Signal Detection Theory: Sensitivity, Decisions, and Response Bias

Author: Angela Bogdanova · Published: October 16, 2026 · Editorial Policy

Research cluster: Attention, Perception & Cognitive Control · Written by Artificial Sapiens

What Is Signal Detection Theory?

Signal detection theory (SDT) is a quantitative framework for studying decisions made when the available evidence is uncertain. It separates two components that ordinary accuracy scores often confound: sensitivity, or the degree to which the evidence distinguishes a target from its absence, and response criterion, or the amount of evidence required before an observer reports that the target is present. The same observer can detect many signals and produce many false alarms, or report fewer signals and commit fewer false alarms, without an equivalent change in the underlying ability to discriminate. This separation is the central contribution of signal detection theory to experimental psychology (Tanner & Swets, 1954; Stanislaw & Todorov, 1999).

The framework applies to identifying a faint tone in noise, recognizing a previously studied word, deciding whether a medical image contains an abnormality, detecting a hazardous event, and classifying an ambiguous visual object. These tasks differ substantially in their sources of evidence and their consequences, but each requires a response in the face of overlap between evidence generated by two possible states. SDT models the evidence distributions and the rule used to choose a response; it does not presume that every task depends on one specific neural mechanism.

An accessible definition is this: signal detection theory explains observable detection and classification decisions by estimating how distinguishable two evidence distributions are and where an observer places a decision threshold. Its predictions concern patterns of hits, false alarms, misses, and correct rejections across the task, rather than a diagnosis or a fixed capacity inferred from one mistake.

Origins: From Psychophysics to Decision Theory

Classical psychophysics examined how stimulus strength relates to detection and discrimination. A traditional absolute threshold, such as the intensity noticed on half of presentations, summarizes behavior but can change when observers become more willing to say that something was present. Mid-twentieth-century detection research made the decision component explicit. Tanner and Swets described the behavior of observers detecting visual signals against a background, and Green and Swets subsequently consolidated the mathematical and experimental framework in Signal Detection Theory and Psychophysics (Tanner & Swets, 1954; Green & Swets, 1966).

The decisive insight was that two observers with similar sensory evidence may use different decision policies. Instructions emphasizing detection, penalties for missing targets, monetary incentives, and expectations about target frequency can alter the criterion. A measured threshold therefore reflects both the distribution of internal responses and the observer's response strategy. Modern SDT retains this insight while extending it to confidence, memory recognition, clinical classification, animal behavior, and computational studies of decision making (Swets, 1988; Maniscalco & Lau, 2012).

The Four Outcomes of a Detection Decision

In a standard yes–no experiment, each trial has an objectively defined target state: the signal is either present or absent. The participant gives one of two responses: 'signal present' or 'signal absent.' Combining the stimulus state with the response produces a two-by-two table. A hit occurs when the signal is present and the response is yes. A miss occurs when the signal is present and the response is no. A false alarm occurs when the signal is absent but the response is yes. A correct rejection occurs when the signal is absent and the response is no.

These four cells are exhaustive for a completed binary trial with known ground truth, but their interpretation depends on correct labeling and design. A laboratory experiment can independently randomize signal-present and signal-absent trials. In an observational or real-world setting, the true state may be incompletely verified or defined by an imperfect reference standard. When ground truth is uncertain, estimated performance inherits that uncertainty. The four-outcome table remains useful, but the confidence in its conclusions must be justified by the quality of the task (Swets, 1988).

Hit rate and false-alarm rate are conditional probabilities

The hit rate H is the number of hits divided by the number of signal-present trials. The false-alarm rate F is the number of false alarms divided by the number of signal-absent trials. Miss rate equals 1 − H, and correct-rejection rate equals 1 − F. These denominators matter: dividing hits by all trials would mix sensitivity with signal prevalence. A participant might receive 100 target trials and correctly identify 80, giving H = 0.80, while also receiving 100 noise trials and incorrectly responding yes on 20, giving F = 0.20.

Overall percent correct for that balanced example is 80%, but this figure alone omits the trade-off between kinds of errors. With a rare signal, an observer can achieve extremely high overall accuracy simply by responding no on nearly every trial. For a task with only 1% signal-present trials, always saying no yields 99% accuracy while detecting none of the targets. Reporting H and F separately prevents this misleading description of performance.

