The dangerous search is not always the hard one.

Sometimes the target is visible. The observer knows what it looks like. The display stays on the screen until response. Feedback is given. The task is not a brief flash, not a subliminal trick, not a memory test.

And still, when the target becomes rare, people miss it.

This is the unfair part. Rarity sounds like a fact about the world, not a fact about perception. If a target is present, it is present. Why should its base rate change whether the eye finds it?

Because visual search is not only perception. It is a policy.

The observer must decide which items deserve scrutiny, how much evidence is enough to say “present”, and when to stop searching and say “absent”. Low prevalence does not merely make positive trials uncommon. It trains those decisions.

the rare target changes the observer

Seeing Is Only One Part

Visual search begins with an odd asymmetry. A feature target can pop out: the one red item among green items, the one vertical bar among horizontal bars. A conjunction target is different: the red vertical item among red horizontal and green vertical items. Treisman and Gelade’s feature-integration theory made this distinction famous. Features can be registered early and broadly, but binding features into object identities needs focused attention.1

Later visual-search models complicated the picture. Search is not simply “parallel” or “serial”. Attention is guided. Some items are more target-like, more salient, or more consistent with the current template, so they are checked earlier.2 Psychophysical accounts of search often frame the problem as noisy evidence across many locations or items, then ask how an observer combines that evidence into a decision.3

That last phrase is the hinge:

evidence into a decision

The evidence is not the response. An item can have target-like evidence without being called a target. A display can contain a target that is never inspected. A search can stop while useful evidence remains on the table.

Signal detection theory gives the old language for this distinction.4 There is sensitivity: how separated target and distractor evidence are. And there is criterion: how much evidence is required before the observer says yes. Move sensitivity and the distributions separate. Move criterion and the same evidence produces a different pattern of hits, misses, false alarms, and correct rejections.

Visual search adds a second criterion. Besides the criterion for saying “this item is a target”, there is a criterion for saying “I am done looking”.

That second criterion is where rarity gets under the skin of the task.

The Baggage-Screening Result

Wolfe, Horowitz, and Kenner tested an artificial baggage-screening task in which observers searched for tools among overlapping, semi-transparent objects on noisy backgrounds.5 The target prevalence was 50%, 10%, or 1%. The low prevalence condition was deliberately long: at 1%, each observer needed 2,000 trials to get about 20 target-present trials.

The result was not subtle.

At 50% prevalence, target misses were about 7%. At 10%, misses rose to about 16%. At 1%, they rose to about 30%. False alarms were extremely rare. The problem was not that people became reckless and said yes too often. The problem was that they said no too readily.

The reaction times were the clue. In many ordinary search tasks, absent trials take longer than present trials because a target can end the search early, while an absent response requires enough checking to justify giving up. At low prevalence, Wolfe and colleagues found the pattern could invert: absent responses became faster than present responses. The observer was not merely failing to recognize targets. The observer was quitting too soon.

The same paper contains a detail that prevents the easy explanation. Maybe rare targets are missed because the observer has little practice with them. But in a mixed-target experiment, some target was present on half the trials while one specific target type remained very rare. The very rare target was still missed far more often. Prevalence itself mattered, not just the absolute number of positive examples.

Later work made the effect look stubborn. Low target prevalence remained a source of search errors under variants that tried to make rare targets more available or more salient.6 Not every intervention fails, and not every task has the same effect size, but the baseline lesson is hard to ignore: the base rate is part of the task, not a footnote under it.

Missed After Looking, Missed Before Looking

There are two separable routes to a miss.

First, the observer can inspect the target and reject it. This is an item-level criterion problem. If the target evidence is noisy and the criterion is high, the target does not clear the bar.

Second, the observer can never inspect the target. This is a quitting problem. The target may be sitting in the display, but the search terminates before attention reaches it.

Wolfe and Van Wert argued that prevalence can reveal two dissociable decision criteria in visual search.7 In their framing, observers can shift an item criterion and a quitting threshold. These are not the same lever. A stricter item criterion says:

I looked at this, but it was not convincing enough.

A lower quitting threshold says:

I did not find anything quickly, so this is probably absent.

