Random Choice and Market Demand Javier A. Birchenall∗ University of California at Santa Barbara September 10, 2016 Abstract This paper presents comparative statics predictions for a statistical model of choice in which consumption is randomly chosen subject to a linear budget constraint. Market demands are shown to satisfy all the properties of the behavioral model of choice, including symmetry and negative semidefiniteness of the Slutsky matrix. The findings characterize market demands for random choice behaviors that abandon all behavioral postulates of rational choice theory but also for realistic individual choice behaviors representable by choice probabilities. The findings also serve to demonstrate the inconsistency of statistical tests of revealed preferences that rely on rationality indices. Keywords : random choice, market behavior, demand theory JEL classification: D01; D11; C02 Communications to: Javier A. Birchenall Department of Economics, 2127 North Hall University of California, Santa Barbara CA 93106 Phone/fax: (805) 893-5275 [email protected] ∗ I am grateful to audiences at several locations for comments and helpful suggestions. I am specially grateful to Geir Asheim, Ted Bergstrom, Gary Charness, Luis Corchón, John Duffy, Martin Dufwenberg, Enrique Fatas, Raya Feldman, Eric Fisher, Rod Garratt, Zack Grossman, John Hartman, David Hinkley, Marek Kapicka, Tee Kilenthong, Natalia Kovrijnykh, John Ledyard, Steve LeRoy, Gustavo Ponce, ChengZhong Qin, Warren Sanderson, Perry Shapiro, Joel Sobel, Glen Weyl, Eduardo Zambrano, and Giulio Zanella for useful conversations. The usual disclaimer applies. 1 Introduction Modern economic theory is founded on the assumption of rational individual behavior. In the behavioral model of choice, consumer demand functions are derived from a rational preference relationship and maximizing behavior. Demand functions, however, are not exclusive to the behavioral model of choice. Gary Becker examined a statistical model of choice that dispenses with any kind of preferences and maximizing behavior. In Becker [9], individual consumption is randomly (i.e., “irrationally”) chosen subject to a linear budget set. Yet, compensated price changes force individuals to, on average, satisfy the Law of Demand. Becker’s [9] example, while an alternative to the behavioral model of choice, is impractical and very restrictive: it relies on two commodities, random choices drawn from a uniform distribution, and non-satiated demands.1 It is not obvious, for instance, that one would retain any predictive power without these assumptions or that these predictions would apply to random choice procedures of interest to behavioral scientists.2 This paper presents comparative statics predictions for a statistical model of choice based on a general probability distribution function, an arbitrary number of commodities, and for demands that lie in the interior or in the boundary of the budget set. I demonstrate that market demands satisfy the compensated Law of Demand quite generally. Indeed, I show that market demand functions satisfy all the properties of the behavioral model of choice including symmetry and negative semidefiniteness of the Slutsky matrix. This last property is the defining feature of the behavioral model. Market demands in the statistical model of choice are therefore shown to be indistinguishable from those derived from a behavioral model of choice. 1 Hildenbrand ([36], p. 34) remarked “I do not know of any successful alternative for modelling the dependence of [demand] on [prices]. There is, of course, the well-known example of Becker.” He them proceeds to point out some of the crucial limitations of Becker [9]; particularly, the use of uniform distributions. Varian ([63], p. 105) similarly remarked “there seem to be few alternative hypotheses other than Becker’s that can be applied using the same sorts of data used for revealed preference analysis.” 2 The uniform distribution is special; a famous example is Caplin and Spulber [14] where monetary neutrality follows entirely due to their use of uniform distributions. Integrability in two-commodities is also special as it is easily resolved; see Katzner ([38], Theorem 4.1-2). Finally, nonsatiation, e.g., “more is preferred to less,” is responsible for downward sloping indifference curves, for the direction of increasing utility in the indifference map, and for homogeneity and adding up conditions in behavioral models. Demsetz ([20], p. 489) was indeed agnostic about Becker’s [9] findings: “as imaginative and informative as [Becker’s] arguments are, they cannot be sustained in full without appealing to rationality.” 1 I illustrate the theory’s value-added using a number of applications. I place no restrictions on the choice probabilities beyond those implied by probability theory. Thus the comparative statics derived here characterize, not just a number of “irrational” behaviors, but any individual choice behavior representable by a continuous probability distribution over the commodity space. Decision theorists have in effect formulated numerous realistic (i.e., descriptive) random choice models based on axiomatic rules or choice heuristics. Naturally their focus is on individual choice behavior not on market behavior; see, e.g., Gonzalez-Vallejo [32], Luce [43], Roe et al. [54], and Tversky ([59], [60]). This paper finds that these realistic individual choice behaviors share the market predictions of the standard behavioral model of choice.3 The statistical model of choice has been primarily used to derive the statistical power for revealed preference tests that rely on rationality indices; see, e.g., Varian [62] and Echenique et al. [22].4 (Apesteguia and Ballester [5] and Dean and Martin [19] recently derived additional rationality indices.) Quantifying statistical power requires an alternative hypothesis in case the behavioral postulates are incorrect. The statistical model of choice provides an alternative hypothesis to measure statistical power. Power calculations have employed the uniform distribution, the logistic distribution, and random perturbations from observed choices; see, e.g., Bronars [12], Choi et al. [17], and Andreoni et al. [4].5 This paper shows that revealed preference tests are inconsistent and that their inconsistency is not sensitive to the choice of distribution function. For instance, tests based on the amount of money that could be pumped out of non-rational individuals have no power to detect 3 A recent literature in economics studies sophisticated random choices but also at the individual level and not at the market level; see, e.g., Fudenberg et al. [28], Gul and Pesendorfer [34], and Manzini and Mariotti [46]. An early exception that focuses on market behavior is Mossin’s [53] characterization of mean demands in Luce’s [43] model of choice. 4 Tests of individual rationality abound in economics and they are likely to become more prevalent as detailed individual purchase data becomes more accessible. Revealed preference tests have even been applied to nonstandard populations such as animal subjects (Kagel et al. [37] and Chen et al. [16]), psychiatric patients (Cox [18]), children (Harbaugh et al. [35]), and individuals under the influence of alcohol (Burghart et al. [13]) to name a few. 5 Varian ([61], [62], and [63]) provide an authoritative overview of revealed preference theory. Bronars [12] first applied Becker’s [9] model to derive statistical power. Choi et al. [17] considered boundedly rational individuals that follow Luce’s [43] choice axioms. Andreoni et al. [4] considered numerous alternatives. Beatty and Crawford [8] discuss a separate concern based on nonstatistical aspects of rejectability in revealed preference tests. Essentially, if budget sets do not intersect, violations of rationality cannot be detected even in deterministic settings. This concern will not be present here. 2 individual rationality. Additional applications of statistical models of choice are very specialized. Grandmont [33], in the spirit of Becker [9], derived through aggregation market demands of the CobbDouglas type; see also Hildenbrand [36] and Kneip [40]. These demands are consistent with positive income effects and hence ensure equilibrium uniqueness in general equilibrium. I characterize income effects here and show that positive income effects are not general to the statistical model of choice. Gode and Sunder ([29], [30]), in another application, used a double auction mechanism to document high allocative efficiency of markets under statistical choices; see also Duffy [21]. This literature relies extensively but unnecessarily on the uniform distribution used by Becker [9]. Market interactions turn out to serve as a partial substitute for individual rationality in far more general settings than those previously analyzed. This paper also substantiates some of Becker’s [9] remarks about rational economic behavior. Statistical choices, for example, make some of the criticisms of the behavioral assumptions of economics less forceful. This paper shows that the consistency of preferences and maximizing behavior, often refuted by psychologists and experimental economists, are unnecessary for demand theory as its main predictions can be derived using a completely different route. The findings also attest to the independence between positive and normative assessments in economics. In the behavioral model, positive and normative assessments are always intertwined leading to a number of difficulties in the presence of unstable preferences or inconsistent behavior. Choices in a statistical model have no normative content; welfare can be measured (i.e., by the consumer’s surplus) but it has no significance or practical value. Finally, the findings demonstrate by example that the intensive understanding of individual decision-making does not necessarily hold the key to the understanding of aggregate behavior. Statistical choices prompt all sorts of violations of individual rationality; even in a general stochastic rationality sense. Aggregate behavior, however, is consistent with the existence of a rational “representative agent.” A statistical model of choice is more general than random preference models (i.e., models in which individuals maximize randomly chosen preferences). To guarantee that choices can be rationalized by a random preference model, observed choice probabilities must satisfy 3 the axiom of revealed stochastic preference; see McFadden and Richter [50].6 The choice procedure here is entirely statistical so it could, in principle, dispense with any behavioral assumption. In fact, random choices here are not necessarily stochastically rational.7 The paper unfolds as follows. Section 2 generalizes Becker’s [9] example; Section 3 presents the general theorems; Section 4 lists a few examples of random choice models; Section 5 discusses the power of revealed preference tests; and Section 6 concludes. 2 A simple example I first generalize the two-commodity example of Becker [9]. The two commodities are denoted 1 and 2 . Prices = (1 2 ) and income are given and are assumed to be strictly positive. There is a continuum of individuals and their choices must lie in a linear budget set © ª ( ) ≡ (1 2 ) ∈ R2+ : 0 ≤ 1 1 + 2 2 ≤ . (1) The model of choice is statistical: 1 is a random variable with a well-behaved distribution function (1 ) defined over 1 ∈ R+ ; the density function is (1 ) 0 for all 1 ∈ R+ . For technical reasons (i.e., differentiability), I assume that all random variables are absolutely continuous and that their first and second moments are finite. Analytical problems concerning results that hold on sets of measure zero are neglected throughout and all subsets are assumed non-empty and measurable. Throughout the paper, I also maintain the requirement of consumer sovereignty: I assume that (1 ) is invariant with respect to ( ). This assumption is the analog of assuming that tastes are invariant to the economic environment. As in the behavioral 6 The literature that studies random preference models places a special emphasis on econometric applications; see, e.g., Anderson et al. [3]. These models typically rely on an indirect utility function that varies with prices and income. Individuals are subject to optimization errors in the form of a random utility component. This paper is direct with respect to prices and income, and there is no need to assume consistent preferences or maximizing behavior at any stage of the decision process. “Irrational” choices tough admit a random utility representation; see Section 3. 7 McFadden and Richter [50] provides a definite treatment of revealed stochastic preference theory; see also Bandyopadhyay et al. [7], Falmagne [25], Lewbel [42], and McFadden [49]. In the approach taken here, preferences or maximizing behavior are not needed at any stage of the decision process. I also focus on market demands and not on observed choice probabilities, which are motivated by the discreteness of applied random preference models. 