The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Working Paper Series Deep Habits in the New Keynesian Phillips Curve WP 11-08 Thomas A. Lubik Federal Reserve Bank of Richmond Wing Leong Teo University of Nottingham This paper can be downloaded without charge from: http://www.richmondfed.org/publications/ Deep Habits in the New Keynesian Phillips Curve Thomas A. Lubik Federal Reserve Bank of Richmondy Wing Leong Teo University of Nottinghamz November 2011 Working Paper No. 11-08 Abstract We derive and estimate a New Keynesian Phillips curve (NKPC) in a model where consumers are assumed to have deep habits. Habits are deep in the sense that they apply to individual consumption goods instead of aggregate consumption. This alters the NKPC in a fundamental manner as it introduces expected and contemporaneous consumption growth as well as the expected marginal value of future demand as additional driving forces for in‡ation dynamics. We construct the driving process in the deep habits NKPC by using the model’s optimality conditions to impute time series for unobservable variables. The resulting series is considerably more volatile than unit labor cost. General Methods of Moments (GMM) estimation of the NKPC shows an improved …t and a much lower degree of indexation than in the standard NKPC. Our analysis also reveals that the crucial parameters for the performance of the deep habit NKPC are the habit parameter and the substitution elasticity between di¤erentiated products. The results are broadly robust to alternative speci…cations. JEL Classification: E31; E32. Key Words: Phillips curve; GMM; marginal costs; deep habits. The views expressed in this paper are those of the authors and should not be interpreted as those of the Federal Reserve Bank of Richmond or the Federal Reserve System. y Research Department, P.O. Box 27622, Richmond, VA 23261. Tel.: +1-804-697-8246. Email: thomas.lubik@rich.frb.org. z School of Economics, University of Nottingham - Malaysia Campus, Jalan Broga, 43500 Semenyih, Selangor, Malaysia. Tel.: +603-8924-8698. Email: WingLeong.Teo@nottingham.edu.my 1 1 Introduction The New Keynesian Phillips curve (NKPC) is the centerpiece of modern macroeconomic models used for monetary policy analysis. It is derived from the optimal price-setting problem of a monopolistically competitive …rm that operates in an environment where …rms face downward-sloping demand curves. The NKPC, in contrast to earlier accelerationist Phillips curves, is explicitly forward-looking and imposes theoretical restrictions on the comovement of its components. Speci…cally, theory identi…es marginal cost as the main driver of in‡ation dynamics. However, the NKPC faced the early criticism that marginal cost is not observable to the empirical researcher and that the stochastic properties of various proxies do not line up with the properties of the in‡ation process they claim to explain. In a seminal contribution, Galí and Gertler (1999) show that the performance of the NKPC can be improved by introducing backward-looking price-setting or indexation to the e¤ect that lagged in‡ation enters the structural speci…cation. Moreover, they demonstrate that marginal cost is well proxied by unit labor cost under the assumption of perfectly competitive factor markets. Their results established a benchmark in the literature, namely that in‡ation dynamics are explained both by intrinsic factors, such as in‡ation indexation in price-setting, and by external driving forces, such as marginal cost movements. Much of the follow-up research con…rmed their initial …ndings and established their robustness under the chosen modeling environment (for instance, Galí et al., 2005). This paper follows in the steps of more recent research that modi…es the environment in which …rms operate. We introduce deep habits in the preferences of the consumer and derive the corresponding NKPC. Habit formation is deep in the sense that it extends to each individual good of the consumption bundle available to consumers, and not only to the consumption composite. This seemingly simple modi…cation, however, has far-reaching consequences. It implies a downward-sloping demand function that depends on the lagged level of the consumer’s purchases. Since …rms take this demand function as a constraint in their optimal price-setting problem, the time dependence carries over to the NKPC and results in the introduction of future, current, and lagged consumption in this relationship. We thus show that deep habits fundamentally a¤ect the analytical form and interpretation of the driving process and the interaction of marginal cost with in‡ation. We estimate the NKPC under deep habits using generalized methods of moments (GMM) techniques. More speci…cally, our empirical approach is a mixture of calibration 2 and structural estimation. In our benchmark speci…cation, we combine the additional explanatory variables introduced by the deep habits environment with marginal cost into a single driving process. We then impute this unobservable series using data on consumption and real unit labor cost. The weights on the various elements are functions of the model’s structural parameters, which we calibrate. This procedure allows us to compare the driving process of the standard NKPC, namely marginal cost, with the one implied by deep habits. We show that the latter is considerably more volatile than real unit labor cost. This observation is re‡ected in the NKPC estimates for the coe¢ cients on the driving process and the weight on the intrinsic contribution to in‡ation dynamics, which is much smaller than in the standard speci…cation. Moreover, the …t of the deep habits NKPC is much improved over the standard NKPC per typical speci…cation measures in the GMM literature. The representation of the driving process under deep habits involves expectations of both observable variables, such as consumption, and unobservables, speci…cally the marginal value of demand. In order to back out processes for these variables we pursue a parametric approach in that we use the optimality conditions of the model to link the marginal value of demand to observable consumption. In the deep habits environment, this involves an expectational di¤erence equation which we solve forward to express the marginal value of demand as the present value of future consumption growth and marginal cost. We then pursue alternatively a univariate and a multivariate approach to produce forecasts for the latter variables, which are used to construct a synthetic series for these unobservables. Our approach is therefore of a partial equilibrium and limited information nature in that we do not use all potential information available within the general equilibrium context of the full theoretical model. We also study an alternative representation of the deep habits NKPC that uses analytical representations for the unobserved and expectational terms in the driving process. This yields a more reduced-form representation in terms of marginal cost and current consumption growth, which we can estimate directly. The coe¢ cients in this representation are functions of the structural parameters. This representation then allows us to identify the crucial elements for the improved performance of the deep habits model. Not surprisingly, the size of the deep habits parameter is the central element. We show that performance notably improves for values above 0:6. Our GMM-estimate of this parameter is 0:85, which is identical to previous estimates in the literature (see Ravn et al., 2010). The second important element is the degree of substitutability of di¤erentiated products in the consumer’s preferences, which is inversely related to the …rm’s markup. We estimate this to be consis- 3 tent with a markup of 74%, which may be considered high. A robustness analysis shows, however, that the performance of the deep habits NKPC is still very satisfactory for more typical markups of 10%. The modeling environment in our paper draws from, and contributes to, an emerging literature on deep habits, starting with the original contribution by Ravn et al. (2006). Their key insight was that deep habits impart additional internal propagation on a model, on top of what an external habit formation would produce, while at the same time being arguably more plausible. We extend this insight to the speci…cation of the NKPC by showing that at a purely empirical level deep habits add additional regressors to the empirical in‡ation equation. Moreover, Ravn et al. (2006) show that deep habits give rise to countercyclical markups, which translates into procyclical marginal cost in the context of the NKPC. We utilize this insight in the construction of a modi…ed driving process. Ravn et al. (2010) build on their earlier paper and extend it to a New Keynesian framework. They estimate the model using impulse-response function matching for identi…ed monetary policy shocks. Our estimates of the key structural parameters using a GMM approach are virtually identical to theirs, which suggests robustness of the derived insights. Our paper also touches upon a host of other research on the NKPC that is concerned with modifying the notion of the driving process. In the wake of Galí and Gertler (1999), researchers at …rst assessed the robustness of their conclusions with respect to proxies for marginal costs, the instrument set used in the estimation, and the empirical approach being taken.1 Our paper follows the spirit of this research in that we do not modify or speci…cally address the nature of marginal cost. We thus continue to proxy marginal cost with unit labor cost, which is justi…ed under the assumption of competitive factor markets and a standard neoclassical production structure. More recent contributions, however, deviate from this assumption. The key insight is that modi…cations to the production structure change the nature of the driving process in the NKPC either by introducing additional elements, that is, regressors, or by altering the responsiveness of in‡ation to marginal cost movements For instance, this literature introduces search and matching frictions in the labor market (e.g. Krause et al., 2008), or incorporates models with …nished goods inventories that highlight the distinction between marginal costs of production and marginal costs of sales (e.g. Lubik and Teo, 2011). In both approaches, modi…cation of the production side does not improve the performance of the NKPC. The reason is that the driving process in 1 Galí et al. (2005) is an example of the former, Kurmann (2007) an example of the latter. Nason and Smith (2008) give a comprehensive survey of the developments in this literature. 