The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Productivity Shocks and Real Business Cycles Charles L. Evans Working Papers Series issues in Macroeconomics Research Department Federal Reserve Bank of Chicago December 1991 (WP-91 -22) FEDERAL RESERVE BANK OF CHICAGO Productivity Shocks and Real Business Cycles 1* Charles L. Evans Federal Reserve Bank of Chicago May 1989 Revised November 1991 Abstract Productivity shocks play a central role in real business cycles as an exogenous impulse to macroeconomic activity. However, measured Solow/Prescott residuals do not behave as an exogenous impulse. Rather, econometric evidence provided in this paper indicates that (1) money, interest rates, and government spending Granger-cause these impulses; and (2) a substantial component of the variance of these impulses (between one quarter and one half) is attributable to variations in aggregate demand. These results are robust to a number of econometric issues, including measurement errors, specification of the production function, and certain forms of omitted real variables. Address: Charles L. Evans Research Department Federal Reserve Bank of Chicago P.0. Box 834 Chicago, IL 60690-0834 (312) 322-5812 This paper has evolved from Chapter 2 of my Carnegie Mellon Ph.D. dissertation. I thank my committee members, Bennett McCallum (chairman), Martin Eichenbaum, Albert Marcet; also Toni Braun, Robert Clower, Finn Kydland, the editors of this Journal and an anonymous referee for helpful comments. I alone am responsible for any errors. The views expressed in this paper are solely those of the author and do not necessarily represent those of the Federal Reserve Bank of Chicago or the Federal Reserve System. 1 1. Introduction Productivity shocks play a central role in Real Business Cycle theories as an impulse to macroeconomic activity (as in Kydland and Prescott (1982), Hansen (1985), example). and then Altug (1985), and King, Plosser, and Rebelo (1988), for In characterizing the business cycle properties of these models, comparing them with the cyclical properties of the data, these researchers assume that productivity shocks are exogenous and uninfluenced by other economic factors. And yet no evidence currently exists to support this standard Real Business Cycle assumption. Many critics of Real Business Cycle (RBC) exogeneity of procyclical productivity shocks; these shocks to be endogenous. For theories question the indeed, many theories predict example, Summers (1986) argues that empirical measures of the change in total factor productivity are contaminated by labor hoarding phenomena; consequently, aggregate demand impulses can give rise to a procyclical productivity measure. Mankiw (1989) argues that the large growth in total factor productivity from 1939-1944 is interpreted most plausibly as a demand-driven response to the military buildup of World War II. Hall (1988) finds evidence in annual data that cost-based measures of Solow residuals covary with exogenous instruments: to noncompetitive forces. Murphy, he attributes this endogeneity Shleifer, and competitive theories with external increasing returns; Vishny (1989) survey these theories predict that changes in total factor productivity are endogenous and demand-driven. Caballero and Lyons (1990) find evidence in annual data of external increasing returns in manufacturing. According to these criticisms, measures of productivity shocks which are based upon changes in total factor productivity will not be strictly exogenous. This paper investigates several quarterly measures of the impulse to an aggregate productivity shock and asks if these measured Solow residuals can 2 survive simple exogeneity tests. The evidence is inconsistent hypothesis that the impulse to an aggregate productivity consequently, the productivity shock is not exogenous. with the shock is exogenous; Initially, in Section 2, the analysis employs Prescott's (1986) measure of the impulse to aggregate productivity. Money, nominal interest rates, and government consistently provide significant predictive power for this results are economically significant: spending impulse. These about one-quarter of the variance of the productivity impulse can be attributed to aggregate demand shocks.^* The analysis of Sections 3-5 demonstrates that these conclusions are robust to a number of econometric issues. Section 3 considers the possibility of random measurement error in the productivity data: in this case, about one-half of the variance of the productivity impulse can be attributed to aggregate demand shocks. Section 4 considers the possibility of specification errors in the production function; twelve measures of the productivity considered and the exogeneity test results are unchanged. impulse are Section 5 considers the possibility that these results are due to omitted real shocks, along the lines considered by King and Plosser (1984) and Litterman and Weiss (1985). However, the finding that money and nominal interest rates provide predictive power a year in advance of the productivity impulse realization makes this an unlikely explanation. For each possibility, the evidence favors the conclusion that measured aggregate productivity impulses do not behave as a strictly exogenous stochastic process. These findings indicate that the role of productivity shocks in generating economic fluctuations has been overstated in the RBC literature. Further research aimed at identifying and understanding "productivity shocks” may be an important element in the debate between RBC theorists and their critics. 3 2. Are Productivity Shocks Exogenous? Prescott (1986) measures the impulse to the aggregate productivity shock as the change in total factor productivity. Assuming an aggregate Cobb-Douglas production function, [] i the productivity shock z _ can be measured using data on output ^ hours (Y) , labor (N), and the capital stock (K) for a given labor share parameter 0. Assuming that contains a unit root in logarithms leads to: zt = £t where c zt l exp ( / + et ) i “ [2] ^ (L) £t-l + W t is a stationary random variable, /?