Digitized by the Internet Archive in 2011 with funding from Boston Library Consortium Member Libraries http://www.archive.org/details/commitmentvsflexOOamad -, 31 il5 \0 Technology Department of Economics Working Paper Series Massachusetts Institute of COMMITMENT vs. FLEXIBILITY Manuel Amador Ivan Werning George-Marios Angeletos Working Paper 03-40 November] 5, 2003 Room E52-251 50 Memorial Drive Cambridge, MA 02142 This paper can be downloaded without charge from the Social Science Research Network Paper Collection http://ssrn.com/abstract=471 822 at i-^ k =, V V* U.-m MASSACHUSETTS |^''T1TUTE oftechnol. . >' Commitment vs. Flexibility' Manuel Amador"' Ivan Werning-^ George-Marios Angeletos^ November 15. 2003 Abstract This paper studies the optimal trade-off between commitment and ity in flexibil- an intertemporal consumption/savings choice model. Indi\iduals expect to receive relevant information regarding their erating a value for flexibility - own but also expect to generating a value for commitment. situation tastes - gen- from temptations suffer The model combines of preferences for flexibihty introduced by and - the representations Kreps (1979) with its recent antithe- commitment proposed by Gul and Pesendorfer (2002), which nests the h\T>erbolic discounting model. We set up and solve a mechanism design probsis for lem that optimizes over the individual each period. We set of tion takes a simple threshold form Our analysis or the is consumption/saving options available to the characterize the conditions under which the solu- where minimum sa\dngs also relevant for other issues policies are optimal. such as situations with externalities problem faced by a "paternahstic" planner, which may be important thinking about some regulations such as forced minimum for schooling laws. Introduction If people suffer from temptation and self-control problems, what should be done to help them? Most analysis lead to a simple and extreme conclusion: take over the individual's choices completely. 'We'd like to Diamond and optimal to For example, in models with h\-per- thank comments and suggestions from Daron Acemoglu, Andy Atkcson, Peter especially Pablo Werning. 'Stanford GSB tmx, NBER and UTDT §MIT and it is NBER bolic discounting preferences it is desirable to impose a particular savings plan on individuals. Indeed, one commonly articulated justification for government involvement in re- tirement income in modern economies is population would save "inadequately" 1977). the belief that an important fraction of the if left to their own devices Diamond, (e.g. From the workers perspective most pension systems, pay-as-you-go and talized systems alike, effectively of this paper is to see if such capi- impose a minimum saving requirement. One purpose minimum saving policies are optimal in a model where agents suffer from the temptation to "over-consume" In a series of recent papers Gul and Pesendorfer (2001, 2002a,b) have given prefer- ences that value commitment an axiomatic foundation and derived a useful representation theorem. In their representation the individual suffers from temptations and may exert costly self-control. This formalizes the notion that as a way commitment is useful to avoid temptations that either adversely affect choices or require exert- ing costly self-control. On an axiomatic foundation the opposite side of the spectrum, Kreps (1979) provided for preferences for flexibility. His representation theorem shows that they can be represented by including taste shocks into an expected utility framework. Our model combines Kreps' with Gul and Pesendorfer's representations. Our main application modifies the intertemporal taste-shock preference specification introduced by Atkeson and Lucas (1995) to incorporate temptation. In their model the individual has preferences over random consumption streams. Each period an is i.i.d. taste shock realized that affects the individual's desire for current consumption. Importantly, the taste shock at time-t is assumed to be private information. We modify these preferences by assuming that agents suffer from the temptation for higher present consumption. This feature generates a desire The for commitment. informational asymmetry introduces a trade-off between flexibility. Commitment valued because it is valued because it commitment and reduces temptation while flexibility allows the use of the valuable private information. We solve for the optimal incentive compatible allocation that trades-off commitment and One can interpret our solution as describing the optimal In addition to commitment is flexibility. device. Gul and Pesendorfer's framework, models with time-inconsistent preferences, as in Strotz (1956), also generate a value for committment. In particular, the hyperbolic discounting model has proven useful for studying the effects of a temptation to 'over-consume' as well as the desirability of committment devices (Phelps and Pollack, 1968, and Laibson, 1994). As Krusell, Kuruscu and Smith (2002) have pointed out, however, the temptation framework provided by Gul and Pesendorfer effectively generalizes the hyperbolic discounting when model: it results in the limiting case the agent cannot exert any self-control, giving in fully to his temptations. For expositional purposes show that the We we first treat the hyperbolic discounting case in detail results extend to and then Gul and Pesendorfer's framework. begin by considering a simple hyperbolic discounting case with two possible taste shocks. By we solving this case, on the strength of the temptation critically dispersion of the taste shocks. how the optimal illustrate for current allocation consumption depends relative to the For the resulting second-best problem there are two important cases to consider. For low levels of temptation, relative to the dispersion of the taste shocks, optimal to separate the high and low taste shock agents. If too low, then in order to separate them the principal must offer that yield somewhat the temptation too onerous. to the agent's temptation for higher current consumption. Thus, single When The not consumption bundles both bundles provide more present consumption than their counterparts best allocation. is it is temptation is in the first strong enough, separating the agents becomes principal then finds it optimal to bunch both agents: she offers a consumption bundle equal to her optimal uncontingent allocation. This solution resolves the average over-consumption issue at the expense of foregoing flexibility. In this way, the optimal amount of flexibility depends negatively on the strength of the temptation relative to the dispersion of the taste shocks. two shocks are simple and intuitive. Unfortunately, results are not easily generalized. examples where 'money burning' consuming We is These results with with more than two shocks, these show that with three shocks there are robust optimal: it is optimal to have one of the agents in the interior of his budget set. Moreover, any pair of agents. The examples present a wealth of bunching can occur between possibilities with no obvious discernible pattern. Fortunately, strong results are obtained in the case with a continuum of taste shocks. Our main result is a condition on the distribution of taste shocks that is necessary and sufficient for the optimal mechanism to be a simple threshold rule: a minimum savings level imposed, with is The optimal minimum savings level full flexibility allowed above this depends positively on the strength of temptation. Thus, the main insight from the two type case carries over here: the strength of temptation and this We is important because it is flexibility falls with accomplished by increased bunching. is extend the model to include heterogeneity This tion. minimum. in temptation of current consump- reasonable to assume that agents suffer from temptation at varying degrees. Indeed, perhaps some agents do not suffer from temptation at all. dividuals that Allowing for heterogeneity in temptation would imply that those we observe saving higher temptation. However, of a minimum The we saving policy rest of the paper is we show is more less are that the likely to main in- be the ones suffering from result regarding the optimality robust to the introduction of this heterogeneity. organized as follows. In the remainder of the introduction briefly discuss the related literature. Section 1 lays out the basic intertemporal model using the hyperbolic discounting model. Section 2 analyzes this model with two and three taste shocks while Section 3 works with a continuum of shocks. Section 4 extends the analysis to arbitrary finite time horizons results to the case and Section 5 extends the where agents are heterogenous with respect to Section 6 contains the more general case with temptation and their temptation. self-control proposed by