RESEARCH DOSSIER
Laboratory and Field Tests of Microeconomic Choice
Abstract. Laboratory and field experiments transform microeconomic choice into an observable relation between feasible alternatives, information, consequences, institutional rules, and decision makers. Their scientific value rests on the correspondence between the behavior elicited, the theoretical restriction tested, the causal estimand identified, and the population or institution to which the conclusion is applied. This Research Dossier examines the foundations of induced values, revealed preference, risk and reference-dependent choice, social and strategic behavior, experimental demand, field identification, treatment heterogeneity, external validity, replication, and welfare interpretation. It integrates original research in controlled markets, individual decision tasks, cross-society games, labor-market correspondence experiments, microcredit interventions, energy consumption, and field valuation. Transparent analytical examples distinguish violations of stable utility maximization from estimated changes in choice, and distinguish program assignment effects from preference parameters and welfare effects. The central synthesis is that laboratory control and field institutional relevance constitute complementary dimensions of evidential quality. A persuasive test retains the mapping from observed decisions to maintained assumptions, from assignment to estimand, and from the original sample to the intended target. Research designs that explicitly examine these mappings can identify why results differ across populations and environments. The dossier is a critical narrative synthesis with illustrative derivations, rather than a new empirical dataset or a quantitative meta-analysis.
Keywords: Microeconomic choice; experimental economics; revealed preference; induced value theory; incentive compatibility; risk preferences; social preferences; field experiments; randomized evaluation; treatment heterogeneity; external validity; replication; welfare economics
The object of experimental microeconomics
Microeconomics studies structured decisions made under scarcity and the consequences of interaction among decision makers. An experimental investigation of choice constructs or observes a sufficiently specified decision environment to determine which implications of a microeconomic account are consistent with behavior. A household choosing between state-contingent payments, a worker allocating effort, a trader responding to a price, and an employer selecting resumes each makes a decision within a distinct institutional architecture. A shared language of choice makes these phenomena analytically comparable while maintaining their different consequences.
The basic empirical object is a conditional distribution of actions. Let X describe the feasible set, I the available information, C the consequences and institutional conditions, and s a description of the decision maker or relevant history. An observation of action a is a realization of behavior conditional on (X, I, C, s). A theory can restrict this relation through stable preferences, utility maximization, subjective probabilities, strategic expectations, limited attention, reference dependence, or learning. The experimental researcher modifies selected elements of the environment and measures the resulting actions. The identity of an action alone does not determine the mechanism that produced it.
Three test objects deserve separate treatment. A consistency test asks whether a collection of decisions satisfies restrictions implied by a model, such as revealed-preference inequalities. A parameter-estimation exercise identifies magnitudes within a model given maintained assumptions about utility, beliefs, and errors. A causal design estimates the change in an outcome induced by a well-defined assignment, offer, information intervention, price, or rule. These objects intersect, but each possesses a distinct denominator of inference. A significant treatment contrast need not falsify utility maximization; a utility-rationalizable dataset need not identify a particular functional form; a precise estimate of a risk parameter in one task need not predict borrowing in a different institutional environment.
Laboratory economics acquired a particularly clear empirical identity through the controlled investigation of markets. Smith's (1962) experimental study of competitive markets examined trading under specified supply and demand conditions. The scientific achievement was the ability to distinguish the institution through which buyers and sellers interact from the theoretical allocation used to assess its performance. Competitive equilibrium thereby became a benchmark against which a sequence of observed trades could be evaluated. Researchers could observe prices, trading quantities, and deviations from predicted outcomes while retaining explicit knowledge of the constructed trading environment.
The same methodological distinction applies to individual choice. A price-taking demand model describes choices over affordable bundles; a laboratory task can construct known budgets, a field study can randomize prices or access, and transaction data can record purchases under natural constraints. The laboratory typically provides more direct access to the alternatives presented. The field typically supplies consequences and context generated by existing institutions. The observational record may cover a much broader range of situations but leave counterfactual variation unidentified. Scientific strength depends on the match between the design and the research question.
This dossier therefore treats economic rationality as a family of empirically assessable restrictions rather than a single label applied to persons. It treats experiments as instruments for studying observable implications of those restrictions. Its historical examples establish why control, incentive design, variation, and interpretation must be considered jointly. The selected studies are influential primary investigations and methodological contributions chosen for the distinctions they illuminate. This is a purposive narrative review, not an exhaustive systematic search or independently reproduced meta-analysis. Reported empirical findings retain the scope of the associated study.
Choice, budgets, and revealed preference
The classical consumer model begins with a preference relation over bundles and maximization over a budget set. For nonnegative quantities x and positive prices p, a linear budget consists of bundles satisfying p·x ≤ m. Stable, complete, transitive preferences represented by a suitable utility function generate restrictions on observed choices across budgets. The strength of revealed-preference analysis is its ability to test some of these restrictions without committing to a parametric demand equation.
Suppose the observed choice x(t) is made at prices p(t). If another observed bundle x(s) was affordable at the time, x(t) is directly revealed at least as preferred as x(s). The transitive closure of this relation yields indirect revealed preference. Under local nonsatiation and the usual finite linear-budget setting, the generalized axiom of revealed preference prohibits an indirectly revealed preference for one bundle when the reverse comparison is strictly revealed by a cheaper alternative. Varian's (1982) nonparametric demand analysis develops this approach to testing demand data, recovering utility restrictions, and assessing extrapolation.
Consider two hypothetical decisions. At prices (2, 1), the individual chooses (2, 0), spending 4, although (0, 2) costs 2. At prices (1, 2), the individual chooses (0, 2), again spending 4, although (2, 0) costs 2. The cheaper alternative in each observation is the bundle selected in the other. Local nonsatiation then implies incompatible strict rankings. In this example the pattern violates GARP; there is no single locally nonsatiated preference ordering rationalizing those choices on the stated budgets.
Reverse the price vectors while retaining the choices. At prices (1, 2), the first bundle costs 2 and the second costs 4; at prices (2, 1), the reverse applies. Neither alternative is affordable on the other's observed expenditure budget. No revealed-preference cycle follows. This second dataset is consistent with a stable utility account, although it identifies neither the unique utility function nor the consumer's ranking outside the observed choices. Both examples are constructed to expose the logic; neither is participant data.
A finite dataset can also be evaluated by the Afriat inequalities. For utility indices u(t) and positive multipliers λ(t), feasibility of the inequalities
u(s) ≤ u(t) + λ(t) p(t)·(x(s) − x(t))
for all observed pairs supports construction of a concave, monotonic utility representation in the usual setting. The multiplier is not directly observed marginal utility of income. It is an auxiliary quantity in the feasibility system. The existence of such numbers establishes rationalizability conditional on the stated premises; it does not establish psychological introspection or unique preferences.
The degree of violation adds information beyond a binary test. A critical cost efficiency index asks how much expenditure budgets would have to be contracted in revealed-preference comparisons before the data satisfy the relevant consistency restriction. Values nearer one indicate smaller expenditure violations in that particular metric. Index values are influenced by the number and arrangement of budgets, measurement errors, and choice noise. They should be compared under comparable test environments, with attention to the power of each environment to generate contradictions.