4 either speci…cation relies solely on present-value relationships of the cost of, respectively, maintaining long-run employment relationships and of keeping an inventory. By their very nature, present discounted values tend to smooth out the movements imparted by volatile driving processes. In the two cases this virtually negates the e¤ect of additional regressors. This e¤ect is present in our paper, too, but it is compensated by the presence of current consumption growth, as we show in the reduced-form representation of the NKPC. Guerrieri et al. (2010) provide a bridge between these di¤erent aspects of the literature. In an otherwise standard model of a small open economy, they introduce both variable demand elasticities on the preference side as well as a production structure that relies on intermediate inputs from domestic and foreign sources. They derive an NKPC, whose driving process is a function of marginal cost, relative prices and exogenous variations in the aggregate elasticity of …nal products, and …nd that the estimated NKPC provides a better …t than a standard speci…cation, mainly due to the addition of the relative price term. Our paper similarly improves the …t of the NKPC via changes to the demand structure in the form of persistence-inducing deep habits. We now proceed as follows. In the next section, we present our theoretical model and show how the introduction of deep habits modi…es the speci…cation of the NKPC. In Section 3, we discuss our empirical approach, the data being used, and the various speci…cations of the deep habits NKPC that we estimate. Section 4 presents the GMM estimation results for a standard NKPC and our deep habits model. In Section 5, we conduct various robustness checks and consider alternative speci…cations. The …nal section concludes. 2 The Model Our theoretical model is based on Ravn et al. (2010). It describes a New Keynesian monetary economy with utility-maximizing households and pro…t-maximizing …rms. Households consume a bundle of consumption goods, each of which individually is subject to habit formation. Firms are monopolistically competitive and hire labor from the households as input in the production process. They set their prices subject to a downward-sloping demand schedule, which is derived from household preferences, and they are subject to quadratic costs of adjusting nominal prices. It is the latter element that gives rise to an NKPC. In what follows, we describe the full model of the economy, but focus mainly on the relationships that are needed to derive the NKPC. 5 2.1 Households The household sector is described by the decisions of a representative household. Its intertemporal utility function is given by: V 0 = E0 1 X 1 t t=0 0< xt1 1 < 1 is the discount factor, and ht : (1) > 0 is the coe¢ cient of relative risk aversion. ht is the amount of labor supplied by the household to the …rm sector. The core of the deep habits formulation is captured by the term xt . It is a sub-utility function that depends on a CES aggregate of consumption goods i, cit , relative to the previous period’s consumption of goods i: xt = Z 1=(1 1=") 1 (cit 1 1) cit 1=" di : (2) 0 " > 1 is the substitution elasticity between the di¤erentiated goods, while 0 < 1 is the deep habits parameter. In contrast to habits at the aggregate level of consumption, deep habits apply to each individual consumption good and are thus deeply embedded into the utility function. Solving for the household’s optimality conditions is slightly more involved than in the standard case. We note that we can rewrite the household’s problem as an expenditure minimization problem: min Xt = cit Z 1 Pit cit di; 0 subject to the de…nition of the habit stock (2). The …rst-order condition implies the following demand function: cit = Pit Pt Pt = Z where 0 1 " xt + cit 1; (3) 1=(1 ") Pit1 " di (4) is the associated aggregate price index. The demand function has the typical feature that it is downward-sloping in its own relative price Pit =Pt . In addition, deep habits render consumption demand persistent through lagged consumption cit 1. It is speci…cally this feature which changes the nature of the driving process in the NKPC. We can now describe the household’s intertemporal utility optimization problem. The budget constraint of the representative household is given by: Pt xt + {t + Bt = Rt 6 1 Bt 1 + W t ht + t; (5) where {t = Z 1 Pit cit 1 di 0 is the current consumption expenditure required to maintain the habit consumption level from the previous period. Wt is the nominal wage payment for labor services ht . t is the residual pro…t accruing to the household from its ownership of the …rms. Finally, we assume that households have access to a risk-free one period nominal government bond Bt that pays a gross interest rate of Rt . The household maximizes the utility function (1) by choosing sequences of xt , ht , and Bt . The resulting …rst-order conditions are: xt xt = = Wt ; Pt R t Et (6) Pt x Pt+1 t+1 (7) The …rst equation is the standard labor-leisure trade-o¤ a household faces, where the variable that determines the overall level of economic activity is the habit stock xt instead of the usual consumption aggregate. Similarly, the second optimality condition is a standard Euler-equation for intertemporal smoothing with xt instead of consumption. This relationship also de…nes a stochastic discount factor which …rms use to evaluate future pro…t streams. 2.2 Firms The …rm sector is composed of a continuum of monopolistically competitive …rms that face a downward-sloping demand schedule for their product. Demand for …rm i0 s output is given by equation (3), where we associate each good i with a speci…c …rm. Firms choose an optimal price, but they are subject to quadratic costs of price adjustment. We assume that …rms have access to a linear production technology that uses labor as its only input.2 Production is subject to aggregate productivity disturbances At . We write the production function as yit = At hit . Each …rm hires labor input from the representative household for a competitive wage Wt . The …rm has to solve an intertemporal pro…t maximization problem: " # 1 2 X Pit max E0 qt Pit cit Wt hit Pt ct ; t 2 Pit 1 fPit ;cit ;hit g (8) t=0 2 The assumption of constant returns to scale is immaterial for the NKPC as it is independent of the curvature of the production function. Derivations and results for alternative production functions are available from the authors upon request. 7 by choosing sequences of prices Pit , output cit , and labor input hit , subject to the demand function (3) and the production function. The discount factor qt = the household’s ownership of the …rm. t xt =Pt re‡ects > 0 is the price adjustment cost parameter. We assume that …rms only incur this cost when the chosen price path deviates from the weighted in‡ation rate t 1 = ( t 1) , where is steady state in‡ation and 0 1 is the degree of indexation in the targeted in‡ation rate. The …rst order conditions of …rm i are given by: Wt = Pt c t Pt Pit 1 Pit Pit 1 t ct +" y t + Pt (cit Pit y t At ; (9) Pit qt+1 Pt+1 c = Et ; Pt qt Pt t+1 qt+1 Pit+1 cit 1 ) ct cit = Et qt Pit (10) Pit+1 Pit t+1 Pt+1 ct+1 : Pit (11) The …rst condition equates the real wage to the marginal product of the worker, which is simply productivity with linear production . y t is the Lagrange-multiplier on the production function. It can be interpreted as the real marginal cost. To see this, denote the real wage wt = Wt =Pt . Total cost of production is wt hit = wt y t, to yit , to get mct = wt =At = can now be used to eliminate yit At . Take the derivative with respect where the last equality follows from (9). This expression y t from the optimality condition (10), which becomes an expectational di¤erence equation in c t, with the driving variable being marginal cost. The second …rst-order condition connects marginal cost the demand function (3). c t y t with c t, the multiplier on can be interpreted as the marginal value of demand. In the absence of deep habits (when = 0), it equals relative prices minus marginal cost. Under deep habits, however, the persistence of demand for individual goods a¤ects …rms’demand for labor intertemporally. Finally, the third …rst-order condition captures the optimal pricesetting problem of the …rm. We now derive the NKPC from this equation. 