Gul and Pesendorfer (2001,2002a,b). Section 7 studies the case where agents discount exponentially at a different rate than a 'social planner' and preferences are logarithmic. Section 8 diverges to discuss some alternative interpretations and applications of our main results regarding the optimal trade-ofl[ between committment and The final Section concludes. An appendix collects some flexibility. proofs. Related Literature At least since in the Ramsey's (1928) moral appeal economists have long been interested implications of, and justifications for, socially discounting the future at lower rates than individuals. Recently, Caplin and Leahy (2001) discuss a motivation welfare criterion that discounts the future at a lower rate than individuals. (2002) provides another motivation and studies implications of opportunity of a zero social discount rate. In and agents discount the future exponentially. for for a Phelan long-run inequality both these papers the social planner Some papers on social security policies possible "undersaving" by individuals. agents may undersave due have attempted to take into account the Diamond (1977) discussed the case where OLG Feldstein (1985) models to mistakes. agents that discount the future at a higher rate than the social planner and studies the optimal pay-as-you-go system. Laibson (1998) discusses public policies that avoid undersaving in hyperbohc discounting models. Imrohoroglu, Imrohoroglu and Joines (2000) use a model with hyperbolic discounting preferences to perform a quantitative exercise on the welfare (2002) use a effects of pay-as-you-go social security systems. Diamond and Koszegi model with hyperbolic discounting agents to study the policy endogenous retirement choices. effects of O'Donahue and Rabin (2003) advocate studying pa- ternalism normatively by modelling the errors or biases agents may have and applying standard public finance analysis. Finally, several papers discuss trade-offs similar to those emphasized here ious contexts not related to the intertemporal consumption/saving in var- problem that is our focus. Since Weitzman's (1974) provocative paper there has been great interest in the efficiency of the price system compared to a command economy, see (1984) and the references therein. In a recent paper, Athey, Atkeson and Holmstrom Kehoe (2003) study a problem of optimal monetary policy that also features a trade-off between time-consistency and discretion. Sheshinski (2002) models heterogenous agents that make choices over a discrete set of alternatives but are subject to shows that in such a setting reducing the set of alternatives random may be errors and optimal. Laib- son (1994, Chapter 3) considers a moral-hazard model with a hyperbolic-discounting agent and shows that the planner may reward the agent for high output by tilting consumption towards the present. 1 The Basic Model For reasons of exposition we inconsistent. first study a consumer whose preferences are time- Following Strotz (1956), Phelps and Pollack (1968), Laibson (1994) and many others we model the agent for subgame show that all perfect equilibria of the our results go through in each period as different selves and solve game played between when we selves. In section 6 we use the more general framework pro- vided by Gul and Pesendorfer (2001,2002a,b) which, in addition, does not require an game intrapersonal Consider period t = first interpretation. the case with two periods of consumption, from which we evahiate expected arbitrary finite horizons. Ed = so that make \ which Each period agents more current consumption We the agent at time t} the available observable variables. c and The We denote utility for selj-1 where U {) < The N 1. and W {) from periods W () i = 1, ^ U {), utihty for self-0 from periods t = (c) 1 — /3 observed privately by is first after conditioning on and second period consumption 2 with taste shock 9 is {k) and continuously differentiable" and this generality facilitates the extension to 1, + Agents have quasi-geometric discounting: < taste shock consumption: higher 9 4. 9U /? initial taste shock 9, normalized i.i.d. and saving data are increasing, concave notation allows periods in section The and an k, respectively. 9U (c) + [3W /3 1,2, think of the taste shock as a catch-all for the significant variation one actually observes in consumption by an utility of current The valuable. = utihty. Section 4 extends the analysis to receive marginal affects the t 2 W {k) self-t 1 and is a measure of this disagreement or bias. in this respect, there is is . discounts the entire future at rate among disagreement On the different t-selves and the other hand, there is agreement regarding taste shocks: everyone values the effect of 9 in the same way. Below we often associate the value of B to the strength of a 'temptation' for current consumption; thus, we say that temptation An is stronger if 13 is lower. alternative interpretation to 'hyperbolic' discounting only periods 1 and 2. One can simply work with is Although available if we consider the assumption that the correct wel- fare criterion does not discount future utility at the both do so exponentially. is same rate as agents do, although this alternative interpretation is available for 'With exponential CARA utility income shocks are equivalent to taste shocks. ^Note that a taste shock for period i = 2 is not included in this expression. However, only apparent since k cannot depend on 60 and EOo = 1. its absence two-periods we will see that in general sion of the analysis to more periods. In it does not permit a straightforward exten- we section 7 discuss a case in which it does generalize. We investigate the optimal allocation from the point of view of self-0 subject to the constraint that 6 tailoring is private information of self-1. consumption to the taste shock and the The self-1 essential tension 's To solve the allocation preferred by self-0 with total income y we now set up the y. Periods V2 {y) eU (c {9)) + = max / F (9) [OU > 9U m' (k {9)) + c{9) where flexibility self-O. optimal direct truth telling mechanism given Two between constant higher desire for commitment and current consumption. This generates a trade-off between from the point of view of is k (c {9)) (c {9')) {9) < + W {k {6))] + PW {k {9')) y for for all 9,9' eQ (1) 6 G all dF [9) the distribution of the taste shocks with support O. is This problem maximizes, given total resources y, the expected utility from the point of view of self-0 (henceforth: the principal) subject to the constraint that 9 is private information of self-1 (henceforth: the agent). The incentive compatibility it is in agent-^'s self interest to report truthfully, thus obtaining the allocation that is intended for him. constraint (1) ensures that interest rate is normalized to zero for simplicity. The problem above imposes across ^-agent's is ruled out. agent's types. This choice First, may be it many relevance in possible or if In the budget constraints the a budget constraint for each ^ G 0, so that insurance The planner cannot transfer resources across different was motivated by several considerations. possible to argue that the case without insurance situations. This could be the case if pooling risk is is of direct simply not insurance contracts are not available because of other considerations outside the scope of our model. Second, the cardinality of the taste shocks plays a more important role ysis with insurance. The taste shock 9 definitely in an anal- affects ordinal preferences between current and future consumption, c and However, we would k. like to avoid taking a strong stand on whether or not agents with high taste for current consumption also have a higher marginal from utility total resources as the expression 9u +w implictly assumes. Focusing on the case without insurance avoids making our analysis depend strongly on such cardinality assumptions. Third, without temptation as Mirrlees (1971) or (/? = 1) incentive constrained insurance problems such Atkeson and Lucas (1995) are non-trival and the resulting op- timal allocations are not easily characterized. the solutions with temptation (/? < more 1) the optimal allocation without temptation This would make a comparison with In contrast, without insurance difficult. [j3 = 1) is straightforward ~ every agent chooses their tangency point on the budget set - allowing a clearer disentangling of the effects of introducing temptation. Finally, we hope that studying the case without insurance may yield insights into the case with insurance which we are currently pursuing. Once the problem above is solved the optimal allocation for selj-0 solves a standard problem: max {9oU [cq) + (iv2 {yo - co)} CO where yo, Cq and respectively. In ^o represents the initial t — what follows we ignore the 0, income, consumption and taste shock, initial consumption problem and focus on non-trivial periods. 