ASYMMETRIC PHASE SHIFTS IN U.S. INDUSTRIAL PRODUCTION CYCLES Yongsung Chang and Sunoong Hwang* Abstract—We identify cyclical turning points for 74 U.S. manufacturing industries and uncover new empirical regularities: (a) industries tend to comove between expansion and contraction phases over the business cycle; (b) clusters of industry turning points are highly asymmetric between peaks and troughs: troughs are much more concentrated and sharper than peaks; (c) the temporal pattern of phase shifts across industries supports the spillovers through input-output linkages; and (d) macroeconomic shocks, such as unanticipated changes in monetary policy, government spending, oil prices, and financial conditions, are significant drivers of industrial phase shifts.
I.
Introduction
T
HE comovement of industries over the business cycle is a salient feature of market economies (Burns & Mitchell, 1946; Lucas, 1977). The empirical pattern of industrial comovement is of profound importance because it forms the basis for modern (one- or multisector) business cycle models. While there has been a great deal of empirical work on such comovement (Long & Plosser, 1987; Shea, 2002; Foerster, Sarte, & Watson, 2011), most studies have focused on correlations between industrial growth rates. In contrast, relatively little is known about the comovement of phase shifts across industries, while the concentration of cyclical phases is a cornerstone of the classical definition of the cyclical comovement, suggested by Burns and Mitchell (1946, p. 70): “A period in which expansions are concentrated is succeeded by another in which cyclical peaks are concentrated, by another in which contractions are concentrated, by another in which cyclical troughs are concentrated; and this round of events is repeated again and again.” The timing of turning points is of great interest to policymakers, financial analysts, and individual investors. There has been a renewed interest in this type of comovement of economic variables, aided by both the increased availability of data and the development of relevant econometric tools. For example, Harding and Pagan (2006), Chauvet and Piger (2008), and Stock and Watson (2010, 2014) have developed formal procedures to determine the aggregate turning points based on the clusters of turning points in individual series. Received for publication August 2, 2011. Revision accepted for publication November 5, 2013. Editor: Mark W. Watson. * Chang: University of Rochester and Yonsei University; Hwang: Korea Institute for Industrial Economics and Trade. We are grateful to Adrian Pagan, Valerie Ramey, Mark Watson (the editor), two anonymous referees, and participants at various seminars for helpful comments and suggestions. We also thank Martin Burda, Luca Gambetti, Simon Gilchrist, Don Harding, Naohisa Hirakata, Adrian Pagan, Valerie Ramey, Nao Sudo, Christoph Thoenissen, Mark Watson, and Egon Zakrajšek for the support of data and computer codes used in different versions of this paper. The views expressed here are our own and do not necessarily reflect the views of the Korea Institute for Industrial Economics and Trade. Y.C. acknowledges support from the WCU (WCU-R33-10005) program through the Korean National Research Foundation of the Ministry of Education, Science and Technology. A supplemental appendix is available online at http://www.mitpress journals.org/doi/suppl/10.1162/REST_a_00436.
Owyang, Piger, and Wall (2005) and Hamilton and Owyang (2012) examined commonalities and differences across U.S. states in the timing of switches between cyclical phases. Artis, Marcellino, and Proietti (2004) and Altug and Canova (2013), among others, provided synchronization analyses for the business cycles of European countries. Nonetheless, there is so far no systematic study of industry-level cyclical phases. Further, none of the above studies looked for asymmetries in the clusters of turning points between peaks and troughs or examined the determinants of co-occurrence of phase shifts in individual series. In this paper, we take a step toward filling these gaps. More specifically, we provide empirical evidence on (a) the concentration of industry-level phases over the business cycle, (b) the asymmetric features of peak and trough clusters, and (c) the effects of interindustry spillovers and macroeconomic shocks on the concurrence of industry phase shifts. We first identify turning points in the industrial production (IP) indices for 74 manufacturing industries over the period 1972:I to 2011:IV, using a nonparametric dating algorithm proposed by Harding and Pagan (2002). The diffusion and concordance analyses indicate a strong comovement of industries between expansion and contraction phases in sync with the aggregate business cycle. But more interesting from our point of view is the asymmetric concentration of peak and trough clusters: industry troughs are much more concentrated than industry peaks at the corresponding aggregate turning points. We also find that industry peaks tend to lead the aggregate peaks, whereas industry troughs tend to slightly lag the aggregate troughs. Reflecting these distributional features, the list of leading, coincident, and lagging industries changes substantially between the aggregate peaks and troughs. In the literature, there has been a tradition of characterizing the business cycle by sudden stop and slow recovery, dating back to Keynes (1936).1 Our disaggregate evidence on changes in log IP surrounding the aggregate turning points does not support this conventional notion. The result is instead consistent with sharp troughs and round peaks, as in the case of aggregate IP documented by McQueen and Thorley (1993). All of these asymmetries are recurrently observed over the past six recessions in our sample. We next proceed by investigating the determinants of concurrence of phase shifts across industries. We consider two types of explanatory variables: spillovers from inputoutput linkages and macroeconomic shocks.2 We distinguish 1 Keynes (1936, p. 314) wrote: “The substitution of a downward for an upward tendency often takes place suddenly and violently, whereas there is, as a rule, no such sharp turning point when an upward is substituted for a downward tendency.” 2 For the theoretical discussion in support of the importance of inputoutput linkages, see Long and Plosser (1983), Horvath (2000), and Acemoglu et al. (2012); for that of aggregate shocks, see Lucas (1977) and Dupor (1999).
The Review of Economics and Statistics, March 2015, 97(1): 116–133 © 2015 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology doi:10.1162/REST_a_00436
ASYMMETRIC PHASE SHIFTS IN U.S. INDUSTRIAL PRODUCTION CYCLES
upstream (demand-side) and downstream (supply-side) spillovers based on the input-output matrix. A novelty of our approach is that a spillover effect is identified by the change in the probability of an industry experiencing a phase shift resulting from past phase shifts in its neighbor industries. The macroeconomic shocks we consider in the baseline analysis are (a) monetary policy shocks identified using a vector autoregression (VAR), (b) Ramey’s (2011) government spending shocks, (c) Hamilton’s (2003) oil price shocks, and (d) Gilchrist and Zakrajšek’s (2012) financial shocks. By estimating a panel probit random-effects model over 1973:III to 2010:III, we find that all of the above explanatory variables have a significant effect on the occurrences of industry phase shifts, confirming our economic priors. The results are robust to various sensitivity checks and to the sample period starting in the mid-1980s. The main contributions of our work can be summarized as follows. First, we provide a new empirical characterization of industry comovement. To measure the degree of comovement, previous studies have usually relied on the standard correlation coefficient computed with all of the sample observations, without any focus on specific discrete events like phase shifts. As a result, industry comovement has been characterized by a cross-sectional set of scalar values representing the degree of linear relationship between industrial growth rates. Relative to these studies, our paper highlights the dynamic nature of industry comovement and its nonlinearity by demonstrating how the degree of concentration of cyclical phases changes over time and comparing the clusters of peaks and troughs. Second, we offer a new dimension to the analysis of business cycle asymmetries. As is well known (and can be seen in the above quotation), Burns and Mitchell’s definition of the business cycle has two key attributes: the alternation between periods of expansion and contraction and the comovement among disaggregated series. Previous studies of business cycle asymmetry have generally focused on the differences between expansion and contraction phases in the time series properties of just one or a few macroeconomic variables.3 By comparison, focusing on the other defining feature of the business cycle, comovement, our paper shows that substantial asymmetries are also found in the clusters of peaks and troughs in industry cycles. Third, our work complements the existing studies on the effects of interindustry spillovers and macroeconomic shocks on industry comovement. Among previous empirical studies in this area, the closest to our work are Bartelsman, Caballero, and Lyons (1994) and Shea (2002), who used a reduced-form panel data regression and found that both types of determinants are important in explaining observed patterns of comovement.4 The general conclusion of our analysis is in line with these studies. However, our work 3 See
Morley (2009) for an extensive summary. different empirical approaches, Long and Plosser (1987), Conley and Dupor (2003), Foerster et al. (2011), and Holly and Petrella (2012) obtained similar results. 4 Using
117
differs from this literature in several important respects. While the previous studies focus on the effects on industrial growth rates, our emphasis is instead on the effects on the probabilities of industry phase shifts.5 Our approach also enables us to evaluate whether the impacts of the determinants are asymmetric between peaks and troughs, a question that has not been addressed before. Finally, we consider four macroeconomic shocks, whereas earlier work considered only monetary policy and oil price shocks (e.g., Shea, 2002). The new empirical regularities that we uncover will help us to better understand the sources and propagation of business cycles. The observed patterns of comovement can be used as a metric for evaluating the validity of standard multisector models. Our results can motivate new theoretical studies on the mechanisms behind the asymmetric turning point clusters. The knowledge of distributional features of turning points and the list of identified leading industries can be used to improve our forecasting ability for national recessions and recoveries. The remainder of this paper is organized as follows. Section II describes the data and the method for dating turning points in industry cycles. Section III presents the results of diffusion and concordance analyses. Empirical results for asymmetric turning point clusters are given in section IV. In section V, we investigate the determinants of interindustry comovement. Section VI concludes. II.
