Chapter 2: DSGE Models, Solutions and Approximations

This chapter describes some standard Dynamics Stochastic General Equilibrium (DSGE) models which will be used in examples and exercises throughout the book. Since these models do not have a closed form solution except in very special circumstances, we also present a number of methods to obtain approximate solutions to the optimization problems. There is a variety of models currently used in macroeconomics. The majority is based on two simple setups: a competitive structure, where allocations are, in general, Pareto optimal; and a monopolistic competitive structure, where one type of agents can set the price of the goods she supplies and allocationsare suboptimal. Typically, these two structures require dierent approaches to nd an expression for the variables of interest in terms of the exogenous forces and the states. When competitive allocations are Pareto optimal the principle of dynamic programming is typically used and iteration on Bellman equation are employed to compute the value function and the policy rules whenever they are known to exist and to be unique. As we will see, calculating the value function is a complicated enterprise. Two simple specications which deliver closed form solutions do exist, but they are unpalatable for many economic problems. For general preference and technological specications, approximations are employed. Two methods are typically used: quadratic approximations of the return function, and discretization of the dynamic programming problem. When the equilibrium allocations are distorted, one must alter the dynamic programming formulation and in that case the recursive formulation provided by the Bellman equation does not have a clear advantage over a more standard stochastic Lagrangian multipliers methodology, where one produces rst order conditions and uses them, together with the transversality condition, to nd an equilibrium. In general, a solution is hard to nd also with the Lagrangian approach especially when the problem is nonlinear since it involves expectations of future variables. Euler equation methods, which approximate the rst order conditions, the expectational equations or the policy function can be used in these frameworks. Many methods have been developed in the literature. Here we restrict attention to the three widely used approaches: discretization of the state and shock space; log-linear approximations; and parametrizing expectations. For a thorough discussion of the various methodologies see Cooley (1995, chapters 2 and 3) or Marimon and Scott (1999). 29

30 The next two sections illustrate features of the various models and the mechanics of solution methods with the aid of examples and exercises. A comparison between various approaches concludes the chapter.

2.1

Few useful models

It is impossible to provide a thorough description of the models currently used in macroeconomics. Therefore we focus attention to two prototype structures: one involving only real variables and one considering also nominal ones. In each of the two classes we analyze both models with representative and heterogeneous agents and consider both optimal and distorted setups.

2.1.1

A basic Real Business Cycle (RBC) Model

where ct is consumption, Nt is employment (hours) and Kt is capital and E0 E [.|F0 ] is the expectations operator, conditional on the information set F0 . The instantaneous utility function is bounded, twice continuously dierentiable, strictly increasing and strictly concave in all arguments. It depends on ct and ct1 to account for possible habit-persistence in consumption and we assume that 0 < < 1. The maximization of (2.1) is subject to the sequence of constraints ct + Kt+1 (1 T y )f (Kt , Nt , t ) + Tt + (1 )Kt 0 Nt 1

A large portion of the current macroeconomic literature uses versions of the one sector growth model to jointly explain the cyclical and the long-run properties of the data. In the basic setup we consider there is a large number of identical households that live forever and are endowed with one unit of time, which they can allocate to leisure or to work, and K0 unit of productive capital, which depreciates at the rate 0 < < 1 every period. The social planner chooses {ct , Nt , Kt+1 } t=0 to maximize X t u(ct , ct1 , Nt ) (2.1) max E0
t

(2.2) (2.3)

where f (.) is twice continuously dierentiable, strictly increasing and strictly concave in Kt and Nt production technology; t is a technological disturbance, T y is a (constant) income tax rate and Tt lump sum transfers. There is a government which nances a stochastic ow of current expenditure with income taxes and lump sum transfers: expenditure is unproductive and does not yield utility for the agents. We assume a period-by-period balanced budget of the form Gt = T y f (Kt , Nt , t ) Tt ct + Kt+1 (1 )Kt + Gt = f (Kt , Nt , t ) (2.4)

The economy is closed by the resource constraint, which provides a national account identity: (2.5)

Methods for Applied Macro Research 2: Models and Approximations

31

Example 2.1 In (2.4) we have assumed that the government balances the budget at each t. This is not restrictive since consumers in this economy are Ricardian; that is, the addition of government debt in the model does not change the allocations. This is because, if debt B) = is held in equilibrium, it must bear the same rate of return as capital, so that (1 + rt f Et [fk (1 T y ) + (1 )] where fk = K . In other words, debt is a redundant asset and y can be priced by arbitrage, once (, T , fk ) are known. One example where debt matters is considered later on. Exercise 2.1 Decentralize the RBC model so that there is a representative consumer and a representative rm. Assume that the consumer makes the investment decision while the rm hires capital and labor from the consumer. Is it true that decentralized allocations are the same as those obtained in the social planners problem? What conditions need to be satised? Repeat the exercise assuming that the rm makes the investment decision. Exercise 2.2 Set ct1 = 0 in (2.1) and assume T y = 0 t. Rewrite the problem in a recursive format by dening a value function V as the utility value of the optimal plan, given (Kt , t , Gt ). i) Dene the states and the controls for the problem, that is the variables which characterize the state of the economy at each t and the choice variables. ii) Show that the problem described by (2.1)-(2.5) can be equivalently written as: V(K, , G) = max {u([f (K, N, ) + (1 )K G K + ], N ) + E [V(K + , + , G+ )|K, , G]} (2.6) where E (V|.) is the expectation of V conditional on the available information and the superscript 00 +00 indicates future values. (2.6) is the so-called Bellman equation . 1 Nt . iii) Assume u(ct , ct1 , Nt ) = ln ct + ln(1 Nt ) and that GDPt f (Kt , Nt , t ) = t Kt K c Find steady state values for ( GDP , GDP , N ). There are few conditions that need to be satised for a DSGE model to be tted into a Bellman equation format. First, the utility function must be time separable in the contemporaneous control and state variables. Second, the objective function and the constraints have to be such that current decisions aect current and future utilities but not past ones. These conditions are somewhat restrictive but, as we will see, problems can be reformulated so that they are typically satised. Exercise 2.3 Show how to modify Bellman equation (2.6) when T y 6= 0. The solution to (2.6) is typically complicated since the function V is unknown and there is no analytic way to nd it. Had the solution been known, we could have used (2.6) to dene a function h mapping every (K, G, ) into (K + , N ) that gives the maximum. Since V is unknown, methods to prove its existence and uniqueness and to describe its properties have been developed (see e.g. Lucas and Stokey (1989)). These methods implicitly provide a way of computing a solution to (2.6) which we summarize next:
{K + ,N }

32 Algorithm 2.1 1) Choose an initial dierentiable and concave function V0 (K, , G). 2) Compute V1 (K, , G) = max{K + ,N } {u([f (K, , N ) + (1 )K G K + ], N ) + E [V0 (K + , + , G+ )|K, , G]}. 3) Set V0 = V1 and iterate on 2) until |Vl+1 Vl | < , small. 4) When |Vl+1 Vl | < , compute K + = h1 (K, , G) and N = h2 (K, , G). In algorithm 2.1 V can be obtained as the limit of Vl for l . Under some regularity conditions, this limit exists, it is unique and the sequence of iterations dened by algorithm 2.1 achieves it. In general, the iteration process is complicated when the combined number of states and shocks is large. Moreover, unless V0 is appropriately chosen, the iteration process may be time consuming. In a few simple cases the solution to the Bellman equation has a known form. We analyze one of these cases in the next example. Example 2.2 Assume that u(ct , ct1 , Nt ) = ln ct , = 1; the production function has the 1 Nt ; ln t is an AR(1) process with persistence , and Gt = T y = Tt = form GDPt = t Kt 0. Guess that the value function has the form V(K, ) = V0 + V1 ln Kt + V2 ln t . Then Bellman equation maps logarithmic functions into logarithmic ones. Therefore the limit, if it exists, will also have a logarithmic form. Sargent (1986, p. 26) describes in details how to compute the constants V0 , V1 , V2 using the method of undetermined coecients (see later on). In words, we need four steps: (i) substitute the guess in (2.6); (ii) use (2.5) and the guess to nd an expression for K + ; (iii) use the envelope theorem to get an expression for V(K, ) K ; (iv) dierentiate the guess and match coecients with those in (iii). These four 1 steps imply that a candidate solution is K + = t (1 )K 1 and that V1 = 1(1 ) . With these two expressions Bellman equation becomes
1 1 1 V0 +V1 ln Kt +V2 ln t = ln(t Kt t (1 )Kt )+ [V0 +V1 ln(t (1)Kt )+V2 ln t ] (2.7) ln(1 (1 ))+ V1 ln( (1 )) and Matching coecients on the two sides of (2.7) we have V0 = 1

V2 =

1+ V1 1

and these values conrm that the optimal policy is K + = t (1 )K 1 .

Exercise 2.4 Assume that both (t , Gt ) are iid and maintain all the other assumptions of example 2.2 except Gt = Tt . Guess that the value function is V(K, , G) = V0 + V1 ln Kt + V2 ln t + V3 ln Gt . Determine V0 , V1 , V2 , V3 . Show the optimal policy for K + . Two other simple cases of interest where solutions to the Bellman equation can be found analytically are analyzed in the next exercise.

Methods for Applied Macro Research 2: Models and Approximations

33

y Exercise 2.5 i) Suppose that u(ct , ct1 , Nt ) = a0 + a1 ct a2 c2 t and that Gt = Tt = T = 0 t. Show that the value function is of the form V(K, ) = [K, ]0 V2 [K, ] + V0 . Find the 0 2 optimal values of V0 and V2 . (Hint: use the fact that E (e0 t V2 et ) = tr(V2 )E (et et ) = tr (V2 )e 2 where e is the covariance matrix of et and tr(V2 ) is the trace of V2 ). Show that the decision rule for c is linear in K and the one for K + is linear in K and .

ii) Suppose u(ct , ct1 , Nt ) = 1t ; Kt = 1 t and assume that t can take n nite values. Let t evolve according to P (t = i|t1 = i0 ) = pii0 > 0. Assume that there are claims to the output in the form of stocks St with price ps t and dividend sdt . Write down Bellman equation for this problem. Let = 0.9, pii = 0.8, i = 1, . . . , n; pi,i+1 = 0.2 and pii0 = 0 otherwise, n = 3. Show the rst two terms of the value function iterations in this case. Can you guess what the limit is? Part ii) of the exercise 2.5 shows that if the shocks take only a nite number of values and no states exist, an analytical solution to the value function can be typically computed. We can relax some of the assumptions we have made (e.g. we can use a more general law of motion for the shocks), but except for these simple cases, even the most basic stochastic RBC model does not have a closed form solution. As we will see later, existence of a closed form solution is not necessary to estimate the structural parameters (here , , and the parameters of the process for t and Gt ) and to examine its t to the data. However, a solution is needed when one wishes to simulate the model, compare its dynamics with those of the data, and/or perform policy analyses. There is an alternative to Bellman equation to solve simple linear quadratic problems. It involves substituting all the constraints in the utility function (if possible) and maximizing the resulting expression either unconstrained or using a stochastic Lagrangian multiplier approach. We illustrate such an approach with an example. Example 2.3 Suppose a representative consumer P which obtains utility from the services of durable and nondurable goods according to E0 t t (cst t )0 (cst t ) where 0 < < 1, t is a preference shock and consumption services cst satisfy cst = b1 cdt1 + b2 ct where cdt1 is the stock of durable goods, accumulated according to cdt = b3 cdt1 + b4 ct where 0 < b1 , b3 < 1 and 0 < b2 , b4 1 are parameters. Output is produced with the technology f (Kt1 , t ) = (1 )Kt1 + t where 0 < 1 and t is a productivity disturbance, and divided between consumption and investment goods according to b5 ct + b6 invt = GDPt . Physical capital accumulates according to Kt = b7 Kt1 + b8 invt , where 0 < b7 < 1, 0 < b8 1. Using the denition of (cst , cdt , Kt ) and the resource constraint we have cst + cdt = (b1 + b3 )cdt1 + (b2 + b4 ) b6 ((1 )Kt1 + t (Kt b7 Kt1 )) b5 b8 (2.8)

c1

+b4 ) b6 , b11 = b10 b , b12 = b11 b7 and using (2.8) in the utility Letting b9 = (b1 + b3 ), b10 = (b2b 5 8 function the problem can be reformulated as X max E0 t C1 [cdt , Kt ]0 + C2 [cdt1 , kt1 , t , t ]0 )0 (C1 [cdt , Kt ]0 + C2 [cdt1 , kt1 , t , t ]0 ) {cdt ,Kt } t

(2.9)

34
0 C is invertible, and the where C1 = [1, b11 ], C2 = [b9 , b12 + b10 (1 ), b10 , 1]. If C1 1 shocks (t , t ) are known at each t, the rst order condition of the model imply [cdt , Kt ]0 = 0 C )1 (C 0 C ) [cd (C1 1 t1 , Kt1 , t , t ]. Given (cdt , Kt , t , t ), values for cst and ct can be found 1 2 from (2.8) and from consumption services constraint.