2 Two Types In this section we study the optimal commitment with only two occuring with probabilities p and Without temptation, /3 = 1, 1 — there taste shocks, d^ > 6i, p, respectively. is no disagreement between the planner and the agent and we can implement the ex- ante first-best allocation defined by the solution to eU' {cfb {9)) /W {kfb {9)) temptation, so that compatible. /3 is Intuitively, = close if 1 and Cfh (9) enough to 1, + kfb {9) = y. the first-best allocation the disagreement in preferences dispersion of taste shocks then, at the the high shock agent's allocation. first For low enough levels of best, the low is is still incentive small relative to the shock agent would not envy Proposition 1 There is exists a (3* < such that for P 6 1 the first-best allocation [/5*, 1] implementable. Proof. At Define /3* = /5 < 1 1 the incentive constraints are slack at the ex-ante first-best allocation. to be the value of equality at the This result first (3 which the incentive constraint of for best allocation. The on the discrete relies when we study a continuum centive compatible. result follows. diflFerence in taste of shocks in Section For higher levels of temptation, lower i.e. -I- k* (6) = P > 6i/9h separation (b) if P < 9i/9h hunching is optimal, (c) if P = 9i/9h separating First, p* P* > P first best allocation The next proposition is he attained with the — y for 9 (a) if Proof. the is not in- for &gent-9h characterizes Ccises. Proposition 2 The optimwn can always ing with equality: c* {9) 3. /3, to obtain a higher level of current consumption. optimal allocations in such shocks and no longer holds would take the bundle meant agent-^; If oflfered, holds with di optimal, i.e. i.e. We 9h, 9i. c* {9h) c{9i) = > budget constraint hold- have that 9i/9h c* {9i) and k {9h) = c{9h) and k{9i) < < k P* and: {9i) k{9h) and hunching are optimal follows since u{c*{eH))-u{c*{ei)) = Wiy-c*{9i))-W{y~c*{9H)) U'{c{e,)){c{9,)-c{9^)) 'W'{y-c {9h)) where c* is the first (c {9n) - c {9i)) U'{c{9,)) 'W'{y-c {9^)) in c{0k) _ 9h~- best allocation. Now, consider the case where P > P and suppose that an increase 9i and a decrease unchanged increases c{9h) centive compatible because + it c {9h) in k {9h) that holds {9i/P) + k (9^) U {c{9h)) < y. Then U {k (9^)) -\- k {9h) and the objective function. Such a change is in- strictly relaxes the incentive compatibility constraint of the high type pretending to be a low type and leaves the other incentive compatibility constraint unchanged. It follows that we must have c {9h) +k {9h) = y at an optimum. This also shows that separating arguments establish parts (b) optimal is and proving part in this case, (a). Analogous (c). Proposition 2 shows that for /?</?* the resulting non-trivial second-best problem can be separated into essentially two cases. For intermediate 61/ dh < must offer P, levels of temptation, i.e. optimal to separate the agents. In order to separate them the principal it is consumption bundles that yield somewhat to the agent's ex-post desire higher consumption giving them higher consumption in the first period than the for first best. For higher levels of temptation, onerous, U — {) bunching them W {) i.e. /? < 9[/9h, separating the agents is too then optimal at the best uncontingent allocation - with is = this implies c(-) k{-) = bunching resolves the disagreement y/2. amount problem at the expense of flexibility. In this way, the optimal depends negatively on the size of the disagreement relative to the dispersion of the taste shocks as measured by 9i/9h- Proposition 2 also shows that c{9) + k{6) = y. In this sense, not required for optimality. of flexibihty As it is always optimal to consume 'money burning', discuss below, with i.e. all setting c(0/j) + the resources k{9h) more than two types this < is y, is not a foregone conclusion. Figure 1 below shows a typical case that = W () c^-"/ {1-a), U figure shows consumption we (•) , and a = 2, 9^ illustrate these results. = 1.2, 9i in the first period, c (0) also plots the optimal ex-post consumption , for = .8, p = set 111 and y as a function of both types We (i.e. /3. = U (c) = \. The For comparison the full flexibility outcome). Note that these are always higher than the optimal allocation: the principal does manage to lower consumption in the first 10 period. """---^, ex-ante c^ i--^___ 0.54 --..^ 1 "" Q. E ^"^--^^^ c o o ex-anle '~"--~-_._ C|_ 0.48 \ 0.46 0.6 0.8 1 p values Figure The first Optimal 1: first period consumption figure illustrates Proposition 1 best allocation is with two shocks as a function of 2 in the following way. For high consumption in the first period rises as /3. the /3 attainable so the optimal allocation does not vary with this range. For intermediate way the and (c) /? in /? falls. In this principal yields to the agent's desire for higher consumption. For low enough (3 P bunching becomes optimal and = c{6) To summarize, with two types we y/2. are able to characterize the optimal allocation which enjoys nice properties. In particular, the budget constraint holds with equality and we found simple necessary and sufficient conditions for a bunching or separating outcome to be optimal. Unfortunately, with more than two types extending these conclusions forward. For example, with three taste shocks, 9^ > 0^ > Oi-, it is simple to construct robust examples where the optimal solution has the following properties: get constraint for agent is optimal; and h, and 9h (ii) it is is 9m although (3 is satisfied < 9m/9h with remains a optimal; (iii) characterizing the to bunching can occur between show a variety optimum with more than two find conditions optimal allocation i.e. /3 9i > (i) the bud- 'money burning' bunching m 9m/9h where bunching 9m and 9m, with 9h is separated. of possibilities that illustrate the difficulties in Fortunately, with a continuum of types we - sufficient condition for no longer necessary: there are cases with The examples seem section strict inequality not straight- is types. more progress can be made. In the next on the distribution of 9 which allows us to characterize the fully. 11 Continuous Distribution of Types 3 Assume that the distribution of types = [9,6]. We find it convenient = U {c) and w = VV (k) and we u K {w) A' be the inverse functions of To to change variables from {c,k) to {u,'w) where term either pair an U {c) and W{k), and respectively, so and characterize the incentive compatibility constraint (1) in this case consider the problem faced by agent-^ when confronted with a V{9) = m&xl^u(9') + 0'ee the mechanism If C (u) that C {) Let 'allocation'. and convex. are increasing (•) represented by a density / (9) over the interval is is direct mechanism {it. {9) , w [6)): tu(9') [P truth telling then V (9) = |u ~u[9)de + [9) + w (6) and integrating the envelope condition we obtain, ^u {9) + w (see Milgrom and [9) Segal, 2002). be a non-decreasing function of = j^ The 9. Thus, condition is vo (y) (2), particular, it We now u to and the monotonicity of u are know that well (u, u') also requires these two conditions are thus, = max f C (u {9)) + K {w [6)) < convex since the objective function is It is (2) (2) Fudenberg and Tirole, 1989). planner's problem subject to + w{9) Incentive compatibility of necessary for incentive compatibility. also sufficient (e.g. ^u{9) follows that ih (y) is [9u {9) + w {9)] f y and u is linear concave in {9') > u {9) d9, (9) for 9' > and the constraint 9. This problem set is convex. In y. substitute the incentive compatibility constraint (2) into the objective function and the resource constraint, and integrate the objective function by parts. This allows us to simplify the problem by dropping the function iu{9), except its value at n 9 : 9. for Consequently, the maximization below requires finding a function ^ R and a scalar w representing w {9). 12 Continuous Distribution of Types v^ (y) = max i |u(0 + ^ + i A(l - F [9)) - {y-C {u {9))) + ^u {9) -^u{9)-w- K-' u For any feasible allocation u bunch some upper < {9') > u (9) for 9' 9 and u feasible, (9) = tail u{9) it is of agents. is [9)] u ^u{9)d9 > /3) f {9) cW j^ >9 always feasible to modify the allocation so as to That feasible for we now show that To gain some it is the allocation is, any 9. + ^w +w < (59 = u tail is (9) for always share the ordinal preferences That is, the indifference curves are equivalent. Informally, these agents can > In contrast, agents with 9 their preferences. {9) always optimal. intuition, note that agents with 9 and 9/ [5u u given by u Thus bunching the upper of the planner with a higher taste shock equal to 9/13. 