Characterizing Non Classical Models of Intertemporal Choice by Present Values

Umut Keskin October 23, 2012

Abstract We provide characterization of decreasing impatience and variation aversion in terms of present values of individuals. Earlier work on these models provided their axiomatization in terms of preference conditions. In a nite and discrete time setting, we show that the specic functional forms representing these behaviors can be justied also by the behavior of an intuitive and familiar tool, the present value. The reason for the choice of these particular models is that they do not t into the standard discounted utility model, and are usually called anomalies or irrationalities. We argue against such labeling and present our results with the hope that this present value setting makes it easier to attribute more rationality to these models.

Introduction

At a time when political correctness was apparently not a requirement for academic papers, Carver (1918) called behavior that is not rational, eternally feminine. Needless to say, we do not sympathize with his attitude of attributing any behavior that is not cerebral, to a particular sex without any supporting evidence. Yet it still remains the case that individual behavior, that is not in line with a benchmark economic decision model has always received a somewhat negative reaction from the supporters of that particular model. The case of Carver (1918) shows that it has been so
1

even as early as about a hundred years ago. In this paper we take two such behaviors in the context of intertemporal decision making, variation aversion (Gilboa 1989) and decreasing impatience (Prelec 2004), and give their axiomatization in terms of peoples perception of future income changes from a current time periods point of view, i.e. the present value. In this paper we dene present value as follows: Assume that the agent is endowed with a stream of payments spanned over a nite and discrete amount of time. Given this endowment, suppose we change the amount in one future period only. Then we ask how much change in current period outcome would make the agent indierent to the distorted scheme. We call the answer the present value of the amount by which the future endowment changes. Previous work on these models axiomatized them in terms of preference conditions. We chose characterizations in terms of present value instead for several reasons. Firstly, many experiments testing validity of intertemporal choice models are carried out inevitably from a todays point of view. For example subjects are asked to submit their evaluations of some future monetary changes. The responses they give are direct revelations of how they perceive future outcomes as of today, which is basically how we dene present value. Therefore as an empirical tool, present value is a natural way to draw conclusions from these experiments since ecision makers are already familiar with the concept. Also we believe that the behavioral conditions we state on present value are more compact and more easily grasped most of the time hence simplies the analysis and the testing of the theories to a great extent. Moreover, from a theoretical point of view, characterizing the representation of outcomes spanned over multiple periods of time just by looking at the particular interaction between two periods -today and a future timeis an improvement.

A lengthy list of economic situations that involve intertemporal decision making would be redundant. It suces to say that a careful observer should notice its extensive use in modern economic theory which points out to its undeniable importance for economics. Most problems in intertemporal choice come down to the question of how streams of outcomes spread over the future are (or should be) evaluated. In this study we take these streams to be monetary payments but our methods can be generalized to valuation of streams of nonmonetary objects by eliciting the willingness to pay for these objects (See SOME REFERENCE HERE!!). This basic question of how to make evaluations in dynamic contexts dates back to early economic theory -such as Adam Smith (1776)- but the answer came quite later by Samuelson (1937). He suggested the standard discounted utility (SDU) model; one which is the most prominent in economics when intertemporal decisions are involved. According to this model, a sequence of outcomes {f0 , . . . , fn } received in periods {0, . . . , n} is evaluated by its discounted utility,
n i=0

i u(fi ). In

this formulation, u is the instantaneous utility function and is the discount factor which is usually taken to be less than one so that future outcomes are weighted less and less as time goes by. Koopmans (1960) axiomatized SDU where he laid out the necessary and sucient conditions that an individuals preference relation must satisfy to be representable by this particular function. To this day, most of economic theory has taken these conditions to be a set of normative standards for rationality in intertemporal decision making. However Koopmans (1960) did not make any case for rationality of the SDU in his paper, indeed he called discounting short sidedness for it assigns less value to future payments. He also said it is a form of lack of imagination. Pigou (1920) used the term faulty telescopic faculty to describe preference for present. Although not refereing

