Errors, fast and slow

Human errors in cognitive, attentional, and decision-making tasks are sometimes faster than correct responses, and sometimes slower, even for the same fixed task and experimental implementation. Several existing models can fit response time distributions exhibiting these phenomena. However, it is hard to predict ex ante (i.e., before data collection) when errors will be fast or slow. Relying on 20 different datasets comprising 31 experiments from different domains, we empirically validate a simple nonparametric model which successfully predicts when errors will be faster or slower than correct responses. The predictions also include a generalized Stroop effect, as well as error rate differences. The model applies to generalized conflict tasks, where the interaction of multiple processes determines behavior, and makes predictions which depend on whether those processes are in alignment or conflict in a given trial, which can be determined before data collection (e.g., congruent vs. incongruent trials). This yields new testable hypotheses which are overwhelmingly supported in the data. The model's predictions can also be seen as a test of whether process multiplicity is a reasonable assumption in a given task.