2.3 Deriving the NKPC The …rst step is to impose a symmetric equilibrium. That is, we assume in line with the literature that all …rms behave identically and are ex-post homogeneous. This amounts to erasing the …rm-speci…c subscripts i, which simpli…es the above expressions considerably. We now de…ne the aggregate, consumption-based (gross) in‡ation rate t = Pt =Pt 1. Substituting in the stochastic discount factor qt results in the following expression: " ct xt + t( t t ) ct = ct + Et 8 xt+1 xt t+1 ( t+1 t+1 ) ct+1 : (12) This is an expectational di¤erence equation in cross-products of in‡ation t and consump- tion ct . We identify as driving forces the terms involving xt and the marginal value of demand c t. Our goal is now to re-write this expression in terms of marginal cost and potentially other variables. As is common in the literature, we consider a linearized version in terms of deviations from the steady state. Denote the (log-) deviation of a variable zt from its steady state z as zet = log zt log z. We can now linearize (12) around its steady state, whereby we note that the resulting relationship is independent of the steady-state in‡ation rate up to …rst order, since …rms face price adjustment cost only to the extent that their prices deviate from the aggregate price path t. The expression for marginal value of demand c t can also be linearized around its steady state. We substitute these into the linearized form of (12) and collect terms.3 The NKPC in the model with deep habits is: et = Et et+1 + 1+ et 1+ 1 ft 1 mc + where the coe¢ cients are given by: 1 = "(1 ) (1 (1+ ) ) ; 2 = It is straightforward to verify that itive if " > (1 (1 ) =(1 ) = (1 1 ); (1+ 2, 2 Et 3 4 e ct+1 (1+ )1 (1+ ) = and 3 3 ; ec 4 Et t+1 ; e ct 4 = (1+ (13) ): are strictly non-negative. 1 is pos- ). We impose this condition henceforth. The critical value ) increases monotonically with .4 For 2 (0; 0:99), the critical value is below 2 for changes as + = 0, the critical value equals 1. For = 0:99. Figure 1 shows how (1 varies between 0 and 0.99 with ) = (1 set at 0.99. There are a few observations to make. First, in the absence of deep habits, when it can be easily veri…ed that 2 = 3 = ) 4 = 0, = 0. The relationship thus reduces to the standard NKPC derived in Galí and Gertler (1999): et = 1+ Et et+1 + 1+ et 1 + " 1 mc f t; (1 + ) (14) so that the speci…cation with deep habits cleanly nests the standard speci…cation. Second, the introduction of deep habits a¤ects the conditional responsiveness of in‡ation to marginal cost. It is straightforward to show that 1 < " 1 (1+ ) , the standard NKPC-coe¢ cient. Ceteris paribus in‡ation under the deep habits formulation is less reliant on marginal cost as driving process. 3 4 Details of the derivation are given in the Appendix. It is easy to show that the derivative of (1 ) =(1 9 ) with respect to is positive. The third observation is that deep habits add additional terms to the NKPC, above and beyond marginal cost, to wit, expected and current consumption growth and the expected marginal value of future demand. They stem from the fact that …rms have to consider the e¤ect of current pricing decisions on future demand through its feedback via the persistence of demand. Contemporaneous consumption growth e ct engenders future consumption growth via deep habits formation, which …rms encourage by lowering their prices. Higher c expected marginal value of future demand Et et+1 reduces current in‡ation since it creates an incentive for …rms to lower their prices in order to capture future market share (Ravn et al., 2010). In contrast, higher expected future consumption growth raises current in‡ation as …rms do not have to lower their prices to generate an increase in future demand. We …nd it useful for comparison with the standard NKPC to rewrite equation (13) in a slightly di¤erent way by factoring out a common coe¢ cient and by grouping the relevant terms together: et = 1+ We refer to Et et+1 + = coe¢ cients are ~ i " 1 (1+ ) 1 1+ et 1 + h ~ mc ~ ct+1 1 f t + 2 Et e ~ 3 e ct i ~ Et e c 4 t+1 : (15) > 0 as the NKPC-coe¢ cient. The expressions for the responsei, i = 1; :::; 4 which preserves the sign restrictions that we impose on the original coe¢ cients. Factoring out the coe¢ cient Note that if thus allows us to de…ne the driving process det as: det = ~ 1 mc f t + ~ 2 Et e ct+1 ~ 3 e ct ~ Et e c : 4 t+1 (16) = 0, we have det = mc f t , and the speci…cation reduces to the standard NKPC. We treat (15) as our benchmark speci…cation for the NKPC under deep habits. Summarizing the additional regressors in terms of a driving process allows us to compare it directly to the driving process in the standard NKPC, namely marginal cost. The remainder of our paper is concerned with computing this driving process. The key challenge is to determine c the behavior of the unobserved term Et e . Once we derive a series for det we can then t+1 assess the performance of the NKPC under deep habits in a limited information setting. 3 In‡ation Dynamics and Deep Habits: A Limited Information Approach We now proceed to a formal empirical analysis of the NKPC under deep habits. We pursue a limited information approach in that we do not use all the information available in the full 10 general equilibrium model that embeds the NKPC. To be more precise, we do not impose the cross-equation and cross-coe¢ cient restrictions on the comovement of the endogenous variables that the full model would prescribe. We thus treat the NKPC as a moment condition which we estimate with a GMM approach. We begin with a short description of the data and our estimation method. We then describe how to deal with unobservable variables in the formulation of the driving process by backing them out of intertemporal optimality conditions. We apply two methods: First, our benchmark speci…cation, which treats observable, but exogenous processes as separate univariate processes, and, second, a VAR-based method. 3.1 Data and Empirical Approach We extract quarterly data from the FRED database at the Federal Reserve Bank of St. Louis. Our sample period ranges from 1955:1 to 2011:2, but we also consider a sub-sample from 1984:1 onwards, which covers the Great Moderation during which the behavior of many macroeconomic time series changed. Output and consumption are constructed by dividing real GDP in chained dollars (GDPC96 in FRED) and real consumption in chained dollar (PCECC96) by the civilian non-institutional population aged 16 and over (CNP16OV). The GDP implicit price de‡ator is our measure of Pt (GDPDEF). Real unit labor cost is constructed by dividing nominal unit labor cost of the nonfarm business sector (ULCNFB) by the price de‡ator. We remove a linear trend from GDP and consumption. We also use compensation per hour in the nonfarm business sector (COMPNFB) as a measure of the nominal wage to construct wage in‡ation, which we then use as an instrument in the GMM estimation. Table 1 reports some moments of the data series. Over the full sample period, GDP is more volatile than real unit labor cost, which has been used in the literature as a proxy for marginal cost. Concentrating on the sub-sample from 1984 on, we …nd, however, that the volatility of GDP drops, while that of the marginal cost proxy increases, to the e¤ect that the latter becomes now more volatile than the former. In‡ation and marginal cost are mildly positively correlated, both for the full sample and for the sub-sample. This pattern is almost a requirement for the validity of the NKPC as the logic of the relationship suggests that increases in marginal cost should drive up in‡ation. We also note that the correlation pattern between GDP and real unit labor costs is negative for the full sample but quite positive for the period of the Great Moderation. Finally, we report second moments for consumption growth as it appears in the driving process of the deep habits speci…cation. It 11 comoves negatively with in‡ation over both the full sample and sub-sample. Since current consumption growth a¤ects in‡ation negatively per equation (16), this suggests that deep habits have a role to play in explaining in‡ation dynamics. Let zt denote a vector of variables observed at time t. The NKPC then de…nes a h i e set of orthogonality conditions: Et et e e d zt = 0. Given these t f t+1 b t 1 conditions, we can estimate the model using a GMM approach. To aid comparison with the recent literature, we use the same set of instruments as Galí et al. (2005). Speci…cally, we use 4 lags of in‡ation and 2 lags of the regressors and wage in‡ation as instruments. The weighting matrix is computed from the estimated heteroskedasticity- and autocorrelationadjusted (HAC) covariance matrix, where the number of lags in the HAC estimation is chosen based on the criterion in Andrews (1991). We consider two empirical speci…cations: …rst, a reduced-form version which estimates only the coe¢ cients in the moment conditions. The focus here is on the relative importance of the backward and forward in‡ation terms, and thus the degree of intrinsic price dynamics, and on the NKPC-coe¢ cient , which captures the strength of the transmission mechanism between the real and nominal side and indicates the presumed degree of price stickiness. The second version attempts to estimate the underlying structural parameters of the model embedded in the reduced-form coe¢ cients. 