Time Discounting and Time Preference: A Critical Review
Author(s): Shane Frederick, George Loewenstein and Ted O'Donoghue Source: Journal of Economic Literature, Vol. 40, No. 2 (Jun., 2002), pp. 351-401 Published by: American Economic Association Stable URL: http://www.jstor.org/stable/2698382 . Accessed: 10/04/2014 02:22 Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp . JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org. . American Economic Association is collaborating with JSTOR to digitize, preserve and extend access to Journal of Economic Literature. http://www.jstor.org This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Journal of Economic Literature Vol. XL (June 2002), pp. 351-401 T ime Discounting and T ime Preference: A C ritical R ev iew SHA NE FR EDER IC K, GEOR GE LOEWENST EIN, and T ED O'DONOGHUE' 1. Introd uction I NT ER T EMPOR A L C HOIC ES-d ecisions inv olv ing trad eoffs among costs and benefits occurring at d ifferent times- are important and ubiquitous. Such d eci- sions not only affect one's health, w ealth, and happiness, but, may also, as A d am Smith first recognized , d etermine the economic prosperity of nations. In this paper, w e rev iew empirical research on intertemporal choice, and present an ov erv iew of recent theoretical formula- tions that incorporate insights gained from this research. Economists' attention to intertempo- ral choice began early in the history of the d iscipline. Not long after A d am Smith called attention to the impor- tance of intertemporal choice for the w ealth of nations, the Scottish economist John R ae w as examining the sociologi- cal and psychological d eterminants of these choices. In section 2, w e briefly rev iew the perspectiv es on intertempo- ral choice of R ae and nineteenth- and early tw entieth-century economists, and d escribe how these early perspectiv es interpreted intertemporal choice as the joint prod uct of many conflicting psychological motiv es. A ll of this changed w hen Paul Sam- uelson proposed the d iscounted -utility (DU) mod el in 1937. Despite Samuel- son's manifest reserv ations about the normativ e and d escriptiv e v alid ity of the formulation he had proposed , the DU mod el w as accepted almost in- stantly, not only as a v alid normativ e stand ard for public policies (e.g., in cost- benefit analyses), but as a d escriptiv ely accurate representation of actual behav - ior. A central assumption of the DU mod el is that all of the d isparate mo- tiv es und erlying intertemporal choice can be cond ensed into a single parameter- the d isc.ount rate. In section 3 w e exam- ine this and many other assumptions und erlying the DU mod el. We d o not present an axiomatic d eriv ation of the mod el, but instead focus on those features that highlight the implicit psychological assumptions und erlying the mod el. 1 Fred erick: Sloan School of Management, Mas- sachusetts Institute of T echnology. Loew enstein: Department of Social and Decision Sciences, C arnegie Mellon Univ ersity. O'Donoghue: De- partment of Economics, C ornell Univ ersity. We thank C olin C amerer, Dav id Laibson, John McMillan, Drazen Prelec, Daniel R ead , Nachum Sicherman, Duncan Simester, and three anony- mous referees for useful comments. We thank C ara. Barber, R osa Blackw ood , Mand ar Oak, and R osa Stipanov ic for research assistance. For finan- cial support, Fred erick and Loew enstein thank the Integrated Stud y of the Human Dimensions of Global C hange at C arnegie Mellon Univ ersity (NSF Grant SBR -9521914), and O'Donoghue thanks the National Science Found ation (A w ard SE'S-0078796). 351 This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 352 Journal of Economic Literature, Vol. XL (June 2002) Samuelson's reserv ations about the d escriptiv e v alid ity of the DU mod el w ere justified . Section 4 rev iew s the grow ing list of "DU anomalies"- patterns of choice that are inconsistent w ith the mod el's theoretical pred ic- tions. Virtually ev ery assumption und er- lying the DU mod el has been tested and found to be d escriptiv ely inv alid in at least some situations. Moreov er, as w e d iscuss at the end of the section, these anomalies are not anomalies in the sense that they are regard ed as errors by the people w ho commit them. Unlike many of the better-know n expected - utility anomalies, the DU anomalies d o not necessarily v iolate any stand ard or principle that people believ e they should uphold . T he insights about intertemporal choice gleaned from this empirical re- search hav e led to the proposal of nu- merous alternativ e theoretical mod els, w hich w e rev iew in section 5. Some of these mod ify the d iscount function, per- mitting, for example, d eclining d iscount rates or "hyperbolic d iscounting." Oth- ers introd uce ad d itional arguments into the utility function, such as the utility of anticipation. Still others d epart from the DU mod el more rad ically, by in- clud ing, for instance, systematic mis- pred ictions of future utility. Many of these new theories rev iv e psychological consid erations d iscussed by R ae and other early economists that w ere extin- guished w ith the ad option of the DU mod el and its expression of intertem- poral preferences in terms of a single parameter. In section 6, w e rev iew attempts to estimate d iscount rates. While the DU mod el assumes that people are charac- terized by a single d iscount rate, this literature rev eals spectacular v ariation across (and ev en w ithin) stud ies. T he failure of this research to conv erge to- w ard any agreed -upon av erage d iscount rate stems partly from d ifferences in elicitation proced ures. But it also stems from the faulty assumption that the v ar- ied consid erations that are relev ant in intertemporal choices apply equally to d ifferent choices and thus that they can all be sensibly represented by a single d iscount rate. T hroughout the paper, w e stress the importance of d istinguishing among the v aried consid erations that und erlie in- tertemporal choices. We d istinguish time d iscounting from time preference. We use the term time d iscounting broad ly to encompass any reason for caring less about a future consequence, includ ing factors that d iminish the ex- pected utility generated by a future consequence, such as uncertainty or changing tastes. We use the term time preference to refer, more specifically, to the preference for immed iate utility ov er d elayed utility. In section 7, w e push this theme further, by examining w hether time preference itself might consist of d istinct psychological traits that can be separately analyzed . Section 8 conclud es. 2. Historical Origins of the Discounted Utility Mod el T he historical d ev elopments that cul- minated in the formulation of the DU mod el help to explain the mod el's limi- tations. Each of the major figures in the d ev elopment of the DU mod el-John R ae, Eugen v on B6hm-Baw erk, Irv ing Fisher, and Paul Samuelson-built upon the theoretical framew ork of his pred ecessors, d raw ing on little more than introspection and personal obser- v ation. When the DU mod el ev entually became entrenched as the d ominant theoretical framew ork for mod eling in- tertemporal choice, it w as d ue largely to its simplicity and its resemblance to the familiar compound interest formula, This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 353 and not as a result of empirical research d emonstrating its v alid ity. Intertemporal choice became firmly established as a d istinct topic in 1834, w ith John R ae's publication of T he So- ciological T heory of C apital. Like A d am Smith, R ae sought to d etermine w hy w ealth d iffered among nations. Smith had argued that national w ealth w as d e- termined by the amount of labor allo- cated to the prod uction of capital, but R ae recognized that this account w as in- complete because it failed to explain the d eterminants of this allocation. In R ae's v iew , the missing element w as "the effectiv e d esire of accumulation"-a psychological factor that d iffered across countries and d etermined a society's lev el of sav ing and inv estment. A long w ith inv enting the topic of in- tertemporal choice, R ae also prod uced the first in-d epth d iscussion of the psy- chological motiv es und erlying inter- temporal choice. R ae believ ed that intertemporal-choice behav ior w as the joint prod uct of factors that either pro- moted or limited the effectiv e d esire of accumulation. T he tw o main factors that promoted the effectiv e d esire of accumulation w ere the bequest motiv e ("the prev alence throughout the society of the social and benev olent affections," p. 58) and the propensity to exercise self-restraint ("the extent of the intel- lectual pow ers, and the consequent prev alence of habits of reflection, and prud ence, in the mind s of the mem- bers of society," p. 58). One limiting factor w as the uncertainty of human life: When engaged in safe occupations, and liv ing in healthy countries, men are much more apt to be frugal, than in unhealthy, or hazard ous occupations, and in climates pernicious to hu- man life. Sailors and sold iers are prod igals. In the West Ind ies, New Orleans, the East Ind ies, the expend iture of the inhabitants is profuse. T he same people, coming to resid e in the healthy parts of Europe, and not get- ting into the v ortex of extrav agant fashion, liv e economically. War and pestilence hav e alw ays w aste and luxury, among the other ev ils that follow in their train. (R ae 1834, p. 57) A second factor that limited the ef- fectiv e d esire of accumulation w as the excitement prod uced by the prospect of immed iate consumption, and the con- comitant d iscomfort of d eferring such av ailable gratifications: Such pleasures as may now be enjoyed gener- ally aw aken a passion strongly prompting to the partaking of them. T he actual presence of the immed iate object of d esire in the mind by exciting the attention, seems to rouse all the faculties, as it w ere to fix their v iew on it, and lead s them to a v ery liv ely conception of the enjoyments w hich it offers to their instant possession. (R ae 1834, p. 120) A mong the four factors that R ae id en- tified as the joint d eterminants of time preference, one can glimpse tw o fund a- mentally d ifferent v iew s. One, w hich w as later championed by William S. Jev ons (1888) and his son, Herbert S. Jev ons (1905), assumes that people care only about their immed iate utility, and ex- plains farsighted behav ior by postulat- ing utility from the anticipation of future consumption. On this v iew , d e- ferral of gratification w ill occur only if it prod uces an increase in "anticipal" utility that more than compensates for the d ecrease in immed iate consumption utility. T he second perspectiv e assumes equal treatment of present and future (zero d iscounting) as the natural base- line for behav ior, and attributes the ov erw eighting of the present to the miseries prod uced by the self-d enial required to d elay gratification. N. W. Senior, the best-know n ad v ocate of this "abstinence" perspectiv e, w rote, "T o abstain from the enjoyment w hich is in our pow er, or to seek d istant rather than immed iate results, are among the most painful exertions of the human w ill" (Senior 1836, p. 60). This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 354 Journal of Economic Literature, Vol. XL (June 2002) T he anticipatory-utility and absti- nence perspectiv es share the id ea that intertemporal trad eoffs d epend on im- med iate feelings-in one case, the im- med iate pleasure of anticipation, and in the other, the immed iate d iscomfort of self-d enial. T he tw o perspectiv es, how - ev er, explain v ariability in intertemporal- choice behav ior in d ifferent w ays. T he anticipatory-utility perspectiv e attrib- utes v ariations in intertemporal-choice behav ior to d ifferences in people's abilities to imagine the future and to d ifferences in situations that promote or inhibit such mental images. T he ab- stinence perspectiv e, on the other hand , explains v ariations in intertemporal- choice behav ior on the basis of ind iv id - ual and situational d ifferences in the psychological d iscomfort associated w ith self-d enial. In this v iew , one should observ e high rates of time d iscounting by people w ho find it painful to d elay gratification, and in situations in w hich d eferral is generally painful e.g., w hen one is, as R ae w ord ed it, in the "actual presence of the immed iate object of d esire." Eugen v on Bohm-Baw erk, the next major figure in the d ev elopment of the economic perspectiv e on intertemporal choice, ad d ed a new motiv e to the list proposed by R ae, Jev ons, and Senior, arguing that humans suffer from a systematic tend ency to und erestimate future w ants: It may be that w e possess inad equate pow er to imagine and to abstract, or that w e are not w illing to put forth the necessary effort, but in any ev ent w e limn a more or less incom- plete picture of our future w ants and espe- cially of the remotely d istant ones. A nd then there are all those w ants that nev er come to mind at all. (Bohm-Baw erk 1889, pp. 268-69)2 B6hm-Baw erk's analysis of time pref- erence, like those of his pred ecessors, w as heav ily psychological, and much of his v oluminous treatise, C apital and Interest, w as d ev oted to d iscussions of the psychological constituents of time preference. How ev er, w hereas the early v iew s of R ae, Senior, and Jev ons ex- plained intertemporal choices in terms of motiv es that are uniquely associated w ith time, B6hm-Baw erk began mod el- ing intertemporal choice in the same terms as other economic trad eoffs-as a "technical" d ecision about allocating re- sources (to oneself) ov er d ifferent points in time, much as one w ould allocate resources betw een any tw o competing interests, such as housing and food . B6hm-Baw erk's treatment of inter- temporal choice as an allocation of con- sumption among time period s w as for- malized a d ecad e later by the A merican economist Irv ing Fisher (1930). Fisher plotted the intertemporal consumption d ecision on a tw o-good ind ifference d iagram, w ith consumption in the cur- rent year on the abscissa, and consump- tion in the follow ing year on the ord i- nate. T his representation mad e clear that a person's observ ed (marginal) rate of time preference-the marginal rate of substitution at her chosen con- sumpti-on bund le-d epend s on tw o consid erations: time preference and d i- minishing marginal utility. Many econo- mists hav e subsequently expressed d is- comfort w ith using the term "time preference" to includ e the effects of d if- ferential marginal utility arising from unequal consumption lev els betw een time period s (see in particular Mancur Olson and Martin Bailey 1981). In Fisher's formulation, pure time prefer- ence can be interpreted as the marginal 2 In a frequently cited passage from T he Eco- nomics of Welfare, A rthur Pigou (1920) proposed a similar account of time preference, suggesting that it results from a type of cognitiv e illusion: "our telescopic faculty is d efectiv e, and w e, therefore, see future pleasures, as it w ere, on a d iminished scale." This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 355 rate of substitution on the d iagonal, w here consumption is equal in both period s. Fisher's w ritings, like those of his pred ecessors, includ ed extensiv e d iscus- sions of the psychological d eterminants of time preference. Like B6hm-Baw erk, he ld ifferentiated "objectiv e factors," such as projected future w ealth and risk, from "personal factors." Fisher's list of personal factors includ ed the four d escribed by R ae, "foresight" (the abil- ity to imagine future w ants-the inv erse of the d eficit that Bohm-Baw erk postu- lated ), and "fashion," w hich Fisher be- liev ed to be "of v ast importance . . . in its influence both on the rate of interest and on the d istribution of w ealth itself." (Fisher 1930, p. 88): T he most fitful of the causes at w ork is prob- ably fashion. T his at the present time acts, on the one hand , to stimulate men to sav e and become millionaires, and , on the other hand , to stimulate millionaires to liv e in an ostentatious manner. (Fisher 1930, p. 87) Hence, in the early part of the tw en- tieth century, "time preference" w as v iew ed as an amalgamation of v arious intertemporal motiv es. While the DU mod el cond enses these motiv es into the d iscount rate, w e w ill argue that resur- recting these d istinct motiv es is crucial for und erstand ing intertemporal choices. 3. T he Discounted Utility Mod el In 1937, Paul Samuelson introd uced the DU mod el in a fiv e-page article titled "A Note on Measurement of Util- ity." Samuelson's paper w as intend ed to offer a generalized mod el of intertem- poral choice that w as applicable to mul- tiple time period s (Fisher's graphical ind ifference-curv e analysis w as d ifficult to extend to more than tw o time peri- od s) and to make the point that repre- senting intertemporal trad eoffs re- quired a card inal measure of utility. But in Samuelson's simplified mod el, all the psychological concerns d iscussed ov er the prev ious century w ere compressed into a single parameter, the d iscount rate. T he DU mod el specifies a d ecision maker's intertemporal preferences ov er consumption profiles (C t,.. .,C T ). Und er the usual assumptions (completeness, transitiv ity, and continuity), such pref- erences can be represented by an in- tertemporal utility function Ut(C t,...,C T ). T he DU mod el goes further, by as- suming that a person's intertemporal utility function can be d escribed by the follow ing special functional form: T - t Ut(C t,...,C T ) = E D(k)u(ct+ k) k=O , k w here D(k) = l+p In this formulation, U(C t+k) is often inter- preted as the person's card inal instanta- neous utility function-her w ell-being in period t + k-and D(k) is often inter- preted as the person's d iscount func- tion-the relativ e w eight she attaches, in period t, to her w ell-being in period t + k. p represents the ind iv id ual's pure rate of time preference (her d iscount rate), w hich is meant to reflect the collectiv e effects of the "psychological" motiv es d iscussed in section 2.3 Samuelson d id not end orse the DU mod el as a normativ e mod el of in- tertemporal choice, noting that "any connection betw een utility as d iscussed here and any w elfare concept is d is- av ow ed " (p. 161). He also mad e no claims on behalf of its d escriptiv e v alid - ity, stressing, "It is completely arbitrary to assume that the ind iv id ual behav es so as to maximize an integral of the form env isaged in [the DU mod el]" (p. 159). How ev er, d espite Samuelson's manifest 3T he continuous-time analogue is Ut({C tlt e [t,T ]) = __t e - P( - t)u(ct). For expositional ease, w e shall restrict attention to d iscrete-time throughout. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 356 Journal of Economic Literature, Vol. XL (June 2002) reserv ations, the simplicity and ele- gance of this formulation w as irresist- ible, and the DU mod el w as rapid ly ad opted as the framew ork of choice for analyzing intertemporal d ecisions. T he DU mod el receiv ed a scarcely need ed further boost to its d ominance as the stand ard mod el of intertemporal choice w hen T jalling C . Koopmans (1960) show ed that the mod el could be d eriv ed from a superficially plausible set of axioms. Koopmans, like Samuel- son, d id not argue that the DU mod el w as psychologically or normativ ely plausible; his goal w as only to show that und er some w ell-specified (though ar- guably unrealistic) circumstances, in- d iv id uals w ere logically compelled to possess positiv e time preference. Pro- d ucers of a prod uct, how ev er, cannot d ictate how the prod uct w ill be used , and Koopmans' central technical mes- sage w as largely lost w hile his axiom- atization of the DU mod el helped to cement its popularity and bolster its perceiv ed legitimacy. In the remaind er of this section, w e d escribe some important features of the DU mod el as it is commonly used by economists, and briefly comment on the normativ e and positiv e v alid ity of these assumptions. T hese features d o not rep- resent an axiom system-they are nei- ther necessary nor sufficient cond itions for the DU mod el-but are intend ed to highlight the implicit psychological assumptions und erlying the mod el.4 3.1 Integration of New A lternativ es w ith Existing Plans A central assumption in most mod els of intertemporal choice-includ ing the DU mod el-is that a person ev aluates new alternativ es by integrating them w ith her existing plans. T o illustrate, consid er a person w ith an existing con- sumption plan (C t,...,C T ) w ho is offered an intertemporal-choice prospect X, w hich might be something like an op- tion to giv e up $5000 tod ay to receiv e $10,000 in fiv e years. Integration means that prospect X is not ev aluated in isola- tion, but in light of how it changes the person's aggregate consumption in all future period s. T hus, to ev aluate the prospect X, the person must choose w hat her new consumption path (C 't, .,C fT ) w ould be if she w ere to accept prospect X, and should accept the prospect if Ut(C 't, . .,C T ) > Ut(ct.- * .,C T ). A n alternativ e w ay to und erstand in- tegration is to recognize that intertem- poral prospects alter a person's bud get set. If the person's initial end ow ment is Eo, then accepting prospect X w ould change her end ow ment to Eo u X. Let- ting B(E) d enote the person's bud get set giv en end ow ment E-i.e., the set of consumption streams that are feasible giv en end ow ment E-the DU mod el says that the person should accept prospect X if: T IC -t max (, C t u(c) (C t, ,C T ) e B(Eo U X) Xt yP ) T m T t > 'max I;( u (c). ( C t, ,cT ) e B(Eo) X= t While integration seems normativ ely compelling, it may be too d ifficult to actually d o. A person may not hav e w ell-formed plans about future con- sumption streams, or be unable (or un- w illing) to recompute the new optimal plan ev ery time she makes an intertem- poral choice. Some of the ev id ence w e rev iew below supports the plausible presumption that people ev aluate the results of intertemporal choices ind e- pend ently of any expectations they hav e 4 T here are sev eral d ifferent axiom systems for the DU mod el-in ad d ition to Koopmans, see Peter Fishburn (1970), K. J. Lancaster (1963), R ichard F. Meyer (1976), and Fishburn and A riel R ubinstein (1982). This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 357 regard ing consumption in future time period s. 3.2 Utility Ind epend ence T he DU mod el explicitly assumes that the ov erall v alue-or "global utility"- of a sequence of outcomes is equal to the (d iscounted ) sum of the utilities in each period . Hence, the d istribution of utility across time makes no d ifference beyond that d ictated by d iscounting, w hich (assuming positiv e time prefer- ence) penalizes utility that is experi- enced later. T he assumption of utility ind epend ence has rarely been d iscussed or challenged , but its implications are far from innocuous. It rules out any kind of preference for patterns of utility ov er time-e.g., a preference for a flat utility profile ov er a roller-coaster util- ity profile w ith the same d iscounted utility.5 3.3 C onsumption Ind epend ence T he DU mod el explicitly assumes that a person's w ell-being in period t + k is ind epend ent of her consumption in any other period -i.e., that the marginal rate of substitution betw een consump- tion in period s t and T ' is ind epend ent of consumption in period t". C onsumption ind epend ence is analo- gous to, but fund amentally d ifferent from, the ind epend ence axiom of expected - utility theory. In expected -utility the- ory, the ind epend ence axiom specifies that preferences ov er uncertain pros- pects are not affected by the conse- quences that the prospects share-i.e., that the utility of an experienced out- come is unaffected by other outcomes that one might hav e experienced (but d id not). In intertemporal choice, con- sumption ind epend ence says that pref- erences ov er consumption profiles are not affected by the nature of consump- tion in period s in w hich consumption is id entical in the tw o profiles-i.e., that an outcome's utility is unaffected by outcomes experienced in prior or future period s. For example, consumption in- d epend ence says that a person's prefer- ence betw een an Italian and T hai res- taurant tonight should not d epend on w hether she had Italian last night, nor w hether she expects to hav e it tomor- row . A s the example suggests, and as Samuelson and Koopmans both recog- nized , there is no compelling rationale for such an assumption. Samuelson (1952, p. 674) noted that, "the amount of w ine I d rank yesterd ay and w ill d rink tomorrow can be expected to hav e ef- fects upon my tod ay's ind ifference slope betw een w ine and milk." Simi- larly, Koopmans (1960, p. 292) acknow l- ed ged that, "One cannot claim a high d egree of realism for [the ind epen- d ence assumption], because there is no clear reason w hy complementarity of good s could not extend ov er more than one time period ." 