The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Working Paper Series Labor Supply Shifts and Economic Fluctuations WP 03-07 Yongsung Chang Federal Reserve Bank of Richmond Frank Schorfheide University of Pennsylvania This paper can be downloaded without charge from: http://www.richmondfed.org/publications/ Labor Supply Shifts and Economic Fluctuations * Federal Reserve Bank of Richmond Working Paper 03-07 July 2003 Yongsung Chang Research Department, Federal Reserve Bank of Richmond and Frank Schorfheide Department of Economics, University of Pennsylvania Abstract: We propose a new VAR identification scheme that distinguishes shifts of and movements along the labor demand schedule to identify labor-supply shocks. According to our VAR analysis of post-war U.S. data, labor-supply shifts account for about 30 percent of the variation in hours and about 15 percent of the output fluctuations at business cycle frequencies. To assess the role of labor-supply shifts in a more structural framework, estimates from a dynamic general equilibrium model with stochastic variation in home production technology are compared to those from the VAR. JEL Classification: E32, C52, J22 Key Words: Labor-Supply Shifts, VAR, Home Production, Bayesian Econometrics * Marco Airaudo provided excellent research assistance. We wish to thank Larry Christiano, Frank Diebold, Martin Eichenbaum, John Geweke, Michael Kiley, Richard Rogerson, and Chris Sims for helpful comments and suggestions. Thanks also to seminar participants at the NBER Summer Institute, University of Pennsylvania, Princeton, Rochester, ISBA Regional Meeting, USC, Econometric Society Meetings, the Federal Reserve Bank of Cleveland, and the Board of Governors. The second author gratefully acknowledges financial support from the University of Pennsylvania Research Foundation. The GAUSS programs to implement the empirical analysis are available at http://www.econ.upenn.edu\~schorf. The views expressed herein are those of the authors and do not necessarily reflect those of the Federal Reserve Bank of Richmond or the Federal Reserve System. Chang: Yongsung.Chang@rich.frb.org; Schorfheide: schorf@econ.upenn.edu. 2 1 Introduction A leading question in macroeconomics is the identification of forces that cause the cyclical allocation of time. Modern dynamic general equilibrium analysis empha sizes shifts in labor demand due to technological change. Empirical studies on the decomposition of sources of business cycles by Shapiro and Watson (1988) and Hall (1997) have called for an attention to labor-supply shifts. This paper examines the importance of labor-supply shifts as a source of economic fluctuations. First, we develop and apply a new identification procedure for vector autore gressions (VAR). It decomposes the fluctuations of aggregate hours and output into movements along the labor demand schedule and shifts of the schedule itself. The former is interpreted broadly as response to a labor-supply shock. Our VAR iden tification is based on the notion that an increase in hours due to a labor-supply shock leads to a fall in labor productivity, as the production capacity is fixed in the short-run and the economy operates along the decreasing marginal-product-of-labor schedule. We place a prior distribution on the slope of the short-run labor-demand curve and on the reduced-form VAR parameters and conduct Bayesian inference. Second, we impose additional restrictions by estimating a fully-specified dynamic stochastic general equilibrium model (DSGE). The DSGE model potentially yields a more precise estimate of the relative importance of labor supply shifts. We consider an aggregate home production model (Benhabib, Rogerson, and Wright, 1991, and Greenwood and Hercowitz, 1991)) in which labor-supply shifts are caused by the stochastic variation in home production technology. 3 The main empirical findings can be summarized as follows. Based on the VAR variance decomposition, temporary shifts in labor supply are an important source of hours fluctuations. They account for about 30 percent of the cyclical variation of hours worked. The DSGE model attributes more than 50 percent of the variation of hours to temporary labor-supply shifts. This larger estimate, however, may partly be due to misspecified over-identifying restrictions as the time series fit of the DSGE model is significantly worse than the VAR fit. According to both VAR and DSGE models, labor-supply shocks are less important for aggregate output as they explain only about 15 percent of its variation at business cycle frequencies. Our estimates of the contribution of labor-supply shifts to economic fluctuaΩ tions are somewhat smaller than those reported by Shapiro and Watson (1988) and Hall (1997). Shapiro and Watson (1988) identify labor-supply shocks through the stochastic trend in hours worked. While the empirical evidence on the stationarity of hours worked is not conclusive, we assume hours are stationary, which is conΩ sistent with a large class of dynamic equilibrium models. In Hall (1997), Parkin (1988), and Baxter and King (1991), labor-supply shocks (or preference shocks) are identified as deviations from the optimality condition associated with the labor supply of competitive households. However, these residuals also reflect the extent to which a representative-agent model is inconsistent with aggregate hours and wages, potentially leading to a bigger estimate of labor-supply shifts. While we exploit the labor-market equilibrium as Hall, our identification scheme does not rely on a speΩ cific form of households’ preferences. Instead, it is based on the firms’ production technology through the slope of the marginal-product-of-labor schedule. 