We used permutation tests to assess whether the population CPDs for each predictor at each time-point were significantly larger than expected by chance: We generated an ensemble of 5000 permuted datasets by circularly shifting the predictors relative to the neural activity by a random number of trials drawn independently for each session from the range [0, N] where N is the number of trials in the session. This permutation preserves the autocorrelation across trials in both the neural activity and the predictors but randomizes the relationship between them. We calculated P values for each predictor at each time point as the fraction of permutations for which the permuted datasets had a larger CPD than the true dataset. P values for each predictor were corrected for multiple comparison across time-points using the Benjamini–Hochberg procedure (Benjamini and Hochberg, 1995).