3.4 Stationary Instantaneous Utility When applying the DU mod el to spe- cific problems, it is often assumed that the card inal instantaneous utility func- tion u(c,) is constant across time, so that the w ell-being generated by any activ ity is the same in d ifferent period s. Most economists w ould acknow led ge that sta- tionarity of the instantaneous utility function is not sensible in many situ- ations, because people's preferences d o, in fact, change ov er time in pred ictable 5"Utility ind epend ence" has meaning only if one literally interprets u(ct+k) as w ell-being expe- rienced in period t + k. We believ e that this is, in fact, the common interpretation. For a mod el that relaxes the assumption of utility ind epend ence, see Benjamin Hermalin and A lice Isen (2000), w ho consid er a mod el in w hich w ell-being in period t d epend s on w ell-being in period t - 1- i.e., they assume ut = u(ct, ut-1). See also Daniel Kahneman, Peter Wakker, and R akesh Sarin (1997) w ho propose a set of axioms that w ould justify an assumption of ad d itiv e separability in instantaneous utility. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 358 Journal of Economic Literature, Vol. XL (June 2002) and unpred ictable w ays. T hough this unrealistic assumption is often retained for analytical conv enience, it becomes less d efensible as economists gain insight into how tastes change ov er time (see Loew enstein and A ngner, forthcoming, for a d iscussion of d ifferent sources of preference change).6 3.5 Ind epend ence of Discounting from C onsumption T he DU mod el assumes that the d is- count function is inv ariant across all forms of consumption. T his feature is crucial to the notion of time preference. If people d iscount utility from d ifferent sources at d ifferent rates, then the no- tion of a unitary time preference is meaningless. Instead w e w ould need to label time preference accord ing to the object being d elayed -"banana time preference," "v acation time prefer- ence," and so on. In section 7, w e d is- cuss in more d etail the v alid ity of the assumption that the same rate of time preference applies to all forms of consumption. 3.6 C onstant Discounting and T ime C onsistency A ny d iscount function can be w ritten in the formD(k) = H-Ik I (l,w herepn rep- resents the per-period d iscount rate for period n-that is, the d iscount rate applied betw een period s n and n + 1. Hence, by assuming that the d iscount function takes the form D(k) = K kI the DU mod el assumes a constant per- period d iscount rate (ppn = p for all n). 7 C onstant d iscounting entails an ev en- hand ed ness in the w ay a person ev alu- ates time. It means that d elaying or accelerating tw o d ated outcomes by a common amount should not change preferences betw een the outcomes-if in period t a person prefers X at X to Y at t + d for some t, then in period t she must prefer X at t to Y at t + d for all t. T he assumption of constant d iscounting permits a person's time preference to be summarized as a single d iscount rate. If constant d iscounting d oes not hold , then characterizing one's time preference requires the specification of an entire d iscountfunction. C onstant d iscounting implies that a person's intertemporal preferences are time-consistent, w hich means that later preferences "confirm" earlier prefer- ences. Formally, a person's preferences are time-consistent if, for any tw o con- sumption profiles (C t,.. .,C T ) and (C 't,...,C 'T ), w ith ct = c't, Ut(ct,ct + 1,.. .,C T ) ? Ut(c't,C 't+ 1, ...,C /T ) if and only if Ut+1(ct+1,...,cT ) > Ut+l(c't+1,...,C 'T ).8 For an interesting d is- cussion that questions the normativ e v a- lid ity of constant d iscounting, see Martin A lbrecht and Martin Weber (1995). 3.7 Diminishing Marginal Utility and Positiv e T ime Preference While not core features of the DU mod el, v irtually all analyses of intertem- poral choice assume both d iminishing 6 A s w e d iscuss in section 5, end ogenous prefer- ence changes, d ue to things such as habit forma- tion or reference d epend ence, are best und erstood in terms of consumption interd epend ence and not nonstationary utility. In some situations, nonsta- tionarities clearly play an important role in behav - ior-e.g., Stev en Suranov ic, R obert Gold farb, and T homas Leonard (1999), and O'Donoghue and Mathew R abin (1999a; 2000) d iscuss the impor- tance of nonstationarities in the realm of ad d ictiv e behav ior. 7 A n alternativ e but equiv alent d efinition of con- stant d iscounting is that D(k)/D(k + 1) is ind epen- d ent of k. 8 C onstant d iscounting implies time-consistent preferences only und er the ancillary assumption of stationary d iscounting, for w hich the d is- count function D(k) is the same in all period s. A s a counterexample, if the period -t d iscount function is Dt(k) = (,+p) w hile the k period -t + 1 d iscount function is Dt+l(k) = (r+E1 for some p' ? p, then the person exhibits cons ant d iscounting at both d ates t and t + 1, but nonetheless has time- inconsistent preferences. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 359 marginal utility (that the instantaneous utility function u(ct) is concav e) and posi- tiv e time preference (that the d iscount rate p is positiv e).9 T hese tw o assumptions create opposing forces in intertemporal choice: d iminishing marginal utility mo- tiv ates a person to spread consumption ov er time, w hile positiv e time prefer- ence motiv ates a person to concentrate consumption in the present. Since people d o, in fact, spread con- sumption ov er time, the assumption of d iminishing marginal utility (or some other property that has the same effect) seems strongly justified . T he assump- tion of positiv e time preference, on the other hand , is more questionable. Sev - eral researchers hav e argued for posi- tiv e time preference on logical ground s (Jack Hirshleifer 1970; Koopmans 1960; Koopmans, Peter A . Diamond , and R ichard E. Williamson 1964; Olson and Bailey 1981). T he gist of their argu- ments is that a zero or negativ e time preference, combined w ith a positiv e real rate of return on sav ing, w ould command the infinite d eferral of all consumption.10 But this conclusion as- sumes, unrealistically, that ind iv id uals hav e infinite life-spans and linear (or w eakly concav e) utility functions. Nev er- theless, in econometric analyses of sav - ings and intertemporal substitution, posi- tiv e time preference is sometimes treated as an id entifying restriction w hose v io- lation is interpreted as ev id ence of misspecification. T he most compelling argument sup- porting the-logic of positiv e time pref- erence w as mad e by Derek Parfit (1971; 1976; 1982), w ho contend s that there is no end uring self or "I" ov er time to w hich all future utility can be ascribed , and that a d iminution in psychological connections giv es our d escend ent fu- ture selv es the status of other people- making that utility less than fully "ours" and giv ing us a reason to count it less:11 We care less about our further future . . . because w e know that less of w hat w e are now -less, say, of our present hopes or plans, lov es or id eals-w ill surv iv e into the further future . . . [if] w hat matters hold s to a lesser d egree, it cannot be irrational to care less. (Parfit 1971, p. 99) Parfit's claims are normativ e, not d e- scriptiv e. He is not attempting to ex- plain or pred ict people's intertemporal choices, but is arguing that conclusions about the rationality of time preference must be ground ed in a correct v iew of personal id entity. How ev er, if this is the only compelling normativ e rationale for time d iscounting, it w ould be instruc- tiv e to test for a positiv e relation be- tw een observ ed time d iscounting and changing id entity. Fred erick (2002) cond ucted the only stud y of this type, 9 Discounting is not inherent to the DU mod el, because the mod el could be applied w ith p < 0. How ev er, the inclusion of p in the mod el strongly implies that it may take a v alue other than zero, and the name d iscount rate certainly suggests that it is greater than zero. 10 In the context of intergenerational choice, Koopmans (1967) called this result the parad ox of the ind efinitely postponed splurge. See also Ken- neth J. A rrow (1983), S. C hakrav arty (1962), and R obert M. Solow (1974). 11A s noted by Fred erick (2002), there is much d isagreement about the nature of Parfit's claim. In her rev iew of the philosophical literature, Jennifer Whiting (1986, p. 549) i entifies four d ifferent in- terpretations: (1) the strong absolute claim: that it is irrational for someone to care about their future w elfare, (2) the w eak absolute claim: that there is no rational requirement to care about one's future w elfare, (3) the strong comparativ e claim: that it is irrational to care more about one's ow n future w elfare than about the w elfare of any other per- son, and (4) the w eak comparativ e claim: that one is not rationally required to care more about their future w elfare than about the w elfare of any other person. We believ e that all of these interpretations are too strong, and that Parfit end orses only a w eaker v ersion of the w eak absolute claim. T hat is, he claims only that one is not rationally required to care about one's future w elfare to a d egree that exceed s the d egree of psychological connected ness that obtains betw een one's current self and one's future self. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 360 Journal of Economic Literature, Vol. XL (June 2002) and found no relation betw een mone- tary d iscount rates (as imputed from proced ures such as "I w ould be ind iffer- ent betw een $100 tomorrow and $- in fiv e years") and self-perceiv ed stabil- ity of id entity (as d efined by the follow - ing similarity ratings: "C ompared to now , how similar w ere you fiv e years ago [w ill you be fiv e years from now ]?"), nor d id he find any relation betw een such monetary d iscount rates and the presumed correlates of id entity stability (e.g., the extent to w hich peo- ple agree w ith the statement "I am still embarrassed by stupid things I d id a long time ago"). 4. DU A nomalies Ov er the last tw o d ecad es, empirical research on intertemporal choice has d ocumented v arious inad equacies of the DU mod el as a d escriptiv e mod el of be- hav ior. First, empirically observ ed d is- count rates are not constant ov er time, but appear to d ecline- a pattern often referred to as hyperbolic d iscounting. Furthermore, ev en for a giv en d elay, d iscount rates v ary across d ifferent types of intertemporal choices: gains are d iscounted more than losses, small amounts more than large amounts, and explicit sequences of multiple outcomes are d iscounted d ifferently than outcomes consid ered singly. 4.1 Hyperbolic Discounting T he best d ocumented DU anomaly is hyperbolic d iscounting. T he term "hyperbolic d iscounting" is often used to mean, in our terminology, that a per- son has a d eclining rate of time prefer- ence (in our notation, Pn is d eclining in n), and w e ad opt this meaning here. Sev eral results are usually interpreted as ev id ence for hyperbolic d iscounting. First, w hen subjects are asked to com- pare a smaller-sooner rew ard to a larger-later rew ard (see section 6 for a d escription of these proced ures), the implicit d iscount rate ov er longer time horizons is low er than the implicit d is- count rate ov er shorter time horizons. For example, R ichard T haler (1981) asked subjects to specify the amount of money they w ould require in [one month/one year/ten years] to make them ind ifferent to receiv ing $15 now . T he med ian responses [$20/$50/$100] imply an av erage (annual) d iscount rate of 345 percent ov er a one-month horizon, 120 percent ov er a one-year horizon, and 19 percent ov er a ten-year hori- zon.12 Other researchers hav e found a similar pattern (Uri Benzion, A mnon R apoport, and Joseph Yagil 1989; Gretchen B. C hapman 1996; C hapman and A rthur S. Elstein 1995; John L. Pend er 1996; Daniel A . R ed elmeier and Daniel N. Heller 1993). Second , w hen mathematical functions are explicitly fit to such d ata, a hyper- bolic functional form, w hich imposes d eclining d iscount rates, fits the d ata better than the exponential functional form, w hich imposes constant d iscount rates (Kris N. Kirby 1997; Kirby and Nino Marakov ic 1995; Joel Myerson and Leon- ard Green 1995; How ard R achlin, A nd res R aineri, and Dav id C ross 1991).13 T hird , researchers hav e show n that 12 T hat is, $15 $20*(e-(3.45)(1/12)) = $50*(e-(1.20)(1)) = $100*(e-(O 19)(l0)). While most empirical stud ies re- port av erage d iscount rates ov er a giv en horizon, it is sometimes more useful to d iscuss av erage "per- period " d iscount rates. Framed in these terms, T haler's results imply an av erage (annual) d iscount rate of 345 percent betw een now and one month from now , 100 percent betw een one month from now and one year from now , and 7.7 percent betw een one year from now and ten years from now . T hat is, $15 = $20*(e-(3.45)(1/12)) = $50*(e-(3.45)(1/12) e-(1.00)(11/12) = $100*(e-(3.45)(1/12) e-(1.00)(11/12)e-(0.077)(9)). 13 Sev eral hyperbolic functional forms hav e been proposed : George A inslie (1975) suggested the function D(t) = lit, R ichard Herrnstein (1981) and James Mazur (1987) suggested D(t) = 1/(1 + oct), and George Loew enstein and Drazen Prelec (1992) suggested D(t) = 1/(1 + oct)P/uA . This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Dlonoghue: T ime Discounting 361 preferences betw een tw o d elayed re- w ard s can rev erse in fav or of the more proximate rew ard as the time to both rew ard s d iminishes-e.g., someone may prefer $110 in 31 d ays ov er $100 in 30 d ays, but also prefer $100 now ov er $110 tomorrow . Such "preference re- v ersals" hav e been observ ed both in humans (Green, Nathaniel Fristoe, and Myerson 1994; Kirby and Herrnstein 1995; A nd rew Millar and Douglas Nav arick 1984; Jay Solnick et al. 1980) and in pigeons (A inslie and Herrnstein 1981; Green et al. 1981).14 Fourth, the pattern of d eclining d is- count rates suggested by the stud ies abov e is also ev id ent across stud ies. In section 6, w e summarize stud ies that es- timate d iscount rates. Figure la plots the av erage estimated d iscount factor (= 1/(1 + d iscount rate)) from each of these stud ies against the av erage time horizon for that stud y.15 A s the regres- sion line reflects, the estimated d is- count factor increases w ith the time ho- rizon, w hich means that the d iscount rate d eclines. We note, how ev er, that after exclud ing stud ies w ith v ery short time horizons (one year or less) from the analysis (see figure lb), there is no ev id ence that d iscount rates continue to d ecline. In fact, after exclud ing the stud - ies w ith short time horizons, the corre- lation betw een time horizon and d iscount factor is almost exactly zero (-0.0026). A lthough the collectiv e ev id ence out- lined abov e seems ov erw helmingly to support hyperbolic d iscounting, a re- cent stud y by Daniel R ead (2001) points out that the most common type of ev id ence-the find ing that implicit d iscount rates d ecrease w ith the time horizon-could also be explained by "subad d itiv e d iscounting," w hich means the total amount of d iscounting ov er a temporal interv al increases as the inter- v al is more finely partitioned .16 T o d em- onstrate subad d itiv e d iscounting and d istinguish it from hyperbolic d iscount- ing, R ead elicited d iscount rates for a tw o- year (24-month) interv al and for its three constituent interv als, an eight-month interv al beginning at the same time, an eight-month interv al beginning eight months later, and an eight-month inter- v al beginning sixteen months later. He found that the av erage d iscount rate for the 24-month interv al w as low er than the compound ed av erage d iscount rate ov er the three eight-month subinterv als- a result pred icted by subad d itiv e d is- counting but not pred icted by hyper- bolic d iscounting (or any type of d iscount function, for that matter). Moreov er, there w as no ev id ence that d iscount rates d eclined w ith time, as the d iscount rates for the three eight-month inter- v als w ere approximately equal. Similar empirical results w ere found earlier by J. H. Holcomb and P. S. Nelson (1992), 14 T hese stud ies all d emonstrate preference re- v ersals in the synchronic sense-subjects simulta- neously prefer $100 now ov er $110 tomorrow and prefer $110 in 31 d ays ov er $100 in 30 d ays, w hich is consistent w ith hyperbolic d iscounting. But there seems to be an implicit belief that such pref- erence rev ersals w ould also hold in the d iachronic sense-that if subjects w ho currently prefer $110 in 31 d ays ov er $100 in 30 d ays w ere brought back to the lab thirty d ays later, they w ould prefer $100 at that time ov er $110 one d ay later. Und er the assumption of stationary d iscounting (as d iscussed in footnote 8), synchronic preference rev ersals im- ply d iachronic preference rev ersals. T o the extent that subjects anticipate d iachronic rev ersals and w ant to av oid them, ev id ence of a preference for commitment could also be interpreted as ev id ence for hyperbolic d iscounting (w e d iscuss this issue more in section 5.1.1). 15 In some cases, the d iscount rates w ere com- puted from the med ian respond ent. In other cases, the mean d iscount rate w as used . 16 R ead 's proposal that d iscounting is subad d i- tiv e is compatible w ith analogous results in other d omains. For example, A mos T v ersky and Derek Koehler (1994) found that the total probability as- signed to an ev ent increases the more finely the ev ent is partitioned -e.g., the probability of "d eath by accid ent" is jud ged to be more likely if one separately elicits the probability of "d eath by fire," "d eath by d row ning," "d eath by falling," etc. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 362 Journal of Economic Literature, Vol. XL (June 2002) 1.0 - e 0.8 0.6 El 0.4- $ 0.2 0.0 0 5 10 15 time horizon (years) Figure la. Discount Factor as a Function of T ime Horizon (all stud ies) although they d id not interpret their results the same w ay. If R ead is correct about subad d itiv e d iscounting, its main implication for economic applications may be to prov id e an alternativ e psychological und erpin- ning for using a hyperbolic d iscount function, because most intertemporal d ecisions are based primarily on d is- counting from the present.17 -~1.0 @0 c 0.8 - * o 0.6- e 0.4- = 0.2- *E0.0- 0 5 10 15 time horizon (years) Figure lb. Discount Factor as a Function of T ime Horizon (stud ies w ith av g. horizons > 1 year) 4.2 Other DU A nomalies T he DU mod el not only d ictates that the d iscount rate should be constant for all time period s; it also assumes that the d iscount rate should be the same for all types of good s and all categories of intertemporal d ecisions. T here are sev - eral empirical regularities that appear to contrad ict this assumption, namely: (1) gains are d iscounted more than losses; (2) small amounts are d iscounted more than large amounts; (3) greater d iscounting is show n to av oid d elay of a good than to exped ite its receipt; (4) in choices ov er sequences of outcomes, improv ing sequences are often preferred to d eclining sequences though positiv e time preference d ic- tates the opposite; and (5) in choices ov er sequences, v iolations of ind epen- d ence are perv asiv e, and people seem to prefer spread ing consumption ov er time in a w ay that d iminishing marginal utility alone cannot explain. 4.2.1 T he "Sign Effect" (gains are d iscounted more than losses) Many stud ies hav e conclud ed that gains are d iscounted at a higher rate than losses. For instance, T haler (1981) 17A few stud ies hav e actually found increasing d iscount rates. Fred erick (1999) asked 228 respon- d ents to imagine that they w orked at a job that consisted of both pleasant w ork ("good d ays") and unpleasant w ork i"bad d ays") an to equate the attractiv eness of hav ing ad d itional good d ays this year or in a future year. On av erage, respond ents w ere ind ifferent betw een 20 extra good d ays this year, 21 the follow ing year, or 40 in fiv e years, implying a one-year d iscount rate of 5 percent and a fiv e-year d iscount rate of 15 percent. A possible explanation is that a d esire for improv ement is ev oked more strongly for tw o successiv e years (this year and next) than for tw o separated years (this year and fiv e years hence). R ubinstein (2000) asked stud ents in a political science class to choose betw een the follow ing tw o payment sequences: March 1 June 1 Sept 1 Nov 1 A : $997 $997 $997 $997 A pril 1 Julyl Oct 1 Dec 1 B: $1000 $1000 $1000 $1000 T hen, tw o w eeks later, he asked them to choose betw een $997 on Nov ember 1 and $1000 on December 1. Fifty-four percent of respond ents preferred $997 in Nov ember to $1000 in Decem- Eer, but only 34 percent preferred sequence A to sequence B. T hese tw o results suggest increasing d iscount rates. T o explain them R ubinstein specu- lated that the three more proximate ad d itional ele- ments may hav e masked the d ifferences in the timing of the sequence of d ated amounts, w hile making the d ifferences in amounts more salient. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 363 asked subjects to imagine they had re- ceiv ed a traffic ticket that could be paid either now or later and to state how much they w ould be w illing to pay if payment could be d elayed (by three months, one year, or three years). T he d iscount rates imputed from these an- sw ers w ere much low er than the d iscount rates imputed from comparable questions about monetary gains. T his pattern is prev alent in the literature. Ind eed , in many stud ies, a substantial proportion of sub- jects prefer to incur a loss immed iately rather than d elay it (Benzion, R apoport, and Yagil 1989; Loew enstein 1987; L. D. MacKeigan et al. 1993; Walter Mischel, Joan Grusec, and John C . Masters 1969; R ed elmeier and Heller 1993; J. Frank Yates and R oyce A . Watts 1975). 4.2.2 T he "Magnitud e Effect" (small outcomes are d iscounted more than large ones) Most stud ies that v ary outcome size hav e found that large outcomes are d iscounted at a low er rate than small ones (A inslie and Vard a Haend el 1983; Benzion, R apoport, and Yagil 1989; Green, Fristoe, and Myerson 1994; Green, A strid Fry, and Myerson 1994; Hol- comb and Nelson 1992; Kirby 1997; Kirby and Marakov ic 1995; Kirby, Nancy Petry and Warren Bickel 1999; Loew enstein 1987; R aineri and R achlin 1993; Marjorie K. Shelley 1993; T haler 1981). In T haler's (1981) stud y, for ex- ample, respond ents w ere, on av erage, ind ifferent betw een $15 immed iately and $60 in a year, $250 immed iately and $350 in a year, and $3000 immed i- ately and $4000 in a year, implying d is- count rates of 139 percent, 34 percent, and 29 percent, respectiv ely. 4.2.3 T he "Delay-Speed up" A symmetry Loew enstein (1988) d emonstrated that imputed d iscount rates can be d ramatically affected by w hether the change in d eliv ery time of an outcome is framed as an acceleration or a d elay from some temporal reference point. For example, respond ents w ho d id n't expect to receiv e a VC R for another year w ould pay an av erage of $54 to re- ceiv e it immed iately, but those w ho thought they w ould receiv e it immed i- ately d emand ed an av erage of $126 to d elay its receipt by a year. Benzion, R apoport, and Yagil (1989) and Shelley (1993) replicated Loew enstein's find ings for losses as w ell as gains (respond ents d emand ed more to exped ite payment than they w ould pay to d elay it). 4.2.4 Preference for Improv ing Sequences In stud ies of d iscounting that inv olv e choices betw een tw o outcomes-e.g., X at T v s. Y at T '-positiv e d iscounting is the norm. R esearch examining prefer- ences ov er sequences of outcomes, how - ev er, has generally found that people prefer improv ing sequences to d eclin- ing sequences (for an ov erv iew , see A riely and C armon, in press; Fred erick and Loew enstein 2002; Loew enstein and Prelec 1993). For example, Loew en- stein and Nachum Sicherman (1991) found that, for an otherw ise id entical job, most subjects prefer an increasing w age profile to a d eclining or flat one (see also R obert Frank 1993). C hristo- pher Hsee, R obert P. A belson, and Peter Salov ey (1991) found that an in- creasing salary sequence w as rated as highly as a d ecreasing sequence that conferred much more money. C arol Varey and Kahneman (1992) found that subjects strongly preferred streams of d ecreasing d iscomfort to streams of in- creasing d iscomfort, ev en w hen the ov er- all sum of d iscomfort ov er the interv al w as otherw ise id entical. Loew enstein and Prelec (1993) found that respon- d ents w ho chose betw een sequences of tw o or more ev ents (e.g., d inners or This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 364 Journal of Economic Literature, Vol. XL (June 2002) v acation trips) on consecutiv e w eekend s or consecutiv e months generally pre- ferred to sav e the better thing for last. C hapman (2000) presented respond ents w ith hypothetical sequences of head - ache pain that w ere matched in terms of total pain that either grad ually less- ened or grad ually increased w ith time. Sequence d urations includ ed one hour, one d ay, one month, one year, fiv e years, and tw enty years. For all se- quence d urations, the v ast majority (from 82 percent to 92 percent) of sub- jects preferred the sequence of pain that lessened ov er time. (See also W. T . R oss, Jr. and I. Simonson 1991). 