LFP analysis

As for preprocessing steps (Extended Data Fig. 1-1A), to measure neural activity in brain regions linked to tasks, we conducted standard preprocessing and time frequency (TF) analyses using custom MATLAB scripts and functions from EEGLAB (Ramanathan et al., 2018). (1) Data epoching: we first extracted time points for events of interest during the task (trial start and response). Time-series data were extracted for each electrode, from 2 s before to 5 s after each behavioral marker for each trial and organized into a 3D matrix (electrodes, times, trials). (2) Artifact removal: noisy trials were removed. Trials with >4× the SD in activity (measured across the time dimension) were treated as artifact and discarded. (3) Median reference: activity was then median referenced. At each time point, the “median” activity was calculated across all electrodes and subtracted from each electrode. (4) TF decomposition: a trial by trial TF decomposition (TF decomposition) was calculated using a complex wavelet function implemented within EEGLAB (newtimef function, using Morlet wavelets, with cycles parameter set to: [2, 0.7], frequency window of between 2 and 70 Hz and otherwise default settings used; Delorme and Makeig, 2004). We calculated the analytic amplitude of the signal (using the abs function). (5) Baseline normalization: to measure evoked activity (i.e., change from baseline) we subtracted, for each electrode at each frequency, the mean activity within a baseline window between 1000 and 750 ms before the start of the trial. (6) Trial averaging: we next calculated the average activity across trials for specific trial types (go correct, wait correct or wait incorrect) at each time point and frequency for each electrode, thus creating a 3D matrix (time, frequency, and electrode) for each behavioral session. (8) Comparison across animals: before averaging across sessions/animals, we “z-scored” the data recorded from each behavioral session. This was accomplished by subtracting the mean and dividing by the SD of activity in each electrode (at each frequency) over time. Z-scoring was helpful for normalizing activity measured from different animals before statistical analysis. These preprocessing steps resulted, for each session used in our data analysis, in a 3D TF-electrode (TFE) matrix of dimensions 200 × 139 × 32, which was used for further statistical analyses as described below.

We performed two main types of statistical analyses on the whole-brain TFE data. (1) Trial-type mean: we analyzed the mean and evaluated statistical significance of this mean for each trial type (go correct, wait correct, wait incorrect) for each electrode. Mean was calculated at each time and frequency point for each electrode across the 60 behavioral sessions. Statistical significance was estimated with a one-sample, two-sided t test (t test function in MATLAB), compared with the null hypothesis of Z = 0. Because we had already performed a “baseline” subtraction (as described above), this analysis was essentially capturing whether there was a significant increase or decrease in activity compared with baseline. False discovery rate (FDR)-correction was applied to the entire TF0-electrode matrix (32 electrodes, 200 time points and 137 frequencies) to evaluate statistical significance at this level. (FDR-corrected p value threshold set to 0.05). To visualize significant TF activations or de-activations, non-significant values were set to 0. 2). Trial-type contrast: for many analyses in this article, we were interested in the “contrast” of activity between trial-types (i.e., [go–wait] or [wait–go]). Contrasts were performed by subtracting the mean TFE data matrix for different trial types estimated from each session. We then calculated the average increase between trial types across sessions for either [go–wait] or [wait–go]. Statistical significance was performed with a one-sample, two-sided t test applied to this contrast, using a null-hypothesis of 0 (i.e., no difference between trial types). FDR-correction was performed across all times-frequencies-electrodes for whole-brain correction.

For many of these contrasts, we were interested in understanding where there were significant differences, but only in electrodes that show a significant activation for one trial-type in question. For this reason, when evaluating the [go–wait] contrast, we thresholded the activations based on those time/frequency/electrode points that were also significant in go trials alone. Similarly, for the [wait–go] contrast, the whole-brain activations were thresholded based on those time/frequency/electrode points that were also significant in the wait trials. This enabled the difference maps to be constrained to identify patterns of activation that were significantly related to a trial type in question. We performed this step during the initial stages of identifying significant patterns of activation that seemed most relevant to the behavior at hand (i.e., as a means of screening out significant patterns in the contrast that may be less relevant to the behavior in question).

We calculated the mean analytic amplitude for each electrode across a specified time window and frequency band of interest (described in paper for different analyses). The mean/SEM of that data were calculated across sessions, and we used a one-sample, two-sided t test (again compared with the null hypothesis that Z = 0) to evaluate significance of the means. We applied a Bonferroni correction to the p values from this data (across 32 comparisons if only performed at one frequency band; or across 32*5 if performed across all five frequency bands, as described in results). Results from these follow-up analyses were reported in Extended Data, and were also performed (without correction) at the level of animals (11 animals).