Errors in low prevalence visual search: Easy to produce, hard to cure

Michael J. Van Wert 1, Todd S. Horowitz 1,2, Skyler S. Place 1, Jeremy M. Wolfe 1,2

1 - Brigham & Women's Hospital
2 - Harvard Medical School

In standard visual search paradigms, targets are present on 50% of trials. Many important search tasks (e.g., medical and baggage screening) involve much lower target prevalence. Previously, Wolfe, Horowitz, and Kenner (Nature, 2005) reported that a task yielding 7% miss errors at 50% prevalence produced 30% errors at 1% prevalence. Their task was quite artificial. Here we describe a replication using x-ray images of baggage, as well as two efforts to reduce errors. Stimuli were virtual bags assembled from actual x-ray images of weapons and other objects. Bags contained 3, 6, 12, or 18 objects. Target prevalence was 50% or 1-2%. Observers searched for guns and knives. The CONTROL condition replicated the prior study with these realistic stimuli: miss errors were 18% at 50% prevalence, 40% at 1% prevalence. At low prevalence, target-absent reaction times (RTs) become too fast. Thus, the SPEEDING TICKET condition warned observers when absent RTs dropped below a threshold based on the observer's average target-present RT. These warnings slowed absent RTs by about 700 msec but failed to reduce errors (40% at 2% prevalence). The DUAL OBSERVER condition showed two independent observers the same sequence of bags. If errors were independent, joint error rates would be the product of individual rates. Joint error rates were slightly lower in DUAL OBSERVER conditions. The extra errors produced at low prevalence are not independent between observers. Elevated errors at low prevalence are easy to produce, potentially dangerous, and hard to ameliorate.