9u - Bunching 3.1 9 9 (1 (59 make a case for display a blatant over-desire for current consumption from the principal's point of view, in the sense that there taste shock that would justify these preferences to the planner. Thus, that these agents are bunched since separating them is it is is no intuitive tantamount to increasing some of these agents consumption, yet they are already obviously "over-consuming" The next result shows that bunching goes even further than ^9. Proposition 3 Define E > 9 9 Note that u* {9) = 9p < j39 u* {9p) for 9 and > such that for 9 9p as the lowest value in 9p 9p < (i.e. l39 it all 13 9p: -1 < - P as long as f bunchs > > 0. An optimal allocation u* has agents above 9p) > Proof. The contribution to the objective function from 9 Substituting u = Jg 9p is 1 f {{i-F{e))-e{i-p)f{e})u{9)d9. du + u {9p) and integrating by parts we obtain, I\{i-F{9))-9{1-P)f{9))d9 ^i{Op)- Note that, ((l J' for all 9 -F [§)) > 9p so {1-^3) f 9 optimal to set du it is With two types Proposition 9h/9i < 1//3. = (1 - F{9))9 I or equivalently 0, 2 showed that bunching is - ^ u 3.2 To <0, 9 (9) = strictly the support is essentially 9p unbounded then 9p may = u {9p) full flexibility or bunch Assumption all , for 6* > 9p. optimal whenever when 9h/9i < 9i. not exist. This occurs, for example, with the Pareto distribution. One can show that in this case allow E 9\9>9 Proposition 3 generalizes this result since with two types 1/P then according to the definition If = (0")) d9 it is either optimal to agents depending on the Pareto parameter. A < solve for the optimal allocation for 9 the density / and /3. A. The density f {9) is differentiable and 9p we require the following condition on satisfies J'iO)^ f{e)- 14 2-/3 i-P <9p. all 9 for Assumption A places a negative lower uous and decreasing in as — /3 1 > the lower The /?. off to for 9 < For any density / such that 9f'/f for /3 close enough to For example, and it is Moreover, 1. bound highest lower bound goes on the whole support 0, only bound on the — oo. Note of that elasticity of —2 A is / that attained for is /3 contin- = does not impose the and bound 9p. is A is satisfied assumption A for all bounded from below assumption many densities satisfy /3. trivially satisfied for all density functions that are non-decreasing also holds for the exponential distribution, the log-normal, Pareto and Gamma distributions for a large subset of their parameters. Minimum 3.3 Define u* {9) , w* Saving Policies {9) to be the unconstrained optimum {u* (9) , w* (9)) = r arg 9 max —u + w <^ s.t. Our next result shows that under assumption unconstrained optimum and agents with 9 optimum the left That for 9p. of some point is, for agent-^: C (u) + K [w] < A > agents with 9 9p are bunched < y 9p are offered their at the unconstrained the optimal mechanism offers the whole budget line to (c*,fc*), given by the ex-post unconstrained optimum of the ^p- agent. Define the Lagrangian function L{w,u\A) = ^ui9)+w + + 1 ^j as: [ip-l)9f{9) + {l-Fi9))]u{9)d9 (^K-'{y-C{u{9))) + ^u{9)-(^^ui9}+w^-J^ ^u{9)d9^ dA where the function A patibility constraint. is We the Lagrange multiplier associated with the incentive comrequire the Lagrange multiplier Luenberger, 1969, Chapter Intuitively, the {9) A to be non-decreasing (see 8). Lagrange multiplier A can be thought of as a cumulative distribu- 15 tion function'^. If of constraints A happens to be differentiable with density A then the can be incorporated into the Lagrangian as the famihar integral of the hand side of each constraint product of the left though common approach in many this is a of discontinuity As we and points: 6 (u* (9) , would not be In such cases, working with a density A A indeed discontinuous at two is 9p. = lu* (9)) Proposition 4 The Proof. A may have points applications, in general, shall see, in our case the multiplier Consider the allocation and and the density function X{6). Al- and these mass points are associated with individual constraints that are particularly important. valid. continuum Without {u, {u* {9p) , w) given by (u w* {9p)) for 9 allocation {u, w) is loss of generality set {9) > ((9) w (6)) = {u* {6) , w* [6)) for 9 < 9p 9p. optimal A , if = and only 1. We will if assumption construct A A holds. to be left continuous. Integrating the Lagrangian by parts: L{w,u\A) = {^u{9) + w)A /3 + ^j i{f3-l)ef{9)-F{9) + A{e))u{9)d9 + Note that the Lagrangian is a sum (a-^ J' (y -C on Gateaux ^Except + ^u (^)) dA {9) of integrals of concave functions of to This implies that the Gateaux differential exists and Lemma {u {9))) differentiability in the Appendix) . [9). computed (see the In particular, at the proposed is easily for the integrability condition. Also, for notational purposes, we make A a function, instead of the usual right-continuous convention for distribution functions. 16 and u left-continuous allocation for dL lu, u the Gateaux {w, u; h^, = /7,„|A) + + where x\g g\ is differential r|/i„ (0 y -^ ((/3 is + O - 1) given by: A {6} ef {9) (3) - F (^) + A (^) - 1) /i„ (0) d^ ^ (^-llX[,„,]MA(^) ^ r (I I the indicator function over [6'p,^, i.e. x\q = gi 1 for 6* e \9p,6\ and zero otherwise. The problem tion is is convex, the Lagrangian continuous. It follows that non-decreasing function A u ty, is optimal is c>L(«;,u;/i„,/i„|A) is /ij^ if and only a non-decreasing function u (5) requires there exists if some = (4) <0 (5) and hu that belongs to the convex cone given by Condition alloca- such that: 5L(u),n;M,u|A) for all and the proposed differentiable : that Q, ^^ R} A {9) = (see Luenberger, Using 0. this A' = {w, u Chapter : w E R and u 8). and integrating (3) by parts leads to dL {w, u- h^, /i„| A) = 7 K {d)+ Je[ [9) 1 {0) dK {0) (6) where, 1(0)^^1 HP It 1) Of {e) -F{9) + A follows that condition (4) requires 7 (^) = d^ + for 9 G (9)) Iy [9,9p] (^i , i.e. - 1 j X[e,^^^dA where u is strictly increasing. This implies, A(^) 9 G {9,9p). The proposed = (l-,5)^/(^) + F(^), (7) allocation thus determines a unique candidate multiplier 17 A the separating region in to be non-decreasing. Op) [9, and assumption ^4 is follows that assumption It necessary and sufficient for A A {6) necessary for the proposed is solution to be optimal. We now A over the whole range A specified A prove sufficiency by showing that there exists a non-decrccising multiplier (^) = for 1 in [9_, 9p) so is we only need satisfies (4) to specify the value of A and for [9p, 9j We've (5). and we set this interval. The constructed A To show that A Op. such that the proposed Uku fi is is not continuous, non-decreasing it that remains all jump has an upward is at 9 and a jump show that the jump to at at 9p upward, \-[{l-P)9pf{9p) + F{9p)]>0, which follows from the definition of 9p. immediate since A jumps from at that 7 (^) < for all The proposed 7 < and that 7 9>9, 1 w and wherever u is u, see this, note that ^p if Otherwise, recall that 9p 9. which implies allocation, = to To 7' {Op) = {1 - p) Of and the Lagrange increasing. Using (6) {9) is - = ^ the result is the lowest 9 such {1 -F (0)) < 0. multiplier. A, imply that it follows that (4) and (5) are satisfied. The figure below illustrates the form of the multipher A {0) proof of the proposition. separating with Figure 2: full flexibility The Lagrange 18 multiplier A {0) constructed in the Proposition 4 shows that under assumption simple. It can be implemented by imposing a minimum or equivalently, a A the optimal allocation maximum Such minimum saving level of savings. pervasive part of social security systems around the The next shows the comparative result spect to temptation bunched As the temptation 6p decreases). (i.e. minimum /5. Proof. That is Op = — C {u {9p)) y is i.e. a policies are optimal allocation with re- more types are decreases, /? decreases with , in the allocation. is less flexibility 9p increases with The minimum savings j3. re- j3. weakly increasing follows directly from its definition. To see that decreasing note that Smin solves 9p U' {y and that 9p, when - Sm\n) _ .. interior, solves, ^-^ Combining these, we obtain is consumption, In terms of policies, as the disagreement increases the Proposition 5 The bunching point Smin increases, extremely vi^orld. statics of the savings requirement decreases so there quirement, Smin level of current is = E[9\9>9p]. E [9\9 > 9p] U' [y - s^in) /W {smm) increasing in 6p the result follows from concavity of U = 1- Since E[9\9>9p] and W. Drilling 3.4 In this subsection we study cases where assumption A does not hold and show that the allocation described in Proposition 4 can be improved upon by drilling holes in the separating section where the condition in assumption Suppose we are offering the unconstrained optimum agents and we consider removing the open interval found their tangency within the interval 6b. The moving to 9a critical issue in to 9a versus and falls Oi,. will move A is not satisfied. for a closed interval [9a {9a, Oh). , Oi,] of Agents that previously to one of the two extremes, 9a or evaluating the change in welfare is counting how many agents For a small enough interval, welfare rises