(L) is a polynomial in the lag operator L, and w^ is a mean zero, serially uncorrelated random variable. Prescott's study, € is the measure of technological change. The In Real Business Cycle literature has not taken a firm stand on the stochastic process for e^. Braun Prescott (1989) (1986), assume that Altug (1985), e white is Christiano-Eichenbaum noise; while (1991), Christiano and (1988), King-Plosser-Rebelo (1988), and Eichenbaum-Singleton (1986) allow objects like to be serially correlated. 2 A critical assumption that these papers share is that z^ is an exogenous random variable. policy variables models like These models assume that changes in monetary and fiscal do not alter the distribution of z^; these can usefully "provide a evaluating the importance of other factors ... consequently, well-defined benchmark (e.g., monetary disturbances) actual business-cycle episodes [Long-Plosser(1983, p.68)]." z^ is endogenously-determined, as real Summers (1986) and for in Alternatively, if the models of Murphy-Shleifer-Vishny (1989) imply, then the omission of fiscal and monetary variables distorts the benchmark assessment. [2], the exogeneity of z . requires ^ 4 that In the context of specification be exogenous. Thus, the RBC literature relies upon the exogeneity of but it may be either white noise or a serially correlated random variable. Using [1] and [2], e €t - A can be measured as follows: 6 A log Nt - (1-0) A log Kt - log Yt - / x and e will hereafter be referred to as the productivity impulse. [3] 3 To measure e, Prescott (1986) uses GNP data, an efficiency labor hours series as computed by Hansen (1984), and a capital stock measure which includes the stock of residential housing but excludes the stock of durable consumption goods. calibration purposes, This particular Prescott states that a value of 0=.75 is appropriate. choice requires Prescott uses the value output during the elaboration. In the postwar period when output is defined Since GNP understates the theoretical but labor's compensation is unaffected, postwar period. model, to include the His empirical analysis, however, uses measureof output, and GNP does not include the consumption goods. theoretical 64 since this is the average of labor's share in services of durable consumption goods. GNP as the For services of durable measure of output, labor's share rises to .75 for the This reasoning underlies the value of 0«.75 and Prescott's measure of the productivity impulse e . Given a measure exogeneity of assumption furthermore, the of aggregate productivity RBC models a c, a standard refutable assumption; standard exogeneity testing remains valid even if measures other real shocks are not available. models in which there are two real, that r becomes impulse For example, of consider a class of RBC driving variables, and r^. Suppose follows log Tt - p log Tt_1 + Vt where v^ is a mean zero, random variable. |P|<1 The innovations assumed to constitute a vector white noise process, contemporaneously correlated. and and are and i ^ may be / According to specification [2], past values of 5 v should not help predict e beyond the own past history of e. the productivity impulse e is unpredictable based upon the past values of real variables, context, the of to more than generalization or the omitted real representations for r e can two shock u: be nominal variables, exogeneity measuring refuted without driving variables should be clear. omission of lagged shocks Consequently, and in this v. alternative The linear The critical assumption in [2] is the e (namely, other than , s>l): all of the previously cited RBC papers share this assumption. One way to investigate the exogeneity issue is to conduct a standard, multivariate variables. time series analysis of e and other potential explanatory The following specification is investigated:^ £t = ^ (L) et-i + where /?(L), and a(L) specification a(L) xt-i + are polynomials [u] wt in the Lag operator L. According to [2], x should not provide predictive power for c. A finding that a(L)?*0 in [4] is sufficient to refute the assumption that e is strictly exogenous (for example, see Geweke (1984)).^ The list of variables included in the vector x is: money (Ml), 90-day Treasury Bill rates (TBILL), the the Ml measure of Consumer Price (CPI), real government expenditures (GOVT), and Crude Oil prices (OIL). variables were selected since, in an RBC model, reflect the influence of any omitted variables: typically omitted. productivity index These shocks may all of these variables are The data is quarterly and seasonally adjusted. Four lags of all variables are included in the autoregression [4] . The interest rate variable is measured as the change in Treasury Bill rates; money, government expenditures, the consumer price index, and the crude oil price measured as growth rates (that is, log first-differences).^ periods studied are 1957:11-1983:11 and 1957:11-1978:IV. The two sample The 1983:11 sample period is dictated largely by the availability of Prescott's 6 index are series for € which begins in 1954:IV. sensitivity of the The 1978:IV sample period was chosen to gauge the results to an alternative sample period which did not include the "Volcker experiment" years, 1979-1982. Table 1 reports that Ml, TBILL, CPI, and GOVT individually Granger-cause 6 over the 1983:11 sample period. the statistical significant. always significance 7 8 of 2 The R these for this regression is .47, so results is also quantitatively For both periods, government spending, money and inflation are significant significant at at levels conventional below the levels. 2% level. This Oil suggests prices that are not identifying productivity shocks with past oil price increases may be misleading. 