RESEARCH DOSSIER
Choice Allocation and the Matching Law
Abstract. Choice allocation describes the distribution of behavior among concurrently available alternatives as a function of reinforcement contingencies, temporal conditions, and the costs of action. This fundamental research monograph traces the experimental origins of the matching law from Findley's concurrent-schedule procedures to Herrnstein's proportional relation, Baum's generalized matching formulation, and the subsequent development of melioration, maximization, delay-reduction, molecular choice, and modern quantitative matching theories. It integrates original experimental studies in pigeons, rats, and human participants with methodological analyses of reinforcement ratios, response and time allocation, switching and changeover delays, reinforcer amount and effort, steady-state preference, dynamic adaptation, and response vigor. Particular attention is given to model discrimination, endogenous obtained reinforcement, absolute versus relative response rate, temporal structure, and the evidentiary status of applications to human behavior and applied behavior change. The concluding synthesis specifies a research program for identifying choice mechanisms through experimentally divergent predictions and cross-environment validation, including controlled artificial-agent simulations. This is an evidence-based critical narrative synthesis of identifiable published literature, not a registered systematic review or meta-analysis.
Keywords: choice allocation; matching law; generalized matching law; operant choice; concurrent schedules; relative reinforcement; reinforcer sensitivity; bias; undermatching; melioration; maximization; behavioral economics; changeover delay; concurrent chains; reinforcement learning; behavioral dynamics
Choice Allocation as a Scientific Problem
Choice allocation concerns the distribution of observable behavior across two or more concurrently available courses of action as their consequences change. The principal dependent variable is an allocation of responses, time, or visits; the principal experimental manipulation is a distribution of reinforcers among alternatives. A useful research question specifies the response topographies, access constraints, reinforcement contingencies, and measurement interval. An organism can exhibit strong preference while continuing to sample the less productive alternative, and that allocation can be characterized mathematically without attributing a conscious calculation of utility. The historical core of the field is the relation between relative behavior and relative obtained reinforcement described in Herrnstein (1961).
The matching law is the empirical regularity that, under certain concurrent reinforcement arrangements, relative response rates or relative time allocations approximate relative reinforcement rates. For two alternatives, the strict response form is B1/(B1+B2)=R1/(R1+R2), where B denotes behavior and R denotes obtained reinforcement in a common observation period. This relation specifies a proportional distribution, not necessarily equality between absolute response and reinforcer counts. The generalized relation introduces sensitivity and bias parameters, thereby making systematic departures measurable Baum (1974). Throughout this monograph, a claim of matching refers to a specified schedule, timescale, measure, and fitting procedure.
The scope includes concurrent variable-interval schedules, response switching and changeover constraints, the effects of reward amount and delay, choice acquisition and adaptation, time allocation, steady-state and local models, formal estimation, and research with human and nonhuman populations. Those topics together constitute a coherent quantitative program in experimental learning and behavioral psychology. Its methods connect to behavioral economics and computational modeling when equivalent response alternatives can be specified, but the mere occurrence of a decision does not demonstrate the matching relation. Conceptual proximity to broad judgment and decision-making remains distinct from the operant experimental object defined by reinforcement feedback functions Davison & McCarthy (1988).
This is a critical narrative synthesis of identifiable historical experiments, subsequent experimental comparisons, analytical models, and methodological reviews. Studies were selected for direct relevance to choice allocation, influence on model development, and capacity to discriminate between explanations. The synthesis does not claim an exhaustive registered search, meta-analysis, or newly collected participant data. A paper that describes a model is assessed separately from a paper that tests its predictions, and successful fits to steady-state proportions are distinguished from an explanation of how individual responses are generated McDowell (2005).
Historical Origins of Quantitative Choice Research
The foundations of matching research lie in the experimental analysis of operant behavior, where investigators related response rates to schedules of consequences. Earlier response-strength approaches emphasized the frequency with which an action was reinforced; choice experiments made the surrounding alternatives explicit. Findley (1958) examined preference and switching under concurrent scheduling and established procedures for observing sequences of allocation. This experimental arrangement exposed the importance of available alternatives, schedule interactions, and changes between options rather than treating a focal response as an isolated unit.
The decisive result associated with Herrnstein (1961) came from pigeons responding on concurrent variable-interval schedules of food reinforcement. Across manipulated relative reinforcement rates, the distribution of responses closely tracked the distribution of obtained reinforcers. This observation allowed investigators to formulate a compact quantitative relation whose predictions could be challenged with other schedule structures. The original experiment must be read in its procedural context: pigeons, experimentally arranged contingencies, repeated response opportunities, and switching conditions. Its contribution was a reproducible question and quantitative pattern, not a universal statement about every kind of choice.
In Herrnstein (1970), the matching relation became part of a broader quantitative account of the law of effect. The analysis related response rates under concurrent, single, and multiple schedules to reinforcement obtained for the focal and other activities. It made alternative reinforcement a central explanatory variable and proposed ways to connect relative allocation with absolute response strength. Herrnstein (1974) developed formal properties of matching, including saturation as the focal alternative approaches the entire distribution of reinforcement. These theoretical extensions generated distinct tests because a model accounting for relative ratios need not explain absolute response rates or behavioral contrast.
Research then shifted from an ideal proportional law toward precise descriptions of recurrent deviations. Baum (1974) expressed response ratios as a function of reinforcement ratios and separated bias from sensitivity, furnishing the framework now called generalized matching. Later work increasingly studied the acquisition of preference, switching organization, schedule feedback, delayed outcomes, and mathematical explanations of why the aggregate relation emerges. The historical sequence is therefore cumulative: an empirical regularity was established, generalized mathematically, compared against competing mechanisms, and repeatedly reexamined under altered schedules Davison & McCarthy (1988).
Concurrent Schedules and the Measurement of Allocation
A concurrent schedule makes at least two operants available during an overlapping period, with each operant correlated with a consequence schedule. In the classic concurrent variable-interval–variable-interval arrangement, reinforcement becomes available according to a time-based schedule and is delivered when the corresponding response occurs. The programmed intervals, realized reinforcer availability, and actually obtained reinforcement rate are distinct variables. An organism's switching may alter how many scheduled opportunities become obtained outcomes. Thus a correlation between obtained reinforcers and responses can contain reciprocal effects of the behavioral allocation being measured Findley (1958).
Response allocation is commonly summarized as B1/(B1+B2), while time allocation is T1/(T1+T2). These are related but empirically separable measures: an animal may respond faster on one option while remaining there for less time. Changeover count, visit duration, run length, local rate of responding, and postreinforcement residence time add information about temporal organization. MacDonall (2003) showed that stay and switch dynamics can be analyzed using their reinforcement contingencies, a perspective that shifts attention from aggregate response counts toward the structure of a choice sequence.