3.2 Computing the Driving Process (I): Baseline The baseline speci…cation we intend to estimate is: et = f Et e t+1 + b et 1 + det ; (17) where the driving process det is given by equation (16). The advantage of this speci…cation is that the coe¢ cient estimates are immediately comparable to those from the standard NKPC. The only di¤erence is that allowing for deep habits a¤ects the nature of the driving process, which is no longer marginal costs alone, but a composite of marginal cost, expected and current consumption and the expected marginal value of future demand. Once we have constructed a times series for det , estimating this speci…cation is straightforward. However, this approach presents a few challenges. First, marginal cost is unobservable to the econometrician. We proxy mc f t with real unit labor cost in line with most of the NKPC literature. Second, while we can compute current consumption growth straight from the data, expected consumption growth is unobservable. This can be obtained in several ways. One possibility is to proxy expected consumption using survey data from sources 12 such as the Survey of Consumer Finances. This approach has numerous drawbacks, such as the potential inconsistency of forecast horizons and forecast object between the model and the survey respondents. Second, we can try to back out expected consumption from other equilibrium conditions. However, as it turns out, all possible relationships involve unobservable variables, speci…cally, Lagrange multipliers that also would have to be proxied. We therefore choose to specify a parametric model for consumption, which we use to compute conditional expectations. In this section, we specify univariate processes for the observable variables, while the subsequent section assumes a multi-variate relationship that allows for richer interactions. The third challenge to computing the driving process det is the presence of the term c Et e , which involves the conditional expectation of an unobservable quantity, namely the t+1 marginal value of future demand. We follow the approach of Lubik and Teo (2011) and use an intertemporal equilibrium condition to relate the unobservable expected marginal value of future demand to observables. We …nd such a relationship in the …rm’s …rst-order condition (10). Using the relationship mct = y t, and linearizing (10) around the steady state yields:5 ec = t c Et et+1 Et 1 ( e ct+1 e ct ) [" (1 ) )] mc f t: (1 (18) c This is an expectational di¤erence equation in et which can be solved forward.6 We …nd that: ec = t = 1 X ( )j Et ( )j Et j=0 1 X j=0 1 1 ( e ct+j+1 1 e ct+j ) 2 e ct+j+1 2 [" (1 [" (1 ) ) (1 (1 )] mc f t+j )] mc f t+j e ct : (19) 1 The last equality follows from collecting terms in consumption growth. The marginal value + of demand can now be expressed as a function of observable e ct and as the present dis- counted value of future consumption growth and marginal cost. Proxying the latter by real c unit labor cost, we can now back out et from a parametric model for the two series, e ct and mc f t. As a …rst pass, we assume that both variables follow AR(1)-processes: 5 6 mc ft = ft 1 mc mc The derivation is shown in the Appendix. Since 0 < < 1, the equation has a unique solution. 13 + "mc;t ; (20) where j mc ; cj that Et mc f t+j = e ct = < 1 and "mc;t and " j f t, mc mc c;t e ct c 1 +" c;t ; (21) are i:i:d: random variables with zero mean. Noting we can substitute this into the above expression and solve out the in…nite discounted sum. This results in the following expression: ec = t 2 1 1 1 c 1 2 1 c e ct " (1 ) 1 (1 ) mc mc f t; (22) which is a weighted average of consumption growth and marginal cost. Since the latter can be proxied by unit labor cost, this expression allows us to impute a time series for the c unobservable et = ct f t. c e mc mc The conditional one period-ahead forecast can then be computed by iterating forward one more time: c Et et+1 = c c e ct f t: mc mc mc (23) The sign of the coe¢ cients depends on the size of the deep habits parameter. since we impose " > (1 if > mc is positive ) = (1 ). Furthermore, it is easy to show that c > 0 e c . Whether movements in Et t+1 reinforce or dampen movements of the other variables in the driving process thus depends on the size of the habit parameter and whether consumption growth and marginal cost comove positively or negatively.7 Before we can make further empirical progress, however, our …nal step assigns numerical values to the structural parameters. In our benchmark exercise, we calibrate all parameters required to impute the unobservable series since we focus on the impact of changes in the driving process only. In a robustness exercise below, we show how to use the previous expression to compute a reduced-form representation of the NKPC, which allows us to estimate some of these parameters. The calibrated parameter values are detailed in Table 2. We base our calibration on the estimates in Ravn et al. (2010), which is to the best of our knowledge the …rst empirical study of a deep habits model. We …x the discount factor = 0:99 to be consistent with an annual real interest rate of 4%. We set the coe¢ cient of relative risk aversion to which implies log utility. As an alternative, we consider a value of = 1, = 3, which implies much more risk averse households. We follow again Ravn et al. (2010) in choosing the habit parameter = 0:85. The substitution elasticity between di¤erentiated products ", which can be interpreted as a demand elasticity is set at " = 2:48, based on the empirical estimates in Ravn et al. (2010). In standard models without deep habits this parameter is Recall that the coe¢ cient on the expected marginal value of future demand in (15), ~ 4 , is strictly positive. 7 14 usually …xed at " = 11 to imply a markup of 10% over marginal cost. In our model with deep habits, the steady state markup is given by "(1 "(1 ) ) (1 8 ). Figure 2 plots the steady state markup as " varies for the standard model without deep habits and our deep habits model. and are …xed at 0.99 and 0.85, respectively. The markup is slightly higher for a given value of " in our deep habits model compared to the standard model without deep habits. In our deep habits model, " = 11 implies a steady state markup of 10.6% instead of 10%. The di¤erence arises because for a given value of " the presence of deep habits makes demand less elastic, giving …rms incentive to charge a higher markup. Our benchmark " = 2:48 imposes a steady state markup of 74:2%, which may seem excessive. We discuss this assumption further in the robustness section, where we also investigate alternative values. Finally, we …t separate AR(1)-processes to consumption growth and real unit labor costs. This results in estimates of which satis…es the restriction > c for the coe¢ cient c mc = 0:98 and c = 0:31, to be positive. Figure 3 depicts the constructed driving process and marginal cost. The former has a standard deviation of 5.68%, which is higher than marginal cost. The correlation of the two series is 0:57. This con…rms that the introduction of the additional elements into the NKPC via deep habits renders the driving process more volatile. This is also re‡ected in the less than perfect comovement, since some elements of det enter the driving process with negative signs, as the previous discussion has shown. Nevertheless, what we cannot distinguish at this stage is whether the changed properties of the driving process are simply due to the increased number of regressors or to the changed responsiveness of the coe¢ cients. We attempt to disentangle this further below. 3.3 Computing the Driving Process (II): A VAR Approach In the previous section we used independent AR(1) processes for marginal cost and conc sumption growth as predictors for the behavior of Et et+1 . In order to capture potential additional information in the data, we alternatively pursue a VAR-based approach. Con- sider a generic data vector vt , which contains consumption growth, marginal cost and other variables that we judge useful for forecasting. Assume that vt is described by a VAR: vt = Avt 1 + t .9 The conditional forecast is then given by Et vt+j = Aj vt , j note the extraction vector for some element at of the vector vt as Et ( e ct+j ) = 8 cA jv t and Et (mc f t+j ) = mc A jv . t a, 1,8t. De- so that, for instance, Estimating the VAR and the coe¢ cient See the Appendix for a derivation. The …rst-order speci…cation is without loss of generality since any higher-order VAR can be written in …rst-order companion form. We discuss this speci…cation for expositional expediency. 