9 The significance of interest rates in the 1983:11 period does not hold for the shorter 1978:IV period. McCallum (1983) has argued in a similar context that both Ml and TBILL may reflect monetary policy in an equation such as this. Therefore, a specification which includes both TBILL and Ml appreciably better than one with simply TBILL (or simply Ml). this possibility, significant not be To investigate notice that Ml and TBILL are jointly significant at less than the 1% level in both periods. only TBILL may (and not Ml) (at the 2.5% Further, when only Ml (and not TBILL) or are included in the x-vector, level). Thus, these variables are money and nominal jointly provide significant explanatory power for c. interest rates The results in Table 1 provide evidence against the hypothesis that this measure of the productivity impulse e is exogenous; consequently, the productivity shock z is not exogenous. The quantitative significance of these nonexogeneity investigated by a decomposition of variance analysis. Ml, TBILL, OIL and GOVT, Table 2 reports results can be For a VAR containing c, the percentage of the 16-quarter ahead forecast error variance of e attributable to these variables. Since the own e-innovations account for 70.8% and 68.5% of the variance in e in the 7 1983:11 and 1978:IV samples, the Ml, TBILL, OIL and GOVT innovations jointly account for 29.2% and 31.5% of the variance in e. The lower bound of the 95% confidence interval is 16.6%,^ so the nonexogeneity of e is quantitatively significant. OIL, Taken singly, the lower bounds of the intervals for Ml, TBILL, and GOVT are near zero; innovations are some uncertainty remains about exactly which quantitatively significant. However, following McCallum (1983) in interpreting monetary policy as Ml and TBILL innovations jointly, monetary policy is quantitatively significant for the full sample period. To conclude Prescott's aggregate this measure demand, of section, evidence productivity reflected in Ml, has been shocks is TBILL, and presented not to exogenous. GOVT, show that Changes influence statistically as well as economically significant way. 11 12 e in in a These results alone, however, are insufficient to refute the exogeneity hypothesis. In principle, these results could represent erroneous rejections if certain econometric and theoretical objections are quantitatively important. Sections 3, 4, and 5 tackle the issues of measurement error bias, specification error bias, and a special form of omitted shock bias. In fact, the essential conclusions of this section are unchanged by these considerations.3 3. Measurement Error Analysis The failure of e to pass simple exogeneity tests in Section 2 could be due to measurement errors in the data. Ordinary Least Squares estimator of If e is measured with error, then the /?(L) in estimated standard errors are not consistent, are uninterpretable. exogeneity tests, To assess consider the [4] is not consistent, the and the previous test results the influence of measurement error on the following statistical model of the true productivity impulse (now referred to as e ), the other variables (x), and two error-ridden measures of the productivity impulse (e^ and 8 H0 : A 12(L)-0 ‘t - A U (L) ‘t-l + A 12(L) xt-l + V [5] k x t " A 21(L) et-l + A 22(L> xt-l + "t [6] ‘it - et + B1 <L) vlt [7] e2t “ 't + B2 (L) v2t [8] where A^. (L) and B^(L) are polynomials in the lag operator L, and are the innovations to k € and x^. Economic agents observe and the k productivity impulse e , but the econometrician can only observe true and e^. The random variables v^ and V 2 are mean zero, serially independent measurement errors generated by the data reporting agencies. random measurement errors, independent of k e . When Since this is a model of each of the errors v^ and V 2 is assumed to be the two productivity measures and are constructed with data reported by independent agencies, the errors v^ and V 2 are assumed measurement to be error mutually similar to independent this as well. one have been Models of investigated classical recently by Sargent (1989), Prescott (1986), and Christiano-Eichenbaum (1991). To complete the measurement error model, the relationship between x, and €2 must be clarified. without error: I assume that the test variables x are measured x, v^, and V 2 are jointly independent at all leads and lags. Allowing for measurement errors in x, as well as and data merit. series symmetrically, an analysis with much would treat all Unfortunately, insufficient data on x is available to implement the instrumental variables estimator described below. errors, therefore, To make some progress on the issue of measurement I follow Prescott (1986) and Christiano-Eichenbaum (1991) in treating the data series asymmetrically. Testing the exogeneity hypothesis in this context requires estimation of A^( L ) and its covariance matrix estimator; consistent estimation of A^(L) as well. k If either consistent the latter requires or c^ is used in place of the unobserved e , and OLS is applied to equation [5], the A^( L ) estimator 9 Using e^ as an instrument for e^ in equation [5] , will not be consistent. however, results in consistent estimation and a valid exogeneity test can be conducted. This estimation procedure estimates of B^(L) are not necessary; is semiparametric in the sense that consequently, misspecification of the order of B.(L) is not an issue. l A decomposition of variance analysis of the VAR system [5] and [6] is possible if a consistent estimator innovation vector, is available. of Q, the covariance matrix In fact, for each innovation for and the two error-ridden observations are available given estimates of A^. (L) and the two error-ridden series are orthogonal, Since the measurement errors in e ^ and e ^ and ^t* the error-ridden residual Construction of a consistent covariance these residual series. series will estimator also be orthogonal. is straightforward given 13 Implementing this econometric procedure requires two measures of e measurement errors are arguably independent. Prescott assumes whose that the measurement errors in the growth rates of GNP and the capital stock measure are negligible. He focuses on measurement errors in the labor input, where two independent series are available for total labor hours: efficiency hours (constructed from the Household Survey Gary Hansen's data), and nonagricultural hours from the Survey of Business Establishments. The data for these series are collected by two separate government agencies, measurement errors errors are arguably independent. I also in output by employing the Federal Reserve's Production as a proxy for GNP. total so the consider measurement series for Industrial If the one-sector theoretical economy exhibits balanced growth, then the data's actual sectoral outputs should aggregate to the one-sector aggregate output series. Thus, the growth rates of GNP and IP should be measuring the same theoretical growth rate in output: to the extent that these growth rates differ, this is interpreted as being due to (serially 10 correlated) measurement errors. Finally, the tables below do not report results which allow for measurement errors in the capital stock variable: I am unable to find an independent measure of the capital stock which is highly correlated with the primary measure used in this study. 14 Table 3 presents the Instrumental Variable (IV) exogeneity test results. The results are presented for two cases: (1) assuming that only the growth rate of hours is measured with error (Hours only);^ and (2) assuming that only measured the growth rates of hours and output are with error • f f (Hours/Output). For the Hours only case, € continues to fail the exogeneity •ff CPI and GOVT Granger-cause e test, but the patterns of failure differ. both periods; in TBILL does in only the 1983:11 period; and Ml does in only the •ff 1978:IV period. periods; Included However, Ml and TBILL jointly Granger-cause and when only TBILL (and not Ml) or only Ml in the system, these variables are e in both (and not TBILL) are significant in both periods. Interpreting both Ml and TBILL as instruments of monetary policy sustains the conclusion that monetary policy has influenced the evolution of the * productivity impulse c . •ff For the case of Hours/Output, the evidence of predictability in e weaker. Ml, TBILL and CPI are jointly significant in the 1983:11 period, but not in the 1978:IV period. be due is to a change in This lack of stability across sample periods could monetary policy over the period 1979-82. GOVT •ff Granger-causes e in both periods. For this case, there is some evidence •ff against the exogeneity of € , but the Granger-causality evidence is substantially weaker than in Table 1. A A Given IV estimates of A^. (L) and 0, Table 4 reports decomposition of •ff variance results includes Ml, e , the for TBILL, OIL, true and GOVT. productivity For each impulse, case in a VAR which in both periods, the •jjf percentage of variance in e which 11 is attributable to own innovations is Apparently, in Table 2 the measurement error in e is smaller than in Table 2. being attributed more to the productivity impulse innovations than the other innovations. The confidence intervals tend to be wider when measurement error is accommodated. Nevertheless, aggregate demand variables k contribute between 34-60% of the variance of e ; confidence interval are between 10-43%. and oil prices the lower bounds on the 95% The nonexogeneity evidence here is stronger than in Table 2. Based upon the evidence presented in Tables 3 and 4, the failure of measured productivity impulses to pass simple exogeneity tests is not likely to be due to the presence of classical measurement errors in the productivity data. 4. Specification Error Analysis Another potential criticism of the exogeneity tests measure of the aggregate productivity impulse c. Section 2 might be specific to: capacity technology; utilization. or This In principle, the results in (1) the choice of labor input data; value of the constant labor share parameter 0; the aggregate is the particular (4) the section (3) (2) the the functional form for assumption of a constant rate briefly discusses the results of of a sensitivity analysis. The principal finding is that the results of Section 2 are robust: the strict exogeneity of c is refuted for the 12 measures considered. First, Prescott's measure of e uses Hansen's series as the measure of labor hours. (1984) efficiency hours In principle, the predictability of e could be an artifact of this constructed series. Two alternative aggregate labor hours series, however, are available: the Household Survey measure and the Survey of Business Establishments. Accordingly, alternative measures of e have been computed using the Household and Establishment Survey hours data to 12 address this possibility. Second, function, under the assumption of an aggregate Cobb-Douglas measuring e requires an estimate of labor's share in output The previous measure assumes that 0=.75, just as Prescott did. the three production labor measures, 0 can be however, aggregate Cobb-Douglas production function. estimated (0). For each of directly from the Since theory predicts that labor hours will respond to productivity shocks, consistent estimation requires the use of an instrumental variables estimator. uncorrelated, however, labor hours, If the true impulse is serially a valid set of instruments includes lagged values of capital, and output. Given consistent estimates of 0, appropriate measures of e can be constructed. A third problem production function. may be the assumption an aggregate Cobb-Douglas This criticism can be addressed by computing a standard Solow measure of total factor productivity, weights. of which uses time-varying factor This measure is consistent with any constant-returns-to-scale (CRS) aggregate technology if markets are competitive. Since Real Business Cycle theories typically assume a competitive environment, the Solow residual is an appropriate measure of the productivity impulse for any CRS technology. As Hall (1988) has noted, however, in noncompetitive environments this measure of productivity impulses will not be exogenous. In this case, an exogeneity test failure would be consistent with Hall's findings.