Letting Different Views about Business Cycles Compete
2010; University of Chicago Press; Volume: 24; Issue: 1 Linguagem: Inglês
10.1086/648305
ISSN1537-2642
Autores Tópico(s)Global Financial Crisis and Policies
ResumoPrevious articleNext article FreeLetting Different Views about Business Cycles CompetePaul Beaudry and Bernd LuckePaul BeaudryUniversity of British Columbia and NBER Search for more articles by this author and Bernd LuckeUniversity of Hamburg Search for more articles by this author University of British Columbia and NBERUniversity of HamburgPDFPDF PLUSFull Text Add to favoritesDownload CitationTrack CitationsPermissionsReprints Share onFacebookTwitterLinkedInRedditEmailPrint SectionsMoreI. IntroductionThe rise of real business cycle (RBC) models in the 1980s initiated much controversy about the main driving forces of macroeconomic fluctuations. Some 25 years later, economists have still not reached a consensus on this issue. Shocks to disembodied technology had been singled out by the RBC literature as a central element in business cycle fluctuations. In contrast, a large literature based on new Keynesian models tends to emphasize instead the importance of monetary and other nontechnology shocks in fluctuations. For example, an influential paper by Galí (1999) has suggested that surprise technology shocks may not be an important contributor to business cycle fluctuations. More recently, Fisher (2006) reframed the debate by distinguishing between shocks to disembodied and those to embodied technology. While he found the former to be unimportant indeed, he claimed that shocks to investment-specific technology (IST) are the major source of hours variance. Simultaneously, Beaudry and Portier (2006) suggested expectational shocks reflecting news about future technological developments (referred to as news shocks) as an important force behind macroeconomic fluctuations.In this paper, we aim to assess the relative importance of several candidate explanations of macroeconomic fluctuations by adopting a framework that allows them to compete. Following, among others, Galí (1999) and Fisher (2003, 2006), we use a structural vector autoregressive (SVAR) approach to explore this issue. We depart slightly from these authors by explicitly allowing for cointegration. Within this framework, we explore several alternative identification schemes that allow us to isolate five shocks commonly discussed in the literature. These are as follows: surprise changes to disembodied and to embodied technology, news shocks, monetary policy shocks, and preference shocks.Our benchmark identification scheme imposes only a few long-run restrictions since at least three of the shocks we consider (two surprise technology shocks and the news shocks) may well cause permanent effects. The first identification scheme we propose therefore relies mostly on impact restrictions. For example, the news shock is identified to be orthogonal to measures of total factor productivity (TFP) and the relative price of investment on impact but unrestricted in the long run. However, to illustrate the robustness of our results, we also work with an alternative identification scheme that imposes fewer short-run restrictions and relies more on long-run restrictions.Our baseline vector error correction model (VECM) framework is composed of five variables: measured TFP, the relative price of investment goods, an index of stock prices, hours worked, and the federal funds rate. In accordance with much of the literature, we choose hours of work as our primary measure of aggregate economic activity. We also document the robustness of our results by considering alternative measures of economic activity such as consumption, investment, and output. Following Fisher (2006), we use the relative price of investment to help identify IST shocks. Since standard deflators from the National Income and Product Accounts (NIPA) have been criticized for insufficient quality adjustment, for example, by Gordon (1989), we also work with a measure of the real price of investment based on the work of Cummins and Violante (2002) and adjust investment, output, TFP, and capital stock data accordingly. We do not find that the issue of quality adjustment matters much.Our main findings are as follows. Our two main identification schemes give very similar results. In both cases we find that neither type of surprise technology shock explains more than a small share of activity variance. The dominant force appears to be the news shock, which precedes growth in measured TFP by about 2 years. Monetary shocks, preference shocks, and in some cases surprise TFP shocks play more minor roles but are not negligible. IST shocks, on the other hand, appear to play a negligible role in fluctuations, provided the analysis allows for the possibility of news shocks reflected in stock prices. These results are shown to be robust across various modifications of the underlying dependent variables and the identifying