9 15 b therefore allows us to construct a time series for conditional expectations of the matrix A variables of interest. We can use these expressions in the equation for the expected marginal value of future demand. Following the same steps as above, this yields: c Et et+1 = 1 1 2 + 1 b c Avt : 2 b2 cA [" (1 ) (1 )] b mc A h I i b A 1 vt (24) Given this expression, observed consumption growth e ct , expected consumption growth b t , and our proxy for marginal cost mc f t , we can now construct an imEt e ct+1 = c Av e puted time series for the driving process dt from equation (16). As before, we impose our benchmark calibration, that is, = 0:85, " = 2:48, and = 1. We estimate a VAR(4) in consumption growth, real unit labor cost and output growth to construct expectations. The constructed driving process from the deep habits speci…cation is depicted in Figure 4, together with the marginal cost proxy. The standard deviation of deV AR is 4:82%, whereas t that of real unit labor cost is 3:44%. The correlation of the two series is 0:45. Both numbers are lower than the corresponding values from the baseline speci…cation with independent AR(1) processes for expected marginal cost and consumption growth.10 Nevertheless, the imputed driving process exhibits substantial volatility. Figure 5 depicts the imputed series c for Et et+1 against marginal cost. The contemporaneous correlation of both series is 0:94, while the standard deviation of the expected marginal value of future demand term is 2:27%. Since the latter term enters the driving process (16) with a negative sign, and its own coe¢ cient is positive, the negative correlation thus imparts positive comovement with marginal cost and reinforces its contribution to the driving process. 4 Estimating the NKPC We now provide formal estimates of the NKPC using a GMM approach. Our benchmark speci…cation relies on the use of constructed driving processes. We …rst estimate a standard NKPC, where we use real unit labor cost as a proxy for marginal cost. We then estimate the corresponding NKPC, where the driving process is imputed from the …rst-order conditions of a deep habits model, using the two methods described in Section 3. 10 This is reminiscent of the …nding in Lubik and Teo (2011), where the use of a VAR-based imputation process tends to smooth out the present discounted value much more than simple univariate processes. 16 4.1 The Standard NKPC In order to provide a benchmark for our deep habits speci…cation, we …rst estimate both unrestricted and restricted versions of the standard NKPC with a proxy for marginal cost as in the original model of Galí and Gertler (1999). Speci…cally, we estimate the following standard NKPC speci…cation: et = f Et e t+1 b et 1 + + mc f t: (25) The GMM-estimation results for the standard NKPC are reported in Table 3. The estimates are quite similar to those found in the literature and statistically signi…cant throughout. In the fully unrestricted speci…cation, the coe¢ cient the coe¢ cient b f on expected in‡ation is 0:79 while on lagged in‡ation is 1=5, which is consistent with the …ndings of Galí and Gertler (1999) and subsequent work. The coe¢ cient on marginal cost = 0:004. The J-test for overidentifying restrictions does not reject the speci…cation, as evidenced by a high p-value. When of b is restricted to zero, is estimated to be 0:989, while the estimate f increases by 50 percent to 0:006. At the same time, the p-value for the J-test increases, which suggests that the speci…cation without indexation is preferred. Next, we estimate the structural parameters of the NKPC. From (14), the coe¢ cient on expected in‡ation f = =(1 + Note that when = 0 the speci…cation reduces to the purely forward looking NKPC. The slope coe¢ cient = (" ), while the coe¢ cient on past in‡ation 1)=[ (1 + )]. We impose b = =(1 + ). = 0:99 on the estimation, which is consistent with the implied value from the restricted NKPC estimation. We also note the parameters in the coe¢ cient 0 = (" 1)= are not separately identi…able in this speci…cation as the coe¢ cient simply scales the marginal cost term and appears nowhere 0 else. We therefore only report estimates for and the indexation parameter . The results are in the last line of Table 3. We note that the high p-value of the J-statistic suggests that the cross-coe¢ cient restrictions are informative in the estimation. The estimate corresponds to an implied backward-coe¢ cient b = 0:27 = 0:21, which is consistent with the reduced-form estimate. 4.2 The NKPC with Deep Habits We now estimate the NKPC speci…cation with deep habits: et = f Et e t+1 + b et 1 + det ; (26) where the driving process det is either imputed using independent AR-processes for the observables or from a VAR-based approach. We note again that the speci…cation of the NKPC 17 is such that we only vary the term det . Estimates for the forward-looking coe¢ cient backward-looking coe¢ cient b and the NKPC-coe¢ cient f, the thus allow us to make direct comparisons between the models. We also report results from the structural speci…cation: et = 1+ Et et+1 + 1+ et 1 + " 1 e dt : (1 + ) (27) Not all parameters in this speci…cation are identi…able, however. We therefore focus on the indexation parameter and the price adjustment cost parameter , and …x the remaining parameters. Speci…cally, we set = 0:99 and " = 2:48, following the estimates reported in Ravn et al. (2010). The calibrated parameter values that go into the imputed driving process det are as reported in the previous section. We will consider alternative calibrations in our robustness analysis. Table 4 contains the GMM estimates for the deep habits NKPC when the marginal cost and consumption growth processes are assumed to be independent AR(1)-processes. Compared to the standard NKPC, two observations stand out. First, the degree of indexation and thus the weight on the lagged in‡ation term is much lower for the deep habits speci…cation. In the unrestricted version, NKPC. Restricting b b = 0:10 is barely half as big as in the standard to zero results in an estimate for implied value for the discount factor f = 0:984, which is identical to the = 0:99 in the restricted speci…cation. Second, the p-value for the J-test rises substantially from the unrestricted to the restricted speci…cation. The respective p-values are also much higher than the corresponding values for the standard NKPC. When interpreted as a speci…cation test, this suggests that the deep habits NKPC captures in‡ation dynamics exceedingly well with only a minor degree of intrinsic in‡ation persistence. This is also re‡ected in the structural estimates. The fraction of price-indexing …rms is estimated at a highly signi…cant coe¢ cient of b = 0:11, which translates into a backward-looking = 0:099. Moreover the p-value of the J-statistic is almost one, which sug- gests an excellent …t, based on the information content in the cross-coe¢ cient restrictions. Finally, the estimated NKPC-coe¢ cients are statistically signi…cant and almost twice as large as those for the standard NKPC. We now vary the speci…cation for the driving process and use deVt AR which has been constructed from the VAR-based forecasts. The estimation results are reported in Table 5. Conceptually, the estimates do not di¤er from those based on AR-forecasts. In fact, the …t of the model is improved in the case of the unrestricted speci…cation where the coe¢ cient on lagged in‡ation comes in at a statistically insigni…cant b = 0:034. The estimates for the restricted and structural speci…cation are essentially unchanged from before. This suggests 18 that the performance of the deep habits NKPC does not rest solely on the speci…cation of the forecasting model for variables that are extraneous to the in‡ation dynamics equation.11 The more important aspect is the fact that introducing deep habits imparts additional regressors into the driving process. We will take up this issue again in the robustness section. The conclusion we can draw from our benchmark analysis is straightforward. Deep habits dramatically improve the performance of the NKPC in describing in‡ation dynamics. We demonstrated in the previous section that the implied driving process generated from our model is more volatile than marginal cost. We …nd that this improves the …t of the model in terms of a standard J-type speci…cation test, but it also much reduces, even negates, the role of indexation in price-setting for explaining in‡ation dynamics. This device was introduced by Galí and Gertler (1999) in order to better capture in‡ation persistence through an intrinsic, that is, built-in, mechanism. We show that this role is played by the process for consumption growth in the driving process above and beyond indexation in price-setting. 5 Robustness We assess the robustness of our conclusion in three directions. First, we consider the role that calibration plays in generating the desired stochastic properties of the driving process. In the second exercise, we look at an alternative speci…cation for the driving process. Instead of …rst imputing a time series for the driving process, which is then used as a single regressor, we use the theoretical model restrictions to derive an alternative representation for a second set of observable driving forces. This allows us to decompose the e¤ects on the NKPC into two underlying forces, namely marginal cost and consumption growth. Speci…cally, we use the imputed representations for expected consumption growth and the expected marginal value of future demand, substitute them into the driving process, and thus generate a reduced-form in marginal cost and consumption growth. We then use these as independent regressors in the NKPC. Finally, we also look at the performance of the model for a subsample of our full data set that considers only the period of the Great Moderation from 1984 on. 11 This stands in contrast to the results in Lubik and Teo (2011), where the speci…cation of the forecasting model matters. 