Dating Industry Cycles
A. Data and Algorithm
We use disaggregated data on IP extracted from the board of governors of the Federal Reserve System. The data are quarterly, seasonally adjusted, and taken in logs. The sample period is 1972:I to 2011:IV. In our data, the manufacturing sector is broken down into 74 industries, which correspond roughly to the four-digit level of disaggregation in the 2002 North American Industry Classification System (NAICS).6 We identify turning points in industry cycles using the method proposed by Harding and Pagan (2002), a quarterly variant of the monthly Bry and Boschan (1971) algorithm. This algorithm is applied to the log levels of IP indices and has the following advantages. First, inspecting the behavior of the series in levels is consistent with the practice maintained by the Business Cycles Dating Committee of the National Bureau of Economic Research (NBER), which provides the 5 Note that phase shifts require not only changes in the magnitude of growth rates but also changes in the direction of movement in IP that persist for some time. More important, a peak (trough) can occur only after expansion (recession); in this sense, phase shifts are by definition statedependent phenomena. In contrast, the growth rate of IP changes at all points in time. 6 Seventy industries are defined at the four-digit level. The following four industries are defined at the three-digit level: apparel (NAICS 315), leather and allied products (NAICS 316), printing and related support activities (NAICS 323), and petroleum and coal products (NAICS 324).
118
THE REVIEW OF ECONOMICS AND STATISTICS Table 1.—Frequencies and Durations of Industry Cycles Duration Number Complete Duration of Cycles Cycle Expansion Contraction Asymmetry
NBER cycle Industry cycles Mean Median Maximum Minimum SD
5.0
27.4
23.8
3.8
6.2
10.5 11.0 16.0 3.0 2.5
14.4 13.3 38.7 8.4 4.9
8.9 7.4 31.0 4.1 4.7
5.3 5.0 9.2 2.7 1.3
1.9 1.5 10.3 0.5 1.6
Complete cycles are measured from trough to trough. For calculating the values in the panel headed “Duration,” we first compute the average durations of complete cycles, expansions, and contractions for each industry. Duration asymmetry, which is the average duration of expansion divided by that of contraction, is also measured first for each of the disaggregated industries. Then the summary statistics are calculated across industries.
most authoritative chronology of U.S. business cycles.7 Second, the same classic approach has been widely used in previous studies seeking to establish the stylized facts of the business cycle (see, for just a few of numerous examples, Zarnowitz, 1992; King & Plosser, 1994; Watson, 1994; and Bordo & Haubrich, 2010). Third, it does not require a particular definition of trend components from the raw series, allowing us to avoid the potential problems inherent in detrending methods.8 The implementation of Harding and Pagan (2002) has two stages: 1. Define a peak in a time series {yt }Tt=1 as occurring at time t if yt = max {yt−2 , yt−1 , yt , yt+1 , yt+2 } and a trough as occurring at time t if yt = min {yt−2 , yt−1 , yt , yt+1 , yt+2 }. That is, a peak (trough) occurs at time t if yt is higher (lower) than its two preceding and two succeeding observations. 2. Check whether these peaks and troughs satisfy the predetermined censoring rules described below. Censoring rules make sure that peaks and troughs alternate and that a phase and a complete cycle have minimum durations. If these requirements are not fulfilled, the least pronounced among adjacent turning points is eliminated. Following Harding and Pagan (2002), we set the minimum duration of a phase to be two quarters and that of a cycle to be five quarters.9
we include the corresponding statistics for the aggregate business cycle based on the NBER dates. The number and duration of complete cycles are measured from trough to trough. Employing peak-to-peak measures does not change the general features. Manufacturing industries have experienced more frequent phase shifts than the aggregate economy. During the sample period, 1972:I to 2011:IV, the U.S. economy experienced five trough-to-trough cycles, whereas manufacturing industries on average experienced 10.5 cycles. Consequently, the average duration of complete cycles is much shorter for manufacturing industries (14.4 quarters) than for the aggregate economy (27.4 quarters). Though less pronounced than for the aggregate economy, manufacturing industries also exhibit duration asymmetry between expansion and contraction phases. The duration of expansion is, on average, 1.9 times longer than that of recession at the individual industry level, while the same ratio for the U.S. economy is 6.2. There are large cross-sectional differences in the duration properties of industry cycles. For example, the average duration of complete cycles goes up to 38.7 quarters in the computer and peripheral equipment industry (NAICS 3341), while it drops to 8.4 quarters in the iron and steel products industry (NAICS 3311,2). The semiconductor and other electronic components industry (NAICS 3344) experienced, on average, the longest expansion, with a duration of 31.0 quarters, in sharp contrast to the minimum expansion duration of 4.1 quarters recorded for the apparel industry (NAICS 315). The cross-sectional differences in duration asymmetry are also quite striking. The average duration of expansion, for instance, is 10.3 times longer than that of recession for the computer and peripheral equipment industry (NAICS 3341), while it is just half of that of recession for the apparel industry (NAICS 315). III.
In this section we revisit the phenomenon of industry comovement in terms of the concentration of cyclical phases. To this end, we employ two measures of comovement: diffusion and concordance indices. The diffusion index measures the fraction of industries being in a contraction at a given point in time (Artis et al., 2004), using
B. Frequencies and Durations of Industry Cycles
Table 1 reports the summary statistics for frequencies and durations of the identified industry cycles. For comparison, 7 Nonetheless, it is important to note that the committee does not rely on any fixed dating rules when making its final decisions; see, for details, www.nber.org/cycles/recessions.html. 8 For example, Harvey and Jaeger (1993) and Cogley and Nason (1995) provide analyses of spurious cycles arising from the application of the Hodrick-Prescott filter. Canova (1998) illustrates how the different detrending methods generate different stylized facts of the U.S. business cycle. 9 The minimum duration requirement for a phase also prevents a turning point from occurring in the first and last two quarters of the sample.
Comovement: Diffusion and Concordance
Dt =
N i=1
wit Sit ,
N
wit = 1,
t = 1, . . . , T ,
(1)
i=1
where wit is the weight assigned to the ith industry at time t, Sit is a binary variable taking the value 1 if the ith industry is in a contraction and 0 otherwise, and N is the cross-sectional dimension. For the industry weights, we use equal shares (1/N) for all industries as our benchmark. We also check the robustness of the results to the use of (time-varying) output shares of each industry available from the Federal Reserve
ASYMMETRIC PHASE SHIFTS IN U.S. INDUSTRIAL PRODUCTION CYCLES Figure 1.—The Diffusion Index of Contraction
Table 2.—Concordance Indices
1 Equal weights Output share weights
0.9
119
Mean Median Maximum Minimum SD
0.8 0.7 0.6
Pairwise
NBER
0.604 0.606 0.863 0.319 0.080
0.675 0.669 0.894 0.450 0.086
“Pairwise” measures the concordance between individual industry cycles. “NBER” measures the concordance of industry cycles with the aggregate business cycle whose turning points are determined by the NBER.
0.5 0.4 0.3
of each industry with the aggregate business cycle is measured by
0.2 0.1 0
1975
1980
1985
1990
1995
2000
2005
2010
The diffusion index measures the fraction of industries in contraction at each point in time. The solid and dashed lines distinguish between the indices using equal and output share weights, respectively. The shaded vertical bars indicate U.S. recessions as dated by the NBER.
Board. Note that the fraction of industries in expansion is simply 1 minus the index of diffusion of contraction. Figure 1 displays how the measured diffusion index changes over time. The first thing to notice is that it moves repeatedly from near 0 to near 1 and back again, consistent with Burns and Mitchell’s characterization of comovement: the alternation of periods dominated by expansion and contraction. Also evident is its association with the aggregate business cycle. While the index stays far below 0.5 in most quarters of national expansions, it rises sharply above 0.5 at the onset of national recessions. More precisely, the average fraction of industries in contraction is 33.6% for the NBER expansions and 73.1% for the NBER recessions. Note that using output share weights has little effect on the results. The two NBER recessions in 1973–1975 and 2007–2009 deserve special attention, since the estimated index indicates that the level of industrial production declined in almost all industries during these two national recessions. By comparison, in other NBER recessions—1980, 1981– 1982, 1990–1991, and 2001—the production level continued to increase in more than 20% of industries. The figure also shows that there are several periods (1985, 1995, 2003, and 2006) in which more than 50% of industries experienced a contraction while the U.S. economy as a whole did not. Our second measure of comovement, the concordance index, measures the fraction of time that two cycles are in the same phase over the sample period (Harding & Pagan, 2002). This index can be used in two different ways. First, the concordance of phases between two industries is measured by Ci, j
T 1 = [ Sit Sjt + (1 − Sit )(1 − Sjt ) ], T t=1
(2)
where Sit and Sjt are binary variables indicating contractions of industry i and j, respectively. Similarly, the concordance
Ci,US
T 1 = [ Sit SUS,t + (1 − Sit )(1 − SUS,t )], T t=1
(3)
where SUS,t is a dummy variable indicating the NBER recession dates. Table 2 reports the summary statistics for the concordance indices computed over all 2,701 (= 74×73/2) pairwise combinations of industries (Pairwise) and the 74 pairs between industry cycles and the U.S. business cycle (NBER). The pairwise concordance indices range from 0.319 to 0.863, with a mean of 0.604, suggesting that a randomly chosen pair of industries are in the same cyclical phase about 60.4% of the time. The degree of concordance between individual industries and the aggregate U.S. economy is on average 0.675.10 Taken jointly, the results in this section confirm that in spite of large differences in the duration properties, industries tend to comove between cyclical phases in sync with the aggregate business cycle. IV.