specically to discounting, his criticism can easily be channelled to SDU especially with high discount factors. Also, from a mathematical point of view, violations of SDU are quite similar to violations of expected utility representation in the presence of uncertainty. Yet the violations of expected utility were criticized as being irrational much more frequently and severely than their intertemporal counterparts. We believe that naming the violation of the same mathematical principle in one scenario irrational, and in another not-soirrational is falling into a framing trap which is usually considered to be another form of irrationality by many researchers(WHICH RESEARCHERS? SOME REFERENCE HERE). When Carver (1918) was criticizing eternally feminine behavior as opposed to self maximizing rational one, he made his case on the premise that self maximizing individuals would progress their nations. He did not know that 32 years later a mathematician, A. W. Tucker, would present the famous prisoners dilemma which proved that self maximizing individual behavior does not necessarily aggregate to a socially optimum outcome. As these points suggest, classifying behavior as rational and irrational is quite problematic. Therefore we follow an alternative approach and adopt Richters (1966) notion of rationalizability. In a nutshell, this approach only asks whether there could be any preference relation that results in a particular observed behavior; leaving aside whether this preference is rational in some sense or not. Once presented in the form of perception of future (i.e. in terms of present value behavior), the models mentioned above seem to be less vulnerable to criticism on grounds of rationality compared to preference condition axiomatizations. And this is another important reason why we chose present value characterizations instead of preference conditions. The idea of present value characterization instead of preference

conditions has been used before in Bleichrodt et al (2012). In that paper we characterized decision models that are more commonly used in economics and nance, those that are accepted to be closer to being called rational. This study is the eternally feminine counterpart of our previous work.

The Model

Let S = {0, . . . , n} be the set of time periods with 0 being the current time. At each time point i an individual receives a monetary outcome fi . That makes our object of study the streams of payments {f0 , . . . , fn } over which the agent has a continuous, complete and transitive preference relation . We will sometimes call them prospects or acts. These acts will also be viewed as functions from S to R and we will use both interpretations interchangeably. We will denote the set of all such functions, hence all such payment streams, by F. More formally, F = {f |f : S R}. We will assume that each {i} S is strictly essential i.e., for all i S, and any f F, there exist , R such that i f part of with f i x where is the asymmetric . Moreover, we assume solvability; so, for any f, g F g, and for all i S, there exists R such that i f g. g whenever fi gi for all i S

is called strictly monotone if f

and fj > gj for at least one j S. For any f F, and R, i f will denote the vector obtained by replacing the ith component of f by . Also i j f replaces the ith and jth components of f by and respectively. Our concept of present value is subjective in the sense that it is dened through the preferences of individuals. Denition 2.1. Suppose that an individual with preference relation on F is initially endowed with an income stream f F. Further, suppose that f is changed by units only in period i. Then is
5

called the individuals present value of if it satises the following condition: (f0 + , f1 , . . . fi , . . . , fn ) (f0 , . . . , fi + , . . . , fn ) If we assume solvability, existence of is guaranteed. Then if is strictly monotone, is unique. In its most general form, the PV will depend on f, i, and , and such dependence will be denoted by i (f, ) . In the case where any of these variables is irrelevant for , it will be omitted from the notation. By default. the PV of any amount received in period 0 is itself i.e., 0 (f, ) =

Variation Aversion

Consider the following example taken from Gilboa (1989): Preferences are dened on the set of income streams spread over four periods. In each period, the payment can be high (H) or low (L). Assume that the agent dislikes variation in her periodical payments which might be due to adjustment costs or psychological reasons.1 Then it is quite intuitive to assume the preferences below for an agent with such taste: (H, H, L, L) (L, L, H, H) (H, L, H, L) (L, H, L, H) (1)

The reason is that there is less variation involved in (H, H, L, L) and (L, L, H, H) than in (H, L, H, L) and (L, H, L, H) between periods. A careful analysis of 1 shows that the standard discounted utility model cannot accommodate these preferences. Indeed a closer investigation yields that no additively separable representation could be used in this case. In other words, even a representation of the general form
1 See

3 t=0

ut (xt ), where xt {H, T } cannot explain (1). In this

section we will analyze a model developed by Gilboa (1989) that can

Gilboa (1989) for a more detailed discussion on the justication of such preferences.