assumptions.The paper is organized as follows: Section II presents our structural VECM framework and discusses the identifying assumptions for our two basic identification schemes. Section III describes the database, and Section IV contains the analysis of the benchmark system under both identification strategies. In Section V we modify the system to allow for improved investment good quality adjustment. Further robustness checks are in Section VI. Here we also explore why our results differ from previous studies. Section VII concludes.II. Framework of the AnalysisOur objective is to identify and quantify the relative importance of five shocks that we consider as important contenders for explaining business cycle fluctuations. These shocks are as follows: surprise shocks to TFP, surprise shocks to IST, news about future technology, preference shocks, and monetary shocks. To achieve this goal, we will work mainly with a five-variable structural vector error correction model (SVECM). Specifically, we consider an environment where a K-dimensional vector of observable variables yt is integrated of order one and can be represented as a vector autoregressive (VAR) process of order p < ∞. Allowing for r > 0 cointegrating vectors, the error-correction representation of the process is given by where α and β are K × r matrices of loading coefficients and cointegrating vectors, respectively; the $$\Gamma \prime _{j}s$$, j = 1, … , p − 1, are K × K coefficient matrices; and ut are the reduced-form error terms. These can be thought to be linear combinations of the structural shocks, εt, we are interested in. As is common in the literature, we assume that the covariance matrix of εt is the identity matrix IK. Since the covariance matrix of ut is nonsingular, there exists a nonsingular matrix B such that ut = Bεt. This matrix is not unique, and suitable assumptions must be imposed on its coefficients to identify it. The structural model, a B-model in the sense of Lütkepohl (2005), is then obtained from (1) by applying the Granger representation theorem: where y0 is a vector of initial conditions, $$L\mathrel{\mathop:}= \beta _{\perp }\left[\alpha _{\perp }\prime \left(I_{K}-\sum_{i=1}^{p-1}\Gamma _{i}\right)\beta _{\perp }\right]^{-1}\alpha _{\perp }\prime B$$ is a K × K matrix with rank K-r, α⊥, β⊥ denote orthogonal complements of α, β, respectively, and the matrices Ξ*j, j = 1, … , ∞, are absolutely summable; that is, $$\mathrm{\lim}_{\tau \rightarrow \infty }\Xi ^{*}_{\tau }=0$$. Hence, in terms of structural interpretation, L is the long-run multiplier matrix of the structural shocks εt and B is the corresponding short-run impact matrix. We have to propose and justify (at least) $$K\left(K-1\right)/ 2$$ restrictions on $$B=\mathrel{\mathop:} \left(b_{ij}\right)$$ and $$L=\mathrel{\mathop:} \left(l_{ij}\right)$$ to identify the structural shocks. Thus for K = 5, we need at minimum 10 restrictions to identify the five structural shocks of interest.1Many structural models can be approximated by the type of moving average representation given in (2), for example, most linearized stochastic dynamic general equilibrium models. To set ideas, it is useful to imagine the underlying data-generating process as potentially being derived from a representative agent model where there is a final good sector and an investment good sector, and where technology in each sector is stochastic. Moreover, the representative agent in this model economy is allowed to be subject to stochastic changes in preferences. The idea of technological news in such a setting can be captured by assuming that the representative household learns about productivity innovations before they are effectively implemented in the economy (news shocks can be interpreted as diffusion lags in technology). In a Web appendix (http://www.wiso.uni-hamburg.de/beaudry-lucke), we present an extended RBC model that incorporates all these characteristics. The illustrative model we present in that appendix is also an example of a model that satisfies the type of identification assumptions we will pursue here to recover structural shocks.Many papers, including, for example, a paper by Chari, Kehoe, and McGrattan (2008), question the plausibility of structural VAR methodology being used to identify structural shocks. For this reason, in our Web appendix we use artificial data generated from the structural model to explore whether the identification strategies we use in this paper are likely to allow identification. When the model is calibrated to deliver a variance decomposition similar to that observed in U.S. data, we find that the methodology works well.A priori, it is not obvious that our five shocks of interest can be identified; that is, it is not obvious that there exists a vector y with corresponding B and L matrices that exhibit 10 theoretically plausible restrictions. However, as we will show, by choosing the vector y carefully, the desired identification can be achieved