4.2.5 Violations of Ind epend ence and Preferencefor Spread T he research on preferences ov er se- quences also rev eals strong v iolations of ind epend ence. C onsid er the follow ing pair of questions from Loew enstein and Prelec (1993): Imagine that ov er the next fiv e w eekend s you must d ecid e how to spend your Saturd ay nights. From each pair of sequences of d inners below , circle the one you w ould prefer. "Fancy French" refers to a d inner at a fancy French restaurant. "Fancy Lobster" refers to an exquisite lobster d inner at a four-star restaurant. Ignore sched uling consid erations (e.g., your current plans). first second third fourth fifth w eekend w eekend w eekend w eekend w eekend Option A Fancy Eat at Eat at Eat at Eat at [11%] French home home home home Option B Eat at Eat at Fancy Eat at Eat at [89%] home home French home home Option C Fancy Eat at Eat at Eat at Fancy [49%] French home home home Lobster Option D Eat at Eat at Fancy Eat at Fancy [51%] home home French home Lobster A s d iscussed in section 3.3, consump- tion ind epend ence implies that prefer- ences betw een tw o consumption pro- files should not be affected by the nature of the consumption in period s in w hich consumption is id entical in the tw o profiles. T hus, anyone preferring profile B to profile A (w hich share the fifth period "Eat at Home") should also prefer profile D to profile C (w hich share the fifth period "Fancy Lobster"). A s the d ata rev eal, how ev er, many respond ents v iolated this pred iction, preferring the fancy French d inner on the third w eekend , if that w as the only fancy d inner in the profile, but prefer- ring the fancy French d inner on the first w eekend if the profile contained another fancy d inner. T his result could be explained by the simple d esire to spread consumption ov er time-w hich, in this context, v iolates the d ubious as- sumption of ind epend ence that the DU mod el entails. Loew enstein and Prelec (1993) pro- v id e further ev id ence of such a prefer- ence for spread . Subjects w ere asked to imagine that they w ere giv en tw o cou- pons for fancy ($100) restaurant d in- ners, and w ere asked to ind icate w hen they w ould use them, ignoring consid - erations such as holid ays, birthd ays, and such. Subjects either w ere told that "you can use the coupons at any time betw een tod ay and tw o years from to- d ay" or w ere told nothing about any constraints. Subjects in the tw o-year constraint cond ition actually sched uled both d inners at a later time than those w ho faced no explicit constraint-they d elayed the first d inner for eight w eeks (rather than three) and the second d in- ner for 31 w eeks (rather than thirteen). T his counterintuitiv e result can be ex- plained in terms of a preference for spread if the explicit tw o-year interv al w as greater than the implicit time hori- zon of subjects in the unconstrained group. 4.3 A re T hese "A nomalies" Mistakes? In other d omains of jud gment and choice, many of the famous "effects" This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 365 that hav e been d ocumented are re- gard ed as errors by the people w ho commit them. For example, in the "con- junction fallacy" d iscov ered by T v ersky and Kahneman (1983), many people w ill- w ith some reflection-recognize that a conjunction cannot be more likely than one of its constituents (e.g., that it can't be more likely for Lind a to be a femi- nist bank teller than for her to be "just" a bank teller). In contrast, the patterns of preferences that are re- gard ed as "anomalies" in the context of the DU mod el d o not necessarily v io- late any stand ard or principle that peo- ple believ e they should uphold . Ev en w hen the choice pattern is pointed out to people, they d o not regard them- selv es as hav ing mad e a mistake (and probably hav e not mad e one!). For example, there is no compelling logic that d ictates that one w ho prefers to d elay a French d inner should also pre- fer to d o so w hen that French d inner w ill be closely follow ed by a lobster d inner. Ind eed , it is unclear w hether any of the DU "anomalies" should be regard ed as mistakes. Fred erick and R ead (2002) found ev id ence that the magnitud e ef- fect is more pronounced w hen subjects ev aluate both "small" and "large" amounts than w hen they ev aluate either one. Specifically, the d ifference in the d iscount rates betw een a small amount ($10) and a large amount ($1000) w as larger w hen the tw o jud gments w ere mad e in close succession than w hen they w ere mad e separately. A nalogous results w ere obtained for the sign ef- fect, as the d ifferences in d iscount rates betw een gains and losses w ere slightly larger in a w ithin-subjects d esign, w here respond ents ev aluated d elayed gains and d elayed losses, than in a betw een-subjects d esign w here they ev aluate only gains or only losses. Since respond ents d id not attempt to coord inate their responses to conform to DU's postulates w hen they ev aluated rew ard s of d ifferent sizes, it suggests that they consid er the d ifferent d is- count rates to be normativ ely appropri- ate. Similarly, ev en after Loew enstein and Sicherman (1991) informed respon- d ents that a d ecreasing w age profile ($27,000, $26,000, . . . $23,000) w ould (v ia appropriate sav ing and inv esting) permit strictly more consumption in ev ery period than the correspond ing increasing w age profile w ith an equiv - alent nominal total ($23,000, $24,000, . . . $27,000), respond ents still pre- ferred the increasing sequence. Perhaps they suspected that they could not exercise the required self control to maintain their d esired consumption sequence, or felt a general leeriness about the significance of a d eclining w age, either of w hich could justify that choice. A s these examples illus- trate, many DU "anomalies" exist as "anomalies" only by reference to a mod el that w as constructed w ithout regard to its d escriptiv e v alid ity, and w hich has no compelling normativ e basis. 5. A lternativ e Mod els In response to the anomalies just enumerated , and other intertemporal- choice phenomena that are inconsistent w ith the DU mod el, a v ariety of alter- nate theoretical mod els hav e been d ev eloped . Some mod els attempt to achiev e greater d escriptiv e realism by relaxing the assumption of constant d iscounting. Other mod els incorporate ad d itional consid erations into the in- stantaneous utility function, such as the utility from anticipation. Still others d epart from the DU mod el more rad ically, by includ ing, for instance, systematic mispred ictions of future utility. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 366 Journal of Economic Literature, Vol. XL (June 2002) 5.1 Mod els of Hyperbolic Discounting In the economics literature, R . H. Strotz (1955-56) w as the first to con- sid er alternativ es to exponential d is- counting, seeing "no reason w hy an ind iv id ual should hav e such a special d iscount function" (p. 172). Moreov er, Strotz recognized that for any d iscount function other than exponential, a person w ould hav e time-inconsistent preferences.18 He proposed tw o strate- gies that might be employed by a per- son w ho foresees how her preferences w ill change ov er time: the "strategy of precommitment" (w herein she commits to some plan of action) and the "strat- egy of consistent planning" (w herein she chooses her behav ior ignoring plans that she know s her future selv es w ill not carry out).19 While Strotz d id not posit any specific alternativ e functional forms, he d id suggest that "special attention" be giv en to the case of d eclining d iscount rates. Motiv ated by the ev id ence d iscussed in section 4.1, there has been a recent surge of interest among economists in the implications of d eclining d iscount rates (beginning w ith Dav id Laibson 1994, 1997). T his literature has used a particularly simple functional form w hich captures the essence of hyperbolic d iscounting: D(k) =ib ifk; P fk> 0. T his functional form w as first introd uced by E. S. Phelps and Pollak (1968) to stud y intergenerational altruism, and w as first applied to ind iv id ual d ecision mak- ing by Jon Elster (1979). It assumes that the per-period d iscount rate betw een now and the next period is '- w hereas the per-period d iscount rate betw een any tw o future period s is lA < l P8. Hence, this (J3,6) formulation assumes a d eclining d iscount rate betw een this pe- riod and next, but a constant d iscount rate thereafter. T he (J3,6) formulation is highly tractable, and captures many of the qualitativ e implications of hyperbolic d iscounting. Laibson and his collaborators hav e used the (P8,6) formulation to explore the implications of hyperbolic d iscount- ing for consumption-sav ing behav ior. Hyperbolic d iscounting lead s a person to consume more than she w ould like from a prior perspectiv e (or, equiv a- lently, to und er-sav e). Laibson (1997) explores the role of illiquid assets, such as housing, as an imperfect commit- ment technology, emphasizing how a person could limit ov erconsumption by tying up her w ealth in illiquid assets. Laibson (1998) explores consumption- sav ing d ecisions in a w orld w ithout illiq- uid assets (or any other commitment technology). T hese papers d escribe how hyperbolic d iscounting might explain some stylized empirical facts, such as the excess comov ement of income and consumption, the existence of asset-spe- cific marginal propensities to consume, low lev els of precautionary sav ings, and the correlation of measured lev els of patience w ith age, income, and w ealth. Laibson, A nd rea R epetto, and Jeremy T obacman (1998), and George-Marios A ngeletos et al. (2001) calibrate mod els of consumption-sav ing d ecisions, using both exponential d iscounting and (,3,6) hyperbolic d iscounting. By comparing simulated d ata to real-w orld d ata, they d emonstrate how hyperbolic d iscount- ing can better explain a v ariety of empirical observ ations in the consump- tion-sav ing literature. In particular, 18 Strotz implicitly assumes stationary d iscount- ing. 19 Build ing on Strotz's strategy of consistent planning, some researchers hav e ad d ressed the question of w hether there exists a consistent path or general non-exponential d iscount functions. See in particular R obert Pollak (1968), Bezalel Peleg and Menahem Yaari (1973), and Stev en Gold man (1980). This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 367 A ngeletos et al. (2001) d escribe how hyperbolic d iscounting can explain the coexistence of high preretirement w ealth, low liquid asset hold ings (rela- tiv e to income lev els and illiquid asset hold ings), and high cred it-card d ebt. C arolyn Fischer (1999) and O'Donoghue and R abin (1999c, 2001) hav e applied (,3,6) preferences to pro- crastination, w here hyperbolic d iscount- ing lead s a person to put off an onerous activ ity more than she w ould like from a prior perspectiv e.20 O'Donoghue and R abin (1999c) examine the implications of hyperbolic d iscounting for contract- ing w hen a principal is concerned w ith combating procrastination by an agent. T hey show how incentiv e schemes w ith "d ead lines" may be a useful screening d ev ice to d istinguish efficient d elay from inefficient procrastination. O'Donoghue and R abin (2001) explore procrastina- tion w hen a person must not only choose w hen to complete a task, but also w hich task to complete. T hey show that a person might nev er carry out a v ery easy and v ery good option because they continually plan to carry out an ev en better but more onerous option. For instance, a person might nev er take half an hour to straighten the shelv es in her garage because she persistently plans to take an entire d ay to d o a major cleanup of the entire garage. Extend ing this logic, they show that prov id ing peo- ple w ith new options might make pro- crastination more likely. If the person's only option w ere to straighten the shelv es, she might d o it in a timely manner; but if the person can either straighten the shelv es or d o the major cleanup, she now may d o nothing. O'Donoghue and R abin (1999d ) apply this logic to retirement planning. O'Donoghue and R abin (1999a, 2000), Jonathan Gruber and Botond Koszegi (2000), and Juan D. C arrillo (1999) hav e applied (J3,6) preferences to ad d iction. T hese researchers d e- scribe how hyperbolic d iscounting can lead people to ov erconsume harmful ad d ictiv e prod ucts, and examine the d egree of harm caused by such ov er- consumption. C arrillo and T homas Mariotti (2000) and R oland Benabou and Jean T irole (2000) hav e examined how (13,6) preferences might influence a person's d ecision to acquire informa- tion. If, for example, a person is d ecid - ing w hether to embark on a specific research agend a, she may hav e the op- tion to get feed back from colleagues about its likely fruitfulness. T he stan- d ard economic mod el implies that peo- ple should alw ays choose to acquire this information if it is free. How ev er, C ar- rillo and Mariotti show that hyperbolic d iscounting can lead to "strategic igno- rance"-a person w ith hyperbolic d is- counting w ho is w orried about w ith- d raw ing from an ad v antageous course of action w hen the costs become imminent might choose not to acquire free infor- mation if d oing so increases the risk of bailing out. 5.1. 1 Self A w areness A person w ith time-inconsistent pref- erences may or may not be aw are that her preferences w ill change ov er time. Strotz (1955-56) and Pollak (1968) d iscussed tw o extreme alternativ es. A t one extreme, a person could be com- pletely "naYv e" and believ e that her future preferences w ill be id entical to her current preferences. A t the other extreme, a person could be com- pletely "sophisticated " and correctly pred ict how her preferences w ill change ov er time. While casual observ a- tion and introspection suggest that 20 While not framed in terms of hyperbolic d is- counting, George A kerlof's (1991) mod el of pro- crastination is formally equiv alent to a hyperbolic mod el. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 368 Journal of Economic Literature, Vol. XL (June 2002) people lie somew here in betw een these tw o extremes, behav ioral ev id ence re- gard ing the d egree of aw areness is quite limited . One w ay to id entify sophistication is to look for ev id ence of commitment. Someone w ho suspects that her prefer- ences w ill change ov er time might take steps to eliminate an option that seems inferior now but might tempt her later. For example, someone w ho currently prefers $110 in 31 d ays to $100 in 30 d ays but w ho suspects that in a month she w ill prefer $100 immed iately to $110 tomorrow , might attempt to elimi- nate the $100 rew ard from the later choice set, and thereby bind herself now to receiv e the $110 rew ard in 31 d ays. R eal-w orld examples of commit- ment includ e "C hristmas clubs" or "fat farms." Perhaps the best empirical d emon- stration of a preference for commit- ment w as cond ucted by Dan A riely and Klaus Wertenbroch (2002). In that stud y, MIT executiv e-ed ucation stud - ents had to w rite three short papers for a class and w ere assigned to one of tw o experimental cond itions. In one cond ition, d ead lines for the three pa- pers w ere imposed by the instructor and w ere ev enly spaced across the se- mester. In the other cond ition, each stud ent w as allow ed to set her ow n d ead lines for each of the three papers. In both cond itions, the penalty for d elay w as 1 percent per d ay late, re- gard less of w hether the d ead line w as externally or self-imposed . A lthough stud ents in the free-choice cond ition could hav e mad e all three papers d ue at the end of the semester, many d id , in fact, choose to impose d ead lines on themselv es, suggesting that they ap- preciated the v alue of commitment. Few stud ents chose ev enly spaced d ead lines, how ev er, and those w ho d id not performed w orse in the course than those w ith ev enly spaced d ead - lines (w hether externally imposed or self-imposed ) 21 O'Donoghue and R abin (1999b) ex- amine how people's behav iors d epend on their sophistication about their ow n time inconsistency. Some behav iors, such as using illiquid assets for commit- ment, require some d egree of sophisti- cation. Other behav iors, such as ov er- consumption or procrastination, are more robust to the d egree of aw are- ness, though the d egree of misbehav ior may d epend on the d egree of sophisti- cation. T o und erstand such effects, O'Donoghue and R abin (2001) intro- d uce a formal mod el of partial naiv ete', in w hich a person is aw are that she w ill hav e future self-control problems but und erestimates their magnitud e. T hey show that sev ere procrastination cannot occur und er complete sophistication, but can arise ev en if the person is only a little naiv e. For more d iscussion on self-aw areness, see O'Donoghue and R abin (in press). T he d egree of sophistication v ersus naiv ete has important implications for public policy. If people are sufficiently sophisticated about their ow n self- control problems, prov id ing commit- ment d ev ices may be beneficial. How - ev er, if people are nafv e, policies might be better aimed at either ed u- cating people about loss of control (making them more sophisticated ), or prov id ing incentiv es for people to use commitment d ev ices, ev en if they d on't recognize the need for them. 21 A similar "natural" experiment w as recently cond ucted by the Economic and Social R esearch C ouncil of Great Britain. T hey recently eliminated submission d ead lines and now accept grant pro- posals on a "rolling" basis (though they are still rev iew ed only period ically). In response to this policy change, submissions hav e actually d eclined by about 15-20 percent (d irect correspond ence w ith C hris C asw ill at ESR C ). This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 369 5.2 Mod els T hat Enrich the Instantaneous Utility Function Many d iscounting anomalies, espe- cially those in section 4.2, can be un- d erstood as a misspecification of the instantaneous utility function. Similarly, many of the confound s w e d iscuss in section 6 are caused by researchers at- tributing to the d iscount rate aspects of preference that are more appropriately consid ered as arguments in the instan- taneous utility function. A s a result, alternativ e mod els of intertemporal choice hav e been ad v anced that ad d ad - d itional arguments, such as utility from anticipation, to the instantaneous utility function. 5.2.1 Habit-Formation Mod els James Duesenberry (1952) w as the first economist to propose the id ea of "habit formation"-that the utility from current consumption ("tastes") can be affected by the lev el of past consump- tion. T his id ea w as more formally d ev el- oped by Pollak (1970) and Harl R yd er and Geoffrey Heal (1973). In habit for- mation mod els, the period -T instantane- ous utility function takes the form u(C t;C t - 1,C u - 2,...) w here a2u/ac, act '> 0 for T ' <T . For simplicity, most such mod els assume that all effects of past consumption for current utility enter through a state v ariable. T hat is, they assume that period -T instantaneous util- ity function takes the form u(cr;zt) w here zt is a state v ariable that is in- creasing in past consumption and D2/aC C DZt > 0. Both Pollak (1970) and R yd er and Heal (1973) assume that zt is the exponentially w eighted sum of past consumption, or zt = =ici A lthough habit formation is often said to ind uce a preference for an in- creasing consumption profile, it can, und er some circumstances, lead a per- son to prefer a d ecreasing or ev en non- monotonic consumption profile. T he d i- rection of the effect d epend s on things such as how much one has alread y con- sumed (as reflected in the initial habit stock), and , perhaps most importantly, w hether current consumption increases or d ecreases future utility. In recent years, habit-formation mod - els hav e been used to analyze a v ariety of phenomena. Gary Becker and Kev in Murphy (1988) use a habit-formation mod el to stud y ad d ictiv e activ ities, and in particular to examine the effects of past and future prices on the current consumption of ad d ictiv e prod ucts.22 Habit formation can help explain asset- pricing anomalies such as the equity- premium puzzle (A nd rew A bel 1990; John C ampbell and John C ochrane 1999; George M. C onstantinid es 1990). Incor- porating habit formation into business- cycle mod els can improv e their ability to explain mov ements in asset prices (Urban Jermann 1998; Michele Bold rin, Law rence C hristiano, and Jonas Fisher 2001). Some recent papers hav e show n that habit formation may help explain other empirical puzzles in macro- economics as w ell. Whereas stand ard grow th mod els assume that high sav ing rates cause high grow th, recent ev i- d ence suggests that the causality can run in the opposite d irection. C hristo- pher C arroll, Jod y Ov erland , and Dav id Weil (2000) show that, und er cond itions of habit formation, high grow th rates can cause people to sav e more. Jeffrey Fuhrer (2000) show s how habit forma- tion might explain the recent find ing that aggregate spend ing tend s to hav e a grad ual "hump-shaped " response to 22 For rational-choice mod els build ing on Becker and Murphy's framew ork, see A thanasios Orphanid es and Dav id Zerv os (1995), R uqu Wang (1997), and Suranov ic, Gold farb, and Leonard (1999). For ad d iction mod els that incorporate hyperbolic d iscounting, see O'Donoghue and R abin (1999a, 2000), Gruber and Koszegi (2000), and C arrillo (1999). This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 370 Journal of Economic Literature, Vol. XL (June 2002) v arious shocks. T he key feature of habit formation that d riv es many of these re- sults is that, after a shock, consumption ad justment is sluggish in the short term but not in the long term. 5.2.2 R eference-Point Mod els C losely related to, but conceptually d istinct from, habit-formation mod els are mod els of reference-d epend ent util- ity, w hich incorporate id eas from pros- pect thI eory (Kahneman and T v ersky 1979; T v ersky and Kahneman 1991). A ccord ing to prospect theory, outcomes are ev aluated using a v alue function d e- fined ov er d epartures from a reference point-in our notation, the period -t in- stantaneous utility function takes the form u(ct,rt) = v (ct - rt). T he reference point, rt, might d epend on past con- sumption, expectations, social compari- son, status quo, and such. A second feature of prospect theory is that the v alue function exhibits loss av ersion- negativ e d epartures from one's refer- ence consumption lev el d ecrease utility by a greater amount than positiv e d e- partures increase it. A third feature of prospect theory is that the v alue func- tion exhibits-d iminishing sensitiv ity for both gains and losses, w hich means that the v alue function is concav e ov er gains and conv ex ov er losses.23 Loew enstein and Prelec (1992) ap- plied a specialized v ersion of such a v alue function to intertemporal choice to explain the magnitud e effect, the sign effect, and the d elay-speed up asymmetry. T hey show that if the elas- ticity of the v alue function is increasing in the magnitud e of outcomes, people w ill d iscount smaller magnitud es more than larger magnitud es. Intuitiv ely, the elasticity cond ition captures the insight that people are responsiv e to both d if- ferences and ratios of rew ard amounts. It implies that someone w ho is ind iffer- ent betw een, say, $10 now and $20 in a year should prefer $200 in a year ov er $100 now because the larger rew ard s hav e a greater d ifference (and the same ratio). C onsequently, ev en if a person's time preference is actually constant across outcomes, she w ill be more w ill- ing to w ait for a fixed proportional in- crement w hen rew ard s are larger, and , thus, her imputed d iscount rate w ill be smaller for larger outcomes. Similarly, if the v alue function for losses is more elastic than the v alue function for gains, then people w ill d iscount gains more than losses. Finally, such a mod el helps explain the d elay-speed up asymmetry (Loew enstein 1988). Shifting consump- tion in any d irection is mad e less d esir- able by loss av ersion, since one loses consumption in one period and gains it in another. When d elaying consump- tion, loss av ersion reinforces time d is- counting, creating a pow erful av ersion to d elay. When exped iting consumption, loss av ersion opposes time d iscounting, red ucing the d esirability of speed up (and , occasionally, ev en causing an av ersion to it). Using a reference-d epend ent mod el that assumes loss av ersion in consump- tion, Dav id Bow man, Deborah Mine- hart, and R abin (1999) pred ict that "new s" about one's (stochastic) future income affects one's consumption grow th d ifferently than the stand ard Permanent Income Hypothesis pred icts. A ccord ing to (the log-linear v ersion of) the Permanent Income Hypothesis, changes in future income should not 23 R eference-point mod els sometimes assume there is a d irect effect of the consumption lev el or reference lev el, so that u(c,,r,) = v (cT - r-) + w (c.) or u(cT ,rt) = v (cT - r-) + w (r.). Some habit-formation mod els could be interpreted as reference-point mod els, w here the state v ariable zt is the refer- ence point. Ind eed , many habit-formation mod els, such as Pollak (1970) and C onstantinid es (1990), assume instantaneous utility functions of the form u(cC - z.), although they typically assume neither loss av ersion nor d iminishing sensitiv ity. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 371 affect the rate of consumption grow th. For example, if a person find s out that her permanent income w ill be low er than she formerly thought, she w ould red uce her consumption by, say, 10 per- cent in ev ery period , leav ing her con- sumption grow th unchanged . If, how - ev er, this person w ere loss av erse in current consumption, she w ould be un- w illing to red uce this year's consump- tion by 10 percent-forcing her to re- d uce future consumption by more than 10 percent, and thereby red ucing the grow th rate of her consumption. T w o stud ies by John Shea (1995a,b) support this pred iction. Using both aggregate U.S. d ata and d ata from teachers' unions (in w hich w ages are set one year in ad v ance), Shea find s that consump- tion grow th respond s more strongly to future w age d ecreases than to future w age increases. 5.2.3 Mod els Incorporating Utility from A nticipation Some alternativ e mod els build on the notion of "anticipal" utility d iscussed by the eld er and younger Jev ons. If people d eriv e pleasure not only from current consumption, but also from anticipating future consumption, then current in- stantaneous utility w ill d epend posi- tiv ely on future consumption-that is, the period -t instantaneous utility func- tion w ould take the form u(ct;ct + 1, C t+2,...) w here au/ac, >o for '>t. Loew enstein (1987) ad v anced a formal mod el w hich assumes that a person's in- stantaneous utility is equal to the utility from consumption in that period plus some function of the d iscounted utility of consumption in future period s. Spe- cifically, if w e let v (c) d enote utility from actual consumption, and ' assume this is the same for all period s, then: U(C r;C r + 1,C r + 2,. . ) = V(C ) + C [yV(C + 1) + y2v (c+2) + ...]for somey<1. Loew enstein d escribes how utility from anticipation may play a role in many DU anomalies. Because near-term consumption d eliv ers only consumption utility w hereas future consumption d e- liv ers both consumption utility and an- ticipatory utility, anticipatory utility prov id es a reason to prefer improv e- ment and for getting unpleasant out- comes ov er w ith quickly instead of d elaying them as d iscounting w ould pred ict. It prov id es a possible explana- tion for w hy people d iscount d ifferent good s at d ifferent rates, because utility from anticipation creates a d ow nw ard bias on estimated d iscount rates, and this d ow n- w ard bias is larger for good s that create more anticipatory utility. If, for instance, d read ing future bad outcomes is a stronger emotion than sav oring future good outcomes, w hich seems highly plausible, then utility from anticipation w ould generate a sign effect.24 Finally, anticipatory utility giv es rise to a form of time inconsistency that is quite d ifferent from that w hich arises from hyperbolic d iscounting. Instead of planning to d o the farsighted thing (e.g., sav e money) but subsequently d o- ing the shortsighted thing (splurging), anticipatory utility can cause people to repeated ly plan to consume a good after some d elay that permits pleasurable anticipation, but then to d elay again for the same reason w hen the planned moment of consumption arriv es. Loew enstein's mod el of anticipatory utility applies to d eterministic out- comes. In a recent paper, C aplin and Leahy (2001) point out that many an- ticipatory emotions, such as anxiety or 24Waiting for und esirable outcomes is almost alw ays unpreasant, but w aiting for d esirable out- comes is sometimes pleasurale and sometimes frustrating. Despite the manifest importance for intertemporal choice of these emotions associated w ith w aiting, w e are aw are of no research that has sought to und erstand w hen w aiting for d esirable outcomes is pleasurable or av ersiv e. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 372 Journal of Economic Literature, Vol. XL (June 2002) suspense, are d riv en by uncertainty about the future, and they propose a new mod el that mod ifies expected - utility theory to incorporate such antici- patory emotions. T hey then show that incorporating anxiety into asset-pricing mod els may help explain the equity pre- mium puzzle and the risk-free rate puz- zle, because anxiety creates a taste for risk-free assets and an av ersion to risky assets. Like Loew enstein, C aplin and Leahy emphasize how anticipatory util- ity can lead to time inconsistency. Koszegi (2001) also d iscusses some implications of anticipatory utility. 5.2.4 Visceral Influences A final alternativ e mod el of the utility function incorporates "v isceral" influ- ences such as hunger, sexual d esire, physical pain, crav ings, and such. Loew enstein (1996, 2000b) argues that economics should take more seriously the implications' of such transient fluctuations in tastes. Formally, v isceral influences mean that the person's instantaneous utility function takes the form u(cr,d l) w here d r represents the v ector of v isceral states in period t. Visceral states are (at least to some extent) end ogenous-e.g., a person's current hunger d epend s on how much she has consumed in prev ious period s- and therefore lead to consumption interd epend ence. Visceral influences hav e important implications for intertemporal choice because, by increasing the attractiv e- ness of certain good s or activ ities, they can giv e rise to behav iors that look ex- tremely impatient or ev en impulsiv e. Ind eed , for ev ery v isceral influence, it is easy to think of one or more associ- ated problems of self-control-hunger and d ieting, sexual d esire and v arious "heat-of-the-moment" behav iors, crav - ing and d rug ad d iction, and so on. Vis- ceral influences prov id e an alternate account of the preference rev ersals that are typically attributed to hyperbolic time d iscounting, because the temporal proximity of a rew ard is one of the cues that can activ ate appetitiv e v isceral states (see Laibson 2001; Loew enstein 1996). Other cues-such as spatial prox- imity, the presence of associated smells or sound s, or similarity in current set- ting to historical consumption sites- may also hav e such an effect. T hus, research on v arious types of cues may help to generate new pred ictions about the specific circumstances (other than temporal proximity) that can trigger myopic behav ior. T he fact that v isceral states are end ogenous introd uces issues of state-management (as d iscussed by Loew enstein 1999, and Laibson 2001 und er the rubric of "cue management"). While the mod el (at least the rational v ersion of it) pred icts that a person w ould w ant herself to use d rugs if she w ere to experience a sufficiently strong crav ing, it also pred icts that she might w ant to prev ent ev er experiencing such a strong crav ing. Hence, v isceral influences can giv e rise to a preference for commitment in the sense that the person may w ant to av oid certain situations. Visceral influences may d o more than merely change the instantaneous utility function. First, there is ev id ence that people d on't fully appreciate the effects of v isceral influences, and hence may not react optimally to them (Loew en- stein 1996, 1999, 2000b). When in a hot state, people tend to exaggerate how long the hot state w ill persist, and , w hen in a cold state, people tend to und eresti- mate how much future v isceral influ- ences w ill affect their future behav ior. Second , and perhaps more importantly, people often w ould "prefer" not to re- spond to an intense v isceral factor such as rage, fear, or lust, ev en at the This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 373 moment they are succumbing to its in- fluence. A w ay to und erstand such ef- fects is to apply the d istinction pro- posed by Kahneman (1994) betw een "experienced utility," w hich reflects one's w elfare, and "d ecision utility," w hich reflects the attractiv eness of op- tions as inferred from one's d ecisions. By increasing the d ecision utility of cer- tain types of actions more than the experienced utility of those actions, v is- ceral factors may d riv e a w ed ge be- tw een w hat people d o and w hat makes them happy. Douglas Bernheim and A ntonio R angel (2001) propose a mod el of ad d iction framed in these terms. 5.3 More "Extreme" A lternativ e Perspectiv es T he alternativ e mod els d iscussed abov e mod ify the DU mod el by altering the d iscount function or ad d ing ad d i- tional arguments to the instantaneous utility function. T he alternativ es d is- cussed next inv olv e more rad ical d epartures from the DU mod el. 5.3.1 Projection Bias In many of the alternativ e mod els of utility d iscussed abov e, the person's utility from consumption-her tastes- change ov er time. T o properly make in- tertemporal d ecisions, a person must correctly pred ict how her tastes w ill change. Essentially all economic mod els of changing tastes assume (as econo- mists typically d o) that such pred ictions are correct-that people hav e "rational expectations." How ev er, Loew enstein, O'Donoghue, and R abin (2000) propose that, w hile people may anticipate the qualitativ e nature of their changing preferences, they tend to und erestimate the magnitud e of these changes-a systematic mispred iction they label projection bias. Loew enstein, O'Donoghue, and R abin rev iew a broad array of ev id ence that d emonstrates the prev alence of projec- tion bias, and then mod el it formally. T o illustrate their mod el, consid er pro- jection bias in the realm of habit forma- tion. A s d iscussed abov e, suppose the period -t instantaneous utility function takes the form u(ct;zt), w here zt is a state v ariable that captures the effects of past consumption. Projection bias arises w hen a person w hose current state is zt must pred ict her future utility giv en future state zt. Projection bias implies that the person's pred iction U(c;z, I Zt) w ill lie betw een her true future utility u(ct;zt) and her utility giv en her current state u(ct;zt). A particularly simple functional form is Ui(cr;zr I Zt) = (1 - Oc)u(cr;z) + fcu (C tZt) for some cc E [0,1]. Projection bias may arise w henev er tastes change ov er time, w hether through habit formation, changing ref- erence points, or changes in v isceral states. It can hav e important behav ioral and w elfare implications. For instance, people may und erappreciate the d egree to w hich a present consumption splurge w ill raise their reference consumption lev el, and thereby d ecrease their enjoy- ment of more mod est consumption lev - els in the future. When intertemporal choices are influenced by projection bias, estimates of time preference may be d istorted . 5.3.2 Mental-A ccounting Mod els Some researchers hav e proposed that people d o not treat all money as fungi- ble, but instead assign d ifferent types of expend itures to d ifferent "mental ac- counts" (see T haler 1999 for a recent ov erv iew ). Such mod els can giv e rise to intertemporal behav iors that seem od d w hen v iew ed through the lens of the DU mod el. T haler (1985), for instance, suggests that small amounts of money are cod ed as spend ing money, w hereas larger amounts of money are cod ed as sav ings, and that a person is more This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 374 Journal of Economic Literature, Vol. XL (June 2002) w illing to spend out of the former ac- count. T his accounting rule w ould pre- d ict that people w ill behav e like spend - thrifts for small purchases (e.g., a new pair of shoes), but act more frugally w hen it comes to large purchases (e.g., a new d ining-room table).25 Shlomo Benartzi and T haler (1995) suggest that people treat their financial portfolios as a mental account, and emphasize the importance of how often people "ev alu- ate" this account. T hey argue that if people rev iew their portfolios once a year or so, and if people experience joy or pain from any gains or losses, as as- sumed in Kahneman and T v ersky's (1979) prospect theory, then such "my- opic loss av ersion" represents a plausi- ble explanation for the equity premium puzzle. Prelec and Loew enstein (1998) pro- pose another w ay in w hich mental ac- counting might influence intertemporal choice. T hey posit that payments for consumption confer immed iate d isutil- ity or "pain of paying," and that people keep mental accounts that link the con- sumption of a particular item w ith the payments for it. T hey also assume that people engage in "prospectiv e account- ing." A ccord ing to prospectiv e account- ing, w hen consuming, people think only about current and future payments; past payments d on't cause pain of paying. Likew ise, w hen paying, the pain of pay- ing is buffered only by thoughts of future, but not past, consumption. T he mod el suggests that d ifferent w ays of fi- nancing a purchase can lead to d ifferent d ecisions, ev en hold ing the net present v alue of payments constant. Similarly, a person might hav e d ifferent financing preferences d epend ing on the con- sumption item (e.g., they should prefer to prepay for a v acation that is con- sumed all at once v s. a new car that is consumed ov er many years). T he mod el generates a strong preference for pre- payment (except for d urables), for get- ting paid after rather than before d oing w ork, and for fixed -fee pricing schemes w ith zero marginal costs ov er pay-as- you-go schemes that tightly couple mar- ginal payments to marginal consumption. T he mod el also suggests that interind i- v id ual heterogeneity might arise from d ifferences in the d egree to w hich peo- ple experience the pain of paying rather than d ifferences in time preference. On this v iew , the miser w ho eschew s a fancy restaurant d inner is not d oing so because she explicitly consid ers the d elayed costs of the ind ulgence, but rather because her enjoyment of the d inner w ould be d iminished by the immed iate pain of paying for it. 5.3.3 C hoice Bracketing One important aspect of mental ac- counting is that a person makes at most a few choices at any one time, and gen- erally ignores the relation betw een these choices and other past and future choices. Which choices are consid ered at the same time is a matter of w hat R ead , Loew enstein, and R abin (1999) label "choice bracketing." Intertempo- ral choices, like other choices, can be influenced by the manner in w hich they are bracketed , because d ifferent brack- eting can highlight d ifferent motiv es. T o illustrate, consid er the conflict be- tw een impatience and a preference for improv ement ov er time. Loew enstein and Prelec (1993) d emonstrate that the relativ e importance of these tw o mo- tiv es can be altered by the w ay that 25 While it seems possible that this conceptual- ization could explain the magnitud e effect as w ell, the magnitud e effect is found for v ery "small" amounts (e.g., betw een $2 and $20 in A inslie and Haend el 1983), and for v ery "large amounts" (e.g., betw een, $10,000 and $1,000,000 in R aineri and R achlin 1993). It seems highly unlikely that re- spond ents w ould consistently cod e the low er amounts as spend ing and the higher amounts as sav ings across all of these stud ies. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 375 choices are bracketed . T hey asked one group of subjects to choose betw een hav ing d inner at a fine French restau- rant in one month v s. tw o months. Most subjects chose one month, presumably reflecting impatience. T hey then asked another group to choose betw een eating at home in one month follow ed by eating at the French restaurant in tw o months v s. eating at the French restaurant in one month follow ed by eating at home in tw o months. T he majority now w anted the French d inner in tw o months. For both groups, d inner at home w as the most likely alternativ e to the French d inner, but it w as only w hen the tw o d inners w ere expressed as a sequence that the preference for improv ement became a basis for d ecision. A nalyzing how people frame or bracket choices may help illuminate the issue of w hether a preference for im- prov ement merely reflects the com- bined effect of other motiv es, such as reference d epend ence ror anticipatory utility, or w hether it is something unique. View ed from an integrated d ecision-making perspectiv e, it perhaps seems natural to conclud e that the pref- erence for improv ement is d eriv ativ e of these other concepts, because it is not clear w hy improv ement for its ow n sake should be v aluable. But w hen v iew ed from a choice-bracketing perspectiv e, w herein a person must hav e some choice heuristic for ev aluating sequences, it seems possible that improv ement may be v alued for its ow n sake. Specifically, a preference-for-improv ement choice heuristic may hav e originated from con- sid erations of reference d epend ence or anticipatory utility, but a person using this choice heuristic may come to feel that improv ement for its ow n sake has v alue.26 Loew enstein and Prelec (1993) d e- v elop a (choice-heuristic) mod el for how people ev aluate choices ov er sequences. T hey assume that people consid er a sequence's d iscounted utility, its d egree of improv ement, and its d egree of spread . T he key ingred ients of the mod el are "gestalt" d efinitions for im- prov ement and spread . In other w ord s, they d ev elop a formal measure of the d egree of improv ement and the d egree of spread for any sequence. T hey show that their mod el can explain a w id e range of sequence anomalies, includ ing observ ed v iolations of ind epend ence, and that it pred icts preferences be- tw een sequences much better than other mod els that incorporate similar numbers of free parameters (ev en a mod el w ith an entirely flexible time d iscount function). 5.3.4 Multiple-Self Mod els A n influential school of theorists hav e proposed mod els that v iew intertempo- ral choice as the outcome of a conflict betw een multiple selv es. Most multiple- self mod els postulate myopic selv es w ho are in conflict w ith more farsighted ones, and often d raw analogies betw een intertemporal choice and a v ariety of d ifferent mod els of interpersonal strate- gic interactions. Some mod els (e.g., A inslie and Nick Haslam 1992; T homas 26T hus, to the extent that the preference for improv ement reflects a choice heuristic, it should be susceptible to framing or bracketing effects, because w hat constitutes a sequence is highly sub- jectiv e, as noted by Loew enstein and Prelec 1993 and by John G. Beebe-C enter (1929) sev eral d e- cad es earlier: What enables one to d ecid e w hether a giv en set of affectiv e experiences d oes, or d oes not, constitute a unitary temporal group? . w hat of series inv olv ing experiences of d iffer- ent mod alities- v isual and aud itory ex- periences, for instance? . . . A nd w hat of such complex ev ents as "arising in the morn- ing" or "eating a good meal" or "enjoying a good book?" (Beebe-C enter 1929, p. 67, emphasis ad d ed ) This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 376 Journal of Economic Literature, Vol. XL (June 2002) C . Schelling 1984; Gord on C . Winston 1980) assume that there are tw o agents, one myopic and one farsighted , w ho al- ternately take control of behav ior. T he main problem w ith this approach is that it fails to specify w hy either type of agent emerges w hen it d oes. Further- more, by characterizing the interaction as a battle betw een the tw o agents, these mod els fail to capture an impor- tant asymmetry: farsighted selv es often attempt to control the behav iors of my- opic selv es, but nev er the rev erse. For instance, the farsighted self may pour v od ka d ow n the d rain to prev ent to- morrow 's self from d rinking it, but the myopic self rarely takes steps to ensure that tomorrow 's self w ill hav e access to the alcohol he w ill then crav e. R espond ing, in part, to this problem, T haler and Hersh Shefrin (1981) pro- posed a "planner-d oer" mod el that d raw s upon principal-agent theory. In their mod el, a series of myopic "d oers," w ho care only about their ow n immed i- ate gratification (and hav e no affinity for future or past d oers), interact w ith a unitary "planner" w ho cares equally about the present and future. T he mod el focuses on the strategies em- ployed by the planner to control the behav ior of the d oers. T he mod el high- lights the observ ation, later d iscussed at length by Loew enstein (1996), that the farsighted perspectiv e is often much more constant than the myopic perspec- tiv e. For example, people are often con- sistent in recognizing the need to main- tain a d iet. Yet they period ically v iolate their ow n d esired course of action- often recognizing ev en at the moment of d oing so that they are not behav ing in their ow n self-interest. Yet a third type of multiple-self mod el d raw s connections betw een inter- temporal choice and mod els of multi- person strategic interactions (Elster 1985). T he essential insight that these mod els capture is that, much like coop- eration in a social d ilemma, self-control often requires the cooperation of a se- ries of temporally situated selv es. When one self "d efects" by opting for immed i- ate gratification, the consequence can be a kind of unrav eling or "falling off the w agon" w hen subsequent selv es follow the preced ent. Few of these multiple-self mod els hav e been expressed formally, and ev en few er hav e been used to d eriv e testable implications that go much beyond the intuitions that inspired them in the first place. How ev er, perhaps it is unfair to criticize the mod els for these short- comings. T hese mod els are probably best v iew ed as metaphors intend ed to high- light specific aspects of intertemporal choice. Specifically, multiple-self mod - els hav e been used to make sense of the w id e range of self-control strategies that people use to regulate their ow n future behav ior. Moreov er, these mod - els prov id ed much of the inspiration for more recent formal mod els of sophisti- cated hyperbolic d iscounting (follow ing Laibson 1994, 1997). 5.3.5 T emptation Utility Most mod els of intertemporal choice- ind eed , most mod els of choice in any framew ork-assume that options not chosen are irrelev ant to a person's w ell- being. In a recent paper, Gul and Pesend orfer (2001) posit that people hav e "temptation preferences," w herein they experience d isutility from not choosing the option that is most enjoy- able now . T heir theory implies that a person might be better off if some particularly tempting option w ere not av ailable, ev en if she d oesn't choose that option. A s a result, she may be w ill- ing to pay in ad v ance to eliminate that option, or in other w ord s, she may hav e a preference for commitment. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 377 5.3.6 C onclusion: C ombining Insights from Different Mod els Many behav ioral mod els of intertem- poral choice focus on a single mod ifica- tion to the DU mod el and explore the ad d itional realism prod uced by that single mod ification. But many empirical phenomena reflect the interaction of multiple phenomena. For instance, a preference for improv ement may inter- act w ith hyperbolic d iscounting to pro- d uce preferences for U-shaped sequences- e.g., for jobs that offer a signing bonus and a salary that increases grad ually ov er time. A s d iscussed by Loew enstein and Prelec (1993), in the short term, the preference-for-improv ement motiv e is sw amped by the high d iscount rates, but as the d iscount rate falls ov er time, the preference-for-improv ement motiv e may gain ascend ance and cause a net preference for an increasing payment sequence. A s another example, introd ucing v is- ceral influences into mod els of hyper- bolic d iscounting may more fully account for the phenomenology of impulsiv e choices. Hyperbolic-d iscounting mod els pred ict that people respond especially strongly to immed iate costs and benefits, and v isceral influences hav e pow erful transient effects on immed iate utilities. In combination, the tw o assumptions could explain a w id e range of impulsiv e choices and other self-control phenomena. 