Choi, Fisman, Gale, and Kariv (2007) use state-contingent graphical budgets to study consistency and heterogeneous behavior under uncertainty. Such designs make individual tests feasible because each participant faces a collection of systematically varied budget lines. Choices may be treated as allocations of payoff among states, allowing separate attention to consistency and to the shape of preferences. A sufficiently varied budget design is essential. Repeated observations at nearly identical prices provide limited identifying variation; diverse prices and state-contingent opportunities are more informative.
A related application connects experimental consistency to economic outcomes outside the task. Choi, Kariv, Müller, and Silverman (2014) report in Who Is (More) Rational? an association between a one-standard-deviation increase in measured consistency and 15–19 percent greater household wealth. The finding is an association conditional on their design and analysis. Wealth, education, financial experience, cognitive resources, and task familiarity may influence both performance and household circumstances. The study does not establish that an intervention increasing the consistency index by one standard deviation would itself cause wealth to increase by the reported percentage.
An exact failure of GARP can arise from a stable-preference model being empirically inadequate, a misunderstood price, changing beliefs or reference points, a mistaken click, or an incorrectly measured feasible set. The researcher must keep these explanations available until the design distinguishes them. Repeated measurements, comprehension tasks, error models, and variation in interface or stakes help determine which explanation has empirical support. Treating every inconsistency as the same psychological mechanism sacrifices the information provided by the experiment.
Induced values and incentive compatibility
An induced value is an experimental valuation created through specified consequences. Smith's (1976) induced value theory provides the classic methodological account of using a reward medium and a known payoff structure to construct economically meaningful experimental preferences. If participants value the reward monotonically and its consequences dominate relevant uncontrolled motives, a redemption schedule can implement a ranking of experimental trades. This permits researchers to evaluate an institution under an explicitly designed environment.
Dominance of the reward is substantive. Monetary amounts may be small relative to the value of time, the desire to satisfy the researcher, or the significance of fairness and reputation. Conversely, very large rewards may alter the object of choice through income effects, risk aversion, limited liquidity, or strategic attention. A payment protocol therefore requires an argument about why its consequences support the intended inference. The fact that a task uses money is only the starting point.
An incentive-compatible elicitation mechanism makes truthful reporting optimal under specified assumptions. The single-response method of Becker, DeGroot, and Marschak (1964), introduced in Measuring Utility by a Single-Response Sequential Method, supplies an important foundation for valuation mechanisms using an independently determined transaction price. Its elementary economic logic can be displayed in a simple purchase setting.
A participant has a private monetary valuation v for a good and reports a bid b. A price r is drawn independently of the bid. If b ≥ r, the participant buys the good at r; otherwise no purchase is made. Under quasilinear utility and enforceable transactions, purchase is desirable precisely when v ≥ r. Reporting b = v implements the preferred action for every r. A lower bid risks losing desirable purchases, while a higher bid risks acquiring the good at an undesirable price.
For example, let v = 10. A bid of 8 forgoes positive surplus whenever the price falls strictly between 8 and 10. A bid of 12 exposes the participant to negative surplus when the price lies strictly between 10 and 12. Neither deviation improves the outcome at any price. If the price distribution assigns positive probability to the relevant intervals, the deviations are strictly worse in expected utility. With a degenerate price distribution, multiple bids may implement the same choices; truthful reporting remains weakly optimal, but unique identification is lost.
This derivation is an illustration under stated premises. A participant who cannot borrow to pay an otherwise attractive price has a constrained feasible set; the report may identify effective willingness to pay under liquidity constraints rather than unconstrained valuation. A participant who expects the price distribution to depend on the report no longer faces the intended mechanism. If the transaction concerns a public or socially visible good, additional motives may enter the valuation. If the task is hypothetical, the consequences that produce truth-telling in the derivation are absent.
Mechanism comprehension is therefore an empirical variable. A researcher can demonstrate the pricing rule with practice rounds, use a simple induced-value benchmark, record choices at which comprehension has failed, and compare alternative interfaces. In rural Uganda, Burchardi, de Quidt, Gulesci, Lerva, and Tripodi (2021) directly study field willingness-to-pay elicitation, illustrating the importance of testing incentive mechanisms in the population and task where valuations are measured. A mechanism that is formally well designed can still be difficult to understand; a mechanism that participants readily understand can still lack a truthful equilibrium. The two questions require different forms of validation.
Payment procedures for repeated choice tasks present an additional issue. Paying every decision can create wealth accumulation, portfolio interactions, or learning across rounds. Paying one randomly selected decision can reduce some cross-round interactions, but its interpretation depends on the assumed treatment of compound lotteries and the independence of preferences across tasks. If participants integrate decisions across rounds or exhibit nonstandard probability weighting, random-task payment is not automatically neutral. The research record should specify the draw mechanism, probability of payment, payment timing, actual stakes, and whether participants were informed of these matters before responding.
In field settings, an offer may change several incentives at once. A discount can alter the monetary price and convey information about quality; an application fee can select participants and signal institutional seriousness. A payment can induce participation while changing the composition of participants. The experiment must distinguish the effect of assignment from the structural parameter that researchers wish to recover. The meaning of incentive compatibility is thus inseparable from the architecture of information and enforcement.
Risk, ambiguity, and the architecture of elicitation
Expected utility provides a benchmark for choices among known lotteries. For consequences x(j) with probabilities q(j), expected utility is the probability-weighted sum of utility over consequences. Risk aversion under this account is represented by concavity of utility over the relevant outcome or final wealth. A risk-neutral decision maker compares expected monetary values when utility is linear in money. These are restrictions on preferences within an explicit model of probabilities and outcomes.
A conventional power utility specification takes utility of positive wealth w to be w raised to the power 1 − r, divided by 1 − r, for r ≠ 1; logarithmic utility is the limiting case r = 1. Constant relative risk aversion is a feature of this functional form. Whether an estimated coefficient is meaningful depends on whether the task is evaluated in terms of final wealth, incremental experimental income, consumption, or another argument. The same observed switching decision can imply different coefficients when those background assumptions differ.
A hypothetical example clarifies the role of scale and baseline. A lottery pays 0 or 100 with equal probability. With utility equal to the square root of the total payoff and a zero baseline, expected utility is 5. The certainty equivalent is 25, compared with expected payoff of 50, yielding a risk premium of 25. With initial wealth of 100 evaluated jointly with the lottery, expected utility is (10 + √200)/2. The corresponding certainty equivalent in final wealth is approximately 145.71. The certainty equivalent of the additional payoff is approximately 45.71, and the risk premium is approximately 4.29. The function is unchanged; the baseline changes the curvature relevant to the payoff increments. These are stylized calculations, not measured parameters.
