Signal Detection Theory

Abstract

Signal detection theory is a framework for measuring how an observer separates a faint signal from background noise when the two are confusable. Its central insight is that any detection performance mixes two quantities that raw accuracy conflates: sensitivity, the observer's capacity to tell signal from noise, and the criterion, the standard of evidence demanded before responding that a signal is present. By modelling the evidence as overlapping probability distributions, the theory extracts a bias-free sensitivity index, d′, and a separate criterion measure from the rates of hits and false alarms. The receiver operating characteristic curve traces how those rates covary as the criterion moves, and its shape reveals sensitivity independently of bias. This article surveys the theory's structure, its measures, and its reach across perception, memory, and diagnostic decision making.

Keywords: signal detection theory, sensitivity, response criterion, receiver operating characteristic, d-prime

Signal detection theory recasts every act of detection as a decision made under uncertainty. When a radiologist inspects a mammogram, a witness scans a lineup, or a listener strains for a tone in static, the evidence favouring “signal present” is never decisive; it is a graded quantity that noise alone can also produce. The observer must therefore set a standard and respond “yes” only when the evidence clears it. The theory's enduring contribution was to prise apart two things that ordinary measures of accuracy tangle together: how well the observer can discriminate signal from noise, and how willing the observer is to say “yes.” Wilson Tanner and John Swets imported this apparatus from electrical engineering into psychology in 1954, and the move reshaped how detection, discrimination, recognition, and diagnosis are measured (Tanner & Swets, 1954; Wixted, 2020).

What Signal Detection Theory Is

The theory begins by refusing a tempting but incoherent idea: that there is a fixed sensory threshold below which a stimulus produces no experience and above which it always does. Classical psychophysics leaned on such a threshold, yet observers detect the same faint stimulus on some trials and miss it on others, and they sometimes report a signal when none was presented. Signal detection theory explains this by positing that every trial, whether or not a signal is present, gives rise to some internal evidence, and that this evidence is noisy. On a noise trial the evidence is drawn from one probability distribution; on a signal trial, from a second distribution shifted toward stronger evidence. Because the two distributions overlap, no single value of evidence belongs unambiguously to one or the other (Swets et al., 1961).

The observer's task reduces to a rule: choose a point on the evidence axis, the criterion, and respond “yes, signal” whenever the evidence exceeds it and “no” otherwise. This single decision generates four possible outcomes, crossing what was presented with what was reported. Table 1 names them. Two are correct and two are errors, but the two error types are not interchangeable: a miss and a false alarm have different causes and, usually, different costs.

Table 1
The Four Outcomes of a Yes-No Detection Trial
Note. The hit rate and the false-alarm rate are the two numbers signal detection theory works from; the miss rate is one minus the hit rate, and the correct-rejection rate is one minus the false-alarm rate, so the two columns each supply a single free proportion (Macmillan & Creelman, 2005).

The crucial move is that these four outcomes are governed by two underlying parameters, not one. How far apart the signal and noise distributions lie fixes how many hits an observer can win for a given number of false alarms; that is sensitivity. Where the observer places the criterion decides how the errors divide between misses and false alarms; that is response bias. Raw percent-correct blends the two and can therefore reward a cautious observer with poor eyes and penalize a keen one who happens to say “yes” freely. Separating them is the whole point (Wixted, 2020).

Sensitivity and Criterion: d′ and c

When the two distributions are Gaussian with equal variance, the separation between their means, expressed in units of the common standard deviation, is the sensitivity index d′ (d-prime). It is recovered from the observed rates by converting each to a z score: d′ equals the z-transform of the hit rate minus the z-transform of the false-alarm rate. A d′ of zero means the distributions coincide and the observer is guessing; larger values mean cleaner separation and easier discrimination. Because both rates enter, d′ is unchanged by a shift of the criterion that trades hits for false alarms, which is precisely what makes it a measure of the observer rather than of the observer's willingness to respond (Stanislaw & Todorov, 1999).