Many standard procedures use a changeover delay, a short period after moving to the other alternative during which switching cannot immediately produce reinforcement. Its function is to diminish reinforcement of rapid alternation and make preference sensitive to the ongoing schedules. The particular delay is an experimental parameter, not a neutral convenience. Changing its duration can change switching frequency, local response rates, and apparent conformity to matching. The literature therefore requires explicit reporting of schedule independence, changeover contingencies, stimulus arrangements, and response recording rather than reporting only final percentages Jozefowiez et al. (2009).
Sound experimental comparison normally uses repeated conditions within the same subject, systematic variations of reinforcement ratios, observation periods long enough to characterize adaptation, and counterbalancing or reversal where appropriate. Individual-subject curves preserve information hidden by a group-average line. A fit based on relatively few reinforcement ratios may look impressive even when it cannot predict transfer, switch timing, or subsequent manipulations. The distinction between arranged and obtained reinforcement is especially important when two subjects encounter the same programmed schedule but allocate behavior differently Alsop & Elliffe (1988).
The Strict Matching Relation and Its Interpretation
The two-alternative strict matching equation may be written B1/B2=R1/R2, for positive denominators, or equivalently B1/(B1+B2)=R1/(R1+R2). The algebra states that the ratio of behavior equals the ratio of obtained reinforcement and is invariant to a common multiplication of both obtained rates. For example, if one alternative yields three quarters of obtained reinforcers, strict matching predicts three quarters of the defined behavior measure on that alternative. This numerical illustration follows from the formula and does not constitute an observation from a particular experiment. The equation originated in an experimental research program rather than in a claim that all agents optimize their long-run returns Herrnstein (1961).
The equation must be interpreted with care because its variables are often mutually dependent. Under variable-interval schedules, additional time on an option may change the frequency of reinforcement obtained there; the resulting data points can reflect feedback between the organism and the apparatus. The same measured proportions can also be consistent with several behavioral processes operating at a finer time scale. This is the central identification problem: an equilibrium relation between aggregated quantities does not by itself specify the causal process producing that equilibrium McDowell (2005).
Strict matching makes stronger predictions than mere monotonic preference. It predicts a particular slope and intercept when response ratios are transformed to logarithms and plotted against reinforcement ratios. A strong preference for richer reinforcement is compatible with undermatching, overmatching, or biased allocation. Similarly, nearly perfect matching in a single condition provides weak evidence because a broad range of processes can reproduce that proportion. Adequate examination needs multiple ratios, precision estimates, analysis of bias, and tests outside the conditions used to fit the parameters Baum (1974).
The relative and absolute formulations also demand separation. Herrnstein (1970) proposed equations for response strength that include extraneous reinforcement and an asymptotic rate parameter. Their empirical adequacy must be examined independently from the proportional allocation law. The constancy assumptions governing absolute-rate models have been contested by later experimental and theoretical research McDowell (2005). The most defensible summary is that the strict law identifies a benchmark relation in selected concurrent schedules, while broader accounts of action strength remain hypotheses requiring additional measurements.
Generalized Matching, Sensitivity, and Bias
The generalized matching equation is log(B1/B2)=a log(R1/R2)+log b, where a is the sensitivity of behavioral allocation to relative reinforcement and b is a multiplicative bias. On the original scale, B1/B2=b(R1/R2)^a. Strict matching corresponds to a=1 and b=1. Values of a below one describe undermatching, in which allocation changes less extremely than relative reinforcement; values above one describe overmatching. Values of b other than one describe a stable asymmetry associated with a side, response effort, stimulus property, or preference not absorbed by the chosen reinforcement metric Baum (1974).
Bias and sensitivity answer different experimental questions. If reinforcer ratios are equal, bias may favor one alternative. If reinforcer ratios vary, the slope of the log-ratio function indicates how far response allocation moves with those variations. An apparently flat slope can arise from weak discrimination of alternatives, rapid switching, motivational limits, poor schedule control, or pooling of heterogeneous observations. A model fit cannot assign one cause to the parameter without independently manipulating the suspected mechanism. In Alsop & Elliffe (1988), sensitivity varied with overall reinforcer rate even when the relative rate was manipulated systematically, indicating that background context matters.
Logarithmic estimation also involves choices about zero counts, sparse observations, measurement error, and response dependencies. A zero reinforcer or response rate makes the log ratio undefined; arbitrary pseudocounts can shift estimated slopes. Investigators should report rules for such cases and, where possible, fit probabilistic or count-data alternatives that reflect the observation process. Confidence intervals around a and log b, individual-level estimates, residual diagnostics, and checks across condition order are more informative than a correlation coefficient alone Davison & McCarthy (1988).
Generalized matching has broad descriptive usefulness partly because it explicitly admits deviations. That advantage has an inferential cost: the flexible parameters may fit patterns produced by mechanistically different processes. It should be compared against alternative models with appropriate penalties for flexibility, held-out conditions, or predictive data. A successful generalized-matching line is an empirical regularity at the specified level of aggregation, while theories of melioration, maximization, or dynamic learning offer distinct accounts of its origin Vaughan (1981).
Overall Reinforcement Rate and Context
The absolute density of reinforcement can modify choice even when the ratio between options remains unchanged. In the central experiment by Alsop & Elliffe (1988), six pigeons experienced series of concurrent variable-interval conditions arranged at different overall reinforcement rates while relative reinforcement availability was varied within each series. Both response and time allocation could be described with generalized-matching functions, yet estimated sensitivity declined when the overall reinforcer rate was lower. This finding matters because a strictly ratio-based theory might predict the same preference function whenever the relative ratio is preserved. Absolute opportunity, the temporal grain of feedback, and the discriminability of alternatives must also be considered.
The manipulation in Alsop & Elliffe (1988) offers a useful methodological lesson. Overall rates were arranged across conditions, and preference was measured with both behavioral proportions and changeover rates. Reduced sensitivity was not explained adequately by the reported changes in changeover rates alone; the authors considered accounts involving the discrimination of local reinforcement conditions. Their interpretation is an explanatory proposal, while the observed interaction between overall reinforcement density and preference is the empirical result. The distinction protects the reader from treating one parameter of a fitted equation as a direct physiological or cognitive mechanism.
Models incorporating background reinforcement include the hypothesis that behavior depends on the focal option relative to other sources of reinforcement Herrnstein (1970). In natural settings, competitors extend far beyond the two recorded responses: rest, exploration, social contact, task avoidance, and unmeasured rewards may alter the opportunity structure. This makes laboratory arrangements scientifically valuable, because investigators can specify most immediate competitors. Extrapolation to open environments requires measuring at least the most consequential unobserved alternatives before concluding that a change in preference reflects altered reward sensitivity.
The general methodological consequence is that reinforcement ratio and reinforcement density should be crossed experimentally rather than treated as interchangeable. When possible, studies should also vary session duration and document stability after reversals. A model capable of reproducing response proportions after prolonged exposure may fail to predict the time needed to approach those proportions. Inferences about sensitivity become stronger when the manipulation predicts not only average allocation but also adjustment dynamics, switching, and residual patterns across individuals Davison & McCarthy (1988).
