Targeted-component analysis
As a complementary, ERP-focused approach, we examined the responses for all comparisons of interest in the N1, P2 and P3 time windows at the Fz, Cz and Pz electrodes. The windows were defined after visual inspection of the data by locating the highest negative or positive (depending on the component of interest) peak in the usual latencies and electrodes for each component, and defining a window centered on the peak and adjusted to the width of the component, as reported by previous works (SanMiguel et al., 2013). We observed morphological differences between responses to acquisition sounds and test sounds, thus windows were defined separately for the two types of sounds, based on the peaks observed in either the average of all acquisition sounds or all test sounds. The N1 was measured at the Fz electrode in the window 80-120 ms in acquisition sounds and 110-140 ms in test sounds. The P2 was measured at Cz in the window 180-240 ms in acquisition sounds and 210-270 ms in test sounds. Both components showed reversed polarity at the mastoid electrodes in these windows. The P3 component was measured at the Fz electrode (P3a) and the Pz electrode (P3b), respectively, in the 310-390 ms time-window in acquisition sounds and the 340-400 ms time-window in test sounds.
We ran repeated-measures ANOVAs in order to test for differences on the mean amplitude of each component at the selected electrodes between conditions of interest. Specifically, for the acquisition sounds, we ran a two-way ANOVA with the factors agency (two levels: agent and observer) and learning stage (two levels: early and late) for each ERP component of interest.
We ran two separate ANOVAs for each component for the test sounds. A one-way ANOVA with the factor movement-sound congruency (levels: congruent and incongruent) and a two-way ANOVA with the factors agency and learning stage. We also analysed whether the effects of congruency were modulated by agency and learning stage by using the differences between amplitudes in congruent and incongruent trials as the dependent variable in a two-way ANOVA with the factors agency and learning stage.
Correlation analysis
We aimed to identify electrophysiological markers related to the performance benefits associated with active learning amongst a small set of pre-defined candidate ERP components. Thus, we tested for significant correlations, using Pearson’s coefficient, between each participant’s effect of agency on the percentage of correct responses and the effect of agency on the N1, P2 and P3 in acquisition sounds (amplitude of acquisition sound ERPs in agent condition – observer condition). Given that agency effects on the behavioural data were restricted to the early learning stage (see results), we used only the performance data from the early learning stage (%Correct in agent–%Correct in passive) for this correlation analysis.

Behavioural results

Behavioural results show that active exploration led to faster learning and better memory performance. However, given enough training, passive viewing led to similarly good performance (fig. 2).
A two-way ANOVA was run with the factors agency (agent versus observer) and learning block (1 to 7) and the dependent variable %Correct.
We found a significant main effect of agency [F(1, 22)= 11.865, p < .001, ηp2 = 0.32] and learning block [F(6,132) = 16.534, p < .001, ηp2 = 0.39], and a significant interaction between the two factors [F(6,132) = 6.635, p < .001, ηp2 = 0.20]. Post-hoc t-tests (Bonferroni corrected) showed that memory performance in the agent condition was significantly higher than in the observer condition in the first, second and fourth learning block. The effect of agency on %Correct was inversely correlated with learning block (r(5) = -3.922, p = .01). Post-hoc t-tests showed that the difference between agent and observer was significantly larger in the first learning block than in the last learning block [t(22) = 4.291, p < 0.001, d = 1.105]. The effect of learning was significantly smaller in the agent compared to the observer condition [t(22) = -4.291, p < 0.001, d = -0.656] (subtracting learning block 1 from learning block 7 in agent versus observer condition). This shows that agency accelerates memory encoding for arbitrary audiovisual associations.