4 The paper is organized as follows. Section 2 develops the VAR identification scheme. The home production model is presented in Section 3. Section 4 discusses the econometric estimation and inference. The empirical findings are summarized in Section 5 and Section 6 provides a conclusion. 2 A VAR Model of Labor Market Fluctuations In this section we describe our identification scheme for the VAR using a labor demand and supply framework. The labor-market fluctuations are viewed as a se ries of equilibria generated by competitive households and firms whose tastes and technologies are perturbed by stochastic disturbances. To identify the sources of fluctuations we will fit a VAR and a DSGE model (specifically, an aggregate home production model) to three macroeconomic time series: hours worked, labor pro ductivity, and expenditure on consumer durable goods. Expenditure on consumer durables serves as a proxy for the household’s permanent income. In the context of the home production model, it represents the investment in home capital. 2.1 Identifying Assumptions During the past four decades, labor productivity, spending on consumer durables, and aggregate output exhibited a pronounced trend, whereas aggregate hours and the consumption share did not show an apparent trend. Based on this observation, many dynamic macroeconomic models have been designed to evolve along a balanced growth path. A common stochastic trend in output, consumption, investment, cap- 5 ital, and labor productivity is induced by a labor augmenting technology and hours worked are stationary around this path. The VAR innovations are decomposed into three orthogonal shocks, denoted by �a,t , �b,t , and �z,t . Assumption 1 The shock �z,t has a permanent effect on labor productivity and consumer durables whereas it has no effect on hours in the long run. The shocks �a,t and �b,t have transitory effects on hours, labor productivity, and consumer durables.� The shock �z,t induces a common stochastic trend, and it is subsequently interpreted as permanent technology shock in a DSGE model. At time t, the competitive firms’ inverse labor demand can be written as: Wt = M P Lt = ϕD (Lm,t |Km,t , St ), (1) where Wt represents the real wage rate, M P Lt the marginal product of labor, Lm,t hours employed, Km,t capital stock, and St state of technology at time t.1 Similarly, the inverse labor supply of a competitive household is: Wt = ϕS (Lm,t |Ω(St , Tt )), (2) where Ωt represents endogenous variables that influence the labor supply (e.g., real interest rate, consumption, and wealth) and Tt the exogenous stochastic component of tastes. 1 Subscripts m and h, respectively, denote the market and home sector, consistent with the DSGE model introduced in Section 3. 6 Assumption 2 The shock �b,t has a contemporaneous impact on Tt , but not on St and Km,t . Thus, upon impact this shock shifts the labor-supply curve, but not the labor-demand curve (marginal-product-of-labor schedule). � We will interpret the shock �b,t broadly as a labor-supply shock, such as an unanticipated shift of the preference for leisure or the productivity of non-market activities. The capital stock Km,t is inherited from the previous period and therefore not affected by the current period’s labor-supply shock. Given that produc tion capacity is fixed in the short-run, firms operate along the downward sloping marginal-product-of-labor schedule. The responses of the marginal product of labor and hours worked (both in logs) to a labor-supply shock �b,t have to satisfy the following relationship ∂ ln M P Lt = ∂�b,t � 1 ∂ϕD · ϕD ∂ ln Lm,t � · ∂ ln Lm,t . ∂�b,t (3) The factor in parentheses is the slope of the inverse labor demand function. For example, under a Cobb-Douglas production technology with labor share α one ob tains: ∂ ln Lm,t ∂ ln Pt = (α − 1) , ∂�b,t ∂�b,t (4) where Pt = M P Lt is the labor productivity. Roughly speaking, conditional on the slope of the labor demand function, it is possible to identify �b,t through its joint effect on hours and productivity. The slope of the labor demand schedule itself, however, is imposed and not identifiable. Finally, �a,t , namely a temporary labor demand shock, is identified by assuming that the three structural shocks are 7 orthogonal.2 . 2.2 VAR Specification Define the vector of stationary variables Δyt = [Δ ln Pt , Δ ln Ih,t , ln Lm,t ]� . Moreover, let �t = [�z,t , �a,t , �b,t ]� . The VAR can be expressed in vector error correction form as Δyt = Φ0 + Φvec yt−1 + p−1 � Φi Δyt−i + ut , ut ∼ iid N (0, Σu ). (5) i=1 The reduced form disturbances ut are related to the structural disturbances �t by ut = Φ∗ �˜t , where �˜t is a standardized version of �t with unit variance. According to Assumption 1, the shock �z,t generates a stochastic trend in pro ductivity and expenditures on consumer durables. The two series are cointegrated with cointegration vector λ = [1, −λ21 , 0]� . Instead of restricting λ21 to one we esti mate the parameter in the VAR analysis to allow for a possibly steeper Engle curve for expenditures on durable goods. We do not impose a cointegration relationship � between the cumulative hours of work ( tτ =0 Lm,t ) and labor productivity or con sumer durables. Hence, the rank of Φvec = µλ� is one, which is in fact confirmed by a formal selection based on Bayesian posterior odds. The stochastic trend of yt has 2 Our analysis does not consider other disturbances such as monetary and fiscal policy shocks. For post-war U.S. data, government policy shocks are often considered to be of secondary impor tance in business-cycle analysis. For example, according to King, Plosser, Stock and Watson (1991), permanent nominal shocks account for only a small variation of real variables. The cyclical compo nents of government spending is not highly correlated with output measures – it is less than 0.2 for Hodrick-Prescott filtered data. Also, expanding the list of shocks often invites arbitrary identifying restrictions in the VAR analysis. 