How the Internal Evidence Model Works

The basic model assumes that each trial produces a value on an internal evidence axis. In a faint-tone experiment the evidence might relate to the strength of a noisy auditory representation; in a recognition task it might be a familiarity signal. Because sensory and cognitive responses vary from trial to trial, identical physical conditions need not generate identical evidence values. Noise trials form one evidence distribution and signal-plus-noise trials form another. When the distributions overlap, some trials will produce ambiguous evidence.

The observer places a criterion on this axis. Evidence above the criterion leads to a 'yes' decision, and evidence below it leads to a 'no' decision. The area of the signal distribution above the criterion determines the hit rate. The area of the noise distribution above the same criterion determines the false-alarm rate. Moving the criterion changes the two rates together; separating the distributions makes discrimination easier. The mathematical model is compact, and its assumptions must be evaluated separately from the observable four-outcome counts (Stanislaw & Todorov, 1999).

Noise is variability, not necessarily a distracting sound

In SDT, noise refers to the evidence distribution when the target is absent and, more broadly, to variability that limits discrimination. It can arise from spontaneous sensory activity, physical background variation, changing stimulus properties, perceptual uncertainty, attention fluctuations, memory variability, or imperfect measurement. The model often combines multiple sources of variability into one statistical description. Calling it noise does not establish which physiological or psychological source produced it.

A signal can be genuinely present while remaining difficult to distinguish from the alternative state. Conversely, an unusually strong noise trial may resemble a signal and elicit a false alarm. SDT therefore explains why both error types can occur under fully rational or well-practiced decision policies, depending on uncertainty and the task's costs.

Sensitivity: What d′ Measures

The most widely taught sensitivity index is d′, pronounced 'd-prime.' Under an equal-variance Gaussian model, d′ represents the distance between the mean of the signal and noise evidence distributions in units of their common standard deviation. A larger positive d′ corresponds to better separation and, within this model, stronger ability to distinguish the two states. A d′ near zero means that measured responses do not reliably distinguish signal and noise at the modeled level. Negative estimates can arise from reverse classification, small samples, or data and coding problems, and require contextual examination rather than automatic psychological interpretation (Stanislaw & Todorov, 1999).

For hit rate H and false-alarm rate F strictly between zero and one, d′ = z(H) − z(F), where z is the inverse standard-normal cumulative distribution function. The z transformation maps each probability to the standard-normal value below which that proportion of a distribution lies. Because both response rates contribute, d′ can disentangle a tendency to respond yes from discrimination performance under the model assumptions.

Using H = 0.80 and F = 0.20, z(0.80) is approximately 0.842 and z(0.20) is approximately −0.842, giving d′ ≈ 1.68. That number is a standardized estimate for this task and model. It does not mean the participant will perform with the same sensitivity across modalities, levels of fatigue, stimulus categories, or experimental designs. The parameter is only as interpretable as the conditions, sampling precision, and model fit that support it.

Why d′ is not a universal measure of perceptual capacity

A person's d′ on a particular discrimination task depends on the signal contrast, duration, mask, noise level, instructions, memory demand, and the stability of the criterion. Comparisons are meaningful when the experimental conditions and assumptions support the comparison. SDT can estimate discrimination in perception and memory using analogous mathematics, but equal numerical scores do not prove that the same cognitive mechanism, neural resources, or general aptitude is involved.

An observed change in d′ is compatible with improved encoding, reduced internal variability, altered sustained attention, learning of diagnostic features, and several other mechanisms. Identifying which process changed requires experimental manipulation or additional converging measurements. An intervention that changes a d′ estimate in one laboratory task has therefore demonstrated a local change in performance, not automatic far transfer to everyday attention or clinical functioning.

Response Bias and the Decision Criterion

Response bias in SDT is the location of the decision threshold relative to the evidence distributions. The word bias here describes a measurable response tendency. It does not by itself imply irrationality, prejudice, or the cognitive biases studied in judgment and decision science. A liberal criterion favors saying 'signal present': it tends to increase both hits and false alarms. A conservative criterion favors saying 'signal absent': it tends to decrease both hits and false alarms. The practical value of either policy depends on the task's base rates and the costs associated with the two kinds of error.

