The time course of neural activity predictive of self-initiated movement

Decades of research have shown that early ramping activity is a very reliable antecedent of self-initiated movement. However, the dominant paradigm has been to only analyze data epochs culminating in movement, which discounts how often slow ramping might occur when no movement follows. To address this, we introduced a matched control condition that did not culminate in movement. We recorded electroencephalography data and compared these two conditions using a variety of powerful machine-learning classifiers. Although early ramping activity was present, it did not predict movement. Instead, classification accuracy rose abruptly about 100 ms before movement. By contrast, the traditional approach reproduced the spurious impression of early predictability. Our results resolve a long-standing controversy, showing that the neural commitment to act is a late-stage event consistent with the timing of conscious intention and the ability to inhibit movement. More broadly, our results show that a reliable antecedent is not necessarily a good predictor, underscoring the need for proper control conditions in time series analyses.