Holt and Laury's (2002) risk-aversion and incentive experiment employed structured lottery comparisons to examine choices across payoff conditions. Their study found greater measured risk aversion under higher real stakes than under the lower real-payoff conditions, with different patterns for hypothetical scaling. The result made payment magnitude an empirical issue rather than a procedural detail. In their (2005) new data without order effects, the authors directly addressed order effects in the earlier experimental design. Reading the two articles together prevents one experiment's sequence from being mistaken for an invariant stake-response relationship.
Multiple-price-list tasks offer a practical way to measure risk choices. A sequence of paired lotteries is ordered by the probability of the higher outcome. Under a maintained family of expected-utility preferences, a switching point can interval-identify a curvature parameter. Observed multiple switching points may reflect errors, indifference, limited comprehension, response strategy, or preferences inconsistent with the maintained specification. Deleting such responses mechanically may create a selected analytical sample. A preregistered rule for invalid or multiple-switch responses, plus sensitivity analyses retaining them under an error model, makes the estimand clearer.
State-contingent graphical allocations permit a more flexible design. They vary relative prices of payoffs in states and allow participants to choose anywhere on a budget line. The experiments of Choi et al. (2007) indicate how repeated budget variation supports the separate study of consistency and heterogeneity. There is a difference between estimating risk aversion because a parametric model fits well and establishing that a particular participant satisfies the nonparametric conditions needed by a wider class of utility representations. Reporting both quantities avoids conflation.
Ambiguity adds another dimension. Risk involves a sufficiently specified distribution of consequences. Ambiguity concerns the decision maker's uncertainty over the relevant probability distribution or its reliability. A person who avoids an uncertain payoff can be responding to curvature, pessimistic beliefs, aversion to uncertainty over probabilities, or distrust of the experiment. Identifying one mechanism requires independent variation in known probabilities, uncertain probabilities, information sources, and incentives for reporting beliefs. Even an incentivized belief report can interact with risk preferences if the scoring rule changes lottery payoffs.
The appropriate empirical strategy is to design alternative tasks whose predictions differ under candidate models. An effect replicated across stake levels and interfaces provides stronger evidence of a mechanism than a precise estimate generated by one payoff schedule. However, invariance itself is a testable property, not an assumption granted by repetition. A researcher's substantive goal may be to identify a preference parameter within a narrow institution rather than to measure a stable trait across all domains. Both goals are scientifically coherent when stated explicitly.
Reference dependence and learning
Expected-utility models evaluate consequences through preferences over the outcome space. Reference-dependent accounts allow the value of an outcome to depend on a benchmark, such as an expectation, an existing entitlement, or a previous allocation. Kahneman and Tversky's (1979) Prospect Theory studies systematic departures from expected utility, emphasizing gains and losses relative to a reference point and transformations of decision probabilities. Their original formulation is historically distinct from later cumulative formulations; the dossier uses the original work to motivate identification questions, not to attribute later specifications retroactively.
A simple test can present identical final allocations under different initial descriptions. If the displayed reference differs and all other relevant determinants are fixed, a change in choice is evidence against an account based exclusively on unchanged preferences over those final allocations. But a design must distinguish the reference manipulation from altered beliefs, perceived fairness, ownership rights, transaction costs, and information about quality. A frame can affect several channels at once.
The endowment effect is often examined through willingness to buy versus willingness to accept compensation or through exchange decisions following randomized ownership. These are different measurements. A selling price may incorporate anticipated regret, transaction costs, attachment or a strategic belief that the elicitation process is negotiable. A buyer may face liquidity constraints that the owner does not. The valuation gap alone cannot separate a loss-aversion mechanism from these alternatives.
List's (2003) market experience and anomalies study examines endowment-related behavior in actual trading environments and shows how experience relates to the strength of exchange anomalies in the investigated markets. Traders who differ in experience can also differ in selection, information, professional objectives, and transaction practices. The appropriate conclusion is conditional: market experience is an empirically consequential dimension when assessing the stability of the observed anomaly. Establishing the causal effect of learning requires variation in experience that is independent of preexisting trader characteristics or a credible longitudinal design.
Learning can take multiple forms: comprehension of contractual rules, discovery of one's own preferences, formation of beliefs about prices, adaptation to an interface, strategic learning about other participants, and selection into a repeated environment. A choice pattern that converges toward the benchmark over time does not by itself identify which form of learning occurred. A researcher may distinguish these mechanisms by varying feedback, opportunity to practice, information about others, or exposure to the institution separately.
Reference points and learning also interact with treatment timing. An announcement that a subsidy will end may change expectations before the monetary price actually changes. An initial transfer may establish a benchmark for later allocations. Repeated price offers can form a perceived normal price. If previous tasks shape present choices, randomizing task order identifies an average order-sensitive effect only over the randomized order distribution. It does not make choice history disappear. Experimental analysis should represent the temporal sequence explicitly and preserve the distinction between initial exposure and repeated participation.
A continuum of experimental institutions
Laboratory and field studies differ across several dimensions rather than on a single scale of realism. Harrison and List (2004), in Field Experiments, characterize experimental environments through the subject pool, information, commodity, task, stakes and setting. Their distinctions among conventional laboratory, artefactual field, framed field and natural field experiments make the design choice multidimensional. A study using a recruited nonstudent population in an otherwise standardized experiment addresses a different question from a study embedding random variation in a routine market transaction.
A laboratory can create known budgets that are almost impossible to observe in natural consumption data. It can randomly assign complete information about probabilities, require simultaneous decisions, and implement clean comparisons between market institutions. A natural field experiment can observe choices under ongoing contractual practices, experienced decision makers and genuine alternatives. A framed field experiment can combine a standardized task with familiar commodities or local information. These properties should be selected to answer a precise scientific question.
There are trade-offs. Strong control can remove institutional features necessary for the mechanism of interest. A market experiment without persistent relationships cannot identify the full consequences of relational contracting. Ordinary market behavior may incorporate unobserved credit limits, product quality, search costs, or informal advice; such conditions can complicate the interpretation of a randomized intervention. An experiment can have persuasive identification of an offer effect even when the researcher cannot identify the preferences underlying every purchase.
Research involving real transactions also has ethical constraints. Participants should receive appropriate information about procedures and consequences, and data collection should respect privacy and applicable ethical review. Natural field experiments may involve subjects who do not know that a routine decision is being studied. The permissibility and scientific necessity of such designs depend on oversight, risks, data handling, potential deception, and available alternatives. Scientifically informative interventions should not create unnecessary burdens or collect unrelated personal information.
An effective interdisciplinary strategy connects settings through common empirical restrictions. A laboratory study identifies whether an information manipulation changes choice with payoffs fixed. A field study tests the same manipulation where subjects encounter genuine market options. If responses differ, additional experiments can vary stakes, information sources, familiarity, and social observation independently. Comparing these designs maps the conditions under which the relationship persists. Treating all departures as failures of one setting obscures their scientific content.
Randomization and causal identification
Random assignment creates an explicit mechanism for identifying causal comparisons. Let Y(1) and Y(0) denote potential outcomes under treatment and control, and Z indicate random assignment. Under a well-defined treatment, valid randomization and an appropriate no-interference assumption, the observed outcome equals Y(Z). Independence of Z from the potential outcomes supports identification of the average treatment effect through the difference between treated and control outcome means. The realized difference remains subject to sampling and assignment uncertainty.