8 the form CLR Φ∗ �t ˜t . τ =0 � Since productivity and consumption expenditures have a common trend, the first two rows of the 3 × 3 long-run multiplier matrix CLR are proportional. The structural shocks �˜t are identifiable if the elements of the 3 × 3 matrix Φ∗ can be uniquely determined based on Φ0 , . . . , Φp , Φvec , and Σu . Let Ψ∗ denote the unique lower triangular Cholesky factor of Σu . Any matrix Φ∗ such that Φ∗ Φ�∗ = Σu is an orthonormal transformation of Ψ∗ , that is, Φ∗ = Ψ∗ B for some orthonormal matrix B. Let [A]ij denote the i’th row and j’th column of a matrix A. According to Assumption 1, the shocks �a,t and �b,t only have transitory effects on productivity and consumer expenditures. Thus, the elements [(CLR Ψ∗ )B]12 and [(CLR Ψ∗ )B]13 have to be zero. The contemporary effects of the labor-supply shock �b,t on productivity and hours worked are given by ∂Pt /∂�b,t = [Φ∗ ]13 and ∂Lm,t /∂�b,t = [Φ∗ ]33 . Define C∗ = [1, 0, −(α − 1)]. According to Assumption 2 and Equation (4) the value of [(C∗ Ψ∗ )B]13 has to be zero. These three orthogonality conditions uniquely determine the orthonormal transformation B. 3 A Fully Specified Model Economy The DSGE model provides a more specific interpretation of the three structural shocks and their propagation. It also assists the understanding of the economic intuition behind our identification scheme. The model economy consists of identical, infinitely lived households who maximize the expected discounted lifetime utility 9 defined over consumption Ct and pure leisure IE t �∞ � � β s−t (log Cs + κ log(1 − Lm,s − Lh,s )) , s=t where Lm,t is the fraction of time supplied to the labor market and Lh,t is hours spent on home production. IE t is the time t conditional expectation and β is the discount factor. Consumption is an aggregate of market consumption Cm,t and the consumption of home produced goods Ch,t : υ−1 υ−1 υ υ C(Cm,t , Ch,t ) = [χCm,t + (1 − χ)Ch,tυ ] υ−1 , (6) The substitution elasticity υ captures the households’ willingness to substitute mar ket and home-produced goods. The home production technology exhibits constantreturns-to-scale in home capital Kh,t and labor Lh,t : Ch,t = [ψ(Xh,t Lh,t ) τ −1 τ τ −1 τ + (1 − ψ)Kh,tτ ] τ −1 , (7) where Xh,t is a labor-augmenting home productivity process. We do not restrict the home technology to a Cobb-Douglas function (unlike the market technology below) because some consumer durable goods (e.g., dishwashers and microwaves) are substitutes for time, whereas others (e.g., DVD players) are complements. Our specification reduces to a separable-in-logs utility if ν = τ = 1. The households own Kh,t and the market capital stock Km,t ; and their budget constraint is: Cm,t + Im,t + Ih,t = Wt Lm,t + Rt Km,t , (8) where Rt is the rental rate of market capital and Im,t and Ih,t are capital invest ments. To avoid unreasonably volatile investments in a multi-sector model, capital 10 accumulation is subject to a convex adjustment cost as in Baxter (1996): Kj,t+1 = φ(Ij,t /Kj,t )Kj,t + (1 − δ)Kj,t , j = h, m (9) where δ is the depreciation rate of capital and φ� > 0, φ�� ≤ 0. The representative firm produces the market output Yt according to a CobbDouglas technology in market capital and labor and maximizes the profit each pe riod: max Lm,t ,Km,t 1−α Km,t (Xm,t Lm,t )α − Wt Lm,t − Rt Km,t , (10) where Xm,t represents a labor-augmenting market productivity process. The goods market equilibrium condition is: Yt = Cm,t + Im,t + Ih,t . (11) Market and home productivity are, respectively, Xm,t = exp[zt + at ] and Xh,t = exp[zt + bt ], where zt represents a common trend that follows a random walk with drift: zt = γ + zt−1 + �z,t . (12) The temporary components, at and bt , follow stationary first-order autoregressions: at = ρa at−1 + �a,t , bt = ρb bt−1 + �b,t . (13) We assume that �t = [�z,t , �a,t , �b,t ]� is serially uncorrelated with diagonal covariance matrix Σ� . The diagonal elements are denoted by σz2 , σa2 , and σb2 , respectively. As the VAR in Section 2, the log-linearized DSGE model provides a probabilistic representation for Δyt = [Δ ln Pt , Δ ln Ih,t , Lm,t ]� . The model economy also satisfies 11 all the identifying assumption of the VAR described above. Moreover, the DSGE model imposes additional restrictions which potentially yield more precise estimates of variance decompositions and impulse responses. 4 Econometric Approach The goal of the econometric analysis is to assess the relative importance of laborsupply shocks for the cyclical variation of output and hours.3 The VAR is denoted by M0 and the over-identified log-linearized DSGE model by M1 . To be consistent with the Cobb-Douglas production technology used in the DSGE model, we will assume that under the VAR specification the slope of the inverse labor demand function is also α − 1. Hence, the parameter α appears in both M0 and M1 . The parameters of model Mi , except for α, are stacked in the vector θ(i) , i = 0, 1. θ(0) contains the cointegration parameter λ12 and the non-redundant elements of the reduced-form matrices Φ0 , . . . , Φp , Σu in Equation (5). Variance decompositions and truncated impulse response functions, denoted by the m × 1 vector ϕ, are transformations of the parameters θ(i) and α, that is, ϕ = ϕ˜i (θ(i) , α). Under both M0 and M1 the vector process Δyt has a movingaverage (MA) representation in terms of the standardized structural shocks �˜t : Δyt = µ(θ(i) , α) + ∞ � Cj (θ(i) , α)�˜t−j . (14) j=0 3 A Technical Appendix that summarizes the computational details is available from authors upon request. 