The criterion is captured by its own statistic. The location measure c is the distance of the criterion from the point midway between the two distribution means, so that c equals zero for a neutral observer, is negative for a liberal one who says “yes” readily, and is positive for a conservative one who withholds assent. An equivalent measure, the likelihood ratio β, gives the height of the signal distribution relative to the noise distribution at the criterion; it is the currency in which the optimal criterion is naturally expressed. Figure 1 shows the two distributions, the criterion, and the regions whose areas are the hit and false-alarm rates.

Figure 1

Overlapping Noise and Signal Distributions Split by a Criterion

The demonstration below makes the independence tangible. Setting the separation controls sensitivity; sliding the criterion trades hits against false alarms while leaving d′ fixed, so that the same underlying acuity can produce very different-looking error profiles (Macmillan & Creelman, 2005).

The Receiver Operating Characteristic

Because the criterion can sit anywhere, a single observer with a fixed sensitivity does not produce one hit-and-false-alarm pair but a whole family of them. Plotting the hit rate against the false-alarm rate as the criterion sweeps from strict to lax traces the receiver operating characteristic, or ROC curve. Every point on the curve is the same observer with the same d′ adopting a different bias; movement along the curve is a change of criterion, and movement of the whole curve toward the upper-left corner is a change of sensitivity. The diagonal from corner to corner is chance performance, where hits and false alarms rise together in lockstep and d′ is zero. The more the curve bows toward the top-left, the greater the sensitivity, and the area under it has a clean interpretation: the probability that a randomly chosen signal trial yields stronger evidence than a randomly chosen noise trial (Swets et al., 2000).

The ROC is more than a summary; it is a test of the model. John Egan and colleagues showed that instead of running many sessions at different payoffs to sweep the criterion, an experimenter can ask observers to rate their confidence on each trial and read the whole ROC from a single session, treating each confidence level as an implicit criterion (Egan et al., 1959). The shape that results is diagnostic. Under the equal-variance Gaussian model the ROC is symmetric about the minor diagonal; real recognition-memory ROCs are often asymmetric, which forces the unequal-variance model in which the signal distribution is wider than the noise distribution, and this deviation is itself a substantive finding about the underlying evidence (Wixted, 2020). The same confidence ratings that trace the ROC also support a second-order, or type-2, analysis: treating confidence as a judgment about one's own correctness yields a metacognitive sensitivity, meta-d′, that measures how well confidence tracks accuracy and that can be compared directly against first-order d′ to quantify metacognitive efficiency (Fleming, 2024). The demonstration below builds the curve point by point as the criterion moves and reports the area beneath it.

Payoffs, Priors, and the Optimal Criterion

If sensitivity is fixed by the observer's perceptual system, the criterion is open to choice, and signal detection theory says exactly where a rational observer should place it. The optimal criterion is the likelihood ratio β that maximizes expected value, and it depends on two things outside the senses: the prior probability that a signal will occur, and the payoff structure of the four outcomes. When signals are rare, the optimal criterion rises, because most “yes” responses to weak evidence would be false alarms; when a miss is far costlier than a false alarm, as in screening for a dangerous disease, the criterion falls so that few signals slip through. The theory thus predicts that a fully sensible observer will change response bias as circumstances change, while holding sensitivity constant (Swets et al., 2000).

Human observers track these variables but rarely optimally. In visual search, lowering the prevalence of a target sharply raises the miss rate: when targets are rare, observers adopt a conservative criterion and fail to report targets that are plainly visible when common, a pattern with direct consequences for baggage screening and diagnostic radiology (Wolfe & Van Wert, 2010). More generally, observers are systematically suboptimal in perceptual decisions, deviating from the ideal criterion in lawful ways that a bias-aware analysis can quantify but a raw accuracy score would hide (Rahnev & Denison, 2018). Framing the criterion as a value-sensitive choice also connects detection to the emotional stakes of responding: the same evidence is judged against a stricter or looser standard depending on the costs the observer attaches to each error (Lynn & Barrett, 2014). The demonstration below computes the optimal criterion from a prior and a payoff matrix and shows expected value rising to a peak as the criterion is tuned.