A common index of criterion placement is c = −0.5 × [z(H) + z(F)] for the equal-variance Gaussian yes–no model. A c of zero is located at the midpoint between the modeled means; a negative c corresponds to a more liberal criterion, and a positive c to a more conservative one. This convention defines a neutral geometric reference point, not a universally optimal moral or behavioral position. In the earlier numerical example H = 0.80 and F = 0.20, c equals zero because the standardized hit and false-alarm values cancel (Stanislaw & Todorov, 1999).

At a fixed modeled sensitivity of d′ = 1.68, moving c from 0 to −0.50 predicts approximately H = 0.91 and F = 0.37 rather than H = 0.80 and F = 0.20. The extra detected signals come with additional false alarms because the observer accepts less evidence before responding yes. The values are illustrative predictions of the equal-variance normal model, not empirical findings about a particular population.

Criterion c, likelihood-ratio beta, and optimal decisions

The quantity c is a standardized location index, while the likelihood-ratio criterion compares how probable a specific evidence value is under the signal and noise models. Some textbooks use beta, β, for a likelihood ratio at the threshold; conventions differ in whether the ratio is formulated as signal divided by noise or its inverse, so the definition must be stated before interpreting a value. A likelihood-ratio decision rule can incorporate prior probabilities and the consequences of responses.

When false alarms are expensive and signals are rare, the rational threshold for acting may be high. When misses are extremely costly, a lower threshold may be appropriate despite more false alarms. A medical screening program, safety alarm, and laboratory tone-detection study can therefore rationally choose different response criteria. Decision optimality concerns a specified objective and credible probabilities, while SDT's core descriptive analysis can proceed even when an observer's reward function is unknown (Swets, 1988).

Why Accuracy Alone Can Mislead

Suppose two people both classify 80% of trials correctly. One may correctly detect many targets and also sound frequent false alarms; another may rarely respond yes and miss many targets. If the target is uncommon, overall accuracy will reward conservative responding even when the task's practical purpose is finding every important signal. Conversely, maximizing hits alone may reward saying yes nearly constantly and produce an unacceptable number of false alarms. Sensitivity and criterion describe distinct aspects of the response pattern.

Balanced accuracy averages sensitivity and specificity, and it can improve comparisons under unequal class frequency, but it still summarizes one operating point and does not reveal the entire decision trade-off. Precision, also called positive predictive value, depends on prevalence: it is the proportion of positive classifications that truly contain the target. Signal detection sensitivity d′, hit rate, specificity, predictive value, and overall accuracy answer different questions. A responsible experimental report identifies which quantity is being estimated and why it is relevant.

Receiver Operating Characteristic Curves

A receiver operating characteristic (ROC) curve plots hit rate on the vertical axis against false-alarm rate on the horizontal axis as the decision criterion varies. Each point on the curve represents a particular operating threshold, and a collection of points traces the detector's ability to trade hits against false alarms. A diagonal from the bottom-left to the top-right indicates chance-level discrimination across thresholds, while curves rising toward the upper-left indicate better discrimination. The ROC is especially informative when a single experimental criterion would conceal alternative ways of balancing errors (Swets, 1988).

Researchers may estimate several ROC points by offering confidence categories or manipulating the cost of false alarms and misses. A theoretical ROC from a fitted SDT model can also predict performance at criteria not directly observed. Such extrapolation depends on how well the model describes the data. An ROC curve is therefore both a descriptive tool for observed discrimination and a test of assumptions about the latent evidence distributions.

AUC and its interpretation

The area under the ROC curve (AUC) summarizes discrimination across thresholds. Under common conditions, it can be interpreted as the probability that a randomly selected signal observation receives a higher evidence score than a randomly selected noise observation, with tie conventions handled appropriately. An AUC of 0.5 corresponds to chance-level ranking and an AUC of 1 to perfect ranking. Under the equal-variance Gaussian model AUC = Φ(d′/√2), where Φ is the standard-normal cumulative distribution function. Thus a modeled d′ around 1.68 corresponds to AUC near 0.88.