quite easily by exploiting a set of properties that are common to most contemporary models embodying such shocks. In fact, we will advance two main identification schemes to isolate the shocks of interest. While these two identification schemes share some common restrictions, they will also differ considerably. Since in many models both of these schemes should achieve the same identification, it is of interest to know whether their empirical implementation renders similar results. If the identification schemes do lead to similar results, it will offer support to the claim that we have isolated the shocks of interest. In fact, in the robustness section we will study a broad variety of identification schemes related to the two basic settings and show that our findings are very robust across these schemes.The five observable variables on which we will base our primary analysis are as follows: measured TFP, the inverse of the relative price of investment goods, a stock market index, a measure of economic activity (such as hours worked, investment, consumption, or output), and finally the rate of interest on federal funds. Details on the construction of the variables are discussed in the next section. Intuitively, the reasons we choose these variables are that (i) measured TFP should help identify innovations to disembodied technology; (ii) the value of the stock market should help isolate news about future technological developments; (iii) the federal funds rate should help identify monetary policy shocks; (iv) we need a measure of economic activity since it is our main focus; and finally, (v) since the relative price of investment goods is modeled by most researchers as an indicator of investment-specific technological change, it therefore is likely to be helpful in identifying investment-specific technological shocks.2Since in most of the business cycle literature TFP is considered a driving force, we will exploit this property to help identify shocks. In particular, we will begin by assuming the following properties for the relationship between TFP and the structural shocks.Assumption A1. Only TFP shocks may have contemporaneous effects on TFP.Assumption A2. Preference shocks and monetary shocks have no long-run effects on TFP.Without loss of generality, if we let the order of dependent variables in the vector yt be TFP, inverse of relative investment price, stock price index, activity, and federal funds rate, and let the order of the structural εt shocks be TFP shock, IST shock, news shock, preference shock, and monetary shock, then assumptions A1 and A2 imply the identifying restrictions b12 = b13 = b14 = b15 = 0 and l14 = l15 = 0, respectively.Assumptions A1 and A2 follow directly from common assumptions regarding TFP as a driving force of economic fluctuations. In particular, it is quite natural to assume that the TFP process is independent of preference shocks and monetary shocks both in the short and long run. In addition, in the literature on IST, the process for TFP is generally modeled as independent of innovations in investment-specific technological change. With respect to news about future technological change, by definition, these shocks have no impact effects on TFP (following Beaudry and Portier [2006]) or IST but are allowed to predict long-run movements in TFP. Since measured TFP may be contaminated by changes in the price of capital, we will later explore the effect of dropping the restrictions b12 = 0.The second identification restriction we will impose in this section is that monetary shocks affect economic activity only with delay, as stated under assumption A3. This assumption has been widely used in the literature aimed at identifying the effects of monetary disturbances (cf., e.g., Bagliano and Favero 1998). Since we want to be consistent with this literature, we maintain this assumption. Assumption A3 yields the identifying restriction b45 = 0.Assumption A3. Monetary shocks do not have a contemporaneous effect on economic activity.Assumptions A1, A2, and A3 provide seven restrictions. To identify the five shocks of interest we therefore need at least three additional restrictions. We will begin by suggesting two sets of additional restrictions. In both cases, these restrictions will exploit properties of the relative price of investment. Our first approach is to examine impact restrictions implied by the literature that incorporates investment-specific technological change into macro models. In most of this literature, the final good can be transformed to investment goods using a linear technology, and it is shocks to this linear technology that are referred to as IST shocks. The market implementation of this technology implies that the relative price of investment goods in terms of consumption goods reflects the IST. Since the process for investment specific technological change is modeled as a process driven by one shock, it follows that monetary shocks, preference shocks, and news shocks