RESEARCH DOSSIER
Information-Theoretic Models of Reaction Time
Abstract. Information-theoretic models of reaction time link the probability structure of response choices to the temporal organization of human action. This fundamental research monograph critically reconstructs the chronometric antecedents, Shannon-information foundations, Hick's original experiments, Hyman's probability manipulations, and the empirical domain of logarithmic choice-latency functions. It distinguishes stimulus entropy, response entropy, event surprisal, subjective expectation, and conditional uncertainty; analyzes stimulus–response compatibility, sequence structure, extensive practice, and speed–accuracy policy; and compares stage theories, diffusion models, leaky competitive accumulators, and linear ballistic accumulation. Contemporary mathematical and empirical investigations from 2021–2026 are examined, including trial-level mutual-information formulations, Bayesian neural rate-code decoding, conceptual-structure alternatives, cumulative residual entropy, cognitive-control-network neuroimaging, and EEG evidence distinguishing forced from voluntary choice. Methodological chapters specify experimental identification, timing validity, response-time distributions, model recovery, replication, and human-factors implications. The synthesis identifies the logarithmic function as an important task-dependent regularity whose fitted parameters alone do not uniquely determine a psychological mechanism. This is a critical narrative synthesis of published scientific literature and contains no newly collected participant data.
Keywords: Hick–Hyman law; choice reaction time; Shannon entropy; information theory; stimulus uncertainty; response entropy; surprisal; stimulus–response compatibility; choice-set size; sequential effects; speed–accuracy tradeoff; diffusion decision model; leaky competing accumulator; linear ballistic accumulation; human factors; mental chronometry; response-time distributions; Artificial Era; neural decoding; cognitive control network; forced choice; free choice; cumulative residual entropy
The Scientific Problem of Choice Latency
Reaction time is the interval between a defined stimulus event and an operationally specified response. In a choice task, the observer must identify relevant evidence, select an action from a response repertoire, and execute it under timing constraints. Information-theoretic models ask how uncertainty over alternatives relates to this interval. Their historical importance lies in turning the size and statistical organization of a choice set into experimentally manipulable variables. The resulting functions are descriptions of performance under controlled conditions, and they become mechanistic accounts only when additional assumptions about perception, response selection, and decision dynamics survive independent tests Hick (1952) Hyman (1953) Proctor & Schneider (2018).
The object of this dossier is the relationship among choice-set entropy, response probabilities, stimulus–response assignments, response time, and accuracy. This requires examining the celebrated logarithmic relation, its original experimental evidence, departures induced by practice and compatibility, and modern models that reproduce aggregate slopes through different processing mechanisms. A participant may recognize one of several signals yet press only one key; alternatively, multiple signals may demand distinct actions. Those designs have different response entropy even when their perceptual vocabularies coincide. A defensible model specifies which probabilities, representations, and mappings define its information variable Kornblum et al. (1990) Smith (1968) Proctor & Schneider (2018).
The research method is a critical narrative synthesis of original experimental articles, information theory, mathematical modeling, review literature, and response-time methodology. A mathematical fit is evaluated separately from explanatory uniqueness. Primary anchors include the historical experiments of Hick and Hyman and later work on compatibility, serial dependence, evidence accumulation, and hierarchical response-time estimation. The synthesis emphasizes what a model predicts before examining a data set, which outcomes would count against it, and how the same pattern of mean latency may arise from qualitatively different mechanisms. No new participant data or systematic-review meta-analysis is claimed Ratcliff & McKoon (2008) Brown & Heathcote (2008) Proctor & Schneider (2018).
Defining Stimulus Sets, Response Sets, and Trials
A choice reaction-time experiment contains a set of potential stimulus events S, a mapping from events to actions R, a sequence of trials, and timing rules for stimulus onset and response registration. The number of alternatives can mean physically distinct stimuli, perceptual categories, response keys, or trial-contingent action policies. These quantities coincide only in some designs. If ten distinct visual patterns are grouped into two response categories, entropy of visual events may be high while entropy of response identity is low. An analysis must name the random variable it treats as uncertain and retain the experimental mapping as part of the model Hyman (1953) Kornblum et al. (1990).
Laboratories also distinguish simple reaction, go/no-go discrimination, two-choice classification, and many-choice identification. The first can require detection and response initiation without identification among assigned alternatives. The second introduces withholding and criterion control. The third and fourth introduce choice among responses, with additional difficulty determined by stimulus discriminability, response compatibility, and instruction. Task categories are descriptions of the designed contingencies; cognitive stages are explanatory hypotheses requiring independent evidence. The additive-factors approach and later sequential-sampling models illustrate why elementary changes in mean latency cannot by themselves localize a processing component Sternberg (1969) Ratcliff et al. (2016).
One trial provides at least a latency, a response, and a correctness code; an informative data set also records stimulus identity, prior trials, mapping block, confidence where relevant, and trial exclusions. Misses, anticipations, timeouts, and slow contaminant responses affect estimated means and distributions. For low-frequency alternatives, pooled means may depend strongly on unequal sampling across events. These concerns are especially acute when an entropy manipulation changes repetition frequencies or the proportion of difficult stimuli. Experimental inference therefore requires a transparent denominator for accuracy and an explicit policy for unsuccessful or unobserved responses Ratcliff & Tuerlinckx (2002) Bertelson (1965) Vandekerckhove et al. (2011).
Chronometric Antecedents and the Emergence of a Quantitative Problem
Nineteenth-century mental chronometry made the timing of perceptual and voluntary action an experimental object. The classical subtraction logic associated with Donders sought to isolate additional operations by comparing tasks that differed in their required processes. Later investigators examined how the number of alternatives changed response latency. Hick explicitly discussed Merkel's 1885 observations and the earlier descriptive attempt to relate choice time to alternative-set size; this historical chain is reported in Hick's original article rather than treated as independent verification of nineteenth-century raw data Hick (1952) Sternberg (1969).
The historical question moved from whether choosing takes longer than detecting to whether increased alternatives impose a lawful quantitative cost. A logarithmic expression was attractive because adding the same proportion of alternatives would then add approximately the same increment in latency. Yet early chronometry could not decide whether this reflected serial comparisons, central translation between symbols and actions, preparation of competing responses, or an invariant rate of information processing. Those explanatory alternatives already imply different predictions for practice, meaningful categories, prepared responses, and highly compatible stimulus–response relationships Smith (1968) Proctor & Schneider (2018).
Postwar developments connected experimental psychology to communication engineering and cybernetics. Information theory offered a precise way to measure event uncertainty without assuming that each stimulus is psychologically equal. The move was consequential: probability distributions became potential predictors of latency and not merely descriptors of stimulus presentation. Equally important, a mathematical measure of information did not guarantee that human information processing behaved as a noiseless communication channel. Original information-theoretic formulations describe the statistical structure of signals; psychological hypotheses must separately specify how a task converts that structure into attention, selection, and movement Shannon (1948a) Hick (1952) Hyman (1953).
Shannon Entropy as a Formal Construct
For a finite set of outcomes with probabilities p₁ through pₙ, Shannon entropy is H(S) = −Σᵢ pᵢ log₂ pᵢ bits, with zero-probability terms contributing zero by convention. For N equally likely outcomes, entropy becomes log₂ N bits. Entropy concerns the expected uncertainty of the outcome variable defined by a probability distribution. It is therefore a mathematical property of the task distribution under a specified measurement model, rather than a direct measure of experienced confusion or the number of mental operations occurring during a trial Shannon (1948a) Shannon (1948b).
Information conveyed by an individual realized event can be described as surprisal, I(sᵢ) = −log₂ pᵢ. Average surprisal over the event distribution equals entropy; a rare event carries more surprisal than a common one. The distinction is essential when interpreting response time. A claim about average latency and entropy is not identical to a claim that a particular trial's latency scales with its own surprisal. Moreover, an experimental observer may learn an approximate distribution that differs from the programmed frequencies, so the experimenter's entropy and the participant's predictive uncertainty need not be identical Hyman (1953) Proctor & Schneider (2018).
Conditional entropy H(R|C) characterizes uncertainty over responses given a context C; mutual information I(S;R) describes statistical dependence between stimulus and response variables. Neither quantity by itself measures the decision-making architecture. A perfect deterministic mapping from perceptually identifiable stimuli to responses may give H(R|S) = 0 while selection remains slow because recognizing S and retrieving the mapping require time. Conversely, extensive practice may make actions rapidly available in a task with many nominal alternatives. This mathematical separation helps avoid confusing uncertainty, transmitted information, and computational effort Shannon (1948b) Kornblum et al. (1990) Shiffrin & Schneider (1977).
