RESEARCH DOSSIER
Empirical Counterexamples and Economic Model Validation
Abstract. Economic models acquire scientific standing through the restrictions they impose on observable economic reality and through the quality of the evidence used to evaluate those restrictions. This monograph examines empirical counterexamples as instruments for identifying the boundaries of economic explanation, distinguishing logical counterexamples from statistical discrepancies, causal contradictions, failures of measurement, and failures of transport across institutions and policy regimes. It reconstructs the logic of deterministic, probabilistic, structural, and predictive validation, presents elementary formal demonstrations, and evaluates historically important disputes over minimum wages, decision under risk, household consumption, financial asset pricing, and macroeconomic policy. It connects these disputes to identification, replication, forecast evaluation, and the selection of competing mechanisms. Its central analytical conclusion is that an anomaly is scientifically informative in proportion to the specificity of the restriction it challenges and the independence of the evidence that localizes the failure. An empirically useful model may retain a restricted domain after a wider claim has failed; revisions acquire credibility from distinct, prospective tests. The study is a critical synthesis of established research, supplemented by transparent illustrative derivations rather than a new econometric estimation or a new named theory.
Keywords: Economic methodology; empirical counterexamples; model validation; falsifiability; causal identification; auxiliary assumptions; economic anomalies; minimum wages; expected utility; asset pricing; replication; out-of-sample prediction; external validity; specification testing
Introduction
Economic models make selected features of economies intelligible by imposing structure on the relations among choice, technology, institutions, expectations, prices, and allocations. Their analytical value derives from restrictions that connect assumptions to consequences. A model in which every possible outcome can be explained by choosing a suitable unobserved parameter may organize description, yet its empirical content remains weak until the permitted parameter values or their distribution receive independently justified limits. The empirical evaluation of a model begins with the identification of a claim that could in fact be wrong.
There are several distinct purposes for which economists construct models. A pricing model may be intended to estimate an unobservable marginal valuation. A labor-market model may explain how employment responds to a statutory wage change. A forecasting model may seek to predict inflation over the next four quarters without identifying the effects of a monetary policy intervention. A welfare model may rank social choices conditional on a normative criterion. The standards of evidence for these purposes overlap, but they do not coincide. An extraordinary forecast can arise from a correlation that is useless for intervention; a well-identified local treatment effect may be a poor forecast of the national aggregate; a welfare conclusion may follow from explicit normative premises even when individual preferences are heterogeneous. Validation has a specified object and a specified intended use.
This dossier studies the movement from observed discrepancies to justified model evaluation. It treats Economic Theory as its writing owner and Microeconomics and Economic Methodology as its substantive companion fields. Its scope centers on the evidence that discriminates between empirical models, rather than supplying an encyclopedic treatment of estimation techniques, the history of each substantive theory, or every method of experimentation. Five case studies bring different testing problems into view. Minimum wage research illustrates the comparison of potentially confounded causal effects. Choice anomalies concern the structural restrictions of preference representations. Consumption tests separate predictability statements from the broader claim of intertemporal optimization. Asset pricing turns an apparently modest discrepancy in moments into a demanding restriction on discount factors. Macroeconomic policy emphasizes the possibility that the act of intervention changes the relation previously estimated.
The research method is a purposive comparative synthesis of canonical papers and methodological literature, selected for their capacity to illustrate different kinds of empirical restriction. It is not a systematic review with an exhaustive search protocol, meta-analysis, or replication of proprietary microdata. Numerical examples marked illustrative are constructed in this monograph; historical and empirical figures are attributed to the studies that report them. Where the available evidence is an abstract or institutional record, the discussion remains at the level supported by that evidence. The scientific issue is the logic of empirical adjudication, not a quantitative recommendation about a particular policy.
The historical problem of empirical adjudication
The historical methodological question is how principles derived in abstraction can warrant inferences about a complex economy. The tradition associated with Mill recognized the significance of isolating causal forces whose consequences might be modified by other influences. Later methodological arguments emphasized operational implications, predictive success, or severe opportunities for refutation. Hausman's survey of economic methodology (2018) places these approaches in the context of the increasing centrality of mathematical and experimental models.
A formal system can yield a valid theorem from deliberately idealized premises. A perfectly competitive model, for example, can clarify the relationship between price and marginal cost under price-taking behavior. Its mathematical validity is independent of whether a particular market has important entry barriers. Empirical evaluation begins when the model is assigned to an observed market and the conditions required for the theorem are supplied or measured. Failure of an empirical implication calls into question that application, the auxiliary conditions, or the proposed explanation. The theorem remains a mathematical consequence of its premises.
The predictive emphasis associated with Friedman has an economic attraction: abstraction is unavoidable, and success in predicting consequences provides a reason for using a tractable model. Yet prediction becomes ambiguous when the prediction target and the intervention differ. Consider a retailer whose inventory purchases predict tomorrow's demand. This is a useful forecasting relationship if managers anticipate customer activity. Increasing inventories to induce the same rise in demand would require a distinct causal account. A favorable predictive score establishes evidence for the task represented by the score.
The Popperian distinction between claims that exclude possible observations and claims compatible with all conceivable outcomes introduces a stricter demand. In economics the demand encounters probabilistic predictions, unobserved quantities and ceteris paribus assumptions. A prediction about an expectation does not forbid every individual realization. A claim about the consequences of changing one variable while holding others constant may require an experiment, a natural experiment, a structural argument, or another source of exogenous variation. Empirical criticism becomes rigorous through a statement of the conditions under which the model is exposed to contradiction.
A useful historical lesson is that methodological disagreement concerns the relation between tasks and evidence. A model intended to isolate a mechanism can be evaluated through the reality and strength of the mechanism, while a model intended as a predictive instrument can be compared through prespecified forecast loss. A complete explanation of a market episode may require several mechanisms and an account of their interaction. The choice among these aims precedes the choice of a goodness-of-fit statistic.
The anatomy of a counterexample
A logical counterexample supplies an object satisfying the premises of a general proposition but violating its conclusion. For a proposition of the form “for every x in D, P(x) implies Q(x),” it is enough to exhibit an x in D with P true and Q false. A mathematical counterexample operates entirely within the definitions of a formal system. An empirical counterexample claims that an observed or experimentally constructed case belongs to the domain D, satisfies the relevant operational premises, and conflicts with the prescribed observable implication. Each of these three conditions requires evidence.
