We began by reanalyzing data from a steady-state visually evoked potential (SSVEP) experiment reported by
Vilidaite et al. (2018). Participants viewed arrays of flickering gratings of varying contrasts. In some conditions, a single grating orientation was present flickering at 7 Hz (the target), whereas in other conditions, a high-contrast “mask” was added at right angles to the target gratings and flickering at 5 Hz. The left panel of
Figure 1a shows contrast response functions with and without the mask—the presence of the mask reduces the 7 Hz response to the target (blue squares are below the black circles; significant main effect of mask contrast,
F(1, 99) = 26.52,
p < 0.001). Similarly, the right panel of
Figure 1a shows that the 5 Hz response to the mask was itself suppressed by the presence of high-contrast targets (main effect of target contrast on the mask response,
F(2.92, 288.63) = 46.77,
p < 0.001; note that the data from the masking conditions were not reported by
Vilidaite et al., 2018). At both frequencies, responses were localized to the occipital pole (see insets).
We then performed a time course analysis, in which we analyzed each 11-s trial using a sliding 1-s time window. The top panel of
Figure 1c shows the response at the target frequency (7 Hz) to a single stimulus of 32% contrast (black) and the response at 7 Hz when the 32% contrast mask is added (blue). For comparison, a baseline time course is also shown (gray), which was the target response at 7 Hz when a 5 Hz mask stimulus was shown (therefore controlling for attention, blinking, etc.). Analogous responses are shown at three other frequencies—the mask frequency (5 Hz) and the second harmonics of both target and mask frequencies (14 Hz, 10 Hz), at which strong responses were also found (see spectra in
Figure 1b). The reduction in signal strength when a second component is added at a different orientation frequency illustrates the masking effect. Surprisingly, there was sometimes substantial activity before and after the stimulus was presented, as is especially clear in the baseline condition shown by the gray curves in
Figure 1c. We think the most likely explanation for this is broadband noise from participant movement during the breaks between trials. Since it is approximately equal across conditions, it appears to cancel out in the suppression ratios (
Figure 1d).
Taking the ratio of the two time courses (the target-only time course and the target time course when a mask was present) to calculate a masking index reveals that for 7 Hz targets, masking increases steeply during the first 2 s of stimulus presentation, and then plateaus for several seconds (blue trace in
Figure 1d). A similar pattern is observed for the 5 Hz mask (red trace in
Figure 1d), as well as at the second harmonics, with some variability in the time course across frequencies; for example, at 5 Hz, suppression peaks at around 4 s. The black trace shows the average masking ratio across all four frequencies, which rises steeply for just over 2 s and then stays approximately constant until stimulus offset. We conducted cluster-corrected
t tests between ratios separated by 1,000 ms, testing for an increase in suppression ratio across time (i.e., a one-sided test). Points at y = 1.8 in
Figure 1d indicate time points where the ratio is significantly increasing (i.e., there is significantly more suppression 500 ms after the time point than there was 500 ms before it) and occur up until 2.27 s after stimulus presentation. We also calculated an overall effect size by comparing the amount of suppression during the first 1,000 ms following stimulus onset with that between 2,000 and 3,000 ms, averaged across all temporal frequencies. This effect size (Cohen’s
d = 0.49) indicated a medium-sized effect.
Our initial reanalysis was promising, but the data were noisy despite the large sample size (of
N = 100), because each participant contributed only eight trials (88 s) to each condition. We therefore preregistered two new experiments (see
https://osf.io/4qudc) to investigate these effects in greater detail. These had a similar overall design to the
Vilidaite et al. (2018) study, with some small changes intended to optimize the study (see
Materials and methods). The key differences were that we used shorter trials (because there were few changes in the latter part of the trials shown in
Figure 1d) and also focused all trials into a smaller number of conditions, such that each participant contributed 48 repetitions (288 s of data) to each of four conditions.
Figure 2 summarizes the results of our EEG experiment testing a further 100 adult participants. Averaged EEG waveforms showed a strong oscillatory component at each of the two stimulus flicker frequencies (
Figure 2a), which slightly lagged the driving signal. Signals were well isolated in the Fourier domain (
Figure 2b) and localized to occipital electrodes. Responses at 7 Hz were weaker in the two masking conditions, showing significant changes in response amplitude for both the monocular (
t = 7.56,
df = 87,
p < 0.001) and dichoptic (
t = 11.35,
df = 87,
p < 0.001) masks. Dichoptic masking was significantly stronger than monocular masking (
t = 7.96,
df = 87,
p < 0.001), and a similar pattern was evident at 5 Hz (note that for this experiment, the terms “target” and “mask” are arbitrary, as each component was presented at a single contrast).