it into an optimal Exercise 2.6 Take the model of example 2.3 P but let t0 = 0. Cast 0 + 2y Q y 0 ] subject + y Q y linear regulator problem of the form max E0 t t [y2t Q2 y2 1 t 1 2t 3 1t t 1t to y2t+1 = Q4 y2t + Q5 y1t + Q6 y3t+1 where y3t is a vector of (serially correlated) shocks, y2t a vector of states and y1t a vector of controls. Show the form of Qi , i = 1, . . . , 6. The stochastic Lagrangian approach provides an alternative to Bellman equation to solve stochastic models; it works generally and requires relative weak conditions. For example, u(ct ,ct1 ,Nt ) ,ct1 ,Nt ) f , UN,t = u(ct , the Euler (expectational) letting fN = N , Uc,t = ct Nt equation for capital accumulation in the basic RBC model is Et Uc,t+1 [(1 T y )fk + (1 )] 1 = 0 Uc,t Uc,t 1 = UN,t (1 T y )fN (2.10)

while the intratemporal marginal condition between consumption and labor is: (2.11)

(2.10)-(2.11), the budget constraint and the transversality condition, limt t Kt = 0, need to be solved for (Kt+1 , Nt , ct ), given (Gt , t , Kt ). This is not easy: the system of equations is nonlinear and involves expectations of future variables. In general, no analytical solution exists. Versions of the basic RBC model with additional shocks, alternative inputs in the production function or dierent market structures have been examined extensively in the macroeconomic literature. We consider some of these extensions in the next four exercises. Exercise 2.7 (Utility producing government expenditure) Consider a basic RBC model and suppose that government expenditure provides utility to the agents; that private and public consumption are substitutes in the utility function; and that there is no habit in consumption, e.g. U (ct , ct1 , Gt , Nt ) = (ct + G Gt ) (1 Nt )1 . i) Using steady state relationships describe how private and public consumption are related. Is there some form of crowding out? ii) In a cross section of steady states, is it true that countries which have higher level of government expenditure will also have lower levels of leisure (i.e. is it true that the income eect of distortionary taxation is higher when G is higher)? Exercise 2.8 (Production externalities) In a basic RBC model assume that the production function depends on rm specicR inputs and on the aggregate capital stock, i.e f (Kit , Nit , t , Kt ) = K 1 N , > 0 and K = K di. Kt t it it it t i) Derive the rst order conditions and discuss how to nd optimal allocations. ii) Is it appropriate to use dynamic programming to nd a solution to this problem? What modications do you need to introduce to the standard setup?

Methods for Applied Macro Research 2: Models and Approximations

35

Exercise 2.9 (Non-competitive labor markets) Assume that, in a basic RBC model , there are one-period labor contracts. The contracts set the real wage on the basis of the expected marginal product of labor. Once shocks are realized, and given the contractual real wage, the rm chooses employment to maximize its prots. Write down the contractual wage equation and the optimal labor decision rule by rms. Compare this decision rule with a traditional Phillips curve relationship ln Nt Et1 (ln Nt ) ln pt Et1 (ln pt ). Exercise 2.10 (Capacity utilization) Assume Gt = Tt = T y = 0; that the production function depends on both capital (Kt ) and its utilization (kut ) and it is of the form f (Kt , kut , Nt , t ) = t (Kt kut )1 Nt . This production function allows rms to respond to shocks by varying utilization even when the stock of capital is xed. Assume that the higher is the use, the more 2 capital depreciates. In particular, assume (kut ) = 0 + 1 ku t where 0 , 1 and 2 are parameters. i) Write down the optimality conditions of the rms problem and Bellman equation. ii) Show that, if capital depreciates instantaneously, the solution of this problem is identical to the one of a standard RBC model examined in part ii) of exercise 2.2. One objection often present in the literature concerns the use of Solow residuals as a proxy for technological disturbances. The main reason is that they tend to overstate the variability of these shocks and the may contain not only technology but also important demand shocks. The example below provides a case where this can occur. Example 2.4 Suppose that output is produced with part-time hours (N P ) and full-time 1 1 (NtF ) + t Kt (NtP ) . Typically, hours (N F ) according to the technology GDPt = t Kt Solow accounting proceeds assuming that part-time and full time hours are perfect substitutes 1 (NtF + N P ) so that an and use total hours in the production function, i.e. GDPt = t Kt d estimate of t is obtained via ln t = ln GDPt (1 ) ln Kt ln(NtF + NtP ) where is the d share of labor income. It is easy to see that ln t = ln t +ln((NtF ) +(N P ) ) ln(NtF + N P ) d so that the variance of ln t overestimates the variance of ln t . This is a general problem: whenever a variable is omitted from an estimated equation, the variance of the estimated residuals is at least as large as the variance of the true one. Note also that if NtF > NtP and if NtF is less elastic than NtP to shocks (e.g., if there are dierential cost in adjusting full and part-time hours), ln((NtF ) + (N P ) ) ln(NtF + N P ) > 0. In this situation any (preference) shock which alters the relative composition of N F , N P will induce procyclical labor productivity movements even if t = 0, t. Several examples in this book are concerned with the apparently puzzling correlation between hours (employment) and labor productivity - the so-called Dunlop-Tashis puzzle. What is puzzling is that this correlation is roughly zero in the data while is high and positive in an RBC model . We will study later how demand shocks aect the magnitude of this correlation. In the next example we examine how the presence of government capital alters this correlation when an alternative source of technological disturbance is considered.

36
t , the budget Example 2.5 (Finn) Suppose agents utility is u(ct , ct1 , Nt ) = t 1 y K P B P +B constraint is (1 T )wt Nt + (rt T (rt ))Kt + Tt + (1 + rt )Bt = ct + invt t+1 P P P K and private capital evolves according to Kt+1 = (1 )Kt + invt where T are capital taxes, rB is the net rate on real bonds and rt the net return on private capital. Suppose the P +B G B Government budget constraint is T y wt Nt + T K (rt )Kt t+1 = invt + Tt + (1 + rt )Bt G G where government investments increases government capital according to invt = Kt+1 g . The production function is GDPt = t N (K P )1 (K G ) ; output is used for (1 )Kt private consumption and investment; and 0. This model does not have an analytic solution but some intuition about some of its G . Suppose properties can be gained by analyzing the eects of random variations in invt G that invt is higher than expected. Then, less income is available for private use and, at the same time, more public capital is available in the economy so that the productivity of private factors increases. Which will be the dominant factor depends on the size of the investment increase relative to . If it is small, there will be a positive instantaneous wealth eect so that hours, investment and output decline while consumption and labor productivity increases. If it is large, a negative wealth eect will result in which case hours and output will increase and consumption and labor productivity decreases. In both cases, despite the RBc structure, the contemporaneous correlation between hours and labor productivity is negative.

(c (1N )1 )1

2.1.2

Heterogeneous agent models

Although representative agent models constitute the backbone of current dynamic macroeconomics, the literature has started examining setups where some heterogeneities in either preferences, the income process, or the type of constraints that agent face are allowed for. The presence of heterogeneities does not change the structure of the problem: it is only required that the sum of individual variables match aggregate ones and that the planner problem is appropriately dened. We consider few of these models here. Since the scope is purely illustrative we restrict attention to situations where there are only two types of agents. The generalization to a larger but nite number of agents type is straightforward. Example 2.6 (A two country model with full capital mobility) Consider two countries and one representative agent in each country. Consumers in country i choose sequences for con 1 ]1 P t [cit (1Nit ) sumption, hours, capital and contingent claim holdings to maximize E0 t=0 1 subject to the following constraint cit + X
j

b Kit+1 Bjt+1 pB 1)2 )Kit (2.12) jt Bjt + wit Nit + rit Kit (Kit+1 (1 )Kit ( 2 Kit

where wit Nit is labor income; rit Kit is capital income ; Bjt is a set of Arrow-Debreu one period contingent claims and pB jt is its price; b is an adjustment cost parameter and the depreciation rate of capital. Since nancial markets are complete, agents can insure themselves against all form of idiosyncratic risk.

Methods for Applied Macro Research 2: Models and Approximations

37

We assume that factors of production are immobile. Domestic consumers rent capital and labor to domestic rms which produce an homogeneous intermediate good using a constant returns to scale technology. Domestic markets for factors of production are competitive and intermediate rms maximize prots. Intermediate goods are sold to domestic and foreign nal good producing rms. The resource constraints are:
1 1 1 inty1 t + inty2t = 1t K1t N1t 2 inty1 t 2 + inty2 t

(2.13) (2.14)

1 2t K2 t N2t

1 are exports of goods from country 1 and inty 2 imports from country 2. where inty2 t 1t Final goods consumed by agents are an aggregate of the goods produced by intermediate rms of the two countries. Final goods are assembled with a constant returns to scale 1 1 )1a3 )+(1a )(inty 2 )1a3 )) 1a3 where a 1 and a (1a ) technology GDPit = (ai (intyit i 3 1 2 it measure the domestic content of domestic spending. The resource constraint in the nal goods market is GDPit = cit +invit . The two countries dier in the realizations of technology 2. shocks. We assume ln(it ) is an AR(1) with persistence | | < 1 and variance To map this setup into a Bellman equation assume that there is a social planner who attributes the weights W1 and W2 to the utilities of the agents of the two countries. Let the 1 ]1 P P t [c it (1Nit ) planners objective function be usp (c1t , c2t , N1t , N2t ) = 2 ; i=1 Wi E0 t=0 1 1 , inty 2 , c , N , K let y1t = [intyit , B , i = 1 , 2] , y = [ K , K , B ] and y = it it+1 1t+1 2t 1t 2t 1t 3t it it + + , y3 |y2 , [1t , 2t ]. Then Bellman equation is V(y2 , y3 ) = max{y1 } usp (c1 , c2 , N1 , N2 )+ E V(y2 y3 ) and the constraints are given by (2.12) -(2.14) and the law of motion of the shocks. Clearly, the value function has the same format as in (2.6). Since the functional form for utility is the same in both countries, the utility function of the social planner will also have the same functional form. A solution requires value function iterations, unless some form of approximation is taken. Some information about the properties of the model can be obtained by examining the rst order conditions and the properties of the nal good production function. In fact, we have: 1 2 + p2t intyit cit + invit = p1t intyit p2t T ott = p1t 1 2 nxt = inty2 t T ott inty1t

(2.15) (2.16) (2.17)

(2.15) implies that output of the nal good is allocated to the inputs according to their prices, GDP1t GDP1t p2t = 2 , p1t = inty 1 ; (2.16) gives an expression for the terms of trade and (2.17) inty1 t 1t for the trade balance. Exercise 2.11 Show that the demand functions for the two goods in country one are
1 1 a3 a3 1 a inty1 t = a1 (a1 + (1 a1 ) 3 T ott 1 a3 2 inty1 t = a2 T ott 1 a 1 3 a3 (a1 + (1 a1 ) a3 T ott 1 1

1a3 a3

) )

13 a 13 a
a

GDP1t GDP1t

(1a3 ) a3

38 i) Describe terms of trade relate to the variability of nal goods demands. (1a )(inty2t )a3 show that when the elasticity of substitution between ii) Noting that T ott = a (1inty 1 )1 a3
1

1 is high, any excess of demand in either of the two goods induces domestic and foreign good a 3 small changes in the terms of trade and large changes in the quantities used.