6. Measuring T ime Discounting T he DU mod el assumes that a per- son's time preference can be captured by a single d iscount rate, p. Ov er the past three d ecad es, there hav e been many attempts to measure this rate. Some of these estimates are d eriv ed from observ ations of "real-w orld " be- hav iors (e.g., the choice betw een elec- trical appliances that d iffer in their initial purchase price and long-run op- erating costs). Others are d eriv ed from experimental elicitation proced ures (e.g., respond ents' answ ers to the ques- tion "Which w ould you prefer: $100 tod ay or $150 one year from tod ay?"). T able 1 summarizes the implicit d is- count rates from all stud ies that w e could locate in w hich d iscount rates w ere either d irectly reported or easily computed from the reported d ata. Figure 2 plots the estimated d iscount factor for each stud y against the publi- cation d ate for that stud y, w here the d is- count factor is 6 = 1/(1 + p).27 T his figure rev eals three notew orthy observ ations. First, there is tremend ous v ariability in the estimates (the correspond ing im- plicit annual d iscount rates range from -6 percent to infinity). Second , in con- trast to estimates of physical phenom- ena such as the speed of light, there is no ev id ence of method ological progress; the range of estimates is not shrinking ov er time. T hird , high d iscounting pred ominates, as most of the d ata points are w ell below 1, w hich repre- sents equal w eighting of present and future. In this section, w e prov id e an ov er- v iew and critique of this empirical lit- erature w ith an eye tow ard und er- stand ing these three observ ations. We first d iscuss a v ariety of confound ing factors, such as intertemporal arbitrage, uncertainty, and expectations of chang- ing utility functions. T hese consid era- tions typically are not regard ed as legiti- mate components of time preference per se, but they can affect both experi- mental responses and real-w orld choices. With these confound ing factors in mind , w e then rev iew the proced ures used to estimate d iscount rates. T his section reiterates our general theme: T o truly und erstand intertemporal choices, 27 In some cases, the estimates are computed from the med ian respond ent. In other cases, the authors reported the mean d iscount rate. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions T A BLE 1 EMPIR IC A L EST IMA T ES OF DISC OUNT R A T ES Stud y T ype Good (s) R eal or Hypo? Elicitation Method Maital & Maital 1978 experimental money & coupons hypo. choice Hausman 1979 field money real choice Gateley 1980 field money real choice T haler 1981 experimental money hypo. matching A inslie & Haend el 1983 experimental money real matching Houston 1983 experimental money hypo. other Loew enstein 1987 experimental money & pain hypo. pricing Moore and Viscusi 1988 field life years real choice Benzion et al. 1989 experimental money hypo. matching Viscusi & Moore 1989 field life years real choice Moore & Viscusi 1990a field life years real choice Moore & Viscusi 1990b field life years real choice Shelley 1993 experimental money hypo. matching R ed elmeier & Heller 1993 experimental health hypo. rating C airns 1994 experimental money hypo. choice Shelley 1994 experimental money hypo. rating C hapman & Elstein 1995 experimental money & health hypo. matching Dolan & Gud ex 1995 experimental health hypo. other Dreyfus and Viscusi 1995 field life years real choice Kirby & Marakov ic 1995 experimental money real matching C hapman 1996 experimental money & health hypo. matching Kirby & Marakov ic 1996 experimental money real choice Pend er1996 experimental rice real ohoice Wahlund & Gunnarson 1996 experimental money hypo. matching C airns & v an d er Pol 1997 experimental money hypo. matching Green, Myerson & experimental money hypo. choice McFad d en 1997 Johanneson & Johansson experimental life years hypo. pricing 1997 Kirby 1997 experimental money real pricing Mad d en et al. 1997 experimental money & heroin hypo. choice C hapman & Winquist 1998 experimental money hypo. matching Hold en, Shiferaw & Wik experimental money & corn real matching 1998 C airns & v an d er Pol 1999 experimental health hypo. matching C hapman, Nelson & Hier experimental money & health hypo. choice 1999 C oller & Williams 1999 experimental money real choice Kirby, Petry & Bickel 1999 experimental money real choice v an d er Pol & C airns 1999 experimental health hypo. choice C hesson & Viscusi 2000 experimental money hypo. matching Ganiats et al. 2000 experimental health hypo. choice Hesketh 2000 experimental money hypo. choice v an d er Pol & C airns 2001 experimental health hypo. choice Warner & Pleeter 2001 field money real choice Harrison, Lau & Williams experimental money real choice 2002 This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions T A BLE 1 (C ont.) A nnual Discount Stud y T ime R ange A nnual Discount R ate(s) Factor(s) Maital & Maital 1978 1 year 70% 0.59 Hausman 1979 und efined 5% to 89% 0.95 to 0.53 Gateley 1980 und efined 45% to 300% 0.69 to 0.25 T haler 1981 3 mos. to 10 yrs. 7% to 345% 0.93 to 0.22 A inslie & Haend el 1983 und efined 96000% to 0 0.00 Houston 1983 1 yr. to 20 yrs. 23% 0.81 Loew enstein 1987 immed iately to 10 yrs. -6% to 212% 1.06 to 0.32 Moore and Viscusi 1988 und efined 10% to 12% 0.91 to 0.89 Benzion et al. 1989 6 mos. to 4 yrs. 9% to 60% 0.92 to 0.63 Viscusi & Moore 1989 und efined 11% 0.90 Moore & Viscusi 1990a und efined 2% 0.98 Moore & Viscusi 1990b und efined 1% to 14% 0.99 to 0.88 Shelley 1993 6 mos. to 4 yrs. 8% to 27% 0.93 to 0.79 R ed elmeier & Heller 1993 1 d ay to 10 yrs. 0% 1.00 C airns 1994 5 yrs. to 20 yrs. 14% to 25% 0.88 to 0.80 Shelley 1994 6 mos. to 2 yrs. 4% to 22% 0.96 to 0.82 C hapman & Elstein 1995 6 mos. to 12 yrs. 11% to 263% 0.90 to 0.28 Dolan & Gud ex 1995 1 month to 10 yrs. 0% 1.00 Dreyfus and Viscusi 1995 und efined 11% to 17% 0.90 to 0.85 Kirby & Marakov ic 1995 3 d ays to 29 d ays 3678% to - 0.03 to 0.00 C hapman 1996 1 yr. to 12 yrs. negativ e to 300% 1.01 to 0.25 Kirby & Marakov ic 1996 6 hours to 70 d ays 500% to 1500% 0.17 to 0.06 Pend er 1996 7 mos. to 2 yrs. 26% to 69% 0.79 to 0.59 Wahlund & Gunnarson 1996 1 month to 1 yr. 18% to 158% 0.85 to 0.39 C airns & v an d er Pol 1997 2 yrs. to 19 yrs. 13% to 31% 0.88 to 0.76 Green, Myerson & 3 mos. to 20 yrs. 6% to 111% 0.94 to 0.47 McFad d en 1997 Johanneson & Johansson 6 yrs. to 57 yrs. 0% to 3% 0.97 1997 Kirby 1997 1 d ay to 1 month 159% to 5747% 0.39 to 0.02 Mad d en et al. 1997 1 w eek to 25 yrs. 8% to 0 0.93 to 0.00 C hapman & Winquist 1998 3 months 426% to 2189% 0.19 to 0.04 Hold en, Shiferaw & Wik 1 yr. 28% to 147% 0.78 to 0.40 1998 C airns & v an d er Pol 1999 4 yrs. to 16 yrs. 6% 0.94 C hapman, Nelson & Hier 1 month to 6 mos. 13% to 19000% 0.88 to 0.01 1999 C oller & Williams 1999 1 month to 3 mos. 15% to 25% 0.87 to 0.80 Kirby, Petry & Bickel 1999 7 d ays to 186 d ays 50% to 55700% 0.67 to 0.00 v an d er Pol & C airns 1999 5 yrs. to 13 yrs. 7% 0.93 C hesson & Viscusi 2000 1 year to 25 yrs. 11% 0.90 Ganiats et al. 2000 6 mos. to 20 yrs. negativ e to 116% 1.01 to 0.46 Hesketh 2000 6 mos. to 4 yrs. 4% to 36% 0.96 to 0.74 v an d er Pol & C airns 2001 2 yrs. to 15 yrs. 6% to 9% 0.94 to 0.92 Wamner & Pleeter 2001 immed iately to 22 yrs. 0% to 71% 0 to 0.58 Harrison, Lau & Williams 1 month to 37 mos. 28% 0.78 2002 This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 380 Journal of Economic Literature, Vol. XL (June 2002) 1.0 - * . 4 x 10.8 * 0.60- * * * 0.8 * )S~, 0.4 - $ 0.2 - 0.0 1975 1980 1985 1990 1995 2000 year of publication Figure 2. Discount Factor by Year of Stud y Publication one must recognize the influence of many consid erations besid es pure time preference. 6.1 C onfound ing Factors A w id e v ariety of proced ures hav e been used to estimate d iscount rates, but most apply the same basic ap- proach. Some actual or reported in- tertemporal preference is observ ed , and researchers then compute the d iscount rate that this preference implies, using a "financial" or net present v alue (NPV) calculation. For instance, if a person d emonstrates ind ifference betw een 100 w id gets now and 120 w id gets in one year, the implicit (annual) d iscount rate, p, w ould be 20 percent, because that v alue w ould satisfy the equation 100 = (1/(1 + p))120. Similarly, if a person is ind ifferent betw een an ineffi- cient low -cost appliance and a more efficient one that costs $100 extra but sav es $20 a year in electricity ov er the next ten years, the implicit d iscount rate, p, w ould equal 15.1 percent, be- cause that v alue w ould satisfy the equation 100 = =0 1(1/(l + p)) t20. A lthough this is an extremely w id e- spread approach for measuring d iscount rates, it relies on a v ariety of ad d itional (and usually implicit) assumptions, and is subject to sev eral confound ing factors. 6.1.1 C onsumption R eallocation T he calculation outlined abov e as- sumes a sort of "isolation" in d ecision making. Specifically, it treats the ob- jects of intertemporal choice as d is- crete, unitary, d ated ev ents; it assumes that people entirely "consume" the re- w ard (or penalty) at the moment it is receiv ed , as if it w ere an instantaneous burst of utility. Furthermore, it assumes that people d on't shift consumption around ov er time in anticipation of the receipt of the future rew ard or penalty. T hese assumptions are rarely exactly correct, and may sometimes be bad approximations. C hoosing betw een $50 tod ay v ersus $100 next year, or choos- ing betw een 50 pound s of corn tod ay v ersus 100 pound s next year, are not the same as choosing betw een 50 utils tod ay and 100 utils on the same d ay next year, as the calculations imply. R ather, they are more complex choices betw een the v arious streams of con- sumption that those tw o d ated rew ard s make possible. 6.1.2 Intertemporal A rbitrage In theory, choices betw een trad able rew ard s, such as money, should not re- v eal anything about time preferences. A s Victor Fuchs (1982) and others hav e noted , if capital markets operate effec- tiv ely (if monetary amounts at d ifferent times can be costlessly exchanged at a specified interest rate), choices be- tw een d ated monetary outcomes can be red uced to merely selecting the rew ard w ith the greatest net present v alue (using the market interest rate).28 T o 28 Meyer (1976) expresses this point: ". . . if w e can lend and borrow at the same rate . . ., then w e can simply show that, regard less of the fund a- mental ord erings on the c's [consumption streams], the ind uced ord ering on the x's [se- quences of monetary flow s] is giv en by simple d is- counting at this giv en rate.. . We could say that the mar et assumes command and the market rate prev ails for monetary flow s." This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 381 illustrate, suppose a person prefers $100 now to $200 ten years from now . While this preference could be ex- plained by imputing a d iscount rate on future utility, the person might be choosing the smaller immed iate amount because she believ es that through proper inv estment she can turn it into more than $200 in ten years, and thus enjoy more than $200 w orth of con- sumption at thatfuture time. T he pres- ence of capital markets should cause imputed d iscount rates to conv erge on the market interest rate. Stud ies that impute d iscount rates from choices among trad able rew ard s assume that respond ents ignore oppor- tunities for intertemporal arbitrage, either because they are unaw are of capital markets or unable to exploit them.29 T he latter assumption may sometimes be correct. For instance, in field stud ies of electrical-appliance pur- chases, some subjects may hav e faced borrow ing constraints that prev ented them from purchasing the more expen- siv e energy-efficient appliances. More typically, how ev er, imperfect capital markets cannot explain choices; they cannot explain w hy a person w ho hold s sev eral thousand d ollars in a bank ac- count earning 4-percent interest should prefer $100 tod ay ov er $150 in one year. Because imputed d iscount rates d o not, in fact, conv erge on the prev ail- ing market interest rates, but instead are much higher, it seems that many re- spond ents are neglecting capital mar- kets and basing their choices on some other consid eration, such as time pref- erence or the uncertainty associated w ith d elay. 6.1.3 C oncav e Utility T he stand ard approach to estimating d iscount rates assumes that the utility function is linear in the magnitud e of the choice objects (e.g., amounts of money, pound s of corn, d uration of some health state). If, instead , the utility function for the good in question is concav e, estimates of time preference w ill be biased upw ard . For example, ind ifference betw een $100 this year and $200 next year implies a d ollar d iscount rate of 100 percent. How ev er, if the utility of acquiring $200 is less than tw ice the utility of acquiring $100, the utility d iscount rate w ill be less than 100 percent. T his confound is rarely d iscussed , perhaps because utility is as- sumed to be approximately linear ov er the small amounts of money commonly used in time-preference stud ies. T he ov erw helming ev id ence for reference- d epend ent utility suggests, how ev er, that this assumption may be inv alid - that people may not be integrating the stated amounts w ith their current and future w ealth, and therefore that curv a- ture in the utility function may be substantial ev en for these small amounts (see Ian Bateman et al. 1997; Dav id W. Harless and C olin F. C amerer 1994; Kahneman and T v ersky 1979; R abin 2000; R abin and T haler 2001; T v ersky and Kahneman 1991). T hree techniques could be used to av oid this confound . (1) One could re- quest d irect utility jud gments (e.g., at- tractiv eness ratings) of the same conse- quence at tw o d ifferent times. T hen, the ratio of the attractiv eness rating of 29 A rguments about v iolations of the d iscounted utility mod el assume, as Pend er (1996, pp. 282- 83) notes, "that the results of d iscount rate ex- periments rev eal something about intertemporal preferences d irectly. How ev er, if agents are opti- mizing an intertemporal utility function, their op- portunities for intertemporaY arbitrage are also important in d etermining how they respond to such experiments . . . w hen trad able rew ard s are offered , one must either aband on the assumption that respond ents in experimental stud ies are opti- mizing, or make some assumptions (either implicit or explicit) about the nature of cred it markets. T he implicit assumption in some of the prev ious stud - ies of d iscount rates appears to be that there are no possibilities for intertemporal arbitrage. ... This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 382 Journal of Economic Literature, Vol. XL (June 2002) the d istant outcome to the proximate outcome w ould d irectly rev eal the im- plicit d iscount factor. (2) T o the extent that utility is linear in probability, one can use choices or jud gment tasks in- v olv ing d ifferent probabilities of the same consequence at d ifferent times (A lv in E. R oth and J. Keith Murnighan 1982). Ev id ence that probability is w eighted nonlinearly (see, e.g., Starmer 2000) w ould , of course, cast d oubt on this approach. (3) One can separately elicit the utility function for the good in question, and then use that function to transform outcome amounts into utility amounts, from w hich utility d iscount rates could be computed . T o our know l- ed ge, C hapman (1996) cond ucted the only stud y that attempted to d o this. She found that utility d iscount rates w ere substantially low er than the d ollar d is- count rates, because utility w as strongly concav e ov er the monetary amounts subjects used in the intertemporal choice tasks.30 6.1.4 Uncertainty In experimental stud ies, subjects are typically instructed to assume that d e- layed rew ard s w ill be d eliv ered w ith certainty. It is unclear w hether subjects d o (or can) accept this assumption, because d elay is ord inarily-and perhaps un- av oid ably-associated w ith uncertainty. A similar problem arises for field stud - ies, in w hich it is typically assumed that subjects believ e that future rew ard s, such as energy sav ings, w ill materialize. Because of this subjectiv e (or "epistemic") uncertainty associated w ith d elay, it is d ifficult to d etermine to w hat extent the magnitud e of imputed d iscount rates (or the shape of the d is- count function) is gov erned by time preference per se, v ersus the d iminu- tion in subjectiv e probability associated w ith d elay.31 Empirical ev id ence suggests that in- trod ucing objectiv e (or "aleatory") un- certainty to both current and future re- w ard s can d ramatically affect estimated d iscount rates. For instance, Gid eon Keren and Peter R oelofsma (1995) asked one group of respond ents to choose betw een 100 florins (a Nether- land s unit of currency) immed iately and 110 florins in one month, and another group to choose betw een a 50-percent chance of 100 florins immed iately and a 50-percent chance of 110 florins in one month. While 82 percent preferred the smaller immed iate rew ard w hen both rew ard s w ere certain, only 39 percent preferred the smaller immed iate rew ard w hen both rew ard s w ere uncertain.32 A lso, A lbrecht and Weber (1996) found that the present v alue of a future lottery (e.g., a 50-percent chance of receiv ing 250 d eutsche marks) tend ed to exceed the present v alue of its certainty equiv alent. 6.1.5 Inflation T he stand ard approach assumes that, for instance, $100 now and $100 in fiv e years generate the same lev el of utility at the times they are receiv ed . How ev er, 30 C hapman also found that magnitud e effects w ere much smaller after correcting for utility function curv ature. T his result supports Loew en- stein and Prelec's (1992) explanation of magnitud e effects as resulting from utility function curv ature (see section 5.2.2). 31 T here may be complicated interactions be- tw een risk and d elay, because uncertainty about future receipt complicates and imped es the plan- ning of one's future consumption stream (Michael S pence and R ichard Zeckhauser 1972). For exam- pIe, a 90-percent chance to w in $10,000,000 in fifteen years is w orth much less than a guarantee to receiv e $9,000,000 at that time, because, to the extent that the person cannot insure against the resid ual uncertainty, there is a limit to how much she can ad just her consumption lev el d uring those fifteen years. 32 T his result cannot be explained by a magni- tud e effect on the expected amounts, because 50 percent of a rew ard has a smaller expected v alue, and , accord ing to the magnitud e effect, should be d iscounted more, not less. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 383 inflation prov id es a reason to d ev alue future monetary outcomes, because in the presence of inflation, $100 w orth of consumption now is more v aluable than $100 w orth of consumption in fiv e years. T his confound creates an upw ard bias in estimates of the d iscount rate, and this bias w ill be more or less pro- nounced d epend ing on subjects' ex- periences w ith and expectations about inflation. 6.1.6 Expectations of C hanging Utility A rew ard of $100 now might also gen- erate more utility than the same amount fiv e years hence because a person ex- pects to hav e a larger baseline con- sumption lev el in fiv e years (e.g., d ue to increased w ealth). A s a result, the mar- ginal utility generated by an ad d itional $100 of consumption in fiv e years may be less than the marginal utility gener- ated by an ad d itional $100 of consump- tion now . Like inflation, this confound creates an upw ard bias in estimates of the d iscount rate. 6.1.7 Habit Formation, A nticipatory Utility, and Visceral Influences T o the extent that the d iscount rate is meant to reflect only time preference, and not the confluence of all factors influencing intertemporal choice, the mod ifications to the instantaneous util- ity function d iscussed in section 5 rep- resent ad d itional biasing factors, be- cause they are typically not accounted for w hen the d iscount rate is imputed . For instance, if anticipatory utility moti- v ates one to d elay consumption more than one otherw ise w ould , the imputed d iscount rate w ill be low er than the true d egree of time preference. If a person prefers an increasing consump- tion profile d ue to habit formation, the d iscount rate w ill be biased d ow nw ard . Finally; if the prospect of an immed iate rew ard momentarily stimulates v isceral factors that temporarily increase the person's v aluation of the proximate re- w ard , the d iscount rate could be biased upw ard .33 6.1.8 A n Illustrativ e Example T o illustrate the d ifficulty of sepa- rating time preference per se from these potential confound s, consid er a prototypical stud y by Benzion, R apoport, and Yagil (1989). In this stud y, respon- d ents equated immed iate sums of money and larger d elayed sums (e.g., they specified the rew ard in six months that w ould be as good as getting $1000 im- med iately). In the cov er story for the questionnaire, respond ents w ere asked to imagine that they had earned money (amounts ranged from $40 to $5000), but w hen they arriv ed to receiv e the payment they w ere told that the "financially solid " public institute is "temporarily short of fund s." T hey w ere asked to specify a future amount of money (d e- lays ranged from six months to four years) that w ould make them ind iffer- ent to the amount they had been prom- ised to receiv e immed iately. Surely, the d escription "financially solid " could scarcely be sufficient to allay uncertain- ties that the future rew ard w ould actu- ally be receiv ed (particularly giv en that the institute w as "temporarily" short of fund s), and it seems likely that re- sponses includ ed a substantial "risk premium." Moreov er, the subjects in this stud y had "extensiv e experience w ith . . . a three-d igit inflation rate," 33It is unclear w hether v isceral factors should be consid ered a d eterminant of time preference or a confound ing factor in its estimation. If v isceral factors increase the attractiv eness of an immed iate rew ard w ithout affecting its experienced enjoy- ment (if they increase w anting but not liking), they are probably best v iew ed as a legitimate d eterminant of time perference. If, how ev er, v isceral factors alter the amount of utility that a contemplated proximate rew ard actually d eliv ers, they might best be regard ed as a confound ing factor. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 384 Journal of Economic Literature, Vol. XL (June 2002) and respond ents might w ell hav e con- sid ered inflation w hen generating their responses. Ev en if respond ents assumed no inflation, the real interest rate d ur- ing this time w as positiv e, and they might hav e consid ered intertemporal arbitrage. Finally, respond ents may hav e consid ered that their future w ealth w ould be greater and that the later re- w ard w ould therefore yield less mar- ginal utility. Ind eed , the instructions cued respond ents to consid er this, as they w ere told that the questions d id not hav e correct answ ers, and that the answ ers "might v ary from one ind iv id - ual to another d epend ing on his or her present or future financial assets." Giv en all of these confound ing fac- tors, is it unclear exactly how much of the imputed annual d iscount rates (w hich ranged from 9 percent to 60 per- cent) actually reflected time prefer- ence. It is possible that the responses in this stud y (and others) can be entirely explained in terms of these confound s, and that once these confound s are con- trolled for, no "pure" time preference w ould remain. 