Low prevalence pushes both in plausible directions. If targets are rare, the prior odds of target presence are low, so a conservative observer should require stronger evidence before saying yes. At the same time, most target-absent responses are rewarded. Trial after trial teaches the motor system and the control system that “no” is usually correct. The search can become fast, fluent, and wrong on precisely the trials that matter.

This is not a contradiction. It is two criteria adapting to the same environment, one at the item and one at the exit.

A Toy Screener With Two Thresholds

The toy lab below has a display of noisy items. One target may be present. Every item receives an evidence value. The target, when present, has a higher mean evidence than distractors, controlled by d'. Items are inspected in a guided order: target-like items tend to be checked earlier, but guidance is noisy. The observer says “present” if an inspected item clears an item criterion. If not, the observer quits after a limited number of inspections and says “absent”.

The slider called prevalence pressure controls how much the observer adapts to the base rate. At zero, the policy is fixed across prevalence levels. At high values, low prevalence raises the item criterion and reduces the number of items inspected before quitting.

This is not a reproduction of any experiment. It is a small executable diagram of the two-criterion story.

Hits and target evidence Misses or strict item criterion False alarms Fixed policy Prevalence-adapted policy

Deterministic toy model. The numbers are qualitative, not estimates for airport screening, radiology, security work, or any individual observer.

Start with the default: 2% prevalence, a reasonably discriminable target, and a policy that adapts strongly to prevalence. Misses are not produced because the target has zero signal. They are produced because the observer often inspects too few items, and because the item criterion has moved upward.

Now drag prevalence pressure to zero. The miss rate drops, but the false alarm rate rises. This is the base-rate bargain in miniature. A fixed liberal policy catches more rare targets by treating many target-absent displays as suspicious.

Raise target prevalence toward 50%. The adapted policy becomes more willing to say yes and more willing to keep searching. Misses fall. False alarms rise. The target did not become more visible. The observer changed what a search is allowed to cost.

Lower search caution at 1% or 2% prevalence. The display panel becomes the important one. Rings mark the items inspected before response or quitting. A target can sit in the display and remain uninspected. This is not a failure of the retina. It is a failure of the stopping rule.

No Is an Action

Chun and Wolfe asked a deceptively practical question: how are searches terminated when there is no target present?8 A literal serial model would inspect every item and then quit. But people often terminate blank trials too quickly for that to be the whole story. Guided Search offers a different answer: items carry activations or target-likeness values, and the observer samples promising items until either a target is found or the remaining evidence falls below a quitting threshold.

That model makes the prevalence effect feel less mysterious. Low prevalence does not need to alter early vision. It can alter the value of continuing.

Imagine a searcher whose recent history is:

absent, absent, absent, absent, absent, absent, absent, absent, absent

After that history, a quick “no” is usually reinforced. It feels efficient. It keeps the queue moving. It is correct most of the time. The rare target is the case where that learned efficiency becomes a liability.

This is why “try harder” is not a theory of rare-target misses. The observer may already be trying. The problem is that effort is being allocated under a policy whose feedback is dominated by negative cases.

The Base Rate Trains the Operator

A low-prevalence task is an imbalanced dataset with a human in the loop.

Most trials teach the same response. Most rewards come from saying absent. Most errors that would reveal an over-fast stopping rule are unobserved because the target is absent. The system gets an enormous number of correct rejections and a tiny number of informative misses.

This looks familiar if you come from machine learning. A classifier trained on rare positives can become excellent at the majority class while silently damaging recall. The psychology version is sharper because the human is not just a classifier. The human chooses how long to compute.

In a model, class imbalance can shift a decision threshold. In visual search, class imbalance can also change computation time.

rare positive -> stricter yes criterion
rare positive -> cheaper early stopping
rare positive -> fewer chances to correct the first impression

That is the hidden cost of rare events. They do not merely appear less often. They provide less frequent corrective feedback for the very policy that must catch them.

Fix the Policy, Not Just the Poster

The naive fix is to add more positives. Training with enriched target prevalence can help people see more examples and calibrate what targets look like. But it does not automatically solve the low-prevalence operating problem. If production prevalence is 1%, the production policy still lives in a world where almost every “no” is rewarded.