Distributions of Turning Points
A. Concentration and Skewness
We now ask whether the distributions of industry turning points have the same concentration between the NBER peaks and troughs. To shed light on this issue, we define a turning point cluster as a set of industry turning points whose distances from given NBER turning points are less than a predetermined bound (e.g., eight quarters). Formally, the cluster is defined as follows. Let τPij be the jth peak of industry i and mk be the kth peak in the U.S. business cycle identified by the NBER. Then the kth peak cluster centered around mk is Ψk = τPij | d(mk − τPij )
< d(m − τPij ) for all = k; and d(mk − τPij ) ≤ d¯ , (4)
10 The concordance index has a shortcoming in that it is positively affected by the expected values of the phase indicators. To address this problem, we checked our results using the mean-corrected correlation index proposed by Harding and Pagan (2006). The results also indicated that most industries are positively synchronized with other industries, as well as with the aggregate economy.
120
THE REVIEW OF ECONOMICS AND STATISTICS
Figure 2.—Distributions of Industry Turning Points:All Six Recessions 0.35 peak
trough
Fraction of industries
0.30 0.25 0.20 0.15 0.10 0.05 0.00 -8
-7
-6
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
7
8
Leads (-) and lags (+) at the NBER dates
where d(·) is a measure of distance and d¯ is a predetermined cluster bound.11 Following Harding and Pagan (2006), we choose d¯ = 8 for our quarterly data. Clusters of industry troughs are defined in a similar fashion. Since the U.S. economy experienced six recessions during our sample period, we can construct six pairs of peak and trough clusters using this definition.12 Figure 2 compares the resulting distributions of industry turning points between peak and trough clusters. The horizontal axis denotes the lead (negative) and lag (positive) time over the NBER turning point dates. The vertical axis is the corresponding average fraction of industries, where the average is taken for each lead and lag time separately over the past six NBER peak and trough dates. In calculating the fraction of industries, we divide the number of industries shifting at that point by the total number of industries (74), not by the total number of shifts included in the cluster. Therefore, the choice of the cluster bound does not affect the results in figure 2, except that only the limits of the horizontal axis change with it. Inspection of figure 2 reveals sharp contrasts between the shapes of peak and trough clusters. First, clusters of industry troughs are highly concentrated at the NBER trough date, whereas clusters of industry peaks are much more dispersed. For troughs, 32% of industries on average exit simultaneously from the contraction phase at the NBER trough date. 11 This definition is based on Harding and Pagan (2006). The major difference between their work and ours is that they use this definition to extract the reference cycle dates, while we focus on how the peaks and troughs of industry cycles are distributed around the given NBER dates. 12 It is important to note that the above definition allows an industry turning point to appear in at most one cluster. This restriction limits the maximum lag for the 1980 peak and the maximum lead for the 1981 peak to two quarters and the maximum lag for the 1980 trough and the maximum lead for the 1982 trough to four quarters. In addition, due to data availability, the maximum lead for the 1973 peak is reduced to five quarters. Hence, careful attention needs to be paid to the results for the points that miss observations for some clusters. However, as will be seen from figure 4, which displays the turning point distributions for each NBER recession, our conclusion about the general shapes of clusters does not appear to be sensitive to these partial truncations.
For peaks, just about 13% of industries newly enter the contraction phase at the NBER peak date. Second, peak clusters are skewed toward leads, whereas trough clusters tend to be skewed toward lags. For peaks, the sums of the industry fractions over the left and right sides of the cluster are 80.3% and 39.8%. In trough clusters, the respective ratios are 37.6% and 57.1%.13 According to the Kolmogorov-Smirnov test using exact p values (Higgins, 2004), the null hypothesis of equal distribution between the peak and trough clusters is rejected at the 1% level.14 To check the stability of our results, we repeat the analysis separately for the first three (1973–1975, 1980, 1981–1982) and the last three recessions (1990–1991, 2001, 2007–2009). In the past decade, a large number of studies have found evidence of structural changes in the time series properties (especially volatility) of output growth since the mid-1980s at both the aggregate (Kim & Nelson, 1999; McConnell & Perez-Quiros, 2000) and industry levels (Stock & Watson, 2003; Foerster et al., 2011). Relative to these studies, our results in figure 3 suggest that the general shapes of turning point distributions are quite similar between the recessions before and after the mid-1980s.15 If we look only at trough clusters, the average fraction of industries coinciding with the NBER trough has slightly declined from 33.8% in the first three to 30.2% in the last three recessions. However, the asymmetry between peaks and troughs in the fraction of coincident industries was substantially larger in the last three recessions. In the first three, the fraction of coincident industries was on average 2.0 times higher at troughs than at peaks. The ratio increased to 3.2 in the last three recessions. The skewness feature has been qualitatively unchanged for both peaks and troughs. For peaks, the sum of industry fractions was 2.0 times larger in leads than in lags in both the first and last three recessions. For troughs, it was larger in lags than in leads, by 1.2 times in the first three and by 1.9 times in the last three recessions. Figure 4 shows the turning point distributions for each NBER recession. Although there have been some variations in details, both of the asymmetric patterns have emerged repeatedly in the past six NBER recessions. The fraction of coincident industries has almost always been more than twice as large at the NBER troughs than at the NBER peaks. One exception was the 1981–1982 recession, for which this ratio was 1.1. The maximum ratio was 8.2 for the 2001 recession, 13 Note that for neither the peak nor the trough clusters is the sum of the fractions of industries necessarily equal to 1. This is because some industries do not experience any cyclical turns during the time period spanned by the cluster and some industries experience mutliple turns during the same time period. 14 We adopt the randomized permutation test for exact inference, since our data on lead and lag times are discretely recorded on a quarterly basis. 15 We note, however, that our results do not necessarily conflict with the existing evidence on the Great Moderation—the increased length of expansions and reduced output volatility after the mid-1980s. Our emphasis is that although aggregate recessions became less frequent, once one occurred, clusters of industry turning points were formed with shapes similar to previous ones. This interpretation is also applied to the results of the stability analyses conducted below.
ASYMMETRIC PHASE SHIFTS IN U.S. INDUSTRIAL PRODUCTION CYCLES
121
Figure 3.—Distributions of Industry Turning Points: The First and Last Three Recessions First three
Last three
0.35
0.35 peak trough
0.25 0.20 0.15 0.10 0.05
peak trough
0.30 Fraction of industries
Fraction of industries
0.30
0.25 0.20 0.15 0.10 0.05
0.00
0.00 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
Leads (-) and lags (+) at the NBER dates
Leads (-) and lags (+) at the NBER dates
Figure 4.—Distributions of Industry Turning Points: Each NBER Recession 1973-75 recession
1980 recession
0.45
0.45
Fraction of industries
0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00
0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00
-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
Leads (-) and lags (+) at the NBER dates
Leads (-) and lags (+) at the NBER dates
1990-91 recession
1981-82 recession 0.45
0.45 Fraction of industries
0.35 0.30 0.25 0.20 0.15 0.10 0.05
0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00
0.00
-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
Leads (-) and lags (+) at the NBER dates
Leads (-) and lags (+) at the NBER dates
2001 recession
2007-09 recession
0.45
0.45 peak trough
0.30 0.25 0.20 0.15 0.10 0.05 0.00
0.40 Fraction of industries
Fraction of industries
0.35
peak trough
0.40 Fraction of industries
peak trough
0.40
0.40
peak trough
0.40 Fraction of industries
peak trough
0.40
0.35
peak trough
0.30 0.25 0.20 0.15 0.10 0.05 0.00
-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 Leads (-) and lags (+) at the NBER dates
and for the most recent recession in 2007–2009, this ratio was 2.8. With respect to skewness, industry peaks have always been skewed toward leads except for the NBER peak in 1973. Industry troughs have been skewed toward lags except for the NBER trough in 1980. We performed several additional analyses, the results of which are given in the online appendix. There, we found that neither employing different levels of disaggregation (three and six digits) nor using detrended IP series obtained from
-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 Leads (-) and lags (+) at the NBER dates
the Hodrick-Prescott and bandpass filters yielded qualitatively different results from those discussed here. We also performed a simulation exercise and found that asymmetric distributions of turning points are not necessarily a simple reflection of duration asymmetry—long expansions and short recessions. Overall, our results of asymmetries between peak and trough clusters appear to be quite robust and provide important new information about the characteristics of industry comovement.
122
THE REVIEW OF ECONOMICS AND STATISTICS Table 3.—Transition Probability Current
Previous Peak to peak Leading Coincident Lagging Acyclical Trough to trough Leading Coincident Lagging Acyclical Peak to trough Leading Coincident Lagging Acyclical
Leading
Coincident
Lagging
Acyclical
0.599 0.519 0.547 0.490
0.135 0.058 0.173 0.137
0.167 0.173 0.133 0.177
0.099 0.250 0.147 0.196
0.342 0.271 0.163 0.324
0.317 0.312 0.341 0.189
0.293 0.328 0.357 0.324
0.049 0.090 0.140 0.162
0.226 0.305 0.130 0.273
0.396 0.407 0.210 0.109
0.335 0.254 0.630 0.164
0.044 0.034 0.030 0.455
Each (i, j) cell reports the average percentage of industries moving into group j among the industries having been in group i between two adjacent NBER peak dates (peak to peak), between two adjacent NBER trough dates (trough to trough), and between the start and the end of an NBER recession (peak to trough). The average is taken over all the past NBER recessions in the sample.