Reinforcer Magnitude, Quality, and Integrated Value
Reinforcement is multidimensional. Two alternatives may differ in how often a reward occurs, how much is delivered, how long delivery is delayed, and how the organism values the outcome in the current motivational context. A simple rate-based matching relation using only counts can therefore misrepresent the effective consequences. An investigator may compare matching to numbers of reinforcers with matching to weighted or transformed reinforcement magnitude, but that transformation must be independently motivated and specified before it is fitted to behavior Fantino (1969).
The work of Kyonka (2008) experimentally examined choice in transition while reinforcer rates and magnitudes varied. Its design illustrates why an analysis of terminal preference can overlook distinct effects of the dimensions of reinforcement during adaptation. A larger delivered amount can shift local choice without producing the same temporal pattern as a comparable increase in delivery frequency. The empirical interpretation depends on how conditions were arranged, how reinforcement was obtained, and how rapidly the organism sampled changes. Differential reward value is a scientific construct requiring operationalization rather than a synonym for every observable preference.
Reinforcer quality is especially challenging because a change in the kind of outcome may alter satiation, consumption time, motivational state, and switching cost at once. Comparing pellets with fluid, or different kinds of food, therefore requires separate measures of ingestion and preference when available. One attractive but potentially circular analysis simply defines the 'value' of each outcome from the very choice proportions the value estimate is intended to explain. Independent preference assessments, demand manipulations, and out-of-sample predictions provide stronger tests Rachlin et al. (1981).
The quantitative question is whether separate dimensions combine multiplicatively, additively, or through context-dependent functions. No single combination rule follows directly from the strict matching equation. Concurrent-chains research shows that conditioned stimuli and terminal delays can also govern allocation in ways that immediate primary reinforcement counts do not describe Fantino (1969). Accordingly, investigators should keep rate, amount, delay, quality, and stimulus value conceptually distinct until the particular model and experiment justify combining them.
Response Effort and the Costs of Choice
Every response requires some investment of time or activity, and concurrent options may differ in force, travel, key position, sequence length, or other response requirements. These costs influence which consequences can be contacted. A change in allocation after effort is raised may reflect differential energetic demands, slower response throughput, altered delay to reinforcement, or the loss of an opportunity to respond elsewhere. Effort therefore has to be manipulated with the other schedule parameters held as constant as possible Fantino (1968).
In Fantino (1968), required response rates were experimentally connected with choice, demonstrating that schedule demands belong in a quantitative account of preference. The methodological significance extends beyond a particular result: an experiment that labels two alternatives as equivalent while their response requirements differ cannot interpret the preference ratio solely as sensitivity to reinforcement frequency. Investigators must compare physical effort, time costs, and effective reinforcement rates under the achieved behavior. A difficult response may generate fewer reinforcers precisely because the subject spends less time performing it.
Behavioral economic frameworks express a related problem through opportunity costs and response prices. Maximization models ask which distribution of behavior would yield the largest total payoff under a specified feedback function Rachlin et al. (1981). But a formal optimum depends on the cost function and constraints included in the calculation. Some of the experimental appeal of melioration and other local rules comes from the possibility that observed choice follows proximal differences even when global payoff could be greater under another distribution Vaughan (1981).
In applied work, choice architecture may be changed by reducing the work necessary to contact a constructive reinforcer instead of increasing its frequency. The scientific claim that effort is causally important is strongest when the effort manipulation is experimentally separated from outcome quality and delay. A broad appeal to convenience or friction, however plausible, is insufficient to identify a matching-law mechanism. Response allocation and obtained consequences must be observed directly to establish the claimed operant relation McDowell (2021).
Switching, Changeover Delays, and Visit Structure
Relative response counts summarize behavior over a period but conceal the sequence by which those counts were produced. A subject might alternate rapidly between equal alternatives, switch only after reinforcement, or remain on one option for extended runs while occasionally sampling the other. Identical relative allocations can emerge from these different microstructures. For that reason, changeovers and visit duration are constitutive measurements of concurrent choice research, not optional behavioral curiosities Findley (1958).
A changeover delay is a programmed interval after switching during which responses on the new alternative cannot immediately earn the scheduled consequence. By reducing reinforcement of switching itself, it can change the balance between remaining and exploring. MacDonall (2003) analyzed schedules that reinforced staying and switching within a changeover-delay arrangement, highlighting how procedural definitions create different behavioral currencies. The relevant event may be a visit, a response emitted during a visit, or a transition between alternatives. Each currency has a different denominator and can imply a different form of matching.
Local choice theories sometimes treat every response as an independent selection between alternatives. This assumption becomes questionable when responding is organized in bouts or visits, because the probability of another response on the same option depends strongly on the state generated by preceding actions. History-sensitive models explicitly represent this dependence; summary ratio models generally do not. Shimp (1966) investigated probabilistic reinforcement and momentary choice in pigeons, providing an early challenge to the idea that aggregate proportions suffice for an explanation of response sequences.
Good evidence about switching requires recording timestamps, explicit rules for classifying visits, treatment of very short gaps, and transparency about whether reinforcers remain scheduled after departures. Multiple analyses can be reported: global behavior ratio, time ratio, transition probabilities, run lengths, and hazard of switching after reinforcement. The combination offers ways to distinguish between equivalent aggregate fits. A general law of choice gains explanatory depth when its parameters forecast temporal data from conditions withheld during estimation Jozefowiez et al. (2009).
Time Allocation, Response Rates, and Behavioral Vigor
Matching may be evaluated for response allocation or for time allocation. These measures need not agree because the organism can emit responses at different rates during its visits to competing alternatives. If time allocation matches obtained reinforcement while response allocation undermatches, a difference in local response rates may explain part of the disagreement. The analysis should therefore present time, responses, and reinforcers in common observation windows and define how periods of inactivity are counted Baum (1974).
Behavioral vigor adds a question about the absolute intensity of responding once an alternative is selected. The fraction of responses devoted to an option can remain constant while the sum of all response rates rises or falls. Herrnstein (1970) developed quantitative accounts linking absolute rate to reinforcement within a larger pool of opportunities, though the adequacy of its assumptions has been debated McDowell (2005). The empirical task is to specify whether the same model can predict the proportion of actions and the number of actions per unit time.
Time spent consuming reinforcement, collecting rewards, and changing options may not be available for operant responding. These intervals can introduce mechanical changes into observed response vigor independently of the subject's preference. A schedule that produces more frequent rewards may reduce free time for responses even as it attracts greater choice allocation. Clear denominators and a distinction between session time and available response time help prevent a misleading comparison of vigor across conditions Davison & McCarthy (1988).