explain it. Since separability is violated, Gilboa studied a restricted domain of preferences; those that can be represented by Choquet integral since it allows for non-separability. Inspired by similar concerns in the context of choice under uncertainty, Schmeidler laid out the preference axiomatization of this integral representation that had been developed by Choquet. We will rst present Choquet integral, then discuss the relevance of Schmeidlers axiomatization in Gilboas work. Choquet integral is constructed as follows. Let B be an arbitrary set and B an algebra on B. A nonadditive measure on (B, B) is a set function that satises the properties below: i. () = 0, (B) = 1 ii. If A A B, then (A) (A ). Then given a B measurable bounded function , its Choquet integral is dened as: d =
B R

(({b|(b) > t})1)dt+

R+

(({b|(b) > t}))dt (2)

where the integrals on the right are Riemann integrals. As the intuitive reasoning for the above example suggests, any attempt to model preferences in 1 should take into account the variation in utility terms between periods, |u(xi ) u(xi1 )|. Assuming only those preferences representable by Choquet integral, Gilboa narrowed them down to the following form via his variation preserving sure thing principle.
n

(i u(xi ) + i |u(xi ) u(xi1 )|)

i=0

(3)

So here, discounted utility is adjusted by a weighted sum of utility variations in each period. We take his model as benchmark and give a characterization of in terms of ones perception of present value instead of his preference conditions over the set of all income streams.
7

Since Gilboas model is built on the premises of Schmeidlers axiomatization and since that in turn stemmed from problems in choice under uncertainty, we next give a brief refresher of this theory in relation to the problem at hand. 3.1 Relation to Decision Under Uncertainty

In the context of uncertainty, an agent chooses among a set of uncertain prospects which are simply functions, f , from the state space S, to an outcome set X. The interpretation is that f yields outcome f (s) X if state s S happens. Which of these states will actually occur is unknown to the decision maker. Suppose now that S is nite and X is the set of real numbers, hence S = {s0 , . . . , sn } and X = R. Whether f is an uncertain prospect that pays f (si ) if si occurs, or it is a deterministic scheme of payments f (si ) at time si is just a matter of interpretation and choosing between uncertain prospects and streams of payments are essentially the same thing from a mathematical point of view. Therefore the tools and techniques of decision theory under uncertainty can easily be borrowed to study intertemporal choice problems of the format displayed here. Gilboa did so, and so shall we. Savage (1954) proposed a formal model for the study of choice under uncertainty, i.e. expected utility, which is derived from preference axioms. For a long time expected utility was, and in fact in most of mainstream economics still is, the dominant model used to analyze behavior under uncertainty, and it hinges on the idea of separability between states. This is guaranteed by Savages (1954) sure thing principle (STP). Formally it is dened as follows. Take any A S and suppose that f, g, f , g are acts that satisfy the

conditions below: f (s) = g(s) f (s) = f (s) Then according to STP, f f (s) = g (s) g(s) = g (s) g if and only if f s A, s Ac . g . The rationale

behind the STP is the presumption that as long as f and g coincide on a subset A, the preference between f and g remains the same no matter what these common values are. In other words, replacing the common parts of f and g has no eect on the preference if this replacement still assigns the same values to f and g on A. Note that, with the intertemporal decision making interpretation of Savages model, the STP is violated in 1. Accordingly, Gilboa gave a weaker version of STP, that is variation preserving sure thing principle, which leads to 3. It is similar to STP in that if f and g assume common values on a subset A S, preference between them does not change when both are replaced with a dierent common set of values on A as long as -and this is where it diers from STPthis replacement does not alter the variation in f or g in dierent degrees. P will denote the power set of S. Subsets of S of the form {i, i+1, . . . , j} for i j are called intervals and are denoted by [i, j]. Denition 3.1. Let A = [i, j] S be an interval and f, f , g, g F be such that f (s) = g(s), f (s) = f (s), f (k) = f (k) = g(k) = g (k) Then f f (s) = g (s) g(s) = g (s) s A s Ac

for k = i 1, j + 1.

is said to satisfy variation preserving sure thing principle if g.