A better design separates at least four things:

  1. Perceptual training: show enough varied targets that the template is not brittle.
  2. Criterion calibration: give feedback that makes misses visible, not just correct rejections.
  3. Stopping discipline: audit how long absent and present searches take, not only whether the final button press was correct.
  4. Prevalence-aware evaluation: report recall and false alarms by base-rate condition, not only pooled accuracy.

In operational systems, this can mean controlled target insertion, second reads, adjudication queues, slower review for high-impact negatives, or independent audits that estimate misses rather than counting only discovered positives. The right choice depends on domain costs. A false alarm in one setting is annoying; in another it is expensive, harmful, or politically unacceptable. A miss can be trivial or catastrophic.

The point is not “always lower the threshold”. The point is that the threshold exists, and the system has been teaching it.

A target-present response has a natural endpoint. You found something.

A target-absent response is different. It is a claim about what did not happen. That claim requires a stopping rule, and stopping rules are where institutions hide their assumptions.

How much of the image did the radiologist inspect? How long did the reviewer look at the transaction? How many candidate explanations did the investigator consider? How many logs did the engineer check before saying the outage is not in this subsystem? How many adversarial examples did the evaluator run before saying the model is safe enough?

The negative answer is not free. It is purchased with search.

Rare targets are dangerous because they make the purchase look overpriced most of the time. The system learns to save effort on negatives. The bill arrives on the rare positive.

That is the small moral of the prevalence effect:

absence is a decision, not an observation

And when the thing you seek is rare, the decision to stop becomes part of the thing you are measuring.

  1. Anne M. Treisman and Garry Gelade, “A Feature-Integration Theory of Attention,” Cognitive Psychology 12, no. 1, 1980, 97-136. DOI: 10.1016/0010-0285(80)90005-5. PubMed: 7351125. ↩

  2. Jeremy M. Wolfe and Todd S. Horowitz, “What attributes guide the deployment of visual attention and how do they do it?”, Nature Reviews Neuroscience 5, 2004, 495-501. DOI: 10.1038/nrn1411. PubMed: 15152199. ↩

  3. John Palmer, Preeti Verghese, and Michael Pavel, “The Psychophysics of Visual Search,” Vision Research 40, 2000, 1227-1268. DOI: 10.1016/S0042-6989(99)00244-8. PubMed: 10788638. ↩

  4. David M. Green and John A. Swets, Signal Detection Theory and Psychophysics, Wiley, 1966: Google Books record. See also Neil A. Macmillan and C. Douglas Creelman, Detection Theory: A User’s Guide, 2nd ed., Lawrence Erlbaum, 2005: publisher page. ↩

  5. Jeremy M. Wolfe, Todd S. Horowitz, and Naomi M. Kenner, “Rare items often missed in visual searches,” Nature 435, 2005, 439-440. DOI: 10.1038/435439a. Author PDF with supplementary methods: search.bwh.harvard.edu/new/pubs/Nature_suppl.pdf. ↩

  6. Jeremy M. Wolfe, et al., “Low Target Prevalence Is a Stubborn Source of Errors in Visual Search Tasks,” Journal of Experimental Psychology: General 136, no. 4, 2007, 623-638. DOI: 10.1037/0096-3445.136.4.623. ↩

  7. Jeremy M. Wolfe and Michael J. Van Wert, “Varying Target Prevalence Reveals Two Dissociable Decision Criteria in Visual Search,” Current Biology 20, no. 2, 2010, 121-124. DOI: 10.1016/j.cub.2009.11.066. PubMed: 20079642. Supplement: search.bwh.harvard.edu/new/pubs/Wolfe%2C%20VanWert10%20Supliment.pdf. ↩

  8. Marvin M. Chun and Jeremy M. Wolfe, “Just Say No: How Are Visual Searches Terminated When There Is No Target Present?”, Cognitive Psychology 30, no. 1, 1996, 39-78. DOI: 10.1006/cogp.1996.0002. PubMed: 8635311. ↩