1t

Exercise 2.12 Consider the same two country model of example 2.11 but now assume that nancial markets are incomplete. That is, agents are forced to trade only a one-period bond which is assumed to be in zero net supply (i.e B1t + B2t = 0). How would you solve this problem? What does the assumption of incompleteness imply? Would it make a dierence if agents of country 1 have limited borrowing capabilities, e.g. B1t K1t ? Example 2.7 Interesting insights can be added to a basic RBC model when there are some agents which are not optimizers. For example, suppose that the economy is populated Pby tstandard RBC agents (their fraction in the total population is ) which maximize E0 t u(ct , ct1 , Nt ) subject to the budget constraint ct + invt + Bt+1 = wt Nt + rt Kt + B )B + prf + T where prf are rms prots, T are government transfers and B (1 + rt t t t t t t are real bonds. Suppose that capital accumulates at the rate Kt+1 = (1 )Kt + invt . The remaining 1 agents are myopic and consume all their income every period, that = wt Nt + TtRT . Also, we assume that these agents supply all their work time inis cRT t elastically at each t. Assume that rms decide the real wage and agents supply the hours must demanded by the rm at that wage. To insure that this is the case actual wage wt exceed the competitive wage wt . We assume that wt is increasing in Nt . Then aggregating RT and N A = N +(1 ), where the superscipt A indicates we have cA t t = ct +(1 )ct t aggregate variables. It is clear that in this model the responses of aggregate hours and of consumption to shocks are muted by the presence of myopic consumers. If (1 ) is large, the economy is relative insensitive to shocks (and income eects tend to dominate) while if (1 ) is small the economy resembles the RBC model analyzed so far. Exercise 2.13 (Kiyotaki and Moore) Consider a model with two goods, land La, which is in xed supply, and fruit which is non-storable, and a continuum of two types of agents: farmers of measure 1 and gatherers P of measure . Both types of agents have utilities of the form Et t j cj,t where cj,t is the consumption of fruit of agents type j and where f armers < gatherers . Let pL t be the price of land in terms of fruit and rt the rate of exchange of a unit of fruit today for tomorrow. Both agents have production technologies to produce fruit from land. Farmers use f (Lat )f armer = (b1 + b2 )Lat1 where b1 is the tradeable part and b2 the bruised one (nontradable) while gatherers use f (Lat )gatherer where fgatherer is a decreasing returns to scale function. The budget constraint for the two agents is pL t (Lajt Lajt1 ) + rt Bjt1 + cjt =

f (Lat ) + Bjt where c jt = cjt + b2 Lat1 for farmers and cjt = cjt for gatherers where Bjt are loans and pL t (Lajt Lajt1 ) the value of new land acquisitions. The farmers technology is idiosyncratic so that only farmer i has the skill to produce fruit from it. Note that if no labor is used, fruit output is zero. Farmers cant not precommit to work. The technology of

Methods for Applied Macro Research 2: Models and Approximations

39

gatherers does not require specic skills and all output is tradable. 1 and that for farmers to be able to borrow a i) Show that in equilibrium rt = r = gatherers collateral is required. Show that the maximum amount of borrowing allowed is Bt t+1 . r ii) Show that if there is no aggregate uncertainty, farmers borrow from gatherers up to their 1 maximum, invest in land and consume b2 Lat1 . That is, for farmers Lat = pL r 1 pL (b1 +
t t+1

pL Lat

L 1 L pL t )Lat1 rBt1 where pt r pt+1 is the user cost of land (the down payment needed to L purchase land) and Bt = r1 pL t+1 Lat . Argue that if pt increases, Lat will increase (provided L b1 + pt > rBt1 /Lat1 ) and Bt will also increase so that the higher is the land price, the higher is the net worth of farmers and the more they will borrow.

2.1.3

Monetary Models

The next set of models explicitly includes monetary factors. Finding a role for money in a general equilibrium model is dicult: with a full set of Arrow-Debreu claims, money is a redundant asset. Therefore, frictions of some sort need to be introduced for money to play some role. This means that the allocations produced by the competitive equilibrium are no longer optimal and that Bellman equation needs to be modied to include aggregate constraints. We focus attention on two popular specications: a competitive model with transactional frictions and a monopolistic competitive framework where either sticky prices or sticky wages or both are exogenously imposed. P Example 2.8 (Cooley and Hansen) The representative household maximizes E0 t t u(c1t , c2t , Nt ) where c1t is consumption of a cash good, c2t is consumption of a credit good and Nt Tt t+1 t is hours worked. The budget constraint is: c1t + c2t + invt + Mp wt Nt + rt kt + M pt + pt t where Tt = Mt+1 Mt and pt is the price level. There is a cash-in-advance constraint that forces households to buy c1t with cash. We require pt c1t Mt + Tt and assume that the monetary authority sets ln Mts+1 = ln Mts + ln Mtg where ln Mtg is an AR(1) process with mean , persistence M and variance 2 . Households choose sequences for the two consumption M M goods, for employment, for investment and real balances to satisfy the budget constraint. We assume that shocks are realized at the beginning of each t so agents know their values when they make decisions. The resource constraint is c1t + c2t + invt = f (Kt , Nt , t ) where 2 . Since money is dominated ln t is an AR(1) process with persistence and variance in expected rate of return (by physical investments) the cash-in-advance constraint will be binding and agents will hold just the exact amount of money needed to purchase c1t > 0, money (and prices) grow over time. To map this setup into a stationary When M pt Mt . Then the value function is: problem dene Mt = M s and pt = M s
t t+1

V (K, k, M , , M g ) = U (

1 Tt M + M (M )+ + + , wN + [ r + ( 1 )] k k c , N) 2 pt p p M
+

+ EV (K + , k+ , (M )+ , + , M g )

(2.18)

40
+M 1 Tt + pt and K represents where K + = (1 )K +Inv and k+ = (1 )k +inv with c1 = M Mp the aggregate capital stock. The problem is completed by the consistency conditions K + = h1 (K, , M g ), N = h2 (K, , M g ), p = h3 (K, , M g ). Not much can be done with this model without taking some approximation. However, we can show one of the main features of this setup: monetary disturbances produce expected ination but not liquidity eects. Suppose c2t = 0, t. Then an unexpected increase in Mtg makes agents substitute away from c1t (which is now more expensive) toward credit goods - leisure and investment - which are cheaper. Hence consumption and hours fall while investment increases. With a standard Cobb-Douglas production function output then declines. Also, since positive Mtg shocks increase ination, the nominal interest rate will increase, because both the real rate and expected ination have temporarily increased. Hence, a surprise increase in Mtg does not reduce interest rates (no liquidity eect is generated) nor induces output expansions.

There are several ways to correct for this major shortcoming. For example, introducing one period labor contracts (as we have done in exercise 2.9) does change the response of output to monetary shocks but not necessarily the one of interest rates. The next exercise provides an alternative solution by introducing a loan market, forcing consumers to take decisions before shocks are realized and rms to borrow to nance their wage bill. Exercise 2.14 (Working capital) Consider the same economy of example 2.8 with c2t = 0 t but assume that consumers deposit part of their money balances at the beginning of each t in banks. Assume that deposit decisions are taken before shocks occur and that rms face a working capital constraint, i.e. they have to pay for the factors of production before the receipts from the sale of the goods are received. Consumers maximize utility by choice of P (c (1Nt )1 )1 . consumption, labor, capital and deposits, i.e. max{ct ,Nt ,Kt+1 ,dept } E0 t t t 1 There are three constraints. First, goods must be purchased with money, i.e. ct pt Mt dept + wt Nt . Second, there is a budget constraint Mt+1 = prf1t + prf2t + rt pt Kt + Mt dept + wt Nt ct pt invt pt where prf1t (prf2t ) represent the share of rms (banks) prots and rt is the real return to capital. Third, capital accumulation is subject to an adjustment b Kt+1 ( Kt 1)2 Kt . Firms rent capital and labor and cost b 0, i.e. invt = Kt+1 (1 )Kt 2 borrow cash from the banks to pay for the wage bill. Their problem is max{Kt ,Nt } prf1t = 1 (pt t Kt Nt pt rt Kt (1 + it )wt Nt ) where it is the nominal interest rate. Banks take deposits and lend them together with new money to rms. Prots prf2t are distributed, pro-rata to the households. The monetary authority sets its instrument according to
g a2 a3 1 Mta0 = ia t Yt t Mt

(2.19)

where ai are parameters. For example, if a0 = 0, a1 = 1, the monetary authority uses the nominal interest rate to respond to output and ination and stands ready to provide money when the economy needs it. Let (ln t , ln Mtg ) be AR(1) with persistence , M and 2 , 2 . variances M

Methods for Applied Macro Research 2: Models and Approximations i) Set b = 0. Show that the labor demand and the labor supply are: UN,t = pt wt Et Uc,t+1 pt pt+1 wt it = fN,t pt

41

(2.20)

Argue that labor supply changes in anticipation of ination while labor demand is directly aected by interest rates changes so that output will be positively related to money shocks. U +1 U = Et1 it pc,t . How does this ii) Show that the optimal saving decision satises Et1 pc,t t t+1 compare with the saving decisions of a basic CIA model? t = 1+(1 iii) Show that the money demand can be written as pt GDP Mt /it ) where GDPt =
1 t t Kt Nt . Conclude that velocity pt GDP and the nominal rate are positively related and Mt that a liquidity eect is generated in response to monetary disturbances.

1 using GDPt = t Kt Nt where ln t is a technological disturbance and capital depreciates in one period. Let the quantity of money evolve according to ln Mts+1 = ln Mts + ln Mtg and assume at each t the government takes away Gt units of output. i) Assume Gt = G t. Write down the rst order conditions for the optimization problem t+1 ). of consumers and rms and nd the competitive equilibrium for (ct , Kt+1 , Nt , wt , rt , Mp t ii) Show that in equilibrium employment is independent of the shocks, that output and employment are uncorrelated and that real wages are perfectly correlated with output. iii) Show that monetary disturbances are neutral. Are they also superneutral (i.e. do changes in the growth rate of money have eects on real variables)? iv) Consider an economy with labor contracts where the nominal wage rate is xed one period in advance according to wt = Et1 Mt + ln() ln(m ( )/(1 )) Et1 ln Nt before all the shocks are realized. Show that shocks to technology and lagged monetary disturbances induce positive correlation between real wages and output while contemporaneous monetary disturbances produce negative correlations between these two variables. v) Now assume that Gt is stochastic and set ln Mtg = 0 t. What is the eect of government expenditure shocks on the correlation between real wages and output? Give some intuition for why adding labor contracts (Benassy) or government expenditure (Christiano and Eichenbaum) reduces the correlation between real wages and output found in ii).

P t Exercise 2.15 (Dunlop-Tarshis puzzle) Suppose agents maximize E0 t=0 [ln ct + m Mt+1 Mt+1 +Tt ln pt + N ln(1 Nt )] subject to ct + pt + Kt+1 = wt Nt + rt Kt + ( Mtp ). Let t pt+1 t+1 = pt be the ination rate. Firms rent capital from the households and produce

The nal type of model we consider adds nominal rigidities to a structure where monopolistic competitive rms produce intermediate goods which they sell to competitive nal goods producers. Example 2.9 (Sticky prices) Suppose consumers choose {ct , Nt , Kt+1 , Mt+1 } to maximize P c (1Nt )1 )1 Mt+1 1m + 11 ) subject to the budget constraint pt (ct + invt ) + E0 t t ( t 1 m ( pt )

42 Bt+1 + Mt+1 rt pt Kt + Mt + (1 + it )Bt + wt Nt + prft and the capital accumulation equation b Kt+1 ( Kt 1))2 Kt where b is an adjustment cost parameter. invt = Kt+1R (1 )Kt 2 Here prft = prfit di are prots obtained from owning intermediate rms. There are two types of rms: monopolistic competitive, intermediate good producing rms and perfectly competitive, nal good producing rms. Final goods rms take the continuum of intermediate goods and bundle it up for nal consumption. The production function for nal goods is 1 R1 1+ GDPt = ( 0 intyit p di)1+p where p is the elasticity of substitution between intermediate goods. Prot maximization implies a demand for each input i
intyit GDPt
1+p

it p = (p where pit pt ) R 1 1 is the price of intermediate good i and pt the price of the nal good, pt = ( 0 pit p di)p . Intermediate rms minimize costs and choose prices to maximize prots. Price decisions can not be taken every period: only (1 p ) of the rms are allowed to change prices at t. Their costs minimization problem is min{Kit ,Nit } (rt Kit + wt Nit ) subject to P U +j j 1 intyit = t Kit Nit and their prot maximization problem is max{pit+j } j pc,t p prfit+j t+j +1 where pc,t is the value of a unit of prot, prfit , to shareholders next period, subject to t+1 the demand function from nal goods rms. Here prft+j = (pit+j mcit+j )intyit+j where it +rit Kit are marginal costs. mcit = wit N intyit We assume that the monetary authority uses a rule of the form (2.19). Since only a fraction of the rms can change prices at each t, aggregate prices evolve according to:

p + (1 p ) pt p )p where p t is the common solution (all rms allowed to pt = (p pt1 change prices are identical) to the following optimality condition (dropping the i subscript)

0 = Et

X
j

j j p

Uc,t+j j pt [ mct+j ]intyt+j pt+j 1 + p

(2.21)

where is the steady state ination rate. Hence, rms choose prices so that the discounted marginal revenues equals the discounted marginal costs in expected terms. Note that if p 0 (2.21) reduces to the standard condition that the real wage equals the marginal product of labor. (2.21) is the basis for the so-called New Keynesian Phillips curve (see e.g. Woodford (2003, ch. 3)), an expression that relates current ination to expected future ination and to current marginal costs. To explicitly obtain such a relationship, (2.21) needs to be log-linearized around the steady state. When expressions obtained from exible price solution (i.e the situation where all rms can change prices every period, if they wish) are employed ination becomes a function of the output gap, i.e. the dierence between GDPt and its exible price counterpart. Exercise 2.16 i) Cast the sticky price model of example 2.9 into a Bellman equation formulation. Dene states, controls and the value function. ii) Assuming that rms of type i have weight Wi in the total, write down the rst order conditions of the problem and interpret them. Is it the case that the economy dichotomizes (real vs. nominal variables) as in the case of the cash-in-advance model? Why or why not?