Economic restrictions can be directional, numerical, ordinal, distributional, or invariant. A directional restriction predicts that increasing a binding price floor reduces quantity traded under specified competitive conditions. A numerical restriction specifies the response elasticity or a moment within a range. An ordinal restriction places one bundle or lottery above another. A distributional restriction imposes a relationship on conditional means, variances, quantiles, or dependence. An invariance restriction says that a structural parameter or relationship persists when some aspect of the environment changes. Each form generates different opportunities for contradiction.
An anomaly is an observed pattern that is surprising under a reference model. An anomaly need not be a certified counterexample. If the statistic used to detect it is noisy, a high probability of observing similarly extreme values weakens the case. If the definition of the outcome differs from the model's theoretical quantity, the discrepancy may reflect the mapping between data and theory. If an application lies outside the stipulated domain, the result establishes a boundary of usefulness rather than a violation of the theorem within its intended domain.
A decisive empirical contradiction requires an observable implication, credible measurement, a defensible assignment of the case to the domain, and uncertainty small enough to distinguish the contradictory outcome from the model's permitted range. These requirements explain why the same event has different scientific significance for different models. For a deterministic theory of a precisely observed relationship, a single case can be highly informative. For a probability model that assigns positive density to all real outcomes, the relevant issue is the relative concentration of the distribution and the probability of a test statistic, rather than a literal impossibility.
Counterexamples also differ in their level of localization. A recorded outcome may challenge an entire conjunction of assumptions without distinguishing them. Multiple observations may establish that no parameterization in a restricted family accommodates the data. An experiment designed to vary one mechanism while stabilizing others can localize the discrepancy more sharply. A rival theory acquires stronger support when it predicts the relevant pattern before the researcher uses the observation to choose the rival's parameters.
Formal restrictions and statistical evidence
Let a structural economic model M contain a parameter vector θ, a collection of auxiliary assumptions A, and a measurement relation H connecting latent variables to observations. In a domain D the model implies a family of allowable data distributions:
The empirical question is whether the relevant aspects of the actual distribution are compatible with this family at an appropriate level of approximation. The distinction between a single distribution and a family is crucial. Rejecting one calibration of θ leaves open whether another admissible calibration satisfies the same independent restrictions. A model with an arbitrarily flexible parameter for every observation may attain excellent fit at the cost of empirical discipline.
Moment restrictions and overidentification
Suppose model M implies E[g(Z,θ)]=0, where Z denotes observed data, g is a q-dimensional vector, and θ has p free parameters. An estimation procedure can choose θ by minimizing a quadratic form in the sample moments, with a positive definite weighting matrix. Hansen's GMM analysis (1982) supplies a general econometric foundation for such procedures.
Writing the sample mean as ḡ(θ), the statistic
has an asymptotic chi-square reference distribution with q−p degrees of freedom under the usual correct-specification, identification, regularity, and efficient-weighting conditions. This is the standard overidentification test; it is not a universally valid goodness-of-fit formula. Weak identification, a singular moment covariance, inappropriate dependence corrections, or estimated nuisance quantities can change the relevant inference. An exact-identification case with q=p generally supplies no overidentification degrees of freedom.
In a transparent illustrative example, two population moments require θ=2 and 2θ=6. The first gives θ=2; the second gives θ=3. No scalar satisfies both. The contradiction is between the two restrictions jointly. If sample estimates of these moments are 2 and 6, the same numerical pattern by itself is insufficient to reject the model until sampling errors and their covariance are considered. A defensible test incorporates both uncertainty and the estimated parameter.
Statistical significance and economic magnitude
A small p-value in a valid test indicates that the observed test statistic would be unusual under the specified null distribution. It is not the posterior probability of the null hypothesis. A high p-value indicates an absence of sufficiently strong evidence for rejection in that design; it does not certify the model. The information contained in a confidence interval is often more useful than a dichotomous decision. An interval spanning both a substantial positive and substantial negative elasticity leaves the sign economically unresolved.
Suppose a model predicts an employment elasticity of −0.20. An illustrative estimate of −0.05 with standard error 0.04 gives an approximate 95 percent normal interval of [−0.1284, 0.0284]. Conditional on a valid standard error, this estimate is inconsistent with the point −0.20 at the stated level even though the estimated effect is not statistically distinguishable from zero. A model predicting any elasticity in [−0.30, 0] would face a different test. Statistical evaluation must match the granularity of the claim.
Economic importance requires a tolerance appropriate to the application. A forecasting discrepancy of one unit may be inconsequential for an aggregate annual quantity and decisive for a balance-sheet constraint. For equivalence claims a prespecified tolerance δ can be tested by asking whether the confidence interval is contained in [−δ,δ]. Failure to reject zero does not achieve that result. The tolerance derives from the economic objective and the losses attached to errors.
Multiple restrictions and search
Searching many subgroups, time windows, models, and outcomes and retaining the largest discrepancy changes its reference distribution. For twenty independent tests at nominal size 0.05 under a complete null, the illustrative probability of at least one false rejection is 1−0.95²⁰, about 0.642. Dependence changes this calculation, but the example establishes the cost of unreported search. Exploratory anomalies are valuable for generating new restrictions; their confirmatory evaluation needs a prespecified test, appropriate multiplicity control, or independent observations. The informative question is whether the pattern survives a test whose outcome was not used to construct it.
Measurement and the construction of observable quantities
Economic observations are produced through definitions, records, survey instruments, accounting categories, price indexes, and statistical procedures. Theory refers to quantities such as marginal utility, productivity, expectations, economically relevant labor input, and real wealth. The distance between these concepts and available data is an empirical problem in its own right.
An employment headcount counts persons or positions; hours worked measures labor input over an interval; full-time-equivalent employment converts reported categories into a standardized measure. A firm that reduces weekly hours and increases the number of part-time workers can display increasing headcount alongside declining total hours. An economic model of cost-minimizing adjustment may predict one dimension rather than another. Crosswalks between theoretical quantities and statistical records should therefore be specified before comparing estimates.
Price data involve analogous distinctions. Nominal earnings, purchasing power, relative prices, and total returns answer different questions. A nominal interest rate can increase while the real expected interest rate falls if expected inflation increases enough. A time series measured using today's revised national accounts may contain information that a forecaster operating decades earlier could not have used. Evaluation of historical prediction benefits from vintage data and a transparent information cutoff.
A classical measurement error example clarifies the mechanism. Suppose Y=βX+u with E[u|X]=0 and the researcher observes X*=X+v. If v is mean zero and independent of X and u, the population slope from regressing Y on X* is
For β=2 and equal variances of X and v the observed-regressor slope equals 1. A theory predicting the true structural coefficient 2 can coexist with an observed coefficient 1 under this measurement model. The derivation does not justify assuming classical error in actual survey data; systematic misreporting, correlated errors and censoring require distinct analysis.