AUC is useful for comparing detectors, but it may obscure clinically important behavior in a limited false-alarm region; two ROC curves can cross, and the detector with larger overall AUC may be worse at the operating point that matters for a particular decision. Sampling uncertainty, case mix, verification bias, and the external population also constrain interpretation. Swets's review of diagnostic systems emphasized both the value of ROC analysis and the crucial dependence of its results on testing quality (Swets, 1988).

The zROC and unequal variance

Transforming the observed hit and false-alarm probabilities into z scores yields a zROC. For Gaussian distributions with equal variance, the predicted zROC is a line with slope one. A different slope suggests that equal-variance assumptions may not hold, although sampling noise, mixtures of processes, and trial-to-trial criterion variability can produce related patterns. In recognition-memory experiments, the signal distribution is often modeled as more variable than the lure distribution, motivating unequal-variance SDT and competing explanations of recognition judgments (Wixted, 2007; Mickes et al., 2007).

Different models can fit the same family of behavioral points while attributing them to different latent processes. Neither a curved ROC nor a zROC slope should be treated as a direct photograph of neural evidence. Model comparison requires sufficient confidence-response observations, checks of goodness of fit, uncertainty estimates, and ideally complementary experimental tests.

Experimental Designs and How Researchers Estimate SDT Parameters

A yes–no detection paradigm presents target and no-target trials with controlled timing, randomized order, and documented base rates. It is especially suited to estimating hits and false alarms directly. A two-alternative forced-choice experiment presents two alternatives or intervals and asks which contained the target. Discrimination can sometimes be estimated with less influence from an explicit yes–no criterion, but the mapping between forced-choice accuracy and d′ depends on assumptions about the distributions and dependence of evidence across alternatives. The widely quoted relation d′ = √2 × z(proportion correct) applies to a particular equal-variance, independent-observation two-alternative model, rather than every possible forced-choice procedure (Stanislaw & Todorov, 1999).

Confidence-rating tasks request graded responses, often from 'certainly absent' to 'certainly present.' Multiple response cutoffs can estimate an empirical ROC, support model fitting, and reveal individual differences in calibration. Repeated-measures designs can examine how criterion and sensitivity change with contrast, workload, expectation, incentive, or practice. Researchers must report exactly how many target and noise trials were presented, how missing responses were handled, which exclusions were applied, and whether the task was balanced or naturally imbalanced.

Reliability, uncertainty, and extreme proportions

A finite sample can produce a hit rate of exactly 1 or a false-alarm rate of exactly 0 even when the underlying probabilities are not at those boundaries. Since z(1) and z(0) are infinite, uncorrected Gaussian d′ cannot be computed from these proportions. One established approach adds 0.5 to each of the four cell counts before calculating the two rates, a log-linear correction studied by Hautus. The choice of correction matters, especially in small samples, and should be specified rather than silently applied (Hautus, 1995).

Point estimates are not enough to characterize uncertain discrimination. Confidence intervals, hierarchical models where appropriate, participant-level summaries, and sensitivity analyses help describe precision and heterogeneity. Combining trial counts across participants without considering subject-specific criteria or sensitivity can conceal individual differences and distort group estimates. Pre-registration of primary outcomes, transparent analysis scripts, and independent replication strengthen confidence that a reported change in detection reflects a reproducible effect.

Experimental response criteria may drift

The simplest model treats the decision criterion as fixed within a condition. In practice, observers may learn base rates, adapt to recent targets, become fatigued, or shift strategy during a long session. These changes can alter observed rates and ROC shape. Signal timing, trial difficulty, speed pressure, and whether participants can abstain or give an uncertain response also determine which model is appropriate. Good designs track trial order and response latency where scientifically relevant.

Poorly calibrated monitors, stimulus leakage, missed responses, ambiguous instructions, unequal trial difficulty, or data processing errors can mimic psychological effects. An analysis that reports a precise d′ to several decimal places without documenting these details creates an appearance of certainty unsupported by the experimental design.

Sensitivity, Attention, and Perceptual Learning

Attention can influence discrimination by prioritizing relevant sensory information, reducing the impact of competing inputs, stabilizing evidence accumulation, or altering response expectations. SDT is helpful precisely because a larger number of detected targets could reflect a change in sensory separability or a criterion shift. Investigators may manipulate a cue or attentional load and estimate both quantities instead of interpreting more yes responses as improved perception.