4 model of choice, in which preferences are given and independent of the budget set, the distribution function (1 ) is given and does not vary with ( ). Obviously, observed choices will depend on ( ) because the budget constraint limits choices. Indeed, not all statistical choices are feasible. Due to (1), the maximum quantity of 1 that can be bought is X max ≡ 1 . Choices of 1 are thus subject to 0 ≤ 1 ≤ X max 1 1 . Mean demand for 1 is given by a right-truncated mean formula: ̄1 ( ) ≡ E [1 | 0 ≤ 1 ≤ X max 1 ] = Z max 1 1 0 (1 ) 1 , (X max 1 ) (2) max where (X max 1 ) ≡ Pr{1 : 0 ≤ 1 ≤ X 1 }. As in Becker [9], once 1 is chosen, the remaining income is spent in 2 , i.e., 2 satisfies 2 2 = − 1 1 as long as 0 ≤ 1 ≤ X max 1 . Its mean demand is ̄2 ( ) = − ̄1 ( )1 . 2 (3) A random choice procedure consistent with interior demands is discussed later on. The “Law of Demand.” It is obvious but neither a Hicksian compensation nor standard duality results are possible here because demands do not rely on utility functions. The analysis is carried out in terms of Slutsky compensated demands. I first show that the compensated Law of Demand holds, ¯ ̄1 ( )¯¯ 0, 1 ¯ (4) where ·| denotes the standard Slutsky compensation, = ̄1 ( )1 .8 The relevant derivative needed to evaluate (4) is: 8 ¯ ¯ ¶ µ ¯ ̄1 ( )¯¯ X max E [1 | 0 ≤ 1 ≤ X max 1 ] 1 ¯ . = max ¯ 1 X 1 1 ¯ (5) When 1 changes to 1 +1 mean demand changes to ̄1 (1 +1 2 +), where = ̄1 ( )1 is the Slutsky compensation. Every individual is a set of measure zero so individual demands are not well defined. Income compensations, however, can be thought out as given to every feasible realization of demand, i.e., the truncated mean of 1 1 is ̄1 ( )1 . This means that individuals are not compensated so that everyone can afford mean demands but so that individuals can afford their original bundle. 5 The first term represents the effect of changes in X max on mean demands and the second 1 the effect of a compensated change in 1 on X max 1 . This first term satisfies (X max E [1 | 0 ≤ 1 ≤ X max 1 ] 1 ) = − ̄1 ( )) 0, (X max 1 max X 1 (X max ) 1 (6) by a standard application of Leibniz’s rule for differentiation under the integral sign in (2). The second term in (5) satisfies ¯ ¯ X max − ̄1 ( ) X max 1 ¯ 1 = − 0. 1 ¯ 1 (7) Expressions (5), (6) and (7) yield the following proposition: Proposition 1 Assume that individuals randomly choose 1 and determine 2 as a residual from (1). For any well-behaved continuous probability distribution function (1 ), the compensated Law of Demand holds for ̄1 ( ) and ̄2 ( ). The (compensated) Law of Demand holds due to a statistical fact and an economic fact. First, the truncated mean ̄1 ( ) is less than the truncation point X max 1 , and its value increases as the truncation point increases; see (6). Second, even after taking into account the income compensation, the maximum amount of 1 that can be purchased declines as 1 increases; see (7). Figure 1 illustrates Proposition 1. As the figure shows, some values of 1 are not feasible max so (1 ) must be truncated at X max 1 ; that is, the dark-shaded area to the right of X 1 is not feasible at ( ). When 1 increases, and income is compensated, the truncation max 00 00 point changes to X max X max is no longer 1 1 ; the light-shaded area to the right of X 1 feasible. A compensated increase in 1 thus reduces X max and the mean demand declines 1 from ̄1 ≡ ̄1 (1 2 ) to ̄001 ≡ ̄1 (01 2 0 ).9 9 The only generalizations of Becker [9] I have been able to find are in Sanderson ([55], [56]). These papers noted that an increase in 1 expands the choice set in a first-order stochastic sense. Sanderson ([55], [56]), however, restrict the distribution of consumption opportunities in ways that are not actually needed. Since the relevant choice distribution is (1 ) (X max ), the orderings assumed in Sanderson 1 0 ([55], [56]) are always obtained here: (1 ) (X max ) stochastically dominates (1 ) (X max ) in the 1 1 max 0 first degree sense in [0 X 1 ]. 6 2 2 6 0 e e e 2 e Z eZ e e e Z e r 002 e | Z Z e e | e e Z Z e e Z | (1 ) er Z e 2 Z e e Z | e eZ | e|Z e e | ZZ | e e Z e e | e |Z Z e e Z e| e|Z 001 1 0 Xmax 1 00 Xmax 1 - 1 Xmax 1 Figure 1: Mean demand for 1 , ̄1 ( ), drawn from (1 ). Initial choices are truncated by 0 Xmax = 1 . For uncompensated changes, the truncation point is Xmax . For compensated 1 1 max 00 changes, the truncation point is X1 . Some remarks. (i) Proposition 1 does not depend on particular assumptions about (1 ). The uniform distribution used by Becker [9] yields 1 ̄1 ( ) = (1 ) Z 0 1 1 1 = 1 , 2 1 (8) and ̄2 ( ) = (12)(2 ); these mean demands are exactly the same as the ones derived by Becker [9]. These demands also coincide with those of a rational “representative agent” who maximizes a Cobb-Douglas utility function. (Theorem 2 below provides integrability results that generalize this case.) (ii) Uncompensated price effects satisfy the standard Slutsky equation, ¯ ¶ µ ̄1 ( ) ̄1 ( )¯¯ ̄1 ( ) , = − ̄1 ( ) 1 1 ¯ 7 (9) with income effects given by ̄1 ( ) E [1 | 0 ≤ 1 ≤ X max 1 ] 1 = 0. max X 1 1 (10) As Figure 1 suggests, 1 is on average a normal good. Income effects, however, are not necessarily positive for ̄2 ( ), or in general. A sufficient condition for ̄2 ( ) to be a normal good is that (1 ) is log-concave; see Appendix A. This condition is implicit in Becker [9] since the uniform distribution is log-concave. (iii) Proposition 1 examines mean demands but other aggregators also satisfy the Law −1 of Demand. Particularly, 1 ’s median demand med ( (X max 1 )2) satisfies 1 ( ) ≡ med 1 (X max 1 ( ) 1 ) = 0, max med X 1 2 (1 ( )) (11) which, as in (5), is enough to verify the Law of Demand. The Law of Demand also holds for the mid-range aggregator since this statistic treats (1 ) as a uniform distribution. Similar results cannot be established for the mode, but the mode is not meaningful here. (The mode does not generally depend on X max 1 .) (iv) When 1 increases, the fraction of individual demands (̄1 ( )) − (0) agrees 00 with the compensated Law of Demand but the fraction (X max ) − (̄1 ( )) violates 1 it, as demands increase for these individuals; see Figure 1. These violations have been used to measure the statistical power of revealed preference tests and the economic significance of “irrational” behavior. I discuss this application in detail in Section 5. (v) I focused on 1 but the Law of Demand also holds, on average, for 2 ; see Appendix A. Indeed, (2) and (3) satisfy all the properties of the behavioral model of choice. Homogeneity of degree zero in ( ) follows because X max is a defined in real terms; see 1 (2). By construction, ̄2 ( ) is also homogeneous and adding up conditions are satisfied; see (3). Mean demands are symmetric with respect to cross-price changes so they can be rationalized, e.g., integrated. Integrability, however, is expected because this example considers two commodities and non-interior demands; see, e.g., Mas-Colell et al. ([47], p. 36 and Exercise 2.F.15) and Katzner ([38], Theorem 4.1-2). 8 3 A general framework This section deals with generalizations of the previous example. Further generalizations that consider several new dual (i.e., “irrational”) representations of demands in the behavioral model of choice, a derivation of an aggregate supply curve under random choice behavior, and numerical simulations that examine how ‘large’ the economy needs to be in order to observe consistent results are discussed in an Appendix not for publication. Statistical choices will now take place over 1 commodities using a general probability distribution function. Mean demands will also be allowed in the interior of the budget set. Let ≡ (1 ) denote the commodity vector and let ≡ (1 ) be the corresponding price vector. The budget set generalizes (1), © ª ( ) ≡ ∈ R+ : 0 ≤ · ≤ . (12) The budget constraint (12) may hold as an inequality depending on whether one or more commodities are randomly chosen or determined as a residual. I will separately study the case when the “last” commodity is chosen randomly and when it is determined as a residual. As before, all random variables are absolutely continuous with finite first and second moments. Statistical choices are derived from a well-behaved probability distribution function (1 ) ≡ Pr(1 ≤ 1 ≤ ) with density (1 ). The support of these functions is given by the commodity space R+ . (This is also the support for the behavioral model of choice; see Mas-Colell et al. ([47], p. 18).) To ensure that (12) is satisfied for every realization of , (1 ) must be truncated according to (X max )= Z 0 max 1 Z max (1 −1 ) (1 ) 1 , (13) 0 max where X max ≡ (X max 1 X (1 −1 )) is the vector of maximum feasible consump- max tions and (X max ) ≡ (X max 1 X (1 −1 )). This truncation takes place over 9 the -dimensional budget hyperplane associated with (12). In particular, the vector X max satisfies X max = for = 1 and (1 −1 ) = −1 [X max X max −1 (1 −2 ) − −1 ] , for = 2 . (14) (1 −1 ) is a subset of −1 X max This vector is ordered: X max −1 (1 −2 ) because a positive consumption for commodities limits the maximum amount that can be consumed of commodity , i.e., if = X max (1 −1 ) then X max (1 −1 ) = 0 for all = +1 . For instance, the “last” commodity satisfies X max (1 −1 ) = P−1 − =1 . Mean demands. Mean demand ̄ ( ) ≡ E( |0 ≤ · ≤ ) satisfies ̄ ( ) = Z 0 max 1 Z max (1 −1 ) 0 (1 ) 1 . (X max ) (15) The vector of mean demands is denoted by ̄( ) ≡ (̄1 ( ) ̄ ( )). Notice that mean demands are homogeneous of degree zero in ( ). Homogeneity is easy to verify because ( ) influence ̄( ) only through X max and a proportional change in and leave X max unchanged; see (14). Moreover, if commodity is residual, then = X max (1 −1 ) for any realization of 1 −1 . In this case, the relevant distribution function is (1 X max (1 −1 )) and individual and mean demands add up to income. Notice also that the order of integration is in general irrelevant. The “last” commodity is special, but only when it is residually determined.10 I characterize mean demands next. All derivations are in Appendix A; they are not difficult but tedious as they rely on repeated differentiations under the integral sign. Let ≡ (1 ) and let () denote the vector when is excluded, i.e., () ≡ (1 −1 ). Likewise, let ≡ (1 ) be the vector of differential changes and let () denote this vector when is excluded, i.e., () ≡ (1 −1 ). Finally, let max (() ) denote the density () when X max (() ) takes the place of . That is, 10 Under interior demands, for example, the order of integration in (13) and (15) can be changed without consequence due to Fubini’s Theorem; see, e.g., Fikhtengol’ts ([26], Vol. II, Section 344). (This is generally true for the first − 1 goods.) This change would be somewhat analogous to changing the order in which first order conditions are obtained in the behavioral model. 10 max (() ) ≡ (() X max (() )), which is only a function of () . This density function appears repeatedly in the comparative statics in this section. (a) Own-price effects. As derived in Appendix A, ¯ Z max Z max 2 1 −1 ((−1) ) [ − ̄ ( )] max ( ̄ ( )¯¯ () ) = − max () . ¯ (X ) 0 0 (16) A compensated own-price change is proportional to the negative of the conditional variance of . Slightly abusing notation in the definition of variance, i.e., ignoring the conditioning terms and proportionality factors, let [ ] ≡ E[( − ̄ ( ))2 |() ≤ max X max () = X ]. Then (16) can be written as ¯ ̄ ( )¯¯ = − [ ]. ¯ (17) Thus, for any well-behaved probability distribution function (), a compensated increase in reduces ̄ ( ).11 Figure 2 illustrates the own-price effect for ̄1 ( ) when = 2 and demands are 0 interior. A compensated change in 1 removes the triangular area ( X max X max 1 1 ) and max 0 adds the area ( X max (0)) to the budget set. This ‘pivoting’ of the budget set 2 (0) X 2 max is integrated along X max 2 (1 ) using (1 X 2 (1 )). (In general, changes are integrated max along X max (() ) using (() ).) As in Section 2, the compensated Law of Demand holds because the area that is removed once prices increase favors high values of 1 whereas the area that is added favors low values. (b) Income effects. Income changes satisfy ̄ ( ) = Z 0 max 1 Z max −1 ((−1) ) 0 [ − ̄ ( )] max (() ) () . (X max ) (18) To determine the sign of (18) one needs to make additional assumptions on the distribution function (). In particular, the sign of the previous expression depends on the 11 Mean demand ̄ ( ) when is not residual needs a minor adaptation; see (A5) in Appendix A. 11 2 6 00 Xmax (0) 2 e e e max X2 (0) Q e Q e Q Qe Qe 00 r 2 r Qe Q eQQ e rQ 2 e QQ e Q e Q Q e Q Q e Q e Q e Q 001 00 Xmax 1 1 - 1 Xmax 1 Figure 2: The Law of Demand for interior demands and = 2. (The density (1 2 ) is not represented.) To determine own-price changes, (1 Xmax 2 (1 )) must be integrated max along the “thick” boundary X2 (1 ). association between and . One needs to sign Z max 1 0 Z max −1 ((−1) ) 0 max (() )() Z ≷ ̄ ( ) 0 max 1 Z max −1 ((−1) ) max (() )() . 