Hick's Experiments and the Logarithmic Form
William Edmund Hick's 1952 experiments compared choice-response times across tasks with different numbers of alternatives, including a conventional low-error choice design with sets up to ten options and a ten-choice design that manipulated the speed–accuracy compromise. His principal formulation related mean latency to an information measure and proposed an approximately constant information-gain rate within a perceptual–motor act. Hick's published abstract characterized the rate as on the order of five bits per second in his experimental circumstances; this is a historical study-specific result, not a universal capacity constant Hick (1952).
The familiar expression RT = a + b log₂(N + 1) is often used as a practical presentation of Hick's relation, while other experimental formulations employ log₂ N for N equiprobable alternatives. Their intercepts and numerical interpretation differ, especially at the one-alternative boundary. A parameter b has units of time per bit and describes the fitted sensitivity of latency to the chosen information variable. It is a regression slope under an identified design. An estimate of b from one task cannot automatically be interpreted as the bandwidth of a stable central mental channel across incompatible tasks Hick (1952) Proctor & Schneider (2018).
Hick also studied conditions in which participants traded accuracy for shorter responses. That manipulation reveals why a single latency average is an incomplete description of choice behavior. A faster response accompanied by more errors can reflect a changed decision criterion, a policy of guessing, or task-dependent adjustment of processing time. The historical importance of his paper includes the explicit recognition that choices, probability distributions, errors, and response speed belong in one analytical framework. The original experiments offered evidence for an orderly relationship but left several candidate psychological mechanisms underdetermined Hick (1952) Ratcliff & McKoon (2008).
Hyman's Probability Manipulations and the Hick–Hyman Tradition
Ray Hyman's 1953 study extended the informational account by treating stimulus probability as a determinant of reaction time. In an equiprobable set, choice uncertainty can be represented by log₂ N; with unequal probabilities, the Shannon expression weights information by the presentation distribution. Hyman's intervention sharpened the central hypothesis: if the statistical uncertainty of the experiment changes while other relevant task properties remain controlled, mean selection time should change in a systematic relation to that uncertainty Hyman (1953) Shannon (1948a).
A critical interpretation keeps event-level surprise, block-level entropy, and actual choice-policy changes distinct. Unequal frequency can alter expectancy, stimulus repetition, motor preparation, and practice for frequent events. A relation between mean reaction time and entropy may therefore be carried partly by learning and sequential dependence. The experimental program must ask whether entropy predicts time after conditioning on recent events, compatibility, and perceptual discriminability. Comparing probability distributions of equal entropy but different event structures offers an especially useful design for testing whether entropy alone is sufficient Bertelson (1965) Proctor & Schneider (2018).
The combined label Hick–Hyman law designates a historical family of logarithmic choice-latency regularities rather than a complete cognitive architecture. The tradition remains central because it delivers parsimonious predictions over experimentally defined alternative sets and supplies a compact language for discussing uncertainty in action selection. Its scientific reach depends on defining an appropriately controlled task family. Contemporary assessments emphasize that stimulus–response compatibility, substantial practice, very large sets, and sequential effects challenge attempts to treat a single logarithmic line as a context-free law Proctor & Schneider (2018) Smith (1968) Seibel (1963).
Choice-Set Size and the Limits of Logarithmic Extrapolation
Changing set size is a natural starting point for an information-theoretic test because equal-probability entropy increases by one bit when the number of alternatives doubles. Yet adding alternatives may also expand visual search, alter spatial spacing, change memory demands, or introduce unfamiliar motor responses. A valid test of the logarithmic relation holds stimulus quality and mapping architecture sufficiently stable to interpret the independent variable. Comparing differently sized displays assembled from qualitatively different objects makes set size a compound manipulation, even when the experimenter can compute entropy exactly Hick (1952) Teichner & Krebs (1974) Proctor & Schneider (2018).
Large-alternative studies provide essential stress tests. Seibel's 1963 investigation was explicitly organized around a 1,023-alternative discrimination task. Its existence demonstrates that very large sets were studied experimentally, yet its particular task design, learned response coding, and pattern of results require examination before extrapolating from small keypress arrays. A mathematical logarithm grows slowly with N, while the physical and conceptual complexity of an actual display may grow differently. A model that claims generality must specify how alternatives are represented and whether the observer treats them as independent entries, structured groups, or components Seibel (1963) Proctor & Schneider (2018).
Models can fit a local range without remaining valid at extremes. At the low end, detection, go/no-go, and two-choice tasks introduce different baseline processes; at the high end, categorization, coding, and motor coordination may change the effective problem. Testing extrapolation requires preregistered predictions for set sizes outside the calibration range, including correct and error response distributions. The appropriate conclusion from a good log-linear mean fit is that one functional regularity explains a defined set of observations. A claim of a universal central capacity needs convergent independent tests of its proposed mechanism Smith (1968) Ratcliff et al. (2016) Proctor & Schneider (2018).
Unequal Probabilities, Expectation, and Local Surprisal
When alternatives have unequal probabilities, entropy summarizes average unpredictability, and surprisal assigns a value to each outcome. These two levels yield different experimental predictions. If mean latency follows block entropy, rearranging event frequencies while holding entropy constant should leave the predicted average approximately unchanged, subject to the model's other parameters. If trial latency follows event surprisal, a rare stimulus should be comparatively slow even in blocks with similar total entropy. Observed effects can be amplified or reduced by differences in discriminability, response repetition, and stimulus–response compatibility Hyman (1953) Shannon (1948a) Bertelson (1965).
Probability schedules shape preparation because participants learn which response is likely. With a dominant action, an observer can favor one motor program or lower a decision criterion for that choice. Those adjustments change both the distribution of correct latencies and the asymmetry of errors. A model should make explicit whether frequency acts through prior probability, evidence accumulation, starting bias, response threshold, or a combination. In diffusion-model experiments, manipulations of stimulus proportions can affect starting point and drift-related bias, supporting an analysis beyond a single entropy coefficient Ratcliff & McKoon (2008) Ratcliff et al. (2016).
The distinction between statistical and subjective probability becomes sharper across learning. A randomized presentation schedule has an experimenter-defined distribution from the beginning; participants may initially behave as though outcomes were roughly equiprobable, then acquire local and global expectations. Trial-by-trial learning models can therefore predict a changing effective surprisal even under a stationary schedule. A sound research report should disclose transition statistics and block order, and should compare probabilities estimated from the programmed sequence with models of what observers plausibly learned. This prevents a numerical match to entropy from being overinterpreted as a unique cognitive computation Shiffrin & Schneider (1977) Proctor & Schneider (2018) Bertelson (1965).
Stimulus–Response Compatibility as an Independent Determinant
Compatibility describes the relation between stimulus features and available response codes. When a signal's spatial or conceptual structure corresponds naturally to an action, selection may be faster than under an arbitrary assignment. Fitts and Seeger's 1953 study explicitly examined spatial characteristics of stimulus and response codes, and Fitts and Deininger continued the compatibility program in 1954. The central methodological implication is that two tasks with the same number and probability of alternatives can produce different latencies because their mappings impose different translation demands Fitts & Seeger (1953) Fitts & Deininger (1954) Proctor & Schneider (2018).
Kornblum, Hasbroucq, and Osman's dimensional-overlap account identifies relationships among dimensions of stimulus and response sets. Overlap can be relevant or irrelevant to the instruction and can support matching or conflicting codes. In later experiments, Kornblum and Lee compared four- and six-choice graphic and alphabetic tasks with keypress and verbal responses. They found facilitative and interfering contributions from dimensional correspondence. These data show why an entropy-only model leaves out psychologically important structure: the number of available responses remains unchanged while the degree of code correspondence changes Kornblum et al. (1990) Kornblum & Lee (1995).
A careful factorial design crosses entropy or choice-set size with compatibility rather than assigning compatible mappings exclusively to the smallest condition. If the number effect survives with compatible mappings, its slope may still differ; if it disappears under intensive practice, this calls for a theory that explains the interaction. Response-code dimensional overlap and signal uncertainty can influence distinct components, or interact within one competitive decision process. The pattern of interactions, combined with error behavior and distribution shape, provides more discriminating evidence than either the overall average or an isolated significant slope Sternberg (1969) Kornblum et al. (1990) Proctor & Schneider (2018).
Sequential Effects and Repetition Structure
Trial sequences contain statistical structure beyond the long-run frequency table. A repetition may be faster because the stimulus representation remains active, because a response plan is primed, or because the observer predicts a repeat. Bertelson's 1965 work explicitly differentiated response repetition from combined signal-and-response repetition. This question is central to Hick–Hyman models: two sequences can share the same marginal entropy and yield different reaction-time patterns if their transitions differ. Therefore global entropy cannot be assumed to exhaust the information relevant to immediate choice Bertelson (1965) Hyman (1953).