from those moving from those moving to 9h. 19 Since the relative measure of agents moving to the right versus the the slope of the density function this explains if /' > then upon removing {0a, Oh) to the right than the The proof undesirable. is depends on assumption A. For example, more agents would move As a consequence, such a change left. its role in left of the next result formalizes these ideas. Proposition 6 Suppose an 6 € [9a, [^a,^fc] with db < then removing the interior of the set {{u* 9),] {9)) A assumption If the condition in 9p. ,w allocation {u,w) has {u{9) {9) , w* = ,w* does not hold for 9 G [9]) (u* {6) (&„, 9b)} {9)) for /o7' 9 G improves welfare. Proof. In the appendix. Propositions 6 illustrates by construction why assumption A is necessary for a simple 'threshold rule' to be optimal and gives some insight into this assumption. Of course. Proposition 6 only identifies particular fails. It 4 We seems have not characterized the likely that full improvements whenever assumption optimum when assumption A does not 'money burning' may be optimal in some A hold. cases. Arbitrary Finite Horizons We now show that our results extend to arbitrary finite horizons. We confine ourselves to finite horizons because with any mechanism may yield multiple infinite horizons equilibria in the resulting game. These equilibria that a good equilibrium is upon a deviation. 1996). We involve reputation in the sense sustained by a threat of reverting to a bad equilibria Some authors have questioned equilibria in intrapersonal may games (e.g. the credibility of such reputational Gul and Pesendorfer, 2002a, and Kocherlakota, avoid these issues by focusing on finite horizons. Consider the problem with N < oo periods t = 1, ..., TV where the felicity function U (•) in each period. Let 9^ = ^2, •, ^t) denote the history of shocks up to time A direct mechanism now requires that at time t the agent make reports on the history of shocks r* = {r\,r2, ,rl). The agent's consumption is allowed to depend on the whole history of reports: q {r^,r^~^, r^). A strategy for self-t is a mapping is (6'i, t. ..., from the history of shocks and past reports into current reports: Truth telling requires R^ {9^9*-'^, ...,9^) = 9^ for all 20 t and /?* all histories (6'*,r'~\ ••,'"^)- 9K We first argue that without nisms that at time t optimahty we can loss in require only a report whole history of shocks This 6*. on the current shock r^ and not of the 9t, the case in Atkeson and Lucas (1995) but in their is setup since preferences are time-consistent there is mecha- restrict ourselves to a single player and the argument is straightforward. we have In contrast, in the hyperbolic model A'' players and the difference preferences between these selves can be exploited to punish past deviations. example, an agent at time t that amongst them according to whether there has been a deviation to choose past. In particular, she allocations For between allocations can be asked indifferent is in in the can 'punish' previous deviating agents by selecting the worst from their point of view. Otherwise, if there have been no past deviations, she can 'reward' the truth-telling agents by selecting the allocation preferred by them. Such schemes may make deviations more and are thus generally One way to desirable. remove the the refinement that costly, relaxing the incentive constraints, possibility of these when agents punishment schemes are indifferent to introduce is between several allocations choose the one that maximizes the utility of previous selves. Indeed, Gul and Pesendorfer's (2001,2002a,b) framework, discussed in Section control, the hyperbolic model with We punishment schemes. over which ^-agents are indifferent of self- added refinement. this However, with a continuous distribution to rule out these dehvers, in the limit without 6, is for 9 such a refinement show that at for is not necessary any mechanism the subset most countable. This implies that the probability that future selves will find themselves indifferent is zero so that the threat of using indifference to punish past deviations has no deterrent effect. For any set A of pairs (^i, w) define the optimal correspondence over x E M [x; A) = arg X max {xu + w} {u,w)eA (we allow the possibility that Lemma M (x; (Indifference is M {x, A) is countable). A) has two or more points empty) then we have the following For any A the subset X^ C (set of agents that are indifferent) is at X result. for which most count- able. Proof. The correspondence M {x; A) is monotone 21 in the sense that if xi < X2 and M{xi;A) and (ui.wi) e sense, points at (u2,u'2) € M{x2;A) then ui < which there are more than a single element upward 'jumps'. As with monotonic functions, Thus, in an obvious U2- M in {x; follows easily that it A) represent M A) can (.t; have at most a countable number of such 'jumps'. only on the single crossing property of preferences and not on the This result relies linearity in u and w. We make use of this lemma again in Section These considerations lead us to write the problem with A'^ > 5. 3 remaining periods recursively as follows. N Period Problem vn (y) = max / c,y [OU (c (9)) + v^-i {y [9))] clF {9) J 9U + (3vr,-i {c{9)) [y' [9]) > 9hU c {9) where 112 (•) was defined < + y for 3v^^, all ^ G inefficient we assume, without for all < v^-i is ^Q 9 continuation utility w' {9) xu' {9) = without [y' [9)) i.e. t',v-i optimal shows up loss in optimality in the because can be achieved by 'money i'at-i (y' [9)) for For the simple recursive representation to obtain it is some y' [6) critical that, < y' {9). although the and the agent disagree on the amount of discounting between the current and next period, they both agree on the given by uw-i- This is utility obtained from the next period on, not true in the alternative setup where the principal and the agent both discount exponentially but with different discount factors. case separately in Section For any horizon A'' this only required properties all W {) We treat this 7. problem has exactly the same structure as the two-period problem analyzed previously, with the exception that We 9,9 loss in optimality, that the and incentive constraint. This burning' with the same effect: setting principal (ij (/5)) ex-post optimal given the resources available is objective function any {9) (9)^ in Section 2. In the above formulation mechanism + y' (c to be increasing the previous results apply. Vjsi-i (•) and concave and We sition. 22 summarize has substituted since I'yv-i (•) W {). has these this result in the next propo- Proposition 7 Under assumption A riods can be the optimal allocation vnth a horizon of implemented by imposing a minimum amount of saving St N peperiod {yt) in t. In Proposition 7 the minimum preferences the optimal allocation and y' {9, y) = y' {9) y. It saving is is homogenous independent of Proposition 8 Under assumption the N-period linearly a function of resources follows that the optimal saving rate for each period that nism for is A and U (c) = yt- With in y, so that c {9, y) CRRA = c {9) y mechanism imposes a minimum yt- c^^'^ / {I — a) the optimul m.echa- problem imposes a minim.um saving rate St for each, period t independent ofyt- 4.1 Hidden Savings Another property of the optimal allocation identified in Propositions 4 and 7 is worth mentioning. Suppose agents can save, but not borrow, privately behind the planner's back at the same rate of return as the planner, as The in Cole and Kocherlakota (2001). possibility of this 'hidden saving' reduces the set of allocations that are incentive compatible since the agent has a strictly larger set of possible deviations. Importantly, the mechanism described in Proposition 7 continues to implement the same allocation when we allow agents to save privately, To prove this claim and thus remains optimal. we argue that confronted with the mechanism 7 agents that currently have no private savings would never cumulate private savings. To see this, first find in Proposition optimal to ac- it note that by Proposition 7 the optimal mechanism imposes only a minimum on savings in each period. Thus agent-^ always have the option of saving more observably with the principal than what the allocation recommends, yet by incentive compatibility the agent chooses not Next, note that saving privately on his increasing the amount own can be no better for the agent than of observable savings with the principal. This the principal maximizes the agents utility given the resources at from the point of view of the current is to. self, its is true because disposal. Thus, future wealth accumulated by hidden savings dominated by wealth accumulated with the 23 principal. It follows that agents never find it optimal to save privately and the mechanism implements the same allocation when agents can or cannot save Proposition 9 Under assumption A the