The time course at both flicker frequencies showed an initial onset transient and was then relatively stable for the 6 s of stimulus presentation (
Figures 2c, d). The ratio of target-only to target + mask conditions increased over time (
Figures 2e, f) for both mask types. At 5 Hz, the increase in masking continued for as long as 5 s of stimulus presentation in the monocular condition (
Figure 2e; points at y = 0.8 indicate significantly increasing suppression, which continue until 5.1 s (mon) or 4.2 s (dich)), whereas at 7 Hz, the increase occurred primarily during the first 1.5 s after onset (
Figure 2f; substantial clusters up to 1.5 s (mon) and 1.7 s (dich)). These differences across frequency are consistent with the pilot data (see
Figure 1d). Both monocular and dichoptic masks produced similar time courses of suppression. We calculated an overall effect size comparing suppression in the first 1,000 ms after stimulus onset to the time window from 3,000 to 4,000 ms, pooling over frequency and mask type. This had a value of
d = 0.33. Overall, this second study confirmed that normalization increases during the first few seconds of a steady-state trial and extends this finding to dichoptic mask arrangements.
Next we repeated the experiment on 20 participants using a 248-channel whole-head cryogenic MEG system. Half of the participants had a diagnosis of autism, and the remainder were age- and gender-matched controls. Source localization using an LCMV beamformer algorithm (
Van Veen, Drongelen, Yuchtman, & Suzuki, 1997) showed strong localization of steady-state signals at the occipital pole (see
Figure 3a) and in the Fourier domain (
Figure 3b). Responses from the most responsive V1 vertex showed a similar time course to those of the EEG experiments at both frequencies (
Figures 3c, d) and showed increasing suppression during the first few seconds of stimulus presentation (
Figures 3e, f). The normalization reweighting effect was again clearest at 5 Hz, especially for the dichoptic condition (red curve in
Figure 3e), which increased until 2.5 s. This confirms that the reweighting effects can occur as early as primary visual cortex, consistent with findings from neurophysiology (
Aschner et al., 2018). However the data are more variable than for our EEG experiments and had fewer significant clusters, perhaps owing to a power reduction caused by the smaller sample size for this dataset and greater heterogeneity across frequency. When pooling effects over frequency and condition, the overall effect size (
d = 0.03) was near zero.
Intermodulation responses, at sums and differences of different stimulation frequencies, are another marker of nonlinear interaction (
Cunningham, Baker, & Peirce, 2017;
Regan & Regan, 1988;
Tsai, Wade, & Norcia, 2012). We also calculated the time course of the sum intermodulation terms (at 12 Hz) in our datasets (the difference terms at 2 Hz were negligible).
Figure 4 shows that for both EEG experiments, the intermodulation term increases during the first 1 s of stimulus presentation and then remains approximately constant. The intermodulation response in the MEG data was less clear, consistent with the spectra shown in
Figure 3b. It seems unlikely that intermodulation terms are useful for monitoring the time course of normalization reweighting, and indeed they may derive from a nonlinear process other than suppression, such as exponentiation and signal combination (
Regan & Regan, 1988). Previous work has identified situations in which suppression is constant, but the intermodulation term changes substantially between conditions depending on the extent of signal pooling (
Cunningham et al., 2017).
To investigate whether normalization reweighting effects differ with respect to autistic traits, we then split each dataset (averaged across temporal frequency) using median AQ score (for the EEG experiments) or according to diagnostic group (autism vs. controls) for the MEG data.
Figures 5a–c show distributions of AQ scores for each experiment and indicate for the pilot and EEG data which participants were in the high (purple) and low (green) AQ groups. The median AQ scores were 14 for the pilot data and 18 for the EEG data. In the MEG experiment, AQ scores for the autism group (mean 36.1) and the control group (mean 16.7) were significantly different (
t = 6.00,
df = 14.2,
p < 0.001), with minimal overlap (one participant with an autism diagnosis had an AQ score marginally lower than the highest AQ scores from the control group). These distributions are consistent with previous results for AQ (
Baron-Cohen et al., 2001).
We compared the time course of suppression between groups using a nonparametric cluster correction approach (
Maris & Oostenveld, 2007) to control the Type I error rate. Significant clusters are indicated at y = 0.8 in panels d–h of
Figure 5. Despite some occasionally significant clusters, there is no clear or consistent difference between groups across our three datasets. In particular, none of the significant clusters occur during the first few seconds of stimulus onset, when reweighting takes place. We also compared suppression ratios calculated on Fourier components for the full trial and found no significant effects of autism on suppression strength. For Experiment 1, we assessed suppression at the first and second harmonics separately but also found no AQ-related differences. We therefore conclude that autism/AQ score is not associated with normalization reweighting or the strength of suppression more generally.