Perceptual learning may improve performance for practiced contrasts, orientations, categories, or locations. SDT can quantify local sensitivity gains and changing decision policies over training. Whether learning transfers to new stimuli or daily activities is a separate empirical question. A laboratory increase in d′ should be reported in terms of the practiced experimental operation unless an independent generalization study demonstrates broader consequences.

Scientific treatments of perceptual organization and detection belong in the Perception research article, while this article concentrates on the statistical interpretation of uncertain perceptual decisions. The distinction keeps the experimental measurement question separate from the larger mechanisms of sensory organization.

Signal Detection Theory in Memory and Recognition

In an old–new recognition test, studied items are signals and unstudied lures are the alternative category. A participant may feel varying degrees of familiarity, and the decision threshold determines whether that evidence receives an 'old' response. Hits are correct identifications of previously presented items; false alarms are incorrectly identifying new items as old. SDT thereby separates discrimination of targets and lures from a general tendency to endorse items as familiar.

Recognition experiments often generate confidence ROCs whose shapes are studied using unequal-variance models, threshold models, and dual-process interpretations. Theoretical disagreement centers on which latent structures best account for observed distributions and whether the evidence represents one continuous strength variable or multiple decision processes. Different model fits can support different mechanistic stories even when classification performance is similar (Wixted, 2007).

The same numerical vocabulary transfers between perception and recognition, but theoretical transfer requires discipline: a recognition judgment involves encoding history, memory evidence, semantic relatedness, decision demands, and often confidence. A memory ROC does not, by itself, isolate a particular mnemonic system or establish the presence or absence of conscious recollection.

Clinical Testing, Safety Monitoring, and Prevalence

Diagnostic testing often uses the terms sensitivity and specificity. Here diagnostic sensitivity is the proportion of truly positive cases with a positive test result, corresponding mathematically to hit rate, while specificity is the proportion of truly negative cases with a negative result, corresponding to 1 minus the false-alarm rate. These conditional rates describe an operating threshold. ROC analysis studies performance over many thresholds, while clinical interpretation additionally requires prevalence, population characteristics, benefits, harms, and a trustworthy reference standard (Swets, 1988).

Consider an illustrative screening test with 90% sensitivity and 90% specificity for a condition present in 1% of a population. Among 10,000 people, about 100 have the condition and approximately 90 test positive; among 9,900 without it, about 990 also test positive. The positive predictive value is therefore about 90 divided by 1,080, or 8.3%. These are hypothetical arithmetic values, not performance estimates for any real medical test. They demonstrate that even strong-looking sensitivity and specificity do not ensure that most positive results are true positives when prevalence is low.

In safety monitoring, alarms for rare hazards may be tuned toward sensitivity even when that creates nuisance alerts. In security screening, quality assurance, and human–machine monitoring, both the detector's discriminability and the consequences of errors should be considered. The optimal policy cannot be determined from d′ alone. These applications demonstrate the reach of SDT as decision analysis, not a warrant to transfer laboratory thresholds directly into operational protocols.

Confidence, Metacognition, and Type 2 Signal Detection

Ordinary or Type 1 SDT concerns a decision about the world, such as whether a target was present. Metacognitive or Type 2 analysis concerns the observer's judgment about whether their own earlier decision was correct. Confidence can be informative about accuracy, but it may also vary with evidence strength, response speed, expectations, incentives, or reporting habits. Studying confidence therefore requires additional measurements and assumptions beyond the two rates in an ordinary yes–no task.

Maniscalco and Lau introduced a meta-d′ approach for estimating how effectively confidence distinguishes correct from incorrect decisions while accounting for aspects of the underlying discrimination task. Later work identified interpretive vulnerabilities when distributions depart from assumed forms, so meta-d′ should be read as a model-based estimator, not an infallible measurement of self-awareness (Maniscalco & Lau, 2012; Arnold et al., 2023; Arnold et al., 2024).

Confidence criteria can be tuned for different objectives: maximizing classification of one's errors, maximizing expected reward, or achieving calibrated confidence. The mathematically optimal confidence policy depends on which objective is selected. This approach helps specify questions about metacognitive monitoring without confusing measured certainty with correctness or assuming that an SDT task alone establishes the nature of subjective experience (Maniscalco et al., 2025).