^ Finally, using the entire aggregate capital stock as a measure of the capital input to production implicitly assumes that capacity utilization is constant over the business cycle. Relaxing this assumption is difficult since existing measures of capacity utilization are inappropriate for computing a utilized capital series (see Shapiro (1989) for example). I follow Prescott (1986) in allowing for variable capital utilization through the variations in labor input. Specifically, utilized capital services in production is u^k^, 13 u is the utilization rate, and u^n^. Prescott used a value of a=0.40; selecting a variety of a values left the test results qualitatively unchanged. The Granger-causality and variance decomposition results are similar to the results of Section 2, and so are not reported here to conserve space. four-variable VAR containing £, Ml, TBILL, and GOVT was estimated. A In each of the 12 specifications,^ either Ml, TBILL, or both Granger-causes e at very low significance levels (less than 2.5%); GOVT Granger-causes e in each of the 12 cases also at low significance levels. The predictability of the productivity impulse e is a remarkably robust result. The variance decomposition Granger-causality test results. results mimic the robustness Innovations in Ml, TBILL, of the and GOVT account for between 26-33% of the variance in the 16-quarter ahead forecast error of £. The lower bounds of the 95% confidence intervals are between 12-21%. Thus, the quantitative significance of these variables is also robust across the alternative measures of e . 5 5. Signalling and the Omitted Real Shock Hypothesis The predictability of e can be interpreted plausibly in one of two ways: either (1) changes in money, interest rates, and government spending lead to changes in measured productivity e, or (2) changes in these variables reflect changes in other interpretation, real the shocks omitted findings above are spurious, which real lead to changes shock hypothesis, is in c. that The the latter empirical and a more complete specification of the real shocks in the economy would overturn the results. As I discussed in Section 2, specification [2] rules out many omitted shock hypotheses; however, the RBC literature has featured one important alternative which has not been ruled out so far. King and Plosser (1984) consider an RBC model in which endogenous money can respond to real shocks before output can respond. 14 Specifically, some productivity shocks which occur in period t+1 are revealed in period t; endogenous money and other financial variables respond to this information in period t. Similarly, Litterman and Weiss (1985) describe an economy where economic agents have more information about future aggregate supply shocks than does the econometrician; since financial and monetary variables convey information about these unobserved shocks, real After variables. controlling for nominal variables the unobserved Granger-cause shocks, however, Litterman-Weiss find that real variables are block exogenous with respect to nominal variables. Litterman-Weiss Thus, the apparent importance of nominal variables in the economy is spurious. These examples suggest that the importance of nominal variables for predicting productivity shocks may simply reflect the influence of omitted real shocks, even in the context of specification [2]. To see this in a simple context, suppose that the productivity shock z _ ^ follows the stochastic process: log zt = e^ where and log zt_1 + M + elt + «2 ,t-l €2 t 1 are assumec* to ke mean zero, serially uncorrelated, stationary random variables impulse is both impulses, the spirit of [9] and E [ €^t c2 ^ 1 ^ ^ 0 is permitted. revealed in period t, whereas ^ is revealed in period t-1; however, are realized in period King-Plosser (1984): The t. This economic agents specification is in can anticipate some productivity shocks prior to their realization, while others are completely unanticipated. Define + 62 t 1 an<* note is t^ measured le productivity impulse from equation [3]. In a monetary economy with this aggregate technology, inside money, outside money, stock prices, and nominal interest rates can respond in period t to an impulse (c2t^ period t+l„ In this signalled in period t but not realized until sense, a finding 15 that time t nominal variables Granger-cause e c o u l d be spurious; that is, e could fail Granger-causality tests but be strictly exogenous. In the context of [9] , 6t+} should not be . correlated with money and interest rates which are sufficiently distant in time: growth rate of money and nominal 18 uncorrelated with periods in advance of interest More generally, their rates in this example, the in period specification [4] realization, can be possible signalling factors: et “ ^ (L) et-i + should be some impulses may be revealed p but information available in period t-p should be uncorrelated with of p, t-1 appropriately which becomes For a given choice altered to control for the 19 a(L) xt-P-i + wt [4'] Thus, the exogeneity hypothesis now implies that a(L)=0 in [4']. No a priori information is available to suggest one, unique value for p. Litterman-Weiss (1985) and King-Plosser (1984) each select a model which would set p equal to one period. Since the sample interval for this study is quarterly, and the King-Plosser model could easily refer to yearly decisions, Table 5 reports signalling test results for p= 1, 2, 3, and 4 quarters. In Table 5, the vector of explanatory variables includes Ml, TBILL, and GOVT. First, government spending is not significant at any reasonable level for any choice of p>l. Second, TBILL provides explanatory power as early as four quarters ahead (p=3), and Ml provides explanatory power at seven quarters ahead (p=6, unreported). p=6, unreported). Jointly, Ml and TBILL are always significant (up to Third, when e is computed using 0=.75 and either the Establishment or Household Survey hours, the corresponding results for Table 5 are not appreciably different (again, unreported). If the signalling hypothesis is the 20 correct explanation for the explanatory power of money and interest rates, then productivity impulses must be anticipated 7 quarters ahead: this feature is at variance with every RBC 16 model which has been studied to date. Consequently, the evidence favors an e in a fundamental interpretation in which the nominal variables influence way, not an omitted variable channel such as specification [9], 6. 