m.echanism described when agents can save plern.ents the sam.e allocation privately. m privately. Heterogeneous Temptation 5 Consider now the case where the level of temptation, measured by erogeneity in is Proposition 7 im- f3 captures the commonly more Diamond, If likely to is it is the agents that save the least that be 'undersaving' because of a higher temptation to consume (e.g. 1977). the heterogeneity in temptation were due to permanent differences across indi- viduals then the previous analysis would apply essentially unaltered. their random. Het- held view that the temptation to overconsume not uniform in the population and that are ,i5, /3 at time they would truthfully report time be tailored to their /3 as described above"*. it so that their To explore other If agents knew mechanism could possibilities we assume the other extreme, that differences in temptation are purely idiosyncratic, so that is i.i.d. across time and individuals. Thus, each period 9 and (3 are realized together from a continuous distribution - we do not require independence of 6 and results. We continue to assume that For any set A /3 of available pairs [u, arg < 1 for j3 [3 for our simplicity. w) agents with {6, /3) maximize their utility: max </^ —u + w {u,w)eA yjj Note that this arg implies that max set is we can without identical for all types with the loss in optimality same ratio x = 9/ ^ which assume the allocation depend only on X. To see this note that the allocation a given x only the lemma if the x-agent in section 4 is may depend on indifferent showed that the (5. and /3 independently amongst several pairs of set of x for u, w. for However, which agents are indifferent is course, if agents can only report at f = 1 then one cannot costlessly obtain truthful! reports However, with large enough it is likely that the cost of revealing /3 would be small. *0f on 9 N 24 of measure As a consequence, allowing the zero. allocation to depend on 6 or /3 independently, in addition to x, cannot improve the objective function. Without loss in optimality The we limit ourselves to allocations that are functions of objective function is then: E [eu (x) + w {x)] =E[E [Ou (x) + w (x)| x]] = where a of X and (x) =E F (x) {6\x) and / (x) be its is f [a (x) the density over x. Let u{x) + w (x)] / (x) dx X= [x, x] be the support cumulative distribution. Define Xp as the lowest value such that for x E [a (x)| X > modify our previous assumption Assumption A. For x e A " that 2-«'(x) xf'ix)^ /(x) Xp in the following way. we have [x, Xp], > x] X We x only^. - Note that without heterogeneity a l-a(x)/x = (x) /3x so that assumptions A and A are equivalent in this case. Heterogenous Temptation Problem max {a (x) u + ly (x) = xu (x) / [x) + w (x)} dF (x) <-)M) J subject to xu (x) c{u{x)) u{x) + w (x) + + k{w{x)) < > u (y) for all x u (x) dx 1 > y ^Given /3 < 1 a simpler argument is available. The planner can simply instruct agents with given X to choose the element of the arg max with the lowest u, since the planner has a strict preference for the lowest u element. The argument in the text is similar to the one used to extend the analysis to arbitrary horizons and can be applied to the case where /? > 1 is allowed. 25 The proof of the next result closely follows the proof of proposition 4. Proposition 10 Under assumption A agents with x < Xp optimum and agents with x > Xp strained are offered their uncon- are bunched at the unconstrained optimum for agent Xp. Commitment with 6 Self Control Gul and Pesendorfer (2001,2002a,b) introduced an axiomatic foundation We ences for commitment. review their setup and representation result briefly in how we apply general terms and then describe In their static formulation the primitive P {A) choices, with utility function maXagyip(a) this case then P if is some for a set A is for prefer- it is to our framework. a preferences ordering over sets of over choice sets A. In the classical case utility function P (A) = p defined directly over actions. Note that in reduced to A' without removing the best element, a* from A, In this sense, committment, a preference for smaller sets, not altered. is not valued. To model a preference a set A' that is a strict for commitment they assume a consumer may subset .4, i.e. P {A') > P (.4) such preferences can be represented by two by the and A' utility functions C U A. and strictly prefer They show V that over choices a relation: P {A) = max {p (a) + t{a)} — max a€.4 One can think of when choosing a t (a) t (a) ae.4 — maXag^ t (o) as the cost of self-control suffered instead of argmax^g^i (a). dynamic In a by an agent setting recursive prefer- ences with temptation can be represented similarly (Gul and Pessendorfer (2002a,b). In our framework the action and k. is a choice for current consumption and savings, c In order to nest the hyperbolic preferences and Smith (2001) and model we follow use: (c, k) = On t (c, k) = (p p (c) [On 26 + w {k) (c) + i5w (A:)) Krusell, Kuruscu > where the parameter captures the lack of no self-control and yields —* selects oo the agent has difference is whatever is best for his previous 'selves' objective function {9u{e) maximizes (i.e. converge that in the limit oo we obtain a tie-breaking criteria that whenever an agent The / —^ As fully to his temptation. His preferences essentially by the hyperboHc model. The only to those implied as self control. is indifferent he t). is: + w{e))f{9)de + + pwie))f{9)de {9u{9) (p Je Je, -(p f [9u{9)+l3w{9))f{9)d9 Je, where of that {u, is w) is the allocation for the "temptation agent" and chosen by the "self-control agent" a support larger than 9 Given a set of offered that maximizes 9u + = given by (m, w) f3w with = < 9/3/ 13 = (1 + 4>/3) / (1 is the allocation 9. pairs, the "self-control agent" will /3 w) convenient to define everything for where 9 [9, 9] choose an allocation that maximizes 9u will It is . {u, + 0), while the + (3w. the "self-control" agents choose from the same set choose an allocation "temptation agent" Since both the "temptation" and follows that, it u{9)=u{9l3/P), (8) so that the "self-control agent" 9 acts as a "temptation-agent" with a lower taste shock. Substituting (8) into the objective function (1 The + /" (9u 0) first + /3io (913 /l3\) f {9) d9 obtain, - 4> f {9u {9) + (3w {9)) f (9) d9. term can be shown to equal, (1 Note that h represent (^/3//3) we [9) its = + (/.) {9u {9) 13/ /3 / JiJ3/0 f (9p//3] ^/l3 + f3iu {9)) f (9P/I3) P/pd9. ^ is the density of the corresponding distribution function. 27 ' random variable 9P/^; let H {9) The objective function can be written as, 3 + (1 <^) § r§ rS-e r + {9u (9) /3w (9)) hi9)d9-d {9u (9) + Biv {9)) f (9) d9. Substituting in the incentive constraint, 9u (9) + /3w {9) = u{9)de I + vo, Je where vq {l = {9J3/P)u{9) + (3w{§) parts, we obtain: + (P)^Jf il-H{9))u{9)de-(l>J^ {l-F{9))uie)d9+(^il + where we are taking both h and integrating by {9) = 0, for all 9 > intervals of integration as being and / ^/3//5 (^) = for all 9 < 9. <P)p/l3-cP^vo. from ^ to ^ by letting In addition we require u to be non-decreasing and the budget constraint: vo+ ( u - 9u {9) < (9^ d9 (3K~' {y Definition. Let 9p be the lowest value such that for 9 y_7(l + Assumption A. For 9 € (l We now strained for 9p. 0) ^ (1 [9,9p], - ^ m -<P{1-F we have Denote to > 9p, {9))] d9 < that + ,/.)(/3//3)'/(^/3//3)-0/W>O show that the optimal allocation optimum and - C {u {9))) is to offer bunch types higher than this allocation by {w*,u) as before. 28 all types below 9p their uncon- 9p at the unconstrained optimum The Lagrangian L= is jU{l + <l>)^{l-H{9))-<l>{l-F{9))-il-A{e))\u{e)d9 + {/3K-' / {y-C {u (9)))+ 9ii) dk /3 + l(l+</>)^-0 where A is a non-decreasing Lagrange multipUer for the budget constraint, normahzed = so that A(^) 1. Proposition 11 The Proof. We allocation (l-A(0)No allocation {w*,u) is optimal if and only if assumption follow the proof of Proposition 4 as closely as possible. we A holds. At the proposed have: dLiw,u-h^Ji^\A)= l^{{l + <P)^{l-H{9))-cl){l-F{9))-{l-A{9))]h^{9)d9 + Equation (^P l^^[^f^-^]\\9..e]hudA (5) requires A{9) — {1 + 4>) /),„! A) = 7 (/3//9 — + 1). ;i + 0)|-(/.