Measurement disagreements become productive when data sets represent different collection methods whose biases have different directions or correlates. Independent administrative and survey evidence can help localize systematic mismeasurement. Researchers should compare unit definitions, time coverage, missingness, sampling frames, business exits, treatment timing, and aggregation weights. Replication using the same biased measure can reproduce an estimate while leaving the measurement problem untouched.
Identification of causal contradictions
Many economic model predictions concern what an intervention would do, rather than the association among existing observations. Let Y(1) and Y(0) be potential outcomes with and without treatment. For each unit only one is observed under a consistent treatment definition. The average treatment effect E[Y(1)−Y(0)] is a counterfactual population quantity. Its identification requires sufficient restrictions on how treatment was assigned and how outcomes would otherwise evolve.
Assignment and instruments
Random assignment can make treatment independent of potential outcomes in the assignment mechanism, permitting unbiased comparison in suitable designs. Attrition, interference, noncompliance, missing outcomes and changes in treatment versions may complicate inference about the intended estimand. Deaton and Cartwright (2018) emphasize the need for substantive justification when extending trial evidence beyond the selected sample. Randomization is powerful evidence about assignment within an experiment, not an automatic solution to transport or mechanism identification.
With an instrumental variable Z and treatment D, the Wald ratio is
The potential-outcomes treatment of instrumental variables by Angrist, Imbens, and Rubin (1996) identifies a complier average under conditions including appropriate independence, exclusion, treatment relevance, monotonicity, and consistency. Its estimand is specific to units whose treatment is changed by the instrument. If always-takers and never-takers have different treatment effects, the estimate does not establish the population average. If the instrument directly affects the outcome, the exclusion restriction fails. A weak denominator creates a different statistical difficulty discussed in weak-instrument inference (Andrews, Stock, and Sun, 2019).
Consider an illustrative training subsidy with assignment-induced take-up rates 0.60 and 0.30 and mean annual earnings differing by $900. The ratio is $3,000 for induced participants under the assumptions above. A model predicting $3,000 only for compliers is consistent with the example. A model predicting $3,000 for all eligible people is consistent but less directly tested; a model predicting negative effects for every participant faces a sharper contradiction. The same estimate has different force depending on the mechanism and population named in the model.
Difference-in-differences
For a two-period comparison of treated and comparison groups, let μ_{g,t} denote mean outcomes. The difference-in-differences contrast is
For causal interpretation the relevant untreated potential-outcome trends must be parallel in expectation, possibly conditional on measured covariates; treatment anticipation, spillovers and changing composition require further attention. A constructed example has treated employment increasing from 100 to 104 and comparison employment increasing from 100 to 102. The contrast is +2 workers. If independent evidence showed that the untreated treated-group trend would have been +5 instead, the relevant causal effect would be −1. The observed +2 contrast and −1 causal effect are compatible when parallel trends fails.
Serial correlation can produce overly optimistic precision if regression errors are treated as independent over time. Bertrand, Duflo, and Mullainathan (2004) address inference problems in this design. When treatment timing varies, cohort composition and dynamic effects must be distinguished; the multiple-period extension studied by Callaway and Sant’Anna (2021) is relevant here. Pre-treatment diagnostics provide evidence about design plausibility rather than direct observation of the unrealized post-treatment counterfactual.
Mechanism and outcome
A credible estimate of an effect on Y still need not distinguish among the mechanisms generating it. A positive work incentive may change employment through labor supply, job search, hours, employer recruitment, or a shift in industry composition. Estimating the outcome establishes an effect under the design; isolating a mechanism demands additional restrictions or separately measured intermediate outcomes. Conditioning on a variable affected by treatment can introduce selection into mediation analysis.
Models of different mechanisms may generate observationally equivalent reduced forms in the available experiment. If both predict Y rises by the same amount, the increase cannot discriminate between them. Their differences may be located in wage distributions, turnover, response delays, heterogeneous effects, or behavior outside the original treatment range. A new test should target one of those differences rather than merely reproduce the common prediction.
Labor markets and the minimum wage controversy
Competitive and imperfectly competitive restrictions
A simple competitive labor market describes employers as wage takers. If a firm produces F(L), earns output price p, and pays wage w, an interior profit-maximizing choice satisfies pF'(L)=w. With p fixed and diminishing marginal product F''(L)<0, differentiation gives dL/dw=1/[pF''(L)]<0. A binding increase in the effective wage reduces desired labor in this partial-equilibrium model. This restriction has real economic content because its sign follows from explicit competition, technology, adjustment and price conditions.
A stylized monopsony gives a different restriction. With an upward-sloping wage schedule w(L), total labor expenditure is Lw(L), and marginal expenditure is w(L)+Lw'(L). An employer may choose lower employment than in the competitive benchmark. A suitably placed minimum wage can flatten the relevant marginal expenditure schedule and raise employment over part of the range. A sufficiently high floor eventually constrains demand. Other models introduce search frictions, turnover, employer entry, product-price adjustments, or productivity changes. Their implications depend on the relevant margin and length of adjustment.
For illustration, let marginal revenue product be 20−L and the inverse labor supply w=2+L. A monopsonist has marginal expenditure 2+2L, so L=6 and w=8. Competitive clearing occurs at L=9 and w=11. A minimum wage of 10 supplies up to eight workers at that wage, while labor demand there is ten; the firm optimally employs eight on the flat segment. At a minimum wage of 14, labor demand is six, below the competitive quantity. This constructed example illustrates the nonmonotonic employment implication; it is not an estimate of any labor market.
The New Jersey and Pennsylvania evidence
New Jersey raised its minimum wage from $4.25 to $5.05 per hour on April 1, 1992; the comparison rate in Pennsylvania stayed at $4.25. Card and Krueger's fast-food study (1993 working paper; 1994 journal article) surveyed 410 restaurants before and after the change. Their comparison did not exhibit the employment decline predicted by a simple universal interpretation of the competitive model. The policy provides a case in which an empirically important wage intervention and an informative comparison group challenge a strong directional extrapolation.
Scientific attention centered on the measurement and design of the comparison. Neumark and Wascher's payroll re-evaluation (1995 working paper; 2000 comment) used records from 230 restaurants and reported an adverse effect. Card and Krueger's representative payroll reanalysis (1998 working paper; 2000 reply) reported no differential hours growth in the combined comparison after accounting for differences in reporting intervals. The papers provide a methodological dispute about the evidence required to adjudicate the same episode, not three independent policy events.