12 The population mean µ and the moving average coefficients Cj are model-specific functions of θ(i) and α. Define the vectors Mz = [1, 0, 0]� , Ma = [0, 1, 0]� , and Mb = [0, 0, 1]� . The impulse responses to the shock �˜s,t are given by Cj Ms . The h-th order autocovariance matrix of Δyt can be decomposed into the con tributions of the three structural shocks: � ΓΔy (h) = s∈{z,a,b} ∞ � � (s) ΓΔy (h) = Cj Mk Mk� Cj� +h . (15) s∈{z,a,b} j=max{0,−h} The relative contribution of shock s to the unconditional variance of the j’th element (s) of Δyt is given by the ratio [ΓΔy (0)]jj /[ΓΔy (0)]jj . The spectrum of the stationary process Δyt is SΔy (ω) = � (s) SΔy (ω) = � ∞ � (s) ΓΔy (h)e−ihω (16) s∈{z,a,b} h=−∞ s∈{z,a,b} and represents the contribution of frequency ω to the variance of Δyt . To assess the relative importance of the three shocks at business cycle frequencies we consider the decomposition of �ω ω SΔy (ω)dω, where ω and ω correspond to cycles of 32 and 6 quarters, respectively.4 The likelihood functions for the two models are denoted by p(YT |θ(i) , α, Mi ). 4 According to M0 and M1 the level of output is integrated of order one and its autocovariances (s) do not exist. Let SΔy (ω) denote the three components of the spectrum of output growth. We define the spectrum of the level of output at frequencies ω > 0 as (s) Sy(s) (ω) = lim φ→1 SΔy (ω) . 1+ − 2φcos(ω) φ2 (17) The term 1/[1 + φ2 − 2φcos(ω)] is the power transfer function of the AR(1) filter [1 − φL]−1 , where L denotes the temporal lag operator. 13 We adopt a Bayesian approach and place a prior distribution with density p(θ(i) , α|Mi ) = p(θ(i) |Mi )p(α), i = 0, 1 (18) on the parameters. Equation (18) reflects the assumption that α is a priori inde pendent of θ(0) and θ(1) . Moreover, the prior distribution of α is the same for both models. According to Bayes Theorem the posterior density of the parameters is proportional (∝) to p(θ(i) , α|YT , Mi ) ∝ p(YT |θ(i) , α, Mi )p(θ(i) |Mi )p(α). (19) The likelihood function of the VAR is uninformative about the slope of the inverse demand schedule α − 1 and depends only on the reduced form parameters θ(0) : p(YT |θ(0) , α, M0 ) = p̃(YT |θ(0) , M0 ). (20) Straightforward manipulations using Bayes Theorem can be used to verify that the VAR posterior is the product of the posterior density of the identifiable reduced-form parameters obtained from p̃(YT |θ(0) , M0 ) and the prior density of α:5 p(θ(0) , α|YT , M0 ) = p̃(θ(0) |YT , M0 )p(α). (21) Since ϕ = ϕ˜i (θ(i) , α), Equations (18) and (19) implicitly determine the prior and posterior of variance decompositions and impulse response functions. Rather than attempting to specify a prior on ϕ directly, as in Gordon and Boccanfuso (2001), we 5 Our VAR based inference is a specific example of Bayesian analysis of a nonidentified econo metric model. Poirier (1998) provides a comprehensive survey and many additional examples. 14 use economic intuition derived from assumptions on aggregate preferences, produc tion technologies, and equilibrium relationships to specify the prior for ϕ indirectly. Since the distribution of reduced-form parameters θ(0) is updated based on the sam ple information YT , the implied distribution of ϕ is updated with every observation and we learn about the relative importance of structural shocks and the response of the economy. Our method is explicit about the direction of the parameter space in which learning does not occur. If the dimension of the nonidentifiable component of the parameter vector is low, as in our application, we can assess the robustness of our conclusion by tracing out, for instance, the relative importance of the labor-supply shock as a function of α. A similar approach was used by King and Watson (1992) who plotted their statistics of interest against a one-dimensional variable indexing VAR identification schemes. The VAR identification proposed in this paper is based on the notion that pro ductivity and hours worked move in opposite directions in response to a labor-supply shock. Equation (4) can be qualitatively interpreted as an inequality restriction on the impulse responses: ∂ ln Pt > 0 and ∂�b,t ∂ ln Lm,t <0 ∂�b,t (22) Canova and DeNicolo (2002), Faust (1998), and Uhlig (2003) develop identification and inference procedures based on such inequality constraints. Our approach places a prior distribution on identification schemes that are consistent with (22) and av erages the posterior distribution of population characteristics ϕ over a priori likely 15 values of the unidentifiable parameter α that indexes the identification schemes. 5 Empirical Analysis Both VAR and DSGE models are fitted to post-war quarterly U.S. data on labor productivity, expenditure on consumer durables, and hours worked.6 The sample period ranges from 1964:I to 1997:IV. The overall sample size is T = 136 and the first T∗ = 20 observations are used as training sample to initialize lags and parameterize the prior distributions. 