g if and only if f

Though usually unobjectionable at rst sight, a counterexample in Ellsberg (1961) showed that STP is not as innocent as it may
9

seem. Ellsberg (1961) led to the development of several dierent decision models2 as attempts to capture violations of Savage axioms. Choquet expected utility (CEU) has been a prominent one among such models since its axiomatization by Schmeidler (1989). In this seminal paper Schmeidler used the Anscombe and Aumann (1963) (A-A) setting to analyze choice under uncertainty, employing nonadditive measures developed by Choquet (1955), and this is also the framework Gilboa used with S being interpreted as the set of time points. Leaving aside the details of A-A framework, it suces to mention that they study functions that associate each state with a so called lottery. These are factitious probability distributions over the set of outcomes with objectively known probabilities. Whereas in Savage (1954), this intermediate device is nonexistent and the focus is on functions from the state space to outcomes. Which one of these two settings is more intuitive in the case of uncertainty is a point of discussion outside the scope of this paper. What seems less controvertible is that when adapted to an intertemporal choice problem with no uncertainty, the lottery devices in the A-A setting loses its intuitive appeal. Also, Gilboa (1989) chose A-A framework instead of Savages ...since it[A-A] allows for a nite domain.... This is not a technical necessity anymore, for our eld has progressed to allow for nite domains in Savage settings (Gul (1992), Kbberling o and Wakker (2003)). In Schmeidler, the key preference condition that leads to is the comonotonic independence axiom. Two functions f, g F are called comonotonic if there exist no s, t S such that f (s) > g(s) and f (t) < g(t). Comonotonic Independence holds if for all f, g, h F that are pairwise comonotonic and for all (0, 1), f
2 See

g if and only if f + (1 )h

g + (1 )h.

In addition to Schmeidlers technical assumptions, comonotonic inWakker (2010) for a survey of non expected utility models.

10

dependence gives us 3.1. We nd this condition hard to interpret in our intertemporal setting. Because of these reasons, we will use a dierent setting than Gilboa (1989). Among various characterizations of Choquet representations, Kbberling and Wakker (2003) o better ts our interpretation of the problem. And we explain their framework adapted to our problem next. C F is called a comoncone if it consists of all acts that are comonotonic, ie. C = {f F|f(0) . . . f(n) } for some permutation : S S. For , , , R; we write if there exist a comoncone C, a time point3 j S and functions f, g F such that j f j g and j f j g

where all four income streams above belong to C. Roughly, means that receiving instead of has the same preference eect as receiving instead of . Kbberling and Wakkers (2003) o comonotonic tradeo consistency requires that changing any one of , , or breaks this relation. More precisely, if , then there exist no i S and no f , g F such that i f i g and i f i g if = . The same condition must hold for , and too. Keeping all the structural assumptions, we have the following theorem which is a reformulation of Corollary 10 from Kbberling o and Wakker. Theorem 3.1. (Kbberling and Wakker) o satises comonotonic

tradeo consistency if and only if there exist a continuous utility function u : R R that is unique up to an ane transformation
3 In

the authors original work, this j has to be nonnull in C. Since we assume strict

monotonicity, every j S is nonnull.

11

and a unique (posssibly) nonadditive measure on P f g i u f d u g d

where the integrals are Choquet integrals as introduced in 3.1. We tailor this theorem in the next corollary so that it is more in line with the present value setting we aim to establish in this study. Corollary 3.1. The Choquet representation in Theorem 2.1. holds if and only if for , , , R; if there exist j S and f, g F such that 0j f + (j (f, ))0 0 0j g + (j (g, ))0 0 0j f + (j (f, ))0 0 0j g + (j (g, ))0 0 where j f, j g, j f, j g belong to the same comoncone, then there do not exist i S and f , g F such that 0i f + (i (f , ))0 0 0i g + (i (g , ))0 0 0i f + (i (f , ))0 0 0i g + (i (g , ))0 0 for any = for which i f , i g , i f , i g belong to the same comoncone, and the same holds for , and . Proof. A careful investigation shows that the condition in the corollary is just a restatement of Kbberling and Wakkers comonotonic o tradeo consistency in terms of present value meaasurements, hence it is equivalent to Choquet representation. Like in Gilboa, our main goal will be to narrow these Choquet representations to ones that can explain 1. To this end, the present value condition in Corollary 2.1 is no more context relevant than Schmeidlers comonotonic independence that was used in Gilboa. Yet we still provide it with the sole purpose of maintaining a uniform layout in our characterizations, one that is built by using present values.
12