19 5.1 Alternative Calibration Our …rst robustness check simply looks at the implications of di¤erent parameterizations on the behavior of the driving process. We focus on three parameters, namely the coe¢ cient of relative risk aversion , the demand elasticity ", and the habit parameter . Table 6 contains the GMM estimation results for the imputed driving processes under di¤erent calibrations. We report only estimates for the unrestricted speci…cation where we estimate reduced-form coe¢ cients and where the driving processes are constructed using AR-based forecasts. Estimation results for the restricted version, the structural NKPC which estimates the model’s parameters and the VAR-based driving process o¤er overall consistent results. The …rst experiment documents the sensitivity of the model to the size of the habit parameter. When = 0:15, the NKPC estimates are in between those of the benchmark calibration in Table 4 and the standard NKPC in Table 3. The estimates of f and are larger than those of the standard NKPC but smaller than those of the benchmark calibration of = 0:85. The reverse is true for the estimate of b. Moreover, the standard errors are wider, and the J-test statistic has a lower p-value compared to the results of the benchmark calibration. Going to the other end of the parameter range, when estimate of f = 0:95, the is somewhat larger than that of the benchmark calibration but the J-test statistic has a lower p-value. We also experimented with intermediate values of that the performance of the deep habits NKPC improves notably for values of and …nd above 0:6, the intuition of which we discussed in Section 2. The highest p-value is in fact attained for our benchmark calibration. The second experiment varies , but keeps other parameters at their benchmark values. For = 3, when agents are more risk-averse, the e¤ect is to increase the weight on the forward-looking coe¢ cient. However, the J-test statistic has a lower p-value compared to the benchmark calibration. Finally, we also consider increasing the demand elasticity " to 11, which is the value most commonly used in the calibration literature implying a markup of 10 percent in standard model without deep habits. In this case, the estimates of and f are slightly smaller and the …t of the model worsens compared to the case of the benchmark calibration. Nonetheless, even with " = 11, the J-test statistic for the deep habits models still has a higher p-value than in the case of the standard NKPC. When considering joint variations of the parameters, the strongest role is played, unsurprisingly, by the habit parameter, while small values of " improve performance. This suggest that the benchmark calibration, which has been chosen based on the empirical estimates in Ravn et al. (2010), does provide good estimates of the underlying parameters. 20 5.2 An Alternative Reduced-Form Speci…cation Our benchmark speci…cation relies on an imputed series for the driving process, which we treat as a single regressor. However, the NKPC speci…cation in equation (13) highlights the fact that the introduction of deep habits changes the standard NKPC in two fundamental ways. First, it a¤ects the responsiveness of in‡ation to marginal cost as captured by the coe¢ cient 1, which is di¤erent from the standard NKPC coe¢ cient. Second, it adds additional regressors to the in‡ation equation, namely current and expected consumption growth and the expected marginal value of future demand. Consider the structural representation for the NKPC from equation (13), which we reproduce here for convenience: et = 1+ Et et+1 + 1+ et + 1 ft 1 mc + 2 Et e ct+1 3 c ec 4 Et t+1 ; e ct (28) What prevents direct estimation of this relationship is that Et et+1 is not observable to the econometrician.12 We can, however, use the tools developed in Section 3 to provide analytical expressions for these two components of the driving process. For purposes of exposition, we focus on the univariate representation for the driving forces. These expressions depend on marginal costs and current consumption growth. Substituting them into the NKPC results in a reduced-form speci…cation that only depends on observables. The exact derivations can be found in the Appendix. We thus have: where we treat mc et = and c f Et e t+1 + b et 1 + ft mc mc c e ct ; (29) as reduced-form coe¢ cients. We also consider a representa- tion that factors out the standard NKPC-coe¢ cient and thus imposes the cross-coe¢ cient restrictions implied by theory: 12 et = 1+ Et et+1 + 1+ et 1+ " 1 h~ mc ft (1 + ) 5 ~ 6 i e ct : (30) As we discussed before, there are alternatives to our approach, one of which involves combining future in‡ation and consumption treating them as joint elements in the moment condition. This raises issues of normalization in the estimation, which is well known to be problematic in empirical NKPC models. Moreover, it does not solve the problem with fundamentally unobservable marginal value of demand, for which we would have to use a parametric model in any case. We therefore chose to be fully parsimonious in that we treat all non-in‡ation variables in the NKPC as pure elements of the driving process. A comparison of these additional alternative approaches would be a worthwhile exercise. 21 The coe¢ cients on mc f t and ~ ~ 5 6 e ct are, respectively, given by: = = " (1 ) " 1 " 11 (1 1 1 ) 1+ 1 ( ; mc c) 1 : c We also show in the Appendix that both coe¢ cients are positive for plausible calibrations. Since marginal cost and current consumption growth are negatively correlated, this implies that the latter reinforces the impact of marginal cost on in‡ation. Both speci…cations show that the main e¤ect of deep habits is via introducing an additional regressor in the NKPC, namely consumption growth. We can now estimate (29) without further modi…cations. For a preliminary assessment, Figure 6 depicts the marginal cost series and consumption growth. Clearly, the latter is less volatile (see also Table 1). Any reinforcement of marginal cost on in‡ation dynamics would therefore have to be generated by the relative size of the coe¢ cient c. This is, in fact, borne out by the estimation results in Table 7. The consumption growth coe¢ cient in the unrestricted and the restricted speci…cations (where larger than mc , b = 0) is an order of magnitude which in turn is in line with the estimates from the standard NKPC. This con…rms that capturing in‡ation dynamics is thus a simple matter of adding the correct additional regressor.13 However, the …t of the model as measured by the J-test declines relative to both the standard NKPC as well as our benchmark speci…cation. We now turn to estimating the alternative speci…cation (30), which imposes the crosscoe¢ cient restrictions. We …x benchmark calibration that sets and ", and estimate = 0:99, and . As before, we consider a = 1 and " = 2:48, following the estimates reported in Ravn et al. (2010). We report the estimation results in Table 7, where we also consider variations in key parameters. The estimates for the indexation parameter are identical across the three speci…cations where we vary the …xed parameters. A value of = 0:198 corresponds to b = 0:166, and is thus consistent with the reduced-form estimate. There are minor di¤erences for the other parameters in the two speci…cations that set " = 2:48. In that case, the value of the deep habits parameter is close to 0:85, which is the value estimated by Ravn et al. (2010) with a di¤erent empirical methodology and within the context of a fully-speci…ed general equilibrium model. For the speci…cation with 13 = 3 and " = 11, we note that the habit That this need not be the case is demonstrated by empirical studies of the NKPC that modify factor inputs (e.g. by introducing labor market search and matching frictions as in Krause et al., 2008) or change the structure of product markets (e.g. by introducing …nished goods inventories as in Lubik and Teo, 2011). 22 parameter is close to one and that the price adjustment cost parameter a very high value.14 The latter estimate re‡ects the fact that " and is estimated at are not separately identi…able. The higher calibrated value of the elasticity parameter " then translates into higher implied adjustment cost in order to generate the same implied NKPC-coe¢ cient. At the same time, the increase in is related to the increase in ", which also suggest identi…cation issues regarding these two preference parameters. We investigate this issue further by looking at how the driving process changes with respect to parameters. Figure 7 plots the marginal cost coe¢ cient ~ 5 as well as the relative weight on consumption growth ~ = ~ against the habit parameter over the range [0; 1]. 