0 This expression compares ̄ ( ), the mean value of , to E[ |() ≤ X max () = X max ], which is the mean value of when takes its highest possible value. “Positive association” in the likelihood ratio sense implies that a high value of likely yields a high value for .12 Thus, if and are positively associated, income effects in (18) will be positive; if and are negatively associated, income effects in (18) will be negative. Figures 3 and 4 illustrate this dependence for = 2 and interior demands. Figure 3 considers positively associated variables. At income , mean demands (̄1 ̄2 ) are determined 12 Formally, (1 2 ) is said to be positively likelihood ratio dependent if (01 02 ) (1 2 ) ≥ for 01 1 and 02 2 . Thus it is more likely to observe that 1 and 2 take larger values together and smaller values together than any mixture of these; see, e.g., Figure 3. The multivariate analog requires (1 ) to satisfy a positive likelihood ratio dependence for every pair ( ) when the − 2 remaining variables are fixed; see Shaked and Shanthikumar [57]. The relationship with log-concavity is also discussed by these authors. (01 2 ) (1 02 ), 12 2 0 Xmax (0) 6 2 e e e e max X2 (0) e e e e e (1 2 ) = e e e e ¡ e ¡ ª e e e e e e e 0 r 2 e e e e e 2 r e e e e e e e e 1 01 Xmax 1 - 1 0 Xmax 1 Figure 3: Income effects for 1 are positive when 1 and 2 are “positively associated.” (The density (1 2 ) is represented by its contour.) excluding the shaded area. At 0 , the shaded area becomes relevant to determine demands. Since this area favors high values of 1 , its mean demand increases to ̄01 . Negative income effects are depicted in Figure 4. If 1 and 2 are negatively associated, the new (shaded) area added as income increases favors low values of 1 and this lowers mean values. (c) The Slutsky matrix. Let S( ) denote the Slutsky matrix of substitution effects. I next show that the Slutsky matrix is symmetric and negative semidefinite. A compensated change in on commodity yields ¯ Z max Z max 1 −1 ((−1) ) [ − ̄ ( )][ − ̄ ( )] max ( ̄ ( )¯¯ () ) = − max () , ¯ (X ) 0 0 (19) which can be written as the -th entry of a well-behaved variance-covariance matrix. Several of the properties of the Slutsky matrix will follow from this representation. More specifically, let Σ be the variance-covariance matrix of (() X max (() )) about ̄( ). A typical element of Σ is of the form Σ = [ ] ≡ E[( − ̄ ( ))( − 13 2 0 Xmax (0) 6 2 e e e e max X2 (0) e e e e e (1 2 ) = e e¡ ª e ¡ e 0 e 2 r e ee e e 2 r e e e e e e e e e e e e e e e 01 1 Xmax 1 - 1 0 Xmax 1 Figure 4: Income effects for 1 are negative when 1 and 2 are “negatively associated.” (The density (1 2 ) is represented by its contour.) max ̄ ( ))|() ≤ X max () = X ], which is equivalent to (19) up to a proportionality factor, ¯ ̄ ( )¯¯ S ( ) ≡ = −[ ], ¯ (20) for = 1 − 1, and with = X max (() ) when = ; see Appendix A. The Slutsky matrix can be written as a matrix of second order moments, S( ) = −Σ. Since the variance-covariance matrix Σ is symmetric and positive semidefinite; see, e.g., Fisz ([27], pp. 89-90), the Slutsky matrix S( ) is symmetric and negative semidefinite. Moreover, if is randomly determined and not selected as a residual, then () −̄() ( ) would be linearly independent from X max (() ) − ̄ ( ) making all terms in Σ linearly independent. In this case, Σ would be symmetric and positive definite; see, e.g., Fisz ([27], Theorem 3.6.6). Thus, when demands are interior, ( ) will be symmetric and negative definite instead of just semidefinite. The distinction between positive definiteness and positive semidefiniteness in S( ) is not a curiosity; it will be important for demand integrability. Theorem 1 summarizes the key implications I have presented so far: 14 Theorem 1 (a) Assume that individuals randomly choose subject to (12). For any well-behaved continuous probability distribution function (), mean demands ̄( ) are: linearly homogeneous in ( ); interior, · ̄( ) ; and their Slutsky matrix S( ) is symmetric and negative definite. (b) Assume that individuals randomly choose () subject to (12) and determine the “last” commodity as a residual. For any well-behaved continuous probability distribution function (), mean demands ̄( ) are: linearly homogeneous in ( ); add up to income, · ̄( ) = ; and their Slutsky matrix S( ) is symmetric and negative semidefinite. Economically, the difference between (a) and (b) in Theorem 1 is that if the “last” commodity is a residual, (individual and mean) demands would add up to income so would be redundant. This redundancy of the “last” commodity is of the usual kind: knowledge of () and (12) is enough to determine individual and mean demands. In the behavioral model, homogeneity and adding up also imply that the negative semidefiniteness of the Slutsky matrix cannot be extended to negative definiteness; see, e.g., Mas-Colell et al. ([47], Proposition 2.F.3). In the statistical model, the linear dependence implicit when the “last” commodity is a residual makes a linear combination of () , i.e., = X max (() ). This gives rise to positive definite but singular variance-covariance matrix; see, e.g., Fisz ([27], Theorem 3.6.6). The intuition for the symmetry and negative semidefiniteness of the Slutsky matrix is transparent in (20). This transparency should not detract from the fact that symmetry and negative semidefiniteness are striking findings. For once, negative semidefiniteness establishes the compensated Law of Demand under general conditions including interior demands; see Mas-Colell et al. ([47], pp. 34-35) for uses of S( ) in the behavioral model of choice. More importantly, symmetry and negative semidefiniteness in the Slutsky matrix S( ) are exhaustive properties of the behavioral model of choice; see, e.g., Mas-Colell et al. ([47], pp. 75-76). Mean demands ̄( ) can therefore be rationalized as being the result of the maximization of some utility function: 15 Theorem 2 (a) Under the conditions of Theorem 1(a), there exists some continuous, non-decreasing, → R such that ̄( ) and ̄0 ( ) ≡ and quasi-concave utility function : R+1 + − · ̄( ) are the unique solution to max0 ≥0 {( 0 ) : · + 0 = }. (b) Under the conditions of Theorem 1(b), there exist some continuous, non-decreasing, and quasi-concave utility function : R+ → R such that ̄( ) is the unique solution to max≥0 {() : · = }. Theorem 2 shows that mean demands can be rationalized even when they are interior. To do so, Theorem 2(a) treats ̄( ) as an incomplete demand system. Epstein [24] characterized integrability in incomplete systems; see also LaFrance and Hanemann [41]. Incomplete systems are more economically and mathematically complex than complete systems because one must make economic assumptions about how to complete the system. (Theorem 2(b) relies on a standard complete demand system; see Katzner ([38], Chap. 4).) Suppose for illustration that incompleteness is due to unobserved commodities. In this case, one has to limit the way observed prices of unobserved commodities influence observed demands. A “residual” commodity 0 with 0 = 1 is the simplest way to complete the system. (Some remarks about the interpretation of 0 in the present context are provided below.) Incomplete demands also require a symmetric and negative definite Slutsky matrix. This mathematical requirement cannot be weakened for semidefiniteness; see Epstein ([24], Example 1). Some remarks. (i) Random choices clearly violate individual rationality. On average, however, behavior is consistent with a rational “representative agent.” The utility function that rationalizes ̄( ) even has a dual random representation. (An Appendix not for publication elaborates on this kind of duality principle. In particular, I show that any consumer demand function in the behavioral model of choice can be represented as the outcome of random choice behavior.) In case (b) of the previous theorems, for example, E[|0 ≤ · = ] = argmax {() : 0 ≤ · = )}. ∈R + 16 (21) The utility function () and its associated welfare measures (i.e., the consumer’s surplus) have no content in the statistical model of choice. The process by which choices are revealed, even the choices themselves, provide no basis for normative statements. The difficulty with normative assessments is not the standard problem that the welfare function of the “representative agent” is in conflict with the preferences of the disaggregated individuals, as in the traditional treatment of the normative “representative agent” discussed, e.g., in Mas-Colell et al. ([47], Section 4D). The issue here is more fundamental. Theorem 2 demonstrates that well-behaved preferences would emerge even when individual choices have no preferential basis and there is no maximizing behavior. As the dual representation (21) shows, well-behaved preferences can be recovered from purely random choice data.13 As Mas-Colell et al. ([47], pp. 121-122) note, however, “[t]he moral of all this is clear: The existence of preferences that explain behavior is not enough to attach them any welfare significance. For the latter, it is also necessary that these preferences exist for the right reasons.” (ii) Positive statements about price and income effects have empirical content. They are, however, unable to discriminate between the statistical and the behavioral model of choice. At the market level, both models are equally consistent with, for example, a symmetric and negative semidefinite Slutsky matrix. Even at the individual level it will be difficult to discriminate between a statistical model of choice and a behavioral model. The behavioral model of choice is often assessed by applying revealed preference tests to individual choice data. To make these tests operational, the literature typically adds “noise” to individual demands. The behavioral model of choice is deterministic so ‘small’ violations of behavioral assumptions are traditionally attributed to a stochastic component in the decision process (i.e., mental errors) or to measurement error in prices or individual choices. In statistical tests of the behavioral assumptions, individual demands are assumed to have a “noise” component; see, e.g., Choi et al. [17], Echenique et al. [22], Lewbel [42], and Varian [62]. In the statistical model of choice, individual demands do have a “rational” and a “noise” component. Deviations from “rational” behavior are in fact bounded. Let denote = 13 This point was discussed by Becker [9] in the context of an ‘as if’ justification for rationality and irrationality in economics. Blundell et al. ([11], p. 211) also mentioned this point in passing in their analysis of demand integrability. 17 1 independent sample realizations of an individual’s vector of demands. The sample P average for the individual demand for commodity is defined by ̃ ≡ −1 =1 . Corollary 1 Individual demands in the statistical model of choice have a “rational-plusnoise” representation = ̄( ) + , (22) with ≡ − ̄( ) and E[ ] = 0. Moreover, given 0 and for = 1 , individual demands satisfy ¯ o [ ] n¯ ¯ ¯ Pr ¯̃ − ̄ ( )¯ . 2 (23) (iii) Individual choices can also be represented by “observed” choice probabilities, in the sense of McFadden and Richter [50]. Let denote a subset of ( ). The probability that choices lie in is 1 (|( )) = (X max ) Z (). (24) “Observed” choice probabilities are the starting point of stochastic revealed preference theory. Stochastic revealed preference theory seeks to rationalize observed choice probabilities using a random preference model (see, e.g., footnote 7). Individual choices here are not based on random preferences or maximizing behavior. “Observed” choices, for instance, are not required to satisfy any form of stochastic transitivity.14 Hence, the current statistical model of choice is “irrational” not just in the standard deterministic sense, but also in the more general stochastic sense. Still, Theorem 2 and Corollary 1 imply that individual choices in the statistical model admit a random utility representation 14 There are several notions of stochastic transitivity. Consider , , and under the same feasibility conditions and as binary choices, i.e., a choice between one unit of commodity and one unit of . The proportion of individuals for whom chance selects alternative out of alternatives and is ≡ ( |{ }) = Pr{ }; and are analogously defined. If ≥ 12 and ≥ 12, choice probabilities satisfy: weak stochastic transitivity if ≥ 12, moderate stochastic transitivity if ≥ min{ }, and strong stochastic transitivity if ≥ max{ }; see, e.g., Tversky [59]. Binary choice probabilities can be ordered following basic probability theory; see Appendix A. For instance, ≥ and ≥ if and only if Pr{ } ≥ Pr{ } and Pr{ } ≥ Pr{ }, respectively. Violations of stochastic transitivity can only be ruled out by restricting these probabilities. There is no simple interpretation for these restrictions but none of these conditions is needed in Theorems 1 or 2. 