Methods for Applied Macro Research 2: Models and Approximations

43

iii) Show that if all rms set prices one period in advance, the solution to (2.21) is pit = (1 + p )Et1
Et
Uc,t+j pt+j

pt

1+p p

intyit intyit

1+p Uc,t+j Et1 p pt p t+j

mcit . Conclude that if expectations do not change, rms set

the price of goods as a constant markup over marginal costs. Show that rms put higher weights on the marginal costs in those states where intyit is high. Exercise 2.17 In example 2.9 set Kt = 1, t, assume that prices are set for two periods so that half of the rms change prices at each t. Show that the optimal price setting satises Uc,t ct wt + Uc,t+1 ct+1 wt+1 t+1 p t = (1 + p )( ) 1 pt p Uc,t ct + Uc,t+1 ct+1 t+1
1+p p

(2.22)

+1 where p t is the optimal price, pt the aggregate price level, wt the wage rate and t = pt pt the ination rate. Show that if ln t = ln , ln Mtg = ln M g , t, markup over marginal cost.

Extensions of the model that allow also for sticky wages are straightforward. We ask the reader to study a model with both sticky prices and sticky wages in the next exercise. Exercise 2.18 (Sticky wages) Assume that households are monopolistic competitive in the labor market so that they can choose a wage at which to work. Suppose capital is in xed Mt+1 1m . Suppose supply and that the period utility function is u1 (ct ) + u2 (Nt ) + 11 m ( pt ) that households set nominal wages in a staggered way and that a fraction 1 w can do this every period. When the household is allowed to reset the wage, she maximizes the discounted sum of utilities subject to the budget constraint. i) Show that utility maximization leads to: Et
X j =0 j j w (

j wt U1,t+j + U2,t+j )Nt+j = 0 (1 + w )pt+j

(2.23)

where R is the discount factor, and w is the elasticity of substitution in the labor aggregator Nt = ( Nt (i)1/(1+w ) di)1+w , i [0, 1]. (Note: whenever the wage rate can not be changed wt+j = j wt where is the steady state ination). U2,t t ii) Show that if w = 0, (2.23) reduces to w pt = U1,t . iii) Calculate equilibrium output, real rate and real wage when prices and wages are exible.

Exercise 2.19 (Taylor contracts, Edge) Consider a sticky wage model where there is no capital ( = 1) so that labor demand is Nt = yt and real marginal costs are mct = wt = 1. The market clearing condition is yt = ct . Suppose utility is separable in consumption and t+1 = ct . Suppose ln Mts+1 = real balances. Setting m = 1, the money demand function is Mp t

44 > 0 and assume two period staggered labor ln Mts + ln Mtg , where ln M g is iid with mean M contracts. 1 1 t w t1 w w + wp ) ) where w t is the nominal i) Show that the real wage satises wt = (0.5( w pt ) t wage reset at t. 1 w t1 1 1 1 pt w w w /(2 ( t ) w ))w and that Nit = Nt (( t )/wt ) w = (( ) ii) Show that t = pt p p p t t 1 t1
1

t1 1 w if the wage was set at t and Nit = Nt (( w if the wage was set at t 1. pt1 )/(wt t ) iii) Show that if utility is linear in Nt , monetary shocks have no persistence.

vidual employment and Nt is aggregate employment and c , n , N are parameters. The g Mt1 +Mt Mt1 t consumers budget constraint is M = + f ( N , ) ct and assume that Mtg t pt pt pt 0. It is easy to show that the labor market equilibrium implies is iid with mean M f i Uc,t+1 UN +1 Uc,t+1 Mt = fN (Nt , pt ) and that the demand for money is Et [ M,t ] = Et [ t+1 ] Uc t+1 t+1 where 1 + it is the gross nominal rate on a one-period bond, t the ination rate and f . These two standard conditions are somewhat special in this model. In fact, fM = M/p individual labor supply is downward sloped because it is shifted by changes in the economywide labor supply. Decentralizing in a competitive equilibrium and log linearizing the labor market condition we have c ln ct + n nt (N + n ) ln Nt = ln wt ln pt . Since agents are all equal, the aggregate labor supply will be downward sloping function of the real wage and equal to c ln ct N ln Nt = ln wt ln pt . Hence, a small shift in labor demand increases output and consumption (which is equal to output in equilibrium) makes real wages fall and employment increase. That is, a demand shock can generate procyclical consumption and employment paths. Also, since money enters the production function, an increase in money could shift labor demand as in the working capital model. However, contrary to that case, labor market eects can be large because of the slope of the aggregate labor supply curve, even when money is relative unimportant as productive factor. Example 2.11 (Obstfeld and Rogo ) Consider a structure like the one of example 2.9 where prices are chosen one period in advance, there are two countries and international nancial markets are incomplete, in the sense that only a real bond, denominated in the composite consumption good, is traded. In this economy the domestic nominal interest rate t+1 B B satises i + i1t = Et p1 p1t (i + rt ) where rt is the real rate on internationally traded bonds p1t t+1 and uncovered interest parity holds, i.e. 1 + i1t = Et ner nert (1 + i2t ) where nert = p2t and p1t is the consumption based money price index in country 1. Using the Euler equations for each country we have the international risk sharing condit+1 t+1 ( c2 ] = 0. Hence, while consumption growth needs not be a random tion Et [( c1 c1t ) c2t ) walk, the dierence in adjusted consumption growth is a martingale dierence. 1 iit m it+1 = m cit ( 1+ , i = 1, 2 which The money demand equation for each country is Mp iit ) it c2t implies, using uncovered interest parity and log-linearizing, that M1t M2t 1 (
m

Example 2.10 (Benhabib and Farmer) Consider an economy where output is produced 1 t with labor and real balances i.e. GDPt = (a1 Nt + a2 ( M pt ) ) . Suppose the utility of the (1n ) P 1c nt 1 agents is of the form, u(ct , nt , Nt ) = E0 t t ( c 1c 1n (N n ) ) where nt is indiNt

Methods for Applied Macro Research 2: Models and Approximations

45

d ert where . indicates deviations from the steady state. Hence dierential in c 1t ) + (1 )m n money supplies or consumption levels will make the nominal exchange rate jump to a new equilibrium. Note also that if M1t > M2t the wealth of agents in country 1 will increase and the terms of trade improve. If wealth increases, employment and output fall while prices and the nominal exchange rate increase. Since prices are xed for one period, the terms of trade will temporarily deteriorate. Therefore, there are two opposite eects on the terms of trade, one dominating in the short run and one in the long run.

Variations or renements of the price (wage) technology exist in the literature (see Rotemberg (1984) or Dotsey et al. (1999)). Since these renements are tangential to the scope of this chapter, we invite the interested reader to consult the original sources for details and extensions.

2.2

Approximation methods

Finding a solution to the Bellman equation and calculating the optimal decision rule is, in general, complicated. Bellman equation is a functional equation and a xed point needs to be found in the space of functions. This, as we have seen, requires an iteration process which involves the computation of expectations and the maximization of the value function until convergence is achieved. We have also seen in example 2.2 and exercise 2.4 that when the utility function is quadratic (logarithmic) and time separable and the constraints are linear in the states and the controls, a decision rule for the variables of interest has linear (log-linear) form. In these two cases, given a representation for the candidate solution, the method of undetermined coecients can be used to nd the unknown parameters of the value function and of the policy rules. Quadratic utility functions are not very appealing, however, as they imply implausible behavior for consumption and asset returns. Log-utility functions are easy to manipulate but they are also restrictive regarding the attitude of agents towards risk. Based on a large body of empirical research, the macroeconomic literature currently tends to select a general power specication for preferences. In this case, one has either to iterate on the Bellman equation or to resort to approximations to nd a solution. We have also mentioned that methods for solving general nonlinear expectational equations as those emerging from the rst order conditions of a stochastic Lagrangian multiplier problem are complicated. Therefore, also in this case, approximations need to be employed. This section considers four approximation methods extensively used in the literature. The rst approximates the value function quadratically around a steady state. In the second, the approximation is calculated forcing the states and the exogenous variables to take only a nite number of possible values. This method can be applied to both the value function and to the rst order conditions. The other two approaches approximate the optimal conditions of the problem. In one case a log-linear approximation around the steady state is calculated. In the other, the expectational equations of the model are approximated by nonlinear functions and a solution is obtained by calculating a xed point

46 for the parameters of these nonlinear functions.

2.2.1

Quadratic approximations

Quadratic approximations are easy to compute but work under two restrictive conditions. The rst is that there exists a point - typically, the steady state - around which the approximation can be taken. Although this requirement may appear innocuous, it should be noted that some models do not possess a steady or a stationary state. Furthermore, for problems with inequality constraints (e.g. borrowing or irreversibility constraints) such an approximation is incorrect since the non-stochastic steady state ignores them. The second is that local dynamics are well approximated by linear dierence equations. Consequently, quadratic approximations should not be used when problems involve large perturbations away from the approximation point (e.g. policy shifts), nonlinear dynamic paths or transitional issues. Quadratic approximations can be applied to both value function and Lagrangian multiplier problems. We will discuss applications to value function problems only since the extension to the second type of problems is straightforward. Let Bellman equation be:
+ + , y3 |y2 , y3 ) V(y2 , y3 ) = max u(y1 , y2 , y3 ) + E V(y2 {y1 }

(2.24)

where y2 is a m2 1 vector of the states, y3 is a m3 1 vector of exogenous variables, y1 is + = h(y3 , y1 , y2 ) a m1 1 vector of the controls. Suppose that the constrains are given by y2 + + and that the law of motion of the exogeneous variables is y3 = 3 y3 + where h is a continuous function and a vector of martingale dierence disturbances. Substituting the constraints into (2.24) we have
+ + + ) + E V(y2 , y3 |y2 , y3 ) V(y2 , y3 ) = max u(y2 , y3 , y2
+ {y2 }

(2.25)

+ + Let u (y2 , y3 , y2 ) be the quadratic approximation of u(y2 , y3 , y2 ) around ( y2 , y 3 , y 2 ). If 0 V is quadratic, then (2.25) maps quadratic functions into quadratic functions and the limit value of V (y2 , y3 ) will also be quadratic. Hence, under some regularity conditions, + linear. the solution to the functional equation is quadratic and the decision rules for y2 In practice, when the solution to (2.25) is known to be unique, an approximate solution is found either iterating on (2.25) starting from a quadratic V0 or by guessing a quadratic form for the solution V(y2 , y3 ) = V0 + V1 [y2 , y3 ] + [y2 , y3 ]V2 [y2 , y3 ]0 , and matching coecients.

Exercise 2.20 Consider the basic RBC model with no habit persistence in consumption and utility given by u(ct , ct1 , Nt ) = 1t , no government sector and no taxes and consider the recursive formulation provided by the Bellman equation. i) Compute the steady states and a quadratic approximation to the utility function. iii) Compute the value function assuming that the initial V 0 is quadratic and calculate the optimal decision rule for capital, labor and consumption (you are supposed to do this by hand but if it becomes overwhelming because of algebra, you can adapt one of the computer programs which comes with this book to undertake the iterations).
c1

Methods for Applied Macro Research 2: Models and Approximations

47

It is important to stress that certainty equivalence is required when computing the solution to a quadratic approximation problems. This principle allows us to eliminate the expectation operator from (2.25) and reinsert it in front of all future unknown variables once a solution is found. This operation is possible because the covariance matrix of the shocks does not enter the decision rule. That is, certainty equivalence implies that we can set the covariance matrix of the shocks to zero and replace random variables with their unconditional mean. In general, quadratic approximations are used in situations where the social planner decisions generate competitive equilibrium allocations. When this is not the case the method requires some adaptation to take into account the fact that aggregate variables are distinct from individual ones (see e.g. Hansen and Sargent (1998) or Cooley (1995, chapter 2)) but the same mechanics apply. Approximate quadratic value function problems t into the class of optimal linear regulator problems. Therefore, an approximate solution to the functional equation (2.25) can also be found using methods developed in this literature. One example of an optimal linear regulator problem was encountered in P exercise 2.6. The general setup for such a problem is 0 Q y +2[y , y ]0 Q0 y ] the following: we want to maximize Et t t [y2t , y3t ]0 Q2 [y2t , y3t ]+ y1 2t 3t t 1 1t 3 1 0 y + Q0 y y + Q . The Bellman with respect to y1t , y20 given, subject to y2t+1 = Q0 2 t 1 t 3 t +1 4 5 6 equation for this problem is
+ + 0 V(y2 , y3 ) = max([y2 , y3 ]0 Q2 [y2 , y3 ] + y1 Q1 y1 + 2[y2 , y3 ]0 Q0 3 y1 ]) + E V(y2 , y3 |y2 , y3 ) (2.26) {y1 }

Hansen and Sargent (1998) show that, starting from arbitrary initial conditions, iterations j on (2.26) yield at the j-th step, the quadratic value function Vj = y2 0 Vj 2 y2 + V0 where
+1 j 0 j 0 1 j 0 0 0 Vj = Q2 + Q4 Vj 2 2 Q4 ( Q4 V2 Q5 + Q3 )(Q1 + Q5 V2 Q5 ) ( Q5 V2 Q4 + Q3 ) (2.27) +1 j 0 and Vj = Vj 0 0 + tr (V2 Q6 Q6 ). (2.27) is the so-called matrix Riccati dierence equation which depends on the parameters of the model (i.e. the matrices Qi ), but it does not involve Vj 0 . (2.27) can be used to nd the limit value V2 which, in turns, allows us to compute the limit of the value function. The decision rule which attains the maximum at each iteration j j 0 1 j 0 j j j is y1 t = (Q1 + Q5 V2 Q5 ) ( Q5 V2 Q4 + Q3 )y2t and can be calculated given V2 and the parameters of the model. j While it is common to iterate on (2.27) to nd the limits of Vj 0 , V2 , the reader should be aware that faster algorithms which can produce this limit without iterations are available (see e.g. Hansen, Sargent and McGrattan (1996)).