-(l-A(^)) Using this and integrating L, by (3) parts leads to dL {w, u- /i„, [9) hu {9)+ 7 (0) dh^, {9) (9) where, ^{9)^ It Nil + 4>)t [I- H {9))- <t^{l-F [9))- {I- K{9))\d9+e,j'^(^~^ follows that (4) requires 7 (^i) = for 9 G 29 9,9r, , i.e. where u is strictly increasing. This turn requires,' in A 6 G = (^) - 1 + (1 0) (1 - + // [6)) I - F (1 (10) (9)) A GiA'en the proposed allocation, this defines a unique multiplier [9,9p). separating region non-decreasing. [9, It 9p) and assumption A A assumption follows that necessary and sufficient for is A in the {9) to be necessary for the proposed solution is to be optimal. We now A over the whole range specified A prove sufficiency by showing that there exists a non-decreasing multipher (^) = 1 A for 9p) so [9, Q. such that the proposed iv,u we only need and satisfies (4) to specify the value of A for [9p. 9j (5). We've and we set in this interval. A Note that the constructed decreasing all that remains to is not continuous at is show that the jump To show that A 9p. at 9p is upward which is non- requires: -{l + cP)^{l-H{9)) + <f>il-F{9))<0 This follows from the definition of 9p for all 9 9p = >9, which 9 then the result The proposed 7 = A is is is = - {I + 6) the lowest 9 such that 7 9: 3 § [1 - H {9)) + (f>{l (6*) - F (9)) < < 0. If immediate. allocation, whenever u The next that implies 7' {9p) > for 9p w and increasing. and u, Using multiplier, A, imply that 7 (6) it follows that (4) result establishes a connection and < and that (5) are satisfied. between assumptions A and A showing a weaker requirement. Proposition 12 condition for assumption Proof. Let ^ = $ assumption // the condition for - > {9, e) A holds for il + s) [p holds for [9_,6pj3/ (3], [9, 9p] then assumption = A {e) A is equivalent to l^y f 30 (^^ (e) //?) -/ (0) > then the with /3 (e) = $, {9,e) and ^' (e) $, = (2 - /3 = (1 (^, e) assumption + e) + e) / (1 + (/3 p - = /?) [ep/p) / (1 + ^^ Note that $ e). + ^ {9, 0) = = 0, (2/3/ (e/3//3) + p'f (op/p) 9/p) /?' (e) e)^ Thus: ((2 -^+ e) / (^/5//3) + (1 - /3) /' (^/3//5) ^/3//3) A holding at ^ imphes that (2 - /3) f{9) + {l - P) f'{9)§ > 0. This imphes for e > 0. So if the condition in assumption A /(^) + (1 - /3) f'{9)9 > holds for [9,6pP/P] then $e(^,e) $ {9, 0) we have that 0, if A > for all £ holds for [9, > and 9 G [§,9p]. Given that e^P/P], then /o so that assumption A for ^ G 7 Disagreement on Exponential Discounting [:^, ^p] holds. This section departs a bit from our intra-personal temptation environment. We now consider the case where individuals discount the future exponentially but do so at a different rate than a 'social planner'. Caplin and Leahy (2001) and Phelan (2002) provide motivations for such an assumption. Here we simply explore the implications of such a difference in discount rates. It is important to note that this modification constitutes more than just a de- parture on the form of discounting. Our previous analysis relied on the tensions within an individual due to temptation. In contrast, in the current case we require some paternalistic motivation for the social planner's disagreement with agents. a consequence, some analysis. may view this case as more ad hoc and somehow However, we believe that paternalistic motivations government policies. In may less As worthy of be behind several the next section we discuss other examples of paternalism. For the two period case the analysis requires absolutely no change, only a different 31 interpretation for ,3. The now the objective function discounting, it is difficulty is arises from an assumed chfference that the planner and agent will disagree on more than just on the evaluation of future lifetime in and social whether we can extend the results to longer horizons. is to discount future utility relative to present utility: more in private no longer motivated by time inconsistent preferences. The relevant question that remains A difference in discounting in the incentive constraints versus difficult since utility itself. now there is how much also disagreement This makes a recursive formulation the key simplification was that the same value function appeared the objective function and in the incentive constraints. now we Indeed, require two value functions, one for the planner, v, and one for the agent, v^. Fortunately, in the case with logarithmic utility these two value fmictions are related in a simple way. This allows us to keep track of only one value function, V, rendering the analysis tractable. We can show that all our results go through in this case. Consider for the first the situation with three periods. Let the exponential discount factor agent be given by planner The is P and factor for the social highest utility achievable by the agent in the last two periods some constant for the last B'^. (y) = {I + P) log {y) is + B^ For any homogenous mechanism the planner's value function two periods takes the form: V2{y) The important point ficients. assume the discount unity. vt for for simplicity is = 2\og {y) + B'' that these value functions differ only by constants and coef- As a consequence the correct incentive constraint for the written with either value function. That 9U{c{9)) first is, + pv^ {y{e)) > eu 32 [c {ey)+(5i4 {y (b)) , period can be is equivalent to, du where Ps = /3(1 (c [9]) + kv2 {y {9)) > 0U + /3)/2 < 1 is (c (^~)) + Psv2 {y [e^) a fictitious hyperbolic discount factor three periods to go. Note that the incentive constraint (11) has all (n) , when there are the features of the hyper bohc discounting case. Thus, we can we can write the three period problem disagreement on exponential discounting in the same way as the hyperbolic discounting problem. With k remaining periods the discount generalizes to any finite horizon. must be applied Note that /3k is The arguments factor that is decreasing in k and converges to zero (this last feature the planner not discounting the future at is special to all). Other Interpretations 8 In this section we discuss how our model can be reinterpreted for other applications. Paternalism 8.1 Another interesting application of the model to a paternalistic problem between schooling and leisure choice. In many an adult so that we can interpret paternalism preferences of parents and child. cases the relevant agent literally as Alternatively, other adults child's preference are given represents schoohng time and parameter 9 / by the utility function may be parent's preferences are given by 6U (s) + on schooling W (s) , so that not yet is altruistically legislation. 9U (/) + /3W (s)where vs. s The taste other activities. The represents other valuable uses of time. affects the relative value placed the choice a struggle between the concerned about children without parents and support paternahstic The is more weight is given to schooling time. The s + < I allocation of time 1. In this example is constrained by the time no insurance is possible. 33 endowment normalized to one, Externalities 8.2 Another origin a divergence of preferences between the planner and the agents for when consumption of a ize the effects of their To make good generates positive The [6- (c this precise, last () , k ())) suppose there are two goods, = eU [c {d)) + m when the {k (9)) and /c, that the agent with entire allocation W + fv {9, (c, k)) dF {9) = Thus, we can represent {k {9)) which same (c [6) dF {9) ,k{9)) (12) {1-(3)J 9U (c) + W I {BU is + (c {9)) k. The utilitarian W [k {9))) dF {9) (k) as the relevant utility function for agent-^ from the planner's point of view. Although this agent-6', is is: W= to the c term captures the externality from the consumption of welfare criterion by Agents do not internal- externalities. consumption on other agents but the planner does. taste shock 9 obtains the following utility V is is given by the expression in (12), not the utility actually attained it is an interpretation that leads welfare functional. Conclusion 9 This paper studied the optimal trade-off between commitment and intertemporal consumption/saving model without insurance. flexibility in an In our model, agents expect to receive relevant private information regarding their tastes which creates a demand for flexibility. value commitment. for flexibility But they also expect to suffer The model combined the from temptations, and therefore representation theorems of preferences introduced by Kreps (1979) with the preferences for commitment pro- posed by Gul and Pesendorfer (2002). We setting solved for the optimal solution that trades-off up a mechanism design problem. We commitment and We by showed that under certain conditions the optimal allocation takes the simple threshold form of a ment. flexibility minimum savings require- characterized the condition on the distribution of the shocks under which this result holds, and showed that if this condition 34 is not satisfied, more complex mechanisms might be optimal. Future work with a hope of understanding The model who open is how it may on the case with insurance, will focus affect the results to a variety of other interpretations. cares about an agent but believes the agent is obtained here. A paternalistic principal biased on average in his choices would face a similar trade-off as long as the agent has some private information regarding his tastes that the planner also cares about. We discussed applications to schooling choices by teenagers and situations with externalities. A Proof of Proposition 6 Suppose that we are for 9l offering a segment of the budget between the tangency point line and that of Oh, with associated allocation cl and ch- Define the inchfferent from the allocation c^ and Ch then 9* G {9i,9h) for 9h > 9* that is Upon 9l- removing the interval 9 € {9*,9h) types move to Ch and 9 £ {Oi,0*) types move to Cl allocation. Let A{9h,9l) be the change in utility for the planner of such a move (normalizing income to y = 1 for simplicity) A{9h,9l)= / {9U{c*{9H)) + W{y-c*{9H))}fi9)d9 {eu[c*[9L)) + w{y-e{eL))}f{9)d9 6l Bh {9U{c*{9)) where the function c* [9) is + W{y-c*{9))}f{9)d9 defined implicitly by eU'[c*{9)]^l3W'{y-c*{9)) and 6* {9h, 9l) is (13) then defined by 9* = {9H,eL)U {c* 9* {9h, 0l) {en)) u {c* + I3W {y - {9l)) Notice that A{9l,9l) =0. 