18 ( ) = (̄( )) + , with ≡ (̄( ) + ) − (̄( )) and ≡ − ̄( ). Purely random choice data can therefore be behaviorally represented as a random utility model in which an individual’s utility ( ) results from random perturbations to a well-behaved deterministic indirect utility function (̄( )). (iv) Finally, the “residual” commodity 0 needed for integrability in Theorem 2(a) is not too restrictive. If the number of commodities is large, the “residual” commodity 0 ≡ − · will become insignificant. (Trivially, by Markov’s inequality, Pr{0 } ̄0 ( ), so lim→∞ Pr{0 } = 0.) Moreover, since 0 = 1, 0 could be interpreted as an individual’s cash balance; it could also be interpreted as wasted money or as savings in an intertemporal context. Tests of the behavioral assumptions in economics effectively analyze incomplete demand systems since they are typically carried out in terms of total expenditure and not in terms of income. This means that existing tests could not assign any particular meaning to 0 . As trivial as it might sound, I’m not aware of empirical assessments of non-satiation. Non-satiation is not actually tested as part of conventional revealed preference tests. Empirical analyses typically assume or impose non-satiation to avoid trivial rationalizations of the data; see, e.g., Varian ([61], p. 969). 4 Applications to random choice behaviors The only restrictions needed in Theorems 1 and 2 are those implied by probability theory. One can therefore showcase the applicability of these theorems by considering statistical models of choice that restrict choice probabilities. This section lists a few random choice models formulated for individual decision-making and provides some remarks about the applicability of the previous theorems. Preferential choices. I first consider some classical preferential choice models in which () is derived from behavioral postulates or choice heuristics. I make the dependence on the choice set explicit and use continuous probability distributions.15 (a) Luce’s choice axiom and Tversky’s elimination model. Let (|) be the probabil15 Continuous choice models can be derived as the infinitesimal limit of discrete choice models; see, e.g., Ben-Akiva and Watanatada ([10], p. 327) and McFadden ([48], pp. 311-312). 19 ity that an alternative from is chosen when the set of available alternatives is . In Luce [43], choice probabilities satisfy the independence from irrelevant alternatives: for ⊆ and 0 ⊆ such that ⊆ 0 , (|) = (| 0 ) ( 0 |). This assumption yields a continuous logit formulation Z exp{()} (|) = Z , (25) exp{()} where () is a direct utility function associated with reflexive, transitive, and complete preferences.16 The parameter accommodates extremes in the rationality spectrum; the larger is, the greater the degree of “irrationality.” The limit of (|) as → ∞ is the uniform distribution used by Becker [9], and when → 0, the choice problem becomes a deterministic utility maximization problem; see Anderson et al. ([3], p. 42). Tversky [60] considered a more general choice process in which alternatives are sequentially eliminated. Individuals first choose a subset ⊆ with probability ∫ ( 0 )0 and then choose alternatives within this subset, e.g., choices are of the form (|) = R (|0 )( 0 )0 , where (|0 ) represents the probability of selecting given 0 , and ( 0 ) is a weighting scheme that determines the probability of choosing 0 in , with ∫ ( 0 )0 = 1. Luce [43], for instance, is a special case in Tversky [60]. Moreover, Luce [43] and Tversky [60] can be represented as random preference models; see, e.g., Anderson et al. ([3], Chap. 2). (b) Choice control models. Even more general probabilistic choice behaviors can be seen as outcomes of individual randomization that maximize a utility function that faces a cost of implementing the choice, i.e., a desire for randomization. This probabilistic choice structure has been studied by Machina [44] and more recently by Fudenberg et al. [28]. Let () denote a choice function that specifies the probability of choosing a commodity bundle ∈ ⊆ R+ . Let () be a direct utility function and, as in Fudenberg et al. ([28], p. 2371), let (()) denote “a convex perturbation function that may reward the 16 Random preference models often assume that there is an indirect utility function that depends on prices an income, and that individuals randomly choose among these indirect utilities. The utility functions used in this section are all direct so they are independent of the economic environment. 20 agent for randomizing.” Individual choices, which can be seen as randomly selecting among deterministic bundles ranked according to (), satisfy ∗ (|) ≡ argmax ()≥0 ½Z [()() − (())] : Z ¾ () = 1 , (26) where the feasibility restriction is implied by the probabilistic nature of choice. The presence of a convex cost (()) means that individuals may find random choice preferable to R deterministic choices. In (26), for ⊆ , choice probabilities (|) = ∗ (|) are behaviorally constructed. (c) Difference models. Gonzalez-Vallejo [32] and Roe et al. [54] are examples of multialternative preferential choice models consistent with violations of stochastic rationality. These models are also more general than Luce [43] and Tversky [60]. (Tversky [59] also used proportional difference models to study stochastic intransitivity.) For instance, individual choices in these models are based on pure preferences, inattention, informational updates, and a comparative evaluation of the different alternatives. In a simplified version of Roe et al. [54], for example, individual choices satisfy (|) = Φ( : {() · () ≥ max[() · () : ∈ ]}), where () is an attention weight for the at- tributes = (1 ) and () = 1 (1 ) + + ( ) is a pure preference component perturbed by a stochastic error; see, e.g., Roe et al. ([54], Eq. 1b). In Gonzalez-Vallejo [32], individual choices also have a preferential basis, but the comparative evaluation is based on proportional differences between alternative attributes. Difference models are not based on primitive axioms. Behavioral scientists postulate these models as descriptive representations of actual choices. These models, for instance, are consistent with violations of stochastic regularity and stochastic transitivity along the lines discussed by remark (iii) in Section 3. (d) Satisficing. Consumption alternatives in the previous examples are evaluated based on maximizing behavior. Simon [58] advocated satisficing as a comparative principle of individual choice. In Simon ([58], p. 252), a bundle is satisfactory provided that () ≥ , where is a given aspiration level. In a random choice formulation, individuals would be satisfied with a bundle whose utility value exceeds , e.g., (|) = Φ( : {() ≥ ∈ }). 21 “Irrational” choices. The previous examples have a preferential basis. Individual choices can also be based on ‘pure’ statistical models of choice in which () lacks behavioral foundations. These “irrational” choices typically arise when preferential choices have limited value. Individuals sometimes act impulsively, instinctively, and/or emotionally due to limited decision time or attention.17 Preferences may also be incomplete or individuals might find desirable to follow random choices out of fairness or due to strategic considerations; see, e.g., Moore [52] for a classical application in which individual choices require an “impersonal and relatively uncontrolled process.” Subtle forms of random choice behavior, some still common today (i.e., organized religion, astrology, superstition, divination, etc.), appear to resolve situations that exhibit indecisiveness; see, e.g., Elster [23] for a discussion of random choice in the context of limits of rational behavior.18 As noted by Agranov and Ortoleva ([1], p. 2), random choice might also arise from “preferences originating from the desire to reduce regret, incomplete preferences, difficulty to judge one’s true risk aversion, or other forms of Non-Expected Utility.” Indeed, Agranov and Ortoleva [1] found, in an experimental setting, that random choice is deliberate especially when subjects are dealing with ‘hard’ (i.e., complex) questions. Some remarks. (i) The previous list of examples is obviously non-exhaustive. These examples, however, are more general than Becker [9]; they, for instance, rely on descriptive assumptions that explicitly take into account inconsistent preferences, optimization errors, and bounded rationality. In conjunction with Theorem 1, these examples show how broad the domain of market demand theory can be and how unnecessary it is to rely on the ‘pure’ behavioral model of choice (i.e., on individual rationality) or the ‘pure’ statistical (i.e., “irrational”) model of choice to derive its main testable predictions. 17 These instances motivated Becker [9] and an empirical demand literature on nonhuman or nonstandard populations (see footnote 4). McGrath ([51], p. 433), for example, observed male shoppers in gift shops on Christmas Eve tended “to make large, rapid, spontaneous, and often random purchases.” Becker [9] considered “impulsive good deciders” and Chant [15] “impulsive money deciders” but the difference is irrelevant here since individuals are price-takers. 18 As noted by Aumann ([6], p. 446): “[c]ertain decisions that our individual is asked to make might involve highly hypothetical situations, which he will never face in real life; he might feel that he cannot reach an ‘honest’ decision in such cases. Other decision problems might be extremely complex, too complex for intuitive ‘insight,’ and our individual might prefer to make no decision at all in these problems.” Indifference and indecisiveness though generally have different testable implications in the behavioral model; see, e.g., Mandler [45] and the subsequent literature. 22 (ii) In conjunction with Theorem 2, these examples also show that a rational “representative agent” carries through for a large class of individual behaviors. The emergence of rationality is useful to study markets in which individuals systematically depart from conventional behavioral assumptions. Gode and Sunder [29], for instance, documented high allocative efficiency in a market mechanism (i.e., a double auction) with “zero-intelligence” (ZI) traders characterized by uniformly distributed bids and asks. Gode and Sunder ([29], p. 121), however, explicitly acknowledge that “[t]hese ZI traders are not intended as descriptive models of individual behavior.” The preferential choice examples listed above are descriptive models of individual behavior. The allocative efficiency of markets in which these descriptive behaviors are present has not been systematically studied. Market interactions, however, should also serve as a partial substitute for individual rationality under more realistic individual choice behaviors. (iii) The preferential choice examples listed above recognize that individual choices are context-dependent. The choice set , for instance, is an argument of market demands, as in ̄( |). The fact that individual choices depend on the choice context is not inconsistent with demand theory or with the existence of a rational “representative agent.” Market demands have been characterized holding other things constant and assuming that the budget set constrains choice, i.e., that ( ) ⊂ . The analysis of market demands derived from (|) complements psychological and experimental research that formalizes how individual choices depend on the choice context. This paper does not study the relationship between market demands and the set of alternatives or between the utility function of the “representative agent” (|), the choice set , and the primitive individual utility function (). These questions are beyond the scope of this paper. 5 Statistical power As corollaries to Theorem 1, this section discusses the statistical power and economic significance of tests of the weak axiom of revealed preferences, the most famous testable prediction of the behavioral model of choice. A bundle is weakly revealed preferred to 0 (i.e., satisfies WARP) if whenever · ≥ · 0 it is false that 0 · 0 0 · ; see, e.g., 23 Mas-Colell et al. ([47], Chap. 2) and Echenique et al. ([22], p. 1206). My emphasis here is on WARP, the simplest possible setting for statistical tests of rational behavior. I present some remarks about the generalized axiom of revealed preference (GARP) later on. Tests of revealed preferences measure statistical power against particular alternative hypotheses. As Echenique et al. ([22], p. 1220) notes, it is difficult “to find an acceptable alternative benchmark to rationality under which to measure power.” As an alternative hypothesis, the statistical model of choice requires parametric assumptions on (); see Bronars [12], Choi et al. [17], and Andreoni et al. [4] for specific examples. This section shows that revealed preference tests are inconsistent, and that their inconsistency is insensitive to the choice of the distribution function under the alternative. I consider compensated price changes and formulate the null and the alternative hypotheses as:19 0 : Choices are drawn from a behavioral model of choice in which no ‘true’ violation of revealed preferences exists but there is (additive) measurement error to the consumer’s choices. : Choices are drawn from a statistical model of choice according to a distribution function (). Tests based on the size of the violations. The economic significance of violations of rationality can be measured in terms of the amount of money that can be ‘pumped out’ of non-rational individuals; see Echenique et al. [22]. Before describing the money pump statistic, consider a more general approach. Given a pair of consumption bundles of the form ( ) and (0 0 ), the statistic = · ( − 0 ) + 0 · ( 0 − ) measures the potential profits or losses of a devious “arbitrager” that exploits violations of rationality. In particular, the positive part + ≡ max{0 } and the negative part − ≡ max{− 0} denote the money pump cost and the money drain cost, respectively. Obviously, = + − − so an arbitrager either makes money out of these 19 Tests are uninformative if budget sets do not intersect, even with deterministic data; see Beatty and Crawford [8]. Intersecting budget sets is a standard assumption; see, e.g., the equal marginal utility of income (EMUI) assumption in Echenique et al. [22]. Andreoni et al. [4] discuss this point in great detail. 