Exercise 2.21 Consider the two country model analyzed in example 2.6. i) Take a quadratic approximation to the objective function of the social planner around the steady state and map the problem into a linear regulator problem. ii) Use the matrix Riccati equation to nd a solution to the maximization problem. When quadratic approximations are taken, and a solution is known to exist and to be unique, the method of undetermined coecients could also be used. Although the approach

48 is easy conceptually, it is mechanically cumbersome, even for small problems. If we knew the functional form of the value function (and/or of the decision rule) we could posit a specic parametric representation and use the rst order conditions to solve for the unknowns, as we have done in exercise 2.4. We highlight few steps of the approach in the next example and let the reader ll in the details. Example 2.12 Consider a model where the representative agent chooses sequences for P Mt+1 1 t+1 t+1 ) to maximize E0 t t (c ) where ct is consumption, Mt+1 = Mp (ct , Mp t + pt t t

t+1 are real balances and let t be the ination rate. The budget constraint is ct + Mp = t Mt y y (1 T )wt + pt where T is an income tax, wt is an exogenous income process. We assume that wt and Mt are stochastic. The government budget constraint is Gt = T y wt + Mt+1 Mt which, together with the consumer budget constraint implies the resource conpt

straint, ct + Gt = wt . Let the states of the problem be y2t = (wt , Mt , t ) and the shocks be + + , y3 |y2 , y3 )]. y3t = (wt , Mtg ). Bellman equation is V(y2 , y3 ) = max{c,M } [u(c, M ) + E V(y2 ss ss ss ss Let (c , M , w , ) be the steady state value of consumption, real balances, income and ination. For ss = 1, wss = 1, consumption and real balances are given by css = (1 T y ) 1 ) y 1 ] . A quadratic approximation to the utility function and (M )ss = [ (11 ((1 T )) ss ss 1 , B = is given by B0 + B1 xt + x0 1 t B2 xt where xt = (y2t , Mt+1 ), B0 = (c ) + ((M ) )
) ) [(css )1 (1 T y ); (css ; (css )1 ( ((M ); (css )1 + (1 )((M )ss ) ] and the ss)2 matrix B2 is (1T y ) y )( M ss ) y) (1 T y )2 (( 1 T ( 1 T ss ss 2 ( ) (1T y ) (M )ss ss )1 )] ( )[ + ( c ss ss ss ss 2 ss 2 ( ) ( ) ss ss ss ss ( M ) ( M ) ( M ) ( M ) 2 y (1 T )( ss 2 ) ( ss 2 )[ ss + ss ] ( ss 2 )( + ss ( ss ) (ss )2 )() c c ( ) ( ) ( ) (M )ss 2 ((M )ss )1 y (1 T ) ss ( (ss )2 ) + (css )2
ss 1 ss

where = ( 1)(css )2 . One could guess then a quadratic form for the value function and solve for Qi as we have done before. Alternatively one could guess a policy function (in deviation from steady states) of the form Mt+1 = Q0 + Q1 Mt + Q2 wt + Q3 t and solve for Qi using the rst order conditions of the approximate problem. Exercise 2.22 Find the rst order conditions of the approximate problem of example 2.12. Show the form of Qj , j = 0, 1, 2, 3 using the certainty equivalence principle. When the number of states is large analytic calculation of rst order and second order derivatives of the return function may take quite some time. As an alternative numerical derivatives, which are much faster to calculate and only require the solution of the model u at a pivotal point, could be used. Hence, in example 2.12, one could use e.g. c =
[(1T y )wss +] [(1T y )wss ] 2

for small.

c 1n P c1 Nt t Exercise 2.23 (Ramsey) Suppose households maximize E0 t t ( t 1c + 1n ) where t is a preference shock and c , n are parameters. Suppose the resource constraint is ct + Gt =

Methods for Applied Macro Research 2: Models and Approximations

49

P y b GDPt = t Nt and that the consumer budget constraint is E0 t t p0 t [(1 Tt )GDPt + s0t at 0 and p0 ct ] = 0 where sb t is 0t is a stream of coupon payments promised by the government P t p0 [(G + an Arrow Debreu price at time zero. The government budget constraint is E 0 t t P t 0 bt y ) T GDP ] = 0 . Given a process for G and the present value E p s , a feasible sb t t 0 t 0t 0t t t tax process must satisfy the government budget constraint. Assume that (t , t , sb 0t , Gt ) are random variables with AR(1) representation. Agents choose sequences for consumption and hours and the government selects the tax process preferred by the representative household. The government commits at time 0 to follow the optimal tax system, once and for all. i) Take a quadratic approximation to the problem, calculate the rst order conditions of the household problem and show how to calculate the Arrow-Debreu price p0 t. ii) Show the allocations for ct , Nt and the optimal tax policy Tty . Is it true that the optimal tax rate implies tax smoothing (random walk taxes), regardless of the process for Gt ? Example 2.13 Consider the setup of exercise 2.7 where the utility function is u(ct , Gt , Nt ) = ln(ct + G Gt ) + N (1 Nt ) and where Gt is an AR(1) process with persistence G and vari2 and it is nanced with lump sum taxes. The resource constraint is c + K ance G t t+1 + Gt = 1 Kt Nt t + (1 )Kt where ln t is an AR(1) disturbance with persistence and vari2 . Setting ance G = 0.7, = 0.64, = 0.025, = 0.99, N = 2.8, we have that ss (K/GDP ) = 10.25, (c/GDP )ss = 0.745, (inv/GDP )ss = 0.225, (G/GDP )ss = 0.03 and N ss = 0.235. Approximating quadratically the utility function and linearly the constraint we can use the matrix Riccati equation to nd a solution. Convergence was achieved at iteration 243 and the increment in the value function at the last iteration was 9.41e0 06. The 2 [y2 , y3 ] where y2 = (K, G), y3 = and value function is proportional to [y2 , y3 ]V 1.76e 09 3.08e 07 7.38e 09 V2 = 1.54e 08 0.081 9.38e 08 . The decision rule for y1 = (c, N )0 is 2.14e 06 3.75e 04 8.98e 06 9.06e 10 0.70 2.87e 09 y2t . y1t = 9.32e 10 1.56e 07 2.95e 09

2.2.2

Discretization

As an alternative to quadratic approximations , one could solve the value function problem by discretizing the state space and the space over which the exogenous processes take values. This is the method popularized, for example, by Merha and Prescott (1985). The idea is that the states are forced to lie in the set Y2 = {y21 , . . . y2n1 } and the exogenous processes in the set Y3 = {y31 , . . . y3n2 }. Then the space of possible (y2t , y3t ) combination has n1 n2 points. For simplicity, assume that the process for the exogenous variables is a rst order Markov with transition P (y3t+1 = y3j 0 |y3t = y3j ) = pj 0 j . The value function associated with each pair of states and exogenous processes is V(y2i , y3j ) which is of dimension n1 n2 . Because of the Markov structure of the shocks, iterations on the Bellman equation are easy to calculate. In fact, we have reduced a innite dimensional problem into the problem of mapping of n1 n2 matrices into n1 n2 matrices. The value function can be written as

50 P 2 (T V)(y2i , y3j ) = maxn u(y1 , y2i , y3j )+ n l=1 Vn,l pl,j where y1n is such that h(y1n , y2i , y3j ) = y2n , n = 1, . . . n1 n2 . An illustration of such an approach is given in the next example. Example 2.14 Consider a RBC model where a random stream of government expenditure is nanced by distorting income taxes, labor supply is inelastic and production uses P c1 only capital. The social planner maximizes E0 t t 1t subject to ct + Kt+1 (1

1 2 and )Kt + Gt = (1 T y )Kt , where Gt is an AR(1) with persistence g , variance g (, , T y , , ) are parameters. The social planner chooses {ct , Kt+1 }, given Gt and Kt . The controls are therefore consumption and investment; the state is the stock of capital and goverment expenditure is the exogenous shock. Bellman equation is V(K, G) = y 1 )K GK + ]1 + E (V(K + , G+ |K, G)). Given K0 , dene the opmax{K + } [(1T )K +(1 1 )K GK ] erator T by (T V)(K, G) = max{K + } [(1T )K +(1 + E (V(K + , G+ |K, G). 1 Suppose that the capital stock and government expenditure can take only two values, and let the transition for Gt be pj 0 j . Then the discretization algorithm works as follows:
y 1 + 1

Algorithm 2.2 1) Choose values for (, , , T y , ) and specify the elements of pj,j 0 . 2) Choose an initial 2 2 matrix V(K, G), e.g. V0 = 0. 3) For i, i0 , j, j 0 = 1, 2, calculate (T Vi,i0 )(K, G) = max[ [Vi,i0 pi,i0 + Vi,j 0 pi,j 0 )];
1 (1T y )Ki +(1 )Ki Ki Gi0 )1 ]+ 1 1 y 1 (1T )Ki +(1 )Ki Kj Gi0 ) [ ] + [Vj,i0 pj,i0 + Vj,j 0 pj,j 0 )]. 1

4) Iterate on 3) until e.g. maxi,i0 | T l Vi,i0 T l1 Vi,i0 | , small, l = 2, 3, . . .. Suppose T y = 0.1, = 0.1; = 0.9, = 2, = 0.66, choose G1 = 1.1; G2 = 0.9, K1 = 5.3, K2 = 6.4, p11 = 0.8; p22 = 0.7 , V0 = 0. Then (T V)11 = max1,2 {( = max1,2 (14.38, 0.85) = 14.38. Repeating for the other entries, we have T V = 24.92 3.91 71.63 31.54 2 l , and liml T V = . Implicitly the solution deT V= 21.53 1.10 56.27 1.10 nes the decision rule; for example, from (T V)11 we have that Kt = K1 .

1 1 (1T y )K1 +(1)K1 K1 G1 )1 (1T y )K1 +(1)K1 K2 G1 )1 ); ( )} 1 1

14.38 1.03 12.60 0.81

Clearly, the quality of the approximation depends on the ness of the grid. Therefore, it is a good idea to start from a course grid and after convergence is achieved check whether a ner grid produces dierent results. The discretization approach is well suited for problems of modest dimension (i.e. when the number of state variables and of exogenous processes is small) since constructing a grid which systematically and eectively covers the space of states and shocks is dicult when

Methods for Applied Macro Research 2: Models and Approximations

51

the combined number of states and exogenous processes exceeds 3. For example, when we consider three dimensions and 100 grid points, 1,000,000 evaluations are required in each step. Nevertheless, even with this large number of evaluations, it is easy to leave large portions of the space unexplored. Exercise 2.24 (Search) Suppose an agent has the choice of accepting or rejecting a wage oer. If she has worked at t 1, the oer is wt = b0 + b1 wt1 + et , where et is a shock; if she was searching at t 1 the oer is drawn from some stationary distribution. Having observed wt agents decide whether to work or not (i.e. whether Nt = 0 or Nt = 1). Agents if Nt = 0 where c measures unemployment cant save so that ct = wt if Nt = 1 and ct = c c1 compensations. Agents maximize discounted utility where u(c) = 1t and is a parameter. i) Write down the maximization problem and the rst order conditions. ii) Dene states and controls and write down the Bellman equation for each of the two states of the problem. Suppose et = 0, b0 = 0, b1 = 1; = 0.96 and that wt U(0, 1). Calculate the optimal value function and describe the decision rules. iii) Assume that the agent has now also the option of retiring so that xt = 0 or xt = 1. if Nt = 0, xt = 1 Suppose xt = xt1 if xt1 = 0 and that ct = wt if Nt = 1, xt = 1; ct = c c if Nt = 0, xt = 0 where c is the retirement pay. Write down Bellman equation and ct = and calculate the optimal decision rules. iv) Suppose that the agent has now the option to migrate. For each location i = 1, 2 the i = b + b wl i wage is determined by wt 0 1 t1 + et if the agent has worked at t 1 in location i U(0, i) otherwise. Consumption is c = w if i = i % if i, and by wt t t t t1 and ct = c it 6= it1 where % is a migration cost. Let % = 0.1. Write down the Bellman equation and calculate the optimal decision rules. Exercise 2.25 (Lucas tree model) Consider an economy where innitely lived agents have a random stream of perishable endowments sdt and decide how much to consume and save, where savings can take the form of either stocks or bonds and let u(ct , ct1 , Nt ) = ln ct i) Write down the maximization problem and the rst order conditions. Write down the Bellman equation specifying the states and the controls. ii) Assume process can take only two values sd1 = 6, sd2 = 1 with that the endowment 0.7 0.3 transition . Find the 2 1 vector of value functions, one for each state. 0.2 0.8 iii) Find the policy function for consumption, stock and bond holdings and the pricing functions for stocks and bonds. A discretization approach can also be used to solve the rst order conditions of the model. Hence, it is useful for problems where the value function may not exist. Example 2.15 For general preferences, the Euler equation of exercise 2.25 is
s ps t (sdt )Uc,t = E [Uc,t+1 (pt+1 (sdt+1 ) + sdt+1 )]

(2.28)

sdt = [sdh , sdL ], where we have made explicit the dependence of ps t on sdt . If we assume that P n 1 s 2 = sd and let U p ( sd ) U ; U = use the equilibrium condition ct t i sdi i i i0 =1 sdi Usdi pii0 ,

52 P 1 = (1 P )1 U 2 where P is the (2.28) can be written as Ui1 = Ui2 + i0 pii0 Ui1 0 or U matrix with typical element {pij }. Therefore, given a functional form for the utility it is 2 P Ui 0 where immediate to have that share prices satisfy ps (sdi ) = i0 (I + P + 2 P 2 + . . .)ii0 Usd i 0 the sum is over the (i, i ) elements of the matrix.