35 c* {9h)) + m{y- e {9jy) , (14) The lemma regarding following Lemma. The A {9h,6l) partial of 9^ the partial derivative oi A{9h,0l) n ^n with respect to 6h can be expressed ,.,U' ofn , [9] 9*) is Proof. (c* We Ai as: 89 H defined by. = {y-l3){9-9*)9*f[9*) S[9,9*) Since U' used below. {Oh)) dc* [Oh) [c* P where 5 is > {9h)) and dc-{eH) > d0H 0, f t^e) = then sign (Aj) 9)d9 {t sign {S{9h,9*)) have = {9h, 9l) [9hU - [r + + {c* {9h)) {9h.9j^) U (c* {9U' I {e* {9h. 9l) - [OhU {c* W [y - + (c* U {c* {9h)) {9h)) {9h)) {9l)) c* {9h))\ W + - W + W + W (y - - c* c* - {y [y {9h)) (^//) 89* c* [9^))] - - {y f c* f {9'] W H ^-^^d9 {9h))} f (9) 89* c* {9^))} f f [9*] 89 H {9h)] Combining terms, Ai(^;/,eL) /'" + {r = [eU' {c* {9„)) {9h,Bi^) [U [c* {9l)) Now, from (14) - W - U [y {9h))] f {9) d9) ^K^ W [y - - [c* {9h))\ + c* [9^)) W {y - we have 9U'\c{9)\-W'{y-c{9)) = Substituting above 36 /3-1 9U' [c* 89* c* {9^))} f (9*) 89h Ai(^h,^l) = 9--^9H)f{0)d9^U'{c*{en))^^^ 'O'(0H.eL) + we {e* {9h, 6l) [U (c* {6,)) also have that 9* {9h,9l' - U {c* {9h))] + W{y-c* {9^)) - W {y 39* - c* (9^))} f {9* 39 H from (13) [U {c* {9l)) - U = {W {y - (c* i9H))] c* (0^)) -W{y-c* {9h))} (3 So, ^i{9h,9l) 39* = e* f {9*)\[IT {c* {9h)) -U [c* {9^))] 39 .e„ j^" {^^9h - 9^ f {9)d9^U' Differentiating (14) we [U Using the {9h)) - U {c* {9,))] = - fact that 9U' [c* {9)\ - pW {c* ^ [U {c* 3c* {9h) {c* {9h)) 39 H obtain: 30* ^ H {9h)) - U {c* [9*U' {c* {9h)) {1 - c* {9)) = {9,))\ [Oh - = - 9*] ^W {y ~ this U' 3c* c* {9h))\ 7H) ^^^ imphes [c* {9h)] ^-—^ Substituting back the result follows. From the derivatives lemma we we only need to sign 5 {9h, 9*). Clearly, 5 {9* , 0*) also get that 3S{9,9*) 39 = [l-P]e*f{9*)-{l-/3)9f{9)- / f{9)d9 Notice that dS {9, 9* = 39 3-S{9,9*) = -{2-P)fi9)-{l-/3)9f'{9) id9f 37 = 0. Taking Note that 9^5 (0,r)/(ae)^ does not depend on ^*,just on 5. It ^"gyM < follows that sz.gn ( if and only if - f{0) That is, if A 9-5 holds. Integrating {9, 9*) / [d9f ^ ' twice: red~S[d,9* ren S{9h.B*)= 1-/3 ' / / / , dOdB d9] Thns S{9h,0*) < if >! holds. This impUes then that Ai [9, 9l) < for all ^ ^{9h,9l)= so that ^7^ > -^ and clearly 9i G argmaxe^>9^ => A {Oh, 9l)<0 punching holes into any offered interval The converse is also true: if A ^i, if ; for all In other words is holds; and if 9h and 9^ assumption A holds then not optimal. it is optimal to remove the whole interval. In other words, ^o) A does not hold for some open interval 9 € {di.O^) then the previous calculations show that (6'i, eissumption A,{9;9L)d9 / A {9h^ 9l) > e arg S.t. max A {9h, 9l) ei<9L<9H < This concludes the proof. 38 02 Lemma on B Differentiability Definition. Given a function x,h E spaces. If for T lim Lemma. called the it is X Q C and X and Y are normed the Umit: fl a—>0 exists then ^^ Y, where Q. : a Gateaux [T (x + ah) - T {x)] differential for x,h ^ Q, and is denoted by dT [x] h). Let T{x)= I g{x{e),e)dii{9) Je (9,0,/.;,) is in any measure space (not necessarily Q some space which for T is and a G some for [0, e] is e and (a minimum Suppose g integrability). continuous in x, for > : Q ^ R^ defined (an arbitrary restriction or perhaps a required restriction to ensure measurability gx {-,0) exists a vector space) and x i? or Then all 9. requirement {-,9) is as long as x + concave and a/i G fi for for existence! of course) then: dT{x;h)= f gAx{9),9)h{9)d^t{9) Je if this expression Proof. dT (x; By h) is well defined. definition = lim - a^O a [T {x + ah) - T (x)] = hm [ -[g {x {9) + ah {9) ,9)-g{x (9) = [ , 9)] dp {9) gAx{9),9)h{9)dfL Je lim / "^0 Je since for a < e -[g{x{9) + ah{9) ,9) rv - g{x{e) ,9)]- gAx{9) ,9)h{9) we have -[g{x{9) a < -Jg{x + ah{9) (9) ,9) - g{x{9) ,9)]- g,{x[9) + Eh {6) ,e)-g{x [9) 39 , 9)] - ,9)h{9) g, [x {9) , 9) h {9) d^i{9) by concavity of g. + eh {9) Since g {x (9) sumption it any function / chapter cise 7.26, pg. 192, , 9) + follows that i [g {x (9) integrable. Since gx {x {9) g (x (9) ,6), h (9) eh and [9] g^, (.r integrable We (9) ,9)- g{x if SLP] we have that integrable. is \ 8, is 9) , , {9) , and only 9)] are integrable by as- {9) -g^ {x {9) + eh h 9) , (9) is also integrable [see Exer- if |/| is {x (9) j^ [g h 9) {9) .9) - g (,t {9) , ^)]- then have the conditions for Lebesgue's Domi- nated Convergence Theorem. It follows that: - Hm a (x {9) [g Hm - 'e Q—»o a [g + ah + ah [9] {x {9) by continuity oig^ and (9) ,9) is = h Xi — ^ ,9) + ah Xo and non- decreasing functions , - 9)] g {x {9) , 9)] follows that g, {x {9) - , dT {x; h) = .ri G E ^ fi. for for [0, e] , 9) h (9) Jq in some {9) g^^ is d^L {x {&) , = 9) h ((?) (i/i. ST{xo;h) then the £ Note that the case where convex and so is a E h 9) g, [x {9) convex and that we are interested obvious requirement that xq if g {x {9) its definition. It Remark: Suppose Q and only - > fl is is satisfied if the space of the space of measurable functions. References [1] Athey, Susan; mahty [2] Andy Atkeson and of Transparent Atkeson, Monetary Policy", Staff Andrew and Robert Kehoe Patrick (2003) "On the Opti- Report 326. E. Lucas, Jr. (1992) "On Efficient Distri- bution with Private Information." Review of Economic Studies, 59, no. 3 (July): 427-53. [3] Caplin, Andrew and John Leahy "The Social Discount Rate", mimeo, New York University. [4] Cole, Hai-old L. and tions with Narayana R. Kocherlakota Hidden Income and Hidden Storage" v68, n3 (July): 523-42. 40 , (2001) "Efficient Alloca- Review of Economic Studies, [5] Diamond, Peter (1977) "A Framework for Social Security Analysis", Journal of Public Economics, v8, n3: 275-98. [6] Diamond, Peter and Botand Koszegi (2002) "Quasi-Hyperbolic Discounting and Retirement", mimeo, MIT. [7] Feldstein, Martin S. (1985) "The Optimal Level of Social Security Benefits", Quarterly Journal of Economics, vlO, n2, May: 303-20 [8] Fudenberg, D. and Tirole, [9] Gul, J. , "Game Theory". 1991. MIT Press. Faruk and Wolfgang Pesendorfer, 2001 "Temptation and Self- Control", Econom,etrica, Vol. 69 (6) pp. 1403-35 [10] Gul, Faruk and Wolfgang Pesendorfer, 2002a of [11] consumption" forthcoming , in [12] Holmstrom, Bengt in (1984) Economic Theory, Studies in of Delegation", in Bayesian Bayesian Econometrics, E.: vol. 5, Models edited by Beyer, 115-41. Imrohoroglu, Ayse; Selahattin Imrohoroglu and Douglas H. Joines (2000) "Time Inconsistent Preferences and Social Security", Federal Reserve Minneapolis Discussion Paper [14] "Self-control, revealed pref- prepared for Society for Economic Dynamics. "On the Theory Marcel and Kihlstrom, Richard [13] , and the theory Econometrica. Gul, Faruk and Wolfgang Pesendorfer, 2002b. erence and consumption choice" "Self-control Number 136 Kocherlakota, Narayana 1996 "Reconsideration-Proofness: Infinite Horizon Time Inconsistency" , of Games and Economic A Refinement for Behavior, Vol. 15 (l),pp 33-54 [15] Kreps, David, 1979 "A Preference [16] Krusell, Per; Burhanettin tation [17] for Flexibility", Econometrica, 47: 5 65-576. Kuruscu and Anthony Smith (2001) and Taxation", mimeo. Laibson, David (1994) "Self-Control and Saving", M.I.T. 41 thesis. "Temp- [18] David Laibson, (1998) "Life-Cycle Consumption and Hyperbolic Discount Functions", European Economic Review, Vol. 42, nos. 3-5: 861-871. [19] Luenberger, David G. (1969) "Optimization by Vector Space Methods", John Wiley [20] [21] & Sons, Inc. Mirrlees, James Taxation", The Review (1971) Economic Studies, Model of Sin Taxes", the Theory of Vol. 38, No. 2., Optimum Income pp. 175-208. (2003) "Studying Optimal Pater- mimeo, Phelan, Christopher (2002) "Opportunity and eral [23] in O'Donaghue, Ted and Matthew Rabin nalism, Illustrated by a [22] of "An Exploration UC Berkeley. Social Mobility" , mimeo, Fed- Reserve Bank of Minneapolis. Phelps, Edmund S. and R.A. Pollack (1968) "On Second Best National Savings and Game-Equihbrium Growth", Review of Economic Studies, Vol. 35, No. [24] 2., Ramsey, F.P. vol. 38, [25] pp. 185-199. "A Mathematical Theory of Savings" Economic Journal, 543-559. Sheshinski, Eytan (2002) "Bounded Rationality and Socially Optimal Limits on Choice [26] (1928) in a Self-Selection Weitzman, Martin Studies, v41, n4 : L., Model" mimeo. , (1974) "Prices vs. Quantities", Review of 477-91. 42 Economic kk53 003 Date Due MIT LIBRARIES L 3 9080 02617 791 t e f t !<»-« " Jh W* Wlirv j.aahW'^liifcdl*»3l»JI|gafc