From Detection to Decision: Reach and Neural Basis

The framework's power is that nothing in it is specific to the senses. Any situation in which an observer must sort evidence into two categories under uncertainty is a signal detection problem, and the same two-parameter analysis applies. In recognition memory, “old” versus “new” judgments produce hits and false alarms, memory strength plays the role of internal evidence, and ROC analysis has become the standard tool, with the asymmetry of memory ROCs driving debates about whether recognition rests on a single continuous signal or on separate familiarity and recollection processes (Kellen et al., 2021). In medicine, ROC analysis quantifies the accuracy of a diagnostic test independently of the decision threshold a clinic adopts, letting the sensitivity of a test be assessed separately from the criterion policy applied to its results (Swets et al., 2000). The theory has organized work on eyewitness identification, personnel selection, weather forecasting, and industrial inspection, wherever a yes-no judgment must be graded for both acuity and bias.

The account also has a neural reading. Single-neuron recordings during perceptual decisions show sensory neurons accumulating evidence toward a bound, so that the abstract criterion of signal detection theory corresponds to a firing-rate threshold at which a choice is committed (Gold & Shadlen, 2007). This links the static, trial-level description of detection to the dynamics of how a single decision unfolds in time, and it places the criterion inside the brain as a tunable quantity rather than a mere bookkeeping device. The textbook treatment of these measures and models remains a standard reference for the equal- and unequal-variance cases alike (Wickens, 2002).

Worked Example

Signal detection theory turns a pair of proportions into two interpretable numbers. Suppose an observer in a yes-no task achieves a hit rate of .933 and a false-alarm rate of .309. Percent correct, assuming equal numbers of signal and noise trials, is the average of the hit rate and the correct-rejection rate: (0.933 + 0.691) / 2 = 0.812, a figure that says nothing about whether the observer is acute, biased, or both.

The signal detection analysis separates the two. Converting each rate to a z score gives z(0.933) = 1.5 and z(0.309) = −0.5. Sensitivity is their difference: d′ = 1.5 − (−0.5) = 2.0, a clean two standard deviations of separation between the noise and signal distributions. The criterion location is c = −0.5 × (1.5 + (−0.5)) = −0.5, a negative value marking a liberal observer who leans toward saying “yes” and so collects many hits at the price of a substantial false-alarm rate. Two observers could share this 81% accuracy with entirely different d′ and c; only the signal detection decomposition tells them apart, which is why the analysis, and not the accuracy score, is what the demonstrations above compute (Stanislaw & Todorov, 1999).

Discussion

Signal detection theory succeeded because it replaced a false picture of perception with a workable one. The threshold it displaced treated sensation as all-or-none and left the observer's decisions invisible; the theory made those decisions the object of measurement and, in doing so, turned response bias from a nuisance into a quantity (Tanner & Swets, 1954; Swets et al., 1961). Its two-parameter decomposition is now so routine across psychology that its origin is easy to forget, and the discipline sometimes rediscovers, under other names, the distinction between sensitivity and bias that the theory drew once and for all (Wixted, 2020).

Its limits are the limits of its assumptions. The equal-variance Gaussian model is an idealization; when it fails, as it visibly does for recognition memory, the failure is informative rather than fatal, pointing to unequal variances or to multiple underlying processes (Kellen et al., 2021). The deeper open question is not whether the framework describes the data but why observers set the criteria they do: they are demonstrably suboptimal, and characterizing the lawful shape of that suboptimality is where much current work sits (Rahnev & Denison, 2018). The theory remains, six decades on, the common language in which detection, discrimination, recognition, and diagnosis are all posed.