2.2.3

Log linear Approximations

Log linearizations have been extensively used in recent years following the work of Blanchard and Khan (1980), King, Plosser and Rebelo (1987) and Campbell (1994). Uhlig (1999) has systematized the methodology and provided software useful to solve a variety of problems. King and Watson (1998) and Klein (2001) provided algorithms for singular systems. Log linear approximations are similar, in spirit, to quadratic approximations and the solutions are computed using similar methodologies. The former however may work better when the problem displays some mild non-linearities. The major dierence between the two approaches is that quadratic approximations are typically performed on the return function while log-linear approximations are calculated using the rst order conditions of the problem. Therefore, the latters are useful in situations where, because of distortions, the competitive equilibrium is suboptimal. The basic principles of log linearization are simple. We need a point around which the log-linearization takes place. This could be the steady state or, as in models with frictions, the frictionless solution. Let y = (y1 , y2 , y3 ). The rst order conditions of the problem can be divided into two blocks, the rst containing expectational equations and the second non-expectational equations: 1 = Et [h(yt+1 , yt )] 1 = f (yt , yt1 ) 0 Et [ht+1 yt+1 + ht yt ] 0 = ft yt + ft1 yt1 (2.29) (2.30)

where f (0, 0) = 1; h(0, 0) = 1. Taking rst order Taylor expansions around ( y, y ) = (0, 0) (2.31) (2.32)

f ln h where fj = ln yj and hj = yj . (2.31) and (2.32) form a system of linear expectational equations. Although log linearizing standard models is easy, Uhlig (1999) suggests a set of approximations to calculate (2.31)-(2.32) directly without dierentiations of the h and f eyt where yt is small and using functions. The tricks involve replacing any variable yt with y the following three rules (here a0 is a constant and b1t , b2t small numbers). eb1t +a0 b2t 1 + b1t + a0 b2t . b1t b2t 0. Et [a0 eb1t+1 ] Et [a0 b1t+1 ].

Example 2.16 To illustrate how to use thse rule consider the budget constraint ct + gt + ect + g egt + inveinvt = GDP egdpt and use rule i) to get c (1 + ct ) + invt = GDPt . Use c

Methods for Applied Macro Research 2: Models and Approximations

53

(1 + invt ) GDP (1 + gdpt ) = 0. Then using c + g + inv = GDP we get g (1 + gt ) + k gt + invinvt GDP gdpt = 0. c ct + g Exercise 2.26 Suppose yt and yt+1 are conditionally jointly log-normally distributed ho (yt+1 , yt ))]} where h = log (h) moschedastic processes. Replace (2.29) with 0 = log{Et [exp(h t+1 yt+1 + ht yt ] show that the log linear approximation is 0 Using log h(0, 0) 0.5vart [h Et [ht+1 yt+1 + ht yt )]. What is the dierence between this approximation and the one in (2.31)? Exercise 2.27 Consider an economy where the private production function is GDPt = 1 Nt t Kt 1 /(1 ) ( Nt )2 / where ( Kt ) and ( Nt ) are the average endowment of capital and t ( PK ) opt P opt P opt P opt P Nt 1N 1 t . hours in the economy. Suppose the utility function is t t E (ln( Pc opt ) 1N ( P opt ) Assume that (ln t , ln P opt ) are AR(1) processes with persistence equal tp and 1. i) Show that the rst order conditions of the problem are ct Nt N ( ) P opt P opt P opt ct GDPt P opt P opt+1 GDPt+1 = Et [(1 ) + (1 ) ] Ct+1 Kt+1 = (2.33) (2.34)

ii) Find expressions for the log-linearized production function, the labor market equilibrium, the Euler equation and the budget constraint. iii) Write the log linearized expectational equation in terms of an Euler equation error. Show under what conditions sunspot equilibria obtain (Hint: nd conditions under which there are more unstable roots than state variables). There are several economic models which do not t the setup of (2.29)-(2.30). For example, Rotemberg and Woodford (1997) describe a model where consumption at time t depends on the expectations of variables dated at t + 2 and on. This model can be accommodated in the setup of (2.29)-(2.30) using dummy variables, as the next exercise, shows. In general, restructuring of the timing convention of the variables, or enlarging the vector of states, suces to t these problems into the general framework used here.
= Exercise 2.28 Suppose that (2.29) is 1 = Et [h(yt+2 , yt )]. Create a dummy variable yt yt+1 . Transform the system so that only variables dated at t and t + 1 appear in the expectational equation.

Exercise 2.29 Consider a cash-in-advance model where production requires capital and labor (assume a Cobb-Douglas form), agents maximize utility over consumption and leisure (assume preferences as in exercise 2.14) and assume that there are adjustment costs to invt )invt where f (.) is a twice continuous and dierentiable investment of the form f ( inv t1 0 00 function, f (1) = 0, f > 0, f < 0. Calculate the rst order conditions of the model. Show how to map this problem into (2.29)-(2.30).

54 Exercise 2.30 Consider a sticky price model with Calvo pricing scheme (as in example 2.9) where agents display habit in consumption. In particular, assume a utility function of the form (ct ct1 ) (1 Nt )1 . Derive the rst order conditions of the model and map them into (2.29)-(2.30). Example 2.17 Log-linearizing around the steady state the equilibrium conditions of the model studied in exercise 2.13, and assuming an unexpected change in the productivity of r c L farmers technology (represented by ) lasting one period we have (1 + 1 t % )Lat = + r1 p c t+ = La c t+ 1 for 1 where % is the elasticity of the supply for = 0 and (1 + 1 )La
%

of land with respect to the user costs in the steady state and p L t =

. indicates percentage deviations from the steady state. Solving these two expressions we 1 r c have p L t = % and Lat = 1+ 1 (1 + (r1)% ). Three interesting conclusions follow from these
% %

1 r1 c t, La % r% 1 r(1+ %)

where

solutions. First, if % = 0, temporary shocks have permanent eects on farmers land and on 1 r its price. Second, since 1+ 1 (1 + (r 1)% ) > 1, the eect on land ownership is larger than the

r1 L c shock. Finally note that the static multipliers ( pL L t ) |pt+1 = r% < p t and (Lat ) |Lat+1 = c t . This is occurs because aects the net worth of farmers: a positive reduces the < La value of the obligations and implies larger use of capital by the farmers, therefore magnifying the eect of the shock on land ownership.

Exercise 2.31 Show that the log-linearized rst order conditions of the sticky price model of example 2.9 when Kt = 1, t are 0 = wt + ( iss N ss ) i = ( 1 ( 1 ))( c c ) ( 1 )( 1 )( N N ) t+1 t +1 t +1 t t +1 t 1 + iss 1 N ss Mt+1 (1 ) 1 N ss (1 )(1 ) 1 = ct + Nt it ss pt m 1N m m (1 + iss ) (1 p )(1 p ) Et t+1 = t ( + N ) mct (2.35) p N ss Nt ct 1 N ss

where mct are real marginal costs, p is the probability of not changing the prices, wt is the real wage, is the risk aversion parameter and N ) the labor supply elasticity and the superscript ss refers to steady state. As with quadratic approximations, the solution of the system of equations (2.31)-(2.32) can be obtained in two ways: using the method of the undetermined coecients or nding the saddle-point solution (Vaughans method). The method of undetermined coecients is analogous to the one described in exercise 2.20. Vaughans method works with the state space representation of the system. Both methods require the computation of eigenvalues and eigenvectors. For a discussion of the methods and a comparison with second order dierence equation methods, the reader should consult e.g. the chapter of Uhlig in Marimon

Methods for Applied Macro Research 2: Models and Approximations

55

and Scott (1999) or Klein (2001). Here we briey describe the building blocks of the procedure and highlight important steps with some examples. Let y1t be of dimension m1 1, y2t of dimension m2 1, and y3t of dimension m3 1 and suppose the log linearized rst order conditions, the budget constraint and the law of motion of the exogenous variables be written as: (2.36) 0 = Q1 y2t + Q2 y2t1 + Q3 y1t + Q4 y3t 0 = Et [Q5 y2t+1 + Q6 y2t + Q7 y2t1 + Q8 y1t+1 + Q9 y1t + Q10 y3t+1 + Q11 y3t ] (2.37) 0 = y3t+1 y3t t (2.38) where Q3 is a m4 m1 , m4 m1 matrix and of rank m1 , and has only stable eigenvalues (in (2.29)-(2.30) we had set Q7 = 0. In the method of undetermined coecients one posits a (linear) solution to (2.36)-(2.38) and backs out the unknown parameters. Assume that a solution is given by: y2t = A22 y2t1 + A23 y3t y1t = A12 y2t1 + A13 y3t Uhlig (1999) shows that:
+ a) Letting Z1 = Q8 Q+ 3 Q2 Q6 + Q9 Q3 Q1 , A22 satises the (matrix) quadratic equations: 0 0 = Q0 3 Q1 A22 + Q3 Q2

(2.39) (2.40)

0 =

2 (Q5 Q8 Q+ 3 Q1 )A22

Z1 A22 Q9 Q+ 3 Q2

+ Q7

(2.41) (2.42)

The equilibrium is stable if all eigenvalues of A22 are less than one in absolute value. b) A12 is given by A12 = Q+ 3 (Q1 A22 + Q2 ). c) Given 13 and A23 satisfy: Z2 = (Q5 A22 + Q8 A12 ) and Z3 = Q10 + Q11 , A Im3 Q1 vec(A23 ) vec(Q4 ) Im3 Q3 = 0 Q5 + Im3 (Z2 + Q6 ) 0 Q8 + Im3 Q9 vec(A13 ) vec(Z3 )

where vec(.) is columnwise vectorization; QG 3 is a pseudo inverse of Q3 and satises G = QG and Q QG Q = Q ; Q0 is an (m m ) m matrix whose rows Q Q QG 3 3 3 3 4 1 4 3 3 3 3 3 are a basis for the space of Q0 3 and Im3 is the identity matrix of dimension m3 .