A complete model of choice thus has at least three related outputs: which alternative is visited, how long the organism remains, and how vigorously it responds there. The strict matching equation addresses a specific aggregate relation. Extended dynamic models attempt to explain the process by which allocations and rates coevolve under available reinforcement. These outputs should be evaluated separately before they are treated as parts of one explanatory mechanism Jozefowiez et al. (2009).
Acquisition, Reversal, and Transition Dynamics
Stable allocation is the endpoint of a learning history. A theory that fits the terminal proportions must also explain how the distribution shifts after the richer alternative changes. During reversal, the organism may persist in its previous preference, sample the newly richer option, or change behavior abruptly after a local reinforcer. The relevant observations occur across sessions and within sessions, not simply in an averaged final block. The experiment by Kyonka (2008) demonstrates how independently varying reinforcement rate and magnitude can expose the contribution of current-session conditions to ongoing allocation.
Rapid adaptation and long-run sensitivity need separate outcome measures. One may estimate a within-session trajectory of response ratio, a lag from reinforcement to switching, and the stability of an asymptotic preference. Each can respond differently to a change in rate or magnitude. In Kyonka (2008), four pigeons encountered unpredictable session-level variation in the two reinforcement dimensions; analysis reported significant current-session effects of both. This evidence supports adaptive choice under controlled laboratory conditions but does not by itself establish a single universal updating equation.
A causal description of learning requires a model of how recent consequences influence the probability of future responses. Candidate explanations include recency-weighted reinforcement traces, local payoff comparisons, and rule-based switching. The matching relation often constrains an equilibrium without identifying which of these updates generated it. Shimp (1966) developed a momentary perspective on probabilistically reinforced choice, providing a historical basis for examining individual sequences rather than relying exclusively on stabilized averages.
To compare dynamic models, a researcher can present identical unexpected changes under different preceding histories. If two histories produce equal current reinforcement ratios but distinct switching trajectories, the history variable is theoretically important. Investigators should distinguish the expected payoff programmed by the experimenter, the reward information accessible to the organism, and the response actually observed. The result is a more discriminating research program: an explanation must forecast transitions under novel conditions as well as fit a session's final choice percentages Jozefowiez et al. (2009).
Melioration as a Local Rule of Behavioral Allocation
Melioration denotes movement toward an alternative that currently offers the better local return, with evaluation based on an option's experienced payoff rather than a global optimum calculated over all allocations. In a concurrent setting, a shift can change the experienced reinforcement rates because the organism's own behavior affects the feedback function. Melioration therefore predicts a dynamic tendency that may result in matching under some schedule arrangements and in systematic deviation under others. The key theoretical task is to specify local returns, the observation window, and the rule for reallocating behavior Vaughan (1981).
The experiment in Vaughan (1981) is particularly informative because it manipulated the relationship between allocation and payoff to make rival predictions diverge. In one condition, several theories predicted the same allocation range, and the pigeons' behavior entered that range. In another, melioration predicted movement toward a different range, even though such movement could lower total obtained reinforcement. The reported three birds ultimately moved within the melioration-predicted region. That finding discriminated among the tested proposals in the arranged environment without establishing that every natural choice is driven by local improvement.
Melioration's explanatory appeal arises from the interaction of behavior and contingent consequences. A response rule that adjusts toward recent returns can be implemented without assuming complete knowledge of future reward schedules. But the label alone offers little predictive content until the researcher defines what counts as a local reinforcement rate, how the rate is estimated, and how strongly a new observation updates the allocation. Different parameterizations can converge to different equilibria or respond differently after reversals Davison & McCarthy (1988).
The lesson for contemporary choice research is to design environments in which locally attractive actions and globally best allocations diverge. An apparent matching result obtained when the two theories predict the same behavior cannot discriminate between them. A stronger test holds overall reinforcement opportunities and action requirements under experimental control while varying the shape of the feedback function. Direct temporal measures make it possible to evaluate whether observed movement follows the hypothesized local gradient Vaughan (1981).
Maximization and Behavioral Economics
Maximization models define the preferred allocation as the one that yields the greatest total expected payoff given an explicit schedule and its costs. This is a mathematical optimization statement whose empirical predictions depend on the exact feedback relation. Under some ordinary concurrent variable-interval schedules, maximization and matching can make similar predictions. Under other arrangements, maximizing total reinforcement may require persistent preference for one option or allocation patterns that differ markedly from generalized matching. A theoretical contest is informative only when these alternative predictions are specified before data are examined Rachlin et al. (1981).
Behavioral economic approaches also analyze opportunity costs, substitute reinforcers, response price, and constraints on action. Rachlin et al. (1981) located maximization arguments within behavioral psychology and illustrated why the payoff function must be experimentally meaningful. An analyst should not assume that the maximum is calculated by a participant. Optimization may describe an outcome produced by learning or selection without representing the subject's moment-to-moment process. Conversely, an organism can behave adaptively under limited information without reaching the mathematically highest return available to an omniscient observer.
Apparent failures of maximization can stem from incomplete measurement of costs, imperfect learning, temporal discounting, uncertainty, satiation, or the reinforcement of switching itself. These possible explanations are not interchangeable. A revised maximization account becomes difficult to test when every discrepancy is attributed after the fact to an unmeasured subjective cost. Strong designs independently quantify these costs and then predict allocation under novel trade-offs. The experimental contrast in Vaughan (1981) remains significant precisely because it created conditions in which the relevant models made divergent predictions.
Matching and economic maximization thus occupy different explanatory levels. Matching is ordinarily a descriptive relation among observed allocations and obtained reinforcement; maximization evaluates allocations against a proposed objective function. A situation can satisfy both or only one, depending on the environment. Generalized matching may remain a useful summary when a maximization model is the better account of outcomes, and vice versa. This distinction avoids treating competing mathematical forms as philosophical identities for a choosing organism Davison & McCarthy (1988).
Momentary Maximizing and Molecular Choice
Molecular accounts seek to predict the next response or switch from the current local history of reinforcement, time, and stimulus conditions. In contrast, a molar matching curve summarizes the relation over an extended interval. A molecular model may generate the molar relation as an emergent consequence, yet different molecular rules can produce nearly identical aggregate allocation. Shimp (1966) provided a classic experimental analysis of probabilistically reinforced choice and helped establish the relevance of response-by-response contingencies.
One family of momentary choice hypotheses supposes that an action is selected on the basis of the currently expected likelihood or return of immediate reinforcement. The informational requirements of that rule must be specified: does the organism estimate time since the last reinforcer on each option, observe a change in stimulus state, or track average local payoff? Its predictions may differ according to whether variable-interval schedules operate independently and whether scheduled rewards accumulate while an option is not being visited. These procedural details often distinguish plausible local accounts more sharply than an overall matching coefficient Jozefowiez et al. (2009).