B. Leading, Coincident, and Lagging Industries
Based on the identified turning points for disaggregated industries, we classify industries into leading, coincident, lagging, and acyclical groups for each of the NBER peak and trough dates. Leading (lagging) industries are those whose turning points come earlier (later) than the NBER turning points. Consistent with the previous clustering analysis, we restrict the maximum lead and lag times to eight quarters. Coincident industries are those whose turning points coincide with the NBER turning points. If an industry does not experience a turning point during the time period spanned by the cluster, we define it as acyclical.16 Suppose we observed an industry that declined before the current national recession. How likely is it that the industry will exit the recession before the aggregate economy? Will it lead the next national recession again? More generally, how persistent are the identities of leading, coincident, and lagging industries? This question is important because it may shed additional light on the extent to which the impulse and propagation mechanisms are all alike over the course of business cycles (Lucas, 1977). In addition, if industries move between the four groups in a purely random way, the knowledge of the past leading industries will be of little use in forecasting future turning points of the national business cycle. Table 3 reports the average probabilities that an industry moves from group i to group j between two adjacent NBER peaks (peak to peak), between two adjacent NBER troughs (trough to trough), and between the start and the end of an 16 Note
that an industry may experience multiple peaks (troughs) during the time period spanned by a peak (trough) cluster. In that case, we use minimum distance criterion; that is, for instance, if an industry exhibits two peaks, of which one is marked in the left and the other is marked in the right half of the cluster, and if the lead time is shorter than the lag time, we classify the industry as leading.
NBER recession (peak to trough).17 The average is taken over all the recessions in our sample. Because there are four groups, the estimated transition probabilities are given by a 4 × 4 matrix. The elements of each row sum to 1, and the diagonal elements represent the persistence of identity for each group. The results in table 3 show that for peaks, we find strong persistence among leading industries; an industry that led the current national peak will again lead the next national peak with an average probability of 0.599. But we find little persistence among coincident (0.058) and lagging (0.133) industries. This finding is consistent with the fact that many manufacturing industries tend to lead the aggregate peaks. By comparison, for troughs, the average persistence of leading industries is reduced to 0.342, whereas that of coincident and lagging industries rises to 0.312 and 0.357, respectively. This is also consistent with our previous finding that troughs are more concentrated than peaks and slightly skewed to lags. Regarding the question of whether an industry that leads the national recession also leads the national recovery, we again find strong persistence among coincident (0.407) and lagging industries (0.630), but much less persistence for leading industries (0.226). Table 4 lists the industries that have led, lagged, and coincided with the business cycle on more than three occasions over the past six NBER peak dates (more than 50% of the time).18 For the NBER peaks, 30 (nineteen durables and eleven nondurables) of the 74 industries are defined as leading industries according to this 50% cutoff rule.19 We find that three industries—lime and gypsum products (NAICS 3274), cutlery and handtools (NAICS 3322), and motor vehicles (NAICS 3361)—have led all the NBER peaks in the past six recessions. By comparison, when the same cutoff is used, no industry is defined as coincident, and only four industries are defined as lagging. The corresponding lists for the NBER trough dates are reported in table 5. Here, the number of leading industries is reduced to four, whereas that of coincident and lagging industries increases to nine and thirteen, respectively. Based on the results in tables 4 and 5, we can also easily construct the lists of industries that pass the 50% cutoff for both NBER peaks and troughs. The results are quite intriguing. In this case, only the sawmills and wood preservation industry (NAICS 3211) is defined as leading. Due to the results for peaks, there are only three lagging industries and no coincident industry. These results point to another aspect of asymmetries between peak and trough clusters: they are asymmetric not only in the general shape but also in 17 Our analysis can easily be extended to deal with transitions across three or more NBER turning point dates. We do not take that step in this paper for conciseness. 18 The 50% criterion is arbitrary and chosen for explanatory purposes. We leave rigorous statistical analysis to follow-up studies. 19 The classification between durables and nondurables is based on the definition by the Federal Reserve Board.
ASYMMETRIC PHASE SHIFTS IN U.S. INDUSTRIAL PRODUCTION CYCLES
123
Table 4.—Leading, Coincident, and Lagging Industries at the NBER Peak Dates Leads (−) and Lags (+) NAICS
Type
Leading industries 3274 D 3322 D 3361 D 3371 D 3351 D 3255 ND 3253 ND 3352 D 3212 D 3325 D 3362 D 3219 D 3279 D 3252 ND 324 ND 3131 ND 3372,9 D 3273 D 3221 ND 3133 ND 3315 D 3272 D 3343 D 3118 ND 3334 D 3261 ND 3122 ND 3149 ND 3211 D 3363 D Lagging industries 3113 ND 3345 D 3121 ND 3111 ND
Description
Probability
Mean
SD
Lime and gypsum product Cutlery and handtool Motor vehicle Household furniture and kitchen cabinet Electric lighting equipment Paint, coating, and adhesive Pesticide, fertilizer, and other agriculture chemical manufacturing Household appliance Veneer, plywood, and engineered wood product Hardware Motor vehicle body and trailer Other wood product Other nonmetallic mineral product Resin, synthetic rubber Petroleum and coal products Fiber, yarn, and thread mills Office and other furniture Cement and concrete product Pulp, paper, and paperboard mills Textile and fabric finishing mills Foundries Glass and glass product Audio and video equipment Bakeries and tortilla Ventilation, heating, air-conditioning, and refrigeration equipment Plastics product Tobacco Other textile product mills Sawmills and wood preservation Motor vehicle parts
1.00 1.00 1.00 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67
−2.67 −3.00 −3.00 −1.80 −2.40 −2.40 −2.60 −3.00 −3.60 −4.40 −4.40 −4.60 −2.25 −2.25 −2.50 −2.50 −2.50 −3.25 −3.25 −3.75 −3.75 −3.75 −4.00 −4.00 −4.25 −4.25 −4.25 −4.50 −4.50 −4.50
2.73 1.41 1.79 1.30 1.34 1.52 2.07 2.35 2.07 2.07 2.70 2.61 1.50 1.89 0.58 1.00 1.29 1.50 2.63 1.26 2.50 2.75 1.83 2.94 2.06 2.50 3.78 1.73 2.08 2.08
Sugar and confectionery product Navigational, measuring, electromedical, and control instruments Beverage Animal food
0.67 0.67 0.67 0.67
1.50 2.25 4.00 4.25
1.00 1.50 1.83 2.06
D: durables; ND: nondurables; Probability: percentage of industries belonging to the group over the past six NBER peaks; mean and SD: the mean and the standard deviation of the lengths of leads (lags) when the industry leads (lags) the NBER peak.
the detailed composition of leading, coincident, and lagging industries.20 C. Sharpness Asymmetry
Thus far, we have focused on the timing distributions of industry turning points. This section explores the difference between peak and trough clusters in the sharpness of changes in log IP surrounding the NBER turning points. As mentioned in section I, McQueen and Thorley (1993) find evidence of rounded peaks and sharp troughs using the aggregate IP index over the period 1947–1990. We extend their work to disaggregated industries and to the more recent sample period, 1972–2011. Following McQueen and Thorley (1993), we characterize the sharpness of IP change surrounding an NBER turning 20 In the online appendix, we repeat the analyses in tables 3 to 5 separately for the first three and last three recessions. The results indicate that while there have been considerable variations over time in the detailed composition of each group of industries, our general conclusions are robust: peaks and troughs are highly asymmetric in both the persistence of industry group identities and the composition of leading, coincident, and lagging industries.
point using the mean absolute difference between the IP growth rate during the two quarters ending in the NBER date and that during the following two quarters. Sharpness asymmetry is then measured by the difference in sharpness between the NBER troughs and peaks. The results of these measures are reported in table 6, in which three points are noteworthy. First, troughs are sharper than peaks not only at the aggregate level (McQueen & Thorley, 1993) but also at the disaggregated industry level. Over the past six recessions, the cross-industry mean sharpness is 0.063 at the NBER peaks and 0.129 at the NBER troughs, suggesting that troughs are about twice as sharp as peaks. Second, the average IP growth rate is small and negative (−0.002) in the last two quarters of expansions, whereas it is positive (0.036) in the first two quarters of expansions. In contractions, the decline of log IP is larger in the later stages (−0.080) than in the early stages (−0.046).21 Both of these patterns contribute positively to 21 Previous studies using aggregate measures of output also have found that expansion is faster in its early stages and contraction is steeper in its later stages; see Sichel (1994) and Lam (2004), among others. Our results are compatible with this view.
124
THE REVIEW OF ECONOMICS AND STATISTICS Table 5.—Leading, Coincident, and Lagging Industries at the NBER Trough Dates Leads (−) and Lags (+) NAICS
Type
Leading industries 3366 D 3211 D 3113 ND 3364 D Coincident industries 3149 ND 3132 ND 3261 ND 3272 D 3279 D 3315 D 3325 D 3371 D 3311,2 D Lagging industries 3333,9 D 3345 D 3353 D 3118 ND 3256 ND 3328 D 3336 D 3335 D 3331 D 3122 ND 3121 ND 3111 ND 3365 D
Description
Probability
Mean
SD
Ship and boat building Sawmills and wood preservation Sugar and confectionery product Aerospace product and parts
0.67 0.67 0.67 0.67
−1.50 −2.00 −2.00 −3.75
1.00 1.16 2.00 2.06
Other textile product mills Fabric mills Plastics product Glass and glass product Other nonmetallic mineral product Foundries Hardware Household furniture and kitchen cabinet Iron and steel products
0.83 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Commercial and service industry machinery Navigation, measuring, electromedical, and control instruments Electrical equipment Bakeries and tortilla Soap, cleaning compound, and toilet preparation Coating, engraving, heat treating Engine, turbine, and power transmission equip. Metalworking machinery Agriculture, construction, and mining machinery Tobacco Beverage Animal food Railroad rolling stock
0.83 0.83 0.83 0.83 0.83 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67
1.80 2.20 2.80 3.00 3.00 1.00 1.25 2.00 2.75 3.00 3.25 3.75 4.50
1.30 1.30 2.39 2.45 2.55 0.00 0.50 0.82 2.06 1.41 2.22 0.50 2.38
See the note to table 4.