Exercise 2.32 Suppose m4 = m1 (there are as many expectational equations as endogenous state variables, a case typically encountered in practice). Show the form of the Aij matrices 1 in this case (Hint: note that here QG 3 = Q3 ). Suppose in addition that m2 = m3 = 1. Show the form of Aij . Example 2.18 Consider a RBC model with an intermediate monopolistic competitive sector. Let the prots in rm i be prfit = mkt intyt where mkt = (pit mcit ) is the markup. If

56 the utility function is of the form u(ct , ct1 , Nt ) = 1t + N (1 Nt ), the dynamics depend on the markup only via the steady state. For this model the log linearized conditions are 0 = (inv/GDP )ss invt (c/GDP )ss ct + gdpt
ss c1

0 = (inv/K ) invt kt+1 + (1 )kt 0 = (1 )kt gdpt + nt + t 0 = ct + gdpt Nt rss rt 0 = kt + gdpt ss mk (1 )(GDP/K )ss 0 = Et [ct+1 + rt+1 + ct ] t+1 = t + 1t+1

(2.43) (2.44) (2.45) (2.46) (2.47) (2.48) (2.49)

where (inv/GDP )ss and (c/GDP )ss are the steady state investment and consumption to state markup. Letting output ratios, rss is the steady state real rate and mkss steady 0 (1 )K ss ; Q3 = y1t = (ct , gdpt , Nt , rt , invt ), y2t = kt , y3t = t , we have Q2 = 1 0 ss D C ss GDP ss 0 0 invss 0 0 0 0 invss 0 ; Q1 = [0, K ss , 0, 0, 0]0 ; Q4 = [0, 0, 1, 0, 0]0 ; Q8 = 1 0 0 1 1 0 0 ss ss 0 r 0 0 D [, 0, 0, 1, 0]; Q9 = [, 0, 0, 0, 0], = [ ]; Q5 = Q6 = Q7 = Q10 = Q11 = [0] where Dss = mkss (1 )(GDP/K )ss . It is important to stress that the method of undetermined coecients properly works only when the state space is chosen to be of minimal size, that is, no redundant state variables are included. If this is not the case, A22 may have zero eigenvalues and this will produce bubble solutions. Computationally, the major diculty is to nd a solution to the matrix equation (2.42). The toolkit of Uhlig (1999) recasts it into a generalized eigenvalue-eigenvector problem. Klein (2000) calculates a solution using the generalized Shur decomposition. When applied to some of the problems described in this chapter, the two approaches yield similar solutions. Exercise 2.33 Consider a model where the representative agent maximizes E0
(
Mt+1 1 m ) pt

where c and m are parameters, subject to the resource constraint ct + Kt+1 + 1m Mt+1 1 t Nt + (1 )Kt + M pt = t Kt pt where ln t is an AR(1) process with persistence and t+1 standard error . Let Mt+1 = Mp be real balances, t the ination rate, rt the rental rate t

1c t ct t 1c +

Methods for Applied Macro Research 2: Models and Approximations

57

0 and standard of capital and assume ln Mts+1 = ln Mts + ln Mtg where ln Mtg has mean M error M . i) Verify that the rst order conditions are
Nt + (1 ) rt = (1 )t Kt ct+1 c ) rt+1 ] 1 = Et [ ( ct ct+1 c c (Mt+1 )m c = 1 + Et [ ( ) t+1 ] t ct

(2.50)

1 Mt t Kt+1 (1 )Kt + (GDPt Gt ) + M Nt and assume pt + Tt , pt ct where GDPt = t Kt g s that the monetary authority sets ln Mt+1 = ln Mt + ait where a is a parameter and it is Mt the nominal interest rate. The government budget constraint is Gt + Mt+1 = Tt . Let pt g [ln Gt , ln t , ln Mt ] be a vector of random disturbances. t+1 . Derive the optimization conditions and the i) Assume a binding CIA constraint, ct = Mp t equation determining the nominal interest rate. ii) Compute a log-linear approximation around the steady states of the rst order conditions and of the budget constraint, of the production function, of the CIA constraint, of the equilibrium pricing equation for nominal bonds and the of government budget constraint. t+1 , it ) rst while iii) Show that the system is recursive and can be solved for (Nt , Kt , Mp t
t+1 , it ), where (GDPt , ct , t , Tt ) can be solved in a second stage as a function of (Nt , Kt , Mp t t is the Lagrangian multiplier on the private budget constraint. t iv) Write down the system of dierence equations for (Nt , Kt , M pt , it ). Guess a linear solution g (in deviation from steady states) in Kt and [ln Gt , ln t , ln Mt ] and nd the coecients. v) Assume prices are set one period in advance as a function of the states and of past shocks, i.e. pt = a0 + a1 Kt + a21 ln Gt1 + a22 ln t1 + a23 ln Mtg1 . What is the state vector in this t case? What is the most likely guess for (Nt , Kt , M pt , it )? Use the method of undetermined coecients to nd a solution.

Cast these equations into the form of equations (2.36)-(2.38). ii) Log linearize (2.50), the resource constraint, and the law of motion of the shocks. iii) Guess that a solution for [Kt+1 , ct , rt , Mt+1 ] is linear in (Kt , Mt , t , Mtg ). Determine the coecients of the linear relationship. P Mt+1 t Exercise 2.34 Suppose agents maximize E0 t=0 u(ct , 1 Nt ) subject to ct + pt +

The next example shows the form of the log-linearized solution of a version of the sticky price-sticky wage model described in exercise 2.18. Example 2.19 Assume that capital is xed so that the only variable factor in production P c (1Nt )1 )1 Mt+1 1m m + 1 ). Set N ss = is labor and the utility function is E0 t t ( t 1 m ( pt ) c c )ss = 0.8 where ( GDP )ss is the share of 0.33, = 0.66, ss = 1.005, = 0.99, ( GDP ss ss consumption in GDP, N is hours worked and is gross ination in the steady states, is exponent of labor in the production function, is the discount factor. These choices

58 imply, for example, that in steady-state the gross real interest rate is 1.01, output is 0.46, real balances 0.37 and the real (fully exible) wage 0.88. We select the degree of price and wage rigidity to be the same and set p = w = 0.75. Given the quarterly frequency of the model, this choice implies that on average rms (consumers) change their price (wage) every three quarters. Also, we choose the elasticity of money demand m = 7. In the monetary policy rule we set a2 = 1.0; a1 = 0.5; a3 = 0.1, a0 = 0. Finally, t and Mtg are AR(1) processes with persistence 0.95. The decision rules for real wage, output, interest rates, real balances and ation, in terms of lagged real wages and the two shocks are in w bt 0.5823 0.0005 0.0012 y # " 0.0008 bt bt 0.5571 0.2756 b 0.9595 0.0416 bt1 + . it = g 0.0128 d w d 0.0427 0.1351 Mt Mt 0.1386 0.7812 0.0025 0.1050 bt Two features of this approximate solution are worth commenting upon. First, there is very little feedback from the state to the endogenous variables, except for output. This implies that the propagation properties of the model are limited. Second, monetary disturbances have little contemporaneous impact on all variables except interest rates and real balances. These two observations imply that monetary disturbances have negligible real eects. This is conrmed by standard statistics. For example, technology shocks explain about 99 percent of the variance of output at the four years horizon and monetary shocks the rest. This model also misses the sign of few important contemporaneous correlations. For example, using detrended US data the correlation between output and ination is 0.35. For the model, the correlation is -0.89. P Exercise 2.35 (Delivery lag) Suppose the representative agent maximizes E0 t t [ln ct 1 Nt and assume one period delivery lag, i.e. Kt+1 = l Nt ] subject to ct + it t Kt 1 1 (1 )Kt + invt1 . The Euler equation for this model is Et [c t+1 (1 )GDPt+1 Kt+1 ] + 1 1 1 (1 )c t ct1 = 0. Log linearize the system and nd a solution assuming that the states are Kt and c t = ct1 . Vaughans method, popularized by Blanchard and Kahn (1980) and King, Plosser and Rebelo (1987), takes a slightly dierent approach. First, using the state space representation for the (log)-linearized version of the model, it eliminates the expectation operator either assuming certainty equivalence or substituting the expectations with the actual values of the variables plus an expectational error. Second, it uses the law of motion of the exogenous variables, the linearized solution for the state variables and the costate (the Lagrangian multiplier) to create a system of rst order dierence equations (if the model delivers higher order dynamics, the dummy variable trick described in exercise 2.28 can be used to get the system in the required form). Third, it computes an eigenvalue-eigenvector decomposition on the matrix governing the dynamics of the system and divides the roots into explosive and stable ones. Then, it solves forward those roots which are unstable and uses the restrictions implied by the stability condition to derive the law of motion for the Lagrangian multiplier

Methods for Applied Macro Research 2: Models and Approximations

59

(and the expectational error, if needed). Once these are found, the solutions are plugged into the equations for the states and a solution for the controls is computed. Suppose that the log-linearized system is t = AEt t+1 where t = [y1t , y2t , y3t , y4t ], y2t and y1t are, as usual, the states and the controls, y4t are the costates and y3t are the shocks and partition t = [1t , 2t ]. Let A = PVP 1 be the eigenvalue-eigenvector decomposition of A. Since the matrix A is symptletic, the eigenvalues come in reciprocal pairs when 1 V1 is a matrix with eigenvalues greater than one distinct. Let V = diag(V 1 , V1 ), where 1 1 P P 1 11 12 in modulus and P 1 = 1 1 . Multiplying both sides by A , using certainty P21 P22 equivalence and iterating forward we have " # j 1 1 1t+j V1 0 1t + P12 2t P11 =P (2.51) 1 1 j 2t+j P21 1t + P22 2t 0 V1 We want to solve (2.51) under the condition that 2t+j goes to zero as j starting from some 20 . Since the components of V1 exceed unity, this is possible only by making 1 1 1 ) P21 1t Q1t so the terms multiplying V1 equal to zero. This implies 2t = (P22 that (2.51) is # " j 1 1 QP11 V1 (P11 1t + P12 2t ) Q1t+j (2.52) = j 1 1 2t+j (P11 1t + P12 2t ) P21 V1

1 which also implies Q = P21 P11 . Note that, for quadratic problem, Q is the same as the limit of the Riccati equation (2.27).

Example 2.20 The basic RBC model with labor-leisure choice, no habit, Gt = Tt = T y = 1 Nt and utility function u(ct , ct1 , Nt ) = ln ct + 0, production function f (Kt , Nt , t ) = t Kt 1 N (1 Nt ) when log linearized, delivers the representation t = A 0 A1 Et t+1 where t = t ] (since there is a one-to-one relationship between ct , Nt and t we can solve t t, N t , [ ct , K out of the system) where . indicates percentage deviations from steady states and 1 1 1 1 1 0 0 0 A0 = N ss N ss ; ( c )ss (1 )( N ss ) + ( 1 ) ( ) ( ) ss ss ss K K K K 0 0 0 0 0 0 0 ss ss ss 1 (1 )( N ss ) (1 )( N ss ) (1 )( N ss ) K K K . A1 = 0 1 0 0 0 0 0 1
1 1 where P is a matrix whose columns are the eigenvectors of A1 A Let A 1 0 A1 = PVP 0 and V contains, on the diagonal, the eigenvalues. Then 1 P 1 t t+1 t = V Et t+1 V Et P it = vi Et i,t+

(2.53)

Since V is diagonal, there are four independent equations which can be solved forward, i.e. i = 1, . . . , 4 (2.54)

60 Since one of the conditions describes the law of motion of the technology shocks, one of the 1 eigenvalues is (the inverse of the persistence of technology shocks). One other condition describes the intratemporal eciency condition (see equation (2.11)): since this is a static relationship the eigenvalue corresponding to this equation is zero. The other two conditions, the Euler equation for capital accumulation (equation (2.10)) and the resource constraint (equation (2.5)) produce two eigenvalues: one above and one below one. The stable solution is associated with the vi > 1 since it for vi < 1. Hence for (2.54) to hold for each t in the stable case, it must be that it = 0 for all i < 1. Assuming = 0.99, = 0.64, = 0.025, n = 3, the resulting steady states are 0.79; K ss = 10.9; N ss = , GDP ss = 1.06 and css = 0.29 1.062 0 0 0 2.18 0.048 0.048 24.26 0 1.05 0 0 0 0 23.01 Et 0 . t = 0 0 0.93 0 2.50 1.36 0.056 1.10 t+1 0 0 0 0 2.62 0.94 0.94 2.62 1 , the last one the intertemporal condition. The remaining two The second row has v2 = rows generate a saddle path. Setting the third and the forth rows to zero (v3 , v4 < 1) we have ct = 0.54Nt +0.02Kt +0.44t and Nt = 2.78ct + Kt +2.78t . The law of motion of the capital stock can be read o the rst equation: Kt+1 = 0.07ct + 1.01Kt + 0.06Nt + 0.10t . Exercise 2.36 (log linearization of a CIA model) P Consider a model, where agents maximize a separable utility function of the form: E0 t=0 u(ct , Nt ) by choices of consumption, hours and nominal money balances subject to the following three constraints: gdpt = t Nt = Gt + ct ct = Mt /pt + M g) Mt+1 = (Mt pt ct ) + pt (yt Gt ) + Mt (M t

(2.55)

The rst constraint states that output is produced with labor; it is random because of a technology shock (t ) and it is used for either private or public consumption (the latter being a random variable). The second constraint states that to purchase consumption real balances is a constant and M g is a are needed. The third describes the accumulation of money: M t mean zero random variable. i) Derive and log-linearize the rst order condition of the problem. State what are the exogenous variables, the endogenous variables and what are the states of the problem. ii) Solve the linear system assuming that the growth rate of the exogenous variables (t , Gt , Mtg ) is an AR(1) process with common parameter . Calculate the equilibrium expressions for ination, output growth and real balances. iii) Suppose you want to price the term structure of nominal bonds. Such bonds cost 1 unit of money at time t and give 1 + it+ units of money at time t + , = 1, 2, . . .. Write the equilibrium conditions to price these bonds. Calculate the log-linear expression of the slope for the term structure between a bond with maturity and a one period bond. iv) Calculate the equilibrium pricing formula and the rate of return for stocks which costs s ps t units of consumption at t and pays dividends pt sdt which can be used for consumption

Methods for Applied Macro Research 2: Models and Approximations

61

only at t + 1 (Hint: the value of dividends at t + 1 is ps t sdt /pt+1 ). Calculate a log-linear expression for the equity premium (the dierence between the nominal return on stocks and the nominal return on a 1 period bond). v) Simulate the responses of the slope of term structure and of the equity premium to a unitary shock in the technology (t ), in government expenditure (Gt ) and in money growth (Mtg ). Is the pattern of responses economically sensible? Exercise 2.37 Consider the sticky price model with the capital utilization setup analyzed in exercise 2.10 but without adjustment costs to capital. Log linearize the model and compute output responses to monetary shocks (still assume the monetary rule 2.19). How does the specication compare in terms of persistence and amplitude of real responses to the standard one, without capacity utilization, but with capital adjustment costs? Exercise 2.38 Consider two versions of the basic RBC model described in example 2.20 with capital adjustment costs: one with and one without habit persistence in consumption. Compare the responses of output, employment, consumption and investment to technology shocks in the two cases. Intuitively explain the dierences you found.