should have no contemporaneous effects on the relative price of investment goods. This feature is captured by assumption B1.Assumption B1. News shocks, preferences shocks, and monetary shocks have no contemporaneous effects on the relative price of investment.Assumption B1 implies the restrictions b23 = b24 = b25 = 0. The combination of assumptions A1, A2, A3, and B1 provides sufficient theoretical restrictions for isolating the five shocks of interest. As we shall later show, adding the restriction b21 = 0 to this system—which becomes an overidentifying restriction—is not rejected by the data and does not alter results. We will refer to the identifying scheme embodying assumptions A1, A2, A3, and B1 as ID1. The restrictions associated with ID1 are summarized below, where the set of restrictions on matrices B and L is shown explicitly: Here, starred entries denote unrestricted elements of B and L. Note that under ID1 the news shock is identified by postulating zero effects on both types of technology on impact but allowing for unrestricted long-run effects. Thus, under this identification scheme news can be news about both TFP and IST innovations. Similarly, under ID1, the notion of a preference shock can be given a far more general interpretation than the term may suggest. For example, our identification strategy is compatible with the preference shocks representing any kind of temporary nonmonetary demand shocks (e.g., increases in government spending or foreign demand) or with changes in market structure (e.g., transitory changes in markups). It is also compatible with nontechnology expectational shocks (e.g., socially inefficient market rushes in the sense of Beaudry, Collard, and Portier [2006] or even sunspot shocks and bubbles). Thus, while we label this shock a "preference" shock, the rather weak identifying assumptions for this shock allow it to stand in for any nonmonetary shock that is orthogonal to technology on impact and has no long-run effect on TFP. One of the attractive features of ID1 is that it mainly relies on impact restrictions and therefore is less likely subject to the criticism presented in Chari et al. (2008) regarding the use of long-run restrictions.Most models that incorporate IST assume that the relative price of investment reacts only to investment-specific technological shocks. Our identification scheme ID1 imposes considerably weaker restrictions; for example, the relative price of investment can react to any shock with a lag. Nevertheless, ID1 might be criticized for ruling out that news, preference, or money shocks change the relative price of investment on impact. For example, if it is the case that there are adjustment costs associated with investment, then the relative price of investment may vary in the short run with any shock that increases investment. If this is the case, assumption B1 would not be valid. For this reason, it appears desirable to search for an alternative identification scheme that is not subject to this criticism.An alternative means to identify the shocks of interest is to drop assumption B1 and instead focus on long-run restrictions that models impose on the relative price of investment. This approach is very similar to that proposed in Fisher (2006). In most of the literature incorporating investment-specific technological change, investment-specific shocks are the sole driver of the long-run behavior of the relative price of investment goods. This property will also hold in models where there are adjustment costs to investment and therefore the previous criticism does not apply. Hence it is natural, at a minimum, to assume that monetary shocks and preference shocks do not affect the relative price of investment in the long run.Assumption C1 expresses this property. In addition, we could impose that news and TFP shocks do not affect the long-run behavior of the relative price of investment, since this would be consistent with the idea that only IST shocks drive the long-run behavior of the relative price of investment. However, instead of imposing these additional restrictions, we will examine whether such properties are supported by the data. In particular, we want to allow news shocks to potentially contain information about future changes in the relative price of investment since there is no a priori reason to eliminate such a possibility. As for TFP shocks, we will show that the additional restriction in which TFP shocks do not affect the relative price of investment in the long run is easily accepted by the data.Assumption C1. Preference shocks and monetary shocks have no long-run effects on the relative price of investment.Assumption C1 implies the identification restrictions l24 = l25 = 0. If we combine assumptions A1, A2, and C1, this is insufficient to identify the five shocks of interest since there is nothing that differentiates a news shock from an investment-specific shock. Another common long-run property that characterizes investment