6 5 The graphs are conditional on " = 2:48. Analytically, the coe¢ cient on marginal cost is not a¤ected by , which only enters the weight on consumption growth (see equation 30). The bottom graph therefore contains two lines for di¤erent values of the weight on marginal cost decreases. For obtains. ~ is fairly inelastic to changes in 5 = 1; 3. As increases, = 0, the weight is one and the standard NKPC until about 0.6, after which it declines rapidly. Nevertheless, the weight on marginal cost drops below 0:5 only for extremely high values of the habit parameter. The relative weight on consumption growth is the ‡ip-side of this. For > 0:6, ~ = ~ exceeds 1 and increases exponentially afterwards. For values larger 6 5 than 0.9, the weight on consumption is 10 times that on marginal cost. Higher values of increase the relative weight on consumption growth, but the two curves are close enough to not make this the dominant e¤ect. The previous graphs were plotted conditional on " = 2:48. We now study how the response coe¢ cients change with variations in that parameter for a given . The respective graphs are in Figure 8. We …x = 0:85, and also plot variations to . As before, the weight on marginal cost is not a¤ected by . As " rises over the range [2; 11], the weight on marginal cost increases only slightly. At the same time, the relative weight on consumption growth decreases by an order of 10 over the range. A higher value of increases the relative weight on consumption growth as in Figure 7. Note that the relative weights on consumption growth are all larger than 1 in Figure 8. This analysis shows where the improvement in …t over the standard NKPC is coming from. It is not simply the addition of another regressor, but the fact that deep habits increase the responsiveness of in‡ation to movements in consumption growth by an order of magnitude without a large countervailing e¤ect from a reduced importance of marginal cost. This result is, however, predicated on two requirements. First, the demand elasticity " has 14 We do not report the case of impose the parameter restriction. = 1 and " = 11 since the estimate of 23 would exceed 1 if we did not to be small enough, while the degree of deep habits has to be large enough, as is the case for the estimates of Ravn et al. (2010) and for our estimates using the alternative speci…cation of this section. Although a high degree of habits reduces the weight on marginal costs, this is more than compensated by the increase in the relative weight of consumption growth. The second requirement is that consumption growth has to have the right statistical properties. Although it is markedly less volatile than marginal costs, this is more than compensated by the response coe¢ cient in terms of its overall impact. However, for the ampli…cation e¤ect on marginal cost dynamics, consumption growth has to comove negatively with the marginal cost proxy since consumption growth shows up in the NKPC with a negative sign as per equation (29). Otherwise, deep habits would dampen the e¤ects of marginal cost. Our …nal robustness check investigates such a case. 5.3 Sub-Sample Analysis We conclude our robustness analysis by estimating the benchmark speci…cation for a subsample period that starts in 1984, which covers the period of the Great Moderation. Table 1 shows that the behavior of the series we use in this paper has in fact changed. The volatility of all variables is smaller in the sub-sample than in the full sample with the exception of real unit labor cost. This need not have a dramatic e¤ect on our estimates since the decline in consumption growth volatility is compensated by increased volatility of the marginal cost proxy. More detrimental is the change in comovement pattern between these two series; to wit, the contemporaneous correlation between mct and ct is a positive 0:23 over the sub- sample. Since we show above that the latter enters the theoretical NKPC with a negative coe¢ cient, this is likely to counter the e¤ects of marginal cost. Our concerns are only partially borne out by the estimates reported in Table 8. The standard NKPC estimates in Panel A show a shift towards a stronger weight on forwardlooking behavior compared to the full sample and an overall better …t, although the structural estimates impart a value for the indexation parameter that is much higher than in the benchmark sample. Panel B shows the estimates of the deep habits NKPC using the benchmark calibration with VAR-based forecasts. The estimates do not show dramatic di¤erences to those in Table 5. The relative …t of the sub-sample estimation is worse, but still much improved over the standard NKPC. However, standard errors of the estimates are surprisingly large, technically rendering the NKPC coe¢ cient statistically insigni…cant at the 10%-level. This arguably re‡ects the changing pattern of the comovement between the regressors. Moreover, it may also re‡ect that unit labor cost may not be the best proxy 24 for marginal cost for this sub-sample period.15 Nevertheless, even the sub-sample analysis shows that deep habits are a central component to explaining in‡ation dynamics per the NKPC. 6 Conclusion We show in this paper that deep habits in preferences are an essential element in understanding in‡ation dynamics. Compared to a standard version of the NKPC, a deep habits speci…cation is an improvement in terms of …t and in terms of smaller standard errors of the estimated parameters. The estimated NKPC under deep habits also puts much less weight on lagged in‡ation. This suggests a lower degree of intrinsic in‡ation persistence, where the required propagation is derived from the properties of the imputed driving process. The impact of deep habits on the latter stems from two in‡uences. First, the model implies additional regressors, speci…cally consumption growth and a marginal value of demand term. This in and of itself produces a better …t, but we also show that a large part of the improved performance is due to the altered responsiveness of in‡ation to the coe¢ cients in the NKPC. Deep habits therefore preserve the standard transmission mechanism from marginal cost movements to in‡ation, but reinforce this through additional feedback. The main concern about the validity of our results stems from the partial equilibrium nature of our analysis, that is, we only estimate a single equation that should naturally be seen as a part of a larger general equilibrium model. This is re‡ected in two aspects. First, the elements of the driving process are treated as exogenous regressors, albeit ones that still su¤er from endogeneity problems. This requires the use of instrumental variables in the estimation, that may themselves be of dubious quality. Estimation of the full equilibrium model with, for instance, likelihood-based methods obviates this problem since the likelihood function and the application of the Kalman-…lter automatically constructs the optimal instruments. The drawback of a systems-approach, however, is the possibility of misspeci…cation. Resolving this issue is outside the scope of this paper. We take comfort, however, from the empirical estimates of Ravn et al. (2010), which are close to ours, despite a di¤erent empirical method that does utilize more information. The second problematic aspect lies in the way we impute the unobservable variables. We rely on present-value computations that are known to impose weak restrictions on the imputed variables. It would therefore be a useful exercise to consider alternative methods for backing out unobservables. 15 See the discussion in Galí et al. (2005) and Nason and Smith (2008). 25 References [1] Andrews, Donald W.K. (1991): “Heteroskedasticity and Autocorrelation Consistent Covariance Matrix Estimation”. Econometrica, 59(3), 817-858. [2] Galí, Jordi and Mark Gertler (1999): “In‡ation Dynamics: A Structural Econometric Analysis”. Journal of Monetary Economics, 44(2), 195-222. [3] Galí, Jordi, Mark Gertler, and J. David Lopez-Salido (2005): “Robustness of Estimates of the Hybrid New Keynesian Phillips Curve”. Journal of Monetary Economics, 52, 1107-1118. [4] Guerrieri, Luca, Christopher Gust, and J. David Lopez-Salido (2010): “International Competition and In‡ation: A New Keynesian Perspective”. American Economic Journal: Macroeconomics, 2, 247-280. [5] Krause, Michael U., J. David Lopez-Salido, and Thomas A. Lubik (2008): “Do Search Frictions Matter for In‡ation Dynamics?” European Economic Review, 52(8), 14641479. [6] Kurmann, André (2007): “VAR-based Estimation of Euler Equations with An Application to New Keynesian Pricing”. Journal of Economic Dynamics and Control, 31(3), 767-796. [7] Lubik, Thomas A., and Wing Leong Teo (2011): “Inventories, In‡ation Dynamics, and the New Keynesian Phillips Curve”. Forthcoming, European Economic Review. [8] Nason, James M., and Gregor W. Smith (2008): “The New Keynesian Phillips Curve: Lessons from Single-Equation Econometric Estimation”. Federal Reserve Bank of Richmond Economic Quarterly, 94(4), 361-395. [9] Ravn, Morten O., Stephanie Schmitt-Grohé, and Martín Uribe (2006): “Deep Habits”. Review of Economic Studies, 73(1), 195-218. [10] Ravn, Morten O., Stephanie Schmitt-Grohé, Martín Uribe, and Lenno Uuskula (2010): “Deep Habits and the Dynamic E¤ects of Monetary Policy Shocks”. Journal of the Japanese and the International Economies, 24, 236-258. 26 Table 1. Business Cycle Statistics Sample P eriod : 1955 : 1 2011 : 2 Variable s.d.(%) Cross-Correlation GDP Inf lation RU LC Cons:Growth y 1 3.97 0.58 3.44 0.71 0.15 1 rulc -0.20 0.24 1 c 0.14 -0.28 -0.09 1 Sample P eriod : 1984 : 1 2011 : 2 Variable s.d.