Local stay–switch analyses provide a second route. In MacDonall (2003), run lengths and visit durations were considered in relation to reinforcers for staying and switching. Experimental changes involving changeover delays helped evaluate this local model. Because those measurements retain sequence information, they can test predictions that response-ratio matching alone leaves ambiguous. Still, a model developed for rats under particular schedule arrangements requires cross-procedure testing before being treated as an explanation of human choices, delayed outcomes, or complex behavioral sequences.
A strong contemporary study should preserve raw event times, show behavior both as aggregate ratios and local transitions, and compare models on held-out switching histories. It can also test whether parameters inferred from one temporal aggregation remain stable at another. If a model fits only because an arbitrary time window was chosen, its psychological interpretation is fragile. The program of linking molecular and molar levels remains important because it asks how repeated local interactions produce durable patterns of preference McDowell (2021).
Formal, Classic, and Modern Matching Theories
The name matching theory has been used for more than the empirical proportional law. Classic theoretical formulations attempted to derive absolute response rates and other outcomes from fixed constraints, including a constant overall rate-related parameter. McDowell (2005) argued that key assumptions of that classic formulation conflict with experimental data and that generalized, power-function approaches possess different logical properties. The criticism concerns a theoretical package, not the mere observation that relative response ratios can approximately covary with relative reinforcement.
An especially informative direct test is Dallery et al. (2005). Six human participants responded on concurrent variable-interval schedules for monetary reinforcement while the amount associated with one alternative changed. A formal account and a modern theory were assessed against response and reinforcement rates. Although the formal equations accounted for much of the variance, their residuals showed a systematic relation to predicted response rates; the modern theory's equations provided a more satisfactory description in that experiment. Model fit quality therefore cannot be judged by variance explained alone.
The distinction between theories is also visible in the interpretation of parameters. A biased generalized matching function can describe allocation while allowing absolute output to change. A theory predicting how total response rate should behave under altered reinforcement magnitude must make further assumptions about response strength, background activities, and asymptotic parameters. Dallery et al. (2005) evaluated these kinds of predictions in human participants, offering convergent but bounded evidence for a modern account. The sample and laboratory task matter for generality.
Formal assessment benefits from explicit lists of shared and divergent predictions. A single curve can be treated as a descriptive approximation; a mechanistic model must explain the curve, and an absolute-rate theory must additionally predict levels of responding. Such distinctions reduce disputes caused by using 'matching' to refer simultaneously to data, an equation, a process, and a broad philosophical perspective on behavior. The history reviewed by McDowell (2005) illustrates the scientific value of separating these claims.
Concurrent Chains and Conditioned Reinforcement
Concurrent-chains schedules extend elementary concurrent choice by requiring an initial-link action to produce entry into a terminal link, where primary reinforcement follows under another contingency. The choice response therefore selects a predictive stimulus and a future route to reward. The experimentally important variable is not merely how often food arrives while the initial links are present, but how entering one terminal link changes expected delay and opportunity for primary reinforcement. Fantino (1969) tested this configuration with pigeons and found that initial-link choice proportions did not always match terminal-link reinforcement proportions.
In Fantino (1969), the initial links were varied while a terminal-link reinforcement-rate asymmetry was maintained. Matching occurred under one set of initial-link conditions and failed under other conditions. The observed pattern motivated delay-reduction theory, which relates preference to the relative reduction in expected time to primary reinforcement signaled by terminal-link entry. The theory thus assigns explanatory importance to conditioned stimuli and timing. Its support is limited by schedule structures and by how well its timing assumptions describe the participant's actual environmental information.
Further work by Fantino & Davison (1983) tested a richer set of concurrent-chains arrangements. Six pigeons completed 56 conditions with changes in the initial- and terminal-link schedules. The investigators found that a delay-reduction formulation accounted relatively well for some conditions but fit others poorly, especially when initial links were unequal. This result is theoretically informative because it limits a simple universal application of the model. One fitted relation may generalize across selected schedule manipulations and fail under others without invalidating the broader importance of terminal-link delay.
Choice can also be measured at the level of immediate transitions among terminal stimuli. Fantino & Royalty (1987) examined molecular processes under concurrent chains, including comparisons of independent and interdependent initial-link schedules. The evidence supports studying how the timing and structure of links control responding rather than interpreting every choice solely through the frequency of unconditioned rewards. Models of conditioned reinforcement must specify the predictions carried by stimuli, how those predictions are learned, and the effect of competing opportunities over time.
Delayed Outcomes and Temporal Context
Delay changes what reinforcement means for choice. An option that supplies more outcomes per unit time may require waiting through a longer terminal interval, while a less productive option may deliver a sooner immediate event. Under concurrent chains, the probability of obtaining reinforcement and the time until it arrives interact in experimentally measurable ways. The results of Fantino (1969) show that the delay reduction signaled by a transition can influence initial-link choice, indicating a need for models that represent temporal context in addition to counted reward events.
One analytical challenge is to specify the time origin against which delay is measured. Time from a choice response to entry into a terminal link differs from time between successive primary reinforcers. A temporal model that uses one denominator cannot be compared fairly with data recorded using another without transforming its prediction. Fantino & Davison (1983) shows how initial-link and terminal-link structure may jointly alter allocation. Its tests of delay reduction highlight the possibility of fit deterioration under particular schedule combinations.
Timing models may operate with noisy, nonlinear internal representations. Jozefowiez et al. (2009) proposed a behavioral economic model in which temporal representation and payoff information together guide choice, while also modeling broader interval-timing phenomena. This provides a candidate explanation connecting choice and psychophysical timing. The proposal's mathematical assumptions remain model claims to be tested experimentally. It is particularly valuable when a seemingly inconsistent preference pattern can be resolved by specifying what temporal information is available at the decision point.
Temporal research should distinguish delay discounting in decisions about explicit future outcomes from the momentary adjustment of free-operant choice under ongoing concurrent contingencies. Both involve temporal consequences, but their response structures and informational demands differ. A subject choosing between displayed amounts is not automatically operating in the same experimental system as a pigeon responding on concurrent variable-interval keys. Cross-paradigm comparison demands shared operational definitions and evidence of comparable mechanisms Davison & McCarthy (1988).
Human Choice and Cross-Species Generalization
Evidence from nonhuman species established the matching relation under tightly specified reinforcement schedules, but claims about human behavior require experiments with human participants. Humans can interpret instructions, verbalize rules, track contingencies, and vary their response strategies across sessions. Such capacities may influence both observed choice and the meaning of a fitted matching parameter. A comparison across species should therefore identify functional commonalities while retaining differences in reward currency, prior experience, verbal context, and experimental apparatus Dallery et al. (2005).
In Dallery et al. (2005), six human subjects pressed a lever for monetary reinforcers under five concurrent variable-interval schedules. Varying the amount of reinforcement on one alternative created a direct test of absolute-rate predictions from contrasting matching theories. The study observed that the formal equations had systematic residual patterns despite high overall fit, while the modern formulation provided a better account of the tested data. This supports the relevance of quantitative operant models to human behavior under controlled laboratory conditions; generalization to complex everyday decisions requires additional evidence.