Current Directions

Three lines extend the classical framework. The first embeds the static criterion in a dynamic, neural account: sequential-sampling models treat the evidence as accumulating over time to a bound, unifying the yes-no decision of signal detection theory with the reaction-time structure of the choice and grounding the criterion in the firing of evidence-accumulating neurons (Gold & Shadlen, 2007). The second confronts systematic suboptimality head-on, cataloguing the ways human criteria depart from the ideal observer and building models in which those departures are themselves predictable, so that bias becomes a signal about the decision process rather than error to be averaged away (Rahnev & Denison, 2018).

A third strand widens the theory's scope to affect and ecological validity. Reframing detection around the costs and benefits an organism attaches to each outcome links the criterion to emotion and motivation, and clarifies why the same physical evidence is judged differently when the stakes change (Lynn & Barrett, 2014). Applied work continues to press the prevalence problem, seeking display and training interventions that keep the criterion from drifting when targets are rare, with real consequences for screening and diagnosis (Wolfe & Van Wert, 2010). Across all three, the direction is the same: from a fixed pair of numbers toward a process account of how a criterion is set, adjusted, and mis-set.

Glossary

Bias (response bias).

An observer's general tendency to favour one response over the other independent of sensitivity; in signal detection theory it is the placement of the criterion, measured by c or β.

Correct rejection.

A trial on which no signal is present and the observer correctly responds “no”; its rate is one minus the false-alarm rate.

Criterion location (c).

A bias measure equal to the signed distance of the criterion from the midpoint between the distribution means; zero is neutral, negative is liberal, positive is conservative.

Criterion.

The point on the internal-evidence axis above which the observer responds “signal present”; its location sets the trade-off between hits and false alarms without changing sensitivity.

d′ (d-prime).

The sensitivity index: the separation between the noise and signal distribution means in standard-deviation units, computed as the z-transformed hit rate minus the z-transformed false-alarm rate.

False alarm.

A trial on which no signal is present but the observer responds “yes”; its rate is one of the two proportions the analysis works from.

Hit.

A trial on which a signal is present and the observer correctly responds “yes”; the hit rate paired with the false-alarm rate fixes both d′ and the criterion.

Ideal observer.

A hypothetical decision maker that sets the criterion to maximize expected value given the priors and payoffs; the benchmark against which human suboptimality is measured.

Likelihood ratio (β).

The ratio of the signal to the noise distribution height at the criterion; the natural unit for the optimal criterion, which rises as signals become rarer or false alarms costlier.

Miss.

A trial on which a signal is present but the observer responds “no”; its rate is one minus the hit rate, and its cost drives the criterion down when misses are dangerous.

Noise distribution.

The probability distribution of internal evidence on trials when no signal is present; its overlap with the signal distribution is what makes detection a decision.

Prevalence.

The prior probability that a signal occurs; low prevalence raises the optimal criterion and, in practice, inflates miss rates in tasks such as screening and search.

Receiver operating characteristic (ROC).

The plot of hit rate against false-alarm rate across all criteria for a fixed sensitivity; its bow toward the upper-left corner indexes d′ and the area beneath it summarizes discriminability.

Sensitivity.

The observer's capacity to discriminate signal from noise, reflected in the separation of the two distributions and quantified by d′ independently of response bias.

Signal distribution.

The probability distribution of internal evidence on trials when a signal is present; it is shifted toward stronger evidence than the noise distribution and, in memory, is often wider.

Threshold (classical).

The pre-detection-theory notion of a fixed sensory boundary below which no stimulus is perceived; signal detection theory replaced it with a movable decision criterion over continuous, noisy evidence.

Unequal-variance model.

A version of the theory in which the signal distribution has a larger standard deviation than the noise distribution, needed to fit the asymmetric ROC curves seen in recognition memory.

