Predicting “When” in Discourse Engages the Human Dorsal Auditory Stream: An fMRI Study Using Naturalistic Stories

Experimental design and stimuli.

To examine the neural correlates of predictive processing of a referent across a discourse, we created 20 stories in German, each of ∼2 min in length (±10 s; mean and SD of story length 306 (13) words, 23 (4) sentences). Within each story, we embedded the following sentence triplets: Sentence A introduced a subject referent in a manipulated context (described below), which was subsequently mentioned again in Sentence C. To tease apart the neural response to the referent in Sentence C from the response to its mention in Sentence A, an intervening Sentence B of variable length (mean: 4.30 s; SD: 0.96 s) acted as a natural jitter. Consider the following example (translated from the German original):

(1) “_… [The engineer pushed the pharmacist quickly back into the car,]_A [because due to the traffic one could not stay for long in the narrow street.]_B [The engineer sped off immediately.]_{C …}”

In (1), Sentences A–C are indicated by square brackets and subscripts, with “the engineer” in Sentence C constituting the first mention of that referent after Sentence A. For convenience, we will refer to Sentence A as the “context sentence” in the following and the first subsequent mention of the subject (i.e., “the engineer” in Sentence C) as the “referent”. The measured responses concern the critical event of the referent.

Crucially, the way in which “the engineer” is introduced into the discourse in the context sentence can be manipulated to modulate the referent's predictability in the upcoming discourse. To draw attention to the sentence participant, which will be referred to in the upcoming discourse, the grammatical construction of passive voice is preferred in many languages (Primus, 1999; Gildea, 2012; Bornkessel-Schlesewsky and Schlesewsky, 2014). Text analyses across different languages (Givón, 1983, 1994) have demonstrated that subjects, especially the subjects of a passive sentence, have a higher tendency to reoccur in the following discourse as opposed to other grammatical roles (e.g., in “Peter hit Bill” and “Peter was hit by Bill,” “Peter” as the grammatical subject has a higher likelihood of mention in the following discourse than “Bill” and this likelihood is increased even further via the use of a passive construction, which involves an atypical subject choice). Therefore, we manipulated the context sentence to include a passive construction as in (2), after which the referent is more predictable because it has been introduced with a clear structural cue. The action of the sentence needed to be changed to keep the same entity in the subject position of an active versus passive voice: in (1) the engineer pushed the pharmacist, whereas in (2) the pharmacist pushed the engineer. We chose this design to keep the referent the same; in this way we could compare BOLD responses to the same referent, which had been introduced in a different discourse context.

(2) “_… [then the engineer was pushed quickly into the car by the pharmacist,]_A [because due to the traffic one could not stay for long in the narrow street.]_B [The engineer sped off immediately.]_{C …”}

Importantly, referent tracking within a discourse is a multidimensional process. Thus, while some discourse-related factors, such as passivization, affect a referent's likelihood of subsequent mention (i.e., which of several possible referents the speaker will choose to refer to), others affect the choice of the referring expression used (e.g., whether the speaker will remention a previously introduced referent using a full noun phrase, such as the engineer, or a pronoun, such as he/she) (e.g., Kehler et al., 2008; Chiriacescu and von Heusinger, 2010; Kaiser, 2010). Use of a pronoun is typically viewed as indexing the high salience or accessibility of a referent, often conceptualized via the cognitive metaphor of “high activation.” Intuitively, this reflects the fact that a pronoun is used when the identity of a referent is salient enough to not warrant repetition (i.e., “he” or “she” suffices to for the hearer to realize that the speaker is referring to “the engineer”). Recent Bayesian approaches to referent tracking in discourse (e.g., Kehler et al., 2008; Kaiser, 2010) thus emphasize the need to distinguish the likelihood of referent remention, P(referent), the predictive component examined here, from the likelihood that, should a referent be rementioned, it will be expressed as a pronoun, P(pronoun|referent), a factor expressing the salience of a referent rather than its predictability. Referent salience, as indicated by P(pronoun|referent), is influenced by multiple grammatical and semantic factors (e.g., Arnold, 2001).