Duflo, Glennerster, and Kremer (2007), in Using Randomization in Development Economics Research, develop methodological considerations that connect assignment, design, statistical power, implementation, imperfect compliance and externalities. The toolkit's significance for microeconomic choice is that assignment to a price, information state, credit offer, contractual rule or institutional access can yield an identified change in observed behavior. The assignment effect is not necessarily an estimate of demand elasticity, a preference parameter, or consumption welfare.
The analysis begins by specifying the units. A person can be the subject of measurement while a village, school, store, or neighborhood is the unit of random assignment. Outcomes may be observed at another level. A village-level intervention with household-level outcomes creates within-village correlation and limited independent assignment units. Treating every observed household as independently randomized understates uncertainty.
For a simplified illustration of statistical power, consider two equally sized independent arms with continuous outcomes and common variance. An approximate two-sided 5 percent significance test with 80 percent power for a standardized effect d uses n ≈ 2(1.96 + 0.84)²/d² observations per arm. At d = 0.20, this gives approximately 392 observations per arm. The calculation uses normal approximations and assumes no attrition, no covariate adjustment, and independent units. In a clustered design with 20 observations per cluster and intracluster correlation 0.05, the elementary equal-size design effect is 1 + 19 × 0.05 = 1.95. An actual design requires cluster-level power calculations, attention to the number of clusters, and assumptions consistent with the intended estimator.
Stratified or blocked assignment can improve precision and protect against important imbalances across known population divisions. The analysis should account for the assignment structure. Covariate adjustment can increase precision under appropriate conditions, but covariate selection after inspecting outcomes creates statistical discretion. The prespecified estimator should identify the principal contrast, the outcome definition, variance calculation, and planned covariates or adjustment strategy.
Balance checks are useful diagnostics for implementation and data quality. Randomization does not guarantee exact equality of all baseline characteristics in a realized sample. Conversely, a lack of statistically significant imbalances does not prove that randomization was implemented properly. Assignment records, eligibility rules, prespecified blocking variables, attrition patterns and the integrity of the treatment delivery process provide more direct information about design validity.
The measurement itself needs a defensible interpretation. A price experiment can count purchases, quantities, timing of purchase, substitutions, or expenditure. These outcomes reflect distinct margins of response. A subsidy offered before an application deadline identifies a particular change in the choice environment. It may also shift information or beliefs about product quality. The estimate is meaningful as the effect of that policy bundle; further decomposition requires additional variation.
A model-based structural analysis may use experimental variation to estimate preferences or constraints. It must specify the choice model, scale normalization, error structure, beliefs, availability of alternatives and functional-form assumptions. A randomized reduced-form estimate can validate restrictions on a structural model or supply moments for estimation. Randomness in assignment does not validate the mathematical functional form itself. Good empirical practice reports both the identified treatment contrast and the additional assumptions that make a structural interpretation possible.
Noncompliance, attrition, and interference
Assignment and realized treatment frequently differ. Individuals offered credit may decline; people assigned to information may fail to read it; firms offered training may not send workers. The intent-to-treat effect estimates the causal consequences of assignment or the offer under the given implementation conditions. It retains the randomization and includes imperfect uptake as part of the policy mechanism.
Consider an illustrative offer that increases the probability of using a service by 0.30 and increases an outcome probability by 0.06. The ratio of the two assignment effects is 0.20. Under random assignment, exclusion of a direct assignment effect on the outcome, monotonicity of treatment receipt, a nonzero first stage, and well-defined treatment and interference conditions, this ratio is interpreted as a local average treatment effect for compliers. It is not the average effect for all eligible people. If merely receiving an offer changes expectations or behavior independently of actual uptake, the exclusion restriction fails even when the instrument strongly changes receipt.
An intention-to-treat contrast is often the more directly policy-relevant estimate because a government or organization can control an offer but not compel voluntary use. The complier effect can reveal the magnitude for the group whose choices change because of the assignment. The two answer complementary questions. Separating them also protects against comparing treated participants with untreated participants selected by their own preferences, wealth or willingness to participate.
Attrition can create selection even when the original assignment is randomized. If outcome measurement is missing more often in one arm, the observed comparison may no longer represent the full assigned sample. Researchers should report missingness by assignment, reasons for loss, and consequences of different assumptions about unobserved outcomes. When an outcome is bounded between zero and one, a simple worst-case bound can assign all missing outcomes in each arm first to zero and then to one, using the original assigned denominator. Combining these intervals yields a transparent range for the treatment contrast without assuming missing outcomes resemble observed outcomes. Tighter bounds require additional restrictions.
Interference occurs when the treatment of one unit affects another's outcome. An information intervention can spread through social networks. A credit intervention can change local prices and informal borrowing. A job-search intervention can affect competition for vacancies. The usual individual-level causal contrast may therefore depend on the allocation of treatment to other people. Defining potential outcomes by exposure to neighbors' assignments, or randomizing treatment saturation across clusters, can identify separate direct and spillover effects under an explicit exposure model.
The effect of scaling a program is correspondingly different from the effect of adding one treated participant while holding the environment fixed. Muralidharan and Niehaus (2017) analyze Experimentation at Scale and distinguish dimensions of scale relevant to evaluation. Scaling changes implementation organizations, targeting, market prices, peer interactions, or political incentives. Large numbers of treated participants do not guarantee that the treatment resembles a national policy, while small experiments may provide strong evidence for partial-equilibrium mechanisms.
These issues have a common root: the treatment must be defined at the level where consequences are generated. A contract offer, exposure to information, realized receipt, institutional saturation, and equilibrium adjustment each constitute a distinct treatment object. Causal estimates are interpretable when the assignment and exposure structure specify which object is being measured. A research report should make that mapping accessible before presenting treatment effects.
Heterogeneity, prediction, and transport
Microeconomic choice varies across constraints, beliefs, preferences, and institutional conditions. Treatment effect heterogeneity is variation in causal responses under a specified treatment, rather than simply variation in baseline outcomes. Two consumers may have very different expenditure levels while responding similarly to a price intervention. Conversely, two individuals with similar baseline purchases may respond differently because their budget constraints, expectations, or substitution opportunities differ.
A studied population is characterized by its distribution of effect modifiers. If the conditional average treatment effect for characteristics x is τ(x), the average effect in a population with distribution F is the integral of τ(x) against F. In a target population with a different distribution G, transport of the conditional relationship yields a different aggregate even if τ(x) itself remains stable. Reweighting requires adequate overlap between the populations, correct identification of relevant effect modifiers and a defensible assumption of conditional transportability.
A hypothetical two-group example illustrates the economic significance of composition. Let an intervention change purchase probability by +0.10 in group A and −0.05 in group B. In a source population that is 60 percent A and 40 percent B, the average effect is 0.60 × 0.10 + 0.40 × (−0.05) = 0.04. In a target population that is 20 percent A and 80 percent B, the corresponding average is −0.02. The sign reverses with composition. No individual's stipulated response has changed. These figures are an illustrative arithmetic construction, not empirical estimates.