There are two distinct potential errors in interpreting the controversy. Counting each paper as an independent treatment realization overstates the evidence. Treating a disagreement between data sources as if it settled all wage-policy effects bypasses the question of what each sample represents. A successful adjudication requires comparable outcomes, consistent coding, sampling-frame information, uncertainty estimates that reflect the intervention level, and explicit consideration of the treatment and control economies.
Even a reliable positive point estimate does not automatically demonstrate monopsony: productivity, prices, demand, selection of surviving firms, and other adjustment channels may generate a similar pattern. It does challenge the empirical claim that all affected employers must reduce employment by a material amount when a binding minimum rises, assuming the design identifies the relevant effect. Mechanism-specific predictions have greater discriminating force than broad labels assigned after estimation.
Broader evidence and adjustment margins
Dube, Lester, and Reich (2010) compare contiguous counties on opposite sides of U.S. state boundaries, reporting no adverse employment effect in their study. This broadens the geographic comparison and makes shared local economic conditions relevant to the design. Administrative boundaries, however, do not eliminate spillovers, policy endogeneity, or cross-border commuting.
The low-wage jobs analysis of Cengiz et al. (2019) evaluates 138 state-level changes between 1979 and 2016. Its wage-distribution research design shifts attention toward the number of jobs in bins below and above the new wage floor, rather than an undifferentiated aggregate. Jobs near a wage threshold can disappear from one bin because their pay has risen into another bin. The validation question concerns the accounting of employment and wages across that distribution, with an explicit concern for the time horizon.
Germany supplies a different institutional setting. Dustmann et al.'s reallocation study (2022) examines the nationwide minimum wage introduced in 2015 and reports movement of workers toward larger, higher-paying and more productive establishments. Employment totals alone conceal such compositional changes. The finding gives economic relevance to models permitting worker and firm reallocation, alongside alternative mechanisms.
The sound comparative conclusion is conditional. The labor-market evidence constrains models imposing a universal substantial negative response across the studied regimes. It does not establish a common effect of every conceivable wage floor. The ratio of the floor to prevailing wages, the speed of adjustment, compliance, market power, affected sectors, price pass-through, and worker mobility should enter the empirical specification. The original competitive theorem identifies a meaningful channel; empirical evaluation identifies when the channel dominates the aggregate effect.
Choice under risk and calibration anomalies
Expected utility represents preferences over objective lotteries by probabilities multiplying utility of outcomes. Under the relevant axioms a lottery L with outcomes x_i and probabilities p_i is evaluated by Σ p_i u(x_i). The representation imposes invariance to certain common changes in lotteries. When preferences are interpreted as monetary final-wealth utility, concavity gives aversion to mean-preserving spreads under standard conditions. Descriptive use of the representation extends its domain from mathematical rational choice to actual behavior.
Independence and a constructed common-consequence test
Consider four illustrative lotteries. A pays 100 with certainty. B pays 500 with probability 0.10, 100 with probability 0.89, and zero with probability 0.01. C pays 100 with probability 0.11 and zero otherwise. D pays 500 with probability 0.10 and zero otherwise. Normalize u(0)=0. If A is preferred to B, expected utility requires 0.11u(100)>0.10u(500). If D is preferred to C, it requires the reverse strict inequality. No single such utility function represents these two strict choices jointly, under unchanged probability interpretation and consequence utility.
This is a constructed algebraic illustration of the common-consequence logic associated with the Allais problem. It does not report a sample frequency or reproduce an original experiment. To interpret an observed choice pattern as an empirical violation, the study must distinguish within-person responses from aggregate fractions, ensure comprehension of probabilities, and specify the consequences and elicitation mechanism. Different individuals can satisfy expected utility while an investigator compares unrelated aggregate majority choices. Individual-level evidence and the appropriate aggregation rule are essential.
Kahneman and Tversky's Prospect Theory (1979) develops a descriptive alternative in response to patterns of choice under risk. A reference-dependent account can accommodate patterns that a stable final-wealth expectation fails to reproduce. Scientific comparison also requires testing the additional flexibility of the alternative: reference points should be specified, probability weighting should be constrained, and predictions should be evaluated on choices excluded from estimation. Post hoc explanations are weak if the reference point can be set separately to account for every observation.
Calibration across stake sizes
Rabin's calibration theorem (2000) formalizes a different problem. If concave expected utility over total wealth consistently attributes appreciable aversion to small favorable gambles to wealth curvature across a suitable range of initial wealth, the required curvature can imply implausible risk aversion over larger gambles. The force of the argument lies in the joint restrictions across stake sizes and wealth levels. Rejection of one lottery by one decision maker at one wealth level does not establish those broad premises.
For an intuitive local approximation, let initial wealth be w and an additive gamble X have a small variance. With twice differentiable utility and small risks, the risk premium is approximately half the local absolute risk aversion −u''(w)/u'(w) times Var(X). A large premium for a fixed small risk requires correspondingly high local curvature. Extending that curvature across wealth states is an additional substantive claim. The theorem sharpens the cost of that extension beyond what a single Taylor approximation can establish.
Possible explanatory revisions include reference dependence, narrow framing, probability distortions, unobserved background risk, transaction costs, learning, or stochastic response. Each changes some aspect of the observable implications. A revision that assigns every unexplained choice to an idiosyncratic error makes a probability model highly flexible. It becomes testable when noise distributions, stability and prediction criteria are fixed and evaluated independently.
The normative implications are distinct. A descriptive violation of an axiom raises a question about the adequacy of a positive representation of behavior. A normative defense of the axiom as a criterion of consistency requires an independent argument. Empirical prevalence cannot by itself establish which preferences ought to govern public welfare evaluation. A scientific account can identify the descriptive failure while maintaining the separate status of the normative judgment.
Consumption and intertemporal restrictions
Intertemporal consumption models connect present choice to expected income, wealth, real returns, preferences and borrowing opportunities. In a standard interior model with an asset offering real gross return R_{t+1}, the Euler restriction is
This implication assumes an optimizing decision maker, a well-defined information set, appropriate asset access and interiority. Borrowing constraints may change equality into an inequality or make a different asset relevant. Aggregation across households may destroy a relation derived for the marginal utility of an individual.
Under quadratic utility, a constant certain return with βR=1 and the usual interior conditions, the Euler relation implies E_t[C_{t+1}]=C_t. Thus the innovation in consumption has conditional mean zero. The model does not say that consumption is constant over time; news can generate changes. It says that information in the specified conditioning set cannot systematically predict the innovation. The exact random-walk statement is a special-case restriction, not a general property of all lifetime optimization models.
