulating anxiety on these parameters.
Overall, our results add to the idea that non-belief is more than merely a lack of belief. Rather, in addition to research on the psychological function of non-belief, Non-Believers here were associated with distinct neural networks. Further, we note that neural traits, though highly stable, are not immutable. Various techniques, including neurofeedback, meditation, and repeated task training, can effect changes to cortical volume or cortical baseline activation in targeted brain regions (Ghaziri et al., 2013;Lazar et al., 2005;Takeuchi et al., 2010;Taya, Sun, Babiloni, Thakor & Bezerianos, 2015). As such, future research could further explore if these techniques can also effect changes in EEG microstate function (see also Hernandez, Rieger, Baenninger, Brandeis & Koenig, 2016. As the number of Non-Believers increase, a better understanding of the associated capabilities and shortcomings, as well as ways in which to improve on negative outcomes, may become increasingly important.