Static SDT and Models of Evidence Accumulation

The classical SDT model often summarizes evidence available at a single decision point. Many real decisions evolve over time, and their speed is informative. Sequential probability ratio tests and diffusion-style models describe how evidence can accumulate across multiple samples until a boundary is reached. Such models are closely related to the statistical logic of SDT while making explicit predictions about response time and the temporal dynamics of choice.

A tutorial comparison of signal detection and sequential probability ratio analysis identifies their shared decision-theoretic foundations and different evidence-sampling procedures. Dynamic models can integrate speed–accuracy trade-offs, stopping rules, and sometimes confidence, while a static d′ and criterion estimate remains informative for a simpler task (Griffith et al., 2021; Pleskac & Busemeyer, 2010).

More complicated models are valuable when the available behavioral observations can identify their additional parameters. A model with many plausible processes but weak experimental constraints may fit a response distribution without producing a unique explanation. Model selection should therefore be driven by the scientific question, the design, and predictive adequacy rather than complexity alone.

Assumptions, Alternative Models, and Scientific Limits

The standard textbook expressions for d′ and c require the equal-variance Gaussian account of evidence. Real distributions may be skewed, multimodal, mixtures of trial types, or characterized by unequal variance. When those conditions are violated, the same calculated indices may no longer correspond to their intended simple geometric interpretation. ROC shape, rating data, model checks, and comparison with alternatives help identify potential violations. A published analysis of unequal-variance, random-coefficient, and mixture models demonstrates that superficially similar results may admit different statistical explanations (DeCarlo, 2010).

SDT also separates quantities statistically under its assumptions, but the parameters should not be assigned to pure mental faculties by definition. An increase in d′ could reflect better sensory encoding, memory, attention, task learning, or differences in stimulus distribution. Criterion shifts may include response strategy, changing expectation, motivation, fatigue, or imperfect comprehension of instructions. Independent manipulation and convergent data remain essential for causal inference.

Cross-study comparisons are especially fragile when participants differ in sample characteristics, practice, signal intensity, payoff structure, prevalence, device quality, and response deadlines. In applied research, selection bias and uncertain reference labels further limit transportability. A careful account identifies established mathematics, observed behavioral regularities, model-dependent inference, and unresolved mechanistic questions as distinct levels of scientific claim.

What Signal Detection Theory Can and Cannot Establish

SDT can establish, given an appropriate design and model, whether observed discrimination differs across conditions and whether response criteria shift. It can quantify error trade-offs, make conditional predictions for new thresholds, and help compare classifiers at a defined operating point. It is one of the basic measurement languages connecting perception, recognition memory, experimental attention, and decision analysis.

SDT cannot determine clinical diagnosis from a low experimental score, identify a unique brain circuit from d′ alone, prove that a training intervention generalizes outside its task, or resolve the philosophical status of experience. A person who misses targets in one experiment has produced a behavioral observation under specific conditions. Clinical impairment requires additional evidence about functioning, context, duration, and recognized diagnostic criteria. Researchers should resist describing task-specific signal detection parameters as tests of intelligence or personal worth.

Frequently Asked Questions

What is the difference between sensitivity and response bias?

Sensitivity measures the modeled distinguishability of signal and noise evidence, while response bias describes the rule for choosing when to say yes. A liberal criterion increases both hits and false alarms; a conservative one generally lowers both. A higher d′ means more separated distributions under the assumed model. The indices therefore answer different questions about the same detection behavior.

Can someone have high sensitivity and still make many false alarms?

Yes. A highly discriminating observer may still adopt a very liberal criterion when missing a signal has severe consequences. Sensitivity and the decision policy can vary partly independently; the observed error profile depends on both. Interpreting false alarms always requires the task's incentives, prevalence, and available evidence.

Is d′ the same as accuracy, IQ, attention span, or diagnosis?

No. d′ is a model-based sensitivity estimate for a specified discrimination procedure. Accuracy summarizes correct responses at one operating point; attention span is a much broader and variably defined construct; clinical diagnosis follows separate standards. No general psychological or medical label follows from one d′ value.

