Short answer: base rate neglect happens when a striking new signal receives too much weight while the event’s prior frequency receives too little. A detector can be accurate and still produce mostly false alarms when the event it seeks is rare.
That distinction matters in fraud monitoring, quality control, security alerts, hiring, and diagnostic reasoning. The useful question is not only “How accurate is the signal?” It is also “How common was the event before this signal appeared?”
Base rate neglect in one security example
Suppose a monitoring system examines 100,000 account events. Only 0.2% are genuine attacks. The system detects 95% of attacks and incorrectly flags 1% of normal events.
| Event group | Count | Flagged |
|---|---|---|
| Genuine attacks | 200 | 190 true alerts |
| Normal events | 99,800 | 998 false alerts |
| All alerts | — | 1,188 |
Among 1,188 alerts, 190 are genuine. Therefore, the probability that an alert represents a real attack is about 16%, not 95%. The detector is not useless: it shrinks 100,000 events to a list that investigators can prioritize. Yet an alert is evidence to examine, not a final verdict.

Why a good detector can have a low predictive value
Three quantities do different jobs:
- Base rate: how common the event is in the relevant population before new evidence.
- Sensitivity or hit rate: how often the signal appears when the event is truly present.
- False-positive rate: how often the signal appears when the event is absent.
People often reverse a conditional probability. “The signal appears in 95% of attacks” means P(signal|attack) = 95%. It does not mean P(attack|signal) = 95%. To reverse the direction, we must account for both the base rate and every route by which the signal can appear.
The arithmetic becomes much easier when percentages turn into counts. Start with a population large enough to avoid fractions. Split it into event and non-event groups. Apply the hit rate to the first group and the false-positive rate to the second. Finally, divide true alerts by all alerts.
Base rate neglect and Bayes’ rule
Bayes’ rule expresses the update:
P(A|B) = P(B|A) × P(A) ÷ P(B)
P(A) is the prior or base rate. P(B|A) describes the evidence if A is true. P(B) counts the overall chance of seeing the evidence, including false alarms. The NIST statistical handbook explains how prior information and current data combine to produce an updated assessment.
This equation does not demand blind faith in a single prior. Analysts should ask where the base rate came from, whether the population has changed, and how uncertain the estimate is. A prior taken from all users may be wrong for a carefully selected subgroup. Conversely, a high-risk subgroup should not define the rate for everyone.
If the notation feels abstract, revisit percentage calculations and then work with natural frequencies. For uncertainty across many variables, Monte Carlo simulation offers a complementary tool; it does not replace the need for sensible inputs.
Why the mind discounts prior probability
A vivid match feels diagnostic
A description can resemble a familiar stereotype so closely that we treat resemblance as proof. This is the representativeness shortcut. It is efficient in ordinary life, but it can obscure how many alternatives exist.
Large background groups stay invisible
False positives come from the much larger non-event group. Because that group lacks a dramatic story, it fades from attention. A simple two-row table restores the missing denominator.
The reported metric answers a different question
Vendors may emphasize recall, sensitivity, or accuracy. Each can be legitimate, but none alone answers “How many alerts are real?” A useful evaluation asks for the confusion matrix, the population, and the operating threshold.
Tversky and Kahneman’s classic article on judgment under uncertainty described representativeness, availability, and anchoring as economical mental shortcuts that can also create predictable errors. A 2022 study found that base-rate neglect generalizes across a broader set of problems, while also showing substantial differences between individuals.
Where base rates change real decisions
Fraud and cybersecurity
Rare-event detectors create a workload as well as a score. Teams need escalation rules, secondary evidence, and feedback from resolved cases. Otherwise, too many false alarms can exhaust reviewers and hide the truly urgent cases.
Manufacturing and quality control
A camera may flag surface defects with high sensitivity. If defects are uncommon, human inspection can still receive many clean parts. The practical goal is not a perfect headline metric; it is an operating point that balances missed defects, rework, delay, and inspection cost.
Hiring and admissions
A test result can add information without defining a person. Selection rates, subgroup differences, measurement error, and the cost of false rejection all matter. High-stakes decisions deserve multiple relevant signals and a process that allows review.
Health information
Probabilistic reasoning can help clinicians integrate findings, but an educational calculation is not personal medical advice. The National Academies’ account of the diagnostic process explains that sensitivity, specificity, prior probability, harms, benefits, and patient preferences all inform decisions. Individual results should be discussed with a qualified professional who knows the context.
A practical evidence checklist
- Name the target event. Avoid vague labels such as “risky” or “suspicious.”
- Choose the relevant population. The right denominator matches the place, time, and selection process.
- Find the prior frequency. Record its source and uncertainty.
- Separate hit rate from predictive value. They answer different conditional questions.
- Count false alarms. Apply the false-positive rate to the non-event group.
- Translate percentages into people or cases. Use 1,000, 10,000, or 100,000.
- Check for distribution shift. A model may encounter a new population after deployment.
- Plan the next step. Decide what evidence confirms, dismisses, or escalates an alert.
This checklist fits within broader risk-assessment techniques. It also helps future-proof automated decisions: as data, behavior, and model outputs change, teams should monitor not only accuracy but alert volume, subgroup effects, review capacity, and consequences.
Common mistakes—and better replacements
Mistake: treating 99% accuracy as 99% certainty. Better: ask which errors the accuracy measure combines and how imbalanced the classes are.
Mistake: importing a base rate from another population. Better: compare sampling rules, dates, locations, and definitions before reusing a rate.
Mistake: ignoring the cost of follow-up. Better: count how many alerts require human time and what happens after an alert.
Mistake: assuming a low posterior probability means “ignore it.” Better: connect probability to consequences. A 16% chance may justify immediate review when the harm is severe, while a much higher chance may still need confirmation before an irreversible action.
Quick review quiz
- A test catches 98% of true cases. Is a positive result 98% likely to be true?
No. The base rate and false-positive rate are still required. - What is the denominator when calculating the share of alerts that are genuine?
All alerts: true positives plus false positives. - Why do rare events create a special challenge?
The large non-event group can generate more false alarms than the small event group generates true alerts. - What does P(signal|event) describe?
The probability of seeing the signal when the event is present. - What does P(event|signal) describe?
The updated probability of the event after observing the signal. - Why use natural frequencies?
Counts make the competing sources of a positive result visible. - Can base rates change?
Yes. Population, time, selection, and behavior can all change them. - Does a false alarm make a detector worthless?
No. Its value depends on how it narrows the search and supports an appropriate follow-up process.
Conclusion
Base rate neglect is a mismatch between an attention-grabbing signal and the quieter probability that came before it. Sound reasoning keeps both in view. Start with the relevant population, count true and false alerts, update the prior with evidence, and match the next action to the stakes.
For another worked path through probability concepts, see everyday math formulas and examples. The companion Indonesian lesson approaches the same problem through a different worked example: Base Rate Fallacy: Mengapa Angka Akurat Bisa Menyesatkan.