The use of human samples changes the methods available to investigators. Instructions can be experimentally standardized or varied, subjective expectations can be queried, and comprehension can be assessed. But subjective reports should be considered auxiliary measurements rather than direct substitutes for schedule contact and observed reinforcement history. Within-subject reversals, multiple reward magnitudes, and comparisons of instructed versus experienced probabilities can help distinguish contingency learning from rule following. The best evidence connects the reported rule to changes in measured behavior under controlled feedback McDowell (2021).
Cross-species generalization is strongest when the independent variables and functional dependencies align despite different surface responses. It is weaker when a convenient matching equation is fitted retrospectively to unrelated choices. Similar proportional distributions can appear because of institutional rules, social constraints, response habits, or even arithmetic properties of compositional data. The task is to determine whether manipulating relative reinforcement causally changes allocation in the predicted direction and whether alternative explanatory variables have been controlled Davison & McCarthy (1988).
Applied Behavior Change, Alternatives, and Resurgence
Applied behavior analysis frequently aims to change the allocation between a behavior causing harm and an alternative response that produces desirable consequences. Differential reinforcement procedures can make a safer or more socially useful action more likely by changing the relative accessibility of reinforcement. Matching provides a conceptual and quantitative framework for examining those competing consequences, although successful intervention also depends on learning history, generalization across settings, and the feasible alternatives available to the individual McDowell (2021).
One particular challenge is resurgence: a previously reduced behavior can return when reinforcement for an alternative response is reduced or discontinued. Greer & Shahan (2019) reviewed a quantitative 'Resurgence as Choice' approach that extends choice-based accounts to explain these returns under worsened alternative reinforcement. The contribution is a model and synthesis of relevant evidence, not a license to assume identical mechanisms in every clinical or educational case. Historical reinforcement and the sequence of interventions matter; an observed change in problem behavior requires direct functional assessment.
The intervention implications are structural. Maintaining reinforcement for a constructive alternative may reduce the relative advantage of an older behavior, while abrupt withdrawal may change the distribution of available consequences. Treatment programs therefore need to consider how schedules will operate after training ends. An intervention that works under intensive laboratory or clinic reinforcement may face different competition in ordinary environments. The review by Greer & Shahan (2019) frames durability as a choice problem and identifies research needs concerning relapse and maintenance.
Ethically appropriate translation is grounded in measured behavior, informed participation where applicable, and attention to the individual's broader well-being. Matching-law reasoning cannot replace an assessment of the social meaning of behavior or justify the coercive management of choices. In this monograph applied implications remain research-level examples, and no individual clinical recommendations are inferred from them. The scientific contribution is the ability to test environmental contingencies systematically and compare predicted allocation with observed outcomes McDowell (2021).
Comparative Choice and Ecological Validity
Experimental analysis has used pigeons, rats, and human participants in arrangements that vary in motor requirements, reward type, and access to competing options. This diversity creates opportunities to test functional regularities and difficulties of equivalence. Pigeon key pecks, rat lever presses, and human monetary responses are experimentally defined actions, but they occur under different motivational and sensory conditions. The matching relation becomes an explanatory candidate only after the comparison identifies which features of the reinforcement contingency are preserved Findley (1958).
Choice in daily environments is usually distributed across many options, whose availability changes over time and whose outcomes are partly delayed or socially mediated. This complicates the definition of an 'obtained reinforcement rate'. Two experimentally visible alternatives may constitute a small fraction of an individual's feasible activities. In the quantitative account of Herrnstein (1970), other sources of reinforcement already mattered for absolute response strength, providing an early reason to avoid treating a two-key laboratory system as a literal inventory of all meaningful choices.
Environmental validity can be improved through repeated naturalistic measurements of observable actions and known consequences, but the increased realism may reduce experimental control. When an investigator measures app use, work shifts, or social contact, the observation of a response often fails to identify the outcome that maintains it. Researchers should resist assigning a generic reinforcement label to every subsequent event. Prospective interventions, randomized changes in contingencies, or carefully documented natural experiments provide stronger grounds for causal claims than correlational histories alone McDowell (2021).
Comparative generality is therefore an empirical question with several layers. The algebraic fit of response ratios is the weakest layer; predictable changes following controlled manipulation are stronger; stability of estimated parameters across task structure and participants strengthens generality further. Finally, models that predict transitions, switching, and absolute response rates in new settings offer a more ambitious standard. The distinctive contribution of comparative research is to identify the environmental conditions under which matching is robust and the conditions where other formulations are needed Davison & McCarthy (1988).
Experimental Design and Operational Definition
Experimental tests of matching begin with a precise description of which alternatives are concurrently available, what response constitutes selection, which schedule determines each consequence, and what is recorded as obtained reinforcement. If alternatives are not simultaneously available, the experimental situation may instead require a multiple-schedule or sequential-decision analysis. If the task permits long periods of inactivity, the investigator must specify how this time enters allocation denominators. The standard concurrent variable-interval preparation remains useful because it affords controlled manipulations of relative reinforcement and repeated observations within participants Herrnstein (1961).
Strong within-subject designs vary the reinforcement ratio over several conditions, include order controls and reversals, and document the criterion for stable responding. Conditions should be long enough to estimate aggregate rates, but investigators also need to preserve the transition data rather than discard the early learning period. Programmed and obtained reinforcers should be separately tabulated, as should session length, responses, time on each option, and switching counts. The methods used by Alsop & Elliffe (1988) illustrate the importance of manipulating overall and relative reward rates separately.
Construct validity depends on relating observed data to the term being claimed. A ratio of two button presses is a measure of response allocation; it becomes evidence about reinforcement sensitivity only after the contingencies are shown and systematically varied. Likewise, 'bias' is a model parameter describing asymmetry under a reference condition; it is not inherently a stable personality trait or a subjective preference. Effort and timing manipulations can help distinguish functional sources of bias, while generalized matching provides the descriptive baseline against which alternative hypotheses are compared Baum (1974).
Experimental ethics and transparency also matter. Human participants require appropriate consent and fair, comprehensible reward procedures; animals require species-specific welfare safeguards and approved protocols. The monograph reports published empirical work and makes no new claims about ethical approval. Replication packages should contain the schedule algorithm, pseudorandom sequence generation where relevant, changeover-delay settings, event data structure, and analysis code. These details are often essential to reproduce a reinforcement feedback function and its measured behavior Davison & McCarthy (1988).
Statistical Estimation and Quantitative Model Comparison
The generalized matching equation is often estimated by regressing log response ratio on log obtained reinforcement ratio across conditions. Its slope estimates sensitivity a, and its intercept estimates log bias b. This representation is compact and interpretable, but the variance of estimated log ratios depends on event counts, serial dependence, and possible near-zero denominators. Standard ordinary least squares assumptions may be strained when observations differ substantially in precision. An analysis should disclose the treatment of zero cells, model weights, repeated observations, and individual effects Baum (1974).