Key Researchers

C. Douglas Creelman (1933-2014). Psychophysicist at the University of Toronto; with Macmillan he wrote Detection Theory: A User's Guide, the handbook that made d′ and criterion estimation standard practice across experimental psychology. Detection Theory (2005)

David M. Green (1932-2022). Auditory psychophysicist at UC San Diego, Harvard, and the University of Florida; with Swets he wrote the 1966 monograph that carried signal detection theory from radar engineering into hearing science. Wikidata

Neil A. Macmillan (contemporary). Professor Emeritus of Psychology at Brooklyn College, CUNY; with Creelman he authored the standard modern reference on detection theory and its measures. Detection Theory (2005)

Dobromir Rahnev (contemporary). Associate Professor of Psychology at the Georgia Institute of Technology; with Denison he synthesized the evidence that perceptual decisions are systematically suboptimal relative to the ideal observer. Faculty Page - ORCID

Michael N. Shadlen (b. 1959). Professor of Neuroscience at Columbia University and Howard Hughes Medical Institute investigator; with Gold he established how sensory neurons accumulate evidence to a bound, giving the detection criterion a neural realization. Faculty Page - ORCID

John A. Swets (1928-2016). Psychophysicist at BBN Technologies and Harvard University; a co-founder of the psychological application of signal detection theory who later championed ROC analysis for diagnostic decisions in medicine. Wikipedia

Wilson P. Tanner (1912-1977). Engineer and psychophysicist at the University of Michigan; with Swets he authored the 1954 paper that recast visual detection as a statistical decision, importing the theory of signal detectability into psychology. Psychological Review (1954)

John T. Wixted (b. 1958). Distinguished Professor of Psychology at the University of California, San Diego; he recovered the early history of signal detection theory and applied its ROC machinery to recognition memory and eyewitness identification. Faculty Page - ORCID

Frequently Asked Questions

What is signal detection theory?
Signal detection theory is a framework for measuring how an observer distinguishes a signal from background noise when the two overlap. It models each trial as producing noisy internal evidence and treats the observer's report as a decision made by comparing that evidence to a criterion, separating perceptual sensitivity from response bias (Tanner & Swets, 1954).

What is the difference between sensitivity and bias?
Sensitivity is how well an observer can tell signal from noise, fixed by the separation of the two evidence distributions and measured by d′. Bias is how willing the observer is to respond “yes,” fixed by the placement of the criterion. Two observers with identical sensitivity can differ in bias, and raw accuracy cannot tell them apart (Macmillan & Creelman, 2005).

What is d-prime?
The sensitivity index d′ is the distance between the noise and signal distributions measured in standard-deviation units. It is computed as the z-transformed hit rate minus the z-transformed false-alarm rate, and it stays constant when the observer shifts the criterion, which is what makes it a pure measure of discriminability (Stanislaw & Todorov, 1999).

What are hits, misses, false alarms, and correct rejections?
They are the four outcomes of a yes-no trial, crossing whether a signal was present with whether the observer said “yes.” A hit and a correct rejection are the two correct responses; a miss and a false alarm are the two errors. Signal detection theory works from the hit rate and the false-alarm rate (Wixted, 2020).

What is an ROC curve?
The receiver operating characteristic curve plots the hit rate against the false-alarm rate across every possible criterion for a fixed sensitivity. It bows toward the upper-left corner as sensitivity rises, its diagonal marks chance, and the area beneath it is the probability that a signal trial yields stronger evidence than a noise trial (Swets et al., 2000).

How do rewards and probabilities change the criterion?
The optimal criterion depends on the prior probability of a signal and on the payoffs for each outcome. Rare signals push the criterion up, and a costly miss pushes it down, so a sensible observer changes response bias as circumstances change while leaving sensitivity untouched (Swets et al., 2000).

Why do rare targets get missed more often?
When targets are rare, observers adopt a conservative criterion and withhold “yes” responses, so they miss targets that they detect easily when targets are common. This prevalence effect has direct consequences for baggage screening and diagnostic radiology (Wolfe & Van Wert, 2010).

Where is signal detection theory used beyond perception?
Anywhere a yes-no judgment is made under uncertainty. It is the standard analysis for recognition memory, medical diagnosis, eyewitness identification, weather forecasting, and industrial inspection, because in each the same two parameters, sensitivity and criterion, govern performance (Kellen et al., 2021).

References