Multivariate pattern analysis
MVPA was applied to decode patterns of neural activity associated with different numbers of choices. A backward decoding classification algorithm (linear discriminant analysis) was used, with all electrodes as features. To ensure reliability and interpretability of the results, the choice set sizes of 4 and 8 were grouped together to create the small choice set condition, and 12 and 16 were grouped together to create the large choice set condition11Decoding was also conducted with four distinct choice set sizes, with results presented in the Supplementary Materials.. Before performing MVPA, the epochs were down sampled to 50 Hz to minimize computation time. A 10-fold cross-validation procedure was the applied using within-class and between-class balancing with the Amsterdam Decoding and Modeling toolbox (Fahrenfort et al., 2018). In this procedure, the trials were randomized and divided into 10 equal-sized folds. Nine folds were used for training, while the remaining fold was used for testing. This process was repeated 10 times, ensuring that each fold served as the test set once. To ensure the impartiality of the classifier training, we implemented within-class balancing by undersampling. This procedure involved the random selection of trials from conditions with a surplus of trials to harmonize the conditions with fewer trials, thereby equalizing the count of trials within each class. Additionally, between-class balancing using undersampling was employed to mitigate the potential of the classifier from developing a bias towards the overrepresented class during training, as unbalanced designs can often result in asymmetrical trial counts. The performance of the classifier was assessed using the area under the curve (AUC) (Hand & Till, 2001).
Temporal generalization analysis was conducted to assess the stability of a representation across different time points. A classifier trained on a specific time point was tested on all other time points (King & Dehaene, 2014). The resulting temporal generalization matrix was used to identify periods of stability. Additionally, the product of the classifier weights and the data covariance matrix was calculated and spatially normalized for each participant to obtain the activation patterns (Haufe et al., 2014). Cluster-based nonparametric statistical tests (two-tailed cluster-permutation, alpha p< 0.05, cluster alpha p < 0.05, N permutations = 5000) were used to evaluate the MVPA results using FieldTrip (Oostenveld et al., 2011).
Results