Hall's stochastic consumption study (1978) examines such predictability implications of the life-cycle/permanent-income approach. The indexed results distinguish the predictive role of real disposable income from an index of stock prices. The mixed findings demonstrate the significance of specifying the conditioning information and the restricted empirical implication being evaluated.
Campbell and Mankiw's consumption and income study (1989) interprets aggregate time-series evidence using two groups whose consumption behavior differs, one responding to current income and another following a forward-looking specification. The mixed interpretation illustrates an alternative to a representative-consumer restriction. Aggregate excess sensitivity can be a consequence of heterogeneous constraints or behavior, but the aggregate relationship does not identify the populations or mechanisms by itself.
Suppose, illustratively, that 40 percent of an income-weighted aggregate consists of households consuming each additional dollar of a predictable income increment, and the remainder has zero immediate response under the simplified forward-looking restriction. A $100 predictable increase generates aggregate incremental consumption of $40 under those assumptions. The same aggregate response could be generated by all households spending forty cents per dollar. Distinguishing the two explanations requires distributional data, a design that distinguishes constrained from unconstrained households, or further predictions about transitions between states.
Consumption modeling also illustrates the importance of distinguishing aggregate growth from individual outcomes. A rise in a representative consumption series may reflect changes in population composition. The mean of marginal utilities need not equal marginal utility evaluated at mean consumption. The available aggregate moments impose useful restrictions on models of aggregation, but they do not license treating a representative agent as an empirically observed household.
Asset pricing and joint model restrictions
Asset pricing provides a particularly clear example of theoretical restrictions that can be confronted with moments of observable returns. In an arbitrage-free setting admitting a stochastic discount factor m, appropriately normalized traded gross returns R satisfy
A candidate consumption-based structural model provides a particular m through preferences and consumption growth, subject to market participation and other assumptions. Two assets with excess return X=R_i−R_f imply E[mX]=0 when both satisfy the pricing relation. Expanding the covariance yields
If the discount factor has positive mean, Cauchy–Schwarz gives the restriction
when σ(X)>0. A nonzero expected excess return requires sufficiently variable discount factors, regardless of whether the researcher favors a particular utility specification. The derivation is an elementary bound under stated moment conditions, not a novel theorem.
Mehra and Prescott's equity premium puzzle (1985) concerns restrictions on equity and short-term safe returns imposed by a class of equilibrium models, compared with U.S. data from 1889 to 1978. Their empirical discrepancy concerns jointly attainable asset-return moments, rather than a claim that a particular stock earns a high return in every year. The model's implication depends on preference parameters and the process for aggregate consumption, together with assumptions about the representative consumer and the asset payoffs.
Hansen and Jagannathan's discount-factor bounds (1991) use financial returns to constrain admissible discount-factor moments. To illustrate the logic with constructed quantities, let the mean annual excess return be 0.06 and its standard deviation be 0.20. The ratio is 0.30. A proposed discount factor with coefficient of variation 0.10 cannot satisfy this moment restriction for the stated population moments. With a finite historical sample, uncertainty in both estimates and their dependence must be included before announcing rejection.
A common revision is to change the process determining the discount factor by admitting long-run risks, rare disasters, habit formation, heterogeneous participation or market incompleteness. Such changes may permit a larger price of risk, yet the scientific evaluation requires additional implications. A model fitting the mean equity premium while implying implausible safe rates, consumption responses, wealth distributions or price-dividend behavior still faces joint restrictions. An added parameter can relax a discrepancy and simultaneously create new empirically testable relations.
The asset-pricing case clarifies why multiple outcomes are especially valuable. A free parameter chosen to match one mean may remove one constraint without improving independent moments. The set of jointly attainable moments is the meaningful comparison. Researchers should distinguish unconditional time-series means from conditional expectations and avoid interpreting realized returns as direct observations of ex ante premia. A long sample covers many economic regimes and can remain statistically limited for rare risks.
Policy regimes and structural invariance
Economic agents respond to institutions, prices, budgets, information and expectations. The parameters of a reduced-form relationship estimated under one policy regime may summarize the equilibrium behavior induced by that regime. Policy evaluation therefore requires identifying which relations survive a proposed change and which reflect old equilibrium choices.
Lucas's policy evaluation critique (1976) directs attention to changes in behavioral relationships caused by changes in policy rules. A Federal Reserve research record discussing rational expectations and econometric policy evaluation corroborates the role of the expectation mechanism. The argument applies when agents' decision rules depend on the policy environment. It is a conditional demand for structural invariance, rather than a claim that every statistical parameter changes under every intervention.
An illustrative two-equation system makes this issue explicit. Let private choice x_t=a E_t[p_{t+1}]+u_t, where the expectation depends on a rule p_{t+1}=ρp_t+ε_{t+1} and mean-zero future innovations. Under information sufficient to observe p_t, the observable relation is x_t=aρp_t+u_t. A regression coefficient estimating aρ under ρ=0.8 may predict the old regime accurately. Changing ρ to 0.2 changes the coefficient if a remains fixed. If a=2 the old coefficient is 1.6 and the new one 0.4. Using 1.6 for the new regime entails a fourfold slope error even when the structural response a is unchanged.
The example identifies the counterexample to a particular invariance assumption. It does not prove that the expectation equation is correct in any actual economy. Alternative learning processes, imperfect information, institutional lags or a change in a itself can produce different predictions. An empirical test would require variation in the policy rule, measurements of expectations where possible, and evidence about the time horizon over which adjustment occurs.
Structural models attempt to explain policy-dependent behavior using deeper preferences, technology, constraints and information. Their parameters can also depend on institutional changes or on the population being analyzed. A tax reform can alter occupational selection, innovation, compliance or household formation. Assuming that all relevant preference and technology parameters are invariant is a substantive hypothesis; identifying which ones are stable is part of the research.
Forecast evaluation within a regime and counterfactual evaluation across regimes should be documented separately. A model validated for routine short-run predictions under a central bank's established operating procedure has earned evidence for that purpose. Its credibility under a new policy framework requires additional validation. Conversely a structural model with credible local policy responses may need a broader forecasting system for disturbances it intentionally abstracts from.
Prediction and model comparison
Models compete for specific tasks under specific evaluation criteria. For predictions ŷ_{i,t} of outcome y_t, a loss function L(y_t,ŷ_{i,t}) defines the cost of errors. A squared-error score emphasizes large deviations; absolute error reduces their relative influence; asymmetric losses can express unequal costs of overshooting and undershooting. The scientific comparison is conditional on the declared loss and the information available at the forecast origin.