specific shocks in most models is that such shocks do not determine the long-run behavior of TFP. This property is expressed in assumption C2.Assumption C2. IST shocks do not have a long-run effect on TFP.Assumption C2 implies l12 = 0. As we already noted in the parallel case of the B-matrix, we might also have used the analogous restriction l21 = 0 (TFP shocks do not have a long-run effect on IST). We keep this in mind as an overidentifying restriction to be tested in the robustness section below.Our second identification scheme, which we will refer to as ID2, will be comprised of assumptions A1, A2, A3, C1, and C2. Note that this identification scheme (which we will refer to as ID2) does not place any restriction on the short-run behavior of the relative price of investment and therefore is not subject to our previous criticism.Summing up, the just identifying restrictions for ID2 are as follows: 1Since the long-run matrix is singular, 10 restrictions may not be sufficient for identification.2For example, Greenwood, Hercowitz, and Krusell (1997) and most others model IST as different vintages of capital goods. A new vintage has the property that a more productive capital good can be produced at the resource cost of one consumption good than in previous vintages. Hence, the price of investment goods in constant (base year) quality declines over time relative to the price of consumption goods. If the capital stock, Kt, is measured in constant quality investment goods, It, the capital accumulation equation is where Vt is the inverse of the relative price of investment goods. Since we are interested in identifying shocks to IST, we will include the ratio of the consumer price index to an investment price index in the SVECM.III. DataWe will estimate our SVECM model using quarterly data from different sources. For the economic activity variables, we use seasonally adjusted data for gross domestic product, y, personal consumption expenditures, c, and gross private nonresidential investment, i, from the National Income and Product Accounts (NIPA) of the Bureau of Economic Analysis (BEA), table 1.1.5. These variables are expressed in real terms using standard NIPA deflators taken from the same source (table 1.1.9). Hours of the nonfarm business sectors, h, are drawn from the U.S. Basic Economics Database. All variables are in logs and y, c, i, and h are in per capita form using civilian noninstitutional population, ages 16 and over. TFP data, tfp, are constructed using data on capital services for the private nonfarm business sector published by the Bureau of Labor Statistics (BLS). We multiply capital services by the capacity utilization rate in manufacturing drawn from the Federal Reserve Statistical Release G.17. For TFP construction, hours and real GDP series are also for the nonfarm business sector, the latter taken from NIPA table 1.3.5. The capital share is set at 0.31, the mean over the sample compiled by the BLS.To check robustness, we also construct a set of quality-adjusted (QA) variables. To this end, we use QA deflators for total investment and equipment as used in Fisher (2003, 2006). The one drawback of the QA data is that they are available only on a shorter time span. To construct a QA capital stock we use the perpetual inventory method with fixed nonresidential investment deflated by the QA deflator for total investment. This deflator results in lower estimates of real investment prior to 2000 because all capital goods are measured in constant year 2000 quality. The resulting real investment series is denoted iq. The capital stock starting value is taken from the private nonresidential fixed assets series published by the BEA (table 4.1). Depreciation is set at 0.025 per quarter. We measure the real GDP series yq in consumption units (deflator for nondurables and services), also in the construction of QA total factor productivity, tfpq.The inverse of the real price of fixed nonresidential investment, pi, is the log-difference of the NIPA deflator for consumption and the respective NIPA investment price index. An alternative measure, denoted pieq, uses the deflator for nondurables and services consumption and for QA equipment investment instead.3 Real per capita stock prices, sp, are derived as the log-difference between the Standard and Poors 500 Index (SP500), the population series, and the NIPA consumption deflator. In the case of QA variables we use the deflator for nondurables and services and denote this series spq. The short-run nominal interest rate, int, is the H15 effective rate on federal funds.Capital services and capital stock are available only at annual frequency. They are converted to quarterly data assuming constant growth rates within each year. Stock prices and the federal funds rate were retrieved at monthly frequency from Global Insight; the quarterly values are the monthly averages. The sample size is 1955.1–2007.2 for NIPA variables and 1955.1–2000.4 for all variables that rely on the