Why do researchers need false-alarm trials?

Without noise-only observations, it is difficult to tell whether a person detected more true signals or simply said yes more frequently. The false-alarm rate provides the second conditional probability needed for the familiar yes–no SDT estimates. A different paradigm may use additional assumptions or alternative methods, which must be made explicit.

What is a good d′ or AUC score?

There is no universal cutoff. A satisfactory value depends on the task's difficulty, risk, measurement reliability, comparative baseline, and permissible false-alarm rate. Even a high AUC may be operationally insufficient where errors have unusually severe costs. Numerical interpretations require a defined reference setting.

How are signal detection theory and psychophysics related?

Psychophysics studies the relations between physical stimuli and perceptual judgments. SDT provides a decision-sensitive analytic framework for many psychophysical tasks by distinguishing discriminability from the criterion for reporting a signal. General experimental methods, thresholds, sensory scaling, and stimulus-response functions form a broader research territory.

Related Articles

Perception: How the Mind Interprets Sensory Information explains the sensory and interpretive processes that supply the evidence underlying a detection judgment.

Sensation vs Perception: Differences and Everyday Examples develops the distinction between sensory transduction and the perception of meaningful objects and events.

Selective Attention: How the Mind Filters Information examines how selection and filtering alter which information becomes available for judgment.

References

Arnold, D. H., Clendinen, M., Johnston, A., Lee, A. L. F., & Yarrow, K. (2024). The precision test of metacognitive sensitivity and confidence criteria. Consciousness and Cognition, 123, 103728. DOI.

Arnold, D. H., Johnston, A., Adie, J., & Yarrow, K. (2023). On why we lack confidence in some signal-detection-based analyses of confidence. Consciousness and Cognition, 113, 103532. DOI.

DeCarlo, L. T. (2010). On the statistical and theoretical basis of signal detection theory and extensions: Unequal variance, random coefficient, and mixture models. Journal of Mathematical Psychology, 54(3), 304–313. DOI.

Galvin, S. J., Podd, J. V., Drga, V., & Whitmore, J. (2003). Type 2 tasks in the theory of signal detectability: Discrimination between correct and incorrect decisions. Psychonomic Bulletin & Review, 10, 843–876. PubMed.

Green, D. M., & Swets, J. A. (1966). Signal Detection Theory and Psychophysics. Wiley. WorldCat.

Griffith, T., Baker, S.-A., & Lepora, N. F. (2021). The statistics of optimal decision making: Exploring the relationship between signal detection theory and sequential analysis. Journal of Mathematical Psychology, 103, 102544. DOI.

Hautus, M. J. (1995). Corrections for extreme proportions and their biasing effects on estimated values of d′. Behavior Research Methods, Instruments, & Computers, 27, 46–51. DOI.

Maniscalco, B., & Lau, H. (2012). A signal detection theoretic approach for estimating metacognitive sensitivity from confidence ratings. Consciousness and Cognition, 21(1), 422–430. DOI.

Maniscalco, B., Charles, L., & Peters, M. A. K. (2025). Optimal metacognitive decision strategies in signal detection theory. Psychonomic Bulletin & Review, 32, 1041–1069. DOI.

Mickes, L., Wixted, J. T., & Wais, P. E. (2007). A direct test of the unequal-variance signal detection model of recognition memory. Psychonomic Bulletin & Review, 14, 858–865. DOI.

Pleskac, T. J., & Busemeyer, J. R. (2010). Two-stage dynamic signal detection: A theory of choice, decision time, and confidence. Psychological Review, 117(3), 864–901. DOI.

Stanislaw, H., & Todorov, N. (1999). Calculation of signal detection theory measures. Behavior Research Methods, Instruments, & Computers, 31(1), 137–149. DOI.

Swets, J. A. (1988). Measuring the accuracy of diagnostic systems. Science, 240(4857), 1285–1293. DOI.

Tanner, W. P., Jr., & Swets, J. A. (1954). A decision-making theory of visual detection. Psychological Review, 61(6), 401–409. DOI.

Wixted, J. T. (2007). Dual-process theory and signal-detection theory of recognition memory. Psychological Review, 114(1), 152–176. DOI.