Table 6.—Changes in IP Growth Rates Surrounding the NBER Peak and Trough Dates NBER Recessions All Six Sharpness of peaks −0.002 DP,−2 −0.046 DP,+2 0.063 SP Sharpness of troughs −0.080 DT ,−2 0.036 DT ,+2 0.129 ST Sharpness asymmetry 0.066 ST − SP
First Three
Last Three
1973– 1975
1980
1981– 1982
1990– 1991
2001
2007– 2009
0.010 −0.051 0.071
−0.014 −0.041 0.055
0.024 −0.005 0.042
0.003 −0.082 0.087
0.001 −0.068 0.083
−0.002 −0.051 0.058
−0.033 −0.038 0.054
−0.007 −0.033 0.052
−0.086 0.046 0.146
−0.074 0.025 0.111
−0.139 0.036 0.193
−0.082 0.046 0.132
−0.038 0.057 0.112
−0.051 0.027 0.086
−0.040 0.023 0.079
−0.131 0.027 0.170
0.075
0.057
0.152
0.045
0.029
0.028
0.024
0.118
Letting DiP,−2 and DiP,+2 denote, respectively, the two-quarter growth rates of industry i before and after the NBER peaks, DP,−2 = mean (DiP,−2 ), DP,+2 = mean (DiP,+2 ), and SP = mean (|DiP,−2 − DiP,+2 |). DT ,−2 , DT ,+2 , and ST are defined in a similar way for the NBER troughs. Sharpness asymmetry is ST minus SP . The null of symmetric sharpness is rejected by the Welch t-test at the 5% level for all cases in this table.
more sharpness at troughs than at peaks. Third, the results are qualitatively the same in both the first three and last three recessions, although the degree of sharpness asymmetry slightly declined from 0.075 to 0.057.22 Furthermore, for 22 This decline was mainly due to the slower growth rates in the last three recovery phases; the cross-industry mean growth rate during the two quarters after an NBER trough decreased from 0.046 in the first three to 0.025 in the last three recessions. It is important to note, however, that the growth rate in the recovery phase is just one of four parts defining our measure of sharpness asymmetry—the two two-quarter growth rates before and after the peak and those before and after the trough. Putting all parts together, the growth rates changes surrounding the NBER troughs (0.111) far exceeded those surrounding the NBER peaks (0.055) in the last three recessions as in the first three.
each of the past six NBER recessions, the Welch t-test rejects at the 5% level the null of no sharpness asymmetry in favor of more sharpness at troughs. For example, looking at the 2007– 2009 recession, the sharpness of the trough (0.170) was more than three times larger than that of the peak (0.052). This was because at the peak, the two-quarter IP growth rate of −0.007 was followed by that of −0.033, while at the trough, a very large negative growth rate (−0.131) was replaced by a positive, albeit small, growth rate (0.027). To further explore the effect on sharpness asymmetry of the above-discussed asymmetric concentration of turning points, table 7 compares the sharpness of coincident industries with that of noncoincident—leading, lagging,
ASYMMETRIC PHASE SHIFTS IN U.S. INDUSTRIAL PRODUCTION CYCLES
125
Table 7.—Sharpness of Coincident and Noncoincident Industries NBER Recessions All Six Coincident industries 0.089 SP 0.212 ST 0.122 ST − SP Noncoincident industries 0.058 SP 0.089 ST 0.030 ST − SP
First Three
Last Three
1973– 1975
1980
1981– 1982
1990– 1991
2001
2007– 2009
0.094 0.244 0.149
0.085 0.180 0.095
0.058 0.348 0.290
0.110 0.193 0.083
0.115 0.190 0.075
0.093 0.156 0.064
0.078 0.113 0.036
0.084 0.270 0.186
0.065 0.094 0.029
0.051 0.083 0.032
0.039 0.105 0.065
0.082 0.089 0.007
0.075 0.089 0.014
0.051 0.056 0.005
0.054 0.061 0.007
0.049 0.133 0.084
Noncoincident industries include leading, lagging, and acyclical industries. Industry groupings are based on the results in section IVB.
and acyclical—industries.23 Our logic is straightforward: if coincident industries show more sharpness than noncoincident ones at the NBER turning points, the higher fraction of coincident industries at troughs than at peaks can positively affect the difference between troughs and peaks in the cross-industry mean of sharpness. As seen in table 7, the data support this prediction: coincident industries exhibited sharper changes than noncoincident industries for all of the past NBER turning points. However, this table also shows that the sharpness asymmetry is still evident among both groups of industries. Therefore, we can decompose the cross-industry mean sharpness asymmetry into two parts: one resulting from the difference between troughs and peaks in the fraction of coincident industries and the other due to the sharpness asymmetry inherent in individual groups. The decomposition is achieved by ST − SP =
2
1 (wg,T − wg,P ) × (Sg,P + Sg,T ) 2 g=1 Composition effect
+
2 1 (Sg,T − Sg,P ) × (wg,P + wg,T ), 2 g=1 Group-level sharpness asymmetry
(5)
where g distinguishes coincident industries (g = 1) from noncoincident (g = 2); P and T , respectively, indicate the NBER peaks and troughs; w denotes the fraction of industries belonging to the group; and S denotes the sharpness of the group. Table 8 reports each term from this decomposition estimated separately for each of the past NBER recessions. According to the decomposition, it is the individual-grouplevel characteristic that accounts for the lion’s share of the sharpness asymmetry observed at the cross-industry mean estimates. However, concentration asymmetry also plays a significant role. Its contribution is on average 21.5%, with the lowest of 6.6% for the 1981–1982 recession and the largest of 47.4% for the 2001 recession. 23 In the online appendix, we present the results of sharpness for leading, lagging, and acyclical industries.
Table 8.—Decomposition of Sharpness Asymmetry NBER Sharpness Group-Level Sharpness Recessions Asymmetry Composition Effect (%) Asymmetry (%) All six First three Last three 1973–1975 1980 1981–1982 1990–1991 2001 2007–2009
0.066 0.075 0.057 0.152 0.045 0.029 0.028 0.024 0.118
0.014 0.016 0.012 0.030 0.016 0.002 0.011 0.011 0.015
(21.5) (21.3) (21.8) (19.8) (35.5) (6.6) (37.1) (47.4) (12.8)
0.052 0.059 0.044 0.122 0.029 0.027 0.018 0.013 0.102
(78.5) (78.7) (78.2) (80.2) (64.5) (93.4) (62.9) (52.6) (87.2)
The cross-industry mean sharpness asymmetry (“Sharpness Asymmetry”) is decomposed into two parts: the part ascribed to the difference between troughs and peaks in the fraction of coincident industries (“Composition Effect”), and the part attributed to the sharpness asymmetry observed at the individualgroup level (“Group-Level Sharpness Asymmetry”). The values in parentheses are the share of sharpness asymmetry (in percentage terms) explained by each source.
V.
Determinants of Comovement
In this section we take up the question of what determines the industry comovement. Specifically, we ask whether common macroeconomic shocks and interindustry linkages, emphasized by the existing literature as two main sources of the comovement, have a significant effect on the concurrence of industry turning points. We also examine whether the effects are (a)symmetric between peaks and troughs. A. Empirical Model
For the occurrence of peaks, the empirical model is dit∗ = xit θ + uit =α+
p¯ k¯ k=1 p=1
dit = 1(dit∗ > 0),
zk,i,t−p βk,p +
p¯ ¯
ξ,t−p γ,p + uit ,
=1 p=1
for i = 1, . . . , N and t = 1, . . . , T , (6)
where dit∗ is a latent unobservable variable; dit is the observed binary outcome taking the value 1 if industry i is at a peak at time t and 0 otherwise; 1(·) denotes an indicator function equal to 1 if the condition in parentheses is satisfied and 0 otherwise; xit is a vector of observable covariates; θ is a vector of corresponding coefficients; and uit is the error term. The observable explanatory variables include the lags of two types of spillover effects from the peaks in the neighbor industries, zk,i,t−p , k = 1, 2, and the lags of four macroeconomic shocks
126
THE REVIEW OF ECONOMICS AND STATISTICS
common to all industries, ξ,t−p , = 1, . . . , 4. These variables are lagged by one to eight quarters (p = 1, . . . , 8) to avoid a potential endogeneity problem and ensure sufficient length of propagation. We explain details of the explanatory variables in the next section. The model for troughs can be specified in a similar way. Our baseline model is the standard random-effects panel probit model, assuming i.i.d.N(0, στ2 )
uit = τi + it , τi ∼ and it ∼ i.i.d.N(0, 1) (7) conditional on xi = (xi1 , . . . , xiT ), where τi is an industry-specific time-invariant component that captures unobserved heterogeneity in the duration of expansion phases and it is an idiosyncratic disturbance that changes across t as well as i. It is also assumed that τi and it are independent of each other. In section VD, we show that the main results are not affected by different assumptions about the error term (e.g., models with serially correlated disturbances or those with fixed effects). Letting Φ and φ denote, respectively, the standard normal distribution and density functions, the model parameters are estimated by maximizing the conditional likelihood,
T
N ∞ L= Prob (dit |xit , τi ; θ) φ(τi )dτi , (8) i=1
−∞
t=1
where Prob(dit |xit , τi ; θ) = Φ(xit θ + τi )dit [1 − Φ(xit θ + τi )]1−dit . We use the maximum simulated likelihood estimation to approximate the integral in equation (8) (see Train, 2009). Since dit = 0 with probability 1 for some t simply because of the censoring rule used for detecting turning points, we discard the 0 responses at those points and compute the likelihood function for the remaining binary responses.24 B. Explanatory Variables
The explanatory variables are grouped into two categories. The first group consists of the weighted averages of spillover effects from other industries’ phase shifts, constructed as zk,i,t−p = wk,ij dj,t−p , (9) j =i
where dj,t−p is a dummy variable assigning the value 1 to industry j’s peak (in the case of the peak equation) or trough (in the case of the trough equation) having occurred at time t − p (p = 1, . . . , 8) and wk,ij (k = 1, 2) is a weight capturing the importance of industry j for industry i. Following Bartelsman et al. (1994) and Shea (2002), we distinguish two types of spillover effects depending on their origin. The first is from output users (upstream, or demand 24 To be more concrete, the time points at which a peak (trough) cannot occur by the censoring rule are the first and last two quarters of the sample, the first four quarters after a peak (trough), one quarter after a trough (peak), and the quarters in a contraction (expansion) phase.