The data structure suggests alternatives. Binomial-like models can treat choices as counts conditional on the total number of defined responses, although their independence assumptions may fail for bouted responding. Hierarchical models can partially pool person-level sensitivity and bias without hiding heterogeneity. Time-series, state-space, or point-process models may be more appropriate for timestamps and transition hazards. The choice of model should follow the sampling mechanism and scientific question, not the desire to reproduce a historically familiar equation at any cost MacDonall (2003).
Evidence for a theory improves when rival models are compared on predictions they do not share. In Dallery et al. (2005), models explaining much of the variance were further differentiated through residual structure. This illustrates why high explained variance can coexist with systematic predictive error. Researchers should predefine plausible candidates, assess parameter identifiability, inspect residuals across reinforcement levels, and evaluate predictive performance on conditions withheld from fitting. A more flexible model may fit better while supplying a weaker explanation unless its parameters and generalization are independently constrained.
The inferential limitations are especially pronounced when obtained reinforcers are themselves influenced by allocation. A regression linking obtained reinforcement ratio with response ratio may partially capture an endogenous system. Randomized schedule parameters, feedback-function simulation, and explicit generative models help investigate this problem. Sensitivity analysis should consider programmed versus obtained rewards and evaluate whether parameter estimates are stable across choice and reinforcement aggregation windows. A fitted line is a valuable descriptive summary, but causality requires a defensible intervention on the schedule McDowell (2005).
Strength of Evidence, Replication, and Contradictory Results
The core evidence for matching comes from decades of controlled concurrent-schedule experiments, beginning with Herrnstein (1961) and extended by work on bias, timing, reward amount, and switching. Its strength lies in clear independent variables, observable behavioral measures, and repeatable manipulation of alternatives. Its limits include narrow laboratory environments, limited participant numbers in many individual-subject experiments, parameter variation across schedule formats, and the difficulty of identifying process from steady-state proportions. Recognition of these limits improves the precision of the theory rather than erasing its empirical achievements.
Three historically important findings illustrate the field's capacity for self-correction. Baum (1974) showed that bias and undermatching needed formal representation. Alsop & Elliffe (1988) found changes in reinforcement sensitivity with overall reinforcement rate, challenging an invariant ratio-only account. Fantino & Davison (1983) found that a delay-reduction model fit some concurrent-chains configurations better than others. These findings concern different variables and tasks, but all reject the expectation that one unconstrained equation will explain every allocation pattern without reference to procedural details.
More direct theory comparison strengthened the evidence base. Vaughan (1981) arranged conditions in which local improvement and global optimization could diverge, and Dallery et al. (2005) assessed equations against human response data including residual patterns. MacDonall (2003) introduced fine-grained stay–switch outcomes into the model comparison. These studies provide explanatory leverage beyond retrospective curve fitting. Still, their exact sample sizes, species, reinforcers, and schedule parameters restrict the conclusions. Independent tests should reproduce the causal manipulation as well as the statistical finding.
A responsible synthesis therefore separates well-established regularities from open mechanistic questions. Strong evidence supports an often orderly relationship between relative obtained reinforcement and allocation under specified concurrent schedules, with measurable departures described by generalized matching. More tentative claims concern the uniqueness of any local learning algorithm, direct transfer of parameter values between species, and the explanation of unrestricted human decisions. The critical overview by McDowell (2021) demonstrates continuing efforts to connect basic choice science with applied situations while keeping the status of alternative theories visible.
Computational Models and Artificial Choice Systems
Computational modeling can make the assumptions of matching research explicit. An artificial agent may be supplied with two concurrently available actions, stochastic reward schedules, switching costs, and a learning rule for estimating local returns. The investigator can then compare the resulting response ratios and switching trajectories with empirical data. Such simulations can separate outcomes produced by schedule mechanics from outcomes requiring a specific learning mechanism. A matching-looking equilibrium is therefore a phenomenon to be explained by a computational model, not evidence that the model itself has human-like motives Vaughan (1981).
An informative simulation suite would include independent variable-interval schedules, schedules that depend on the agent's own allocation, abrupt reversals, manipulations of overall reward rate, delayed terminal links, and added response costs. Parameters fitted under one condition should predict choices under held-out environments. The tests used in Alsop & Elliffe (1988), Kyonka (2008), and Fantino & Davison (1983) supply historically grounded examples of diagnostic manipulations. A model that matches choices under a stationary schedule but fails at abrupt reversals would illuminate the difference between equilibrium description and adaptive control.
Artificial reinforcement-learning systems often express learning through estimated action values and policy-update rules. Functional comparison with biological subjects is justified when the inputs, available actions, outcome information, and measurements are aligned. It does not follow from a shared proportion of choices that the agent uses the same cognitive or neural process as a pigeon or human participant. The Artificial Era motivates cross-system research on allocation, adaptive decision policies, and resource constraints, with operational equivalence and predictive validation as the essential scientific requirements Jozefowiez et al. (2009).
The most productive contribution of artificial models is counterfactual experimentation. Investigators can impose known learning rules, introduce controlled memory limits, perturb the agent's temporal discounting, or change the feedback function while monitoring every action and internal state. By demonstrating which algorithms generate matching and which generate characteristic deviations, simulation can guide new biological experiments. This creates a falsifiable bridge between mathematical descriptions and mechanistic accounts without converting a shared behavioral statistic into an unwarranted assertion of psychological identity McDowell (2005).
Open Research Problems and an Integrated Agenda
The first open problem is mechanistic identification. Several different response-selection processes can generate similar molar matching curves. Future research should therefore build diagnostic tests around switching, visit duration, dependence on recent reward, and adaptation after unanticipated changes. A model's success should be judged against experiments specifically designed to separate its predictions from those of rivals. The history of melioration tests Vaughan (1981) and local stay–switch analyses MacDonall (2003) offers methodological examples of this approach.
The second problem is scale and context. Choice allocation is defined over particular alternatives and observational intervals, yet organisms move across changing environments containing many unmeasured activities. The problem is to discover whether ratios and bias parameters are stable under changes in option number, reward quality, overall density, and temporal scale. Findings from Alsop & Elliffe (1988) and Fantino (1969) show that context can alter apparently simple choice functions. A strong generalization program would preregister which parameters should remain invariant and which should change.
The third problem is the relation between quantitative prediction and meaningful application. Differential reinforcement and response costs can reshape choice, but a stable laboratory allocation may not survive environmental disruption or fading of reinforcement. The resurgence framework discussed by Greer & Shahan (2019) illustrates why applied durability requires explicit modeling of histories and competing actions. Future work should combine repeated observations with ethical, practically relevant intervention measures, while preserving the distinction between model predictions and individually validated procedures.