5.1 Priors The prior distribution used in the estimation of the DSGE model is summarized in columns 3 to 5 of Table 1. The shapes of the densities are chosen to match the domain of the structural parameters. The prior means for labor share, discount rate, productivity growth, capital depreciation, and the steady state ratio of home to market investment are respectively α ¯ = 0.666, β¯ = 0.993, γ¯ = 0.004, δ¯ = 0.025, and I¯h /I¯m = 0.7. These values can be justified based on the training sample and 6 Real gross domestic product (GDPQ), consumption of consumer durables (GCDQ), employed civilian labor force (LHEM), civilian noninstitutional population 20 years and older (PM20 and PF20) are extracted from the DRI·WEFA database. Define P OP Q = 1E6 ∗ (P F 20 + P M 20), Yt = GDP Q/P OP Q and Ih,t = GCDQ/P OP Q. Average weekly hours, private non-agricultural establishments (EEU00500005) are obtained from the Bureau of Labor Statistics. Annual hours worked at monthly frequency are Lm,t = 52 ∗ EEU 00500005 ∗ LHEM / P OP Q and converted to quarterly frequency by simple averaging. Labor productivity is Pt = Yt /Lm,t . 16 are commonly used in the literature. Hence, we use fairly small standard deviations for the distributions of these parameters. Prior means for the steady state hours, ¯ m = 0.33 and L ¯ h = 0.25, are obtained from the Michigan Time Use Survey. A L larger standard deviation is allowed for Lh , as the hours spent on home work may be measured with a greater uncertainty. We allow for large standard deviations in the prior distributions of home tech nology parameters because they are not easy to infer. The prior means for the substitution elasticities, ν̄ = 1 and τ̄ = 1, correspond to a one-sector model with separable-in-logs utility. The prior mean of the labor share ψ in the home produc tion function is also set to 0.666. The weight on leisure χ in the utility function is implicitly determined by the other parameters. The steady state adjustment costs are assumed to be zero and the elasticity of the investment/capital ratio with re spect to Tobin’s q, η = (|(I ∗ /K ∗ )φ�� /φ� |−1 ) is estimated. The prior mean for η is 100, implying a small adjustment cost, with a large standard deviation of 100. We use diffuse priors for the exogenous technology processes at , bt , and zt . Finally, we introduce two additional parameters ξ1 and ξ2 to adjust the normalization of total hours to one and to capture the average growth rate differential between labor pro ductivity and home investment in the data. The structural parameters are assumed to be a priori independent of each other. The training-sample is used to construct a conjugate prior for the VAR parame ters conditional on the cointegration parameter λ21 . While the DSGE model implies that λ21 = 1, we relax that restriction and choose the prior λ21 ∼ N (1, 0.0252 ). The prior for α is the same as in the DSGE model analysis (Table 1). Posterior odds 17 were used to select the lag-length p = 2. 5.2 Parameter Estimation and Time Series Fit The posterior means and standard errors of the parameters of the DSGE model are reported in columns 6 and 7 of Table 1.7 The estimated substitution elastic ity between market goods and home goods, υ, is 2.302, slightly higher than those of Rupert, Rogerson, and Wright (1995) and McGrattan, Rogerson, and Wright (1997). The substitution elasticity between capital and labor in home production, τ , is 2.446 suggesting that goods and time are substitutes in home production activ ity. The estimated labor share ψ in the home technology is 0.753 and the fraction of hours spent on home production activity Lh is 0.170. The temporary home pro duction shock is somewhat more persistent than the market shock: ρ̂a = 0.745 and ρ̂b = 0.865. The 90 percent posterior confidence interval for the correlation (con ditional on time t − 1 information) between market productivity ln Xm,t and home productivity ln Xh,t ranges from 0.18 to 0.37, somewhat lower than the values used in the literature (e.g., 0.67 in Benhabib, Rogerson and Wright, 1991 and 1 in Greenwood and Hercowitz, 1991). Finally, the adjustment cost parameter estimate η̂ is 30.70, implying a small adjustment cost in capital accumulation. To assess the relative time series fit of the VAR and the DSGE model we compute 7 While McGrattan, Rogerson, and Wright (1997) also estimate home production models based on aggregate time series, our analysis distinguishes itself from theirs in several dimensions. First, we focus on the comparison to a structural VAR, particularly, the variance decomposition. Second, microeconomic evidence is incorporated through prior distributions in our Bayesian estimation. Third, we are able to uncover the comovement of innovations to market and home productivity. 18 marginal data densities � p(YT |Mi ) = p(YT |θ(i) , α, Mi )p(θ(i) , α|Mi )d(θ(i) , α) (23) conditional on the training sample 1964:I to 1968:IV. The log-marginal data density can be interpreted as a measure of one-step-ahead predictive performance ln p(YT |Mi ) = �T t=T∗ p(yt |Yt , Mi ). The values are ln p(YT |M0 ) = 1087.2 for the VAR and ln p(YT |M1 ) = 999.6 for the DSGE model, implying that for a wide range of prior probabilities the posterior probability of the DSGE model is essentially zero.8 To shed more light on the poor fit of the DSGE model, we computed in-sample, root-mean-squarederrors (RMSE) at the posterior mode estimates. The RMSE’s for the growth rates of output and consumer durable expenditures are very similar for the two models, whereas the RMSE of hours is substantially higher for the DSGE model: 0.0080 versus 0.0059 for the VAR. 5.3 Variance Decompositions and Impulse Responses Our main interest is to unveil the sources of cyclical variation in hours and output. Table 2 presents the variance decomposition (posterior means and 90 percent confi dence intervals) for output and hours at business cycle frequencies, namely cycles of 6 to 32 quarters. Labor-supply shifts play an important role for the fluctuations of hours. The shock �b accounts for almost 30 percent (posterior mean) of the fluctu ations in hours