Genuine out-of-sample evidence
An in-sample fit rewards adaptation to the training observations. Out-of-sample evaluation separates model estimation from subsequent measurement of errors. For time series, rolling or expanding training windows with a chronological evaluation period maintain the information order. Randomly shuffling observations can leak future information. For cross-sectional applications, splitting individuals who share a household, establishment, or institutional shock may create a similar leakage problem.
Suppose three out-of-sample realization values are 10, 12 and 8. Models A and B predict (9,11,9) and (10,9,8). Under squared loss A produces (1,1,1) with average 1, while B produces (0,9,0) with average 3. Under absolute loss A produces (1,1,1) and B produces (0,3,0), both averaging 1. A ranking under one loss does not necessarily persist under another. Three errors are used here for arithmetic illustration, not inference.
The predictive accuracy comparison developed by Diebold and Mariano (1994 working paper; 1995 publication) treats the mean loss differential as the object of inference. For d_t=L_A,t−L_B,t, a zero expected differential is a testable claim; time dependence affects its sampling variance. Overlapping multi-period forecasts, recursively estimated parameters and small samples require attention to the appropriate reference distribution. Comparing nested models can create additional nonstandard inference issues.
Giacomini and White's conditional predictive ability tests (2006) address the evaluation of forecasting procedures in settings where misspecification may persist. Conditional comparison matters when models perform differently in distinct information states. A model can have equal unconditional mean loss to its rival and still produce systematically worse forecasts in recessions or in low-liquidity conditions. A validation report should distinguish average performance, regime-specific loss, and performance relevant to intended decisions.
Benchmarks and instability
A benchmark should be a credible alternative with the same forecast information, horizon and target. If seasonal patterns account for much of a data series, comparing a sophisticated model with an intentionally poor constant forecast overstates the evidence. Useful baselines include persistence, seasonal persistence, simple linear predictors, or institutional forecast procedures appropriate to the target. All competitors must receive equivalent information access.
Prediction intervals and density forecasts require evaluation of calibration and sharpness. A nominal 90 percent interval should cover outcomes at the appropriate rate under the prediction law and relevant dependence structure. A very wide interval may attain coverage while providing weak decision support. Proper scoring rules can reward a density forecast that assigns probability to realized outcomes without obtaining artificial success by indiscriminate spreading. Calibration across relevant states, rather than an average alone, matters to risk-sensitive applications.
Following a structural break, historical cross-validation may be a poor guide to a new environment. The response is to specify candidate change points or state variables on economic grounds, use appropriately timed holdouts, and examine model errors before and after change. Retuning after every failed forecast may improve fit but obstruct falsifiable evaluation unless future data are retained for a new test. Transparent versioning separates the model that generated a forecast from the model revised in response to its failure.
Replication and robustness
A reproducibility exercise seeks to regenerate a result using the same source data, code and computational procedures. A replication collects or uses distinct observations to investigate the same substantive claim. A robustness exercise tests sensitivity to alternative analytically defensible decisions about samples, variables, designs, priors or model forms. These exercises answer related questions and should be named precisely.
Camerer et al.'s economics replication study (2016) examined eighteen laboratory experiments from two prominent journals. Eleven of the eighteen replications produced a significant effect in the same direction as the original; the average relative effect size was reported as 66 percent. Eleven divided by eighteen equals approximately 61.1 percent. The first statistic evaluates the study's stated replication criterion; the second concerns effect magnitudes. Neither identifies the replication probability of every published economic paper. The sample and the criterion delimit the result.
A replication result should be evaluated by magnitude, uncertainty, design comparability, implementation fidelity and whether it addresses the same estimand. Two studies may yield different significance decisions even if their estimates are mutually compatible. Conversely, two significant positive estimates may differ enough in magnitude to have distinct policy implications. A direct test of the difference, an uncertainty-aware synthesis, or an explicit equivalence margin makes the comparison sharper.
Robustness is strongest when alternative specifications are motivated by the problem. A researcher can compare outcome measurements, account for clustering, include a justified competing trend, or use another plausibly exogenous source of treatment variation. Running an enormous number of arbitrary specifications and reporting a fraction with a desired sign has limited value without evidence about which specifications identify the intended quantity. Validity of the design cannot be replaced by numerical majority voting.
Prespecified hypotheses and analysis plans, disclosure of exploratory analyses, complete variable definitions, observable data-processing decisions and versioned code permit more independent evaluation. Restricted administrative data may prevent open distribution of raw records; documented access conditions and synthetic test fixtures can still clarify the computational procedure, while independent reproduction awaits authorized data access. A synthetic example verifies arithmetic or implementation behavior, not a reported empirical estimate.
An important distinction concerns replicating a failure. Independent confirmation that a model mispredicts a particular moment gives greater credibility to the anomaly. Independent testing of a proposed repair on a different outcome gives greater credibility to the repair. These are different experimental achievements. A repair that succeeds only on the data used to motivate it preserves the need for prospective evaluation.
Institutional transport and comparative evidence
Economic regularities arise in environments shaped by technology, market structure, legal rules, social practices, policy instruments and enforcement. An empirical prediction established under one configuration can travel to another when the relevant mechanisms, support conditions and parameter relationships are sufficiently similar or their changes can be accounted for. External validity is therefore an empirical and theoretical question about environments.
A simple transport calculation illustrates the problem. Suppose a treatment effect is 4 for group A and −2 for group B. A study population containing 75 percent A and 25 percent B has average effect 2.5. A target population containing 25 percent A and 75 percent B has average effect −0.5. Both averages follow from the same group-specific effects with different composition. The change of sign does not refute group-level invariance. It refutes a claim of universal invariance of the unadjusted aggregate. Reweighting is justified only when the relevant groups are observed, target support overlaps, and conditional effects actually transport.
The wage-policy evidence offers a practical application. The Pennsylvania comparison, later U.S. border comparisons, wage-distribution estimates and the German nationwide reform involve differences in treatment scale, labor-market institutions, outcomes, sampling and mobility. Their estimates cannot be pooled as if they were repeated draws from the same policy experiment. They can nevertheless evaluate models predicting similarities across these environments, especially models that explicitly incorporate the affected wage distribution and adjustment margins.
Institutional comparison should use the legally effective treatment, rather than a label alone. A statutory floor can vary in exemptions, enforcement, sectoral coverage and compliance. A nominal tax rate can differ from a marginal effective rate because of thresholds and interactions with benefits. A monetary policy announcement can arise under different credibility and expectation conditions. Each example shows the importance of establishing the operational exposure and the incentive margin named by the model.