In our study, we always used full noun phrases rather than pronouns to refer back to the salient/nonsalient entity. Thus, while pronominalization is an important correlate of referent salience, it was not manipulated here. There were several reasons for this. First, the main focus of our study was on the neural correlates of predicting referent remention. Referent salience was included as a contrasting factor rather than one to be focused on in detail in its own right. Second, the comparison between pronouns and full noun phrases would have been rendered extremely difficult by the differences in length between the two (∼100 vs 600 ms, respectively). Thus, rather than manipulating the form with which a referent of a particular salience is rementioned, we manipulated the salience of the referent itself, and hence the likelihood of a full noun phrase being used in the event of a remention. In the spirit also adopted for our active/passive voice manipulation, we thereby kept the critical position constant while only manipulating the context leading up to it.

We manipulated referent salience via the causality of the event within which the critical referent was introduced. Highly causal events (e.g., push) tend to involve a highly prototypical event instigator (Agent) as well as a high prototypicality of the participant affected by the event (Patient). For low-causality events (e.g., hold in high esteem), the prototypicality of event participants is reduced. According to Dowty (1991), prototypical Agent properties are as follows: 1. volitional involvement in the event or state; 2. sentence (and/or perception); 3. causing an event or change of state in another participant; 4. movement (relative to the position of another participant); and 5. exists independently of the event named by the verb. By contrast, prototypical Patient properties are as follows: 1. undergoes change of state; 2. incremental theme; 3. causally affected by another participant; 4. stationary relative to movement of another participant; 5. does not exist independently of the event, or not at all. Low causality versions of examples (1) and (2) are given in examples (3) and (4).

(3) “_… [The pharmacist held the engineer in very high esteem.]_A [They knew each other for a long time so they had developed a strong/intimate friendship.]_B [The pharmacist was waiting in the car while the engineer was getting the tools from his apartment.]_{C …}”

(4) “_… [The pharmacist was held in very high esteem by the engineer.]_A [They knew each other for a long time so they had developed a strong/intimate friendship.]_B [The pharmacist was waiting in the car while the engineer was getting the tools from his apartment.]_{C …}”

High prototypicality of Agent and Patient participants increases referent salience, as indexed, for example, by pronominalization (e.g., Brown, 1983; for a recent replication in German, see Stein, 2016; Stevenson et al., 1994; for related findings, see Arnold, 2001). Additional evidence stems from visual word eye-tracking: Pyykkönen et al. (2010) used this technique to demonstrate that, even for young (3-year-old) children, both Agent and Patient are more salient after a highly causal action (e.g., a pushing event as in 1 and 2) than in a low causal situation (e.g., being held in high esteem as in examples 3 and 4 from the present study). By examining the total number of looks to Agent and Patient pictures, the authors found that children focused more frequently on the visual referents of high causality compared with low causality sentences. This result provides converging support for the link between verb causality and referent salience.

Together, these two manipulations resulted in a 2 × 2 design, including the factors voice (active vs passive) and causality (high vs low) with the following 4 conditions: AH: active voice and high causality (as in hit), AL: active voice and low causality (as in hold in high esteem), PH: passive voice and high causality (as in was hit), PL: passive voice and low causality (as in was held in high esteem). A schematic representation of the design is shown in Table 1. For our hypotheses, see the following subsection.

Each of the 20 stories created included one instance of each of the four conditions; thus, four critical events on the referent (Sentence C) after the manipulated context sentence (Sentence A). The order of the conditions within a story was controlled such that all possible orders were realized. Furthermore, we controlled for the distance between the conditions to ensure that they did not always occur at the same time points across stories. For each story, there was a twin story in which the same verbs occurred but in the opposite voice, thus yielding 40 story stimuli in total.

The 40 stories were assigned to two experimental lists of 20 stories each such that only one version of each story was included in a single list. Each participant heard one list (randomized for each participant); thus, each participant heard one of the two versions of each story. The stories also contained additional manipulations, such as passages requiring Theory of Mind (ToM) processing. The ToM results from these datasets were previously published in Kandylaki et al. (2015).

Stimuli were spoken by a professionally trained female speaker of German at a normal speech rate. We recorded the stimuli in a sound proof EEG laboratory cabin with a sampling rate of 44.1 kHz and a 16 bit (mono) sample size. For sampling, we used the sound recording and analysis software Amadeus Pro (version 1.5.3, HairerSoft) and an Electret microphone (Beyerdynamic MC930C). The audio files of all stories are available in an online repository (Kandylaki, 2016b). A description of how the stories were constructed can be found in Kandylaki (2016a).

Hypotheses.

