This protocol describes partial wavelet transform coherence (pWTC) for calculating the time-lagged pattern of interpersonal neural synchronization (INS) to infer the direction and temporal pattern of information flow during social interaction. The effectiveness of pWTC in removing the confounds of signal autocorrelation on INS was proved by two experiments.
Social interaction is of vital importance for human beings. While the hyperscanning approach has been extensively used to study interpersonal neural synchronization (INS) during social interactions, functional near-infrared spectroscopy (fNIRS) is one of the most popular techniques for hyperscanning naturalistic social interactions because of its relatively high spatial resolution, sound anatomical localization, and exceptionally high tolerance of motion artifacts. Previous fNIRS-based hyperscanning studies usually calculate a time-lagged INS using wavelet transform coherence (WTC) to describe the direction and temporal pattern of information flow between individuals. However, the results of this method might be confounded by the autocorrelation effect of the fNIRS signal of each individual. For addressing this issue, a method termed partial wavelet transform coherence (pWTC) was introduced, which aimed to remove the autocorrelation effect and maintain the high temporal-spectrum resolution of the fNIRS signal. In this study, a simulation experiment was performed first to show the effectiveness of the pWTC in removing the impact of autocorrelation on INS. Then, step-by-step guidance was offered on the operation of the pWTC based on the fNIRS dataset from a social interaction experiment. Additionally, a comparison between the pWTC method and the traditional WTC method and that between the pWTC method and the Granger causality (GC) method was drawn. The results showed that pWTC could be used to determine the INS difference between different experimental conditions and INS’s directional and temporal pattern between individuals during naturalistic social interactions. Moreover, it provides better temporal and frequency resolution than the traditional WTC and better flexibility than the GC method. Thus, pWTC is a strong candidate for inferring the direction and temporal pattern of information flow between individuals during naturalistic social interactions.
Social interaction is of vital importance for human beings1,2. For understanding the dual-brain neurocognitive mechanism of social interaction, the hyperscanning approach has recently been extensively used, showing that the patterns of interpersonal neural synchronization (INS) can well characterize the social interaction process3,4,5,6,7,8,9,10,11,12,13,14. Among recent studies, an interesting finding is that the role difference of individuals in a dyad may lead to a time-lagged pattern of INS, i.e., INS occurs when the brain activity of one individual lags behind that of another individual by seconds, such as that from listeners to speakers5,9, from leaders to followers4, from teachers to students8, from mothers to children13,15, and from women to men in a romantic couple6. Most importantly, there is a good correspondence between the interval of the time-lagged INS and that of social interaction behaviors, such as between teachers questioning and students answering8 or between parenting behaviors of mothers and compliance behaviors of children15. Thus, time-lagged INS may reflect a directional information flow from one individual to another, as proposed in a recent hierarchical model for interpersonal verbal communication16.
Previously, the time-lagged INS was mainly calculated on the functional near-infrared spectroscopy (fNIRS) signal because of its relatively high spatial resolution, sound anatomical localization, and exceptionally high tolerance of motion artifacts17 when studying naturalistic social interactions. Moreover, to precisely characterize the correspondence between the neural time lag and the behavioral time lag during social interaction, it is essential to obtain the INS strength for each time lag (e.g., from no time lag to a time lag of 10 s). For this purpose, previously, the wavelet transform coherence (WTC) procedure was extensively applied after shifting the brain signal of one individual forward or backward relative to that of another individual5,6,18. When using this traditional WTC procedure for fNIRS signals, there is a potential challenge because the observed time-lagged INS may be confounded by the autocorrelation effect of the fNIRS signal for an individual19,20,21. For example, during a dyadic social interaction process, the signal of participant A at time point t may be synchronized with that of participant B at the same time point. Meanwhile, the signal of participant A at time point t may be synchronized with that of participant A at a later time point t+1 because of the autocorrelation effect. Therefore, a spurious time-lagged INS may occur between the signal of participant A at time point t and that of participant B at time point t+1.
Mihanović and his colleagues22 first introduced a method termed partial wavelet transform coherence (pWTC), and then applied it in marine science23,24. The original purpose of this method was to control the exogenous confounding noise when estimating the coherence of two signals. Here, to address the autocorrelation issue in the fNIRS hyperscanning data, the pWTC method was extended to calculate time-lagged INS on the fNIRS signal. Precisely, a time-lagged INS (and a directional information flow) from participant A to participant B can be calculated using the equation below (Equation 1)23.
Here, it is assumed that there are two signals, A and B, from participants A and B, respectively. The occurrence of signal B always precedes that of signal A with a time lag of n, where WTC (At, Bt+n) is the traditional time-lagged WTC. WTC (At, At+n) is the autocorrelated WTC in participant A. WTC (At, Bt) is the time-aligned WTC at time point t between participant A and B. * is the complex conjugate operator (Figure 1A).