24 transactions when + 0 (and there is a violation of the weak axiom) or loses money when − 0 (and the weak axiom is satisfied). Echenique et al. [22] focused on the money pump but I will describe the asymptotic properties of tests based on and + . Since mean demands are well-behaved in the statistical model of choice, the power of tests based on these statistics agree almost completely. Consider first a test based on . (The case of + is presented below.) Under the null, = ( − 0 ) · ( ( ) − (0 0 )) + ( − 0 ) · ( − 0 ), where ( ) and (0 0 ) represent the unobserved ‘true’ choices and and 0 are distributed according to Φ( 2 ). The appropriate test is one-sided and rejects the null if the -statistic exceeds a critical test-value. The distribution of the -statistic under the null cannot be calculated without additional assumptions on the ‘true’ choices. As first explored by Varian [61], under the null of no ‘true’ violations, there is an upper bound statistic given by = ( − 0 ) · ( − 0 ), with ≥ and with normally distributed with 2 ≡ 2 · k − 0 k2 · 2 . Given a significance level and a value for 2 , one can calculate a critical test-value 0 from the normal distribution according to Pr{ |0 } = Φ( : { }) = . The probability of a Type I error is thus no greater than , i.e., Pr{ |0 } ≤ . The power function of tests based on the -statistic is Pr{ | }. Power based on the critical values from the -statistic cannot exceed the power of a test based on the -statistic, i.e., Pr{ | } ≥ Pr{ | }. To characterize asymptotic power, let { : = 1 } denote the sample data. The observations are obtained at ( ) and the observations at (0 0 ). Let ≡ ( ) = ( − 0 ) · ( − ) for 6= denote the sample -statistic for the ( )-pair. The -statistic can be computed for a total of ( − 1)2 pairwise sample comparisons. The sample average of the -statistic P is ̃ ≡ −1 ()6= , or ̃ = −1 X ()6= ( − 0 ) · ( − ). (27) The behavior of the -statistic under the alternative is standard: ̃ converges (weakly) to ̄ ≡ ( − 0 ) · (̄( ) − ̄(0 0 )) ≤ 0 (i.e., ̃ → ̄ when → ∞), where the inequality 25 follows from Theorem 1. Moreover, given 2 ≡ 2 · k − 0 k2 · Σ, with Σ as the variance- covariance matrix of and [̃ ] = 2 , Chebyshev’s inequality bounds the power function: Pr{̃ | } 2 . 2 + ( − ̄ ) (28) This upper bound depends on the first two moments of the -statistic under the alternative (̄ 2 ): the more “rational” the statistical model, the less powerful the test would be, i.e., lower values of ̄ and/or 2 (a small price change k − 0 k and/or a small variance in ) yield lower statistical power. Tests based on the money pump statistic + are related to the previous more general test statistic. In particular, the + -statistic under the null is + = max{0 ( − 0 ) · ( ( ) − (0 0 )) + ( − 0 ) · ( − 0 )} and the critical values for the upper bound statistic + ≡ max{0 } are given by + = max{0 ( − 0 ) · ( − 0 )}. The critical values are defined analogously as Pr{+ + |0 } = Φ(+ : {+ + }) = , which relies on a truncated normal distribution. The power function of tests based on the money pump + -statistic is Pr{+ + | }. As expected, statistical power for the and the + statistics decreases with the size of the critical area and + ; see, e.g., (28). Also, for any finite sample of size , the power function of tests based on the money pump statistic + cannot be smaller than the power of tests based on the -statistic, i.e., Pr{̃+ + | } ≥ Pr{̃ + | } since ̃+ = ̃ + ̃− by definition. In the limit as → ∞, however, the -statistic and even the money pump statistic + lose all their ability to detect whether or not individuals behave rationally: Corollary 2 For any well-behaved distribution function (), the asymptotic power of revealed preference tests that rely on the -statistic or the money pump + -statistic is zero. That is, lim→∞ Pr{̃ | } = lim →∞ Pr{̃+ + | } = 0. Tests based on the -statistic or the money pump cost + are inconsistent. A statistical test is consistent if its power against an alternative hypothesis tends to one as → ∞. (This is a standard statistical definition and a minimal requirement in hypothesis testing.) The inconsistency of these tests is not due the use of a particular distribution function 26 under the alternative hypothesis . Asymptotic power is actually not sensitive to the distribution function (). This means that at any level of significance , there is no wellbehaved distribution function () that generates a consistent test of individual rationality based on the - or the + -statistics. Tests based on the number of violations. Statistical power has also been computed for tests based on the number of revealed preference violations; see Bronars [12]. Assume without loss of any generality that 0 . Consider the bundles ( ) and (0 0 ) but study them in relation to ̄( ): the test involves pairwise comparisons between and ̄( ), and between 0 and ̄( ). There are two kinds of violations: for some individuals 0 · 0 might exceed 0 · ̄( ) thus revealing a violation of 0 being weakly preferred to ̄( ), while for some others · ̄( ) might exceed · thus revealing a violation of ̄( ) over .20 The sample distributions of each of these kinds of violations are Π̃ + { } ≡ #{ · ( − ̄( ))}+ #{− · ( − ̄( ))}+ , and Π̃ , { } ≡ − (29) so the sample distribution of a violation of either the first or the second kind (or of both) is Π̃ {( )} = Π̃ + { } + Π̃− { } − Π̃+ { }Π̃− { }. Under the null, the probability of observing a ‘true’ violation should be zero. The appropriate test is also one-sided and rejects the null if the number of observed violations exceeds some critical test-value. Under the null, Π+ { 0 |0 } = Φ(0 : { (0 0 ) + 0 ̄( )}) and Π− {|0 } = Φ( : {̄( ) ( ) + }). The probability of a Type I error can be determined (or at least bounded) as before so I assume that Π{( 0 )|0 } ≤ , where is some significance level. The power function of the test is Π{( 0 )| }. Corollary 3 For a given well-behaved distribution function (), the asymptotic power of revealed preference tests that rely on the Π-statistic is lim Π̃ {( 0 )| } = 1 − →∞ (̄( ))[ (X max ) − (̄( ))] . (X max 00 ) (X max ) 20 (30) In the case of the two goods depicted in Figure 1, these violations lie in the familiar segments [2 ̄1 ] 00 and [̄1 X max ]. An indifference curve tangent to these segments would have to be concave to the origin; 1 an inconsistency with the weak axiom. 27 Tests based on the number of violations are also inconsistent. Asymptotic power Π{( 0 )| } equals one if and only if (̄( )) = 0 or (̄( )) = (X max ). If mean demands are interior in the statistical model of choice, asymptotic power will necessarily be less than one. As in the case of the - and the + -statistics, at any level of significance , there is no well-behaved distribution function () that generates a consistent test of individual rationality based on the Π-statistic. In contrast to tests based on the - and + -statistics, asymptotic power for tests based on the Π-statistic is sensitive to the distribution function in the alternative hypothesis. On one hand, power under a (truncated) symmetric distribution about ̄( ) is larger than power under asymmetric distributions, all else equal. Under a symmetric distribution, asymptotic power is 1 − (X max )[4 (X max 00 )]; power would be further enhanced if (X max ) does not exceed (X max 00 ) by a large margin. Asymptotic power would be at most 34 if (X max ) = (X max 00 ). This upper bound might still represent non-trivial power. Asymptotic power, on the other hand, would equal zero if (X max 00 ) = (X max )4 and (̄( )) = (X max )2. In this case, tests are also completely uninformative. Some remarks. (i) Corollary 2 analyzes the average money pump cost. Echenique et al. [22] also studied the median money pump cost. I have not studied median demands since there is no general way to define the median for the multivariate case. (One approach is to generalize the notion of univariate conditional median using marginal distributions, as in (11).) Median demands in Section 2 also satisfy the compensated Law of Demand. Thus the asymptotic power of tests of revealed preferences based on the median behavior of the and + -statistics is also likely to be small and possibly zero. (ii) I have examined cycles of length two and not the more general cycles associated with the generalized axiom of revealed preferences (GARP). A data set satisfies GARP if for each pair of bundles and with = 1 if · ≥ · then it is false that · · . Given a sample sequence of observations of the form 1 , 2 , ..., , P P one can compute a generalization of the - and + -statistics as T̃ = −1 =1 =1 · ( − +1 ), with +1 = 1 , and T̃+ = max{0 T̃ }.21 For any finite sample of size 21 Power calculations for tests of GARP in terms of the number of violations cannot be derived in closedform, even for the uniform distribution used by Becker [9]. The literature relies on Monte Carlo simulations to measure power in these more general cases; see, e.g., Bronars ([12], p. 695). 28 , the power of tests based on the generalized statistics T̃ and T̃+ cannot be smaller than the power of tests based on the and + -statistic since longer cycles allow for more revealed preference violations. Under the alternative hypothesis, however, T̃ → T̄ ≡ P +1 +1 )) with ̄(+1 +1 ) = ̄(1 1 ), so one should not =1 · (̄( ) − ̄( see any arbitrage opportunity as mean demands ̄( ) satisfy GARP, i.e., they are the outcome of maximizing a quasi-concave utility function (Theorem 2). Even the generalized money pump T̃+ -statistic will have a difficult time identifying individual rational behavior since the T+ -statistic has no statistical power, i.e., lim →∞ Pr{T̃+ 0| } = 0. (iii) Although I have focused on particular test statistics, it is possible to construct additional statistical tests of the behavioral axioms using rationality indices. Afriat’s efficiency index measures the maximum “margin of error” across violations of the weak axiom (Echenique et al. [22] and Varian [62]), the minimum cost index measures the monetary cost of breaking all revealed preference violations in a given data set (see Dean and Martin [19]), and the swaps indices measure the number of “choice swaps” required to rationalize choice inconsistencies in a given data set (Apesteguia and Ballester [5]). Even the Slutsky matrix can be used to assess individual rationality (Aguiar and Serrano [2]). These and similar rationality indices are unlikely to yield consistent statistical tests. For a statistical test of individual rationality to be consistent, the test statistic should be able to perfectly discriminate between the statistical and the behavioral model of choice when the number of observations is unlimited. That is, a consistent statistical test of individual rationality requires that, for a given choice of () in , individuals in the statistical model of choice eventually make only “irrational” choices, i.e., all random choices should violate the behavioral axioms. This statistical requirement clearly contradicts Theorem 1. (iv) The “rational-plus-noise” representation of individual demands in (22) implies that the null and alternative hypotheses are economically indistinguishable. This means that the revealed preference tests formulated here are inherently unrevealing. There are, however, some important statistical issues with the previous tests: First, consistency in hypothesis testing is defined in a large-sample limit. In some simple numerical examples available in an Appendix not for publication, however, I find that random choice behavior leads to well-behaved demand functions for realistic sample 29 sizes as well. Therefore, the inconsistency discussed here is not only of theoretical interest, but it is a practical concern for tests of revealed preferences and for tests of individual rationality.22 Second, revealed preference theory is non-statistical: economic theory does not impose any meaningful restriction on the stochastic behavior of individual demands under the null hypothesis 0 . This is clearly a very important analytical gap and the most serious limitation for statistical testing of the behavioral model of choice. One can assume that the behavioral model is misspecified: prices may be the ones mismeasured (as in Echenique et al. [22]) or the measurement error may be multiplicative; or there may be a stochastic component in the decision process instead of measurement error (see remark (ii) after Theorem 2). These alternative specifications change the meaning of ‘small’ failures under the null hypothesis but they are not likely to reverse the test’s inability to detect individual rationality since these additional specifications do not affect the alternative hypothesis . Third, the null and the alternative hypotheses are not mutually exclusive because 0 and are not standard statistical hypotheses, i.e., these hypotheses are not nested. (This is why statistical power is not bounded by size, as in standard statistical hypothesis testing.) In general, standard hypothesis testing might not be appropriate for testing revealed preference theory. In standard hypothesis testing, for example, only two hypotheses are considered and they are treated asymmetrically, i.e., rejection of the behavioral model of choice does not necessarily imply that one should accept the statistical model of choice. One might consider separate families of hypotheses, as in classical non-nested tests, or an alternative approach based on model selection. These approaches are beyond the scope of the present paper. Non-nested hypothesis testing and model selection, however, could confront observed choice data using several behavioral and statistical models.23 22 Echenique et al. ([22], p. 1220) noted that their test procedure had “low power” in terms of the number of violations of GARP: the number of violations under in statistical choices drawn from a uniform distribution (as in Bronars [12]) is “much lower than the number of GARP violations observed in the actual choice data.” 23 It seems possible, for example, to use Vuong [64] non-nested tests to formulate the null hypothesis that the behavioral and the statistical models of choice are indistinguishable against the alternative that one model of choice is closer to the ‘true’ data generating process. 