For the active/passive voice manipulation, assumed to modulate the predictability of a referent's remention, we expected to observe suppressed BOLD response levels for highly predicted referents (introduced via a passive sentence) versus less strongly predicted referents (introduced via an active sentence). This is in line with the well-known attenuation of neural activity in response to predicted sensory signals, as observed in multiple sensory domains (e.g., vision: Sylvester et al., 2008; audition: Curio et al., 2000; and somatosensation: Blakemore et al., 1998) and modeled in predictive coding architectures (e.g., Rao and Ballard, 1999; Friston, 2010). We further hypothesized that BOLD response differences at the position of the referent would manifest themselves most prominently in the dorsal auditory stream, effectively reflecting the processing of discourse-level referent sequences, as posited by the view that the dorsal stream processes temporal sequences in a hierarchically organized manner corresponding to temporal integration windows of successively increasing length (Bornkessel-Schlesewsky et al., 2015; Hasson et al., 2015). More specifically, this hypothesis is based on the assumption that passive-voice-induced predictions are not restricted to positing that the referent will reoccur at some unspecified time point in the future discourse. Rather, they can be considerably more precise about when the predicted remention of the referent might occur. Rementioned referents are discourse “topics,” which according to a robust linguistic generalization, tend to precede nontopical information and often occur sentence-initially (e.g., Clark and Haviland, 1977; for cross-linguistic confirmation, see Siewierska, 1993; for discussion from a sentence processing perspective, see e.g., Bock, 1982; Gernsbacher and Hargreaves, 1988). Thus, the predictability of a referent with a high likelihood of remention should increase at every new sentence beginning, cued in natural, auditory linguistic stimuli (as used in the current study) via the prosodic cues that signal a sentence boundary. As this amounts to a discourse-based prediction about when in the current sequence of linguistic input the referent is assumed to reoccur, we assume that these types of predictions should correlate with activation changes in the dorsal stream on account of that stream's importance for sequence processing at differing timescales (Bornkessel-Schlesewsky et al., 2015).

For the causality manipulation, assumed to modulate referent salience/accessibility, there are several possible competing hypotheses, depending on the assumed neurobiological implementation of referent salience. One possibility is that salience affects referent processing in a similar manner to effects of attention on stimulus processing, namely, by increasing rather than attenuating sensory signals (for a recent review, see Kok et al., 2012). In Friston and colleagues' active inference-based approach, attention is conceptualized as the mechanism that optimizes precision of sensory signals, which modulates the synaptic gain for prediction errors (Friston, 2009; Feldman and Friston, 2010). As demonstrated empirically by Kok et al. (2012), this can lead to a reversal of sensory attenuation for high precision stimuli (i.e., higher BOLD amplitudes for predicted vs unpredicted stimuli). From this perspective, we should expect to observe an interaction between voice and causality, with BOLD activity attenuated for referents in passive versus active low causality conditions, while this pattern should be reversed for the high causality conditions.

Alternatively, it may be the case that referent salience should not be conceptualized as akin to an attention-based modulation in neurobiological terms. Rather, it may reflect a complementary type of predictive process to that associated with the voice manipulation. Recall from the experimental design section that, rather than affecting the likelihood of a referent's remention, salience affects the form in which the referent will be expressed if it is mentioned again (e.g., as a pronoun; compare Kehler et al., 2008; Chiriacescu and von Heusinger, 2010; Kaiser, 2010). Consequently, it affects predictions of “what” (i.e., which stimulus type is expected), whereas voice affects predictions of “when” the referent might reoccur, as motivated above (compare Arnal and Giraud, 2012; Friston and Buzsáki, 2016), “topics” typically occur at the beginning of a sentence. From the perspective that our voice and causality manipulations both lead to predictive processing (albeit differing along the dimensions of “when” vs “what”), both should give rise to the sensory attenuation expected for predicted stimuli. As described above, however, no pronominal continuations were used in the present study to ensure that the critical referent was identical across all conditions. Thus, because all rementions were in the form of full noun phrases, the preference for a pronominal remention of a highly salient referent is never confirmed in our study. Therefore, a “what”-based prediction approach would lead us to expect higher BOLD responses (reflecting higher prediction errors) in the high versus low causality conditions for noun phrase remention of the referent.