Transport can also fail because an intervention becomes large enough to change equilibrium prices and resource allocation. A local program with available capacity and negligible market feedback may have different consequences at national scale. Spillovers, congestion, firm entry, changes in tax financing and interactions with other institutions can modify the original mechanism. Scaling evidence needs variation in treatment saturation, structural accounting, or credible assumptions about equilibrium adjustment.
The comparative method is most informative when it records both invariant and varying components. If labor-demand curvature remains stable across settings but labor-supply elasticity differs, a change in aggregate employment response may be expected under one common structural account. If a model attributes the result to a parameter that changes only after observing each separate failure, its transport claim weakens. A useful model specifies which changes are predicted and which parameters need independent re-estimation.
Analytical synthesis and open research questions
Across the historical cases, a common analytical pattern emerges. The model imposes restrictions on quantities that are accessible through a particular measurement or experimental procedure. The observation challenges either the restriction itself or the applicability of the conditions that generated it. Scientific progress occurs when additional evidence separates these possibilities and constrains the available responses. This is a synthesis of established inferential practice, rather than a proposal of a new named doctrine.
Localizing failures
A scientific report can identify several locations for a discrepancy. The mapping from theoretical quantities to observed data may be inaccurate. The comparison group or assignment mechanism may fail to identify the target. The parameters estimated for one group or period may not describe another. The mechanism itself may be inadequate. An alternative substantive explanation may imply the observed phenomenon while the original cannot accommodate it under admissible parameter values. These locations are distinguished empirically, not merely by giving them different names.
The minimum wage controversy shows the importance of measurement and causal identification before mechanism inference. The constructed lottery example shows that certain preference patterns can create a direct algebraic incompatibility once the relevance of the axioms is established. Consumption analysis highlights joint dependence on information, aggregation and constraints. Asset pricing supplies cross-equation moment restrictions that limit parameter rescue. Policy analysis emphasizes the dependence of reduced-form relations on regimes. Predictive testing supplies a task-specific benchmark and a distinct way to evaluate revised procedures.
| Case | Restriction under evaluation | Evidence that sharpens adjudication |
|---|---|---|
| Labor-market wage intervention | Sign and magnitude of employment response under stated conditions | Comparison trends, payroll and survey crosswalk, hours, wage bins, reallocation |
| Decision under risk | Compatible rankings across linked lotteries | Within-person choices, prespecified consequences, elicitation and independent stakes |
| Consumption | Conditional predictability and Euler restrictions | Information vintages, household constraints, asset participation and aggregation |
| Asset pricing | Jointly feasible return and discount-factor moments | Multiple returns, consumption measures, conditional moments, independent observations |
| Policy regime | Invariance of a response relation | Rule changes, expectation evidence, timing and heterogeneous adjustment |
| Prediction | Equal or bounded loss relative to a comparator | Chronological holdout, common information, dependence-aware uncertainty |
The author's substantive interpretation is that the evidential force of a counterexample depends on what remains constant across the comparison. A design that changes only a relevant treatment margin gives sharper information than one that changes the outcome definition, population and intervention simultaneously. A restriction that remains inconsistent after independent measurements and justified recalibration faces stronger pressure than a discrepancy resolved by a verified coding difference. An empirically credible repair accounts for the original discrepancy while also making a successful prediction beyond the information used to design it.
Revision and retained scientific value
A model can remain useful after a claim of unrestricted generality is contradicted. The appropriate action is to narrow its application domain, change independently refutable assumptions, introduce a measured mechanism, or replace the model for the relevant task. Domain restriction should carry an economic explanation and observable scope conditions. The statement “the model works when the model works” leaves the relevant domain unidentified.
Consider a family of wage models that permits both competitive and imperfectly competitive employment responses. If the degree of employer market power is measured independently, the family can predict where responses will differ. If that degree is selected separately after every observed outcome, the family accommodates both signs without discriminating power. Adding mechanisms increases explanatory reach only in conjunction with independent restrictions or successful new predictions.
A counterexample can therefore increase scientific knowledge even when the original abstraction is retained. A competitive theorem still explains one causal channel. The empirical dispute establishes the conditions under which additional channels matter for application. Validation is an accumulation of restricted and transferable warrants, with each warrant tied to an intended task and evidentiary basis.
Research practice and unresolved questions
Future research should improve the connection between measurement models and economic estimands. Administrative and survey records frequently encode different units, time periods and institutional definitions. Studies designed to compare those instruments under common conditions could help identify whether anomalies originate in measurement or behavior.
Another research question concerns mechanism discrimination under heterogeneous treatment effects. The identified effect of a policy within a particular population may be compatible with several structural explanations. Additional independent experiments, saturation designs and linked intermediate outcomes can distinguish competing accounts, while preserving the distinction between treatment effect estimation and mediation.
A third question concerns strategic responses to scientific measurement and policy. Economic agents can adapt to eligibility thresholds, published performance measures or expectations about future rules. If the act of defining an observable changes behavior, repeated historical measurement may conceal instability relevant to intervention. The combination of institutional detail, real-time evidence and transparent model updating becomes essential.
Finally, uncertainty about the data-generating process is itself relevant to decision support. A single estimated model may give a narrow prediction interval conditional on its own assumptions while alternative credible models imply widely different outcomes. Reporting this sensitivity is particularly important when extrapolating beyond historical support. Model comparison should preserve differences in objective, population and scope while identifying results robust across the admissible alternatives.
Conclusion
The scientific purpose of empirical counterexamples is to determine what economic models have successfully explained, which predictions fail within justified domains, and what additional evidence can discriminate among candidate revisions. Mathematical validity, measurement adequacy, causal identification, predictive accuracy and institutional transport constitute different dimensions of evaluation. Treating them as interchangeable obscures both the successes and failures of economic theory.
The studies reviewed here show why economic model validation requires explicit rather than implicit empirical commitments. The minimum wage literature challenges broad employment predictions while exposing measurement and identification issues. Risk-choice anomalies constrain stable expected-utility representations under specific behavioral assumptions. Consumption and asset-pricing studies demonstrate the informational power of joint restrictions. Policy-regime analysis reveals the distinction between historical fit and intervention invariance. Replication and out-of-sample comparisons extend evaluation to independent observations and prespecified losses.