according to the VAR, and more than half according to the DSGE model. The relative contribution of the labor-supply shocks to output fluctuations 8 The posterior probability of model Mi is πi,T = probability. � πi,0 p(YT |Mi ) , πj,0 p(YT |Mj ) j=0,1 where πi,0 is its prior 19 is somewhat small albeit non-negligible; they account for about 15 percent of the output variation. The posterior confidence intervals indicate that the VAR model based variance decompositions are associated with more posterior uncertainty than the DSGE decompositions. For instance, the 90 percent confidence intervals for the contribution of �b,t to hours fluctuations are [0, 0.277] (VAR) and [0.083, 0.186] (DSGE), respec tively. As pointed out by Faust and Leeper (1997), VAR variance decompositions based on long-run restrictions are associated with a high degree of uncertainty. Con ditional on the over-identifying restrictions embodied in the DSGE model, however, one obtains fairly sharp estimates. While VAR and DSGE model analysis broadly agree upon the decomposition of output fluctuations, there are discrepancies in terms of the variance decomposition of hours. According to the marginal data densities and the RMSEs, the time series fit of the VAR is much better than the fit of the DSGE model, in particular in terms of hours worked. Hence, the DSGE-based decomposition probably overestimates the contribution of labor-supply shocks to the fluctuations of hours. The finding that the DSGE model analysis attributes only about 7 percent of the hours fluctuation to the permanent technology shock �z,t is partly due to the balanced-growth-path property of this class of models; common technology shocks tend to shift both labor demand and supply in a similar magnitude, leaving hours almost unaffected. The VAR-based point estimates and confidence intervals reported in rows 1-3 and 10-12 of Table 2 were computed based on the reduced form parameters θ(0) and the slope of the labor demand schedule, α − 1. Conditional on the VAR, because 20 the data provide no information about α, the inference is potentially sensitive to the choice of the prior p(α). Moreover, our identifying assumption for labor-supply shocks exploits the notion that the production capacity is fixed in the short run and that there exists a stable relationship between labor productivity and hours employed. This premise may be violated if firms heavily rely on the factor utilization. For example, if there is a significant variation in the level workers’ effort (e.g., labor hoarding in Burnside, Eichenbaum, and Rebelo, 1993) in the face of exogenous shifts in labor supply, our identifying restriction is no longer valid. However, if the capital utilization is time-varying and the cost of intense utilization results in a faster depreciation of capital, our identifying restriction is still appropriate. In this case, the restriction has to be modified to accommodate the systematic capital utilization; the slope of the labor demand schedule is smaller in absolute value than α − 1 (see Appendix). To assess the robustness of our VAR-based analysis we also report the contribution of the labor-supply shock to output and hours variation conditional on α = 0.3, 0.6, and 0.9 in Table 2. The share of �b lies between 23 and 38 percent for hours and 12 to 20 percent for output. As we move to a higher value of α, i.e., utilization of capital is less costly, the importance of the labor-supply shock is reinforced. Overall, both the VAR and DSGE model analysis suggests that the labor-supply shocks play an important role for economic fluctuations, especially for hours worked. Our estimates are, however, smaller than the previous estimates. Shaprio and Wat son (1988) find that 60 percent of the cyclical variation in hours is due to permanent shifts in labor supply. In Shapiro and Watson, labor-supply shocks are identified by 21 the stochastic trends in hours worked. While the empirical evidence on the station arity of hours is not conclusive, we assume hours to be stationary. This is consistent with a large class of DSGE models. Hall (1997) attributes almost the entire variation of hours to preference shocks, which he identifies as deviations from the optimal ity condition associated with the labor supply of a competitive household. Hall’s finding is broadly in line with our DSGE model analysis as both approaches impose more restrictions on the short-run labor market equilibrium and, as a result, require bigger shifts in labor supply. However, as mentioned above, the weak time series fit of the DSGE model indicates that some of these over-identifying restrictions are potentially misspecified and that the corresponding findings have to be interpreted with caution. Our VAR identification scheme does not rely on a specific form of household’s preference. Instead, it is based on the firm’s production technology through the slope of marginal-product-of-labor schedule. To assess whether the structural shocks identified from the VAR conform with our economic intuition, Figure 1 depicts one-standard-deviation impulse responses of labor productivity, expenditure on consumer durable goods (investment in home capital), and market hours to the three structural shocks.9 The graphs show the re sponses from the DSGE model (solid lines) and those from the VAR along with the 90 percent confidence interval (dotted lines). Looking at the first row, in response to a permanent shock, labor productivity both in the DSGE model and VAR ap proach the new steady state at a similar pace. Spending on consumer durables also 9 The signs of the responses are normalized as follows: the initial responses of productivity to �z,t , productivity to �a,t , and hours to �b,t are positive, positive, and negative, respectively. 