Finally, when-based and what-based predictions should manifest themselves in separable networks. When-based predictions (voice) are expected to engender attenuated BOLD signals in regions associated with the processing of implicit timing (e.g., inferior parietal and premotor cortices) (for review, see Coull et al., 2011), these regions overlap with the dorsal auditory stream, thus rendering this hypothesis compatible with the first hypothesis formulated for the voice manipulation above. For what-based predictions (causality) by contrast, regions of expected activation change are somewhat more difficult to pinpoint given that the predictions in question involve a considerably higher level of abstraction than basic acoustic expectations. We might thus expect to observe activation in the ventral auditory stream in view of its importance in processing linguistic “auditory objects” (Rauschecker and Scott, 2009; DeWitt and Rauschecker, 2012; Bornkessel-Schlesewsky et al., 2015). Based on work in nonhuman primate audition (e.g., Rauschecker, 1998; Rauschecker and Tian, 2000), this perspective assumes that the ventral stream categorizes the speech input into perceptual and, at higher levels of abstraction, conceptual units: linguistic auditory objects. Thus, the ventral stream maps complex spectrotemporal patterns in the auditory input to meanings (Hickok and Poeppel, 2007; Saur et al., 2008; Rauschecker and Scott, 2009) rather than displaying the sensitivity to linguistic sequences manifested by the dorsal stream (Bornkessel-Schlesewsky and Schlesewsky, 2013; Bornkessel-Schlesewsky et al., 2015). Accordingly, auditory objects may provide an adequate approximation of the content of a “what”-based prediction.

Pretest.

We tested the stories before the imaging study to ensure their naturalness and the efficacy of the experimental manipulation. The pretests had two aims: (1) to test for the general properties of the stories (naturalness, probability, plausibility, and comprehensibility); and (2) to ensure that highly causal and low causal verbs indeed differed as intended in the manipulation.

In an internet-based questionnaire ratings from 177 participants were obtained (mean: 23.69 years, SD: 4.38 years, range: 18–58 years). We advertised the questionnaire through a students' mailing list. We discarded data of participants who did not fulfill the language criteria (monolingually raised German native speakers). Because of the auditory presentation of the stories, we strongly encouraged the participants to use earphones for completing the questionnaire. To keep the amount of time needed for questionnaire completion at ∼20 min, each participant was asked to rate two stories, which were presented in four or five consecutive parts (depending on the story). Naturalness, probability, plausibility, and comprehensibility of the stimuli were assessed through the following questions (translated from the German original): “How natural was this passage?”; “How probable is the event that was described in the passage?”; “How often does this happen?”; “How comprehensible was this passage?”. When applicable (i.e., when the current part of the story included a causal manipulation), they also rated the agent and undergoer features (Dowty, 1991) with the Kako (2006) questions translated into German. These properties have been tested before using offline (Kako, 2006) and online paradigms (Pyykkönen et al., 2010). Ratings were collected on a 4-point scale, on which 1 was the lowest value and 4 the highest value. Table 2 shows the mean and SD of all ratings for all conditions.

The results of the questionnaire pretest were analyzed using linear mixed effects models (package lme4, Bates et al., 2014) in R (Team, 2014). To assess the effects of the different factors on the ratings, we used a forward model selection procedure within which we used likelihood ratio tests to compare a base model, including only an intercept with successively more complex models, including voice and causality as fixed factors. Participants and stories were modeled as random factors (Barr et al., 2013). Only random intercepts by participant and story were included due to convergence problems with more complex random effects structures. For simplicity, in what follows, we will refer to the model with only main effect of causality plus random factors as “causality model,” to the model with only main effect of voice plus random factors as “voice model,” to the model with only random factors as “null model,” to the model with main effects of both causality and voice as “main effects model” and to the model of the interaction of causality and voice as “interaction model.” All reported p values are outputs of the ANOVA function in R, which we used for pairwise model comparisons.

The mixed-model analyses showed no significant differences between conditions for naturalness, probability, plausibility, and comprehensibility. The null model was the minimal adequate model for the tested general properties of the stimuli (for the p values of the model comparisons, see Table 3).

For the causality ratings, we found that the causality model provided the best fit for the agent features volition, sentience, causation, as well as for motion of undergoer (for the inferential statistics, see Table 3). For motion of agent, there was a marginal improvement of fit for the voice model compared with the null model (Table 3), but between the causality and the main effects model, there was no significant difference (p = 0.127). For change agent and change undergoer, we found that the model, including main effects of both causality and voice, showed a significantly improved fit over the simpler models (for comparison of main effect of causality and main effect of voice with null model, see Table 3) and the more complex models (change agent: comparing causality model and main effects model p = 0.033, comparing voice model and main effects model p < 0.001, comparing main effects to interaction model p = 0.939; change undergoer: comparing causality model to main effects model p = 0.027, comparing voice model to main effects model p < 0.001, comparing main effects to interaction model p = 0.755).