30 6 Some final remarks This paper has rigorously derived a well-behaved consumer demand system from purely random choice behavior. This paper’s findings, colloquially and loosely, represent the economic equivalent of the Shakespearean Monkey Theorem. The interpretation of these findings is open and depends on whether one views the glass as half full or half empty. On one hand, the findings are reassuring. Psychologists and experimental economists have raised countless objections against the consistency of preferences and maximizing behavior. These objections, as valid as they are, are not very forceful since all the testable predictions of the behavioral model of choice can be reached by an alternative route that completely abandons these assumptions. In this defense of the behavioral model, which confirms and extends the one advanced by Becker [9], behavioral objections are largely misguided because the predictive power of demand theory depends primarily on how budget sets change and not on the psychological or neurological process of reasoning involved in an individual’s decision making. In effect, the comparative statics derived in this paper imply that a large class of realistic random choice behaviors rooted in decision theory and aware of systematic departures from individual rationality have the same market predictions of the standard behavioral model of choice. On the other hand, the fact that purely random choices are as predictive as sophisticated rational choices is troublesome; especially for empirical studies of revealed preferences and welfare, and for statistical tests of the behavioral postulates of rational choice theory. The essence of revealed preference theory is to deduce properties of an individual’s preferences from observed choices. But statistical tests of revealed preferences are inherently unrevealing since even a naive (i.e., statistical) model of choice will pass these tests with flying colors. Moreover, and this is a special difficulty for demand integrability and welfare analyses, successfully recovering an individual’s preferences from observed choice data does not necessarily represent a meaningful exercise. Observed choices could, in principle, be fully rational yet completely lack normative content. The present findings, in short, largely diminish the economists’ widely held confidence in the behavioral model of choice. 31 References [1] Agranov, M. and Ortovela P. (2016) “Stochastic Choice and Preferences for Randomization,” Journal of Political Economy, forthcoming. [2] Aguiar, V.H. and Serrano, R. (2014) “Slutsky Matrix Norms and the Size of Bounded Rationality,” Working Paper, Brown University. [3] Anderson, S.P., De Palma, A., and Thisse, J.F. (1992) Discrete Choice Theory of Product Differentiation, Cambridge: MIT Press. [4] Andreoni, J., Gillen, B., and Harbaugh, W.T. (2013) “The Power of Revealed Preference Tests: Ex-Post Evaluation of Experimental Design,” UCSD Working Paper. [5] Apesteguia, J. and Ballester, M.A. (2015) “A Measure of Rationality and Welfare,” Journal of Political Economy, Vol. 123(6), 1278-1310. [6] Aumann, R. (1962) “Utility Theory without the Completeness Axiom,” Econometrica, Vol. 30(3), 445-462. [7] Bandyopadhyay, T., Dasgupta, I., and Pattanaik, P.K. (1999) “Stochastic Revealed Preference and the Theory of Demand,” Journal of Economic Theory, Vol. 84(1), 95-110. [8] Beatty, T.K.M., and Crawford, I.A. (2011) “How Demanding is the Revealed Preference Approach to Demand?,” American Economic Review, Vol. 101(4), 2782-2795. [9] Becker, G.S. (1962) “Irrational Behavior and Economic Theory,” Journal of Political Economy, Vol. 70(1), 1-13. [10] Ben-Akiva, M. and Watanatada, T. (1981) “Application of a Continuous Spatial Choice Logit Model,” in Manski, C.F. and McFadden, D. (Eds.) Structural Analysis of Discrete Choice Data with Econometric Applications, Cambridge: MIT Press. 32 [11] Blundell, R.W., Browning, M., and Crawford, I.A. (2003) “Nonparametric Engel Curves and Revealed Preference,” Econometrica, Vol. 71(1), 205-240. [12] Bronars, S.G. (1987) “The Power of Nonparametric Tests of Preference Maximization,” Econometrica, Vol. 55(3), 693-698. [13] Burghart, D.R., Glimcher, P., and Lazzaro, S.C. (2013) “An Expected Utility Maximizer Walks Into a Bar...,” Journal of Risk and Uncertainty, Vol. 46(3), 215-246. [14] Caplin, A. and Spulber, D.F. (1987) “Menu Costs and the Neutrality of Money,” Quarterly Journal of Economics, Vol. 102(4), 703-725. [15] Chant, J.F. (1963) “Irrational Behavior and Economic Theory: A Comment,” Journal of Political Economy, Vol. 71(5), 505-510. [16] Chen, M. K., Lakshminarayanan, V., and Santos, L.R. (2006) “How Basic are Behavioral Biases? Evidence from Capuchin Monkey Trading Behavior,” Journal of Political Economy, Vol 114(3), 517-537. [17] Choi, S., Kariv, S., Müller, W., and Silverman, D. (2012) “Who Is (More) Rational?,” American Economic Review, Vol. 104(6), 1518-1550. [18] Cox, J.C. (1997) “On Testing the Utility Hypothesis,” Economic Journal, Vol. 107(443), 1054-1078. [19] Dean, M. and Martin, D. (2016) “Measuring Rationality with the Minimum Cost of Revealed Preference Violations,” Review of Economics and Statistics, forthcoming. [20] Demsetz, H. (1996) “Rationality, Evolution, and Acquisitiveness,” Economic Inquiry, Vol. 34(3), 484-495. [21] Duffy, J. (2006) “Agent-Based Models and Human Subject Experiments,” in Tesfatsion, L. and Judd, K.L. (Eds.) Handbook of Computational Economics, Vol. 2: Agent-Based Computational Economics, Amsterdam: Elsevier. 33 [22] Echenique, F., Lee, S., and and Shum, M. (2011) “The Money Pump as a Measure of Revealed Preference Violations,” Journal of Political Economy, Vol. 119(6), 1201-1223. [23] Elster, J. (1989) Solomonic Judgements: Studies in the Limitations of Rationality, Cambridge: Cambridge University Press. [24] Epstein, L.G. (1982) “Integrability of Incomplete Systems of Demand Functions,” Review of Economic Studies, Vol. 49(3), 411-425. [25] Falmagne, J.C. (1978) “A Representation Theorem for Finite Random Scale Systems,” Journal of Mathematical Psychology, Vol. 18(1), 52-72. [26] Fikhtengol’ts, G.M. (1965) The Fundamentals of Mathematical Analysis, London: Pergamon Press. [27] Fisz, M. (1963) Probability Theory and Mathematical Statistics, New York: Wiley. [28] Fudenberg, D., Iijima, R., and Strzalecki, T. (2015) “Stochastic Choice and Revealed Perturbed Utility,” Econometrica, Vol. 83(6), 2371-2409. [29] Gode, D.K. and Sunder, S. (1993) “Allocative Efficiency of Markets with ZeroIntelligence Traders: Market as a Partial Substitute for Individual Rationality,” Journal of Political Economy, Vol. 101(1), 119-137. [30] Gode, D.K. and Sunder, S. (1997) “What Makes Markets Allocationally Efficient?,” Quarterly Journal of Economics, Vol. 112(2), 603-630. [31] Goldberger, A. (1983) “Abnormal Selection Bias,” in Karlin, S., Amemiya, T., and Goodman, L. (Eds.) Studies in Econometrics, Time Series and Multivariate Statistics, New York: Academic Press. [32] Gonzalez-Vallejo, C. (2002) “Making Trade-Offs: A Probabilistic and Context- Sensitive Model of Choice Behavior,” Psychological Review, Vol. 109(1), 137-155. 34 [33] Grandmont, J.M. (1992) “Transformation of the Commodity Space, Behavioral Heterogeneity, and the Aggregation Problem,” Journal of Economic Theory, Vol. 57(1), 1-35. [34] Gul, F. and Pesendorfer, W. (2006) “Random Expected Utility,” Econometrica, Vol. 74(1), 121—146. [35] Harbaugh, W., Krause, K., and Berry, T. (2001), “GARP for Kids: On the Development of Rational Choice Behavior,” American Economics Review, 91(5), 1539-1545. [36] Hildenbrand, W. (1994) Market Demand: Theory and Empirical Evidence, Princeton: Princeton University Press. [37] Kagel, J.H., Battalio, R.C., and Green, L. (1995) Economic Choice Theory: An Experimental Analysis of Animal Behavior, Cambridge: Cambridge University Press. [38] Katzner, D.W. (1970) Static Demand Theory, New York: MacMillan. [39] Khuri, A.L. (2002) Advanced Calculus with Applications in Statistics, New York: Wiley. [40] Kneip, A. (1999) “Behavioral Heterogeneity and Structural Properties of Aggregate Demand,” Journal of Mathematical Economics, Vol. 31(1), 49-79. [41] LaFrance, J.T. and Hanemann, W.M. (1989) “The Dual Structure of Incomplete Demand Systems,” American Journal of Agricultural Economics, Vol. 71(2), 262-274 [42] Lewbel, A., (2001) “Demand Systems with and without Errors,” American Economic Review, Vol. 91(3), 611-618. [43] Luce, R.D. (1959) Individual Choice Behavior: A Theoretical Analysis, New York: Wiley. [44] Machina, M.J. (1985) “Stochastic Choice Functions Generated from Deterministic Preferences over Lotteries,” Economic Journal, Vol. 95, 575-594. 35 [45] Mandler, M. (2004) “Indifference and Incompleteness Distinguished by Rational Trade,” Games and Economic Behavior, Vol. 67(1), 300-314. [46] Manzini, P. and Mariotti, M. (2014) “Stochastic Choice and Consideration Sets,” Econometrica, Vol. 82(3), 1153—1176. [47] Mas-Colell, A., Whinston, M. and Green, J. (1995) Microeconomic Theory, Oxford: Oxford University Press. [48] McFadden, D. (1976) “The Mathematical Theory of Demand Models,” in Stopher, P.R. and Meyburg, A.H. (Eds.) Behavioral Travel Demand Models, Lexington, MA: Lexington Books. [49] McFadden, D. (2005) “Revealed Stochastic Preference: A Synthesis,” Economic Theory, Vol. 26(2), 245-264. [50] McFadden, D. and Richter, M.K. (1990) “Stochastic Rationality and Revealed Stochastic Preference,” in Chipman, J., McFadden, D., and Richter, M.K. (Eds.) Preferences, Uncertainty, and Optimality, Essays in Honor of Leo Hurwicz, Boulder: Westview Press. [51] McGrath, M. A. (1989) “An Ethnography of a Gift Store: Wrappings, Trappings and Rapture,” Journal of Retailing, Vol. 65(4), 421-449. [52] Moore, O.K. (1957) “Divination—A New Perspective,” American Anthropologist, Vol. 59(1), 69-74. [53] Mossin, A. (1968) “Elements of a Stochastic Theory of Consumption,” Swedish Journal of Economics, Vol. 70(4), 200-220. [54] Roe, R.M., Busemeyer, J.R., and Townsend, J.T. (2001) “Multialternative Decision Field Theory: A Dynamic Connectionist Model of Decision Making,” Psychological Review, Vol. 108(2), 370-392. 36 [55] Sanderson, W.C. (1974) “Does the Theory of Demand Need the Maximum Principle?,” in David, P.A. and Reder, N.W. (Eds.) Nations and Households in Economic Growth: Essays in Honor of Moses Abramovitz, New York: Academic Press. [56] Sanderson, W.C. (1980) “A Nonutilitarian Economic Model of Fertility and Female Labor Force Participation,” Revue économique, Vol. 31(6), 1045-1080. [57] Shaked, M. and Shanthikumar, J.G. (2007) Stochastic Orders, New York: Springer. [58] Simon, H.A. (1957) Models of Man. Social and Rational, New York: Wiley. [59] Tversky, A. (1969) “Intransitivity of Preferences,” Psychological Review, Vol. 76(1), 31-48. [60] Tversky, A. (1972) “Elimination by Aspects: A Theory of Choice,” Psychological Review, Vol. 79(4), 281-299. [61] Varian, H.R. (1982) “The Nonparametric Approach to Demand Analysis,” Econometrica, Vol. 50(4), 945-973. [62] Varian, H.R. (1990) “Goodness-of-Fit in Optimizing Models,” Journal of Econometrics, Vol 46(1-2), 125-140. [63] Varian, H.R. (2006) “Revealed Preference,” in Szenberg, M., Ramrattan, L., and Gottesman, A. (Eds.) Samuelsonian Economics and the Twenty-First Century, pages 99-115. Oxford University Press. [64] Vuong, Q.H. (1989) “Likelihood Ratio Tests for Model Selection and Non-nested Hypotheses,” Econometrica, Vol 57(2), 307-333. 