The fourth problem concerns theoretical integration. Reinforcement sensitivity, timing, locally estimated payoff, and global optimization are often developed in separate formalisms. Computational comparisons can clarify which distinctions are essential and which arise from different measurement conventions. McDowell (2021) argues for connecting empirical matching with dynamic accounts of behavior; Jozefowiez et al. (2009) models links between timing and payoff. An integrated research program should ask for cross-task predictions, identifiable parameters, mechanistic perturbations, and open replication materials.
Conclusion
Choice allocation and the matching law constitute a fundamental experimental research program on the quantitative organization of behavior among concurrent alternatives. The strict matching equation relates relative behavior to relative obtained reinforcement, while generalized matching measures sensitivity and bias when observed allocation departs from strict proportionality. The historical progression from the experimental work of Findley (1958) and Herrnstein (1961) to the formal analysis of Baum (1974) established a vocabulary for asking precise questions about preference and environmental contingencies.
The subsequent research record demonstrates that reinforcement density, amount, timing, response effort, changeover contingencies, and learning history influence the observed distribution. Studies of concurrent chains Fantino & Davison (1983), local response sequences MacDonall (2003), transitions Kyonka (2008), and human model comparisons Dallery et al. (2005) show why a single fitted steady-state ratio cannot serve as a complete process account. Theoretical alternatives—including melioration, maximization, dynamic matching, and timing-sensitive models—earn explanatory authority through distinctive predictions tested against appropriate evidence.
The synthesis yields a practical scientific standard: define alternatives and consequences precisely, retain event-level measurements, distinguish programmed from obtained reinforcement, estimate allocation functions with uncertainty, compare mechanistic models on divergent predictions, and document generalization limits. Quantitative regularities remain powerful when presented at the level the data actually support. Behavioral choice can be investigated systematically without requiring that every response express a conscious optimization or that matching solve every problem of human decision making Davison & McCarthy (1988).
The enduring significance of the matching-law tradition lies in its integration of experimental control with formal description. As psychological research extends into complex social settings, adaptive computing, and new forms of human–artificial interaction, the relevant question remains empirically answerable: under which specified consequences, constraints, and histories do choosing systems distribute their actions in predictable ways? The most productive future lies in connecting these patterns to measured learning processes and testing the resulting predictions across environments McDowell (2021).
References
- Alsop, B., & Elliffe, D. (1988). Concurrent-schedule performance: Effects of relative and overall reinforcer rate. Journal of the Experimental Analysis of Behavior, 49(1), 21–36. https://doi.org/10.1901/jeab.1988.49-21
- Baum, W. M. (1974). On two types of deviation from the matching law: Bias and undermatching. Journal of the Experimental Analysis of Behavior, 22(1), 231–242. https://doi.org/10.1901/jeab.1974.22-231
- Dallery, J., Soto, P. L., & McDowell, J. J. (2005). A test of the formal and modern theories of matching. Journal of the Experimental Analysis of Behavior, 84(2), 129–145. https://doi.org/10.1901/jeab.2005.108-04
- Davison, M., & McCarthy, D. (1988). The matching law: A research review. Lawrence Erlbaum Associates. https://www.routledge.com/The-Matching-Law-A-Research-Review/Davison-McCarthy/p/book/9781138194311
- Fantino, E. (1968). Effects of required rates of responding upon choice. Journal of the Experimental Analysis of Behavior, 11(1), 15–22. https://doi.org/10.1901/jeab.1968.11-15
- Fantino, E. (1969). Choice and rate of reinforcement. Journal of the Experimental Analysis of Behavior, 12(5), 723–730. https://doi.org/10.1901/jeab.1969.12-723
- Fantino, E., & Davison, M. (1983). Choice: Some quantitative relations. Journal of the Experimental Analysis of Behavior, 40(1), 1–13. https://doi.org/10.1901/jeab.1983.40-1
- Fantino, E., & Royalty, P. (1987). A molecular analysis of choice on concurrent-chains schedules. Journal of the Experimental Analysis of Behavior, 48(1), 145–159. https://doi.org/10.1901/jeab.1987.48-145
- Findley, J. D. (1958). Preference and switching under concurrent scheduling. Journal of the Experimental Analysis of Behavior, 1(2), 123–144. https://doi.org/10.1901/jeab.1958.1-123
- Greer, B. D., & Shahan, T. A. (2019). Resurgence as choice: Implications for promoting durable behavior change. Journal of Applied Behavior Analysis, 52(3), 816–846. https://doi.org/10.1002/jaba.573
- Herrnstein, R. J. (1961). Relative and absolute strength of response as a function of frequency of reinforcement. Journal of the Experimental Analysis of Behavior, 4(3), 267–272. https://doi.org/10.1901/jeab.1961.4-267
- Herrnstein, R. J. (1970). On the law of effect. Journal of the Experimental Analysis of Behavior, 13(2), 243–266. https://doi.org/10.1901/jeab.1970.13-243
- Herrnstein, R. J. (1974). Formal properties of the matching law. Journal of the Experimental Analysis of Behavior, 21(1), 159–164. https://doi.org/10.1901/jeab.1974.21-159
- Jozefowiez, J., Staddon, J. E. R., & Cerutti, D. T. (2009). The behavioral economics of choice and interval timing. Psychological Review, 116(3), 519–539. https://doi.org/10.1037/a0016171
- Kyonka, E. G. E. (2008). The matching law and effects of reinforcer rate and magnitude on choice in transition. Behavioural Processes, 78(2), 210–216. https://doi.org/10.1016/j.beproc.2007.12.003
- MacDonall, J. S. (2003). Reinforcing staying and switching while using a changeover delay. Journal of the Experimental Analysis of Behavior, 79(2), 219–232. https://doi.org/10.1901/jeab.2003.79-219
- McDowell, J. J. (2005). On the classic and modern theories of matching. Journal of the Experimental Analysis of Behavior, 84(1), 111–127. https://doi.org/10.1901/jeab.2005.59-04
- McDowell, J. J. (2021). Empirical matching, matching theory, and an evolutionary theory of behavior dynamics in clinical application. Perspectives on Behavior Science, 44(4), 561–580. https://doi.org/10.1007/s40614-021-00296-w
- Rachlin, H., Battalio, R., Kagel, J., & Green, L. (1981). Maximization theory in behavioral psychology. Behavioral and Brain Sciences, 4(3), 371–388. https://doi.org/10.1017/S0140525X00009407
- Shimp, C. P. (1966). Probabilistically reinforced choice behavior in pigeons. Journal of the Experimental Analysis of Behavior, 9(4), 443–455. https://doi.org/10.1901/jeab.1966.9-443
- Vaughan, W., Jr. (1981). Melioration, matching, and maximization. Journal of the Experimental Analysis of Behavior, 36(2), 141–149. https://doi.org/10.1901/jeab.1981.36-141