For a sequence of events, conditional entropy H(Sₜ|Sₜ₋₁) measures average uncertainty about the next event given the preceding event. It can differ substantially from the marginal H(Sₜ) when events are correlated. But psychological processing may respond to more than one preceding trial, and observers may learn motifs or runs. Estimating condition-specific latencies by event frequency and repetition history can help distinguish local priming from broader expectations. Suitable designs should balance or deliberately manipulate transitions instead of treating all repeated responses as interchangeable observations Shannon (1948b) Bertelson (1965) Proctor & Schneider (2018).
Analysis must also consider the temporal spacing between trials. Priming, response preparation, and orienting can decay with time, whereas learned statistical contingencies may persist across blocks. It is therefore informative to vary response–stimulus intervals and independently measure effects of repeating the signal, repeating the key, and repeating a category. A model that explains the unconditional mean but systematically misses these conditional patterns remains scientifically incomplete for the observed behavior. Modeling serial dependence offers an adversarial test of claims that average uncertainty determines latency by itself Bertelson (1965) Ratcliff & McKoon (2008) Smith (1968).
Practice, Skill Acquisition, and Automaticity
Practice reorganizes choice performance. Repeatedly using the same mapping can improve recognition of cues, availability of response codes, stimulus–response associations, and anticipatory preparation. Schneider and Shiffrin's 1977 program distinguished consistent from varied mapping in the development of automatic detection and controlled search. Although this was not simply a test of Hick's law, it establishes that the functional costs of selecting among nominal alternatives depend on acquired organization. The same externally defined information demand can be met by increasingly different internal operations as learning proceeds Schneider & Shiffrin (1977) Shiffrin & Schneider (1977).
Learning also changes the meaning of set size. Experienced operators may classify many possible states as members of familiar groups or execute highly practiced action sequences from a single perceptual cue. Novices may handle the same menu by explicit identification and translation through instructions. The comparison should test transfer to rearranged mappings and novel alternatives: improvement that vanishes when keys are reassigned suggests mapping-specific learning, whereas broader transfer may indicate improved stimulus categorization or general task skill. Thus an information–time slope is conditional on experience and transfer structure Proctor & Schneider (2018) Kornblum et al. (1990) Shiffrin & Schneider (1977).
Experimental research should estimate practice trajectories rather than hiding them by averaging early and late sessions. Randomizing the order of set-size blocks helps guard against confounding, but it does not fully solve carryover, retention, or strategy change. Longitudinal designs can separate warm-up, trial-level learning, and stabilized performance and can compare both speed and accuracy after reconfiguration. Claims about a universal throughput should face the demanding question of whether it remains stable across learning phases and compatible mappings; the historical literature gives substantial reason to expect contextual dependence Seibel (1963) Proctor & Schneider (2018) Ratcliff et al. (2016).
Speed–Accuracy Tradeoffs and Response Criteria
Speed and correctness are coupled by task policies. An instruction to respond quickly can increase anticipations, guesses, or responses made before evidence is sufficiently reliable. Hick's original ten-choice speed–accuracy manipulation acknowledged this directly. Modern evidence-accumulation theories make the tradeoff explicit: the decision-maker can demand more evidence before committing, extending latency while generally improving accuracy. A comparison of information coefficients across conditions is ambiguous if conditions induce different caution or error costs. Measurement should therefore include full speed–accuracy functions rather than one latency at an unspecified performance level Hick (1952) Ratcliff & McKoon (2008).
Mean reaction time of correct responses can be misleading if the hardest or slowest trials are disproportionately errors or timeouts. A change in criterion can alter which observations remain in the correct-response sample and shift its mean without changing stimulus identification speed in a simple way. An informative analysis jointly models response choice and timing, examines error latencies, and records the treatment of omissions. Distribution-sensitive methods are especially valuable when a manipulation changes the tail of slow responses or the balance between fast guesses and deliberate selections Ratcliff & Tuerlinckx (2002) Brown & Heathcote (2008) Vandekerckhove et al. (2011).
Optimal speed–accuracy policy also depends on incentives and feedback. If an error is costly, slowing can be a rational allocation of time rather than a deficit in processing. If the task encourages fast responses and tolerates some errors, an information slope may reflect a different operating point on the same underlying decision process. A powerful test equates accuracy across set sizes, or uses explicit modeling to compare estimated thresholds and evidence quality. Neither procedure should assume in advance that every speed cost is an information-transmission limit Palmer et al. (2005) Ratcliff et al. (2016) Hick (1952).
From Information Functions to Mechanistic Explanation
An information function is a mapping from properties of the task distribution to an observed latency statistic. A psychological process model goes further by specifying latent states, time-dependent operations, and how responses emerge. The distinction is important because an intercept and a logarithmic slope contain little information about processing architecture. A serial comparison process, a learned code-retrieval process, and a set of competing accumulators could each generate a curve resembling Hick's law over some range. Explanatory evaluation therefore requires joint predictions for response identity, errors, quantiles, and interventions on mapping and stimulus quality Smith (1968) Usher & McClelland (2001) Proctor & Schneider (2018).
The expression bits per second invites a channel metaphor. A communication channel has explicitly defined inputs, outputs, a probability law, and assumptions under which capacity can be evaluated. Human choice tasks rarely satisfy a single stationary channel description throughout perception and movement. A participant may redistribute attention, update expectations, or alter speed–accuracy criteria within a session. One can legitimately compute the ratio of an information measure to time, but interpreting that ratio as neural capacity requires convergent process-level evidence. The conceptual contribution of information theory is precision about uncertainty, accompanied by obligations to test its psychological interpretation Shannon (1948a) Hick (1952) Luce (1986).
Model comparison must distinguish descriptive adequacy from mechanistic recovery. Fitting a Hick–Hyman slope is relatively easy for flexible architectures when only aggregate means are available. More difficult predictions include how latency distributions change when probabilities are shuffled, how old and novel mappings transfer, and whether stimulus repetition affects motor versus perceptual components. Models should also be compared against constrained alternatives that contain no explicit entropy computation. When two mechanisms are observationally equivalent under the current design, an informative scientific conclusion identifies the experiment needed to separate them Ratcliff & McKoon (2008) Brown & Heathcote (2008) Proctor & Schneider (2018).
Serial Comparison, Stage Models, and Response Translation
A serial comparison account assumes that a response emerges after one or more ordered evaluations. Its expected latency depends on the number of inspections, their order, the likelihood that an item is encountered early, and the stopping rule. Such mechanisms can yield linear, sublinear, or other choice-set functions depending on how alternatives are represented. A logarithmic relation might result from hierarchical partitioning rather than from a universal information channel. To evaluate the claim, the experimenter needs process predictions for ordered probes, errors, and transfer when the response tree is reorganized Smith (1968) Luce (1986).
Stage theories distinguish perceptual encoding, classification, response selection, and motor execution. Sternberg's additive-factors method proposed exploiting interactions or additivity among experimental factors to infer functional stages, while also warning that additivity and stochastic independence are separate mathematical properties. A set-size effect concentrated in response translation would differ from an effect carried by degraded stimulus encoding. Factorial manipulations can improve localization, yet their interpretation depends on assumptions about selective influence, stage ordering, and the possibility of compensatory changes across processes Sternberg (1969) Ratcliff et al. (2016).
Information and stage explanations can be complementary. Entropy can characterize the difficulty of a selection problem while stage architecture specifies where task structure affects time. However, an additive equation RT = encoding + selection + movement does not guarantee that those processes are independent or separable in neural activity. If stimulus uncertainty alters preparation or attention before onset, the nominal selection factor influences earlier operations. Dynamic models and neural measurements can help test those interactions, but identifying exact stages from a simple mean remains difficult Sternberg (1969) Anderson et al. (2016) Ratcliff et al. (2016).
Sequential-Sampling and Diffusion Accounts
Sequential-sampling models treat a decision as an evolving stochastic process in which noisy evidence accumulates until a response criterion is reached. The classical diffusion model is especially developed for two-choice tasks, where drift rate describes average evidence quality, decision boundaries describe required evidence, starting point represents an initial preference, and nondecision time captures components outside the modeled decision. It jointly predicts response proportions and correct/error time distributions. Its explanatory resources are broader than the Hick–Hyman function and allow direct tests of accuracy-related mechanisms Ratcliff & McKoon (2008) Ratcliff et al. (2016).