When effect modifiers are unavailable or when treatment delivery changes across institutions, reweighting does not resolve transport. The conditional mechanism may itself differ. Price discounts delivered by a familiar local organization may produce different beliefs from equally sized discounts delivered by an unfamiliar organization. Regulatory constraints, contract enforcement, and outside options can change the feasible sets. In such cases, the target effect requires new evidence, a defensible structural extrapolation, or explicit uncertainty.
Allcott (2015), in Site Selection Bias in Program Evaluation, analyzes 111 randomized energy-program trials involving 8.6 million U.S. households, finding that results from the initial trials overpredicted subsequent-site effects. The case is especially valuable because randomization within sites and nonrandom selection of sites are different statistical processes. Credible local estimates can coexist with misleading predictions for the sequence of future implementations. Site recruitment, targeting, and the institutions delivering an intervention are relevant empirical variables.
Heterogeneity analysis also raises multiplicity problems. Testing interactions with dozens of baseline characteristics produces opportunities for chance findings. An analysis plan should distinguish theoretically motivated moderators from exploratory discovery. Partial pooling, cross-validation or held-out samples can improve the stability of prediction, although their interpretation is not equivalent to a preregistered confirmatory hypothesis test. A subgroup estimate should be accompanied by uncertainty for both its magnitude and its difference from other groups.
Predictive accuracy and causal interpretability are separate criteria. A baseline feature may forecast treatment responsiveness without constituting the mechanism that produces it. Intervening on that feature may have no corresponding effect. A credible prediction for targeting can be useful without a full account of why the feature is predictive. A scientifically explanatory result requires an additional design or restriction that identifies the mechanism.
The strongest transport design uses planned replication across contexts with explicit variation in plausible effect modifiers. A common measurement and intervention protocol supplies a comparison baseline; carefully chosen local adaptations protect institutional relevance. Researchers should record those adaptations rather than pool results under identical treatment labels. The resulting evidence supports conditional generalizations with known scope, a more informative achievement than an undifferentiated overall mean.
Evidence from labor, credit, consumption, and effort
Labor-market correspondence experiments
Fictitious application studies can construct comparable applications while varying a signal of interest. Bertrand and Mullainathan (2004), in their labor-market correspondence experiment, randomly assigned race-associated names to resumes sent to employers in Boston and Chicago. Their reported result is that resumes with White-associated names received about 50 percent more callbacks than those with African American-associated names. The statement describes a relative callback difference within the study, rather than a difference of 50 percentage points.
For illustration, a baseline callback rate of 8 percent and a relative increase of 50 percent imply a comparison rate of 12 percent and an absolute difference of four percentage points. Those hypothetical rates are not the study's measured rates. Their purpose is to distinguish the statistical scales on which the reported result could be communicated.
The treatment is the name assigned to the application. It can influence perceived race and potentially other inferred characteristics. The randomized comparison identifies the effect of that signal bundle on the callback outcome. It does not observe final hiring, wages, job performance, or the complete causal pathway through employer beliefs. Distinguishing taste-based from statistical discrimination requires additional evidence about how decision makers respond to information about productivity, uncertainty, screening costs and perceived group membership.
The design highlights a central advantage of field experimentation: employers respond within their actual recruitment process, producing a meaningful outcome under routine institutional conditions. Its limitation is that a naturally embedded choice captures a stage of a more complex process. An economic interpretation can connect the estimate to models of screening, but the experiment alone does not identify all parameters of those models.
Credit and household constraints
Randomized credit access examines how households respond to changes in financial opportunities. Banerjee, Duflo, Glennerster, and Kinnan (2015), in The Miracle of Microfinance?, study a microcredit expansion assigned to 52 of 104 neighborhoods in Hyderabad, India. The reported increase in microcredit take-up was 8.4 percentage points; the study found no statistically significant increase in consumption from the intervention in its reported initial comparison. Business activity and investment outcomes present additional dimensions of the response.
The intervention changes access to an institution rather than simply randomizing a preference for borrowing. Uptake reflects eligibility, expectations about repayment, perceived profitability, social obligations and outside credit options. A modest average uptake change can coexist with substantial effects on a subset of households. A null average consumption estimate is compatible with heterogeneous gains and losses or changes in the timing and composition of expenditure. It is not proof that borrowing had no economic consequences for every participant.
Later availability of credit in the comparison neighborhoods changes the interpretation of follow-up contrasts. Researchers must describe whether the later estimate concerns differential duration of exposure, early access, cumulative consequences, or continuing access versus none. Maintaining the original random assignment may still support the effect of the historical intervention path; it does not recreate the original untreated opportunity set.
Field credit evidence thus connects behavioral choice with institutional constraints. A theory explaining credit demand through liquidity and expected return should generate predictions about borrowing, investment, repayment and substitution across lenders. The evaluation of a lending program measures selected outcomes and can test those predictions when the model's other conditions are explicit. It does not by itself establish a general ranking of financial institutions.
Consumption and behavioral interventions
Energy consumption provides an example of repeated decisions embedded in routine household activity, where a randomized information or feedback intervention can generate measured changes without constructing artificial purchase tasks. The Allcott study discussed above illustrates variation in treatment effects across sites. Such effects may depend on baseline consumption, housing characteristics, electricity prices, program design and household attention. Aggregate energy use is an outcome of multiple actions and constraints, including technology and weather, so a treatment effect on consumption does not isolate one decision motive.
A useful design connects aggregate measurements to intermediate mechanisms. If an intervention provides information about comparable households, researchers may examine beliefs about typical consumption, subsequent search or investments, and changes in usage during periods when discretionary demand is high. These measurements can strengthen an interpretation, but conditioning on post-treatment behavior can introduce selection. Mechanism studies should randomize relevant components or use explicit mediation assumptions.
The practical research question concerns the incremental effect of a specified informational program in a specified population. A structural behavioral account can then explain why this effect varies. Distinguishing the two allows both policy evaluation and theory development to proceed without overstating the evidence.
Effort and incentives
DellaVigna and Pope (2018), in What Motivates Effort?, compare 18 treatment conditions in a real-effort task and examine expert forecasts. Their evidence illustrates how several incentives and behavioral motivations can be tested against a common performance outcome. Monetary piece rates, social considerations, framing and other features can be compared within a consistent task architecture; forecasts provide an additional measure of prior expectations about the relative effectiveness of these interventions.
Effort is an economic choice with opportunity costs. A piece rate can increase the return to effort; it can also convey information about task expectations or the employer's intent. A very small payment may be interpreted differently depending on whether the baseline involves voluntary activity, routine work, or an established employment relation. The absence of a particular pattern in one task does not resolve its existence in all employment institutions.
Experimental effort studies are especially informative when the production function is measured. A worker's observed output combines effort, ability, task comprehension, fatigue and technological constraints. A treatment affecting output may change effort or the productivity of effort. Randomizing a productivity aid separately from compensation, measuring time and accuracy, and using repeated tasks under controlled conditions can help distinguish these mechanisms.