6.2 Proced ures for Measuring Discount R ates We d iscussed abov e sev eral con- found ing factors that greatly complicate the assignment of a d iscount rate to a particular choice or jud gment. With these confound s in mind , w e next d is- cuss the method s that hav e been used to measure d iscount rates. Broad ly, these method s can be d iv id ed into tw o categories: field stud ies, in w hich d is- count rates are inferred from economic d ecisions that people make in their or- d inary life, and experimental stud ies, in w hich people are asked to ev aluate styl- ized intertemporal prospects inv olv ing real or hypothetical outcomes. T he d if- ferent proced ures are each subject to the confound s d iscussed abov e, and , as w e shall d iscuss, are also influenced by a v ariety of other factors that are theoretically irrelev ant, but w hich can greatly affect the imputed d iscount rate. 6.2.1 Field Stud ies Some researchers hav e estimated d is- count rates by id entifying real-w orld behav iors that inv olv e trad eoffs be- tw een the near future and more d istant future. Early stud ies of this type exam- ined consumers' choices among d iffer- ent mod els of electrical appliances, w hich presented purchasers w ith a trad eoff betw een the immed iate pur- chase price and the long-term costs of running the appliance (as d etermined by its energy efficiency). In these stud ies, the d iscount rates implied by consum- ers' choices v astly exceed ed market in- terest rates and d iffered substantially across prod uct categories. T he implicit d iscount rate w as 17-20 percent for air cond itioners (Jerry Hausman 1979); 102 percent for gas w ater heaters, 138 per- cent for freezers, 243 percent for elec- tric w ater heaters (H. R ud erman, M. D. Lev ine, and J. E. McMahon 1987); and from 45 percent to 300 percent for refrigerators, d epend ing on assump- tions mad e about the cost of electricity (Dermot Gately 1980).34 34T hese find ings illustrate how people seem to ignore intertemporal arbitrage. A s Hausman (1979) noted , it d oes not make sense for anyone w ith positiv e sav ings to d iscount future energy sav - ings at rates higher than the market interest rate. One possible explanation for these results is that people are liquid ity constrained . C onsistent w ith such an account, Hausman found that the d iscount rate v aried marked ly w ith income-it w as 39 per- cent for household s w ith und er $10,000 of income, but just 8.9 percent for household s earning be- tw een $25,000 and $35,000. How ev er, conflicting w ith this find ing, a stud y by Douglas Houston (1983) that presented ind iv id uals w ith a d ecision of w hether to purchase a hypothetical "energy- sav ing" d ev ice, found that income "played no sta- tistically significant role in explaining the lev el of d iscount rate." This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 385 A nother set of stud ies imputes d is- count rates from w age-risk trad eoffs, in w hich ind iv id uals d ecid e w hether to accept a riskier job w ith a higher salary. Such d ecisions inv olv e a trad eoff be- tw een quality of life and expected length of life. T he more that future utility is d iscounted , the less important is length of life, making risky but high-paying jobs more attractiv e. From such trad e- offs, W. Kip Viscusi and Michael Moore (1989) conclud ed that w orkers' implicit d iscount rate w ith respect to future life years w as approximately 11 percent. Later, using d ifferent econometric ap- proaches w ith the same d ata set, Moore and Viscusi (1990a) estimated the d is- count rates to be around 2 percent, and Moore and Viscusi (1990b) conclud ed that the d iscount rate w as somew here betw een 1 percent and 14 percent. Mark Dreyfus and Viscusi (1995) ap- plied a similar approach to auto-safety d ecisions and estimated d iscount rates ranging from 11 percent to 17 percent. In the macroeconomics literature, re- searchers hav e imputed d iscount rates by estimating structural mod els of life- cycle sav ing behav ior. For instance, Emily Law rence (1991) used Euler equations to estimate household time preferences across d ifferent socioeco- nomic groups. She estimated the d is- count rate of med ian-income house- hold s to be betw een 4 percent and 13 percent d epend ing on the specification. C hristopher C arroll (1997) criticizes Euler-equation estimation on the ground s that most household s tend to engage mainly in "buffer-stock" sav ing early in their liv es-they sav e primarily to be prepared for emergencies-and only cond uct "retirement" sav ing later on. R ecent papers hav e estimated rich, calibrated , stochastic mod els in w hich household s cond uct buffer-stock sav ing early in life and retirement sav ing later in life. Using this approach, C arroll and A nd rew Samw ick (1997) report point estimates for the d iscount rate ranging from 5 percent to 14 percent, and Pierre-Oliv ier Gourinchas and Jonathan Parker (2001) report point estimates of 4.0-4.5 percent. Field stud ies of this type hav e the ad v antage of not assum- ing isolation, because integrated d eci- sion making is built into the mod el. But such estimates often d epend heav ily on the myriad assumptions includ ed in the structural mod el.35 R ecently, John Warner and Saul Pleeter (2001) analyzed d ecisions mad e by U.S. military serv icemen. A s part of military d ow nsizing, ov er 60,000 mili- tary employees w ere giv en the choice betw een a one-time, lump-sum pay- ment and an annuity payment. T he sizes of the payments d epend ed on the em- ployee's current salary and number of years of serv ice-e.g., an "E-5" w ith nine years of serv ice could choose be- tw een $22,283 now v s. $3,714 ev ery year for eighteen years. In general, the present v alue of the annuity payment equaled the lump-sum payment for a d iscount rate of 17.5 percent. A lthough the interest rate w as only 7 percent at the time of these d ecisions, ov er half of all military officers and ov er 90 percent of enlisted personnel chose the lump- sum payment.36 T his stud y is particu- larly compelling in terms of cred ibility of rew ard d eliv ery, magnitud e of stakes, and number of subjects.37 35 T hese macroeconomics stud ies are not in- clud ed in the tables and figures, w hich focus pri- marily on ind iv id ual lev el choice d ata. 36 It should be noted , how ev er, that the guaran- teed payments in the annuity program w ere not ind exed for inflation, w hich av eraged 4.2 percent d uring the four years preced ing this choice. 37 Warner and Pleeter (2001) noted that if ev eryone had chosen the annuity payment, the present v alue of all payments w ould hav e been $4.2 billion. Giv en the choices, how ev er, the present v alue of the gov ernment payout w as just 2.5 billion. T hus, offering the lump-sum alternativ e sav ed the fed eral gov ernment $1.7 billion d ollars. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 386 Journal of Economic Literature, Vol. XL (June 2002) T he benefit of field stud ies, as com- pared w ith experimental stud ies, is their high ecological v alid ity. T here is no concern about w hether estimated d iscount rates w ould apply to real be- hav ior because they are estimated from such behav ior. But field stud ies are sub- ject to ad d itional confound s d ue to the complexity of real-w orld d ecisions and the inability to control for some impor- tant factors. For example, the high d is- count rates implied by the w id espread use of inefficient electrical appliances might not result from the d iscounting of future cost sav ings per se, but from other consid erations, includ ing: (1) a lack of information among consumers about the cost sav ings of the more effi- cient appliances; (2) a d isbelief among consumers that the cost sav ings w ill be as great as promised ; (3) a lack of ex- pertise in translating av ailable informa- tion into economically efficient d eci- sions; or (4) hid d en costs of the more efficient appliances, such as red uced conv enience or reliability, or, in the case of light bulbs, because the more effi- cient bulbs generate a less aesthetically pleasing light spectra.38 6.2.2 Experimental Stud ies Giv en the d ifficulties of interpreting field d ata, the most common method ol- ogy for eliciting d iscount rates is to so- licit "paper-and -pencil" responses to the prospect of real and hypothetical re- w ard s and penalties. Four experimental proced ures are commonly used : choice tasks, matching tasks, pricing tasks, and ratings tasks. C hoice tasks are the most common experimental method for eliciting d is- count rates. In a typical choice task, subjects are asked to choose betw een a smaller, more immed iate rew ard and a larger, more d elayed rew ard . Of course, a single choice betw een tw o intertem- poral options only rev eals an upper or low er bound on the d iscount rate-for example, if a person prefers 100 units of something tod ay ov er 120 units a year from tod ay, the choice merely im- plies a d iscount rate of at least 20 per- cent per year. T o id entify the d iscount rate more precisely, researchers often present subjects w ith a series of choices that v ary the d elay or the amount of the rew ard s. Some stud ies use real rew ard s, includ ing money, rice, and corn. Other stud ies use hypothetical rew ard s, includ - ing monetary gains and losses, and more or less satisfying jobs av ailable at d ifferent times. (See table 1 for a list of the proced ures and rew ard s used in the d ifferent stud ies.) Like all experimental elicitation pro- ced ures, the results from choice tasks can be affected by proced ural nuances. A prev alent problem is an anchoring effect: w hen respond ents are asked to make multiple choices betw een imme- d iate and d elayed rew ard s, the first choice they face often influences sub- sequent choices. For instance, people w ould be more prone to choose $120 next year ov er $100 immed iately if they first chose betw een $100 immed iately and $103 next year than if they first chose betw een $100 immed iately and $140 next year. In general, imputed d is- count rates tend to be biased in the d i- rection of the d iscount rate that w ould equate the first pair of options to w hich they are exposed (see Donald Green et al. 1998). A nchoring effects can be minimized by using titration proced ures that expose respond ents to a series of opposing anchors-e.g., (1) $100 tod ay or $101 in one year? (2) $100 tod ay or $10,000 in one year? (3) $100 tod ay or $105 in one year? and so on. Because titration proced ures typically only offer 38 For a criticism of the hid d en-costs explana- tion, how ev er, see Jonathan Koomey and A lan Sanstad (1994) and R ichard How arth and Sanstad (1995). This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 387 choices betw een an immed iate rew ard and a greater future rew ard , how ev er, ev en these proced ures communicate to respond ents that they should be d is- counting, and potentially bias d iscount rates upw ard . Matching tasks are another popular method for eliciting d iscount rates. In matching tasks, respond ents "fill in the blank" to equate tw o intertemporal options (e.g., $100 now = in one year). Matching tasks hav e been cond ucted w ith real and hypothetical monetary outcomes and w ith hypotheti- cal av ersiv e health cond itions (again see table 1 for a list of the proced ures and rew ard s used in d ifferent stud ies). Matching tasks hav e tw o ad v antages ov er choice tasks. First, because sub- jects rev eal an ind ifference point, an exact d iscount rate can be imputed from a single response. Second , because the intertemporal options are not fully specified , there is no anchoring prob- lem and no suggestion of an expected d iscount rate (or range of d iscount rates). T hus, unlike choice tasks, matching tasks cannot be accused of simply recov ering the expectations of the experimenters that guid ed the experimental d esign. A lthough matching tasks hav e some ad v antages ov er choice tasks, there are reasons to be suspicious of the re- sponses obtained . First, responses often appear to be gov erned by the applica- tion of some simple rule rather than by time preference. For example, w hen people are asked to state the amount in n years that equals $100 tod ay, a v ery common response is $100*n. Second , the responses are often v ery "coarse"- often multiples of tw o or ten of the im- med iate rew ard , suggesting that respon- d ents d o not (or cannot) think v ery carefully about the task. T hird , and most importantly, there are large d iffer- ences in imputed d iscount rates among sev eral theoretically equiv alent proce- d ures. T w o intertemporal options could be equated or matched in one of four w ays: R espond ents could be asked to specify (1) the amount of a d elayed re- w ard that w ould make it as attractiv e as a giv en immed iate rew ard (w hich is the most common technique); (2) the amount of an immed iate rew ard that makes it as attractiv e as a giv en d elayed rew ard (A lbrecht and Weber 1996); (3) the maximum length of time they w ould be w illing to w ait to receiv e a larger re- w ard in lieu of an immed iately av ailable smaller rew ard (A inslie and Haend el 1983; R oelofsma 1994); or (4) the latest d ate at w hich they w ould accept a smaller rew ard in lieu of receiv ing a larger rew ard at a specified d ate that is later still. While there is no theoretical basis for preferring one of these method s ov er any other, the small amount of empiri- cal ev id ence comparing d ifferent meth- od s suggests that they yield v ery d iffer- ent d iscount rates. R oelofsma (1994) found that implicit d iscount rates v aried tremend ously d epend ing on w hether re- spond ents matched on amount or time. One group of subjects w as asked to in- d icate how much compensation they w ould d emand to allow a purchased bi- cycle to be d eliv ered nine months late. T he med ian response w as 250 florins. A nother group w as asked how long they w ould be w illing to d elay d eliv ery of the bicycle in exchange for 250 florins. T he mean response w as only three w eeks, implying a d iscount rate that is tw elv e times higher. Fred erick and R ead (2002) found that implicit d iscount rates w ere d ramatically higher w hen respond ents generated the future rew ard that w ould equal a specified current rew ard than w hen they generated a current rew ard that w ould equal a specified future re- w ard . Specifically, w hen respond ents w ere asked to state the amount in thirty years that w ould be as good as getting This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 388 Journal of Economic Literature, Vol. XL (June 2002) $100 tod ay, the med ian response w as $10,000 (implying that a future d ollar is 1/100th as v aluable), but w hen asked to specify the amount tod ay that is as good as getting $100 in thirty years, the me- d ian response w as $50 (implying that a future d ollar is 1/2 as v aluable). T w o other experimental proced ures inv olv e rating or pricing temporal pros- pects. In rating tasks, each respond ent ev aluates an outcome occurring at a particular time by rating its attractiv e- ness or av ersiv eness. In pricing tasks, each respond ent specifies a w illingness to pay to obtain (or av oid ) some real or hypothetical outcome occurring at a particular time, such as a monetary re- w ard , d inner coupons, an electric shock, or an extra year ad d ed to the end of one's life. (Once again, see table 1 for a list-of the proced ures and rew ard s used in the d ifferent stud ies.) R ating and pricing tasks d iffer from choice and match- ing tasks in one important respect. Whereas choice and matching tasks call attention to time (because each respon- d ent ev aluates tw o outcomes occurring at tw o d ifferent times), rating and pricing tasks permit time to be manipulated be- tw een subjects (because a single respon- d ent may ev aluate either the immed iate or d elayed outcome, by itself). Loew enstein (1988) found that the timing of an outcome is much less im- portant (d iscount rates are much low er) w hen respond ents ev aluate a single out- come at a particular time than w hen they compare tw o outcomes occurring at d ifferent times, or specify the v alue of d elaying or accelerating an outcome. In one stud y, for example, tw o groups of stud ents w ere asked how much they w ould pay for a $100 gift certificate at the restaurant of their choice. One group w as told that the gift certificate w as v alid immed iately. T he other w as told it could be used beginning six months from now . T here w as no signifi- cant d ifference in the v aluation of the tw o certificates betw een the tw o groups, w hich implies negligible d iscounting. How ev er, w hen asked how much they w ould pay [hav e to be paid ] to use it six months earlier [later], the timing be- came important-the d elay group w as w illing to pay $10 to exped ite receipt of the d elayed certificate, w hile the imme- d iate group d emand ed $23 to d elay the receipt of a certificate they expected to be able to use immed iately.39 A nother important d esign choice in experimental stud ies is w hether to use real or hypothetical rew ard s. T he use of real rew ard s is generally d esirable for obv ious reasons, but hypothetical re- w ard s actually hav e some ad v antages in this d omain. In stud ies inv olv ing hypo- thetical rew ard s, respond ents can be presented w ith a w id e range of rew ard amounts, includ ing losses and large gains, both of w hich are generally infea- sible in stud ies inv olv ing real outcomes. T he d isad v antage of hypothetical choice d ata is the uncertainty about w hether people are motiv ated to, or capable of, accurately pred icting w hat they w ould d o if outcomes w ere real. T o our know led ge, only tw o stud ies hav e compared d iscounting betw een real and hypothetical rew ard s. Kirby and Marakov ic (1995) asked subjects to state the immed iate amount that w ould make them ind ifferent to some fixed d e- layed amount (d elayed rew ard sizes w ere $14.75, $17.25, $21.00, $24.50, $28.50; d elays w ere 3, 7, 13, 17, 23, and 29 d ays). One group of subjects an- sw ered all thirty permutations for real rew ard s, and another group of subjects 39R ating tasks (and probably pricing tasks as w ell) are subject to anchoring effects. Shelley and T homas Omer (1996), Mary Kay Stev enson (1992), and others hav e found that a giv en d elay (e.g., six months) prod uces greater time d iscounting w hen it is consid ered alongsid e shorter d elays (e.g., one month) than w hen it is consid ered alongsid e longer d elays (e.g., three years). This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 389 answ ered all thirty permutations for hypothetical rew ard s. Discount rates w ere low er for hypothetical rew ard s.40 Maribeth C oller and Melonie Williams (1999) asked subjects to choose be- tw een $500 payable in one month and $500 + $x payable in three months, w here $x w as v aried from $1.67 to $90.94 across fifteen d ifferent choices. In one cond ition, all choices w ere hypo- thetical; in fiv e other cond itions, one person w as rand omly chosen to receiv e her preferred outcome for one of her fifteen choices. T he raw d ata suggest again that d iscount rates w ere consid - erably low er in the hypothetical cond i- tion, although they suggest that this conclusion is not supported after con- trolling for censored d ata, d emographic d ifferences, and heterosked asticity (across d emographic d ifferences and across treatmentsS).41 T hus, there is, as of yet, no clear ev id ence that hypotheti- cal rew ard s are d iscounted d ifferently than real rew ard s.42 6.3 C onclusion: What Is T ime Preference? Figure 2 rev eals spectacular d isagree- ment among d ozens of stud ies that all purport to be measuring time prefer- ence. T his lack of agreement likely re- flects the fact that the v arious elicita- tion proced ures used to measure time preference consistently fail to isolate time preference, and instead reflect, to v arying d egrees, a blend of both pure time preference and other theoretically d istinct consid erations, includ ing: (a) intertemporal arbitrage, w hen trad eable rew ard s are used ; (b) concav e utility; (c) uncertainty that the future rew ard or penalty w ill actually obtain; (d ) inflation, w hen nominal monetary amounts are used ; (e) expectations of changing utility; and (f) consid erations of habit formation, anticipatory utility, and v isceral influences. Figure 2 also rev eals a pred ominance of high implicit d iscount rates-d is- count rates w ell abov e market interest rates. T his consistent find ing may also be d ue to the presence of the v arious extra-time-preference consid erations listed abov e, because nearly all of these w ork to bias imputed d iscount rates upw ard - only habit formation and anticipatory utility bias estimates d ow nw ard . If these confound ing factors w ere ad equately controlled , w e suspect that many in- tertemporal choices or jud gments w ould imply much low er-ind eed , possibly ev en zero-rates of time preference. Our d iscussion in this section high- lights the conceptual and semantic am- biguity about w hat the concept of "time preference" ought to includ e-about w hat properly counts as time prefer- ence per se and w hat ought to be called something else (for further d iscussion, 40T he tw o results w ere not strictly comparable, how ev er, because they used a d ifferent proced ure for the real rew ard s than for the hypothetical re- w ard s. A n auction proced ure w as used for the real-rew ard s group only. Subjects w ere told that w hoev er, of three subjects, stated the low est im- med iate amount w ould receiv e the immed iate amount, and the other tw o subjects w ould receiv e the d elayed amount. Optimal behav ior in such a situation inv olv es ov erbid d ing. Since this creates a d ow nw ard bias in d iscount rates for the real- rew ard s group, how ev er, it d oes not explain aw ay the find ing that real d iscount rates w ere higher than hypothetical d iscount rates. 41 It is hard to und erstand w hich control elimi- nates the d ifferences that are apparent in the raw d ata. It w ould seem not to be the d emographic d ifferences per se, because the hypotheticafcond i- tion had a "substantially higher proportion of non- w hite participants" (p. 121) and "non-w hites on av - erage rev eal d iscount rates that are nearly 21 percentage points higher than those rev ealed by w hites" (p. 122). 42 T here has been consid erable recent d ebate outsid e of the context of intertemporal choice about w hether hypothetical choices are repre- sentativ e of d ecisions w ith real consequences. T he general conclusion from this d ebate is that the tw o method s typically yield qualitativ ely similar results (see C amerer and R obin Hogarth 1999 for a re- cent rev iew ), though systematic d ifferences hav e been observ ed in some stud ies (R onald C ummings, Glenn Harrison, and Elisabet R utstrom 1995; Yoram Kroll, Haim Lev y, and R apoport 1988). This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 390 Journal of Economic Literature, Vol. XL (June 2002) future consequence future utility is confers less utility less important - -.7 - - -- - - - - --- i- -- -. uncertainty opportunitv costs iupw -. .r changing tastes increased w xealth . it .. A mount WVeighting of utility of utility Figure 3. see Fred erick 1999). We hav e argued here that many of the reasons for caring w hen something occurs (e.g., uncer- tainty or utility of anticipation) are not time preference, because they pertain to the expected amount of utility conse- quences confer, and not to the w eight giv en to the utility of d ifferent moments (see figure 3, ad apted from Fred erick 1999). How ev er, it is not obv ious w here to d raw the line betw een factors that operate through utilities and factors that make up time preference. Hopefully, economists w ill ev entually achiev e a consensus about w hat is includ ed in, and exclud ed from, the concept of time preference. Until then, d raw ing attention to the ambiguity of the concept w ill hopefully improv e the quality of d iscourse by increasing aw are- ness that, in d iscussions about time preference, d ifferent people may be using the same term to refer to significantly d ifferent und erlying constructs.43 7. Utnpacking T ime Preference A s d etailed in section 2, early tw entieth- century economists' conceptions of inter- temporal choice includ ed d etailed accounts of d isparate und erlying psy- chological motiv es. With the ad v ent of the DU mod el in 1937, how ev er, economists eschew ed consid erations of specific motiv es, proceed ing as if all in- tertemporal behav ior could be explained by the unitary construct of time prefer- ence. In sections 5 and 6, w e highlighted sev eral factors that influence intertem- poral d ecisions, but w hich w ould not be consid ered time preference as the term is ord inarily used . In this section, w e turn our focus inw ard and question w hether ev en time preference itself should be regard ed as a unitary construct. Issues of this type are hotly d ebated in psychology. For example, psycholo- gists d ebate the usefulness of conceptu- alizing intelligence in terms of a single unitary "g" factor. T ypically, a posited psychological construct (or "trait") is consid ered useful only if it satisfies three criteria: (1) it remains relativ ely constant across time w ithin a particular ind iv id ual; (2) it pred icts behav ior across a w id e range of situations, and (3) d ifferent measures of it correlate highly w ith one another. T he concept of intelligence satisfies these criteria fairly w ell.44 First, performance in tests of 43 Not only d o people use the same term to re- fer to d ifferent concepts (or sets of concepts), they also use d ifferent terms to represent the same concept. T he w elter of terms used in d iscussions of intertemporal choice includ e: d iscount factor, d iscount rate, marginal priv ate rate of d iscount, social d iscount rate, utility d iscount rate, marginal social rate of d iscount, pure d iscounting, time preference, subjectiv e rate of time preference, pure time preference, marginal rate of time pref- erence, social rate of time preference, ov erall time preference, impatience, time bias, temporal orien- tation, consumption rate of interest, time positiv ity inclination, and "the pure futurity effect." John Broome (1995, pp. 128-29) notes that some of the controv ersv about d iscounting results from d iffer- ences in how the term is used : "On the face of it . . . tv pical economists and typical philosophers seem to d isagree. But actualty I think there is more misund erstand ing here than d isagreement . . . When economists and philosophers think of d iscounting, they typically think of d iscounting d if- ferent things. Economists typicallv d iscount the sorts of good s that are bought and sold in markets [w hereas] philosophers are tv pically thinking of a more fund amental good , people's w ell-being . . . It is perfectlv consistent to d iscount commod ities and not w ell-being." 