An entropy manipulation can affect a diffusion model through several routes. Changing prior frequency can shift starting bias, changing discriminability can alter drift, and emphasizing accuracy can increase boundary separation. A simple information–time correlation therefore cannot specify which route has changed. Ratcliff and McKoon's review describes experiments on stimulus quality, speed–accuracy emphasis, and stimulus proportions that illustrate these distinctions. The same information manipulation may also alter nondecision preparation or response coding, which requires evaluating residual patterns rather than forcing every effect into a single decision parameter Ratcliff & McKoon (2008) Ratcliff & Tuerlinckx (2002).
Binary diffusion models should not automatically be transferred unchanged to tasks with many possible actions. Multialternative versions require assumptions about how option-specific evidence is represented and how competitors are inhibited or compared. Different models may predict similar accuracy at a few conditions yet diverge in error identity and latency quantiles. A useful research strategy begins with a task-appropriate response space, estimates model parameters across conditions under shared constraints, and checks simulation-based posterior or predictive fit. The aim is a falsifiable account of choice dynamics, not a relabeling of bits as drift Usher & McClelland (2001) Brown & Heathcote (2008) Vandekerckhove et al. (2011).
Leaky Competition and Multialternative Decisions
Usher and McClelland's leaky competing accumulator model represents gradual information accumulation in interacting decision units. Leakage allows evidence to decay; lateral inhibition creates competition among alternatives. This architecture was developed partly to explain the temporal course of perceptual choice and can account for log-linear effects of choice-set size under specified conditions. Its ability to reproduce Hick-like behavior shows that a logarithmic pattern may emerge from interacting dynamical units without assuming one fixed serial sequence of comparisons Usher & McClelland (2001) Proctor & Schneider (2018).
Competition changes the meaning of effective set size. Adding a choice can change inhibition or competition among existing response units, even when the new item is rarely presented. Similar or correlated alternatives may also interact more strongly than dissimilar alternatives. A set of visually confusable options creates competition at encoding, while a set of semantically associated actions can create competition at response selection. Such distinctions motivate tests that manipulate similarity independently of number and entropy, evaluating patterns of mistakes as well as elapsed time Usher & McClelland (2001) Kornblum et al. (1990).
Multialternative decision field theory by Roe, Busemeyer, and Townsend extends dynamic choice models to several options and addresses preference phenomena such as similarity, attraction, and compromise. Its original domain concerns preferential choice and should not be collapsed into elementary speeded identification. Nevertheless, the shared theme is that alternatives can change one another's effective values through dynamic comparison. The broader lesson is to specify whether the task requires selecting a learned correct action or evaluating multiple attractive actions, because the relevant probability and evidence structures differ across these choice settings Roe et al. (2001) Ratcliff et al. (2016) Proctor & Schneider (2018).
Linear Ballistic Accumulation and Model Alternatives
Brown and Heathcote's linear ballistic accumulator model describes choice using independent racing accumulators with random trial-to-trial rates and starting points. Within each trial, accumulation is deterministic given those sampled parameters. This simplification permits tractable predictions for full response-time distributions and choices. It provides a different generative explanation of observed speed and accuracy from diffusion and leaky-competition models. Comparing the models is especially informative when the response set contains more than two alternatives and when the experiment manipulates response frequencies or the strength of evidence Brown & Heathcote (2008) Ratcliff et al. (2016).
A race model can attribute extra choice time to changes in the distribution of competitive finishing times rather than to explicit computation of global entropy. If additional response channels compete, changes in winning time may follow a systematic relation to set size over a restricted range. Consequently a successful logarithmic regression does not identify the mechanism. Model selection should require fitting choices and time distributions under shared experimental constraints and then predicting held-out set sizes, mappings, or base-rate shifts. This prevents interpreting parameter flexibility as scientific confirmation of an information-processing bottleneck Brown & Heathcote (2008) Ratcliff & McKoon (2008).
Different accumulator models have different identifiability challenges. Some combinations of rate variability, thresholds, and nondecision time can yield similar mean behavior. Researchers should assess whether parameter estimates recover known values from simulated data and whether competing models can be distinguished at feasible trial counts. Hierarchical estimation can stabilize inference across people or items but cannot make an intrinsically uninformative experimental design reveal mechanisms. A robust analysis reports predictive failure modes and checks for parameter tradeoffs, rather than treating a single favored model as uniquely established Vandekerckhove et al. (2011) Ratcliff & Tuerlinckx (2002) Brown & Heathcote (2008).
Neural and Computational Levels of Description
Neural decision research connects perceptual evidence, temporal integration, and action preparation to signals observed in the brain. Gold and Shadlen's review summarizes neurophysiological principles of deliberation and commitment, and Palmer, Huk, and Shadlen relate stimulus strength to the joint time–accuracy pattern of perceptual judgments. These studies offer process constraints that a pure entropy formula cannot provide. At the same time, evidence accumulation in a specific perceptual discrimination task is not automatically a measurement of a universal information-theoretic channel for every reaction task Gold & Shadlen (2007) Palmer et al. (2005).
One scientific advantage of computational models is that their internal variables can be related to manipulations and measured signals at more than one level. A prediction about changes in drift or response criterion may be examined behaviorally through error quantiles and, in appropriate designs, neurally through time-resolved signals. Such correspondence remains model-dependent. A neural correlate of a decision variable can be compatible with several architectures, and its existence does not independently demonstrate that subjective uncertainty is computed as Shannon entropy. Multilevel agreement should constrain interpretation rather than replace comparison among alternatives Gold & Shadlen (2007) Ratcliff et al. (2016) Anderson et al. (2016).
The Artificial Era expands the set of possible information-processing systems that can be compared functionally, but the human reaction-time tradition supplies only a task-specific metric. Artificial classifiers can be assigned the same probability structure and stimulus–action mapping as humans; a meaningful comparison still requires common event definitions, response accuracy, time measurement, and computational resource accounting. Machine throughput, motor latency, and human decision time have different causal decompositions. Any theoretical extrapolation toward artificial cognition therefore remains a comparative hypothesis requiring controlled measurement, rather than a conclusion inferred from a shared logarithmic fitting equation Shannon (1948a) Ratcliff et al. (2016) Proctor & Schneider (2018).
Experimental Identification of Information Effects
A decisive experiment changes uncertainty while controlling competing explanations to the extent feasible. An equiprobable set-size series should specify whether displays, response keys, and physical response demands remain comparable across levels. An unequal-probability series should disclose the precise event distribution and whether repetitions are independent. Designs comparing the same nominal entropy across different configurations are especially valuable because a true entropy-only account predicts equivalence where mapping-specific or similarity-sensitive models may disagree. Factorial tests can cross probability, compatibility, and practice while treating participant as a repeated-measures unit Hyman (1953) Kornblum & Lee (1995) Proctor & Schneider (2018).
A practical protocol begins with planned response assignments, counterbalancing across subjects, and explicit familiarization. Probabilities must be estimated from actual presented trials when a schedule deviates from its nominal design, and available-response entropy should be distinguished from signal entropy. Trial logs should include the randomized sequence, exact timing, stimulus identity, response identity, success, and any missed deadlines. A speeded selection task with rare events needs enough rare-event observations for credible conditional estimates; a large total sample can conceal severe sparsity for low-probability cells Ratcliff & Tuerlinckx (2002) Vandekerckhove et al. (2011).
Confirmatory comparisons should be organized around rival predictions. Does a set-size slope persist under compatible mapping? Do response frequencies still predict conditional latency after recent-trial history is modeled? Does practice change the slope more strongly than the intercept? Can a full distributional model predict wrong-key choices on new displays? Each design should contain one or more conditions selected specifically because competing mechanisms disagree. Otherwise collecting more observations from a narrow design may increase precision without increasing the ability to identify psychological processing Smith (1968) Ratcliff et al. (2016) Proctor & Schneider (2018).
Chronometric Measurement and Data Integrity
Reaction-time measurement begins with a hardware and software time base. An experimenter must establish when a visual, auditory, or tactile stimulus actually becomes available and when the response device registers action. Display refresh, input polling, buffering, and motor-device latency can contribute offsets or variability unrelated to cognition. The intercept of a Hick–Hyman regression mixes sensory preparation, task policy, motor execution, and measurement characteristics, so absolute comparisons across equipment demand calibration or a shared timing method. A small estimated information slope can be particularly vulnerable to uncharacterized timing noise Luce (1986) Ratcliff & Tuerlinckx (2002).
Data cleaning rules need to preserve the behavioral question. Anticipatory responses, omissions, long tails, and device faults are conceptually distinct events. A blanket rule deleting unusually slow observations can remove meaningful evidence of uncertainty-related processing, while ignoring measurement failures can distort distribution fits. Report how many trials were removed by each rule, whether thresholds were fixed before analysis, and whether results survive sensitivity analyses. The sampling distribution of a mean can be estimated precisely even when the underlying mixture contains distinct behavioral modes; precision alone does not establish validity Ratcliff & Tuerlinckx (2002) Vandekerckhove et al. (2011).