22 increases permanently. Market hours rises instantaneously and slowly returns to its steady state in the DSGE model, whereas its response is hump-shaped in the VAR. The VAR responses to a temporary market productivity shock (second row of the Figure) closely traces those from the DSGE model except for the delayed response of hours in the VAR. Finally, in response to a temporary home productivity shock, labor productivity rises in both the VAR and DSGE model. Home investment rises immediately in the DSGE model whereas it exhibits a slow hump-shape response. In sum, the VAR responses, by and large, conform with the economic intuition pro vided by the DSGE model. However, hours responses are delayed in the VAR for about 2 quarters, and the short-run dynamics of expenditure on consumer durables are somewhat different from the DSGE model prediction. Based on the competitive labor market equilibrium, we identify exogenous shifts in labor supply. Yet the proposed identification scheme allows a more general – also an alternative – interpretation than labor-supply shocks. As an illustrative example, consider a model economy with sticky prices where firms have to produce goods to meet their demand. In this economy, the labor demand is no longer a simple reflection of the marginal product of labor. It is instead jointly determined by the demand for goods and the output-labor elasticity from the production technology. Suppose now there is an increase in the demand for goods that is not caused by a productivity shift. This will lead to an increase in the demand for labor at a given level of production capacity. The joint behavior of labor productivity and hours is still dictated by the marginal-product-of-labor schedule.10 10 In this event, the real wage will increase given the upward sloping labor supply curve. However, 23 5.4 Evolution of Latent Technology Processes According to the home production model, recessions may occur because agents find it optimal to allocate more time in non-market activities. In our DSGE model the attractiveness of non-market activity is measured by the latent home technology process.11 We plot three technology indices in Figure 2 together with the NBER business cycle peaks and troughs. All five recessions during the sample period are associated with low levels of market productivity. Two business cycle troughs, March 1975 and November 1982, coincide with unusually high productivity of nonmarket activities. The strong interpretation of this finding is that unusually high non-market productivity or preference shift has contributed to a low employment and output. A weaker interpretation is that the economic downturns in March 1975 and November 1982 cannot be solely explained by an adverse technology shock in the market. 6 Conclusion We investigate the sources of economic fluctuations in the context of a dynamic general equilibrium. A new VAR identification scheme which distinguishes between shifts of and movements along the marginal product of labor is proposed to identify labor productivity falls as employed hours increases, justifying our use of labor productivity instead of a wage series under this more general interpretation. 11 Ingram, Kocherlakota and Savin (1997) recover non-market variables (non-market consumption, non-market hours, and leisure) based on the households’ optimality conditions and times series of market variables, whereas we obtain the time series of the latent home technology. 24 three structural disturbances: temporary labor-supply shifts, temporary labor de mand shifts, and permanent productivity shocks that eventually move both demand and supply. According to the variance decomposition from the VAR, the laborsupply shifts are an important driving force of the cyclical fluctuation of hours, as they account for about 30 percent of the variation. For output fluctuations at business cycle frequencies, the role of labor-supply shifts is modest as their relative contribution is about 15 percent. To assess the importance of labor-supply shifts in the context of an equilibrium model, a home-production model with stochastic variation in non-market technology is estimated. While the DSGE model based decomposition of output resembles the VAR results, the structural model attributes a higher fraction of hours fluctuations to the labor-supply shifts than the VAR. This result, however, is partly due to misspecified over-identifying restrictions of the DSGE model. 25 Appendix: Labor Demand with Variable Capital Utiliza tion Consider a Cobb-Douglas production function with inputs in capital services and hours: Yt = (ut Km,t )1−α (Xm,t Lm,t )α , where ut represents the utilization of the capital stock. Suppose the intensive use of capital results in a fast depreciation. At the cost of a more complicated notation, we could work with an alternative decentralization scheme in which firms make decisions on accumulation. However, since both decentralizations are essentially identical, as in the main text, suppose the firm rents the capital from the households. Yet the firm has to compensate households for faster depreciation when the capital is utilized more intensively: max Lm,t ,Km,t ,ut (ut Km,t )1−α (Xm,t Lm,t )α − Wt Lm,t − (Rt + δ(ut ))Km,t . For illustrative purposes, assume that the elasticity of depreciation is constant: δ(ut ) = δ0 uλ+1 t λ+1 , where λ > 0. As λ → ∞, the utilization is held constant and the depreciation rate is fixed. The first order conditions of the profit maximization problem with respect to Lm,t and ut imply that the inverse labor demand schedule still depends on the predetermined capital stock and the market productivity shocks only. However, its slope changes: ∂ ln Lm,t ∂ ln Wt = µ(α − 1) , ∂�b,t ∂�b,t µ= λ ≤ 1. λ+α 26 Therefore, the proposed identification scheme is still valid but the slope of the labor demand schedule is smaller than in the constant utilization case, reflecting an extra margin for the firm to exploit. 