2.2.4

Parametrizing expectations

The method of parametrizing expectations has been suggested by Marcet (1989) and developed by Marcet and Lorenzoni (1999)). With this approach the approximation is globally valid (as opposed to valid only around a particular point). Therefore, with such a solution we can undertake experiments which are, e.g., far away from the steady state or unusual from the historical point of view. The approach has two advantages: rst, it can be used when inequality constraints are present. Second, it has a built-in mechanism that allows us to check whether a candidate solution satises the rst order condition of the problem. Therefore, we can implicitly examine the accuracy of the approximation. The essence of the method is simple. First, one approximates the expectational equations of the problem using a vector of functions }, i.e. Et [h(y2t+1 , y2t , y3t+1 , y3t )] }(, y2t , y3t ) where y2t and y3t are known at t and is a vector of (nuisance) parameters. Polynomial, trigonometric, logistic or other simple functions which are known to have good approximation properties can be used. In the second step, one estimates by minimizing a measure of distance between Et [h(y2t+1 (), y2t (), y3t+1 , y3t )] and }(, y2t (), y3t ) where y2t () are simulated states from the approximate solution. Let be the minimizer. Dene Q() = argmin{ } |Et [h(y2t+1 (), y2t (), y3t+1 , y3t )] }( , y2t , y3t )|q (2.56)

some q > 0. The method then looks for a xed point of Q()i.e. a such that Q( ) = 0. Example 2.21 Consider a basic RBC model with inelastic labor supply, utility given by u(ct ) =
c1 t 1

where is a parameter, budget constraint given by ct + Kt+1 + Gt = (1

62
1 + (1 )Kt + Tt and let (ln t , ln Gt ) be AR processes with persistence ( , g ) T y )t Kt and unit variance. The expectational (Euler) equation is y = Et [c c t t+1 ((1 T )t+1 (1 )Kt+1 + (1 ))]

(2.57)

where is the rate of time preferences. We wish to approximate the expression on the right hand side of (2.57) with a function }(Kt , t , Gt , ), where is a set of parameters. Then the method requires ve steps: Algorithm 2.3 1) Select (, T y , , , g , , ). Generate (t , Gt ), t = 1, . . . T , choose an initial 0 . 2) Given a functional form for } calculate ct (0 ) from (2.57) with }(0 , kt , t , Gt ), in y 0 place of Et [c t+1 ((1 T )t+1 (1 )Kt+1 + (1 ))] and Kt+1 ( ) from the resource constraint. Do this for every t - this produces a time series for ct (0 ) and Kt+1 (0 ). 3) Run a nonlinear regression using simulated ct (0 ), Kt+1 (0 ) of }(, Kt (0 ), t , Gt ) on ct+1 (0 ) ((1 T y )t+1 (1 )Kt+1 (0 ) + (1 ). Call the resulting nonlinear estimator 0 and with this 0 construct Q(0, ). 4) Set 1 = (1 %)0 + %Q(0 ) where % (0, 1]. 5) Repeat steps 2)-4) until |L L1 | , small or until Q(L ) 0. When convergence is achieved }( , Kt , t , Gt ) is the required approximating function. Since the method does not specify how to choose }, it is typical to start with a simple function (a rst order polynomial, a Chebishev function which includes only a linear term, or a trigonometric function) and then check the robustness of the solution using more complex functions (e.g a higher order polynomial). For the model of this example, setting = 2, T y = 0.15, = 0.1, g = = 0.95, = 0.66, = 0.99, q = 2 and choosing } = exp(ln 1 + 2 ln Kt + 3 ln t + 4 ln Gt ), 100 iterations of the above algorithm led to the following optimal approximating values 1 = 0.0780, 2 = 0.0008, 3 = 0.0306, 4 = 0.007 and with these values Q() = 0.000008. Example 2.22 We show how to apply the method when inequality constraints are present. Consider a small open economy which nances current account decits by issuing one period < 0. The so that Bt B nominal bonds. Assume that there is a borrowing constraint B Euler equation for debt accumulation is
c t Et [ct+1 (1 + rt ) t+1 )] = 0

(2.58)

where rt is the exogenous world real interest rate, t the Lagrangian multiplier on the bor ) = 0. To nd a solution rowing constraint and the Kuhn-Tucker conditions is t (Bt B ) = 0 and calculate ct and Bt assuming t = 0. use 0 = ct }(, rt , t , ct ) and t (Bt B If Bt > B set Bt = B , nd from the rst equation and ct from the budget constraint. Do this for every t. Then nd 0 , generate 1 and repeat until convergence. In essence, t is treated as an additional variable, to be solved for in the model.

Methods for Applied Macro Research 2: Models and Approximations

63

Exercise 2.39 (CIA with taxes) a model where agents maximize a separable utilP Consider t ity function of the form: E0 t=0 (c ln(c1t ) + (1 c ) ln(c2t ) N (1 Nt )) by choices of consumption of cash and credit goods, leisure, nominal money balances and investments, 0 < < 1. Suppose that the household is endowed with K0 units of capital and one unit of time. The household receives income from capital and labor which is used to nance consumption purchases, investments and holdings of money and government bonds. c1t is the cash-good and needs to be purchased with money, c2t is the credit good. Output is produced with capital and labor by a single competitive rm with constant returns to scale technology and 1 is the share of capital. In addition, the government nances a stochastic ow of expenditure by issuing currency, taxing labor income with a marginal tax rate Tty and issuing nominal bonds, which pay an interest rate it . Assume that money supply evolves according to ln Mts+1 = ln Mts + ln Mtg . Suppose agents start at time t with holdings of money Mt and bonds Bt . Assume that all the uncertainty is resolved at the beginning of each t. i) Write down the optimization problem mentioning the states and the constraints and calculate the rst order conditions (Hint: you need to make the economy stationary). ii) Solve the model parametrizing the expectations and using a rst order polynomial. iii) Describe the eects of an iid shock in Tty on real variables, prices and interest rates, when Bt adjusts to satisfy the government budget constraint. Would your answer change if you keep Bt xed and let instead Gt change to satisfy the government constraint? Exercise 2.40 Suppose in the basic RBC model that u(ct , ct1 , Nt ) =
1 Kt t (ct ct1 )1 , 1

Tt =

= 0 and the production function is GDPt = Provide a parametrized expectation algorithm to solve this model (Hint: there are two state variables in the Euler equation). Example 2.23 (Marcet and Den Haan) As mentioned, the method of parametrizing expectations has a built-in mechanism to check the accuracy of the approximation. In fact, the simulated time series must satisfy the Euler equation whenever the approximation is appropriate. As we will describe in more details in chapter 5, this implies that vt = [g (y1t+1 (), y1t (), y2t+1 (), y2t (), y3t+1 , y3t )] %], where % is a constant and g is determined by the model (e.g. it is the Euler equation (2.57)), must be such that E (vt h(zt )) = 0 where zt is any variable in the information set at time t; h is a q 1 vector of continuous differential functions and yit (), i = 1, 2 are simulated time series (generated after a solution is obtained with the optimal conditions, when T is large, P ). Under regularity 1 P 0W ( 1 2 ( ), where is equal to the dimension v h ( z )) v h ( z )) S = T (T t t t t t t T of the Euler conditions times the dimension of h and W is a weighting matrix. For the example 2.21, the rst order approximation is suciently accurate since S has a p-value of 0.36. One should also be aware of two drawbacks of this approach. First, the iterations dened by algorithm 2.3 may led nowhere since the xed point problem does not dene a contraction operator. In other words, there is no insurance that the distances between the actual and approximating function will get smaller as the number of iterations grows. Second, the

Ty

64 method relies on the suciency of the Euler equations. That is, if the utility function is not strictly concave, the solution that the algorithm delivers may not be appropriate.

2.2.5

A Comparison of methods

There exists some literature comparing various approximation approaches. For example, the special issue of the Journal of Business and Economic Statistics of July 1991 provides a thorough comparison of how various approaches perform in approximating the decision rules of a particular version of the one sector growth model for which analytic solutions are available. Some extra evidence is in Fernandez-Villaverde and Rubio-Ramirez (2003). In general, little is known about the properties of various methods in specic applications. Experience suggests that even for models with AR(1) dynamics (i.e. models without habit, adjustment costs of investment, etc.), simulated series may display somewhat dierent dynamics depending on the approximation methods used. For more complicated models no evidence is available. Therefore caution should be employed in interpreting the dynamics obtained approximating such models with any of the methods outlined in this chapter. Exercise 2.41 (Growth with corruption) Consider a representative agent who maximizes P c1 E0 t t 1t by choices of consumption ct , capital Kt+1 and bribes brt subject to ct + Kt+1 = (1 Tty )Nt wt + rt Kt brt + (1 )Kt y Tty = Tte (1 a ln brt ) + T0

(2.59) (2.60)

where wt is nominal wage income, Tty is the income tax rate, Tte is an exogenously given y is the part of the tax rate which is unchanged by bribes, and (, a, ) are tax rate, T0 G) = parameters. The technology is owned by the rm and it is given by f (Kt , Nt , t , Kt 1 G ) where 0, K is the capital stock and N the hours. Government capital t Kt Nt (Kt t t G G = (1 )K G + N w T y . The resource constraint for the economy is Kt is given by Kt t t t t +1 G + br = f (K , N , ) + (1 )(K + K G ) and ( , T e ) are independent ct + Kt+1 + Kt t t t t t t t t +1 2 , 2 ). AR(1) processes, with persistence ( , e ) and variances ( e i) Dene a competitive equilibrium for the economy and calculate the rst order conditions. 2 = 2 = 1. Write ii) Assume = 2, a = 0.03, = 0.96, = 0.10, e = = 0.95 and set e down the Lagrangian, take a quadratic approximation of utility and a linear approximation of the constraints and nd the decision rule for the variables of interest. iii) Assume that (t , Tte ) and the capital stock can take only two values (say, high and low). Solve the model discretizing the state space (Hint: use the fact that shocks are independent and the values of the AR parameter to construct the transition matrix for the shocks). iv) Take a log linear approximation of the rst order conditions of the problem. Solve the model using either the method of undetermined coecients or Vaugans approach. v) Use the parametrized expectations method with a rst order power function to nd a global solution. vi) Compare the decision rules for consumption, investment and bribes across methods. Are they similar? Are they dierent? In what?

Methods for Applied Macro Research 2: Models and Approximations

1 1 (c ) t Nt 1

65

Exercise 2.42 Consider an economy where preferences are described by u(ct , ct1 , Nt ) = which accumulates capital according to Kt+1 = (1 )Kt + invt where is the depreciation rate. Assume that the production function is Cobb-Douglas in hours (Nt ), k N L capital (Kt ), and land (Lat ) and of the form GDPt = t Kt Nt La t . Suppose individual agents have the ability to borrow and to trade land so that their budget constraint is B L ct + Kt+1 + Bt+1 + pL t Lat+1 GDPt + (1 )Kt + (1 + rt )Bt + pt Lat and suppose that L there is a borrowing constraint of the form pt Lat Bt+1 0 where pi t is the price of land in terms of consumption goods ss (i) Show that in the steady state the borrowing constraint is binding if ( GDP K ) + (1 ) > B ). Give conditions which insure that the constraint is always binding. (1 + rt (ii) Describe the dynamics of the variables of the model following a technology shock in three cases: (a) borrowing constraint never binds, (b) borrowing constraint always binds, (c) borrowing constraints binds at some t. (Hint: be careful with the last case and use an approximation method which allows the comparison across cases). (iii) Is it true that the presence of (collateralized) borrowing constraints amplies and stretches out the transmission of technology shocks?

66