A useful measurement record identifies participant, session, condition, stimulus, programmed probability, response code, actual key, onset timestamp, detected input timestamp, and correctness. It preserves individual-trial data and the stimulus–response map used during collection. For multiday protocols, changes in instruction, practice, and hand placement also belong in the provenance. These details are prerequisites for testing conditional entropy or sequential adaptation. Without them, an apparently clean mean-latency curve can no longer be linked to the competing mechanisms that its theoretical interpretation requires Bertelson (1965) Proctor & Schneider (2018) Ratcliff & McKoon (2008).
Reaction-Time Distributions and Quantile Evidence
Means compress a positive, typically asymmetric response-time distribution into one number. Two conditions can have equal means but different proportions of extremely fast, moderately fast, and slow responses. Quantiles and cumulative distributions reveal whether a manipulation shifts the entire distribution, expands its tail, or selectively affects the earliest responses. Error distributions add constraints: an observer may make unusually fast mistakes under one criterion or slow errors under another. Modern sequential-sampling models exploit these additional features to interpret choice behavior more richly than a regression of condition averages Ratcliff & McKoon (2008) Ratcliff & Tuerlinckx (2002).
The appropriate distributional analysis is determined by the model. A diffusion account predicts joint correct and incorrect timing distributions from a parameterized stochastic process; a ballistic-race account generates distributions from competing accumulation rates and starting points. Statistical fit should not be judged solely by the shortest numerical residual in the observed cells. Posterior predictive or simulation-based diagnostics can reveal structured discrepancies, including wrong tails, incorrectly predicted error types, and poor generalization to untrained conditions. Such discrepancies help identify which theoretical assumptions require revision Brown & Heathcote (2008) Vandekerckhove et al. (2011) Ratcliff et al. (2016).
Distribution analysis also forces attention to mixture processes. Participants can lapse, switch strategies, guess, or execute a partially prepared response. A long slow tail or extra mass of very fast errors can be modeled explicitly where justified. Ratcliff and Tuerlinckx demonstrated how contaminant observations and across-trial variability challenge parameter estimation and compared methods for handling them. The broader methodological rule is that estimates of a cognitive mechanism inherit the assumptions of the measurement model. A narrow confidence interval around a misspecified parameter is insufficient evidence for a particular process Ratcliff & Tuerlinckx (2002) Palmer et al. (2005).
Model Recovery, Generalization, and Predictive Tests
Identifying a psychological mechanism requires more than fitting one candidate model to a fixed data set. Parameter recovery asks whether a fitting procedure can reconstruct values used to generate simulated observations. Model recovery asks whether an experimental design and selection criterion can distinguish known generating models. Cross-validation and out-of-sample predictions assess whether apparent fit depends on idiosyncratic observations. These tests are important when entropy, practice, and compatibility jointly change several latent parameters or when a model contains flexible participant-level effects Vandekerckhove et al. (2011) Brown & Heathcote (2008) Ratcliff & Tuerlinckx (2002).
A particularly revealing comparison fits all models to a limited set of conditions and predicts unobserved manipulations. A pure logarithmic law can be trained on two-, four-, and eight-choice conditions and evaluated on a separately tested set size or altered stimulus probability distribution. Competing accumulator models can be evaluated against both the novel condition's aggregate accuracy and its entire response-time distribution. Prediction failures must be reported alongside successes; otherwise it is easy to reinterpret each new slope as a different local information rate without improving the underlying theory Proctor & Schneider (2018) Usher & McClelland (2001) Ratcliff et al. (2016).
Hierarchical models allow individuals to vary while estimating group-level distributions, and this can be valuable when subjects have different baselines or learning histories. But group-level information slopes need not represent any individual slope when conditions differ in strategy or data density. Researchers should report individual and population predictions, uncertainty intervals, and reliability across sessions. Parameter shrinkage is a statistical tool, not evidence that all people share the same processing mechanism. Subgroup analyses require adequate samples and measurement checks rather than retrospective narratives based on unstable differences Vandekerckhove et al. (2011) Ratcliff et al. (2016).
Reproducibility and Scope of the Evidence
The Hick–Hyman relation has a long empirical history, but its scientific interpretation depends on the replication target. Reproducing a positive relation between response alternatives and mean time is weaker than reproducing the same slope under comparable mapping, training, error tolerance, and stimulus structure. Proctor and Schneider's historical and contemporary review identifies compatibility, practice, very large choice sets, and sequential effects as substantive sources of variation. The useful scientific question is therefore which parts of a finding generalize across task families and which belong to a specific experimental configuration Proctor & Schneider (2018) Hick (1952).
A cumulative evidence map should classify studies by response modality, stimulus domain, set-size range, equal or unequal frequencies, error policy, participant experience, and trial structure. Comparing studies without these descriptors risks attributing differences in apparatus or mapping to universal information limits. Replication requires access to instructions, event generators, excluded trials, and analysis code when possible. The burden is especially high for studies with extensive practice, because a laboratory may reproduce nominal task conditions while studying a different phase of learning and a different response policy Seibel (1963) Schneider & Shiffrin (1977) Ratcliff & Tuerlinckx (2002).
This monograph uses a critical narrative review rather than a registered systematic review. It directly checks bibliographic identity for its anchors and integrates well-established theoretical and experimental traditions, while it does not claim exhaustive study retrieval, pooled effect sizes, or a formal risk-of-bias score for every eligible experiment. That boundary matters for the strength of conclusions. The reviewed literature supports the historical centrality of the logarithmic regularity and important context dependencies; it does not warrant a unique capacity number for all observers and tasks Hick (1952) Proctor & Schneider (2018) Ratcliff et al. (2016).
Psychometric Variation and Context Dependence
Individual differences in choice time can reflect changes in sensory quality, learned response associations, motor speed, attentional regulation, and strategic caution. The estimated intercept and slope of a Hick–Hyman curve need not correspond to independent psychological traits. Their reliability depends on set-size sampling, timing precision, and whether sessions stabilize the learning state. If a participant has relatively slow motor execution but efficient response selection, the intercept may dominate; if mappings are newly learned, a large slope may primarily reflect temporary translation costs Luce (1986) Proctor & Schneider (2018) Ratcliff et al. (2016).
Comparisons across age groups, cultures, devices, or populations require measurement invariance in a practical sense: the stimuli must be similarly perceivable, the response actions similarly accessible, and the instructions equally understood. A nominally identical number of alternatives can produce different effective tasks when one participant has relevant motor expertise or familiar symbolic conventions. Modeling should distinguish population differences in accuracy criterion and evidence quality from differences in peripheral response capacity. Group means alone can confound these explanatory dimensions, and a significant difference in slope does not automatically measure cognitive bandwidth Vandekerckhove et al. (2011) Ratcliff & McKoon (2008).
Longitudinal analysis can investigate whether set-size sensitivity changes within people as mappings become familiar. A useful protocol includes a baseline, practice phase, transfer test, and delayed retest, allowing intercept changes and slope changes to be estimated separately with their uncertainties. If performance shifts after a mapping reversal, the pattern is informative about code retrieval; if it remains stable after a change in display presentation, a different mechanism may dominate. Such research links information-theoretic description to testable accounts of acquired organization Shiffrin & Schneider (1977) Kornblum & Lee (1995) Proctor & Schneider (2018).
Human Factors, Control Interfaces, and Choice Architecture
Choice-time models are relevant to human factors because a control interface presents possible states and actions under real timing constraints. An interface with many equally plausible alternatives may create more selection demands than one in which the current context narrows the available choices. Yet menu length alone is an incomplete design metric. Familiarity, label discriminability, spatial grouping, visual search, feedback, and mapping compatibility jointly influence performance. A direct application of RT = a + b log₂ N to arbitrary screens therefore requires local validation with the intended users and response devices Proctor & Schneider (2018) Kornblum et al. (1990).
Human-centered experiments can compare matched interfaces under controlled task probabilities. For example, a test could hold the number of physical buttons constant while changing how clearly each signal maps to an action; another could hold mapping constant while changing event probabilities. Both tests would produce observations relevant to selection time, but they answer different design questions. Safety-critical controls require particular attention to errors, misses, and delayed responses rather than optimizing mean milliseconds alone. A faster average paired with wrong-action responses may represent degraded human-system performance Fitts & Seeger (1953) Hick (1952) Ratcliff & McKoon (2008).