Figure 1: Overview of pWTC. (A) The logic of the pWTC. There are two signals A and B, within a dyad. The occurrence of A always follows that of B with a lag n. A gray box is a wavelet window at a certain time point t or t+n. Based on the pWTC equation (represented in the figure), three WTCs need to be calculated: the time-lagged WTC of At+n and Bt; the autocorrelated WTC in participant A of At and At+n; and the time-aligned WTC at timepoint t, At and Bt. (B)The layout of optode probe sets. CH11 was placed at T3, and CH25 was placed at T4 following the international 10-20 system27,28. Please click here to view a larger version of this figure.
This protocol first introduced a simulation experiment to demonstrate how well the pWTC resolves the autocorrelation challenge. Then, it explained how to conduct pWTC in a step-by-step way based on an empirical experiment of naturalistic social interactions. Here, a communication context was used to introduce the method. This is because, previously, the time-lagged INS was usually calculated in a naturalistic communication context3,4,6,8,13,15,18. Additionally, a comparison between the pWTC and the traditional WTC and validation with the Granger causality (GC) test were also conducted.
The human experiment protocol was approved by the Institutional Review Board and Ethics Committee of the State Key Laboratory of Cognitive Neuroscience and Learning at Beijing Normal University. All participants gave written informed consent before the experiment began.
1. The simulation experiment
2. The empirical experiment
Simulation results
The results showed that the time-lagged INSWTC with autocorrelation was significantly higher than the time-lagged INSWTC without autocorrelation (t(1998) = 4.696, p < 0.001) and time-lagged INSpWTC (t(1998) = 5.098, p < 0.001). Additionally, there was no significant difference between time-lagged INSWTC without autocorrelation and INSpWTC (t(1998) = 1.573, p = 0.114, Figure 2A). These results indicate that pWTC can effectively remove the impact of the autocorrelation effect on INS. Additionally, when the WTC value was set to be close to 0 or 1, the time-lagged INSpWTC still showed reliable results when the WTC value was away from 0 or 1 (Supplementary Figure 2).
Empirical experiment results
INS pattern using the traditional WTC method
The results showed that at 0.04-0.09 Hz,INSWTCin the sensorimotor cortex (SMC, CH20) of both women and men was significantly higher in the supportive topic than in the conflict topic when the brain activity of men lagged behind that of women by 2 s, 4 s, and 6 s (2 s: t(21) = 3.551, p = 0.0019; lag 4 s: t(21) = 3.837, p = 0.0009; lag 6 s: t(21) = 3.725, p = 0.0013). Additionally, at 0.4-0.6 Hz, INSWTC in the SMC was significantly higher in the conflict topic than in the supportive topic when men's brain activity lagged behind women's by 4 s (t(21) = 2.828, p = 0.01, Figure 2B).
Additionally, to compare the direction of INSWTC in different topics, a topic (supportive, conflict) x direction (women to men, men to women) ANOVA was first conducted on INSWTC of the SMC under a 2-6 s time lag. The 0.04-0.09 Hz results did not show any significant interaction effects at any time lag (ps > 0.05). For the 0.4-0.6 Hz frequency range, the results showed that the interaction effect was marginally significant (F(1, 21) = 3.23, p = 0.086). Pairwise comparisons showed that INSWTC from women to men was significantly higher in the conflict topic than in the supportive topic (M.D. = 0.014, S.E. = 0.005, p = 0.015), whereas INSWTC from men to women did not differ significantly between topics (M.D. = 0.002, S.E. = 0.006, p = 0.695).
Finally, to test the impact of autocorrelation on the results of traditional time-lagged INSWTC, INSWTC was compared between WTC(Wt, Mt+4) and WTC(Mt, Mt+4) at 0.04-0.09 Hz and 0.4-0.6 Hz, respectively. Note that the INSWTC of WTC(Mt, Mt+4) reflects autocorrelation. The results showed that at the 0.4-0.6 Hz, there was no significant difference between the INSWTC of WTC(Wt, Mt+4) and that of WTC(Mt, Mt+4) (t(21) = 0.336, p = 0.740). At 0.04-0.09 Hz, the INSWTC of WTC(Mt, Mt+4) was significantly higher than that of WTC (Wt, Mt+4) (t(21) = 4.064, p < 0.001). A comparison was also conducted between the frequency ranges of 0.04-0.09 Hz and 0.4-0.6 Hz regarding INSWTC of WTC(Mt, Mt+4). The results showed that the INSWTC of WTC(Mt, Mt+4) was significantly higher at 0.04-0.09 Hz than at the 0.4-0.6 Hz (t(21) = 5.421, p < 0.001). These results indicate that the time-lagged INSWTC was affected by autocorrelation in both the low- and high-frequency ranges, but the impact was larger for the lower-frequency range than for the higher-frequency range.