Acoustic analysis.

To determine whether our critical conditions differed in prosodic terms, we performed an acoustic analysis of the context sentence (Sentence A) and the referent (beginning of Sentence C). We extracted measurements for duration, sound intensity, and average frequency (pitch), as these are the parameters used to modulate intonation in speech production. Table 4 presents the descriptive statistics of all measures. After checking the distribution of the measures using the allfitdistr function in MATLAB (The MathWorks), we discovered that we could not fit general or generalized linear models to our data because the data followed different distributions that cannot be modeled with current glmer options in R. Therefore, we performed the inferential statistics using the Kruskal–Wallis test in which the dependent variable is numeric (matched our measures) and the independent variable is a factor. Table 5 shows the results of the Kruskal–Wallis tests.

For the context sentence, the acoustic parameters of our critical conditions did not differ significantly with regard to average intensity and average frequency. They differed in duration, showing effects of both voice and causality. The voice effect is expected given that the passive sentences were ∼500 ms longer than the active sentences. syntactical structure crossed with the causality of the described event. Importantly, however, we used the intervening sentence between context sentence and referent (Sentence B) to introduce a natural jitter, thus allowing us to separate the brain response to the referent from that to the context sentence. Therefore, it appears unlikely that any activation differences related to the acoustic parameters of the context sentence would have been carried over to the processing of the referent.

For the referent, duration and average intensity did not reveal any significant effects of voice or causality. For average frequency, voice did not have a significant effect, whereas causality showed a marginal significance for frequency differences. This marginal effect of causality on frequency involved differences in the order of 2.5% of the average frequency. It is arguable whether such small differences are perceivable in speech. According to Rietveld and Gussenhoven (1985), differences of <3 semitones are generally not perceivable in speech comprehension. The average frequency of high causality referents differed less than a semitone from that of low causality referents; therefore, the marginal effects in the acoustic analysis cannot have influenced the behavioral results. However, from this, it does not necessarily follow that the brain is insensitive to such differences. Also, other acoustical features (such as the F₀ dynamics or the F₀ variance, which have not been analyzed here) might have driven a part of variance of the BOLD response, thereby possibly confounding the prosodic aspects with the conditions of interest. However, in contrast to the highly controlled materials used in traditional studies, the contextually rich, naturalistic auditory stimuli used in the present study render it inherently difficult, if not impossible, to fully control for prosodic differences between conditions.

Imaging procedure and behavioral data acquisition.

Participants completed a training session before entering the scanner for the fMRI session. In the training session, they listened to two stories (not part of the experimental stimuli) and answered to two questions after each story. The fMRI session included 20 stories and 40 questions (two after each story). The stories were presented auditorily through MRI-compatible earphones, and the questions were presented visually, along with their possible answers. For each participant, we optimized loudness before the start of the experiment. The presentation order of the stories was randomly assigned for each participant, to avoid sequence effects. The stories were presented in 4 blocks of 5 stories each. Between two blocks, there was a break of 45 s. During the break, the visual message “Short break!” was shown in the middle of the screen while the scanner was still running.

Each story trial was structured as follows: Before the start of each story, a fixation cross was shown in the middle of the screen for 500 ms. The cross was then replaced by a fixation point to indicate the beginning of the story which lasted ∼2 min. After the story, we included a jitter between 1.5 and 3 s (duration assigned randomly) followed by the visual presentation of the first question (referring to the immediately previous story). The question asked either about a specific person or object mentioned in the story, which were involved in a different situation than the ones described by the current critical manipulations (e.g., “How many sandwiches were left for the woman, after the man had finished eating?”). The whole question was presented at once, centered and toward the top of the screen for 5 s before the possible answers appeared toward the bottom of the screen for a maximum duration of 3 s, clearly separated from each other; each answer was marked with an index letter a (always on the left) and b (always on the right side of the screen below the question) (example answers for the previous question: “One” vs “Three”). After the first question, a second question followed with the same presentation procedure and similar type of content as the first question. Participants gave their answers by pressing the left or right button on a customized button box, fixed to their left leg, with their left middle or index finger accordingly. The left hand was chosen as a response hand to minimize left hemispheric artifacts, which could overlap with language processing (Callan et al., 2004). The position of the correct answer was counterbalanced across the experiment. All visual stimuli (cross, fixation point, questions and answers) were presented in dark gray foreground on light gray background.

