Sensation and Perception
Signal Detection Theory
Separating sensitivity from bias — how observers decide under noise
Signal detection theory (SDT) decomposes a detection judgment into two components: sensitivity (how well the observer discriminates signal from noise) and bias (how willing they are to say "yes"). Developed by Tanner and Swets in the 1950s from radar engineering, SDT replaced the older "absolute threshold" concept. Every detection produces four possible outcomes: hit, miss, false alarm, correct rejection. The sensitivity index d-prime captures pure ability; the criterion captures willingness to respond. SDT has transformed medical screening (mammography), eyewitness identification, lie detection, and any task involving discrimination under noise.
- FoundersWilson Tanner and John Swets (1954)
- OriginRadar signal detection in WWII
- Core measuresd-prime (sensitivity), criterion (bias)
- Four outcomesHit, miss, false alarm, correct rejection
- ROC curvePlot of hit rate vs false alarm rate across criteria
- ApplicationsMedical screening, eyewitness ID, lie detection, memory
Interactive visualization
Press play, or step through manually. The visualization is yours to drive — try it before reading on.
Watch the 60-second explainer
A condensed visual walkthrough — narrated, captioned, under a minute.
Why signal detection theory matters
- Medical screening. Mammography, COVID tests, biopsies all involve hit/false-alarm trade-offs.
- Eyewitness identification. Sequential vs simultaneous lineups change criterion, not sensitivity.
- Memory research. Recognition memory analyzed via d-prime separates encoding strength from response bias.
- Air traffic control. Operators must detect rare events under heavy noise.
- Industrial inspection. Defect detection on production lines uses ROC analysis.
- Lie detection. Polygraph and behavioral cues evaluated as noisy signals.
- Machine learning. ROC and AUC are standard classifier metrics.
Common misconceptions
- Hit rate is sensitivity. Without false alarm rate, you can't separate sensitivity from bias.
- A liberal observer is more sensitive. They have more hits but also more false alarms — d' may be identical.
- Detection has a fixed threshold. Threshold theory was replaced; sensitivity is continuous.
- Better tests have lower false alarm rates. Better tests have higher d-prime; false alarms depend on criterion.
- Observers can't change bias. Bias responds to payoffs, base rates, and instructions.
- ROC is just a plot. ROC encodes the entire performance trade-off; AUC summarizes it bias-free.
Frequently asked questions
What's d-prime?
D-prime (d') is the standardized distance between the means of the noise and signal-plus-noise distributions, in standard deviation units. It measures pure sensitivity, independent of response bias. A d' of 0 means the observer cannot tell signal from noise; d' of 1 is moderate sensitivity; d' above 3 is excellent. Computed as z(hit rate) - z(false alarm rate). Two observers with very different "yes" rates can have identical sensitivity.
What's the criterion?
The decision threshold the observer adopts. A liberal criterion (low threshold) produces many "yes" responses — high hits but also high false alarms. A conservative criterion (high threshold) produces few "yes" responses — fewer false alarms but also fewer hits. Observers can shift the criterion based on payoffs (rewards for hits, costs of misses) and base rates. Sensitivity is fixed; bias is choosable.
What are the four outcome types?
Cross-tabulating actual signal presence with response: (1) Hit — signal present, said yes. (2) Miss — signal present, said no. (3) False alarm — signal absent, said yes. (4) Correct rejection — signal absent, said no. SDT analyzes the proportions in each cell. The "hit rate" is hits / (hits + misses); the "false alarm rate" is false alarms / (false alarms + correct rejections).
What's an ROC curve?
Receiver Operating Characteristic. Plot hit rate (y) against false alarm rate (x) as the criterion shifts from very conservative (origin) to very liberal (top-right). For a given sensitivity, the curve bows away from the diagonal. The area under the ROC (AUC) is a bias-free sensitivity measure. ROC curves are standard for comparing classifiers — medical tests, fingerprint match systems, machine learning models.
Why does it matter for medical screening?
Mammography misses cancers (false negatives) and flags benign lesions (false positives). Adjusting the criterion changes the trade-off but not the sensitivity. Reducing missed cancers requires lowering the criterion — at the cost of more biopsies for benign findings. The ROC curve quantifies the screening test's intrinsic quality. Different countries make different criterion choices based on cost-benefit.
How does it apply to eyewitness identification?
Lineup decisions face the same hits/misses/false-alarms structure. Wells and Lindsay's research showed that simultaneous lineups (suspects together) elicit a more liberal criterion than sequential ones (one at a time), affecting false-identification rates. The reform recommendation — sequential, double-blind administration — is a criterion-shifting intervention. SDT analysis of lineups has informed legal reform in multiple US states.
How is it different from old threshold theory?
Classical psychophysics assumed a fixed sensory threshold below which signals are undetectable, above which detection is perfect. SDT replaced this with continuous probabilistic distributions plus a decision criterion. This explains why hit rates vary with payoff (which threshold theory could not), why d-prime is constant across criterion shifts, and why the same observer can appear "sensitive" or "insensitive" depending on bias.