3.2

Main Result

For notational convenience, we will use i and i instead of (i) and (i) below. Assuming that is represented by Choquet integra, Gilboa gave the following characterization of 3: Theorem 3.2. satises variation preserving sure thing principle takes the following form f F
n

if and only if there exist unique , : S R such that the choquet integral representing u f d =
i=0

(i u(xi ) + i |u(xi ) u(xi1 )|)

where |i | |i | + |i+1 | i < n |n | |n | = 0 Our main result shows that 3 can be characterized by present values. As mentioned earlier we will assume that the preference relation can be represented by Choquet integral. Proposition 3.1. The following are equivalent: (i) The representing Choquet integral for form:
n

is of the following

u f d =
i=0

(i u(xi ) + i |u(xi ) u(xi1 )|)

for all f F. (ii) Present value depends on , i, f0 , f1 , fi1 , fi and fi+1 : = i (, f0 , f1 , fi1 , fi , fi+1 ) Proof. Recall that present value, is dened through the equality (f0 + )0 fi f = f0 (fi + )i f.
13

(4)

Suppose that the representing function for i < n, we have

is as in (i). Then for

0 u(f0 + ) + 0 |u(f0 + )| + 1 |u(f1 ) u(f0 + )| + i u(fi )+ i |u(fi ) u(fi1 )| + i+1 |u(fi+1 ) u(fi )| = 0 u(f0 ) + 0 |u(f0 )| + 1 |u(f1 ) u(f0 )| + i u(fi + )+ i |u(fi + ) u(fi1 )| + i+1 |u(fi+1 ) u(fi + )|, (5) showing that (ii) holds is a matter of solving (4) for for dierent cases. We believe that this messy arithmetic is not crucial for our purposes. Therefore we will present the result for only one illustrative case and leave the rest as an exercise for curious readers. One particular such case is: > 0, u(f0 ) 0, fi fi1 , fi+1 fi + , and (6)

Di (u(fi + ) u(fi )) u(f1 ) u(f0 ) (7) D0 where Di = i + i i+1 , and D0 = 0 + 0 0 . Note that under the assumptions made in (6), D0 = 0, for otherwise we would have Di = 0 too and this would lead (4) to hold true for innitely many values of ; which violates monotonicity. Hence, Di /D0 is well dened. Then with the assumptions made in (6) and (7), we have Di [u(fi + ) u(fi )]) f0 . (8) D0 In (8), depends on f0 , fi and explicitly and on i through Di = u1 (u(f0 ) + in the formula. Dependence on f1 , fi+1 and fi1 is due to the conditions assumed in (6) and (7). The case for i = n is investigated similarly, only noting that we dened fn+1 = 0. Conversely assume that = i (, f0 , f1 , fi1 , fi , fi+1 ). We will show that satises variation preserving sure thing principle which
14

will imply that it can be represented by the function in our proposition. Let A = [i, j] S be an interval and f, f , g, g F be such that f (s) = g(s), f (s) = f (s), f (s) = g (s) g(s) = g (s) s A s Ac

f (k) = f (k) = g(k) = g (k)

for k = i 1, j + 1. g. First we will consider

Assume that j + 1 < n and suppose f

indierence, f g. We discount fk s to period 0 one by one. For this purpose, we dene the recursive sequence of pk s as follows: p0 = f 0 p1 = p0 + n (fn , p0 , f1 , fn1 , 0, 0) p2 = p1 + n1 (fn1 , p1 , f1 , fn2 , 0, 0) . . . pk = pk1 + n+1k (fn+1k , pk1 , f1 , fnk , 0, 0). . . . for 0 k n (j + 2). Also, let p1 = pn(j+2) + i2 (fi2 , pn(j+2) , f1 , fi3 , 0, fi1 ) p2 = p1 + i3 (fi3 , p1 , f1 , fi4 , 0, 0) p3 = p2 + i4 (fi4 , p2 , f1 , fi5 , 0, 0) . . . pi3 = pi4 + 2 (f2 , pi4 , f1 , f1 , 0, 0) pi2 = pi3 + 1 (f1 , pi3 , 0, pi3 , 0, 0). We apply the same procedure to g. We call the resulting values p and p ; analogous to p and p dened above. For notational simplicity,

