Detection falls off when the event rate is low

What it is

Poorer detection under low event rates is actually two effects layered together, and they are often conflated. The vigilance decrement is the drop in detection over time on watch even when the proportion of events stays fixed. The low-prevalence effect is lower detection whenever targets make up a small share of all trials, regardless of how long the person has been watching. Industrial monitoring usually has both at once — alarms are rare by nature, and shifts are long — so field performance is often worse than either mechanism alone would predict.

Why it happens

In signal-detection terms, performance is set by two independent quantities: sensitivity (whether the signal can be told apart from noise at all) and criterion (how much evidence is required before calling "signal"). Two explanatory paths lead to different fixes:

• Criterion shift. As the prior probability of an event drops, observers unconsciously raise the criterion — the subjective cost of a false call (looking trigger-happy) starts to outweigh the cost of a miss, especially when the miss has no immediate visible consequence. Sensitivity is untouched here; only the threshold moves.
• Resource depletion. Tasks that require holding a value in memory and comparing against it (a successive discrimination — "is this reading higher than the last one?" — rather than a simultaneous one, "which of these two channels differs?") show a genuine decline in sensitivity as time on watch accumulates. No amount of criterion adjustment fixes this kind.

Which mechanism dominates depends on whether the task is successive or simultaneous discrimination, and that in turn decides the remedy: rebalance the payoff and reporting incentives, or shorten the unbroken watch period and rotate more often. Artificially raising the apparent event rate can pull the criterion back down, but if the injected signals are indistinguishable from real alarms, operators eventually notice that "the alarm is often fake," and the criterion drifts the other way — creating a new miss or sluggish-response problem instead of solving the original one.

Studying it

Signal-detection analysis converts hit rate and false-alarm rate into two independent measures — sensitivity (d′) and criterion (c or β) — and separates a criterion shift from a true sensitivity loss by manipulating the payoff matrix for misses versus false alarms and watching whether the criterion moves in response. Classic paradigms sit in the sustained-attention task lineage and are typically built as either successive or simultaneous discrimination.

One methodological caution: a single lab session usually runs under an hour, and to collect enough data in that window the designed event rate is often set far above what any real post actually sees. That kind of data validates the mechanism but is systematically optimistic if used to estimate real-world detection rates. Establishing whether the effect is present at a specific post requires splitting the same operators' logged responses into early-shift and late-shift segments, not reading off a single session's average.

Where it stops holding

• It only applies when the true event rate is genuinely low; if a post's anomalies are not actually rare (frequent findings during routine rounds, say), something else is the more likely explanation.
• Criterion shift can be partly offset by incentives and training, but resource-depletion sensitivity loss cannot — if the task is successive discrimination, the only real lever is shortening the unbroken watch, not reminders to "stay sharp."
• Lab watch periods are usually under an hour; whether the same relationship holds across a full eight-to-twelve-hour shift is not established, since rotation, breaks, and task-switching interrupt the accumulation.
• This describes a systematic drop in detection probability with event rate and time on watch, not any single miss — one bad call is not evidence for or against it.

Applying it

• Do not set the event rate in training or simulation much higher than the real post experiences: trainees who get used to an inflated rate will see their criterion drift conservative again once on the job, making the decrement worse, not better.
• Where simulated signals are needed to keep the apparent rate up, keep a design-level record that separates them from genuine alarms (revealed only in debrief) so operators are never left guessing, mid-shift, whether an alarm is real.
• Match the fix to task structure: shorten the unbroken watch and rotate more often for successive-discrimination posts; rebalance the cost of misses versus false calls and reinforce it through training for simultaneous-discrimination posts.
• How to check: pull historical alarm-response logs and compare hit rate and response latency between the first two hours of a shift and the remainder, alongside an estimate of the actual event rate in each window. A late-shift drop in hit rate with an unchanged event rate points to a vigilance decrement; a uniformly low hit rate across both windows points to a low-prevalence criterion shift — and the two call for different fixes.

Related

• Same group: Y1.04.2 Limits of passive visual monitoring · Y1.04.3 Active engagement in supervisory control
• Nearby: Y2.09 Alarm fatigue and false-alarm cost · Y6.05 Fatigue, Shift Work, and Circadian Rhythm
• Search terms: vigilance decrement · low-prevalence effect · signal detection theory · Mackworth clock test