21 Conclusions The results above demonstrate that productivity shocks as measured by Solow/Prescott processes. methods Money, do not nominal behave as interest strictly rates, exogenous and stochastic government spending individually and jointly Granger-cause various measures of the impulses these shocks. hypothesis These results are not due to Classical measurement errors. The that investigated, and this result is no evidence due has to been omitted found to real economically significant: their factors support Furthermore, the influence of money, interest rates, and is to innovations the has been hypothesis. government spending account for between one-quarter and one-half of the forecast error variance in e at the 16-quarter forecast horizon. The lower one-quarter value orthogonalization of the innovations is computed under an RBC in the absence of measurement errors; the upper one-half value, after accounting for measurement errors. As a whole, these results cast a shadow over the current generation of RBC models which assume strictly exogenous productivity shocks and exclude any interesting role for aggregate demand shocks or other supply shocks. At a minimum, these results imply that the RBC literature to date has overstated the importance of productivity shocks for economic fluctuations. which may be consistent with the evidence presented here Two theories are the labor hoarding model of Burnside, Eichenbaum, and Rebelo (1990) and the productive externality model of Baxter and King (1990). According to both models, conventionally measured Solow/Prescott residuals are not exogenous. 17 In these models prices are perfectly flexible, so the empirical finding that money and interest rates Granger-cause productivity shocks would presumably be explained as reverse causation as in King and Plosser (1984). Alternatively, if prices were assumed to be sticky in these types of economies, these Granger-causality findings would be explained as direct causality. To discriminate among these various theories as well as further assess the role of productivity shocks, researchers should investigate economic structures which jointly predict the stylized facts of business cycles and endogenous Solow residuals. Data Appendix Many of the data series used in this study are directly available from the CITIBASE data base (their CITIBASE labels are in []): price index expenditures GNP, less shelter [GGE82]; OIL, [PUXHS]; [GNP82]; Establishment survey and the Capital Stock [KRH72, KN72]. (1984). (federal) government IP, Industrial Production [LPMHU], Household Survey [IP]; [LHOURS]; The Efficiency hours data is from Hansen The Ml (money) and TBILL (90-day Treasury Bill rates) data are the same as in Eichenbaum-Singleton (1986). real the producer price index for crude oil [PW561]; real gross national product Labor hours data: GOVT, CPI, the consumer 18 Table 1: The Predictability of Prescott1s Productivity Impulse0 £t = [4] ^ (L) et-i + q(l) xt-i + wt Marginal Significance Levels for Testing Hq : _ b X- vector a. 1957:11 - 1983:11 a(l 1957:11 - 1978:IV Ml TBILL CPI GOVT OIL .0033 .0183 .0003 .0005 .8895 .0172 .1628 .0193 .0019 .1455 Ml, TBILL Ml, TBILL, CPI .0000 .0000 .0001 .0001 b. Ml alone* .0003+ .0002 c. TBILL alone* .0048 .0209+ a Four lagged values of c and X are used in the autoregression. The marginal significance levels can be interpreted in the following manner: for Ml in the period 1957:11-1978:IV, the marginal level .0172 indicates that the Null Hypothesis of a(L)«0 (with respect to the Ml components of X) would be rejected at significance levels of 1.72% and higher. ^The vector autoregression includes Ml, TBILL, CPI, GOVT, and OIL as components of the X-vector. The line "Ml, TBILL" reports marginal significance levels for testing the joint hypotheses that the Ml and TBILL coefficients are a block zero vector. Similarly for "Ml, TBILL, CPI." * Other elements in the X-vector are: GOVT, OIL, and CPI. + OIL is significant at the 5% significance level. 19 Table 2: Decomposition of Variance Results a Percentage of Variance in Prescott's Productivity Impulse e Explained by Innovations in Vector Autoregression [4]: Point Estimates and 95% Confidence Intervals Components of X-vector 1957:11 ^ 1978:IV 70.8 (58.2, 83.4) 8.2 ( 2.5, 14.0) 7.7 ( 0.4, 15.1) 2.4 ( 0.0, 5.6) 10.8 ( 0.0, 21.9) e Ml TBILL OIL GOVT Ml, TBILL*3 u OIL, GOVT 1957:11 - 1983:11 68.5 (55.0, 82.1) 6.5 ( 0.9, 12.1) 9.0 ( 0.0, 18.4) 4.2 ( 1.1, 7.2) 11.8 ( 0.0, 25.4) 15.9 ( 6.5, 25.3) 13.2 ( 1.9, 24.5) 15.5 ( 3.9, 27.0) 16.0 ( 2.9, 29.1) The order of orthogonalization is in the order of the variables listed. forecast horizon is 16 quarters. The ^The line "Ml, TBILL" reports the percentage of variance jointly explained by Ml and TBILL innovations. The point estimate is the simple sum of the individual Ml and TBILL percentages; however, the 95% confidence interval requires more extensive calculations (see footnote #11 in the text). Similarly for "OIL, GOVT." 20 Table 3: The Predictability of Prescott 's Productivity Impulse5 in the Presence of Classical Measurement Errors et _ An (L) V l [5] + A12(L) Xt-1 + W t Marginal Significance Levels for Testing: H : 1957:11 - 1983:11 _ b X- vector a. 0 Hours Only Hours/Output = l 1957:11 - 1978:IV 0 Hours Only Hours /Output Ml TBILL CPI GOVT OIL .0699 .0004 .0000 .0369 .7428 .7455 .2286 .0338 .0145 .7518 .0092 .2533 .0005 .0327 .0780 .8240 .7554 .3136 .0137 .0405 Ml, TBILL Ml, TBILL, CPI .0000 .0000 .1210 .0004 .0000 .0000 .5305 .3160 b. Ml alone* .0191 .5404 .0056 .4493 c. TBILL alone* .0015 .0840 .0458 .4137 Four lagged values of c and X are used in [5] , and 8 lags are used in computing the Newey-West heteroskedasticity-autocorrelation consistent covariance