side), and the second is from input suppliers (downstream, or supply side). Let mij be the value of a commodity (in producers’ prices) produced by industry i and used in industry j. Then the upstream spillover variable is constructed using mij w1,ij = , j=i mij which indicates the importance of industry j as a user of the product of industry i. Similarly, the downstream spillover variable uses mji w2,ij = , j=i mji which captures the importance of industry j as an input supplier to industry i. We compute these weights using the detailed benchmark input-output accounts available from the Bureau of Economic Analysis (BEA). In contrast to IP data, which are disaggregated by the NAICS system, the inputoutput tables for the years prior to 1997 are available based on only the Standard Industry Classification (SIC) system. Since there is no easy way to convert them to the NAICS codes, we use constant weights drawn from the input-output table for 1997. The second group of explanatory variables consists of four macroeconomic shocks: monetary policy, government spending, oil price, and financial shocks. We identify monetary policy shocks as shocks to the quarterly federal funds rate in a standard monetary VAR model using a recursive identification scheme (Christiano, Eichenbaum, & Evans, 1999).25 For government spending shocks, we use Ramey’s (2011) measure of defense news shocks, defined as the present discounted value of the expected changes in future defense spending divided by the previous quarter’s nominal GDP. The oil price shocks we use are the nominal net oil price increase constructed following Hamilton (2003): the maximum of 0 and the log difference between the crude oil price for the current quarter and its previous peak in the past three years. For financial shocks, we use Gilchrist and Zakrajšek’s (2012) shocks to the excess bond premium, a component of corporate bond credit spreads that is not explained by the expected default risk of issuers. We replicated their VAR model to make the shocks orthogonal to the current state of the economy. In section VD below and the online appendix, we consider different measures of these four macroeconomic shocks and confirm the robustness of our results.26 25 We use a four-variable VAR system that includes the log of real GDP, the log of the GDP deflator, the log of the commodity price index, and the federal funds rate, in this order. As is standard, we estimate the VAR using four lags of the variables and a constant. The federal funds rate shocks are estimated over the period 1973:III–2010:III, the sample period of financial shocks we use. A Cholesky decomposition is used under the assumption that while monetary policy reacts to current economic conditions, the economy responds to changes in the federal funds rate only with a lag. It is worth noting that our measure of monetary policy does not capture the unconventional monetary policy—for example, large-scale asset purchases—implemented in recent years. 26 See the online appendix for a detailed description of the data used and how we construct the macroeconomic shocks.
ASYMMETRIC PHASE SHIFTS IN U.S. INDUSTRIAL PRODUCTION CYCLES
The estimation period is 1973:III–2010:III according to the availability of the Gilchrist and Zakrajšek (2012) financial shocks. In section VD, we will check whether the results are different when we restrict the sample to the post-1984 period. All explanatory variables are normalized to unit variance after setting the mean to 0 in order to facilitate comparison across shocks. C. Results
Direct interpretation of the parameters is difficult in a binary response model because the model is expressed as a nonlinear function of covariates. Therefore, our discussion of the estimation results will be based on the average marginal effects (AME), given by
N T 1 1 ∂ Prob (dit = 1|xit , τi ; θ) . N i=1 T t=1 ∂xit Again we compute this function only for the observations not censored. To save space, the detailed estimates of this function and the underlying coefficients are provided in the online appendix. The summarized results are plotted in figure 5, which shows the cumulative impacts of a 1 standard deviation increase in the explanatory variables, together with 1 standard error bands.27 Table 9 reports the estimated standard deviations of the explanatory variables. Panels A and B in figure 5 show that both the upstream and downstream spillover effects are significant: an industry’s probability of experiencing a phase shift increases with the number of output users and input suppliers having experienced the same type of phase shifts. However, their relative importance is somewhat asymmetric between peaks and troughs. For peaks, while the upstream effect increases steadily until it reaches 9.7%p (percentage point) after eight quarters, the downstream effect begins to weaken after reaching 4.3%p in the fifth quarter and becomes insignificant after seven quarters. Thus, for industries entering a recession, the upstream effect is more persistent and stronger than the downstream effect. By contrast, when exiting the recession, industries are more affected by their input suppliers than by output users: for troughs, the downstream effect is significantly positive in all eight quarters, with the maximum of 5.5%p attained in the seventh quarter. Relative to this, the upstream effect on troughs is much smaller, shorter lived, and less significant. After reaching 2.8%p at its maximum in the third quarter, the upstream effect on troughs begins to decline and eventually becomes negative. Panel C displays the estimated cumulative effects of monetary policy shocks. The results are generally consistent with theoretical predictions for the contractionary effect of tight monetary policy: an unanticipated increase in the federal funds rate increases the probability of a peak (the end of 27 Because no lags of the dependent variable are included in the model, the cumulative effect after m quarters is just the sum of the coefficients on the first m lags of the explanatory variables.
127
expansion) and decreases the probability of a trough (the end of contraction). Both impacts are statistically and economically significant. In expansions, an increase in the federal funds rate of 1 standard deviation, approximately equal to 1 percentage point as reported in table 9, increases the probability of terminating the expansion by 6.0%p eight quarters after the shock. In recessions, the effect of the monetary policy shock is delayed by four quarters. The maximum effect is a fall in the probability of exiting the recession by 7.9%p seven quarters after the shock. Panel D shows that a positive government spending shock has a significant expansionary effect as expected. Following a 1 standard deviation increase in the Ramey (2011) defense spending news series, which is equivalent to 1.2% of the previous quarter’s GDP, the probability of a peak decreases by 2.6%p for the next five quarters and then returns to its preshock level. The impact is faster and stronger in periods of recession. In response to a government spending increase of the same magnitude, the probability of a trough increases by 4.3%p for the first three quarters before it reverts to the preshock level. As shown in panel E, oil price shocks are also an important determinant of industry phase shifts. In expansions, a 1 standard deviation increase in the Hamilton (2003) series— an increase of 6.65% above the peak in the previous three years—increases the likelihood of transition from expansion to contraction, with the maximum effect of 4.9%p six quarters after the shock. In recessions, it lowers the probability of exiting the recession, with its maximum effect of −9.2%p in the third quarter.28 Finally, the effects of financial shocks are presented in panel F. According to Gilchrist and Zakrajšek (2012), an increase in the excess bond premium indicates a reduction in the effective supply of credit. Therefore, a positive shock to the premium is expected to have a negative output effect through the financial accelerator mechanism (Kiyotaki & Moore, 1997; Bernanke, Gertler, & Gilchrist, 1999; Christiano, Motto, & Rostagno, 2013). The empirical results in panel F confirm this prediction: in expansions, an increase of 1 standard deviation in the excess bond premium—94 basis points—leads to a more than 3%p increase in the probability of the expansion ending eight quarters after the shock. In recessions, the probability of going back to recovery drops by 7.5%p for the next seven quarters. In sum, both the input-output linkages and the four macroeconomic shocks we consider are important determinants of industry comovement. All of them are statistically significant and conform with our economic priors. The novelty of our study is to shed new light on their importance for the cross-industry averages of the probability of cyclical 28 Our results for the effects of oil price shocks do not necessarily conflict with the conventional view that an oil price increase has a larger output effect than an oil price decrease (see Hamilton, 2003). The reason is clear because an asymmetry related to the direction of an oil price change does not imply and is not implied by an asymmetry related to the responses of peaks and troughs to a given change in oil price.