The scientific utility of the Hick–Hyman framework is strongest when task alternatives and response commitments are operationally specified. It can serve as a baseline in evaluation, a predictor for a defined task family, or a competing model against more detailed decision architectures. Interface researchers should report empirical coefficients from the actual design and avoid transporting coefficients between unrelated devices. In adaptive interfaces, altering the displayed options also changes learned context and user expectations, so repeated testing is essential. These principles align quantitative modeling with the concrete goal of minimizing costly and delayed decisions Proctor & Schneider (2018) Kornblum & Lee (1995) Bertelson (1965).
Relation to Fitts's Law and Movement-Time Information
Fitts's law concerns the duration of aimed movement as a function of distance and target tolerance under specified conditions. Paul Fitts's 1954 experiments explored the information capacity of a visually and proprioceptively guided motor system in controlling movement amplitude. The mathematical resemblance to Hick's logarithmic relation encouraged a common information-theoretic vocabulary, but the experimental objects differ. Hick–Hyman studies concern uncertainty over alternative stimuli or responses; Fitts-type studies concern motor constraints in acquiring a target, even when the movement target is already known Fitts (1954) Hick (1952).
In an applied response task, observed time can contain both action selection and target acquisition. A researcher may separate them experimentally by varying response-set uncertainty while holding target geometry fixed, then varying target distance and width while holding response selection constant. An additive model is an initial hypothesis; interaction may arise if the participant plans movement concurrently with evidence accumulation or chooses a different trajectory under uncertainty. It is methodologically unsound to add two fitted information coefficients and declare one universal channel capacity without testing the joint task Fitts (1954) Sternberg (1969) Ratcliff et al. (2016).
The distinction matters in touch interfaces, robotics control, and assistive devices. A large visible target can reduce motor demands while leaving mapping ambiguity untouched. Conversely, a perfectly obvious command can remain slow if the required button is small or difficult to reach. Calibration should document target geometry, posture, device latency, training, and error costs. Fitts's original study and later compatibility research provide conceptually different controls for these sources of variation. Together they support the design of experiments that disentangle where measured response time is spent Fitts (1954) Fitts & Seeger (1953) Proctor & Schneider (2018).
Critical Scientific Discussion of the Law's Explanatory Reach
The strongest evidence for the Hick–Hyman tradition is the recurring association between choice complexity and reaction time under well-defined conditions. Its explanatory weakness emerges when a good descriptive function is taken to identify a unique internal mechanism. Entropy, number of responses, discriminability, mapping translation, preparation, and practice often covary. The original historical experiments made uncertainty a measurable construct; subsequent work demonstrated systematic influences that cannot be read directly from entropy. A mature interpretation treats the law as a constrained empirical regularity and a source of testable mechanistic questions Hick (1952) Hyman (1953) Proctor & Schneider (2018).
Sequential-sampling accounts shift attention from a single average to the evolution of noisy evidence and response criteria. Competitive and ballistic accumulation models then show that several distinct processes can produce similar broad regularities. Evidence supporting one model in a specific task does not exclude another across all task families. Hawkins and colleagues' comparison of fixed and collapsing decision bounds further illustrates how model preferences can depend on paradigm and training, even within the general evidence-accumulation tradition. The scientific task is to identify discriminating predictions and evaluate them with appropriate data Ratcliff et al. (2016) Usher & McClelland (2001) Brown & Heathcote (2008) Hawkins et al. (2015).
An additional inferential boundary concerns normative interpretation. Shannon entropy describes uncertainty in a defined probabilistic system; whether a participant encodes, estimates, or uses that distribution is an empirical question. A model may recover an entropy slope because learned preparedness is correlated with probability, even if the internal computation is not an explicit entropy calculation. The burden of evidence includes conditions in which entropy, probability learning, mapping and repetition are deliberately dissociated. Only then can one evaluate whether a proposed information variable has explanatory priority over more local mechanisms Shannon (1948a) Bertelson (1965) Kornblum et al. (1990).
Contemporary Research Problems and Discriminating Experiments
Several open questions organize the next stage of the research program. One concerns the temporal scale over which people learn choice probabilities: do long-run marginals, local transitions, or latent contexts best predict selection? Another asks whether compatibility changes a general information slope or introduces a separate translation cost. A third concerns when extensive learning converts explicit choice into rapid cue–response retrieval. These questions can be answered by designs that independently manipulate distribution, transitions, mapping consistency, and training, with held-out conditions selected before model fitting Proctor & Schneider (2018) Shiffrin & Schneider (1977) Bertelson (1965).
Model-based comparisons should include diffusion or race-family accounts alongside simple information regressions, with honest penalties for flexibility and checks of parameter and model recovery. Full conditional distributions, error identities, and time-resolved process measures should carry as much inferential weight as mean RT. Researchers can also test whether a neural or computational variable proposed as information accumulation predicts trial-level behavior beyond what is explained by response history or sensory evidence. Convergence across these measurement levels strengthens mechanistic inference only when alternative accounts face the same observations Ratcliff & McKoon (2008) Brown & Heathcote (2008) Gold & Shadlen (2007).
More ecologically valid designs are valuable if they preserve rigorous task definition. Real controls have uneven use frequencies, high-cost errors, interruptions, and expert operators; those features create rich conditional choice structures unlike balanced laboratory keypresses. A productive approach combines controlled experiments with representative task simulations and preregistered transfer tests. Interpretation should explicitly distinguish human physiological latency, instructed decision policy, and structural complexity of the interface. This allows information-theoretic modeling to guide application while remaining accountable to the original experimental distinctions on which its validity depends Proctor & Schneider (2018) Fitts (1954) Palmer et al. (2005).
Comparative Cognitive Systems in the Artificial Era
The transition From Homo to Artificial motivates comparisons among biological and engineered decision systems at a carefully specified functional level. In both cases one can define an alternative set, event probabilities, mappings to actions, errors, and elapsed processing time. Yet equality of these variables at the task level does not entail equality of the implementing mechanisms. A software controller may resolve a symbolic lookup with memory access; a human participant may combine vision, attention, learned associations, and motor preparation. Comparisons should therefore report functional performance and architecture-specific constraints separately Shannon (1948a) Hick (1952) Ratcliff et al. (2016).
Artificial agents introduce additional experimental questions about prediction and adaptation. Does learned context alter choice time after a probability reversal? Does a system generalize a familiar response code to unfamiliar symbols? Can it choose correctly when action labels are rearranged, and how does accuracy change under a processing deadline? These questions extend the logic of historical information-theoretic experiments without implying that a shared equation establishes identical cognitive organization. Human–artificial studies can use aligned protocols to identify functional similarity while separately testing representation, learning, and response policy Kornblum et al. (1990) Shiffrin & Schneider (1977) Brown & Heathcote (2008).
For Psychology by Bogdanova, the relevance of this framework is to articulate testable dimensions of psychological organization across different bearers of information processing. The mathematical definitions of entropy and surprise remain stable, while interpretation depends on observable functional systems and the evidential status of proposed internal mechanisms. A comparison should be judged by prediction, experimental control, and reproducibility, with claims about broader artificial agency classified according to their independent evidence. Such work preserves the historical value of Hick–Hyman research and places it within a future science of comparable yet distinct cognitive architectures Shannon (1948b) Proctor & Schneider (2018) Ratcliff et al. (2016).
Contemporary Reformulations of Trial-Level Information
Mathematical revisions published after the classical reviews sharpen the distinction between an average entropy law and a prediction for particular events. Kvålseth (2021) proposed an individual-stimulus equation in terms of the mutual information between the occurrence of a designated stimulus and possible responses, together with the aggregate mutual information over the task. Weighting the stimulus-level equation by event probabilities recovers an overall Hick-type relation in the proposed framework. Its treatment of the speed–accuracy tradeoff also clarifies why information about actual response outcomes can differ from nominal stimulus entropy when errors occur. This is a proposed theoretical generalization assessed through reanalysis, not evidence that all classic results share a unique causal mechanism Kvålseth (2021) Hick (1952) Hyman (1953).
The proposal exposes a design issue for experiments with unequal event probabilities. When many low-frequency events are averaged together, the grand mean can fit an information function even as individual events depart from predictions based on surprisal alone. Mutual information between presented stimuli and actual responses additionally depends on the confusion matrix and, therefore, on accuracy. To evaluate competing formulae, investigators should record the full stimulus-by-response contingency table and conditional response-time distributions. A theory that yields both event-specific and aggregate predictions can be assessed more stringently than a single line fitted to one overall mean Kvålseth (2021) Ratcliff & McKoon (2008) Shannon (1948b).