15

let pi2 = and pi2 = . Using the denition of present value, f (p1 , f1 , . . . , fn1 , 0) (p2 , f1 , . . . , fn2 , 0, 0). Iteratively discounting each fk , we obtain f (, 0, . . . , 0, fi1 , . . . , fj+2 , 0 . . . , 0) Similarly, g ( , 0, . . . , 0, gi1 , . . . , gj+2 , 0 . . . , 0). (10) (9)

Rewriting and gj+2 as = +( ) and gj+2 = fj+2 +(gj+2 fj+2 ) respectively and using (9) and (10), we get ( + ( ), 0, . . . , 0, fi1 , . . . , fj+2 , 0 . . . , 0) ( , 0, . . . , 0, gi1 , . . . , gj+1 , fj+2 + (gj+2 fj+2 ), 0 . . . , 0). Then using the fact that gs = fs for s = i 1, . . . , j + 1 results in ( + ( ), 0, . . . , 0, fi1 , . . . , fj+2 , 0 . . . , 0) ( , 0, . . . , 0, fi1 , . . . , fj+1 , fj+2 + (gj+2 fj+2 ), 0 . . . , 0). By the denition of present value, this means j+2 (gj+2 fj+2 , , 0, fj+1 , fj+2 , 0) = . Since j+2 (.) is independent of fi , . . . , fj , we can replace these by fi , . . . , f j : ( + ( ), 0, . . . , 0, fi1 , fi , . . . , fj , fj+1 , fj+2 , 0 . . . , 0) ( , 0, . . . , 0, fi1 , fi , . . . , fj , fj+1 , fj+2 + (gj+2 fj+2 ), 0 . . . , 0). Noting again that gs = fs for s = i 1, . . . , j + 1, we have ( + ( ), 0, . . . , 0, fi1 , fi , . . . , fj , fj+1 , fj+2 , 0 . . . , 0) ( , 0, . . . , 0, gi1 , gi , . . . , gj , gj+1 , fj+2 + (gj+2 fj+2 ), 0 . . . , 0).
16

Recall that we constructed and by discounting each fs and gs to period 0. Now we do the reverse operations and forward components of them back to their original positions. This gives us ( + ( ), 0, . . . , 0, fi1 , fi , . . . , fj , fj+1 , fj+2 , 0 . . . , 0) f and ( , 0, . . . , 0, gi1 , gi , . . . , gj , gj+1 , fj+2 + (gj+2 fj+2 ), 0 . . . , 0) g Hence f g as we wanted to show. If f g, then by our structural assumptions we can nd K > 0, g. such that f (g0 + K, g1 , . . . , gn ). Afterwards, we repeat the same argument above to show that f If j +1 = n, then we start rolling fk s and gk s back to present time from period i 1 to obtain and . Once and are dened as such, the rest of the proof follows the same steps as above, and once again we obtain f g.

In the SDU model, it is typical that the agents PV depend on the amount given in the future, his current income, his income at the period that change happens and also how far the change is. So for an agent who is variation averse in addition to these factors, next period income and the income right before and after the future distortion aect the perception of present value. And this is all that is needed to characterize such an agent.

Decreasing Impatience

In addition to its additively separable form, one of the most distinctive features of SDU is that the degree of impatience for a given
17

length of period stays the same no matter how far this period is from today. More formally, if i 0 i+k 0 for some i S, then i 0 i +k 0 for all i S as long as i + k n. For example, if you are indierent between $10 today and $11 in one week; then according to SDU, you should be in dierent between $10 in eight weeks and $11 in nine weeks. However, this property of SDU has also been challenged on empirical grounds (Thaler (1981), Loewenstein and Prelec (1992)). These studies show that stationarity property of SDU is violated. People prefer sooner but worse outcomes to better but later outcomes when the comparisons are made for near future but this preference reverses as the comparisons are delayed and people accept to wait more to obtain the better outcome. One way to explain such preferences has been to assert that people tend to decrease their rate of discount for farther away future outcomes. A particular functional form that incorporated this idea is the hyperbolic discounting utility function which is also one of the most commonly used models as alternatives to SDU. In this model, any stream f F is evaluated by u(f0 ) +
iS\{0}