matrix estimator. ^See footnote b in Table 1. q "Hours Only": IV estimation assumes that only the Hours series contains measurement error; "Hours/Output": IV estimation assumes that the Hours and Output series contain measurement error. *Other elements in the X-vector are: 21 GOVT, OIL, CPI. Table 4: Decomposition of Variance Results in the Presence ofc Classical Measurement Errors Percentage of Variance in Prescott's Productivity Impulse Explained by Innovations in the Vector Autoregression [5]: Point Estimates and 95% Confidence Intervals 1957:11 - 1983:11 Components of X-vector * 6 Ml TBILL OIL GOVT Ml, TBILLd OIL, GOVT 1957:11 - 1978:IV Hours Only Hours/Output Hours Onlv Hours /Output 47.5 (30.1, 65.0) 16.1 ( 8.0, 24.2) 13.7 ( 1.2, 26.2) 4.8 ( 0.0, 10.7) 17.9 ( 0.0, 41.3) 66.0 (42.8, 89.3) 9.5 ( 0.0, 20.8) 5.9 ( 0.0, 12.0) 3.2 ( 0.0, 9.7) 15.5 ( 0.0, 36.7) 39.8 (23.5, 56.1) 13.8 ( 3.5, 24.0) 15.1 ( 0.0, 31.2) 8.4 ( 0.0, 17.2) 22.9 ( 0.0, 49.5) 50.8 (22.4, 79.2) 8.0 ( 0.0, 23.0) 7.1 ( 0.0, 21.7) 4.8 ( 0.0, 11.3) 29.2 ( 0.0, 58.5) 29.8 (14.9, 44.7) 22.7 ( 0.0, 46.7) 15.4 3.4, 27.4) ( 18.7 ( 0.0, 39.6) 28.9 ( 9.5, 48.3) 31.3 ( 7.2, 55.4) 15.1 ( 0.0, 38.7) 34.0 ( 6.7, 61.3) c l The order of orthogonalization is in the order of the variables listed. forecast horizon is 16 quarters. b c ’ See the corresponding footnotes in Table 3. ^See footnote b in Table 2. 22 T 1 Table 5: . a Testing the Signalling Hypothesis £t “ <L) V p - i + wt [4'] Marginal Significance Levels for Testing H q : X-vector^ d ~ 1 0 = 2 o = 3 a(L)~0 o = 4 Ml .1536 .0311 .0294 .0013 TBILL .0278 .0008 .0008 .1497 GOVT .8907 .6924 .7703 .9462 Ml, TBILLC .0000 .0000 .0000 .0000 aThe productivity impulse e is Prescott's measure, the sample 1957:11 - 1983:11, and four lags are used in the estimation. period The elements of the X-vector are Ml, TBILL, and GOVT. The Null hypothesis is that the block of coefficients associated with Ml and TBILL are jointly zero. 23 References Altug, S., 1985, Gestation lags and the business cycle: an empirical analysis, manuscript, University of Minnesota. Baxter, M. and R. King, 1990, Productive externalities and cyclical volatility, Rochester Center for Economic Research, working paper no. 245. Boschen, J. and L. Mills, 1988, Tests of the relation between money and output in the real business cycle model, Journal of Monetary Economics 22, 355-374. Braun, R . , 1989, Taxes and postwar U.S. business cycles, manuscript, University of Virginia. Burnside, C., M. Eichenbaum, and S. Rebelo, 1990, Labor hoarding and the business cycle, manuscript, Northwestern University. Caballero, R . , and R. Lyons, 1990, The role of external economies in U.S. manufacturing, manuscript, Columbia University. Christiano, L., 1988, Why does inventory investment fluctuate so much? Journal of Monetary Economics 21, 247-280. Christiano, L. and M. Eichenbaum, 1991, Current real business cycle theories and aggregate labor market fluctuations, forthcoming in the American Economic Review. Costello, D., 1989, A cross-country, cross-industry comparison of the behavior of Solow residuals, manuscript, University of Rochester. Eichenbaum, M. and K. Singleton, 1986, Do equilibrium real business cycle theories explain postwar U.S. business cycles? in: S. Fischer, editor, NBER macroeconomics annual 1986 (MIT Press) 91-135. Engle, R . , D. Hendry, and J. Richard, 1983, Exogeneity, Econometrica 51, 277-304. Geweke, J., 1984, Inference and causality in economic time series models, Handbook of econometrics 2, 1101-1144. Hall, R . , 1988, The relationship between price and marginal cost in U.S. industry, Journal of Political Economy 96, 921-947. Hansen, G., 1984, Fluctuations in total hours worked: a study using efficiency units, manuscript, University of Minnesota. Hansen, G., 1985, Indivisible labor and the business cycle, Journal of Monetary Economics 16, 309-327. Hansen, G. and T. Sargent, 1988, Straight time and overtime in equilibrium, Journal of Monetary Economics 21, 281-308. Hansen, L., 1982, Large sample properties of generalized method of moments estimators, Econometrica 50, 1029-1054. 24 King, R. and C. Plosser, 1984, Money, credit, and prices in a real business cycle, American Economic Review 74, 363-380. King, R . , C. Plosser, and S. Rebelo, 1988, Production, growth, and business cycles, Journal of Monetary Economics 21, 309-342. Kydland, F. and E. Prescott, 1982, Time to build and aggregate fluctuations, Econometrica 50, 1345-1370. Litterman, R. and L. Weiss, 1985, Money, real interest rates, and output: a reinterpretation of postwar U.S. data, Econometrica 53, 129-156. Long, J. and C. Plosser, 1983, Real business cycles, Journal of Political Economy 91, 39-69. Mankiw, N.G., 1989, Real business cycles: a new Keynesian perspective, Journal of Economic Perspectives 3, 79-90. McCallum, B., 1983, A reconsideration of Sims' evidence concerning monetarism, Economic Letters 13, 167-171. McCallum, B . , 1989, Real business cycle models, in: R. Barro, editor, Modern business cycle theory (Harvard University Press), 16-50. Murphy, K. , A. Shleifer, and R. Vishny, 1989, Building blocks of market clearing business cycle models, in: 0. Blanchard and S. Fischer, editors, NBER macroeconomics annual 1989 (MIT Press), 247-287. Newey, W. and K. West, 1987, A simple, positive definite, heteroskedasticity and autocorrelation consistent covariance matrix, Econometrica 55, 703-708. Prescott, E., 1986, Theory ahead of business cycle measurement, Carnegie-Rochester Conference Series on Public Policy 27 (Autumn), 11-44. Runkle, D . , 1987, Vector autoregressions and reality, Journal of Business and Economic Statistics 5, 437-442. Sargent, T. , 1989, Two models of measurements and the investment accelerator, Journal of Political Economy 97, 251-287. Shapiro, M. 1989, Assessing Federal Reserve measures of capacity and utilization, manuscript, Yale University. Summers, L . , 1986, Some skeptical observations on real business cycle theory, Quarterly Review 10 (Federal Reserve Bank of Minneapolis, Minneapolis), 23-27. White, H . , 1980, A heteroskedasticity consistent covariance matrix estimator, Econometrica 48, 817-838. 