Signal Detection Theory

In brief

Signal detection theory (SDT) formalizes a problem classical psychophysics had gotten wrong for a century: when someone reports "yes, I detected it," how much of that answer reflects what they perceived, and how much reflects how willing they were to say yes at all. Wilson P. Tanner and John A. Swets stated the theory in 1954 in "A Decision-Making Theory of Visual Detection" (Psychological Review), building on parallel radar engineering work published the same year by Peterson, Birdsall, and Fox. David Green and Swets gave it its canonical statement in Signal Detection Theory and Psychophysics (1966). The theory splits every detection judgment into two independent numbers: sensitivity, how well a signal can be told apart from background noise, and criterion, how cautious or liberal the observer is about saying yes when the evidence is ambiguous. Two people, or the same person on two different days, can have identical sensitivity and still perform very differently because they set different criteria.

The full treatment

The problem it answers

Gustav Fechner's nineteenth century psychophysics assumed a fixed sensory threshold: below a certain stimulus intensity, nothing is perceived; above it, detection is automatic. World War Two radar operators showed the assumption was false. The same operator, watching the same screen for the same faint blip, reported far more detections when told that missing an incoming aircraft was catastrophic than when told that false alarms wasted scarce resources, even though the physical signal never changed. Classical threshold theory had no way to represent this: it treated detection as a fact about the stimulus and the eye, not a decision made under uncertainty. Tanner and Swets borrowed the statistical framework Neyman and Pearson had developed for hypothesis testing and applied it directly to perception.

How it works

SDT pictures both noise-alone and signal-plus-noise trials as producing an internal sense of evidence strength that varies trial to trial, drawn as two overlapping bell curves on one axis. Noise alone centers lower, signal plus noise centers higher, and the two overlap because noise sometimes looks strong and signal sometimes looks weak. The observer sets a criterion, a cutoff on that axis, and says yes whenever the evidence exceeds it, so the criterion's placement, not the physical stimulus alone, decides how often the observer says yes. Four outcomes follow from crossing the true state of the world against the response: a hit (signal present, says yes), a miss (signal present, says no), a false alarm (signal absent, says yes), and a correct rejection (signal absent, says no). Sensitivity is captured by d prime, the standardized distance between the two distributions, computed as the z-score of the hit rate minus the z-score of the false alarm rate. Bias is captured separately, as beta, the likelihood ratio at the criterion, or as c, the criterion's distance from the neutral point. Sweep the criterion from strict to lenient and plot hit rates against false alarm rates, and the result is a receiver operating characteristic curve. A fixed d' traces one such curve, and the area beneath it is itself a bias-free measure of sensitivity.

What it claims

The theory's central claim is that sensitivity and bias are logically separable, and that any single hit rate or raw accuracy score, taken alone, confounds them. A high hit rate can mean sharp perception or a reckless criterion that also yields many false alarms; a low hit rate can mean poor perception or an overly cautious criterion. Only by examining hits and false alarms together, or better, the full ROC curve traced across several criteria, can the two be pulled apart.

The key study or demonstration

Tanner and Swets ran detection experiments with faint tones embedded in noise and showed that manipulating the payoff for hits against the penalty for false alarms, or changing the prior probability of a signal, shifted hit and false alarm rates together exactly as a moving criterion predicts, while underlying d' stayed roughly constant. This dissociation, criterion moves with incentives and base rates while sensitivity does not, was the theory's decisive break from threshold accounts, replicated since across vision, hearing, and memory.

Related distinctions

SDT displaced high-threshold theory, the older view that a fixed proportion of signal trials cross a true perceptual threshold while the rest are pure guesses; SDT's continuous account fits the data better in most domains studied. Within SDT, Type 1 analysis covers the accuracy of the original judgment; Type 2 SDT extends the same logic to confidence, measuring how well stated confidence discriminates a person's own correct answers from errors.

Lineage

SDT descends from two independent 1954 sources: Tanner and Swets's psychological formulation and Peterson, Birdsall, and Fox's engineering paper "The Theory of Signal Detectability," both products of the University of Michigan's Electronic Defense Group, which had studied radar operator performance during and after the war. Both drew on the Neyman-Pearson framework of 1933 for choosing between statistical hypotheses, recasting detection as a choice between "noise alone" and "signal plus noise." It replaced, rather than extended, Fechner's threshold psychophysics of the 1860s.

The strongest case for it

SDT's strength is that it works across an unusually wide range of tasks without new machinery each time. The core dissociation, that identical sensitivity produces different behavior under different incentives and base rates, has been confirmed in vision, hearing, memory, and clinical judgment. It converts a folk complaint, that a radiologist finds too many false positives, into a precise measurement: is her eye worse, or is her threshold for calling "cancer" simply lower than a colleague's? The ROC curve it introduced is now infrastructure well beyond psychology: Charles Metz's work made it the standard tool for evaluating diagnostic tests in radiology, and ROC-AUC is the standard metric for evaluating classifiers in machine learning. Few mid-century psychological theories have exported this cleanly into engineering and medicine.

The strongest case against it

The simplest version of SDT assumes the noise and signal distributions are both normal and of equal variance, an assumption that often fails: empirical ROC curves from recognition memory are typically asymmetric in a way equal-variance SDT cannot produce, a mismatch usually patched by giving the signal distribution a larger variance, an add-on rather than a prediction. More seriously, Andrew Yonelinas argued in a 1994 paper that recognition memory ROC curves show a signature, a curved segment combined with a threshold-like discontinuity, that single-process SDT cannot fit, and that fits far better under a dual-process account: a graded familiarity signal SDT describes well, plus a separate, threshold-like recollection process it does not. SDT loyalists reply that a single unequal-variance process fits adequately; the argument is unsettled, but it exposes a real vulnerability, since with enough free parameters SDT can be bent to fit almost any ROC shape. Critics also note that real observers rarely hold one fixed criterion across a session; it drifts with recent trials in ways the basic model ignores.

Where it stands now

SDT is now default infrastructure rather than a live controversy. It underlies how diagnostic radiology validates screening tools, how memory researchers report recognition performance, and how John Wixted and Laura Mickes argued, in papers through the 2010s aimed at courts and forensic psychology, that eyewitness identification should be evaluated with full ROC curves rather than a single confidence-accuracy number. Outside psychology, its central export, the ROC curve, is now more often associated with machine learning than with its psychophysical origin.

Test yourself

Think of a recent moment when you had to decide whether something was really there: a strange noise at night, a hint of sarcasm in a text, a symptom worth calling a doctor about. Ask which part of your reaction was your actual read of the evidence, and which part was how cautious or bold you tend to be about crying wolf. Those are two different numbers, and confusing them is the most common way to misjudge your own judgment.