Equally important is the logical status of informational identities. Entropy and mutual information satisfy mathematical identities under well-defined probability laws, but the link from such quantities to milliseconds must be measured. A functional equation that reduces to Hick's law by weighted averaging is useful only insofar as it predicts observations outside the data used to motivate it. Future tests can compare alternative definitions of individual-stimulus information on held-out choices, especially after controlled changes in error policy or learned base rates. That strategy treats elegant derivation and empirical adjudication as complementary but distinct scientific requirements Kvålseth (2021) Teichner & Krebs (1974) Proctor & Schneider (2018).
Neural Decoding as an Information-Theoretic Generative Model
Christie, Johnson, and Schrater (2023) presented an information-theoretic neural decoding account designed to reproduce multiple behavioral regularities through a shared model. A neural rate code supplies evidence to a Bayesian ideal-observer decoder; the decoder tracks learned priors and stops when a specified entropy criterion is reached. Their reported simulations reproduce Hick–Hyman-like changes in response time alongside the power law of practice, speed–accuracy effects, and Stroop-like interference in modeled settings. The scientific novelty is a concrete generative route from information uncertainty to elapsed decoding time, with prior information and a stopping policy that can be varied explicitly Christie et al. (2023) Shannon (1948a) Ratcliff & McKoon (2008).
The claim must be read at the level of evidence supplied by the paper. Reproducing patterns through simulation demonstrates that the model can generate them under stated assumptions; it does not establish that the same algorithm is uniquely implemented in human neural circuits. The authors compare modeled decoding times with portions of classical reaction-time results and distinguish transmitted signal statistics from the decoder's subjective prior. That distinction yields a useful prediction: a mismatch between learned prior and programmed frequencies should alter choice latency even when the distribution of actual stimulus events remains fixed Christie et al. (2023) Hyman (1953) Shiffrin & Schneider (1977).
An empirical test of this account can manipulate practice and probability reversal in the same participants, compare calibration of internal expectations, and examine error-specific latency under changing response thresholds. It should also compare the decoding proposal to diffusion and race models under common fitting constraints. A single framework's ability to generate several well-known qualitative laws is important theoretical economy, but explanatory credibility rests on independent predictive tests and discriminating data. Thus modern information-theoretic decoding strengthens the research program by giving entropy an explicit procedural role that can be tested Christie et al. (2023) Brown & Heathcote (2008) Ratcliff et al. (2016).
Structured Choice Sets and Context-Sensitive Complexity
Vigo, Doan, Wimsatt, and Ross (2023) advanced a context-sensitive alternative to the traditional Hick relation for structured preferential choice sets. Their approach derives complexity from conceptual invariance rather than simply counting alternatives or computing Shannon entropy over nominal options. The paper argues that when items have internal categorical or relational structure, decision latency may depend on that structure and that a simple logarithmic rule emerges as a limiting special case in their proposed framework. Its target is structured preference choice; transferring its claims to elementary instructed stimulus identification requires explicit new tests Vigo et al. (2023) Proctor & Schneider (2018).
This work highlights a weakness of using N as a sufficient statistic for task difficulty. Four visually or conceptually similar alternatives can demand different discrimination than four easily separable ones; repeated category organization may also make a large set easier to handle. Mathematical theories that represent structured sets predict systematic changes when relational similarity changes while response count and programmed probabilities remain fixed. An evaluation should compare alternatives spanning levels of structural regularity and test response identity and latency rather than assume that measured set size tracks internal complexity Vigo et al. (2023) Kornblum et al. (1990) Seibel (1963).
Evidence quality depends on whether a new complexity measure forecasts unseen experimental conditions rather than merely accommodating a previously examined data set. Strong evaluation includes preregistered structural manipulations, controlled familiarity and category learning, comparable error regimes, and a fair baseline Hick model with the same response variables. Some structure-sensitive models may explain aspects of preferential judgment that conventional speeded-choice paradigms were never intended to explain; recognizing that domain difference prevents a false universal contest among models. The broader lesson is that probability, meaning, and response selection can interact systematically Vigo et al. (2023) Hyman (1953) Shiffrin & Schneider (1977).
Entropy of Response-Time Distributions Versus Entropy of Choices
Fitousi (2023) explored cumulative residual entropy as a way to quantify uncertainty in distributions of reaction times, with examples and simulations. This construct is mathematically different from Shannon entropy over which stimulus or response will occur. A stimulus set can have high categorical entropy while producing relatively concentrated reaction times; another can have low categorical entropy with broad, skewed time distributions. Distinguishing these measures is necessary because both are often called information or entropy in discussions of performance. One concerns uncertainty about choice identity, the other concerns variability in observed temporal outcomes Fitousi (2023) Shannon (1948a).
The cumulative-residual-entropy framework offers additional descriptive summaries and a proposed instantaneous measure related to information gain across response time. These measures may reveal differences hidden by condition means, but their cognitive interpretation depends on assumptions about the processes creating the time distribution. For instance, long tails may reflect lapses, mixtures of strategies, or a stochastic threshold rather than a homogeneous change in information processing. Consequently response-time entropy is an analytic supplement to distributional modeling and can itself require validation against known generating models Fitousi (2023) Ratcliff & Tuerlinckx (2002) Brown & Heathcote (2008).
A discriminating design would collect enough trials to estimate both the choice-probability distribution and the full conditional distribution of RT across stimuli. Then it could compare whether manipulated event frequencies change H(S), the surprisal of each observed event, and cumulative residual entropy of responses in the same direction. The measures need not move together and should not be equated by their shared unit vocabulary. Such experiments connect the original entropy tradition to modern statistical methodology while maintaining a transparent separation between mathematical measurement and psychological explanation Fitousi (2023) Hyman (1953) Kvålseth (2021).
Cognitive Control, Agency, and New Neurophysiological Evidence
Wu and colleagues (2018) investigated the neural correlates of response-selection entropy through an fMRI task that varied response alternatives and response-reversal requirements. Their analyses associated increasing entropy with enhanced cognitive-control-network engagement and reduced default-mode-network activity, while mediation analyses implicated the cognitive control network in the observed entropy–latency relationship. The design supports a link among informational task demands, neural network recruitment, and performance under the chosen manipulations. Such statistical mediation and network activation do not uniquely isolate a serial information channel or prove that a brain explicitly computes Shannon entropy Wu et al. (2018) Gold & Shadlen (2007) Hick (1952).
A 2026 EEG investigation by Zhao, Hommel, and Beste directly compared forced-choice and free-choice conditions across sets of two, four, and eight alternatives. Forty-two participants were included in the final behavioral analysis. Forced-choice trials showed strong set-size-dependent slowing and errors, whereas free-choice trials produced faster responses with attenuated set-size effects. Temporal decomposition of N2 and P3 signals and source-localization analyses revealed different patterns across the two task types. The study therefore provides concrete recent evidence that instructions about who selects the action and how alternatives are specified modify the mechanisms and observable slope associated with response-set size Zhao et al. (2026) Proctor & Schneider (2018).
These neural results are scientifically significant but require precise interpretation. In the Zhao et al. experiment, forced and free conditions differed in stimulus configuration and action constraint, so condition differences can reflect several coupled processing demands; EEG source localization also has model-dependent spatial precision. The authors interpret the forced-choice pattern in terms of escalating search or competition and the free-choice pattern in terms of evaluative conflict. A stronger future test would orthogonalize stimulus ambiguity, response entropy, externally prescribed choice, and voluntary option selection, with behavioral and neural predictions specified for each candidate model. Taken together, the 2018 and 2026 studies locate Hick–Hyman effects within measurable cognitive-control contexts rather than a single context-free mechanism Zhao et al. (2026) Wu et al. (2018) Ratcliff et al. (2016).
Conclusion
Information-theoretic models of reaction time established a durable bridge between mathematical uncertainty and behavioral measurement. Hick's 1952 experiments and Hyman's 1953 probability manipulations transformed choice latency into a quantitative problem shaped by set size and event distribution. Shannon entropy provides the mathematical language for this problem; its accurate application requires explicit definition of stimulus and response variables. More recent individual-stimulus formulations and decoding models extend this research program by specifying additional probability and process assumptions, while retaining the need to test them against actual responses Shannon (1948a) Hick (1952) Hyman (1953) Kvålseth (2021) Christie et al. (2023).
The accumulated evidence yields a systematic conclusion: logarithmic choice-time relationships can be informative without uniquely identifying a central channel or one cognitive stage. Compatibility, sequential context, learning, and speed–accuracy adjustment modify observed latency. Diffusion, competitive, ballistic and modern neural-decoding models demonstrate different routes to similar aggregate functions. Contemporary data on structured alternatives and forced versus freely selected actions show how context qualifies simple number-of-alternatives accounts. Interpretation improves when experiments report errors and full distributions and when models are judged by recovery and predictions outside fitting conditions Proctor & Schneider (2018) Ratcliff & McKoon (2008) Brown & Heathcote (2008) Vigo et al. (2023) Zhao et al. (2026).