44 Debates remain, how ev er, about w hether trad itional measures exclud e important d imen- sions, and w hether a multid imensional account of This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 391 cognitiv e ability at early ages correlates highly w ith performance on such tests at all subsequent ages. Second , cogni- tiv e ability (as measured by such tests) pred icts a w id e range of important life outcomes, such as criminal behav ior and income. T hird , abilities that w e re- gard as expressions of intelligence correlate strongly w ith each other. Ind eed , w hen d iscussing the construction of intelligence tests, Herrnstein and C harles Murray (1994, p. 3) note that, "It turned out to be nearly impossible to d ev ise items that plausibly measured some cognitiv e skill [w hich] w ere not positiv ely corre- lated w ith other items that plausibly measured some cognitiv e skill." T he posited construct of time prefer- ence d oes not fare as w ell by these cri- teria. First, no longitud inal stud ies hav e been cond ucted to permit any conclu- sions about the temporal stability of time preference.45 Second , correlations betw een v arious measures of time pref- erence or betw een measures of time preference and plausible real-w orld expressions of it are mod est, at best. C hapman and Elstein (1995) and C hap- man, R ichard Nelson, and Daniel Hier (1999) found only w eak correlations betw een d iscount rates for money and for health, and C hapman and Elstein (1995) found almost no correlation be- tw een d iscount rates for losses and for gains. Fuchs (1982) found no correlation betw een a prototypical measure of time preference (e.g., "Would you choose $1500 now or $4000 in fiv e years?") and other behav iors that w ould plausibly be affected by time preference (e.g., smok- ing, cred it-card d ebt, seat-belt use, and the frequency of exercise and d ental checkups). Nor d id he find much corre- lation among any of these reported be- hav iors (see also Nyhus 1995).46 C hap- man and Elliot C oups (1999) found that corporate employees w ho chose to re- ceiv e an influenza v accination d id hav e significantly low er d iscount rates (as in- ferred from a matching task w ith mone- tary losses), but found no relation betw een v accination behav ior and hypothetical questions inv olv ing health outcomes. Lalith Munasinghe and Sicherman (2000) found that smokers tend to inv est less in human capital (they hav e flatter w age profiles), and many others hav e found that for stylized intertemporal choices among monetary rew ard s, heroin ad d icts hav e higher d is- count rates (e.g., Leanne A lv os, R . A . Gregson, and Michael R oss 1993; Kirby, Petry, and Bickel 1999; Gregory Mad - d en et al. 1997; T homas Murphy and A lan De Wolfe 1986; Petry, Bickel, and Martha A rnett 1998). A lthough the ev id ence in fav or of a single construct of time preference is hard ly compelling, the low cross- behav ior correlations d o not necessarily intelligence w ould hav e ev en greater explanatory pow er. R obert Sternberg (1985), for example, ar- gues that intelligence is usefully d ecomposed into three d imensions: (1) analytical intelligence, w hich includ es the ability to id entify problems, compute strategies, and monitor solutions, and is measured w ell by existing IQ tests; (2) creativ e intelligence, w hich reflects the ability to generate problem-solv ing options, and (3) practical intelli- gence, w hich inv olv es the ability to implement problem-solv ing options. 45 A lthough there hav e been no longitud inal stud ies of time preference per se, Mischel and his colleagues d id find that a child 's capacity to d elay gratification w as significantly correlated w ith other v ariables assessed d ecad es later, includ ing aca- d emic achiev ement and self esteem (Ozlem A yd uk et al. 2000; Mischel, Yuichi Shod a, and Peake 1988; Shod a, Mischel, and Peake 1990). Of course, this prov id es ev id ence for construct v alid ity only to the extent that one v iew s these other v ariables as expressions of time preference. We also note that w hile there is little ev id ence that intertempo- ral behav iors are stable ov er long period s, there is some ev id ence that time preference is not strictly constant ov er time for all people. Heroin ad d icts d iscount both d rugs and money more steeply w hen they are crav ing heroin than w hen they are not (Louis Giord ano et al. 2001). 46A similar lack of intraind iv id ual consistency has been observ ed in risk-taking (Kennet MacC rimmon and Donald Wehrung 1990). This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 392 Journal of Economic Literature, Vol. XL (June 2002) d isprov e the existence of time prefer- ence. Suppose, for example, that some- one expresses low d iscount rates on a conv entional elicitation task, yet ind i- cates that she rarely exercises. While it is possible that this inconsistency re- flects true heterogeneity in the d egree to w hich she d iscounts d ifferent types of utility, perhaps she rarely exercises because she is so busy at w ork earning money for her future or because she simply cares much more about her fu- ture finances than her future card iov as- cular cond ition. Or, perhaps she d oesn't believ e that exercise improv es health. A s this example suggests, many factors could w ork to erod e cross-behav ior cor- relations, and thus, such low correlations d o not mean that there can be no single unitary time preference, und erlying all intertemporal choices (the intertempo- ral analog to hypothesized construct of "g" in analyses of cognitiv e performance). How ev er, notw ithstand ing this d is- claimer, in our v iew the cumulativ e ev i- d ence raises serious d oubts about w hether there is, in fact, such a construct-a sta- ble factor that operates id entically on, and applies equally to, all sources of utility.47 T o better und erstand the pattern of correlations in implied d iscount rates across d ifferent types of intertemporal behav iors, w e may need to unpack time preference itself into more fund amental motiv es, as illustrated by the segmenta- tion of the d elta component of figure 3. Loew enstein et al. (2001) hav e pro- posed three specific constituent mo- tiv es, w hich they labeled impulsiv ity (the d egree to w hich an ind iv id ual acts in a spontaneous, unplanned fashion), compulsiv ity (the tend ency to make plans and stick w ith them), and inhibi- tion (the ability to inhibit the automatic or "knee-jerk" response to the appetites and emotions that trigger impulsiv e be- hav ior).48 Preliminary ev id ence sug- gests that these subd imensions of time preference can be measured reliably. Moreov er, the d ifferent subd imensions pred ict d ifferent behav iors in a highly sensible w ay. For example, repetitiv e behav iors such as flossing one's teeth, exercising, paying one's bills on time, and arriv ing on time at meetings w ere all pred icted best by. the compulsiv ity subd imension. Viscerally d riv en behav - iors, such as reacting aggressiv ely to someone in a car w ho honks at you at a red light, w ere best pred icted by impul- siv ity (positiv ely) and behav ioral inhibi- tion (negativ ely). Money-related behav - iors such as sav ing money, hav ing unpaid cred it-card balances, or being maxed out on one or more cred it card s w ere best pred icted by conv entional measures of d iscount rates (but impul- siv ity and compulsiv ity w ere also highly significant pred ictors). C learly, further research is need ed to ev aluate w hether time preference is best v iew ed as a unitary construct or a composite of more basic constituent motiv es. Further efforts hopefully w ill be informed by recent d iscov eries of neuroscientists, w ho hav e id entified re- gions of the brain w hose d amage lead s to extreme myopia (A ntonio R . Damasio 1994) and areas that seem to play an important role in suppressing the be- hav ioral expression of urges (Joseph E. 47 Note that one can also ov erestimate the strength of the relationship betw een measured time preference and time-related behav iors or be- tw een d ifferent time-related behav iors if these v ariables are related to characteristics such as in- telligence, social class, or social conformity, that are not ad equately measured and controlled for. 48 R ecent research by R oy Baumeister, T od d Heatherton, and Diane T ice (1994) suggests that such "behav ioral inhibition" requires an expend i- ture of mental effort that, like other forms of effort, d raw s on limited resources-a "pool" of w illpow er (Loew enstein 2000a). T heir research show s that behav ioral inhibition in one d omain (e.g., refraining from eating d esirable food ) re- d uces the ability to exert w illpow er in another d o- main (e.g., completing a taxing mental or physical task). This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 393 LeDoux 1996). If some behav iors are best pred icted by impulsiv ity, some by compulsiv ity, some by behav ioral inhi- bition, and so on, it may be w orth the effort to measure preferences at this lev el and to d ev elop mod els that treat these components separately. Of course, such multid imensional perspectiv es w ill inev itably be more d ifficult to opera- tionalize than formulations like the DU mod el, w hich represent time preference as a unid imensional construct. 8. C onclusions T he DU mod el, w hich continues to be w id ely used by economists, has little empirical support. Ev en its d ev elopers- Samuelson, w ho originally proposed the mod el, and Koopmans, w ho prov id ed the first axiomatic d eriv ation-had con- cerns about its d escriptiv e realism, and it w as nev er empirically v alid ated as the appropriate mod el for intertemporal choice. Ind eed , v irtually ev ery core and ancillary assumption of the DU mod el has been called into question by empiri- cal ev id ence collected in the past tw o d ecad es. T he insights from this empiri- cal research hav e spaw ned new theories of intertemporal choice that rev iv e many of the psychological consid erations d is- cussed by early stud ents of intertempo- ral choice-consid erations that w ere ef- fectiv ely d ismissed w ith the introd uction of the DU mod el. A d d itionally, some of the most recent theories show that in- tertemporal behav iors may be d ramatically influenced by people's lev el of und er- stand ing of how their preferences change-by their "metaknow led ge" about their preferences (see, e.g., O'Donoghue and R abin 1999b; Loew enstein, O'Donoghue, and R abin 2000). While the DU mod el assumes that in- tertemporal preferences can be charac- terized by a single d iscount rate, the large empirical literature d ev oted to measuring d iscount rates has failed to establish any stable estimate. T here is extraord inary v ariation across stud ies, and sometimes ev en w ithin stud ies. T his failure is partly d ue to v ariations in the d egree to w hich the stud ies take ac- count of factors that confound the com- putation of d iscount rates (e.g., uncer- tainty about the d eliv ery of future outcomes or nonlinearity in the utility function). But the spectacular cross- stud y d ifferences in d iscount rates also reflect the d iv ersity of consid erations that are relev ant in intertemporal choices and that legitimately affect d if- ferent types of intertemporal choices d ifferently. T hus, there is no reason to expect that d iscount rates should be consistent across d ifferent choices. T he id ea that intertemporal choices reflect an interplay of d isparate and often competing psychological motiv es w as commonplace in the w ritings of early tw entieth-century economists. We believ e that this approach should be resurrected . R eintrod ucing the multiple- motiv es approach to intertemporal choice w ill help us to better und erstand and better explain the intertemporal choices w e observ e in the real w orld . For instance, it permits more scope for und erstand ing ind iv id ual d ifferences (e.g., w hy one person is a spend thrift w hile his neighbor is a miser, or w hy one person d oes d rugs w hile her brother d oes not), because people may d iffer in the d egree to w hich they ex- perience anticipatory utility or are influenced by v isceral factors. T he multiple-motiv e approach may be ev en more important for und erstand ing intra-ind iv id ual d ifferences. When one looks at the behav ior of a single ind iv id - ual across d ifferent d omains, there is often a w id e range of apparent attitud es tow ard the future. Someone may smoke heav ily, but carefully stud y the returns of v arious retirement packages. A nother This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 394 Journal of Economic Literature, Vol. XL (June 2002) may squirrel money aw ay w hile, at the same time, giv ing little thought to elec- trical efficiency w hen purchasing an air cond itioner. Someone else may d ev ote tw o d ecad es of his life to establishing a career, and then jeopard ize this long- term inv estment for some highly tran- sient pleasure. Since the DU mod el as- sumes a unitary d iscount rate that applies to all acts of consumption, such intra-ind iv id ual heterogeneities pose a theoretical challenge. T he multiple- motiv e approach, by contrast, allow s us to read ily interpret such d ifferences in terms of more narrow , more legitimate, and more stable constructs-e.g., the d egree to w hich people are skeptical of promises, experience anticipatory util- ity, are influenced by v isceral factors, or are able to correctly pred ict their future utility. T he multiple-motiv e approach may sound excessiv ely open-end ed . We hav e d escribed a v ariety of consid erations that researchers could potentially incor- porate into their analyses. Includ ing ev ery consid eration w ould be far too complicated , w hile picking and choos- ing w hich consid erations to incorporate may leav e one open to charges of being ad hoc. How , then, should economists proceed ? We believ e that economists should proceed as they typically d o. Economics has alw ays been both an art and a sci- ence. Economists are forced to intuit, to the best of their abilities, w hich con- sid erations are likely to be important in a particular d omain and w hich are likely to be largely irrelev ant. When econo- mists mod el labor supply, for instance, they typically d o so w ith a utility func- tion that incorporates consumption and leisure, but w hen they mod el inv est- ment d ecisions, they typically assume that preferences are d efined ov er w ealth. Similarly, a researcher inv esti- gating charitable giv ing might use a utility function that incorporates altru- ism but not risk av ersion or time prefer- ence, w hereas someone stud ying inv es- tor behav ior is unlikely to use a utility function that incorporates altruism. For each d omain, economists choose the utility function that is best able to in- corporate the essential consid erations for that d omain, and then ev aluate w hether the inclusion of specific con- sid erations improv es the pred ictiv e or explanatory pow er of a mod el. T he same approach can be applied to multiple-motiv e mod els of intertemporal choice. For d rug ad d iction, for exam- ple, habit formation, v isceral factors, and hyperbolic d iscounting seem likely to play a prominent role. For extend ed experiences, such as health states, ca- reers, and long v acations, the prefer- ence for improv ement is likely to come into play. For brief, v iv id experiences, such as w ed d ings or criminal sanctions, utility from anticipation may be an important d eterminant of behav ior. In sum, w e believ e that economists' und erstand ing of intertemporal choices w ill progress most rapid ly by continuing to import insights from psychology, by relinquishing the assumption that the key to und erstand ing intertemporal choices is find ing the right d iscount rate (or ev en the right d iscount func- tion), and by read opting the v iew that intertemporal choices reflect many d is- tinct consid erations and often inv olv e the interplay of sev eral competing mo- tiv es. Since d ifferent motiv es may be ev oked to d ifferent d egrees by d ifferent situations (and by d ifferent d escriptions of the same situation), d ev eloping d e- scriptiv ely ad equate mod els of in- tertemporal choice w ill not be easy. But w e hope this paper w ill help. R EFER ENC ES A bel, A nd rew . 1990. "A sset Prices Und er Habit Formation and C atching Up w ith the Joneses," A mer. Econ. R ev . 80, pp. 38-42. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 395 A inslie, George. 1975. "Specious R ew ard : A Be- hav ioral T heory of Impulsiv eness and Impulse C ontrol," Psych. Bull. 82:4, pp. 463-96. A inslie, George and Vard a Haend el. 1983. "T he Motiv es of the Will," in Etiologic A spects of A l- cohol and Drug A buse. E. Gottheil, K. Durley, T . Skod ola and H. Waxman, ed s. Springfield , IL: C harles C . T homas, pp. 119-40. A inslie, George and Nick Haslam. 1992. "Hyper- bolic Discounting," in C hoice Ov er T ime. George Loew enstein and Jon Elster, ed s. NY: R ussell Sage, pp. 57-92. A inslie, George and R ichard J. Herrnstein. 1981. "Preference R ev ersal and Delayed R einforcement," A nimal Learning Behav ior, 9:4, pp. 476-82. A kerlof, George A . 1991. "Procrastination and Obed ience," A mer. Econ. R ev . 81:2, pp. 1-19. A lbrecht, Martin and Martin Weber. 1995. "Hy- perbolic Discounting Mod els in Prescriptiv e T heory of Intertemporal C hoice," Zeitschrift Fur Wirtschafts- U Sozialw issenschaften, 115:S, pp. 535-68. . 1996. "T he R esolution of Uncertainty: A n Experimental Stud y," J. Inst. T heoretical Econ. 152:4, pp. 593-607. A lv os, Leanne; R . A . Gregson and Michael W. R oss. 1993. "Future T ime Perspectiv e in C ur- rent and Prev ious Injecting Drug Users," Drug A lcohol Depend . 31, pp. 193-97. A ngeletos, George-Marios; Dav id Laibson, A nd rea R epetto, Jeremy T obacman and Stephen Wein- berg. 2001. "T he Hyperbolic C onsumption Mod el: C alibration, Simulation, and Empirical Ev aluation," J. Econ. Perspect. 15:3, pp. 47-68. A riely, Daniel and Ziv C armon. 2002. "Prefer- ences ov er Sequences of Outcomes," in T ime and Decision: Economic and Psychological Per- spectiv es on Intertemporal C hoice. George Loew enstein, Daniel R ead and R oy Baumeister, ed s. NY: R ussell Sage (in press). A riely, Daniel and Klaus Wertenbroch. 2002. "Procrastination, Dead lines, and Performance: Using Precommitment to R egulate One's Be- hav ior," Psych. Sci. (in press). A rrow , Kenneth J. 1983. "T he T rad e-Off Betw een Grow th and Equity," in Social C hoice and Jus- tice: C ollected Papers of Kenneth J. A rrow . Kenneth J. A rrow , ed . C ambrid ge, MA : Belknap Press, pp. 190-200. A yd uk, Ozem; R od olfo Mend oza-Denton, Walter Mischel, G. Dow ney, Philip K. Peake and Monica R od riguez. 2000. "R egulating the Inter- personal Self: Strategic Self-R egulation for C oping w ith R ejection Sensitiv ity," J. Personal- ity Social Psych. 79:5, pp. 776-92. Bateman, Ian; A listair Munro, Bruce R hod es, C hris Starmer and R obert Sugd en. 1997. "A T est of the T heory of R eference-Depend ent Preferences," Quart. J. Econ. 112:2, pp. 479- 505. Baumeister, R oy F.; T od d F. Heatherton and Di- ane M. T ice. 1994. Losing C ontrol: How and Why People Fail at Self-R egulation. San Diego: A cad emic Press. Becker, Gary A nd Kev in M. Murphy. 1988. "A T heory of R ational A d d iction," J. Polit. Econ. 96:4, pp. 675-701. Beebe-C enter, John G. 1929. "T he Law of A ffec- tiv e Equilibrium," A mer. J. Psych. 41, pp. 54- 69. Benabou, R oland and Jean T irole. 2000. "Self- C onfid ence: Intrapersonal Strategies," Prince- ton U. d iscuss. paper 209. Benartzi, Shlomo and R ichard H. T haler. 1995. "Myopic Loss A v ersion and the Equity Pre- mium Puzzle," Quart. J. Econ. 110:1, pp. 73-92. Benzion, Uri; A mnon R apoport and Joseph Yagil. 1989. "Discount R ates Inferred From Deci- sions: A n Experimental Stud y," Management Sci. 35, pp. 270-84. Bernheim, Douglas and A ntonio R angel. 2001. "A d d iction, C ond itioning, and the Visceral Brain," Stanford U. Bohm-Baw erk, Eugen Von. (1889), 1970. C apital and Interest. South Holland : Libertarian Press. Bold rin, Michele; Law rence C hristiano and Jonas Fisher. 2001. "Habit Persistence, A sset R e- turns, and the Business C ycle," A mer. Econ. R ev . 91, pp. 149-66. Bow man, Dav id ; Deborah Minehart and Matthew R abin. 1999. "Loss A v ersion in a C onsumption- Sav ings Mod el," J. Econ. Behav . Org. 38:2, pp. 155-78. Broome, John. 1995. "Discounting the Future," Philosophy and Public A ffairs 20, pp. 128-56. C airns, John A . 1992. "Discounting and Health Benefits," Health Econ. 1, pp. 76-79. . 1994. "Valuing Future Benefits," Health Econ. 3, pp. 221-29. C airns, John A . and Marjon M. v an d er Pol. 1997. "C onstant and Decreasing T iming A v ersion for Sav ing Liv es," Social Sci. Med . 45:11, pp. 1653- 59. . 1999. "Do People Value T heir Ow n Fu- ture Health Differently T han Others' Future Health?" Med . Decision Making, 19:4, pp. 466- 72. C amerer, C olin F. and R obin M. Hogarth. 1999. "T he Effects of Financial Incentiv es in Experi- ments: A R ev iew and C apital-Labor Prod uction Framew ork," J. R isk Uncertainty 19, pp. 7-42. C ampbell, John and John C ochrane. 1999. "By Force of Habit: A C onsumption-Based Explana- tion of A ggregate Stock Market Behav ior," J. Polit. Econ. 107, pp. 205-51. C aplin, A nd rew and John Leahy. 2001. "Psycho- logical Expected Utility T heory A nd A nticipa- tory Feelings," Quart. J. Econ. 166, pp. 55-79. C arrillo, Juan D. 1999. "Self-C ontroI, Mod erate C onsumption, and C rav ing," C EPR d iscuss. paper 2017. C arrillo, Juan D. and T homas Mariotti. 2000. "Strategic Ignorance as a Self-Disciplining Dev ice," R ev . Econ. Stud . 67:3, pp. 529-44. C arroll, C hristopher. 1997. "Buffer-Stock Sav ing and the Life C ycle/Permanent Income Hy- pothesis," Quart. J. Econ. 112, pp. 1-55. C arroll, C hristopher; Jod y Ov erland and Dav id This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 396 Journal of Economic Literature, Vol. XL (June 2002) Weil. 2000. "Sav ing and Grow th w ith Habit Formation," A mer. Econ. R ev . 90, pp. 341-55. C arroll, C hristopher and A nd rew Samw ick. 1997. "T he Nature of Precautionary Wealth," J. Monet. Econ. 40, pp. 41-71. C hakrav arty, S. 1962. "T he Existence of an Opti- mum Sav ings Program," Econometrica 30:1, pp. 178-87. C hapman, Gretchen B. 2000. "Preferences for Im- prov ing and Declining Sequences of Health Outcomes," J. Behav . Decision Making, 13, pp. 203-18. . 1996. "T emporal Discounting and Utility for Health and Money," J. Exper. Psych: Learn- ing, Memory, C ognition 22:3, pp. 771-91. C hapman, Gretchen B. and Elliot J. C oups. 1996. "T ime Preferences and Prev entiv e Health Be- hav ior: A cceptance of the Influenza Vaccine," Med . Decision Making 19:3, pp. 307-14. C hapman, Gretchen B. and A rthur S. Elstein. 1995. "Valuing the Future: T emporal Discount- ing of Health and Money," Med . Decision Making 15:4, pp. 373-86. C hapman, Gretchen; R ichard Nelson and Daniel B. Hier. 1999. "Familiarity and T ime Prefer- ences: Decision Making about T reatments for Migraine Head aches and C rohn's Disease," J. Exper. Psych.: A pplied 5:1, pp. 17-34. C hapman, Gretchen B. and Jennifer R . Winquist. 1998. "T he Magnitud e Effect: T emporal Dis- count R ates and R estaurant T ips," Psychonomic Bull. R ev . 5:1, pp. 119-23. C hesson, Harrell and W. Kip Viscusi. 2000. "T he Heterogeneity of T ime-R isk T rad eoffs," J. Be- hav . Decision Making 13, pp. 251-58. C oller, Maribeth and Melonie B. Williams. 1999. "Eliciting Ind iv id ual Discount R ates," Exper. Econ. 2, pp. 107-27. C onstantinid es, George M. 1990. "Habit Forma- tion: A R esolution of the Equity Premium Puz- zle," J. Polit. Econ. 98:3, pp. 519-43. C ummings, R onald G.; Glenn W. Harrison and E. Elisabet R utstrom. 1995. "Homegrow n Values and Hypothetical Surv eys: Is the Dichotomous C hoice A pproach Incentiv e-C ompatible?" A mer. Econ. R ev . 