QA deflators.3Fisher (2003) states that the relative price of QA equipment may be a better measure of IST than the relative price of quality adjusted total investment.IV. The Benchmark SystemOur first set of results is based on the five-variable system consisting of tfp, pi, sp, an activity measure, and int. The only deterministic series in the VAR is a constant. If the activity is x, we call this the NIPA_x system. Using Akaike's information criterion (AIC) to determine the appropriate lag length, six lags are recommended for NIPA_h and NIPA_c, three lags for NIPA_i, and nine lags for NIPA_y. However, as figure 1 shows, six lags seem to be a reasonable specification for all these systems. For the sake of maximum comparability we therefore estimate all systems with six lags (i.e., five lags in differences).Fig. 1. AICs for lags 1–10, NIPA system with activity h, i, y, or c View Large Image Download PowerPointTurning to cointegration properties, one might expect from theory that the NIPA systems are driven by two stochastic trends representing disembodied and investment-specific technical progress. Johansen tests for cointegration (using six lags in levels) generally give support for this conjecture, finding evidence of either two or three cointegrating vectors. As three cointegrating vectors are consistent with our prior of having two stochastic trends in the system, we will assume three cointegrating vectors in the benchmark system and consider the possibility of only two cointegrating vectors when we study robustness.We proceed by estimating a VECM for the NIPA_h system, which will be our benchmark. We impose three cointegrating vectors and five lags in differences. Note that we do not assume that all variables in this system have a unit root. The stationarity properties of hours, in particular, has been the subject of much debate (see Christiano, Eichenbaum, and Vigfusson 2004). These authors show that the maintained assumption on whether or not hours have a unit root implies vastly different conclusions for its response to technological innovations if VARs in differences are used. In a VECM framework, by contrast, we do not need to impose any assumptions on the stationarity properties of hours, for if hours were in fact stationary, one of the cointegrating vectors would give nonzero weight only to the hours variable, so that the level of hours affects the first differences of the other variables in the VECM via the error correction term. Of course, if hours were trend stationary, the cointegrating combination should allow for a linear trend, but the hours series we use does not seem to have a discernible trend (see fig. 2).Fig. 2. Hours per capita View Large Image Download PowerPointA. Identification ID1We begin by estimating a structural decomposition of the VECM4 using identification scheme ID1. The variables are ordered as tfp, pi, sp, h, int. We compute impulse responses (IR) and forecast error variance decompositions (FEVD).The FEVDs (see fig. 3), show the contributions of the identified structural shocks to the forecast error variances of each dependent variable over a business cycle horizon of 32 quarters. In discussing the results, we will refer to the shocks as the surprise TFP, surprise IST, news, preference, and monetary shocks.Fig. 3. FEVDs of the NIPA_h system, identification ID1. All FEVDs in this paper are available as colored graphs in the electronic NBER working paper version of this paper, Beaudry and Lucke (2009). View Large Image Download PowerPointThe most interesting findings from the FEVDs are the following: First, surprise TFP and IST shocks contribute almost nothing to the variance of hours at all horizons. The single most important contributor to hours variance is the news shock, in our interpretation the anticipation of future technological possibilities. Only in the very short run (the first three quarters) does the preference shock dominate the variance of hours. The monetary shock explains a sizable share (about 20%) of the variance of hours after 2 years, while most of the rest (roughly 70%) is due to the news shock.Second, stock prices are mainly driven by the news shock, accounting for roughly 80% of the variance at all horizons. Much of the remaining variance seems to be due to preference shocks. Again, it is remarkable that "fundamentals" as represented by surprise TFP and IST shocks seem to be quite unimportant for stock prices. This would be consistent with the view that most technological innovations are known before they are implemented on a scale large enough to have a significant impact on the economy. In fact, the FEVDs show that news shocks contribute up to 30% of the variance of tfp at business cycle horizons, and this share increases further as time goes by; for instance, it is 60% after 15 years. Since this finding is, as we will show, very robust across different modifications of our benchmark system, it seems appropriate to infer that the major compo
Referência(s)