37 7 Appendix A: Omitted proofs and derivations Omitted derivations from the simple example. Consider first the Slutsky equation for 1 . Notice that the Slutsky compensation at income is = ̄1 ( )1 . The Slutsky equation (9) follows from the definition. For completeness, notice that ¯ max ¯ X max 1 ¯ = X 1 + ̄1 ( ) . 1 ¯ 1 1 The Slutsky equation (9) is a consequence of this expression as well as (5) and (10). Next consider the mean demand for 2 . Some steps in this case are analogous to the proof of Proposition 1. Consider first the own-price effect. Using equation (3), one has: with µ ¯ ¯ ¶ ¯ µ ̄1 ( ) ¯¯ 1 ̄1 ( )1 ̄2 ( ) ¯¯ (2 ) ¯¯ − = − , 2 ¯ 2 ¯ 2 ¯ 2 22 ¯ ¶ ¯ ̄1 ( )1 ̄1 ( ) ¯¯ E [1 | 0 ≤ 1 ≤ X max (2 ) ¯¯ 1 ] ̄2 ( ) = , = , and max 2 ¯ ¯ 2 2 2 X 1 1 ¯ µ ¶ ̄2 ( ) E [1 | 0 ≤ 1 ≤ X max ̄2 ( )¯¯ 1 ] 0. =− 2 ¯ X max 2 1 The Slutsky equation for 2 follows from simple rearrangements. For instance, notice that the uncompensated own-price effect is ̄2 ( )2 = −̄2 ( )2 which rules out Giffen goods. Income effects, however, can be negative for ̄2 ( ). That is, ̄2 ( ) = ¶ µ 1 E [1 | 0 ≤ 1 ≤ X max 1 ] . 1− max X 1 2 (A1) Corollary 4 Under the conditions of Proposition 1, suppose (1 ) is a log-concave distribution. Then, ̄2 ( ) is a normal good. A1 Proof. The proof relies on (A1) and the following fact about log-concave distributions: 0≤ E [1 | 0 ≤ 1 ≤ X max 1 ] ≤1, max X 1 see Goldberger ([31], Appendix A). Moreover, notice that |S( )| can be written as |S( )| = ̄2 ( ) µ E [1 | 0 ≤ 1 ≤ X max 1 ¯ ¯ ¶ ¯ X max 1 ] ¯ −2 21 ¯ ¯ 11 ¯ ¯ 11 ¯ ¯ , ¯ −12 ¯ with |S( )| = 0. Symmetry and negative semidefiniteness always hold in the special case of two-commodities. Integrability follows trivially from these properties; see Katzner ([38], Theorem 4.1-2). Omitted derivations from the general framework. In order to simplify the derivations in this section, I next present an auxiliary Lemma that will be used repeatedly. The Lemma serves to simplify the calculations of price and income effects. The central implication of the following Lemma is that in order to determine the response in mean demands to changes in prices and income, one only needs to evaluate the changes in the most interior integral. Economically, this result makes sense: recall that X max (() ) = − () · () , which is the only maximum feasible consumption that depends on the entire price vector and income . Lemma 1 Let ( ) ≡ Z 0 max 1 Z max (() ) (), where X max is given in (14). Then, 0 "Z max #¯ ¯ Z max ¯ (() ) 1 ( )¯¯ ¯ = () ¯ () , ¯ ¯ 0 0 ¯ ¶ µ Z max Z max 1 −1 ((−1) ) ¯ X max (() )¯ max = (() X (() )) ¯ () , 0 0 for = 1 . Proof. The proof is a repeated application of Leibniz’s rule for differentiation under the integral sign. The case of = 2 and 1 = 2 = 1 is available in Khuri ([39], pp. 307-308). A2 For the general proof, it is enough to consider changes in 1 since this price enters in all limits of integration. All other price changes can be seen as special cases. # "Z max ¯ ¯ Z max 2 (1 ) (1 −1 ) ¯ ( )¯¯ X max 1 ¯ = ( ) 1 1 1 ¯ 1 ¯ 0 0 1 = max 1 "Z max #¯ Z max Z max ¯ 2 (1 ) (1 −1 ) 1 ¯ + (1 )2 ¯ 1 . ¯ 1 0 0 0 max The first term is zero since X max 2 (X 1 ) = ( − 1 (1 ))2 = 0. Further, as noted max max in the text, since X max = 0 for ≥ 2 also as there would be no 2 (X 1 ) = 0, then X income left for the consumption of these commodities. The second term becomes Z max 1 0 Z + 0 ⎧" ⎨ Z ⎩ max (1 2 ) 3 0 Z max max (1 ) 1 2 0 Z max (1 −1 ) ()2 ⎫ ¯ ⎬ ¯ ⎭ 1 ¯ X max 2 (1 )¯ 1 #¯ Z max ¯ max (1 2 ) 3 (1 −1 ) ¯ ()3 ¯ 1 2 , ¯ 0 0 0 1 # "Z 2 = max (1 ) 2 max max max whose first component is also zero as 3 X max 3 (1 X 2 (1 )) = 2 (X 2 (1 ) − X 2 (1 )) = 0. Thus, as before, the first component evaluates a definite integral over a degenerate interval and this equals zero. The only relevant component is the second one which also needs to be evaluated using Leibniz’s rule. By the way the limits of integration are defined, the derivative operator moves toward high values of . For instance, the step of the sequence of integrals is given by ¯ ( )¯¯ = 0 + + 0 ( − 1 times) 1 ¯ "Z max #¯ Z max Z max ¯ (1 −1 ) 1 (1 −1 ) ¯ () ¯ 1 −1 , + ¯ 1 0 0 0 where the evaluation of the integral with the upper limit of integration X max +1 (1 ) evaluated at = X max (1 −1 ) will also equal zero. Moreover, X max = 0 for all ≥ . A3 The last term in the sequence is "Z max #¯ ¯ Z max ¯ (1 −1 ) 1 ( )¯¯ ¯ = () ¯ 1 −1 , ¯ ¯ 1 1 0 0 which under Leibniz’s rule simply becomes ¯ ¯ ¶ µ Z max Z max 1 −1 (1 −2 ) ¯ X max ( )¯¯ (() )¯ max = ( X ( )) () () ¯ () . 1 ¯ 1 0 0 Proof of Theorem 1. The proof was started in the text and is completed here. To obtain (16), consider (15) and write this expression as ̄ ( ) = ( ) (X max ), with ( ) ≡ (X max Z max 1 max (() ) () , and 0 0 )= Z Z max 1 0 Z max (() ) () , 0 as the numerator and the denominator respectively. Using Lemma 1, notice that ¯ ¯ ¶ µ Z max Z max 1 (() ) ¯ X max ( )¯¯ (() )¯ max = ( X ( )) () () ¯ () , ¯ 0 0 ¯ ¯ ¶ µ Z max Z max 1 (() ) ¯ X max (X max )¯¯ (() )¯ max = (() X (() )) ¯ ¯ () . 0 0 Further, notice that the quotient rule implies ¯ ¯ µ ¯ µ ¶ ¶ 1 ( ) (X max )¯¯ ( )¯¯ ̄ ( )¯¯ , − = ¯ ¯ ¯ (X max ) (X max )2 which can be written in simple terms as ¯ ¯ ¯ ¸ ¶∙ µ ( )¯¯ 1 (X max )¯¯ ̄ ( )¯¯ . = − ̄ ( ) ¯ ¯ (X max ) ¯ A4 This last expression, upon substitution, implies ¯ ¯ ¶ µ max Z max Z max 1 −1 ((−1) ) ¯ (() ) X max ̄ ( )¯¯ (() )¯ = { − ̄ ( )} ¯ () . ¯ (X max ) 0 0 (A2) where, as in the text, max (() ) ≡ (() X max (() )). A compensated change in only needs to be evaluated in terms of its effect on X max (() ). This result implies that price effects are channeled through changes in the maximum feasible consumption. Recall that feasibility implies X max (1 −1 ) = ´ X−1 1 ³ − , =1 (A3) and that income is compensated according to = ̄ ( ) . Then, ¯ ∙µ ¶ ¸ ¯ X max ̄ ( ) − 1 (() ) ¯ − = = . ¯ (A4) Finally, substitution of (A4) into (A2) yields (16). For commodity , (16) implies that the own price effect is given by ¯ µ ¶ Z max Z max 1 −1 ((−1) ) © ª2 max (() ) 1 ̄ ( )¯¯ max () . =− X (() ) − ̄ ( ) ¯ (X max ) 0 0 (A5) The main difference is that has been substituted by X max (() ), which is a (hyper) surface along () .24 To obtain income effects (18), one simply needs to evaluate the appropriate derivative of X max (() ). For instance ( ) = Z 0 max 1 Z max (() ) (() X max (() )) 0 µ ¶ X max (() ) () , and similarly for the denominator. Once these expressions are substituted back into the quotient rule result, one obtains (18). 24 Notice that in the example of Section 2, the own-price effect (5) can be written as ̄1 ( )1 | = −{X max − ̄1 ( )}2 (X max ) (X max )1 , which is the analog of (A5). 1 1 1 A5 Finally, the cross-partial effects (A6) and (A7) also follow using the appropriate derivative of X max (() ). For instance ¯ ¯ ¶ µ Z max Z max max 1 (() ) ¯ X ( ) ( )¯¯ () max ¯ () , = ( X ( )) () () ¯ ¯ 0 0 and similarly for the denominator. These calculations yield ¯ ¯ ¶ µ Z max Z max 1 −1 ((−1) ) ¯ (() ) X max ̄ ( )¯¯ (() )¯ = { − ̄ ( )} ¯ () , ¯ (X max ) 0 0 (A6) where the compensated change in X max (() ) is the same as the one obtained in (A4). Next consider the response of ̄ ( ) to a compensated change in . Following Lemma 1, and repeating the steps just taken, one can show that: ¯ ¯ ¶ µ Z max Z max 1 −1 ((−1) ) ¯ (() ) X max ̄ ( )¯¯ (() )¯ = { − ̄ ( )} max ¯ () , ¯ (X ) 0 0 (A7) ¯ ¯ where X max (() ) = (̄ ( ) − ) . Simple substitutions establish symmetry. The only property that deserves some further comment is the negative semidefiniteness of S( ), which is analogous to the positive semidefiniteness of Σ. I mentioned before that S ( ) can be written as a covariance term given by [ ] = Z max 1 Z max −1 ((−1) ) 0 0 { − ̄ ( )} { − ̄ ( )} ˆ(() )() , where ˆ(() ) is a simple correction since the density (() X max (() )) does not integrate to one. That is, the normalization ˆ(() ) ≡ (() X max (() ))( ) with ( ) ≡ 1 (X max ) ÃZ 0 max 1 Z max −1 ((−1) ) 0 (() X max (() ))() ! , ˆ () ) as a proper density in order to define the matrix of second moments. serves to treat ( Notice that the term ( ) is a positive constant and hence it does not influence any of A6 the conclusions of the analysis. To complete the proof, one only needs to consider standard properties of the variancecovariance matrix. Consider, for instance, a vector and the non-negative quadratic form [ · ( − ̄( ))]2 = · ( − ̄( ))( − ̄( )) · . The conditional expectation (20) satisfies E[ · ( − ̄( ))]2 = · Σ , which only takes non-negative values. Thus Σ is positive semidefinite which is a well-known property of variance-covariance matrices; see Fisz ([27], pp. 89-90). To verify the conditions for positive definiteness, one only needs to check that the “last” commodity is linearly independent from the remaining − 1 commodities because these other commodities have a joint density, i.e., their distribution is non-degenerate; see Fisz ([27], p. 90). Notice that X max (() ) is a linear function of () . That is, for a given realization of demands () , the maximum feasible consumption X max (() ) is linear in () , i.e., X max (() ) = − () · () for all possible realizations of () ; see (14). Notice, however, that is not necessarilly equal to X max (() ). In particular, = X max (() ) for all possible realizations of () only when is selected as a residual. In such case, ̄ ( ) will be linearly related to ̄() ( ) since ̄ ( ) = − () · ̄() ( ). In other words, if is a residual, then Σ is positive semidefinite but not positive definite since · Σ = 0 holds for some 6= 0.25 Preliminaries for integrability conditions. Katzner ([38], Chap. 4) discusses in detail the conditions for integrability in a complete demand system. The integrability of incomplete demand systems is based on Epstein [24]. In terms of notation, let ( ) denote the expenditure function: ( ) ≡ min≥0 { · : () ≥ }. Thus, ( ) = ( ) , (A8) for = 1 and with ( ) as the Hicksian demand for , i.e., ( ) = ̄ ( ( )). Consider the case of = 2. Then, X max (1 ) − ̄2 ( ) = ( − 1 1 )2 − ̄2 ( ). Assume 2 = 2 for all possible realizations of 1 . In this case, 2 is determined as a residual which is the example of Section 2. This implies that ̄2 ( ) = ( − 1 ̄1 ( ))2 ; see, e.g., (3). Then X max (1 ) − 2 ̄2 ( ) becomes −(1 2 )(1 − ̄1 ( )) which is a linear function of (1 − ̄1 ( )). In this case the variance-covariance matrix will be positive semidefinite but not positive definite. 25 X max (1 ) 2 A7 Proof of Theorem 2. Consider first part (b). Symmetry in S( ) is necessary and sufficient for the existence of a solution for the partial differential equation system (A8); see, e.g., Epstein [24]. A negative semidefinite S( ) is necessary and sufficient for the solution of the previous system to be concave in . As Theorem 1 noted, these requirements are met if is determined as a residual. (The utility function can then be recovered from the expenditure function.) For part (a), integrability in incomplete systems requires a negative definite Slutsky matrix. As Theorem 1(a) noted, interior demands satisfy this condition here. The constant of integration in the previous system cannot be uniquely determined in an incomplete system so one must make assumptions about how to complete the system. Assuming the existence of 0 with 0 = 1 accomplishes this in the simplest possible way. In general, let ̄0 ( 0 ) denote the vector of demands for commodities that complete the demand system and let 0 represent their corresponding vector of prices. If ̄( 0 ) = ̄( ) and ̄0 ( 0 ) = ̄0 (0 ), then there is a utility function : R+ → R such that ̄( ) + and ̄0 (0 ) are the solution to max0 ≥0 {( Ψ(0 )) : · + 0 · 0 = }, with Ψ(0 ) linearly homogeneous. Epstein [24] and LaFrance and Hanemann [41] provide additional remarks about the integrability of incomplete demand systems. Proof of Corollary 1. Representing individual demands as a mean-plus-noise random variable = ̄( ) + is standard. Expression (23) in Corollary 1 is simply Chebyshev inequality. Derivation of binary choice probabilities. Statistical choices can be ordered as follows: ≡ Pr{ } = Pr{ } + Pr{ } + Pr{ } , ≡ Pr{ } = Pr{ } + Pr{ } + Pr{ } , ≡ Pr{ } = Pr{ } + Pr{ } + Pr{ } . Then, − = Pr{ } − Pr{ } and − = Pr{ } − Pr{ } which are the statements used in footnote 14 in the text. A8 Proof of Corollary 2. The proof of Corollary 2 follows from the one-sided Cheby- shev inequality (28). Further, lim →∞ Pr{̃+ + | } ≡ lim→∞ Pr{max{0 ̃ } + | }. The limit for the + -statistic follows from the continuous mapping theorem, i.e., lim→∞ Pr{max{0 ̃ } + | } = max{0 Pr{lim→∞ ̃ } + | } = 0. (The max function is continuous although not differentiable.) Proof of Corollary 3. Under , the power components in (29) converge (strongly) to 0 lim Π̃ + { | } = 1 − →∞ (̄( )) (̄( )) − (0) , (A9) max 00 , and lim Π̃− {| } = →∞ (X ) (X max ) where X max 00 corresponds to a budget set (0 0 ). The rest of the proof follows from simple rearrangements. A9 8 Appendix B: Additional results [NOT FOR PUB- LICATION] Random (“dual”) representations. Theorem 2 show that mean demands that arise from individuals randomly choosing their consumptions can be represented as demands that arise