85, pp. 260-66 Damasio, A ntonio R . 1994. Descartes' Error: Emo- tion, R eason, and the Human Brain. NY: G. P. Putnam. Dolan, Paul and C laire Gud ex. 1995. "T ime Pref- erence, Duration and Health State Valuations," Health Econ. 4, pp. 289-99. Dreyfus, Mark K. and W. Kip Viscusi. 1995. "R ates Of T ime Preference and C onsumer Valuations of A utomobile Safety and Fuel Effi- ciency,"J. Law Econ. 38:1, pp. 79-105. Duesenberry, James. 1952. Income, Sav ing, and the T heory of C onsumer Behav ior. C ambrid ge, MA : Harv ard U. Press. Elster, Jon, .1979. Ulysses and the Sirens: Stud ies in R ationality and Irrationality. C ambrid ge, UK: C ambrid ge U. Press. I_ . 1985. "Weakness of Will and the Free- R id er Problem," Econ. Philosophy, 1, pp. 231-65. Fischer, C arolyn. 1999. "R ead T his Paper Ev en Later: Procrastination w ith T ime-Inconsistent Preferences," R esources for the Future d iscuss. paper 99-20. Fishburn, Peter C . 1970. Utility T heory and Deci- sion Making. NY: Wiley. Fishburn, Peter C . and A riel R ubinstein. 1982. "T ime Preference," Int. Econ. R ev . 23:2, pp. 677-94. Fisher, Irv ing. 1930. T he T heory of Interest. NY: Macmillan. Frank, R obert. 1993. "Wages, Seniority, and the Demand for R ising C onsumption Profiles," J. Econ. Behav . Org. 21, pp. 251-76. Fred erick, Shane. 1999. "Discounting, T ime Prefer- ence, and Id entity," Ph.D. T hesis, Dept. Social & De- cision Sci., C arnegie Mellon U. . 2002. "T ime Preference and Personal Id entity," in T ime and Decision: Economic and Psychological Perspectiv es on Intertemporal C hoice. George Loew enstein, Daniel R ead and R oy Baumeister, ed s. NY: R ussell Sage (in press). Fred erick, Shane and George Loew enstein. 2002. "T he Psychology of Sequence Preferences," w ork. paper, Sloan School, MIT . Fred erick, Shane and Daniel R ead . 2002. "T he Empirical and Normativ e Status of Hyperbolic Discounting and Other DU A nomalies," w ork. paper, MIT and Lond on School Econ. Fuchs, Victor. 1982. "T ime Preferences and Health: A n Exploratory Stud y," in Economic A s- pects of Health. Victor Fuchs, ed . C hicago: U. C hicago Press, pp. 93-120. Fuhrer, Jeffrey. 2000. "Habit Formation in C on- sumption and Its Implications for Monetary- Policy Mod els," A mer. Econ. R ev . 90, pp. 367- 90. Ganiats, T heod ore G.; R ichard T . C arson, R obert M. Hamm, Scott B. C antor, Walton Sumner, Stephen J. Spann, Michael Hagen and C hristo- pher Miller. 2000. "Health Status and Prefer- ences: Population-Based T ime Preferences for Future Health Outcome," Med ical Decision Making: A n Int. J. 20:3, pp. 263-70. Gately, Dermot. 1980. "Ind iv id ual Discount R ates and the Purchase and Utilization of Energy- Using Durables: C omment," Bell J. Econ. 11, pp. 373-74. Giord ano, Louis A .; Warren Bickel, George Loew enstein, Eric Jacobs, Lisa Marsch and Gary J. Bad ger. 2001. "Opioid Depriv ation A f- fects How Opioid -De end ent Outpatients Dis- count the Value of Delayed Heroin and Money," w ork. paper, U. Vermont Burlington, Psychiatry Dept. Substance A buse T reatment C enter. Gold man, Stev en M. 1980. "C onsistent Plans," R ev . Econ. Stud . 47:3, pp. 533-37. Gourinchas, Pierre-Oliv ier and Jonathan Parker. 2001. "T he Empirical Importance of Precau- tionary Sav ing," A mer. Econ. R ev . 91:2, pp. 406-12. Green, Donald ; Karen Jacow itz, Daniel Kahneman This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 397 and Daniel Mcfad d en. 1998. "R eferend um C on- tingent Valuation, A nchoring, and Willingness to Pay for Public Good s," R esource Energy Econ. 20, pp. 85-116. Green, Leonard ; E. B. Fischer, Jr., Stev en Perlow and Lisa Sherman. 1981. "Preference R ev ersal and Self C ontrol: C hoice as a Function of R e- w ard A mount and Delay," Behav . A nal. Letters 1:1, pp. 43-51. Green, Leonard ; Nathanael Fristoe and Joel Myer- son. 1994. "T emporal Discounting and Prefer- ence R ev ersals in C hoice Betw een Delayed Outcomes," Psychonomic Bull. R ev . 1:3, pp. 383-89. Green, Leonard ; A strid Fry and Joel Myerson. 1994. "Discounting of Delayed R ew ard s: A Life-Span C omparison," Psychological Sci. 5:1, pp. 33-36. Green, Leonard ; Joel Myerson and Ed w ard McFad d en. 1997. "R ate of T emporal Discount- ing Decreases w ith A mount of R ew ard ," Mem- ory & C ognition 25:5, pp. 715-23. Gruber, Jonathan and Botond Koszegi. 2000. "Is A d d iction 'R ational'? T heory and Ev id ence," NBER w ork. paper 7507. Gul, Faruk and Wolfgang Pesend orfer. 2001. "T emptation and Self-C ontrol," Econometrica 69, pp. 1403-35. Harless, Dav id W. and C olin F. C amerer. 1994. "T he Pred ictiv e Utility of Generalized Expected Util- ity T heories," Econometrica 62:6, pp. 1251- 89. Harrison, Glenn W.; Morten I. Lau and Melonie B. Williams. 2002. "Estimating Ind iv id ual Dis- count R ates in Denmark," A mer. Econ. R ev . 92 (in press). Hausman, Jerry. 1979. "Ind iv id ual Discount R ates and the Purchase and Utilization of Energy- Using Durables," Bell J. Econ. 10:1, pp. 33- 54. Hermalin, Benjamin and A lice Isen. 2000. "T he Effect of A ffect on Economic and Strategic De- cision Making," mimeo, U. C . Berkeley and C ornell U. Herrnstein, R ichard . 1981. "Self-C ontrol as R e- sponse Strength," in Quantification of Stead y- State Operant Behav ior. C hristopher M. Brad - shaw , Elmer Szabad i and C . F. Low e, ed s. Elsev ier/North-Holland . Herrnstein, R ichard J.; George F. Loew enstein, Drazen Prelec and William Vaughan. 1993. "Utility Maximization and Melioration: Inter- nalities in Ind iv id ual C hoice," J. Behav . Deci- sion Making 6:3,pp. 149-85. Herrnstein, R ichard J. and C harles Murray. 1994. T he Bell C urv e: Intelligence and C lass Struc- ture in A merican Life. NY: Free Press. Hesketh, Beryl. 2000. "T ime Perspectiv e in C areer-R elated C hoices: A pplications of T ime- Discounting Principles," J. Vocational Behav . 57, pp. 62-84. Hirshleifer, Jack. 1970. Inv estment, Interest, and C apital. Englew ood C liffs, NJ: Prentice-Hall. Holcomb, J. H. and P. S. Nelson. 1992. "A nother Experimental Look at Ind iv id ual T ime Prefer- ence," R ationality Society 4:2, pp. 199-220. Hold en, Stein T .; Bekele Shiferaw and Mette Wik. 1998. "Pov erty, Market Imperfections and T ime Preferences of R elev ance for Env ironmental Policy?" Env iron. Dev el. Econ. 3, pp. 105-30. Houston, Douglas A . 1983. "Implicit Discount R ates and the Purchaes of Untried , Energy- Sav ing Durable Good s," J. C onsumer R es. 10, pp. 236-46. How arth, R ichard B. and A lan H. Sanstad . 1995. "Discount R ates and Energy Efficiency," C on- temp. Econ. Pol. 13:3, pp. 101-109. Hsee, C hristopher K.; R obert P. A belson and Pe- ter Salov ey. 1991. "T he R elativ e Weighting of Position and Velocity in Satisfaction," Psych. Sci. 2:4, pp. 263-66. Jermann, Urban. 1998. "A sset Pricing in Prod uc- tion Economies," J. Monet. Econ. 41, pp. 257- 75. Jev ons, Herbert S. 1905. Essays on Economics. Lond on: Macmillan. Jev ons, William S. 1888. T he T heory of Political Economy. Lond on: Macmillan. Johannesson, Magnus and Per-Olov Johansson. 1997. "Quality of Life and the WT P for an In- creased Life Expectancy at an A d v anced A ge," J. Public Econ. 65, pp. 219-28. Kahneman, Daniel. 1994. "New C hallenges to the R ationality A ssumption," J. Inst. T heoretical Econ. 150, pp. 18-36. Kahneman, Daniel and A mos T v ersky. 1979. "Prospect T heory: A n A nalysis of Decision Un- d er R isk," Econometrica 47, pp. 263-92. Kahneman, Daniel; Peter Wakker and R akesh Sarin. 1997. "Back to Bentham? Explorations of Experienced Utility," Quart. J. Econ. 112, pp. 375-405. Keren, Gid eon and Peter R oelofsma. 1995. "Im- med iacy and C ertainty in Intertemporal C hoice," Org. Behav . Human Decision Proc. 63:3, pp. 287-97. Kirby, Kris N. 1997. "Bid d ing on the Future: Ev i- d ence A gainst Normativ e Discounting of De- layed R ew ard s," J. Experiment. Psych.: General, 126, pp. 54-70. Kirby, Kris N. and R ichard J. Herrnstein. 1995. "Preference R ev ersals d ue to Myopic Discount- ing of Delayed R ew ard ," Psych. Sci. 6:2, pp. 83-89. Kirby, Kris N. and Nino N. Marakov ic. 1995. "Mod eling Myopic Decisions: Ev id ence for Hy- perbolic Delay-Discounting w ith Subjects and A mounts," Org. Behav . Human Decision Proc. 64, pp. 22-30. . 1996. "Delay-Discounting Probabilistic R ew ard s: R ates Decrease as A mounts Increase," Psychonomic Bull. R ev . 3:1, pp. 100-104. Kirby, Kris N.; Nancy M. Petry and Warren Bickel. 1999. "Heroin A d d icts Hav e Higher Discount R ates for Delayed R ew ard s than Non- Drug-Using C ontrols," J. Exper. Psych.: Gen- eral 128:1, pp. 78-87. Koomey, Jonathan G. and A lan H. Sanstad . 1994. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 398 [ournal of Economic Literature, Vol. XL (June 2002) "T echnical Ev id ence for A ssessing the Perfor- mance of Markets A ffecting Energy Efficiency," Energy Pol. 22:10, pp. 826-32. Koopmans, T jalling C . 1960. "Stationary Ord inal Utility and Impatience," Econometrica 28, pp. 287-309. . 1967. "Objectiv es, C onstraints, and Out- comes in Optimal Grow th Mod els," Econo- metrica 35:1, pp. 1-15. Koopmans, T jalling C .; Peter A . Diamond and R ichard E. Williamson. 1964. "Stationary Utility and T ime Perspectiv e," Econometrica 32, pp. 82-100. Koszegi, Botond . 2001. "Who Has A nticipatory Feelings," w ork paper, econ. d ept. U. C al. Berkeley. Kroll, Yoram; Haim Lev y and A mnon R apoport. 1988. "Experimental T ests of the Separation T heorem and the C apital A sset Pricing Mod el," A mer. Econ. R ev . 78, pp. 500-19. Laibson, Dav id . 1994. "Essays in Hyperbolic Dis- counting," Ph.D. d issertation, MIT . . 1997. "Gold en Eggs and Hyperbolic Dis- counting," Quart. J. Econ. 112, pp. 443-77. _ . 1998. "Life-C ycle C onsumption and Hy- perbolic Discount Functions," Europ. Econ. R ev . 42, pp. 861-71. . 2001. "A C ue-T heory of C onsumption," Quarterly J. Econ. 116, pp. 81-119. Laibson, Dav id ; A nd rea R epetto and Jeremy T o- bacman. 1998. "Self-C ontrol and Sav ing for R e- tirement," Brookings Pap. Econ. A ct. 1, pp. 91- 196. Lancaster, K. J. 1963. "A n A xiomatic T heory of C onsumer T ime Preference," Int. Econ. R ev . 4, pp. 221-31. Law rence, Emily. 1991. "Pov erty and the R ate of T ime Preference: Ev id ence from Panel Data," J. Polit. Econ. 119, pp. 54-77. Led oux, Joseph E. 1996. T he Emotional Brain: T he Mysterious Und erpinnings of Emotional Life. NY: Simon & Schuster. Loew enstein, George. 1987. "A nticipation and the Valuation of Delayed C onsumption," Econ. J. 97, pp. 666-84. . 1988. "Frames of Mind in Intertemporal C hoice," Manage. Sci. 34, pp. 200-14. . 1996. "Out of C ontrol: Visceral Influences on Behav ior," Org. Behav . Human Decision Proc. 65, pp. 272-92. . 1999. "A Visceral A ccount of A d d iction," in Getting Hooked : R ationality and A d d iction. Jon Elster and Ole-Jorgen Skog, ed s. C am- brid ge, UK: C ambrid ge U. Press, pp. 235-64. . 2000a. "Willpow er: A Decision-T heorist's Perspectiv e," Law Philos. 19, pp. 51-76. . 2000b. "Emotions In Economic T heory and Economic Behav ior," A mer. Econ. R ev . Pap. Proceed . 90, pp. 426-32. Loew enstein, George and Erik A ngner. 2002. "Pred icting and Honoring C hanging Prefer- ences," in T ime and Decision: Economic and Psychological Perspectiv es on Intertemporal C hoice. George Loew enstein, Daniel R ead and R oy Baumeister, ed s. NY: R ussell Sage (in press). Loew enstein, George; T ed O'Donoghue and Mat- thew R abin. 2000. "Projection Bias in the Pre- d iction of Future Utility," w ork. paper. Loew enstein, George and Drazen Prelec. 1991. "Negativ e T ime Preference," A mer. Econ. R ev . 81, pp. 347-52. 1992. "A nomalies in Intertemporal C hoice: Ev id ence and an Interpretation," Quart. J. Econ. 107:2, pp. 573-97. . 1993. "Preferences for Sequences of Out- comes," Psych. R ev . 100:1, pp. 91-108. Loew enstein, George and Nachum Sicherman. 1991. "Do Workers Prefer Increasing Wage Profiles?" J. Labor Econ. 9:1, pp. 67-84. Loew enstein, George; R oberto Weber, Janine Flory, Stephen Manuck and Matthew Muld oon. 2001. "Dimensions of T ime Discounting," pre- sented at C onference on Surv ey R esearch on Household Expectations and Preferences, A nn A rbor, Nov . 2-3. Maccrimmon, Kenneth R . and Donald A . Weh- rung. 1990. "C haracteristics of R isk-T aking Executiv es," Manage. Sci. 36:4, pp. 422-35. Mackeigan, L. D.; L. N. Larson, J. R . Draugalis, J. L. Bootman and L. R . Burns. 1993. "T ime Preference for Health Gains v s. Health Losses," Pharmacoecon. 3:5, pp. 374-86. Mad d en, Gregory J.; Nancy M. Petry, Gary J. Bad ger and Warren Bickel. 1997. "Impulsiv e and Self-C ontrol C hoices in Opioid -Depend ent Patients and Non-Drug-Using C ontrol Partici- pants: Drug and Monetary R ew ard s," Exper. C linical Psychopharmacology 5:3, pp. 256- 62. Maital, S. and S. Maital. 1978. "T ime Preference, Delay of Gratification, and Intergenerational T ransmission of Economic Inequality: A Behav - ioral T heory of Income Distribution," in Essays in Labor Market A nalysis. Orley A shenfelter and Wallace Oates, ed s. NY: Wiley. Martin, John L. 2001. "T he A uthoritarian Person- ality, 50 Years Later: What Lessons A re T here for Political Psychology?" Polit. Psych. 22:1, pp. 1-26. Mazur, James E. 1987. "A n A d justment Proced ure for Stud ying Delayed R einforcement," in T he Effect of Delay and Interv ening Ev ents on R ein- forcement Value. Michael L. C ommons, James E. Mazur, John A . Nev in and How ard R achlin, ed s. Hillsd ale, NJ: Erlbaum. Meyer, R ichard F. 1976. "Preferences Ov er T ime," in Decisions w ith Multiple Objectiv es. R alph Keeney and How ard R aiffa, ed s. NY: Wiley, pp. 473-89. Millar, A nd rew and Douglas Nav arick. 1984. "Self- C ontrol and C hoice in Humans: Effects of Vid eo Game Playing as a Positiv e R einforcer," Learning and Motiv ation, 15, pp. 203-18. Mischel, Walter; Joan Grusec and John C . Mas- ters. 1969. "Effects of Expected Delay T ime on Subjectiv e Value of R ew ard s and Punishments," J. Personality Soc. Psych. 11:4, pp. 363-73. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 399 Mischel, Walter; Yuichi Shod a and Philip K. Peake. 1988. "T he Nature of A d olescent C om- petencies Pred icted by Preschool Delay of Gratification," J. Personality Soc. Psych. 54:4, pp. 687-96. Moore, Michael J. and W. Kip Viscusi. 1988. "T he Quantity-A d justed Value of Life," Econ. Inq. 26:3, pp. 369-88. 1990a. "Discounting Env ironmental Health R isks: New Ev id ence and Policy Impli- cations," J. Env iron. Econ. Manage. 18, pp. S51-S62. __ . 1990b. "Mod els for Estimating Discount R ates for Long-T erm Health R isks Using Labor Market Data," J. R isk Uncertainty 3, pp. 381- 401. Munasinghe, Lalith and Nachum Sicherman. 2000. "Why Do Dancers Smoke? T ime Prefer- ence, Occupational C hoice, and Wage Grow th," w ork. paper, C olumbia U. and Barnard C ol- lege. Murphy, T homas J. and A lan S. Dew olfe. 1986. "Future T ime Perspectiv e in A lcoholics, Pro- cess and R eactiv e Schizophrenics, and Nor- mals," Int. J. A d d ictions, 20, pp. 1815-22. Myer, R . F. 1976. "Preferences Ov er T ime," in Decisions w ith Multiple Objectiv es. R . Keeney and H. R aiffa, ed s, pp. 473-89. Myerson, Joel and Leonard Green. 1995. "Dis- counting of Delayed R ew ard s: Mod els of Ind i- v id ual C hoice," J. Exper. A nal. Behav . 64, pp. 263-76. Nisan, Mord ecai and A bram Minkow ich. 1973. "T he Effect of Expected T emporal Distance on R isk T aking," J. Personality Soc. Psych. 25:3, pp. 375-80. Nyhus, E. K. 1995. "Item and Non Item-Specific Sources of Variance in Subiectiv e Discount R ates. A C ross Sectional Stud y," 15th C onfer- ence on Subjectiv e Probability, Utility and De- cision Making, Jerusalem. O'Donoghue, T ed and Matthew R abin. 1999a. "A d d iction and Self C ontrol," in A d d iction: En- tries and Exits. Jon Elster, ed . NY: R ussell Sage, pp. 169-206. . 1999b. "Doing It Now or Later," A mer. Econ. R ev . 89:1, pp. 103-24. . 1999c. "Incentiv es for Procrastinators," Quart. J. Econ. 114:3, Pp. 769-816. . 1999d . "Procrastination in Preparing for R etirement," in Behav ioral Dimensions of R e- tirement Economics. Henry A aron, ed . Brook- ings Institution and R ussell Sage, pp. 125-56. . 2000. "A d d iction and Present-Biased Pref- erences," C ornell U. and U. C . Berkeley. . 2001. "C hoice and Procrastination," Quart. J. Econ. 116:1, pp. 121-60. . 2002. "Self A w areness and Self C ontrol," forthcoming in T ime and Decision: Economic and Psychological Perspectiv es on Intertempo- ral C hoice. George Loew enstein, Daniel R ead , and R oy Baumeister, ed s. NY: R ussell Sage, in press. Olson, Mancur and Martin J. Bailey. 1981. "Posi- tiv e T ime Preference," J. Polit. Econ. 89:1, pp. 1-25. Orphanid es, A thanasios and Dav id Zerv os. 1995. "R ational A d d iction w ith Learning and R egret," J. Polit. Econ. 103:4, pp. 739-58. Parfit, Derek. 1971. "Personal Id entity," Philo- sophical R ev . 80:1, pp. 3-27. - 1976. "Lew is, Perry, and What Matters," in T he Id entities of Persons. A melie 0. R orty, ed . Berkeley: U. C alifornia Press. . 1982. "Personal Id entity and R ationality," Synthese-53, pp. 227-41. Peleg, Bezalel and Menahem E. Yaari. 1973. "On the Existence of a C onsistent C ourse of A ction When T astes A re C hanging," R ev . Econ. Stud . 40:3, pp. 391-401. Pend er, John L. 1996. "Discount R ates and C red it Markets: T heory and Ev id ence from R ural In- d ia," J. Dev el. Econ. 50:2, pp. 257-96. Petry, Nancy M.; Warren Bickel and Martha M. A rnett. 1998. "Shortened T ime Horizons and Insensitiv ity to Future C onsequences in Heroin A d d icts," A d d iction 93, pp. 729-38. Phelps, E. S. and R obert Pollak. 1968. "On Second -Best National Sav ing and Game- Equilibrium Grow th," R ev . Econ. Stud . 35, pp. 185-99. Pigou, A rthur C . 1920. T he Economics of Welfare. Lond on: Macmillan. Pollak, R obert A . 1968. "C onsistent Planning," R ev . Econ. Stud . 35, pp. 201-208. . 1970. "Habit Formation and Dynamic De- mand Functions," J. Polit. Econ. 78:4, pp. 745- 63. Prelec, Drazen and George Loew enstein. 1998. "T he R ed and the Black: Mental A ccounting of Sav ings and Debt," Marketing Sci. 17:1, Pp. 4- 28. R abin, Matthew . 2000. "R isk A v ersion and Expected -Utility T heory: A C alibration T heo- rem," Econometrica 68:5, pp. 1281-92. R abin, Matthew and R ichard H. T haler. 2001. "A nomalies: R isk A v ersion," J. Econ. Perspect. 15:1, pp. 219-32. R achlin, How ard ; A nd res R aineri and Dav id C ross. 1991. "Subjectiv e Probability and De- lay," J. Exper. A nal. Behav . 55:2, pp. 233-44. R ae, John. 1834. T he Sociological T heory of C apital (reprint 1834 ed .). Lond on: Macmil- lan. R aineri, A nd res and How ard R achlin. 1993. "T he Effect of T emporal C onstraints on the Value of Money and Other C ommod ities," J. Behav . De- cision Making 6, pp. 77-94. R ead , Daniel. 2001. "Is T ime-Discounting Hyper- bolic or Subad d itiv e?" J. R isk Uncertainty, 23, ps. 5-32. R ead , Daniel; George F. Loew enstein and Mat- thew R abin. 1999. "C hoice Bracketing," J. R isk Uncertainty 19, pp. 171-97. R ed elmeier, Daniel A . and Daniel N. Heller. 1993. "T ime Preference in Med ical Decision Making and C ost-Effectiv eness A nalysis," Med i- cal Decision Making 13:3, pp. 212-17. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions 400 Journal of Economic Literature, Vol. XL (June 2002) R oelofsma, Peter. 1994. "Intertemporal C hoice," Free U. A msterd am. R oss Jr., W. T . and I. Simonson. 1991. "Ev alu- ations of Pairs of Experiences: A Preference for Happy End ings," J. Behav . Decision Making 4, pp. 155-61. R otA , A lv in E. and J. Keith Murnighan. 1982. "T he R ole of Information in Bargaining: A n Ex- perimental Stud y," Econometrica 50:5, pp. 1123-42. R ubinstein, A riel. 2000. "Is It 'Economics and Psy- chology'? T he C ase of Hyperbolic Discounting," T el A v iv U. and Princeton U. R ud erman, H.; M. D. Lev ine and J. E. Mcmahon. 1987. "T he Behav ior of the Market for Energy Efficiency in R esid ential A ppliances Includ ing Heating and C ooling Equipment," Energy J. 8:1, pp. 101-24. R yd er, Harl E. and Geoffrey M. Heal. 1973. "Op- timal Grow th w ith Intertemporally Depen- d ent Preferences," R ev . Econ. Stud . 40, pp. 1- 33. Samuelson, Paul. 1937. "A Note on Measurement of Utility," R ev . Econ. Stud . 4, pp. 155-61. . 1952. "Probability, Utility, and the Ind e- pend ence A xiom," Econometrica 20:4, pp. 670- 78. Schelling, T homas C . 1984. "Self-C ommand in Practice, in Policy, and in a T heory of R ational C hoice," A mer. Econ. R ev . 74:2, pp. 1-11. Senior, N. W. 1836. A n Outline of the Science of Political Economy. Lond on: C low es & Sons. Shea, John. 1995a. "Myopia, Liquid ity C onstraints, and A ggregate C onsumption," J. Money, C red it, Banking, 27:3, pp. 798-805. . 1995b. "Union C ontracts and the Life- C ycle/Permanent-Income Hypothesis," A mer. Econ. R ev . 85:1, pp. 186-200. Shelley, Marjorie K. 1993. "Outcome Signs, Ques- tion Frames and Discount R ates," Manage. Sci. 39, pp. 806-15. . 1994. "Gain/Loss A symmetry in R isky In- tertemporal C hoice," Org. Behav . Human Deci- sion Proc. 59, pp. 124-59. Shelley, Marjorie K. and T homas C . Omer. 1996. "Intertemporal Framing Issues in Management C ompensation," Org. Behav . Human Decision Proc. 66:1, pp. 42-58. Shod a, Yuichi; Walter Mischel and Philip K. Peake. 1990. "Pred icting A d olescent C ognitiv e and Self-R egulatory C ompetencies from Pre- school Delay of Gratification," Dev elop. Psych. 26:6, pp. 978-86. Solnick, Jay; C atherine Kannenberg, Dav id Ecker- man and Marcus Waller. 1980. "A n Experimen- tal A nalysis of Impulsiv ity and Impulse C ontrol in Humans," Learning and Motiv ation 11, pp. 61-77. Solow , R obert M. 1974. "Intergenerational Equity and Exhaustible R esources," R ev . Econ. Stud . 41:Symposium Econ. Exhaustible R esources, pp. 29-45. Spence, Michael and R ichard Zeckhauser. 1972. "T he Effect of T iming of C onsumption Deci- sions and R esolution of Lotteries on C hoice of Lotteries," Econometrica 40:2, pp. 401- 403. Starmer, C hris. 2000. "Dev elopments in Non- Expected Utility T heory: T he Hunt for a Descriptiv e T heory of C hoice Und er R isk," J. Econ. Lit. 38:2, pp. 332-82. Sternberg, R obert J. 1985. Beyond IQ: A T riarchic T heory of Human Intelligence. NY: C ambrid ge U. Press. Stev enson, Mary Kay. 1992. "T he Impact of T em- poral C ontext and R isk on the Jud ged Value of Future Outcomes," Org. Behav . Human Deci- sion Proc. 52:3, pp. 455-91. Strotz, R . H. 1955-56. "Myopia and Inconsistency in Dynamic Utility Maximization," R ev . Econ. Stud . 23:3, pp. 165-80. Suranov ic, Stev en; R obert Gold farb and T homas C . Leonard . 1999. "A n Economic T heory of C igarette A d d iction," J. Health Econ. 18:1, pp. 1-29. T haler, R ichard H. 1981. "Some Empirical Ev i- d ence on Dynamic Inconsistency," Econ. Let- ters 8, pp. 201-07. . 1985. "Mental A ccounting and C onsumer C hoice," Manage. Sci. 4, pp. 199-214. 1999. "Mental A ccounting Matters," J. Be- hav . Decision Making 12, pp 183-206. T haler, R ichard H. and Hers M. Shefrin. 1981. "A n Economic T heory of Self-C ontrol," J. Polit. Econ. 89:2, pp. 392-410. T v ersky, A mos and Daniel Kahneman. 1983. "Ex- tensional v s. Intuitiv e R easoning": T he C onjunc- tion Fallacy in Probability Jud gment," Psych. R ev . 90, pp. 293-315. . 1991. "Loss A v ersion in R iskless C hoice: A R eference Depend ent Mod el," Quart. J. Econ. 106, pp. 1039-61. T v ersky, A mos and Derek J. Koehler. 1994. "Sup- port T heory: Nonextensional R epresentation of Subjectiv e Probability," Psych. R ev . 101:4, pp. 547-67. v an d er Pol, Marjon M. and John A . C airns. 1999. "Ind iv id ual T ime Preferences for Ow n Health: A pplication of a Dichotomous C hoice Question w ith Follow Up," A ppl. Econ. Letters 6:10, pp. 649-54. . 2001. "Estimating T ime Preferences for Health Using Discrete C hoice Experiments," Social Sci. Med . 52, pp. 1459-70. Varey, C . A . and D. Kahneman. 1992. "Experi- ences Extend ed A cross T ime: Ev aluation of Moments and Episod es," J. Behav . Decision Making 5:3, pp. 169-85. Viscusi, W. Kip and Michael T Moore. 1989. "R ates of T ime Preference anc d Valuation of the Duration of Life," J. Public Econ. 38:3, pp. 297- 317. Wahlund , R ichard and Jonas Gunnarsson. 1996. "Mental Discounting and Financial Strategies," J. Econ. Psych. 17:6, pp. 709-30. Wang, R uqu. 1997. "T he Optimal C onsumption and Quitting of Harmful A d d ictiv e Good s," w ork. paper, Queens U. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions Fred erick, Loew enstein, and O'Donoghue: T ime Discounting 401 Warner, John T . and Saul Pleeter. 2001. "T he Per- sonal Discount R ate: Ev id ence from Military Dow nsizing Programs," A mer. Econ. R ev . 91:1, pp. 33-53. Whiting, Jennifer. 1986. "Friend s and Future Selv es," Philosophical R ev . 95:4, pp. 547-580. Winston, Gord on C . 1980. "A d d iction and Back- slid ing: A T heory of C ompulsiv e C onsumption," J. Econ. Behav . Org. 1, pp. 295-324. Yates, J. Frank and R oyce A . Watts. 1975. "Prefer- ences for Deferred Losses," Org. Behav . Hu- man Perform. 13:2, pp. 294-306. This content downloaded from 128.84.126.57 on Thu, 10 Apr 2014 02:22:25 AM All use subject to JSTOR Terms and Conditions