25 Footnotes ^The empirical approach here (1990) in three ways: different; differs from Hall (1988) and Caballero-Lyons (1) the instruments and identifying restrictions are (2) this paper uses quarterly rather than annual data; Hall-Caballero-Lyons focus exclusively on contemporaneous and (3) correlations, whereas this paper does not. 2 In a trend-stationary economy, the logarithm (or level) of z _ is often ^ assumed to be an exogenous, AR(1) process as in Hansen (1985), Hansen-Sargent (1988), King-Plosser-Rebelo (1988), and McCallum (1989). 3 Referring e as to terminology productivity if the productivity e is serially shock. I will "impulse" correlated. refer is an abuse Nevertheless, repeatedly to e as since the of standard is the "impulse," irrespective of its serial correlation properties. 4 Specifications of [4] which set jS(L)=0 a priori have also been investigated, and the conclusions drawn are similar. ^Weaker forms of exogeneity do not seem appropriate here. Weak exogeneity and predeterminedness are econometric conditions which determine efficient estimation techniques (Engle, Hendry, and Richard (1983)); however, admit specifications for e 26 these conditions, which violate the spirit of RBC models. Alternative investigated. stationary-inducing In particular, transformations the basic of conclusions the of data have been this paper are unchanged for trend-stationary and Hodrick-Prescott transformations of the data (including the productivity variable z^). ^All of the test results reported in this paper have been generated using conditional heteroskedasticity-consistent covariance estimators as suggested by White (1980) and Hansen (1982). 8 In simple autoregressions with only a univariate x-variable, the exogeneity hypothesis fails often. For example, the following variables Granger-cause e in these autoregressions: the monetary base (in the 1983:11 period only), Ml, TBILL, the Federal Funds rate, CPI, GOVT, and OIL. The Trade deficit did not Granger-cause e. 9 This evidence in no way rules out the possibility that oil price changes influence c contemporaneously. ^Confidence intervals were described in Runkle (1987); computed by the normal approximation method the covariance matrix estimator is conditional heteroskedasticity-consistent. ^Confidence intervals around the statistic Q * g^(/3) + g2 ( 3 are comPute^ /) the obvious way, using the fact that Var(Q) * Var[g^(/J)] + Var[g2 (/?)] + 2Cov[gl09),g2 08)]. 12 This conclusion regarding Ml and TBILL continues to hold if the order of orthogonalization is c, OIL, GOVT, Ml, and TBILL. 27 13 For example, let w ^ and and resPect^ve^y• be the two constructed residuals of [5] using Then an estimator for the variance of is the sample covariance between w ^ and w 2t * 14 As an instrument for the capital stock, Costello consumption, but only that data is available (1989) uses electricity annually. As a quarterly instrument, I have tried the production of electricity by utility companies. The correlation between this instrument and the primary capital variable is .37 (in growth rates). When this instrument is employed, the exogeneity hypothesis fails more often than for the case which uses Hours and Output only. "^Prescott and Christiano-Eichenbaum assume that only the logarithm of labor hours is measured with error: A^(L)=0. with their assumptions imply that B^(L)=Bq -Bq L and Instead, I assume that the growth rate of labor hours is measured error, and allow serially correlated. the measurement error process to be arbitrarily Also, A^(L) ^ 0 is permitted. ^Under the assumption that the technology is accurately specified, issues of market power productivity play no explicit shocks. For role in the nonexogeneity example, ina noncompetitive of economy measured where aggregate production takes place according to [1] and [2], if € is correctly measured according to even in the [3], it will survive exogeneity tests presence of market power. ^Twelve measures arise due to the three cases: 0 estimated by 0_ ^; (1) 0®.75; (2) IV; labor hours series (3) time-varying and (4) variable capacity utilization with 0=.75. 28 and the four Solow weights 18 This restriction applies regardless of the propagation mechanisms economy. and in the For example, suppose that the propagation mechanisms lead to m^ ^ ^ being correlated with y^+^, and nt+l’ ^ the techn°l°gy accurately specified and the factors are accurately measured, then A log z _ ^ +^ * / + €t+l# i ^ specification [9], is uncorrelated with m^ ^ and 19 As in [4], serial correlation in e^ can be accommodated. Suppose that the aggregate productivity shock process is given by: log = Z u log z + u "it + u2,t-l + where the {u^} be + n + € are mean-zero and p+l,t-p serially uncorrelated, but the {u.^} may contemporaneously correlated in the period (that is, Efu^^u^ t ^]^0, Efu^u^ t 2^^* in which etc*)# they are realized The ex°geneity tests based upon [4'] are valid for this more general specification of the omitted real shock hypothesis. Also, setting /3(L)=0 a priori leads to essentially the same test results as reported in Table 5. 20 Allowing for conclusions; stronger errors as in Section detrending 3 for both sample procedures also does results. in Section 3 does not alter these in fact, the Granger-causality evidence against exogeneity is than alternative measurement 29 periods. not change Accounting the for qualitative Since productivity shocks contain predictable components, these results are consistent with the existence of numerous sources of economic fluctuations. If nominal variables influence z _ and z _ drives the economy, ^, ^ variables should influence output. significant influence of nominal Boschen-Mills variables on then nominal (1988), however, output. output effect, however, is a challenge for future research. find no Quantifying this Presumably, this will require a structural model which tightly restricts the specifications and lag lengths assumed here and in Boschen-Mills. 30