INS pattern using the pWTC method
The results showed that the difference in INSpWTC between the conflict and supportive topics reached significance at the SMC of both women and men at 0.4-0.6 Hz when male brain activity lagged behind that of women by 4 s (t(21) = 4.224, p = 0.0003). At 0.04-0.09 Hz; however, no significant results were found, nor were their effective results at other frequency ranges (Ps > 0.05, Figure 2C).
An additional ANOVA test was conducted on the INSpWTC of the SMC at 0.4-0.6 Hz. The results showed that the interaction between topic and direction was marginally significant (F(1,21) = 3.48, p = 0.076). Further pairwise comparisons showed that INSpWTC from women to men was significantly higher in the conflict topic than in the supportive topic (M.D. = 0.016, S.E. = 0.004, p = 0.002), whereas INSpWTC from men to women did not differ significantly between topics (M.D. = 0.0007, S.E. = 0.006, p = 0.907, Figure 2D).
INS pattern using the GC method
An ANOVA test was conducted on the INSGC at the SMC within the 0.4-0.6 Hz only. The results showed a significant interaction between topic and direction (F(1,21) = 8.116, p = 0.010). Pairwise analysis showed that INSGC from women to men was significantly higher in the conflict topic than in the supportive topic (MD = 5.50, SE = 2.61, p = 0.043). In contrast, the INSGC from men to women was not significantly different between topics (MD = 1.42, SE = 2.61, p = 0.591, Figure 2E).
Figure 2: Results of the simulation and empirical experiment. (A) The simulation results of three simulated samples. The time-lagged INSWTC with autocorrelation was significantly higher than time-lagged INSWTC without autocorrelation and INSpWTC. There was no significant difference between time-lagged INSWTC without autocorrelation and pWTC. (B) The t-map of INSWTC in the empirical experiment, showing significant context effects within 0.04-0.09 Hz when SMC activity of men lagged behind that of women by 2-6 s. There was also a marginally considerable context effect within 0.4-0.6 Hz when SMC activity of men lagged behind that of women by 4 s. (C) The t-map of INSpWTC, showing a significant context effect within 0.4-0.6 Hz when SMC activity of men lagged behind that of women by 4 s. (D) Comparison of directional INSpWTC at different topics by pWTC. Directional INS from women to men is significantly higher in conflict contexts than in supportive contexts. (E) Validation of directional INS by GC test (INSGC). The resulting pattern of INSGC is similar to INSpWTC. Please click here to view a larger version of this figure.
Supplementary Figure 1: The power spectrum plot for sample rate at 11.1 Hz (blue line) and 55.6 Hz (red line). The power spectrum pattern for the two is quite similar. Please click here to download this File.
Supplementary Figure 2: The pWTC maps of floor and ceil WTC. (A) Left panel: the time-lagged WTC map generated by two same signals, the x-axis is time point, and the y-axis is frequency-band. The mean value of WTC at all points is ~1. Right panel: the pWTC map of two similar signals. The pWTC map is quite similar to the WTC map. (B) Left panel: the time-lagged WTC map generated by two random signals, the x-axis is the time point, and the y-axis is the frequency-band. The mean value of WTC at all points is ~0. Right panel: the pWTC map of two similar signals. The pWTC map is quite similar to the WTC map. Please click here to download this File.
In hyperscanning studies, it is usually essential to describe the directional and temporal patterns of information flow between individuals. Most previous fNIRS hyperscanning studies have used traditional WTC25 to infer these characteristics by calculating the time-lagged INS. However, as one of the intrinsic features of the fNIRS signal20,21, the autocorrelation effect might confound the time-lagged INS. To address this issue, in the protocol herein, a method termed pWTC was introduced22. This method estimates the time-lagged INS after partially out autocorrelation and maintains the advantages of the WTC method. This protocol offers step-by-step guidance on how to conduct pWTC and validates the results of pWTC by comparing its results with those of traditional WTC and GC tests.
The critical steps of applying pWTC in fNIRS-based hyperscanning data are demonstrated in this protocol. Specifically, first, to calculate the time-lagged WTC, the autocorrelated WTC, and time-aligned WTC must be calculated based on the time-lagged fNIRS time series. Next, the pWTC are computed at different time lags according to Equation 1. The results of the pWTC return a time x frequency matrix, and the values in the matrix ranges from 0 to 1. Thus, further statistical tests can be conducted on these values.