128
THE REVIEW OF ECONOMICS AND STATISTICS Figure 5.—Estimated Impacts on the Probabilities of Industry Phase Shifts: Baseline Results A. Upstream spillover Trough Change in probability
Change in probability
Peak 0.10 0.05 0.00 −0.05 −0.10 1
2
3
4 5 Quarters
6
7
0.10 0.05 0.00 −0.05 −0.10
8
1
2
3
4 5 Quarters
6
7
8
6
7
8
6
7
8
6
7
8
6
7
8
6
7
8
B. Downstream spillover Trough Change in probability
Change in probability
Peak 0.10 0.05 0.00 1
2
3
4 5 Quarters
6
7
0.10 0.05 0.00
8
1
2
3
4 5 Quarters
C. Monetary policy shock Trough Change in probability
Change in probability
Peak 0.10 0.05 0.00 −0.05 −0.10 1
2
3
4 5 Quarters
6
7
0.10 0.05 0.00 −0.05 −0.10
8
1
2
3
4 5 Quarters
D. Government spending shock Trough Change in probability
Change in probability
Peak 0.05 0.00 −0.05 1
2
3
4 5 Quarters
6
7
0.05 0.00 −0.05
8
1
2
3
4 5 Quarters
E. Oil price shock Trough Change in probability
Change in probability
Peak 0.10 0.05 0.00 −0.05 −0.10 1
2
3
4 5 Quarters
6
7
0.10 0.05 0.00 −0.05 −0.10
8
1
2
3
4 5 Quarters
F. Financial shock Trough Change in probability
Change in probability
Peak 0.05 0.00 −0.05 −0.10 1
2
3
4 5 Quarters
6
7
8
0.05 0.00 −0.05 −0.10 1
2
3
4 5 Quarters
Sample period: 1973:III–2010:III. The thick solid lines are the cumulative marginal effects of a 1 standard deviation increase in the potential sources of industry comovement on the cross-industry mean probability that a four-digit NAICS industry will experience a turning point in the next m quarters (m = 1, . . . , 8). The left column plots the effects on peaks, and the right column plots the effects on troughs. The thin solid lines represent 1 standard error bands.
turns and, thus, for the concurrence of industry turning points. D. Robustness and Stability
This section addresses the robustness of our findings along a number of dimensions. We consider (a) alternative measures of the macroeconomic shocks, (b) different assumptions about the error term of the estimation model, (c) the use of the input-output matrix for 1977 instead of that for 1997, and
(d) a different sample period starting in 1984:I. The results of (a) through (c) are summarized in figure 6, and the results of (d) are presented in figure 7.29 In figure 6, we first replace monetary policy and oil price shocks with those identified from the Kilian and Lewis (2011) 29 In the online appendix, we perform several further sensitivity analyses. The results support the robustness of our main results to further alternative measures of the macroeconomic shocks, to controlling for the duration dependence of phase-termination probability, and to the use of IP disaggregated at the three-digit level.
ASYMMETRIC PHASE SHIFTS IN U.S. INDUSTRIAL PRODUCTION CYCLES Table 9.—Standard Deviations of Spillover Indices and Macroeconomic Shocks Sample Period
Upstream spillover Peak Trough Downstream spillover Peak Trough Monetary policy shock Government spending shock Oil price shock Financial shock
1973:III–2010:III (A)
1984:I–2010:III (B)
Relative Standard Deviation (B/A)
7.83 8.57
7.77 7.99
0.99 0.93
10.24 11.32 0.94 1.20
10.13 10.38 0.56 1.25
0.99 0.92 0.59 1.04
6.65 0.94
6.29 1.00
0.95 1.06
This table shows the standard deviations of the explanatory variables used in the baseline and the post1984 sample analyses. The standard deviations of the spillover variables are multiplied by 100 for ease of comparison. The standard deviations of macroeconomic shocks are measured in percentage points. The relative standard deviation is the post-1984 sample standard deviation divided by the full-sample standard deviation.
VAR model. This exercise allows us to check the sensitivity of our findings to the use of the growth rate of the real oil price in lieu of Hamilton’s (2003) net increase of the nominal oil price and the identification of the monetary policy shock based on a different VAR specification.30 The results in figure 6 indicate that using these instruments, if anything, tends to slightly reinforce the contractionary impacts of monetary policy and oil price shocks discussed above. We next replace the Ramey (2011) defense news shock with the government spending shock identified from a standard VAR model. To this end, we reestimated the VAR model used in figure 4 of Ramey (2011).31 As figure 6 shows, our results are qualitatively robust to this alternative: an increase in the VAR-based government spending shock also leads to a fall in the probability of a peak and a rise in the probability of a trough. We also repeated the analysis using the financial conditions index constructed by Hatzius et al. (2010) as an alternative measure of financial shocks. Compared to the Gilchrist and Zakrajšek (2012) excess bond premium series, this composite index has the advantage that it covers a wide range of financial indicators, not only the bond market indicator.32 Since increases in this index imply improved financial market conditions as opposed to the credit spread shock used in our baseline analysis, in panel F of figure 6, we present the effects 30 The Kilian and Lewis (2011) VAR model consists of the log difference of the real price of commodities, the log difference of the real price of oil (refiner’s acquisition cost of crude oil, composite), the Chicago Fed’s National Activity Index (CFNAI), the log difference of the consumer price index, and the federal funds rate (in this order). We estimated the model in exactly the same way as described in their paper. 31 Ramey’s (2011) VAR model includes the log of real per capita government spending, the log of real per capita GDP, the log of real per capita consumption of nondurables and services, the log of real per capita private fixed investment, the log of per capita total hours worked, the log of real product wage, and the average marginal income tax rate. See her paper for details. 32 The Hatzius et al. (2010) index is computed as the first principal component of 45 financial series. Among various versions of the indices they construct, we use the index that is purged of the effects of business cycle fluctuations and monetary policy.
129
of a decrease in this index for ease of comparison. The results are consistent with those obtained from the baseline analysis: a worsening of financial market conditions increases the probability of entering a recession and decreases the probability of exiting a recession. In our baseline model with random effects, we implicitly assume that industry-specific factors do not affect the probabilities of phase shifts. This assumption might be too restrictive. To address this problem, we allow industryspecific disturbances, denoted as it in equation (7), to follow a first-order autoregressive (AR(1)) process.33 As is clear in figure 6, the results are fairly robust to this generalization. We also consider a fixed-effects specification to take into account the possibility that the mean durations of expansion phases of industries, which are closely related to their trend growth rates (Harding & Pagan, 2002), may not be independent from the explanatory variables, which account for cyclical movements of industries.34 To correct the bias due to the incidental parameters problem, we estimate the fixed-effects model using the penalized-likelihood-based approximation proposed by Bester and Hansen (2009). Again, the results are very similar to those of the random-effects model. To check the robustness of our results to structural changes in the input-output linkages, we use the 1977 input-output table that is broken down by SIC codes. In order to match this input-output table with IP data, we also make use of the vintage IP data disaggregated by the same SIC codes.35 Despite large time gaps between the two input-output tables and despite the shorter sample period (1973:III–2002:III), the qualitative results remain unchanged.36 Finally, figure 7 compares the results for the full sample in figure 5 with those for the sample 1984:III–2010:III.37 As reported in table 9, while the standard deviations of most of our explanatory variables have remained approximately unchanged, that of the monetary policy shock has fallen from 0.94%p in the full sample to 0.56%p in the post-1984 sample. Thus, for comparison, figure 7 sets the size of each shock equal to that used in figure 5, the 1 standard deviation computed over the full sample. In general, our estimates indicate a remarkable stability in the reactions of turning points 33 We estimate the model using the Geweke-Hajivassiliou-Keane (GHK) simulator; see Lee (1997) for details of the procedure. 34 We note that there is a long-standing debate in economics regarding the relationship between long-term trends and short-term fluctuations in economic variables. 35 The vintage IP data are constructed by Foerster et al. (2011) and available on Mark Watson’s web page. We use the vintage IP data disaggregated into 84 industries. 36 However, there are some notable quantitative differences. In particular, the maximum downstream spillover effect on troughs decreases from 5.5%p in the baseline analysis to 1.7%p in this sensitivity analysis. In addition, for both peaks and troughs, the effects of the government spending shock increase substantially relative to the baseline effects. 37 Unfortunately, we could not make a comparison between the results for the post-1984 sample and those for the pre-1984 sample. The reason is that after dropping the observations censored by the turning point detection algorithm, the number of effective observations for the pre-1984 sample is too small to draw definite conclusions.