i u(fi ). It has been

rst used by Phelps and Pollak (1968) and further developed by Laibson (1997) and employed in many economic applications such as procrastination (ODonoghue and Rabin (1999)), addiction (Gruber and Koszegi (2001)). It is a special case for a more general functional form called general hyperbolic discounting where the discount function is (1 + t)/ ; see Loewenstein and Prelec (1992). And in turn, general hyperbolic discounting is a special case of a broader class of preferences axiomatized by Prelec (2004), the decreasing impatience (DI) preferences. Although commonly used in many studies nowadays, hyperbolic discounting is not the only model that can accommodate the preferences mentioned above (See Rubinstein 2003 for an alternative explanation and a critique of hyperbolic dis-

18

counting). Yet we pursue axiomatizing Prelecs DI mainly because it (indeed mostly its particular case, hyperbolic discounting) has been used more frequently in the literature than other models explaining the particular violations of SDU mentioned above. First we generalize Prelecs (2004) framework to our n period setting. Many studies on hyperbolic discounting and DI deal with simple prospects (i.e. studying preferences between two dated outcomes x and y received in periods t and s, denoted (x, t) and (y, s) respectively) but we study sequences of outcomes instead of isolating two periods. This way we allow for interdependence between periods. Loewenstein and Prelec (1991) describe the need for sequences of outcomes nicely: ...events that take up time cannot be rescheduled without changing the timing of other activities. Then we provide an alternative characterization of his model once again by the notion of present value. Prelec (2004) assumes that preferences are representable by the form
n i=0

(i)u(fi ), where (i) is the discount function and u(fi )

is the utility in period i. Fishburn and Rubinstein (1982) and Harvey (1986) provides preference foundations of this mutiplicatively separable form. (I NEED TO PUT THESE CONDITIONS INTO PRESENT VALUE FORMAT!!! but for the moment just assume it like prelec). We will also assume that limi0 (i) = 0. His main result is that ln((i)) is convex (in other words as i increases, between period discounting becomes smaller) if and only if the agent is decreasingly impatient. DI is dened as follows (generalized to our sequences of outcomes setting) Denition 4.1. DI if (fi+k + )i+k f exhibits DI if for any k > 0, > and f F, fj+k + )j+k (strict fj+k + )j+k ).

(fi + )i f fj + )j f implies (fi+k + )i+k f

We have the following result from Prelec (2004) adapted to our

19

n period setting: Theorem 4.1. (Prelec) The following are equivalent: 1. satises DI.

2. ln((i)) is convex. In accordance with our unied PV characterization system, we provide an alternative axiomatization convex discounting as follows Theorem 4.2. (Prelec) The following are equivalent: 1. depends on i, , f0 and fi and it satises the following for all < : If i (, f0 , fi ) = j ( , f0 , fj ), then i+k (, f0 , fi+k ) > j ( , f0 , fj+k ) 2. ln((i)) is convex. Proof. A careful investigation yields that our present value condition is equivalent to DI, hence the result follows. Therefore for those who are decreasingly impatient, the present value depends on the amount given in the future, their current income, their income at the period that change happens and also how far the change is. And the necessary and sucient behavior of present value for DI agents is as follows: While todays worth of later but better outcome is relatively small compared to the worth of sooner but worse outcome, this changes as time elapses and he leans towards the better outcome since its worth today improves relative to the sooner but worse outcome.

Conclusion

We chose two non standard models in intertemporal decision making namely the decreasing impatience and variation aversion and
20

gave their characterization in terms of subjective present values. We believe that they constitute more intuitive and simpler forms of axioms. We also argued that once presented in terms of present values, these models do not that irrational, a concept which is already quite problematic for even rational models. REFERENCES A problem with BibTex. Once I x it, I will send it.

21