In the demonstration protocol, the representative results of the traditional WTC showed two significant effects at two frequency bands: 0.4-0.6 Hz. However, the impact within the 0.04-0.09 Hz did not survive the threshold in the pWTC results, suggesting that this effect might be confounded by the autocorrelation effect of the fNIRS signal. On the other hand, the results within the 0.4-0.6 Hz range were well replicated by the pWTC method. These results indicate that after removing the autocorrelation effect, pWTC provides more sensitive and specific developments in inferring INS’s directional and temporal patterns between individuals. Another possibility, however, is that pWTC is not susceptible to INS’s directional and temporal patterns in lower frequency ranges than in the higher frequency ranges, resulting in underestimation of the INS effect. Future studies are needed to clarify these possibilities further.
A comparison with the GC test further supports this conclusion. The results of the GC test were quite similar to those of the pWTC, showing important information flow from women to men but not from men to women. There was a slight difference between the results of the GC test and pWTC, i.e., the interaction effect between topic and direction was marginally significant in the results of the pWTC but reached significance in the GC test. This difference may be because the pWTC is calculated at a finer timescale than the GC test. Thus, although both the pWTC and GC tests can provide reliable results when controlling for the autocorrelation effect, the pWTC is advantageous because it is not necessary to make stationary assumptions and holds a high temporal-spectrum structure.
The pWTC method also has its limitations. Similar to the GC test, the causality inferred from pWTC is not a real causality37,38. Instead, it only indicates a temporal relationship between the signals of A and B. This issue should be kept in mind when applying the pWTC method. Second, pWTC only partials out the autocorrelation effect. Thus, other potential concurrent variables, such as shared environments or similar actions, may still impact the results. Consequently, conclusions about the direction and temporal pattern of information flow should be drawn after controlling these confounding factors.
Additionally, there were some complicated issues about fNIRS data preprocessing. Although fNIRS has a high tolerance of head movements, motion artifacts are still the most significant source of the noise39. Large head movements would still lead to a position shift of the optodes, generating motion artifacts such as sharp spike and baseline shifts. To address these issues, many artifacts correction approaches were developed such as spline interpolation40, wavelet-based filtering39, principle component analysis41, and correlation-based signal improvement42, etc. Cooper and his collegues43 have compared these approaches based on real resting-state fNIRS data and found that wavelet-based filtering produced the highest increase in contrast-to-noise ratio. Further, Brigadoi and her collegues44 have also compared these approaches in real linguistic task data and also found that wavelet-based filtering was the most effective approach in correcting motion artifacts. Thus, in this study, wavelet-based filtering was applied and also recommended for future fNIRS hyperscanning studies.
In general, pWTC is a valuable approach in estimating the directional and temporal patterns of information flow during social interaction. More importantly, it is believed that the pWTC method is also suitable for pseudo-hyperscanning studies (i.e., signals of two or multiple brains are not collected simultaneously45,46). In such experiments, although the direction of information flow is fixed, it is also of interest to examine the duration of the time lag between the input of the signal and the process of the signal. Therefore, autocorrelation can also confound the results of the time-lagged INS. In the future, this method can answer many questions in hyperscanning and other interbrain studies. For example, to determine the dominant role in various social relationships, such as teachers and students, doctors and patients, and performers and audiences. Additionally, as pWTC maintains the temporal structures of INS, it is also possible to test the dynamic pattern of INS, such as group attitude convergence.
The authors have nothing to disclose.
This work was supported by the National Natural Science Foundation of China (61977008) and the Young Top Notch Talents of Ten Thousand Talent Program.
fNIRS topography system | Shimadzu Corporation | Shimadzu LABNIRS systen | LABNIRS system contains 40 emitters and 40 detectors for fNIRS signals measurement. In this protocol we used these emitters and detectors created two customized 26-channels probe sets and attached to two caps accroding to 10-20 system. Further, LABNIRS system also contains built-in GUI softwares for data quality check, data convert and data export. |
MATLAB | The MathWorks, Inc. | MATLAB 2019a | In this protocol, several toolboxs and functions bulit in MATLAB were used: SPM12 toolbox was used to normalize the valided MRI data through its GUI. NIRS_SPM toolbox was used to project the MNI coordinates of the probes to the AAL template through its GUI. Homer3 toolbox was used to remove motion artifacts through its function hmrMotionCorrectWavelet with default parameters. Wavelet toolbox was used to compute WTC and pWTC through its function wcoherence. |
MRI scanner | Siemens Healthineers | TRIO 3-Tesla scanner | In this protocol, the MRI scanner was used to obtain MNI coordinates of each channel and optpde. Scan parameters are described in main text. |
customized caps | In this protocol, we first marked two nylon caps with 10-20 system. Then, we made two 26-channels customized optode probes sets. Finally, we attached probes sets to caps aligned with landmarks. |