(%) Cross-Correlation GDP Inf lation RU LC Cons:Growth y 1 3.31 0.25 3.71 0.56 0.16 1 rulc 0.42 0.27 1 c 0.28 -0.11 0.23 1 Table 2. Benchmark Calibrated Parameter Values Parameter c mc De…nition Value Source Discount Factor Risk Aversion Deep Habits Elasticity of Demand AR(1)-coe¢ cient AR(1)-coe¢ cient 0.99 1 0.85 2.48 0.31 0.98 Annual Real Interest Rate Log-utility Ravn et al. (2010) Ravn et al. (2010) Authors’Estimates Authors’Estimates 27 Table 3. GMM Estimates: Standard NKPC Speci…cation Unrestricted NKPC f Restricted NKPC b =0 0.004 (0.001) J(7) 5.428 (0.608) 0.006 (0.002) J(8) 5.813 (0.668) b 0.791 (0.055) 0.197 (0.058) f 0.989 (0.014) 0 Structural NKPC = 0:99 0.271 (0.092) 0.004 (0.001) J(8) 5.333 (0.722) Note: The numbers in parentheses are standard errors. For J-statistics, the numbers in parentheses are p-values. For the structural NKPC, 0 (" 1)= : The instrument set includes 4 lags of in‡ation and 2 lags of marginal cost, output and wage in‡ation. The adjusted sample period for the estimation is 1956Q3 to 2011Q2. Table 4. GMM Estimates: Deep Habits NKPC, AR-based Speci…cation Unrestricted NKPC 0.008 (0.002) J(7) 4.236 (0.752) 0.984 (0.008) 0.010 (0.001) J(8) 3.133 (0.926) 0.110 (0.011) 203.455 (7.377) J(10) 1.068 (0.998) f b 0.857 (0.041) Restricted NKPC b =0 Structural NKPC = 0:99, = 1 " = 2:48, = 0:85 0.105 (0.045) f Note: The numbers in parentheses are standard errors. For J-statistics, the numbers in parentheses are p-values. The instrument set includes 4 lags of in‡ation and 2 lags of constructed driving process, output and wage in‡ation. The adjusted sample period for the estimation is 1956Q3 to 2011Q2. 28 Table 5. GMM Estimates: Deep Habits NKPC, VAR-based Speci…cation Unrestricted NKPC 0.017 (0.002) J(7) 2.251 (0.945) 0.990 (0.011) 0.010 (0.003) J(8) 3.133 (0.926) 0.138 (0.004) 334.825 (10.165) J(8) 1.007 (0.998) f b 0.929 (0.028) Restricted NKPC b =0 Structural NKPC = 0:99 0.034 (0.031) f Note: The numbers in parentheses are standard errors. For J-statistics, the numbers in parentheses are p-values. The instrument set includes 4 lags of in‡ation and 2 lags of constructed driving process, output and wage in‡ation. The adjusted sample period for the estimation is 1956Q3 to 2011Q2. Table 6. Robustness: Alternative Calibration Speci…cation Unrestricted NKPC = 0:15 0.825 (0.059) 0.157 (0.059) 0.005 (0.002) J(7) 6.942 (0.543) Unrestricted NKPC = 0:95 0.902 (0.051) 0.064 (0.053) 0.004 (0.001) 4.615 (0.707) Unrestricted NKPC = 0:60 0.845 (0.058) 0.126 (0.066) 0.006 (0.002) 4.921 (0.670) Unrestricted NKPC =3 0.884 (0.047) 0.071 (0.051) 0.008 (0.002) 4.565 (0.713) Unrestricted NKPC " = 11 0.841 (0.058) 0.130 (0.066) 0.006 (0.002) 4.859 (0.677) f b Note: The numbers in parentheses are standard errors. For J-statistics, the numbers in parentheses are p-values. The instrument set includes 4 lags of in‡ation and 2 lags of constructed driving process, output and wage in‡ation. The adjusted sample period for the estimation is 1956Q3 to 2011Q2. 29 Table 7. Robustness: Alternative Speci…cations Speci…cation Unrestricted NKPC f 0.825 (0.059) Restricted NKPC b =0 b 0.157 (0.059) mc 0.005 (0.002) mc f 1.002 (0.022) 0.004 (0.002) c 0.069 (0.030) J(8) 6.942 (0.543) 0.047 (0.042) J(9) 7.871 (0.547) c Structural NKPC = 0:99 = 1, " = 2:48 0.198 (0.084) 230.699 (62.834) 0.894 (0.038) J(9) 6.710 (0.667) Structural NKPC = 0:99 = 3, " = 2:48 0.198 (0.084) 254.654 (74.851) 0.843 (0.052) J(9) 6.710 (0.667) 0.956 (0.015) J(9) 6.709 (0.667) Structural NKPC = 0:99 = 3, " = 11 0.198 (0.084) 1228.934 306.754 Note: The numbers in parentheses are standard errors. For J-statistics, the numbers in parentheses are p-values. The instrument set includes 4 lags of in‡ation and 2 lags of marginal cost, consumption growth, output and wage in‡ation. The adjusted sample period for the estimation is 1956Q3 to 2011Q2. 30 Table 8. Robustness: Sub-Sample 1984:1-2011:2 Panel A: Standard NKPC Unrestricted NKPC f 0.833 (0.081) Restricted NKPC b =0 Structural NKPC = 0:99 0.004 (0.001) J(7) 4.209 (0.755) 0.005 (0.002) J(8) 5.051 (0.752) b 0.234 (0.052) f 1.146 (0.057) 0 0.003 (0.001) J(8) 2.984 (0.935) 0.009 (0.012) J(7) 3.830 (0.799) 1.073 (0.118) 0.013 (0.009) J(8) 3.881 (0.868) 0.131 (0.159) 139.677 (128.87) J(8) 4.420 (0.817) 0.402 (0.056) Panel B: Deep Habits NKPC Unrestricted NKPC f b 0.988 (0.196) Restricted NKPC b =0 Structural NKPC = 0:99 0.074 (0.140) f Note: The numbers in parentheses are standard errors. For J-statistics, the numbers in parentheses are p-values. The instrument set includes 4 lags of in‡ation and 2 lags of constructed driving process, output and wage in‡ation. The sample period for the estimation is 1984Q1 to 2011Q2. 31 2 1.9 1.8 Critical value 1.7 1.6 1.5 1.4 1.3 1.2 1.1 1 0 0.1 0.2 0.3 0.4 0.5 0.6 Degree of deep habits - Figure 1: Critical value of " for 0.7 θ 1 0.8 0.9 1 to be positive 120 Deep habit S tandard model Steady state markup (in %) 100 80 60 40 20 0 2 3 4 5 6 7 Elasticity of substitution - ε 8 Figure 2: Steady-state markup as " varies 32 9 10 11 30 AR-based driving process Marginal cost Percentage deviation from steady state 25 20 15 10 5 0 -5 -10 -15 -20 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 Figure 3: AR-based constructed driving process and marginal cost 30 VAR-based driving process Marginal cost Percentage deviation from steady state 25 20 15 10 5 0 -5 -10 -15 -20 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 Figure 4: VAR-based constructed driving process and marginal cost 33 2010 c t t+1 Eλ 10 Percentage deviation from steady state Marginal cost 5 0 -5 -10 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 Figure 5: Expected marginal value of future demand and marginal cost 10 Percentage deviation from steady state Consumption growth Marginal cost 5 0 -5 -10 1960 1965 1970 1975 1980 1985 1990 1995 2000 Figure 6: Consumption growth and marginal cost 34 2005 2010 Weight on marginal cost 1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 Relative weight on consumption growth 0.55 0 0.1 0.2 0.3 0.4 0.5 Degree of deep habits - 0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1 θ 90 σ =1 σ =3 80 70 60 50 40 30 20 10 0 0 0.1 0.2 0.3 0.4 0.5 Degree of deep habits - θ Figure 7: Weights on marginal cost and consumption growth as varies Weight on marginal cost 0.87 0.86 0.85 0.84 0.83 0.82 0.81 Relative weight on consumption growth 0.8 2 3 4 5 6 7 Elasticity of substitution - ε 8 9 10 11 25 σ =1 σ =3 20 15 10 5 0 2 3 4 5 6 7 Elasticity of substitution - ε 8 9 10 Figure 8: Weights on marginal cost and consumption growth as " varies 35 11 A Appendix: Analytical Derivations A.1 Derivation of Equation (18) and the Deep Habits NKPC Substituting the de…nitions of the habit stock xt = ct rate " t 1 = c t (ct ct ( t 1) 1 )+ ct 1 and the weighted in‡ation into equation (12), we have: t 1 t ( t 1) ct = ct + xt+1 xt Et t+1 1 t+1 ( t) ct+1 : (A1) Log-linearization of the equation above gives: c ) ~t + " c c (1 1 c~t 1 c~t 1 1 + c (~ t " c ~t 1) = c~ ct + cEt (~ t+1 ~ t ) : (A2) Using the steady-state relation: (1 ) = 1; (A2b) which is obtained from equation (A1), we can simplify equation (A2) as: ~t = 1+ ~t 1 + 1 Et ~ t+1 1+ (1 + ~c + t ) 1 (~ ct Imposing symmetry and substituting the discount factor qt , xt = ct c~t ct 1, 1) : and (A3) y t = mct into equation (10), we have: mct + c t =1+ ct+1 ct ct ct 1 Et c t+1 : (A4) Log-linearization of the equation above gives: mcmc ft + c ~c t c = Et 1 (~ ct+1 c~t c~t + c~t 1) c + ~ t+1 : (A5) Using the steady-state relation: mc = 1 (1 ) c ; (A5b) which is obtained from equation (A4) and (A2b), we can simplify equation (A5) as: ~c = t c Et ~ t+1 Et 1 ( c~t+1 c~t ) (" (1 ) (1 )) mc f t; (A6) which is equation (18) in the main text. Substituting equation (A6) into equation (A3) and rearranging, we then obtain equation (13) in the main text. 36 A.2 Derivation of Equation (22) Substituting (20) and (21) into (19), we get: ec = t = 1 X )j Et ( j=0 + 1 1 X j=0 1 X 2 1 1 2 ( e ct : ( j mc ) (" (1 c~t f t; mc mc j c) 2 1 1 c where: c )) mc ft + (1 1 2 (0; 1) and j mc j 1 c A.3 0 is > 1 2 c ; c 1 = 1 2 (A8) c " (1 ) (1 1 ) : (A9) mc 2 [0; 1), the condition for is the same condition as for )] mc f t+j (A7) 1 c 1 mc Since (1 c~t 1 1 1 = ) 2 ) 2 = [" (1 c~t c j=0 = e ct+j+1 mc > 0 is " > (1 ) = (1 ), which > 0. It is then easy to verify that a su¢ cient condition for c. Steady-State Markup Combining equations (A2b) and (A5b), we obtain: mc = " (1 ) (1 " (1 ) ) : (A10) The steady-state markup is the inverse of steady-state real marginal cost. We therefore have: Steady-state markup = 37 " (1 ) " (1 ) (1 ) : (A11) A.4 Derivation of Equations (29) and (30) Substituting (23) and (21) into (13), we have: et = = Et et+1 + 1+ e ct 3 1+ ( 3 4( Et et+1 + + c e ct c 1+ c 4 et 1+ c 2 et ft 1 mc + 1 + f t) mc mc mc +( 1 + 1 4 e ct ; c) e ct c 2 ::: ft mc mc ) mc (A12) " 1 (1+ ) , which is in the form of equation (29) in the main text. Factoring out we can write equation (A12) as equation (30), where: ~ 5 = = = ~ 6 = = = = ~ +~ 1 4 " (1 ) " " (1 ) " ~ +~ 3 4 1+ " 1 1 1 " 11 1 " 11 ~ is positive if " > (1 5 ~ to be non-negative is 6 ~ c mc mc " ) " (1 " 1 1 ) (1 ) mc 1 mc : 1 mc 11 + ( 1 ) = (1 > + 1 1+ 1+ ) c 2 + (1 1 (1 1 c, 2 1 1 c) c c c 1 c 1 2 1 c " 1 c 2 c c : c ), which we have imposed. A su¢ cient condition for which is the same condition for 38 c 0.