Presentation of stimuli was time-jittered between story and questions and also between first and second question. The whole procedure was implemented and presented using the software package Presentation (Neurobehavioral Systems).

fMRI data acquisition.

During the MR session, a series of echo-planar-images recorded the time course of the subjects' brain activity. Measurements were performed on a 3 tesla MRI system (Trio, A Tim System 3T, Siemens) with a 12-channel head matrix receive coil. Functional images were acquired using a T2*-weighted, single-shot EPI sequence: parallel imaging factor of 2 (GRAPPA), TE = 25 ms, TR = 1450 ms, flip angle 90°, slice thickness 4.0 mm and 0.6 mm gap, matrix 64 × 64, field of view = 224 × 224 mm, in-plane resolution 3.5 × 3.5 mm², bandwidth 2232 Hz/pixel, EPI factor of 64, and an echo spacing of 0.53 ms. Transversal slices oriented to the AC-PC line were gathered in ascending order.

The initial five images were removed from the analyses to avoid saturation and stabilization effects. Head movements of the participants were minimized by using foam paddings.

A whole-head T1-weighted dataset was acquired with a 3d MPRAGE sequence: parallel imaging factor of 2 (GRAPPA), TE = 2.26 ms, TR = 1900 ms, flip angle 9°, 1 mm isometric resolution, 176 sagittal slices, 256 × 256 matrix.

fMRI data analysis.

All analyses for the fMRI data were calculated in SPM8 (Wellcome Trust Centre for Neuroimaging), implemented in MATLAB.

A slice time correction (to the 15th slice) was performed first. Then images were realigned to the first image to correct for head movement artifacts. We then normalized the volumes into standard stereotaxic anatomical MNI space by using the transformation matrix calculated from the first EPI scan of each subject and the EPI template. On the normalized data (resliced voxel size 2 mm³), we applied an 8 mm FWHM Gaussian smoothing kernel to compensate for intersubject anatomical variation.

For the single-subject analysis, we created the design matrix for each individual subject based on the presentation log files because each participant heard the stories in a different order. We modeled the critical events in seconds: mean duration 0.604 s and SD 0.152 s. As factors of no interest, we modeled separately: the rest of the stories excluding Sentence A and the referent on Sentence C (speech), Sentence A, the question, the answer, the button presses and the jitters before each question. The middle sentence between manipulated context and referent (Sentence B) had a mean duration of 4.3 s and a SD of 0.96 s: this was not modeled separately, but it was included in the regressor for speech (which excluded Sentence A and referents). Our baseline consisted of the 45 s pauses between blocks. To remove movement artifacts for each individual session, the realignment parameters were entered as separate regressors in the first-level analysis.

In the group-level analysis, we modeled the BOLD response to the referent in the 2 × 2 factorial design of Table 1 for the first-level conditions: AH versus baseline, AL versus baseline, PH versus baseline, and PL versus baseline. Brain activations were plotted on the anatomical MRIcroN template. We used the cluster extent thresholding algorithm by Slotnick et al. (2003), which implements an FWE correction using a Monte Carlo simulation approach, to correct for multiple comparisons. We set the desired FWE correction for multiple comparisons to p < 0.05 and the assumed voxel type I error to p < 0.005; after 10,000 iterations, our cluster threshold was 72 voxels. For all fMRI results reported here, we thus used an individual voxel threshold of p = 0.005 and cluster extend threshold of 72 voxels in a whole-brain analysis.

For the localization of the effects, we followed a three-step procedure to ensure that the provided neuroanatomical details are as accurate as possible. First, we ran the SPM anatomy toolbox on the results matrices, which outputted the descriptions of the clusters as well as one label for the peak voxel of each cluster. Second, we used AFNI's command line tool “whereami,” which takes MNI coordinates as input and returns labels for the peaks based on current anatomy atlases (e.g., Talairach-Tournoux Atlas, Eickhoff-Zilles cytoarchitectonic maximum probability map on MNI-152 from postmortem analysis, the atlas used by the SPM anatomy toolbox) (Caspers et al., 2006; Eickhoff et al., 2007; Bludau et al., 2014) and the maximum probability maps of Destrieux et al. (2010). Third, we reported the label of the peak voxel as labeled by the SPM anatomy toolbox and a summary of the outputs of different atlases in the description of the anatomical region of each cluster.