27 References Baxter, M., 1996, Are Consumer Durables Important for Business Cycles?, Review of Economics and Statistics 78, 147-155. Baxter, M. and R. King, 1991, Productive Externalities and Business Cycles, Dis cussion paper 53, Institute for Empirical Macroeconomics, Federal Reserve Bank of Minneapolis. Benhabib, J., R. Rogerson, and R. Wright, 1991, Homework in Macroeconomics: Household Production and Aggregate Fluctuations, Journal of Political Econ omy 99, 1166-1187. Burnside, C., M. Eichenbaum, and S. Rebelo, 1993, Labor Hoarding and the Busi ness Cycles, Journal of Political Economy 101, 245-273. Canova, F. and G. DeNicolo, 2002, Monetary Disturbances for Output for Business Cycle Fluctuations in the G-7, Journal of Monetary Economics 49, 1131-1159. Faust, J., 1998, The Robustness of Identified VAR Conclusions about Money, Carnegie Rochester Conference Series 49, 207-244. Faust, J. and E. Leeper, 1997, When Do Long-run Identifying Restrictions Give Reliable Results?, Journal of Business & Economic Statistics 15, 345-353. Gordon, S. and D. Boccanfuso, 2001, Learning from Structural Vector Autoregres sion Models, Manuscript, Universite Laval, Quebec City. 28 Greenwood, J., and Z. Hercowitz, 1991, The Allocation of Capital and Time over the Business Cycle, Journal of Political Economy 99, 1188-1215. Hall, R., 1997, Macroeconomic Fluctuations and the Allocation of Time, Journal of Labor Economics 15, s223-s250. King, G. R., C. I. Plosser, J. H. Stock, and M. Watson, 1991, Stochastic Trends and Economic Fluctuations, American Economic Review 81, 819-840. King, R. and M. Watson, 1992, Testing Long-run Neutrality, NBER Working Paper 4156. Ingram, B., N. Kocherlakota, and N. E. Savin, 1997, Using Theory for Measure ment: An Analysis of the Cyclical Behavior of Home Production, Journal of Monetary Economics 40, 435-456. McGrattan, E., R. Rogerson, and R. Wright, 1997, An Equilibrium Model of the Business Cycle with Household Production and Fiscal Policy, International Economic Review 38, 267-290. Parkin, M., 1988, A Method for Determining whether Parameters in Aggregative Models are Structural, Carnegie-Rochester Conference on Public Policy 29, 215-252. Poirier, D., 1998, Revising Beliefs in Nonidentified Models, Econometric Theory 14, 183-509. Rupert, P., R. Rogerson, and R. Wright, 1995, Estimating Substitution Elasticities in Household Production Models, Economic Theory 6, 179-193. 29 Shapiro, M. and M. Watson, 1988, Sources of Business Cycle Fluctuations, NBER Macroeconomics Annual 1988, 111-48. Uhlig, H., 2003, What are the Effects of Monetary Policy? Evidence from an Ag nostic Identification Procedure, Journal of Monetary Economics, forthcoming. 30 Table 1: Prior and Posterior Distribution for DSGE Model Parameters Parameters Name Range α Prior Posterior Density Mean S.D. Mean S.D. [0,1] Beta 0.666 0.020 0.741 0.018 β [0,1] Beta 0.993 0.002 0.979 0.003 γ IR Normal 0.004 0.0005 0.003 0.0004 δ [0,1] Beta 0.025 0.002 0.017 0.002 Lm [0,1] Beta 0.330 0.020 0.335 0.021 ρa [0,1] Beta 0.800 0.100 0.745 0.034 ρb [0,1] Beta 0.800 0.100 0.865 0.034 Lh [0,1] Beta 0.250 0.050 0.183 0.045 η IR+ Gamma 100.0 100.0 30.70 5.546 ψ [0,1] Beta 0.666 0.100 0.753 0.080 ν IR+ Gamma 1.000 2.000 2.302 0.388 Ih /Im IR+ Gamma 0.700 0.020 0.686 0.020 τ IR+ Gamma 1.000 2.000 2.446 0.536 ξ1 IR Normal 2.960 1.000 3.145 0.006 ξ2 IR Normal 0.000 0.020 0.005 0.0003 σz IR+ InvGamma 0.01∗ 2.000∗ 0.007 0.001 σa IR+ InvGamma 0.01∗ 2.000∗ 0.009 0.001 σb IR+ InvGamma 0.015∗ 2.000∗ 0.021 0.007 Notes: For the Inverse Gamma (u, s) priors we report the parameters u and s. For u = 2 the standard deviation (S.D.) is infinite. The posterior moments are calculated from the output of the Metropolis algorithm. 31 Table 2: Variance Decomposition at Business Cycle Frequencies Variable Model Shock Mean ln Lm,t VAR �z 0.205 [ 0.000 0.529 ] VAR �a 0.510 [ 0.069 0.905 ] VAR �b 0.285 [ 0.005 0.588 ] VAR (α = 0.3) �b 0.230 [ 0.006 0.454 ] VAR (α = 0.6) �b 0.268 [ 0.007 0.551 ] VAR (α = 0.9) �b 0.376 [ 0.021 0.738 ] DSGE �z 0.066 [ 0.006 0.124 ] DSGE �a 0.268 [ 0.132 0.403 ] DSGE �b 0.666 [ 0.526 0.805 ] VAR �z 0.491 [ 0.089 0.911 ] VAR �a 0.367 [ 0.000 0.703 ] VAR �b 0.142 [ 0.000 0.277 ] VAR (α = 0.3) �b 0.118 [ 0.000 0.250 ] VAR (α = 0.6) �b 0.130 [ 0.000 0.256 ] VAR (α = 0.9) �b 0.204 [ 0.000 0.388 ] DSGE �z 0.448 [ 0.339 0.560 ] DSGE �a 0.417 [ 0.311 0.519 ] DSGE �b 0.135 [ 0.083 0.186 ] ln Lm,t ln Ym,t ln Ym,t Conf. Interval Notes: Decomposition of aggregate output ln Yt and market hours ln Lm,t at business cycle frequencies (6 to 32 quarters per cycle). The table reports posterior means and 90 percent confidence intervals. VAR (α = x) signifies that α was fixed at the value x rather than integrated out with respect to its prior distribution. 32 Figure 1: Impulse-response Functions Notes: Figure depicts posterior mean responses for VAR (dashed) and DSGE model (solid). The dotted lines represent pointwise 90 percent Bayesian confidence intervals based on the VAR posterior. 33 Figure 2: Filtered technology processes at , bt , and xt Notes: The posterior mean estimates of the latent technology processes are based on the DSGE model. Solid vertical lines correspond to business cycle peaks, dashed lines denote business cycle troughs (NBER Business Cycle Dating).