Across these examples, the strongest empirical result concerns a precisely described assignment and a defined outcome: an application signal and callback, a credit offer and uptake, an informational intervention and consumption, an incentive treatment and task output. Broader claims about discrimination, preferences, financial development, social comparison or motivation require the theoretical and institutional arguments connecting those outcomes to the proposed mechanism.
Replication, statistical discipline, and research integrity
Replication performs several scientific functions. A direct replication tests whether the designated original effect can be recovered using a sufficiently close design. A conceptual replication changes features of the design to test the persistence of an underlying mechanism. An extension studies the mechanism in a different population or institutional environment. A reproducibility exercise reconstructs the reported analysis from available data and code. These activities should be distinguished because failure in one does not have the same interpretation as failure in another.
Camerer et al. (2016), in Evaluating Replicability of Laboratory Experiments in Economics, examine 18 selected laboratory studies published in leading economics journals during 2011–2014. Eleven replications produced significant effects in the original direction, and the mean replicated effect size was 66 percent of the original estimate. These are outcomes of a specific replication exercise, not a universal failure or success rate for experimental economics. Different replication criteria, study selections, power calculations and estimands would produce different measures.
An original study selected for a significant result is susceptible to estimation error conditional on selection. When the true effect is modest and sampling variation large, the significant estimate can overstate its magnitude. A replication powered on that original estimate may achieve less power for the underlying effect than anticipated. Comparing confidence intervals, estimated effects, and theoretically meaningful effect sizes is more informative than classifying results solely by a significance threshold.
A statistically insignificant result does not establish zero. An equivalence or noninferiority analysis can assess whether an interval excludes effects larger than a specified economically meaningful magnitude, supplied before observing the results. Such tests require a defensible tolerance. For a labor intervention, even a small effect on callbacks may have economic importance at population scale; for a costly treatment, a larger effect may be required to change a decision about implementation.
Preregistration helps record the outcome family, exclusion rules, primary contrast, planned sample size and statistical approach. It does not make every analysis correct. Deviations may be justified by implementation failures or unforeseen measurement problems; reporting them explicitly preserves the distinction between prespecified and exploratory analyses. Publishing code, item definitions, treatment instructions, and appropriately deidentified data facilitates verification subject to consent, confidentiality and lawful access restrictions.
Researcher demand presents a distinct validity challenge. Participants may infer a desired response and act accordingly. Randomization does not prevent this if the treatment signal conveys the desired direction. De Quidt, Haushofer, and Roth (2018), in Measuring and Bounding Experimenter Demand, use structured cues to study and bound demand effects in experimental tasks. This introduces a tractable sensitivity exercise: compare behavior under explicit cues designed to push the outcome in alternative directions, and assess the implications under assumptions about how those cues relate to implicit demand.
The procedure is specific to its assumptions and implementation. Demand may interact with social identity, perceived authority, stakes and context. An explicit cue can differ from a subtle cue embedded in instructions. The researcher should avoid treating a bound estimated in one task as a correction automatically applicable to a different task. Blinding where feasible, neutral scripts, private responses and robustness designs reduce the role of experimenter expectations and help interpret the observed behavior.
Research integrity also requires transparent handling of the measurement scale. A percentage-point change, a relative percentage change, a standardized effect size and a monetary change have different denominators. Converting between them requires the relevant baseline or variance. A paper that reports changes in standardized units can aid comparisons, but economic interpretation usually requires outcomes in their original units as well.
For scientific archives, the data dictionary should record the randomization unit, intervention version, exposure time, task order, payment rule, eligibility conditions, missingness, variable transformations and any conditions of access. These records allow later researchers to distinguish a genuine replication of an estimand from a superficially similar study that altered an essential condition.
Welfare and the limits of behavioral inference
Economic choice supplies information about what people select from feasible options. Welfare analysis supplies a ranking of consequences according to a specified normative or evaluative criterion. A consistent choice pattern may rationalize a stable preference ordering while leaving the welfare implications of an intervention unresolved. An inconsistent pattern may reflect changing objectives, constraints or information, rather than a welfare loss with a uniquely identified magnitude.
For an intervention that reduces the cost of a transaction, observed take-up can help identify demand over the offered range. Estimating consumer surplus requires assumptions about price variation, the relation between willingness to pay and welfare, substitution, income effects, and additional consequences. If the price assignment changes perceived quality or trust, demand changes cannot automatically be integrated into a conventional willingness-to-pay curve.
Consider a constructed example. A program offers a service to 100 people at an administrative cost of 3 monetary units per targeted person. Suppose it generates 10 incremental users, so the total administrative cost is 300 and the cost per incremental user is 30. These figures concern implementation cost relative to a particular causal outcome. A welfare conclusion additionally requires the incremental social benefits, private and public costs, distributional weights, possible displacement of other services, and whether the benefits are already reflected in the chosen outcome. A cost per induced use is neither net social surplus nor a cost-benefit ratio.
The evaluation of behavioral interventions poses further normative questions. A default, reminder or frame may change action without changing underlying opportunities. If choices vary with presentation, the researcher must specify which presentation, informed choice, or experienced consequence is relevant to welfare. A treatment that increases take-up can reduce administrative friction for some people and encourage choices contrary to the informed preferences of others. The direction of the average treatment effect alone does not settle the welfare comparison.
The time horizon matters. Credit may finance investment while imposing repayment obligations; reduced energy consumption may save money while changing comfort; effort incentives may increase output while increasing fatigue. Measuring only the immediate behavior leaves these consequences outside the empirical evaluation. A complete welfare analysis defines relevant outcomes over a suitable horizon and avoids counting transfers as social costs or benefits without identifying the perspective and opportunity costs.
Experiments can contribute to welfare economics by measuring multiple consequences, prices, informed preferences, experienced outcomes and distributional effects. They cannot replace the explicit normative step. The scientific task is to state what has been causally identified and what additional preference, ethical, institutional or equilibrium assumptions connect the result to a welfare ranking.
Comparative synthesis and research agenda
The studies reviewed establish a series of complementary inferential relationships. Controlled budget tasks identify restrictions on choices under known alternatives. Induced-value and independent-price mechanisms create observable environments where the meaning of selected actions can be derived under maintained assumptions. Risk and reference-dependence experiments test the stability of outcome rankings against variation in stakes, frames and histories. Social games distinguish payoff predictions from broader sets of possible motives. Naturalistic assignments in markets and organizations identify intervention effects under existing institutions.
The author's analytical synthesis rests on comparing the scope of these results rather than aggregating them into a single measure of economic rationality. Five questions recur across designs. What is the feasible set? What consequences does the decision maker understand and face? Which benchmark generates the empirical restriction? Which intervention contrast is identified? Which features of the source environment are preserved when the result is applied elsewhere? These questions organize evidence within established experimental and microeconomic methodology; they are not proposed as a new named theory or measurement system.
