Abstract
If all the drivers in your city were bad, would you be better at detecting dangerous events on the road? What if they were all good? In visual search, a low prevalence effect (LPE) has been found, in which observers frequently miss rare targets. These studies have used static stimuli, visible until response. In contrast, detections of road hazards often afford only brief glimpses of complex, dynamic scenes before decisions must be made. We tested the LPE with a novel road hazard detection task. Observers viewed brief (333 ms) video clips of road scenes recorded from dashboard cameras. These preserve the visual complexity of natural driving, while allowing control over event prevalence. In five online experiments (n=16 each), observers viewed our road scene clips and reported whether or not they saw a hazardous event on each trial. Using hazard prevalences of 50% and 4% in separate sessions, we replicate the LPE results from visual search; miss error rates were twice as frequent in the low prevalence condition compared to high prevalence (40 vs 18%, p < .001). This difference was attributable to a more conservative criterion in the low prevalence condition, while sensitivity (d’) was similar between conditions. Furthermore, miss rates increased as hazards became increasingly rare, down to 1% prevalence. Additional experiments showed that these results could not be explained by simple motor errors, since allowing observers to correct their responses did not impact the LPE. Finally, this effect persisted even when observers were explicitly pre-briefed about the LPE, indicating that simple cognitive interventions may not be effective at eliminating it. Together, our results demonstrate that the LPE generalizes to complex perceptual decisions in dynamic natural driving scenes, where observers must monitor and respond to rare hazards.