from the maximization of some utility function. This sub-section provides a partial examination of the “converse” problem and provides two random or “irrational” representations of behavioral demands. The first representation assumes that “irrational” demands are on average feasible whereas the second is individually feasible but it restricts the support of the distribution. Consider a given consumer demand function ( ) for commodity with ( ) ≡ (1 ( ) ( )). Demands ( ) are homogeneous in ( ) and satisfy ·( ) = . These are the basic properties of demand functions in the behavioral model of choice. (i) Average feasibility. This feasibility condition is the same as the one used by Katzner [38] to discuss errors and shocks to individual demand functions. In particular, as in Katzner ([38], p. 161), “if [the consumer] were to choose from the same budget set many times, ‘on average’ he would choose the utility maximizing bundle.” Lemma 2 Let z ≡ and z ≡ (z 1 z ) denote the set of extreme points of the budget set (12). Then, ( ) = (z). Proof. This is just a consequence of homogeneity: demands satisfy (1 ) = (1 1). The goal in the first representation is to find a distribution function ∗ () such that (z) = Z 0 1 Z 0 ∗ () , ∗ (z) (B1) for all = 1 where ∗ (z) ≡ ∗ (z 1 z ) and ∗ () is the joint density of ∗ (). Notice that (B1) integrates random individual demands over a rectangular area instead of B1 over the triangular area given by (12). The derivations needed to construct the first representation rely on an inversion formula for the right-truncated mean. Write (B1) as ∗ (z) (z) = Z 0 µ ¶ ∗ (z 1 z −1 z +1 z ) , which on differentiation with respect to z yields ( (z)z ) ∗ (z)+ (z)( ∗ (z) ) = z ( ∗ (z) ). Let (z) ≡ (z)z denote the derivative function of (z). A con- venient way to write the previous expression is ln ∗ (z) = (z)[z − (z)]. Integration yields ½ Z (z () ) = Φ (z () ) exp − +∞ ∗ µ ¶ ¾ (z () ) , − (z () ) (B2) where ∗ (z () ) ≡ ∗ (z 1 z −1 z +1 z ) and similarly for (z () ) and (z () ), and with Φ (z () ) = Φ (z 1 z −1 z +1 z ) determined such that ∗ () is a distribu- tion function. This distribution function ∗ () can be found by solving the simultaneous equations (B2). (ii) Individual feasibility. A second representation can be constructed using results from convex analysis. Every point in a convex and compact set of finite dimension can be written as a convex combination of the extreme points of the set; see, e.g., Phelps ([2], p. 1). Let be a compact convex subset of R and let z denote the extreme points of . Every point ∈ can be written as a convex combination of z: X = z , =1 where ≥ 0 and P =1 = 1. There is a probabilistic interpretation based on Minkowski’s integral representation; see Phelps [2]: Theorem 3 (Minkowski) Let be a compact convex subset of R . Then, for every point ∈ , there exists a probability measure concentrated in the extreme points of such R that = (). If is a simplex, is unique. B2 Proof. Let be the Dirac measure on the point z , i.e., for every Borel set B, P (B) = 1 if z ∈ B and zero otherwise. Then, () = =1 = 1 and is a measure R R P P with support {z 1 z }. Therefore, () = =1 () = =1 z = . Uniqueness can be established generally for the case of a simplex; see Phelps ([2], Chap. 10). To apply the previous representation it is enough to use Lemma 2 and to notice that the budget set (12) is convex and compact; see Mas-Colell et al. ([47], p. 22). This means that it is possible to construct a probability measure concentrated in z that is always feasible and represents any behavioral demand function (z) as the mean demand from a random choice procedure that selects bundles on the budget hyperplane. Supply and market equilibrium. In this sub-section, I present a simple partial equilibrium analysis in competitive markets to outline the nature of results that random choice may yield in market situations. The purpose of this section is to show that the nice properties of random choice also apply to the analysis of supply curves.26 Consider the market of commodity 1 . The demand side is determined by Proposition 1 and the supply side can be derived based on maximizing behavior or as the outcome of randomization. Let denote the amount of output produced of this commodity. For a given price , profits are − − (), where and () denote the fixed cost and the variable cost function respectively. Factor prices are held constant and costs are wellbehaved. Let Qmin () denote the minimum scale (i.e., the “shut down” point) determined by = (Qmin ())Qmin (). Notice that Qmin () is positive along increasing average costs. Suppose that firms randomly select production levels using a density function (). Each randomly selected production plan must satisfy ≥ Qmin (). Let ̄() denote the mean supply curve. This curve results from the truncation of unfeasible production plans: min ̄() ≡ E[| ≥ Q ()] = 26 Z +∞ () min () 1 − (Qmin ()) , (B3) These discussions are motivated in part by Kirzner’s [1] earlier criticisms to models of irrational behavior. Essentially, Kirzner [1] suggested that leaving out the supply side of the market considerably weakens any conclusion one can draw about random choices. B3 min where 1 − (Q ()) ≡ Z +∞ (). Thus any production plan in [Qmin () +∞) may min () be selected depending on () but any production plan in [0 Qmin ()) must be discarded. The comparative statics properties with respect to prices satisfy: ¶ µ ̄() Qmin () (Qmin ()) min = [̄() − Q ()] 0. 1 − (Qmin ()) (B4) The first two terms in (B4) are positive due to the statistical properties of the lefttruncated mean formula and have interpretations that are analogous to those given in Section 2. The last term contains all the economic information needed to determine the slope of the supply curve under randomization. As I noted before, this term is positive if average costs are increasing and so in this case the mean supply curve will be positively sloped even in the absence of maximizing behavior. For instance, under profit maximization, firms will choose to produce at a level ∗ () such that = ( ∗ ) as long as ∗ ≥ Qmin (); see, e.g., Mas-Colell et al. ([47], Section 5.D). Proposition 1 and (B4) also imply that there is a single equilibrium in the market for commodity 1 which can be identified as the intersection point of the mean demand and supply curves. This finding confronts the criticism of Kirzner [1] and it implies that changes in underlying market conditions affect equilibrium outcomes in the expected way even under “irrational behavior.” Simulation exercises. A key feature in all previous derivations is that mean demand curves are defined from the aggregation of individual choices. In here, I explore some simple simulation exercises whose purpose is to determine how ‘large’ the economy needs to be in order to observe consistent results. I consider a simple uniform distribution and a multivariate log-normal distribution. (i) Uniform distribution. Assume two goods and non-interior choices. This is the setting used in the simple example of Section 2. Assume also that 2 and are constant throughout. At a fixed price level 1 let 1 , = 1 be an i.i.d. sequence of uniform random variables on [0 1 ]. Each represents an individual realization of demand for B4 good 1 . The mean demand is: ̃1 = −1 X 1 , =1 which, by the strong law of large numbers, satisfies ̃1 → ̄1 ( ). Moreover, when prices 1 change, one can trace an uncompensated demand curve whose elasticity should be ε = −1. In the following simulations = 1 and 1 varies from 1 to 2. The results consider two incremental steps. The first is 0005 and the second 005. This means that each individual has 200 realizations of demand in the first case and 20 realizations in the second case. Mean demand is computed for different values of that range from = 1 to = 1 000. In each sample, and for each value of , a log-log linear regression estimates the uncompensated elasticity of the demand curve. The number of cross-samples is 500. Table B1. Simulation results for uniform distribution. Number of individuals aggregated =1 5 10 25 50 100 500 1 000 A. Grid size for 1 of 200 sample points ε̂ −1.0086 −1.0095 −1.0061 −1.0005 −1.0011 −1.0015 −1.0007 −1.0001 std.err. (0.349) (0.100) (0.067) (0.041) (0.029) (0.020) (0.009) (0.006) std.dev. [0.345] [0.102] [0.067] [0.040] [0.029] [0.020] [0.008] [0.006] 2 R 0.0396 0.3343 0.5242 0.7423 0.8538 0.9219 0.9834 0.9917 [0.027] [0.054] [0.043] [0.026] [0.014] [0.008] [0.001] [0.0008] B. Grid size for 1 of 20 sample points ε̂ −1.101 −0.9901 −0.9966 −0.9977 −0.9992 −1.0013 −0.9998 −1.0002 std.err. (0.954) (0.276) (0.188) (0.118) (0.083) (0.058) (0.026) (0.018) std.dev. [1.090] [0.297] [0.208] [0.125] [0.085] [0.061] [0.027] [0.018] 2 R 0.0562 0.3586 0.5497 0.7633 0.8672 0.9316 0.9857 0.9928 [0.109] [0.160] [0.146] [0.075] [0.043] [0.022] [0.004] [0.002] Note: The number of cross-samples is 500. The elasticities are estimated using a linear fit to log[̃1 ] and log[1 ]. The average value of the standard errors across samples is in parentheses. The cross-sample standard deviation for the estimate of the elasticity of demand and the R2 is in brackets. Table B1 shows that ‘individual’ demands are on average negatively sloped with an B5 elasticity consistent with the predicted pattern. The estimates of the elasticity, however, are unreliable when only one individual realization is considered and the sample variation in prices is small. In fact, the elasticity cannot be statistically distinguished from zero in this case. Further, the standard deviation across and between samples is 1.090 and 0.954 respectively. Both estimates suggest that some individuals have a positively-sloped demand curve. Finally, the goodness of fit from the R2 is, on average, about 5 percent. Thus, overall, individual demands are not consistently determined. Consider the two cases in which = 5. In these cases, the goodness of fit increase to over 30 percent and the estimates of the elasticity of demand become (statistically) close to ε̂ = −1. When the sample variation in prices is 200, and = 10, the across sample standard deviation for the elasticity of demand is about 0.06. The goodness of fit and the statistical significance of the estimates also suggest that average demands are precisely estimated. When only 20 sample points for 1 are considered, a similar conclusion follows if the aggregation takes place for 50 individuals. As expected, with = 500 or = 1000, the goodness of fit is well over 98 percent and the elasticity is precisely estimated up to three digits. (ii) Multivariate log-normal distribution. Consider next three goods and non-interior choices. Prices are 1 , 2 , and 3 . Prices 2 = 1 and 3 = 1, and = 3 are constant throughout. 1 varies from 1 to 2. Let (1 2 3 ), = 1 be an i.i.d. sequence of log-normal random variables on ( ). Each represents an individual realization of demand for the three goods 123 . The draws are from a log-normal distribution with mean [−05 −05 −05], variances [2 2 2] and with positive pairwise correlations between all variables of 05. In the following simulations 1 varies in incremental steps of 005, so there are 20 sample points. Mean demand is computed for different values of that range from = 1 to = 10 000. In each sample, and for each value of , a log-log linear regression estimates the compensated elasticity of the demand curve, ε . This elasticity should be negative, although there are no specific numerical values suggested by the theory. Again, the number of cross-samples is 500. Table B2 shows again that ‘individual’ demands are negatively sloped but unreliable. B6 As the number of individuals aggregated increases, mean demands become better behaved especially when more than 100 individuals are considered. The goodness of fit from the R2 increases and the precision of the estimated elasticity of demand also increases. The compensated elasticity is negative, as predicted by Theorem 1. Table B2. Simulation results for multivariate log-normal distribution. Number of individuals aggregated ε̂ std.err. std.dev. R2 =1 10 100 1 000 10 000 −0.0965 −0.2325 −0.2598 −0.2524 −0.2541 (0.723) (0.363) (0.107) (0.0384) (0.022) [2.301] [0.614] [0.184] [0.056] [0.018] 0.0813 0.0869 0.222 0.656 0.858 [0.162] [0.152] [0.213] [0.134] [0.036] Note: The number of cross-samples is 500. The elasticities are estimated using a linear fit regression. The average value of the standard errors across samples is in parentheses. The crosssample standard deviation for the estimate of the elasticity of demand and the R2 is in brackets. In conclusion, Tables B1 and B2 show that consistent mean demand curves do not require an excessively large number of individuals aggregated; this is specially true with enough price variation. That is, while the theoretical analyses rely on a large number of individuals, numerical approximations with fewer individuals still give support to the main theoretical conclusions of the analysis. References [1] Kirzner, I.M. (1962) “Rational Action and Economic Theory,” Journal of Political Economy, Vol. 70(4), 380-385. [2] Phelps, R.P. (2001) Lectures on Choquet’s Theorem, Berlin: Springer. B7
© Copyright 2026 Paperzz