The first relationship concerns the distinction between behavior and preference. An observed action is evidence about preference conditional on feasible alternatives, knowledge and the governing decision model. Revealed preference is powerful because it uses comparative observations to restrict rankings. Its power is attenuated when budgets, available alternatives or outcome perceptions are uncertain. Field studies with rich transaction data can observe consequential behavior while facing these difficulties. Complementary laboratory tasks can resolve specific measurement questions by constructing known alternatives. The combined result is more informative when their choice environments are explicitly connected.
The second relationship concerns incentives and interpretation. The economic derivation of truth-telling proves an optimal action inside a precisely specified mechanism. Empirical validation establishes whether the implemented mechanism produces behavior compatible with that derivation in the intended population. The BDM illustration makes the difference precise: reporting the private valuation weakly dominates deviations under independent pricing and quasilinear consequences. If the participant misunderstands the rule or cannot finance a purchase, the derivation alone does not identify the intended valuation. Field implementation is not an afterthought; it supplies evidence on premises that formal mechanism analysis holds fixed.
The third relationship concerns internal validity and external validity. Randomization within a study identifies an assignment contrast, whereas prediction in a future population requires information about population selection and institutional stability. The hypothetical sign reversal from changing subgroup composition proves that a locally valid average can have the opposite sign in a target population even when conditional effects remain fixed. The empirical site-selection evidence makes the underlying issue consequential for actual program evaluation. A multi-site experimental architecture with documented site recruitment, common outcome definitions and planned heterogeneity tests can help establish the domain of transport.
The fourth relationship concerns explanatory and evaluative claims. A treatment effect is a response to an assignment; its mechanism must be identified with additional restrictions or variation. The social-game example shows why positive transfers can arise under several motives, and the correspondence experiment shows why a name-signal effect does not identify the entire discrimination mechanism. In each case, the appropriate response is a new targeted contrast, not a stronger interpretation of the existing coefficient. Welfare conclusions add the further requirement of a defensible ranking of consequences.
The fifth relationship concerns cumulative evidence. A scientific comparison should preserve the experimental object's identity across studies while documenting changes in measurement, stakes, population, exposure and analysis. Replications can challenge original estimates, reveal unstable contexts or refine error assumptions. A failure to reproduce an effect is information about the conjunction of theory, design and environment. A successful replication strengthens a particular inference under the replicated conditions. A cumulative account progresses by specifying which restrictions survive which changes.
These relationships have practical consequences for research programs. A rigorous test of an individual choice model should combine variation in budgets or lotteries with comprehension checks, a prespecified error treatment and meaningful measures of consistency. A study of an inducement to change behavior should keep assignment, treatment receipt and measured outcomes distinct. A transnational comparison should record differences in purchasing power, institutional opportunity sets, translation and mode of subject recruitment. A program evaluation intended for scale should examine interference and implementation capacity before converting an individual or pilot estimate into a policy prediction.
Several priorities follow. First, linked within-person studies can examine whether observed preference consistency is stable across time, interfaces, stakes and decisions involving real goods. A credible design must handle attrition and state dependence rather than attribute every cross-task change to unstable preferences. Second, bridge studies between laboratory and natural field contexts should vary a small number of well-defined features while retaining a common economic restriction. Third, valuation mechanisms should be validated with induced benchmarks and explicit checks of financing constraints and understanding, especially before using elicited valuations for welfare calculations.
Fourth, heterogeneity and transport should be planned as central design objects. Samples recruited for convenience can identify mechanisms but cannot, without further assumptions, identify target-population means. Diverse sites and prespecified effect modifiers increase the scope of conditional results, while a deliberately held-out implementation can test predictive validity. Fifth, scientific archives should preserve complete research instruments and code when permissible, and records of the original hypothesis, sample exclusions and deviations. An experiment's scientific meaning often depends on small procedural differences invisible in its abstract.
The geographical distribution of examples in this dossier includes studies from the United States, India, Uganda, and cross-society experimental research. The source set does not represent all national economies, labor institutions, or household environments. Direct full-text access varied; bounded claims are tied to the available primary passages or author abstracts. The dossier reports no original microdata, independently executed empirical replication, or exhaustive survey of post-2021 literature. These limitations define the scope of the synthesis and identify where new empirical work would materially extend it.
Conclusion
Laboratory and field tests provide an indispensable connection between theoretical microeconomics and observed economic choice. Their primary contribution is the construction of environments in which choice restrictions and causal responses can be evaluated with specified alternatives, information, consequences, and institutional rules. The resulting evidence supports a differentiated account of rationalizability, preference estimation, incentive response, strategic behavior, and causal intervention effects.
Controlled budgets and induced values make the assumptions behind an empirical restriction unusually transparent. Risk elicitation and reference-dependence experiments show why the scale, presentation and history of consequences matter. Social and strategic experiments expose the joint role of preferences and beliefs. Field experiments identify interventions within the institutional arrangements that generate economically meaningful actions. Replications, validation studies and cross-site comparisons reveal the conditions under which these findings can be maintained and extended.
A mature experimental microeconomics preserves the identity of its estimands. It distinguishes consistency from psychological explanation, truthful elicitation from successful implementation, causal assignment effects from preference parameters, local estimates from transportable claims, and behavioral changes from welfare improvements. Each distinction specifies a substantive research task. Together they sustain the interpretation of laboratory and field results as complementary scientific evidence about choice in actual economic institutions.