The monograph's analytical conclusion is that empirical evidence becomes most informative when it identifies the specific restriction under pressure, measures the relevant quantities credibly, and distinguishes a model revision from the observations that prompted it. Validated economic models earn conditional authority for stated tasks and environments. A scientifically defensible revision preserves successful restricted implications and exposes its new commitments to further evidence. This conception supports both rigorous criticism and cumulative economic explanation without converting a body of methodological judgments into an asserted new theoretical system.
References
- Andrews, Isaiah, James H. Stock, and Liyang Sun. (2019). Weak Instruments in Instrumental Variables Regression: Theory and Practice. Annual Review of Economics 11, 727–753. DOI: 10.1146/annurev-economics-080218-025643. Source record.
- Angrist, Joshua D., Guido W. Imbens, and Donald B. Rubin. (1996). Identification of Causal Effects Using Instrumental Variables. Journal of the American Statistical Association 91(434), 444–455. DOI: 10.1080/01621459.1996.10476902. Source record.
- Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan. (2004). How Much Should We Trust Differences-in-Differences Estimates? Quarterly Journal of Economics 119(1), 249–275. DOI: 10.1162/003355304772839588. Source record.
- Callaway, Brantly, and Pedro H. C. Sant’Anna. (2021). Difference-in-Differences with Multiple Time Periods. Journal of Econometrics 225(2), 200–230. DOI: 10.1016/j.jeconom.2020.12.001. Source record.
- Camerer, Colin F., Anna Dreber, Eskil Forsell, Teck-Hua Ho, Jürgen Huber, Magnus Johannesson, Michael Kirchler, Johan Almenberg, Adam Altmejd, Taizan Chan, Emma Heikensten, Felix Holzmeister, Taisuke Imai, Siri Isaksson, Gideon Nave, Thomas Pfeiffer, Michael Razen, and Hang Wu. (2016). Evaluating Replicability of Laboratory Experiments in Economics. Science 351(6280), 1433–1436. DOI: 10.1126/science.aaf0918. Source record.
- Campbell, John Y., and N. Gregory Mankiw. (1989). Consumption, Income, and Interest Rates: Reinterpreting the Time Series Evidence. NBER Macroeconomics Annual 4, 185–246. MIT Press. DOI: 10.1086/654107. Source record.
- Card, David, and Alan B. Krueger. (1993). Minimum Wages and Employment: A Case Study of the Fast Food Industry in New Jersey and Pennsylvania. NBER Working Paper 4509. https://doi.org/10.3386/w4509. Published in American Economic Review 84(4), 772–793 (1994). Source record.
- Card, David, and Alan B. Krueger. (1998). A Reanalysis of the Effect of the New Jersey Minimum Wage Increase on the Fast-Food Industry with Representative Payroll Data. NBER Working Paper 6386. https://doi.org/10.3386/w6386. Published as Reply in American Economic Review 90(5), 1397–1420 (2000). Source record.
- Cengiz, Doruk, Arindrajit Dube, Attila Lindner, and Ben Zipperer. (2019). The Effect of Minimum Wages on Low-Wage Jobs. Quarterly Journal of Economics 134(3), 1405–1454. DOI: 10.1093/qje/qjz014. Source record.
- Deaton, Angus, and Nancy Cartwright. (2018). Understanding and Misunderstanding Randomized Controlled Trials. Social Science & Medicine 210, 2–21. DOI: 10.1016/j.socscimed.2017.12.005. Source record.
- Diebold, Francis X., and Roberto S. Mariano. (1994). Comparing Predictive Accuracy. NBER Technical Working Paper 169. DOI: 10.3386/t0169. Published in Journal of Business & Economic Statistics 13(3), 253–263 (1995). Source record.
- Dube, Arindrajit, T. William Lester, and Michael Reich. (2010). Minimum Wage Effects Across State Borders: Estimates Using Contiguous Counties. Review of Economics and Statistics 92(4), 945–964. DOI: 10.1162/rest_a_00039. Source record.
- Dustmann, Christian, Attila Lindner, Uta Schönberg, Matthias Umkehrer, and Philipp vom Berge. (2022). Reallocation Effects of the Minimum Wage. Quarterly Journal of Economics 137(1), 267–328. DOI: 10.1093/qje/qjab028. Source record.
- Federal Reserve Bank of Minneapolis. Rational Expectations and Policy Evaluation in Macroeconometric Models. Working paper research record, accessed November 17, 2026. Source record.
- Giacomini, Raffaella, and Halbert White. (2006). Tests of Conditional Predictive Ability. Econometrica 74(6), 1545–1578. DOI: 10.1111/j.1468-0262.2006.00718.x. Source record.
- Hall, Robert E. (1978). Stochastic Implications of the Life Cycle-Permanent Income Hypothesis: Theory and Evidence. Journal of Political Economy 86(6), 971–987. DOI: 10.1086/260724. Source record.
- Hansen, Lars Peter, and Ravi Jagannathan. (1991). Implications of Security Market Data for Models of Dynamic Economies. Journal of Political Economy 99(2), 225–262. DOI: 10.1086/261749. Source record.
- Hansen, Lars Peter. (1982). Large Sample Properties of Generalized Method of Moments Estimators. Econometrica 50(4), 1029–1054. DOI: 10.2307/1912775. Source record.
- Hausman, Daniel M. (2018). Philosophy of Economics. The Stanford Encyclopedia of Philosophy, substantive revision September 4, 2018. Stanford University. Source record.
- Kahneman, Daniel, and Amos Tversky. (1979). Prospect Theory: An Analysis of Decision under Risk. Econometrica 47(2), 263–291. DOI: 10.2307/1914185. Source record.
- Lucas, Robert E., Jr. (1976). Econometric Policy Evaluation: A Critique. Carnegie-Rochester Conference Series on Public Policy 1, 19–46. DOI: 10.1016/S0167-2231(76)80003-6. Source record.
- Mehra, Rajnish, and Edward C. Prescott. (1985). The Equity Premium: A Puzzle. Journal of Monetary Economics 15(2), 145–161. DOI: 10.1016/0304-3932(85)90061-3. Source record.
- Neumark, David, and William Wascher. (1995). The Effect of New Jersey's Minimum Wage Increase on Fast-Food Employment: A Re-Evaluation Using Payroll Records. NBER Working Paper 5224. https://doi.org/10.3386/w5224. Published as Comment in American Economic Review 90(5), 1362–1396 (2000). Source record.
- Rabin, Matthew. (2000). Risk Aversion and Expected-Utility Theory: A Calibration Theorem. Econometrica 68(5), 1281–1292. DOI: 10.1111/1468-0262.00158. Institutional repository version. Source record.