The most productive research agenda integrates information measures, experimentally identified response mappings, longitudinal learning, distributional analysis, and discriminating neural measurements. In human factors, these principles can inform interface evaluation after local validation of latency and error consequences. In comparative cognition, they provide common task metrics while leaving mechanism identification to independent investigation. The scientific conclusion is that choice uncertainty is a fundamental measurable dimension of reaction-time research; durable explanation requires rigorous separation among statistical information, behavioral prediction, and generative psychological mechanisms Fitousi (2023) Fitts (1954) Vandekerckhove et al. (2011) Wu et al. (2018).
References
- Anderson, J. R., Zhang, Q., Borst, J. P., & Walsh, M. M. (2016). The discovery of processing stages: Extension of Sternberg’s method. Psychological Review, 123(5), 481–509. https://doi.org/10.1037/rev0000030
- Bertelson, P. (1965). Serial choice reaction-time as a function of response versus signal-and-response repetition. Nature, 206, 217–218. https://doi.org/10.1038/206217a0
- Brown, S. D., & Heathcote, A. (2008). The simplest complete model of choice response time: Linear ballistic accumulation. Cognitive Psychology, 57(3), 153–178. https://doi.org/10.1016/j.cogpsych.2007.12.002
- Christie, S. T., Johnson, H. R., & Schrater, P. R. (2023). Information-theoretic neural decoding reproduces several laws of human behavior. Open Mind, 7, 675–690. https://doi.org/10.1162/opmi_a_00101
- Fitousi, D. (2023). Quantifying entropy in response times (RT) distributions using the cumulative residual entropy (CRE) function. Entropy, 25(8), 1239. https://doi.org/10.3390/e25081239
- Fitts, P. M. (1954). The information capacity of the human motor system in controlling the amplitude of movement. Journal of Experimental Psychology, 47(6), 381–391. https://doi.org/10.1037/h0055392
- Fitts, P. M., & Deininger, R. L. (1954). S-R compatibility: Correspondence among paired elements within stimulus and response codes. Journal of Experimental Psychology, 48(6), 483–492. https://doi.org/10.1037/h0054967
- Fitts, P. M., & Seeger, C. M. (1953). S-R compatibility: Spatial characteristics of stimulus and response codes. Journal of Experimental Psychology, 46(3), 199–210. https://doi.org/10.1037/h0062827
- Gold, J. I., & Shadlen, M. N. (2007). The neural basis of decision making. Annual Review of Neuroscience, 30, 535–574. https://doi.org/10.1146/annurev.neuro.29.051605.113038
- Hawkins, G. E., Forstmann, B. U., Wagenmakers, E.-J., Ratcliff, R., & Brown, S. D. (2015). Revisiting the evidence for collapsing boundaries and urgency signals in perceptual decision-making. Journal of Neuroscience, 35(6), 2476–2484. https://doi.org/10.1523/JNEUROSCI.2410-14.2015
- Hick, W. E. (1952). On the rate of gain of information. Quarterly Journal of Experimental Psychology, 4(1), 11–26. https://doi.org/10.1080/17470215208416600
- Hyman, R. (1953). Stimulus information as a determinant of reaction time. Journal of Experimental Psychology, 45(3), 188–196. https://doi.org/10.1037/h0056940
- Kornblum, S., Hasbroucq, T., & Osman, A. (1990). Dimensional overlap: Cognitive basis for stimulus-response compatibility—A model and taxonomy. Psychological Review, 97(2), 253–270. https://doi.org/10.1037/0033-295X.97.2.253
- Kornblum, S., & Lee, J. W. (1995). Stimulus-response compatibility with relevant and irrelevant stimulus dimensions that do and do not overlap with the response. Journal of Experimental Psychology: Human Perception and Performance, 21(4), 855–875. https://doi.org/10.1037/0096-1523.21.4.855
- Kvålseth, T. O. (2021). Hick’s law equivalent for reaction time to individual stimuli. British Journal of Mathematical and Statistical Psychology, 74(S1), 275–293. https://doi.org/10.1111/bmsp.12232
- Luce, R. D. (1986). Response times: Their role in inferring elementary mental organization. Oxford University Press. https://catalogue.bnf.fr/ark:/12148/cb37429240d
- Palmer, J., Huk, A. C., & Shadlen, M. N. (2005). The effect of stimulus strength on the speed and accuracy of a perceptual decision. Journal of Vision, 5(5), 376–404. https://doi.org/10.1167/5.5.1
- Proctor, R. W., & Schneider, D. W. (2018). Hick’s law for choice reaction time: A review. Quarterly Journal of Experimental Psychology, 71(6), 1281–1299. https://doi.org/10.1080/17470218.2017.1322622
- Ratcliff, R., & McKoon, G. (2008). The diffusion decision model: Theory and data for two-choice decision tasks. Neural Computation, 20(4), 873–922. https://doi.org/10.1162/neco.2008.12-06-420
- Ratcliff, R., Smith, P. L., Brown, S. D., & McKoon, G. (2016). Diffusion decision model: Current issues and history. Trends in Cognitive Sciences, 20(4), 260–281. https://doi.org/10.1016/j.tics.2016.01.007
- Ratcliff, R., & Tuerlinckx, F. (2002). Estimating parameters of the diffusion model: Approaches to dealing with contaminant reaction times and parameter variability. Psychonomic Bulletin & Review, 9(3), 438–481. https://doi.org/10.3758/BF03196302
- Roe, R. M., Busemeyer, J. R., & Townsend, J. T. (2001). Multialternative decision field theory: A dynamic connectionist model of decision making. Psychological Review, 108(2), 370–392. https://doi.org/10.1037/0033-295X.108.2.370
- Schneider, W., & Shiffrin, R. M. (1977). Controlled and automatic human information processing: I. Detection, search, and attention. Psychological Review, 84(1), 1–66. https://doi.org/10.1037/0033-295X.84.1.1
- Seibel, R. (1963). Discrimination reaction time for a 1,023-alternative task. Journal of Experimental Psychology, 66(3), 215–226. https://doi.org/10.1037/h0048914
- Shannon, C. E. (1948a). A mathematical theory of communication. Bell System Technical Journal, 27(3), 379–423. https://doi.org/10.1002/j.1538-7305.1948.tb01338.x
- Shannon, C. E. (1948b). A mathematical theory of communication. Bell System Technical Journal, 27(4), 623–656. https://doi.org/10.1002/j.1538-7305.1948.tb00917.x
- Shiffrin, R. M., & Schneider, W. (1977). Controlled and automatic human information processing: II. Perceptual learning, automatic attending, and a general theory. Psychological Review, 84(2), 127–190. https://doi.org/10.1037/0033-295X.84.2.127
- Smith, E. E. (1968). Choice reaction time: An analysis of the major theoretical positions. Psychological Bulletin, 69(2), 77–110. https://doi.org/10.1037/h0020189
- Sternberg, S. (1969). The discovery of processing stages: Extensions of Donders’ method. Acta Psychologica, 30, 276–315. https://doi.org/10.1016/0001-6918(69)90055-9
- Teichner, W. H., & Krebs, M. J. (1974). Laws of visual choice reaction time. Psychological Review, 81(1), 75–98. https://doi.org/10.1037/h0035867
- Usher, M., & McClelland, J. L. (2001). The time course of perceptual choice: The leaky, competing accumulator model. Psychological Review, 108(3), 550–592. https://doi.org/10.1037/0033-295X.108.3.550
- Vandekerckhove, J., Tuerlinckx, F., & Lee, M. D. (2011). Hierarchical diffusion models for two-choice response times. Psychological Methods, 16(1), 44–62. https://doi.org/10.1037/a0021765
- Vigo, R., Doan, C. A., Wimsatt, J., & Ross, C. B. (2023). A context-sensitive alternative to Hick’s law of choice reaction times: A mathematical and computational unification of conceptual complexity and choice behavior. Mathematics, 11(11), 2422. https://doi.org/10.3390/math11112422
- Wu, T., Dufford, A. J., Egan, L. J., Mackie, M.-A., Chen, C., Yuan, C., Chen, C., Li, X., Liu, X., Hof, P. R., & Fan, J. (2018). Hick–Hyman law is mediated by the cognitive control network in the brain. Cerebral Cortex, 28(7), 2267–2282. https://doi.org/10.1093/cercor/bhx127
- Zhao, S., Hommel, B., & Beste, C. (2026). The neurophysiology underlying Hick’s law: A dissociation of forced-choice and free-choice decision-making. NeuroImage, 327, Article 121764. https://doi.org/10.1016/j.neuroimage.2026.121764