References
- Allcott, H. (2015). Site Selection Bias in Program Evaluation. Quarterly Journal of Economics, 130(3), 1117–1165. DOI: 10.1093/qje/qjv015. Source record
- Banerjee, A., Duflo, E., Glennerster, R., & Kinnan, C. (2015). The Miracle of Microfinance? Evidence from a Randomized Evaluation. American Economic Journal: Applied Economics, 7(1), 22–53. DOI: 10.1257/app.20130533. Source record
- Becker, G. M., DeGroot, M. H., & Marschak, J. (1964). Measuring Utility by a Single-Response Sequential Method. Behavioral Science, 9(3), 226–232. DOI: 10.1002/bs.3830090304. Source record
- Berg, J., Dickhaut, J., & McCabe, K. (1995). Trust, Reciprocity, and Social History. Games and Economic Behavior, 10(1), 122–142. DOI: 10.1006/game.1995.1027. Source record
- Bertrand, M., & Mullainathan, S. (2004). Are Emily and Greg More Employable than Lakisha and Jamal? A Field Experiment on Labor Market Discrimination. American Economic Review, 94(4), 991–1013. DOI: 10.1257/0002828042002561. Source record
- Burchardi, K. B., de Quidt, J., Gulesci, S., Lerva, B., & Tripodi, S. (2021). Testing Willingness to Pay Elicitation Mechanisms in the Field: Evidence from Uganda. Journal of Development Economics, 152, 102701. DOI: 10.1016/j.jdeveco.2021.102701. Source record
- Camerer, C. F., Dreber, A., Forsell, E., Ho, T.-H., Huber, J., Johannesson, M., Kirchler, M., Almenberg, J., Altmejd, A., Chan, T., Heikensten, E., Holzmeister, F., Imai, T., Isaksson, S., Nave, G., Pfeiffer, T., Razen, M., & Wu, H. (2016). Evaluating Replicability of Laboratory Experiments in Economics. Science, 351(6280), 1433–1436. DOI: 10.1126/science.aaf0918. Source record
- Choi, S., Fisman, R., Gale, D., & Kariv, S. (2007). Consistency and Heterogeneity of Individual Behavior under Uncertainty. American Economic Review, 97(5), 1921–1938. DOI: 10.1257/aer.97.5.1921. Source record
- Choi, S., Kariv, S., Müller, W., & Silverman, D. (2014). Who Is (More) Rational? American Economic Review, 104(6), 1518–1550. DOI: 10.1257/aer.104.6.1518. Source record
- de Quidt, J., Haushofer, J., & Roth, C. (2018). Measuring and Bounding Experimenter Demand. American Economic Review, 108(11), 3266–3302. DOI: 10.1257/aer.20171330. Source record
- DellaVigna, S., & Pope, D. (2018). What Motivates Effort? Evidence and Expert Forecasts. Review of Economic Studies, 85(2), 1029–1069. DOI: 10.1093/restud/rdx033. Source record
- Duflo, E., Glennerster, R., & Kremer, M. (2007). Using Randomization in Development Economics Research: A Toolkit. In T. P. Schultz & J. Strauss (Eds.), Handbook of Development Economics (Vol. 4, pp. 3895–3962). Elsevier. DOI: 10.1016/S1573-4471(07)04061-2. Source record
- Harrison, G. W., & List, J. A. (2004). Field Experiments. Journal of Economic Literature, 42(4), 1009–1055. DOI: 10.1257/0022051043004577. Source record
- Henrich, J., Boyd, R., Bowles, S., Camerer, C., Fehr, E., Gintis, H., & McElreath, R. (2001). In Search of Homo Economicus: Behavioral Experiments in 15 Small-Scale Societies. American Economic Review, 91(2), 73–78. DOI: 10.1257/aer.91.2.73. Source record
- Holt, C. A., & Laury, S. K. (2002). Risk Aversion and Incentive Effects. American Economic Review, 92(5), 1644–1655. DOI: 10.1257/000282802762024700. Source record
- Holt, C. A., & Laury, S. K. (2005). Risk Aversion and Incentive Effects: New Data without Order Effects. American Economic Review, 95(3), 902–912. DOI: 10.1257/0002828054201459. Source record
- Kahneman, D., & Tversky, A. (1979). Prospect Theory: An Analysis of Decision under Risk. Econometrica, 47(2), 263–291. Source record
- Levitt, S. D., & List, J. A. (2007). What Do Laboratory Experiments Measuring Social Preferences Reveal about the Real World? Journal of Economic Perspectives, 21(2), 153–174. DOI: 10.1257/jep.21.2.153. Source record
- List, J. A. (2003). Does Market Experience Eliminate Market Anomalies? Quarterly Journal of Economics, 118(1), 41–71. DOI: 10.1162/00335530360535144. Source record
- Muralidharan, K., & Niehaus, P. (2017). Experimentation at Scale. Journal of Economic Perspectives, 31(4), 103–124. DOI: 10.1257/jep.31.4.103. Source record
- Smith, V. L. (1962). An Experimental Study of Competitive Market Behavior. Journal of Political Economy, 70(2), 111–137. DOI: 10.1086/258609. Source record
- Smith, V. L. (1976). Experimental Economics: Induced Value Theory. American Economic Review, 66(2), 274–279. Source record
- Varian, H. R. (1982). The Nonparametric Approach to Demand Analysis. Econometrica, 50(4), 945–973. Source record
Social choice, strategic behavior, and context
Many microeconomic decisions involve the welfare, beliefs, or actions of other people. A social preference specifies how own and others' consequences enter choice; strategic behavior adds beliefs about the responses of others. A laboratory game can control feasible actions and payments while leaving variation in the extent to which subjects care about reciprocity, fairness, reputation, norms or others' payoffs. A test of a purely self-interested equilibrium is therefore a test of a joint package of preferences, information, strategy, and beliefs.
Berg, Dickhaut, and McCabe (1995) introduce an influential controlled investment game in Trust, Reciprocity, and Social History. One participant decides how much to transfer, the transfer is multiplied, and another participant decides how much to return. The experiment organizes observable investment and reciprocation within a known opportunity structure. Positive transfers have substantive interest because they can reveal behavior beyond a simple self-interested backward-induction benchmark.
For an illustrative version, let the first participant possess 10 units and send t, with 0 ≤ t ≤ 10. The second receives 3t and may return any amount s between 0 and 3t. If both maximize only their own monetary payments, the second chooses s = 0 at every t. Anticipating that response, the first chooses t = 0. This prediction follows from the assumptions of self-interest, the given timing, the feasible choices, and knowledge of the response incentives. A positive t can be consistent with expected reciprocation, altruism, inequality aversion or other objectives. A positive s may reflect reciprocity, fairness, reputation concerns, or experimenter scrutiny. Neither transfer uniquely identifies trust as a structural parameter.
A dictator game removes the recipient's ability to reject or reciprocate, altering the strategic environment. An ultimatum game adds rejection and thus strategic expectations regarding the recipient's response. The games may therefore help separate motives when examined together, particularly with independent belief measurements or manipulations of observability. Their nominal labels should not be treated as direct measures of unconstrained psychological characteristics.
Cross-cultural experimentation exposes the importance of the decision context. Henrich et al. (2001), in experiments in 15 small-scale societies, document differences in social-game behavior across distinct societies. The observed heterogeneity and its reported associations with economic organization support scrutiny of universal behavioral predictions. The evidence does not establish globally representative distributions of preferences or a causal effect of market integration on fairness. It demonstrates the scientific value of testing a common task in institutions with different social and economic histories.
Comparability requires more than translating instructions literally. An equal nominal transfer can have different purchasing power, significance relative to income, or meaning within local exchange norms. Experimental anonymity may be interpreted differently depending on the social network and research setting. A payoff to an unnamed recipient can be imagined as a gift, an obligation, or a random allocation. Researchers should document recruitment, translation and back-translation where appropriate, local stakes, comprehension, and participants' interpretation of the exchange.
Levitt and List (2007) examine laboratory measures of social preferences and the circumstances under which they inform real-world behavior. Scrutiny, stakes, context and selection can influence the relation. The analytical response is to test dimensions of context rather than treat the laboratory result as biased by a single unknown scalar. For example, varying privacy can help distinguish behavior sensitive to observation from behavior stable under private choice. Varying the identity of the recipient can reveal the role of social distance. Replicating under genuine institutional consequences tests a different dimension.
Social choice is also consequential in aggregate. Fairness concerns may affect bargaining, wage setting, voluntary contributions, contracting and reciprocal exchange. Yet the parameter estimated from one game must enter an institutional model with explicit assumptions about counterparties, information and repeated interactions before it can predict aggregate outcomes. The experimental identification of a motive and its equilibrium significance are separate stages of inquiry.