130
THE REVIEW OF ECONOMICS AND STATISTICS Figure 6.—Estimated Impacts on the Probabilities of Industry Phase Shifts: Robustness Analysis A. Upstream spillover Trough Change in probability
Change in probability
Peak 0.10 0.05 0.00 −0.05 1
2
3
4 5 Quarters
6
7
0.10 0.05 0.00 −0.05
8
1
2
3
4 5 Quarters
6
7
8
6
7
8
6
7
8
6
7
8
6
7
8
6
7
8
B. Downstream spillover Trough Change in probability
Change in probability
Peak 0.06 0.04 0.02 0.00 −0.02 1
2
3
4 5 Quarters
6
7
0.06 0.04 0.02 0.00 −0.02
8
1
2
3
4 5 Quarters
C. Monetary policy shock Trough Change in probability
Change in probability
Peak 0.05 0.00 −0.05 −0.10 1
2
3
4 5 Quarters
6
7
0.05 0.00 −0.05 −0.10
8
1
2
3
4 5 Quarters
D. Government spending shock Trough Change in probability
Change in probability
Peak 0.10 0.05 0.00 −0.05 1
2
3
4 5 Quarters
6
7
0.10 0.05 0.00 −0.05
8
1
2
3
4 5 Quarters
E. Oil price shock Trough Change in probability
Change in probability
Peak 0.10 0.05 0.00 −0.05 −0.10 1
2
3
4 5 Quarters
6
7
0.10 0.05 0.00 −0.05 −0.10
8
1
2
3
4 5 Quarters
F. Financial shock Trough Change in probability
Change in probability
Peak 0.05 0.00 −0.05 −0.10 −0.15 1
2
3
4 5 Quarters
KL11
6
7
0.05 0.00 −0.05 −0.10 −0.15
8
GS−VAR
1
HHMSW10
2
RE−AR
3
4 5 Quarters FE
1977IO
KL11: Kilian and Lewis (2011) monetary policy and oil price shocks, 1973:III–2008:II; GS-VAR: shocks to government spending identified in a standard VAR, 1973:III–2008:IV; HHMSW10: Hatzius et al. (2010) financial conditions index, the effects of a decrease in the index, 1973:III–2009:IV; RE–AR: random-effects model with first-order autoregressive errors, 1973:III–2010:III; FE: fixed-effects model, 1973:III–2010:III; 1977IO: the 1977 input-output table and the corresponding data on IP for 84 industries, 1973:III–2002:III.
to our determinants of industry comovement. Even in the cases where there are some departures from the full sample results, the evidence of structural change is only modest.38 In recent years, there have been debates about whether the effects of macroeconomic shocks on the growth rate of output 38 Indeed, in most cases, the maximum estimated effects are slightly larger in the post-1984 sample than in the full sample. A notable exception is the slightly reduced maximum effects of the government spending shock.
have declined since the mid-1980s.39 Our results suggest that regarding the phase shifts occurring at the disaggregated 39 While Boivin and Giannoni (2006) find that monetary policy shocks had a smaller effect on real activity after the early 1980s, Primiceri (2005) and Canova and Gambetti (2009) report little change or an increase in the effect of monetary policy shocks. Changes in output effects of oil price shocks are an ongoing topic of debate as well; see Blanchard and Galí (2010) and Kilian and Lewis (2011) for evidence of a reduced effect, and see, for counterevidence, Hamilton (2009) and Ramey and Vine (2011). For government spending, previous studies generally
ASYMMETRIC PHASE SHIFTS IN U.S. INDUSTRIAL PRODUCTION CYCLES
131
Figure 7.—Estimated Impacts on the Probabilities of Industry Phase Shifts: Post-1984 Sample A. Upstream spillover Trough Change in probability
Change in probability
Peak 0.15 0.10 0.05 0.00 −0.05 −0.10 −0.15 1
2
3
4 5 Quarters
6
7
0.15 0.10 0.05 0.00 −0.05 −0.10 −0.15
8
1
2
3
4 5 Quarters
6
7
8
6
7
8
6
7
8
6
7
8
6
7
8
6
7
8
B. Downstream spillover Trough Change in probability
Change in probability
Peak 0.10 0.05 0.00 1
2
3
4 5 Quarters
6
7
0.10 0.05 0.00
8
1
2
3
4 5 Quarters
C. Monetary policy shock Trough Change in probability
Change in probability
Peak 0.15 0.10 0.05 0.00 −0.05 −0.10 −0.15 1
2
3
4 5 Quarters
6
7
0.15 0.10 0.05 0.00 −0.05 −0.10 −0.15
8
1
2
3
4 5 Quarters
D. Government spending shock Trough Change in probability
Change in probability
Peak 0.05 0.00 −0.05 1
2
3
4 5 Quarters
6
7
0.05 0.00 −0.05
8
1
2
3
4 5 Quarters
E. Oil price shock Trough Change in probability
Change in probability
Peak 0.10 0.05 0.00 −0.05 −0.10 −0.15 1
2
3
4 5 Quarters
6
7
0.10 0.05 0.00 −0.05 −0.10 −0.15
8
1
2
3
4 5 Quarters
F. Financial shock Trough Change in probability
Change in probability
Peak 0.10 0.05 0.00 −0.05 −0.10 1
2
3
4 5 Quarters
6
7
8
Full sample
0.10 0.05 0.00 −0.05 −0.10 1
2
3
4 5 Quarters
Post−1984
Sample period: full sample: 1973:III–2010:III; post-1984: 1984:I–2010:III. The gray shaded areas and thin solid lines represent 1 standard error bands for the full sample and the post-1984 sample periods, respectively.
industry level, the four macroeconomic shocks as well as the interindustry spillover effects have remained important in the post-1984 period. found declines in the effects on real GDP of government spending shocks in the post-1980 period (Perotti, 2005; Bilbiie, Meier, & Müller, 2008). Foerster et al. (2011) find that common factors explain a smaller fraction of the variability in sectoral growth rates after 1984. However, they do not disentangle the estimated factors into specific macroeconomic shocks. See also Stock and Watson (2012) for the roles played by differenct types of macroeconomic shocks in the 2007–2009 recession.
VI.
Summary
The phase shift carries far richer information about the nature of business cycles than a simple correlation. Based on the IP indices for 74 U.S. manufacturing industries, we identify turning points in industry cycles using a nonparametric method developed by Harding and Pagan (2002). We uncover the following new empirical regularities. First, cyclical phases are highly concentrated across industries in sync with the aggregate business cycle. Second, there is
132
THE REVIEW OF ECONOMICS AND STATISTICS
substantial asymmetry in the distribution of turning points between peaks and troughs. Most important, troughs are much more concentrated and sharper than peaks. Third, occurrences of phase shifts across industries strongly support the spillovers through input-output linkages, a core aspect of multisector models. Fourth, we confirm that the standard macroeconomic shocks, such as exogenous changes in the federal funds rate, defense spending, oil prices, and financial conditions, are also important determinants of cyclical turning points. Their effects on industry phase shifts are all statistically significant and conform with our economic priors. REFERENCES Acemoglu, Daron, Vasco M. Carvalho, Asuman Ozdaglar, and Alireza Tahbaz-Salehi, “The Network Origins of Aggregate Fluctuations,” Econometrica 80 (2012), 1977–2016. Altug, Sumru, and Fabio Canova, “Do Institutions and Culture Matter for Business Cycles?” CEPR discussion paper 9382 (2013). Artis, Michael J., Massimiliano Marcellino, and Tommaso Proietti, “Dating Business Cycles: A Methodological Contribution with an Application to the Euro Area,” Oxford Bulletin of Economics and Statistics 66 (2004), 537–565. Bartelsman, Eric J., Ricardo J. Caballero, and Richard K. Lyons, “Customerand Supplier-Driven Externalities,” American Economic Review 84 (1994), 1075–1084. Bernanke, Ben S., Mark Gertler, and Simon Gilchrist, “The Financial Accelerator in a Quantitative Business Cycle Framework” (pp. 1341– 1393), in John B. Taylor and Michael Woodford, eds., Handbook of Macroeconomics (Amsterdam: Elsevier, 1999). Bester, C. Alan, and Christian Hansen, “A Penalty Function Approach to Bias Reduction in Nonlinear Panel Models with Fixed Effects,” Journal of Business and Economic Statistics 27 (2009), 131–148. Bilbiie, Florin O., André Meier, and Gernot J. Müller, “What Accounts for the Changes in U.S. Fiscal Policy Transmission?” Journal of Money, Credit, and Banking 40 (2008), 1439–1469. Blanchard, Olivier J., and Jordi Galí, “The Macroeconomic Effects of Oil Price Shocks: Why Are the 2000s So Different from the 1970s?” (pp. 373–421), in Jordi Galí and Mark Gertler, eds., International Dimensions of Monetary Policy (Chicago: University of Chicago Press, 2010). Boivin, Jean, and Marc P. Giannoni, “Has Monetary Policy Become More Effective?” this review 88 (2006), 445–462. Bordo, Michael D., and Joseph G. Haubrich, “Credit Crises, Money and Contractions: An Historical View,” Journal of Monetary Economics 57 (2010), 1–18. Bry, Gerhard, and Charlotte Boschan, Cyclical Analysis of Time Series: Selected Procedures and Computer Programs (New York: National Bureau of Economic Research, 1971). Burns, Arthur F., and Wesley C. Mitchell, Measuring Business Cycles (New York: National Bureau of Economic Research, 1946). Canova, Fabio, “Detrending and Business Cycle Facts,” Journal of Monetary Economics 41 (1998), 475–512. Canova, Fabio, and Luca Gambetti, “Structural Changes in the US Economy: Is There a Role for Monetary Policy?” Journal of Economic Dynamics and Control 33 (2009), 477–490. Chauvet, Marcelle, and Jeremy Piger, “A Comparison of the Real-Time Performance of Business Cycle Dating Methods,” Journal of Business and Economic Statistics 26 (2008), 42–49. Christiano, Lawrence J., Martin Eichenbaum, and Charles L. Evans, “Monetary Policy Shocks: What Have We Learned and to What End?” (pp. 65–148), in John B. Taylor and Michael Woodford, eds., Handbook of Macroeconomics (Amsterdam: Elsevier, 1999). Christiano, Lawrence, Roberto Motto, and Massimo Rostagno, “Risk Shocks,” NBER working paper 18682 (2013). Cogley, Timothy, and James M. Nason, “Effects of the Hodrick-Prescott Filter on Trend and Difference Stationary Time Series: Implications for Business Cycle Research,” Journal of Economic Dynamics and Control 19 (1995), 253–278.
Conley, Timothy G., and Bill Dupor, “A Spatial Analysis of Sectoral Complementarity,” Journal of Political Economy 111 (2003), 311–352. Dupor, Bill, “Aggregation and Irrelevance in Multi-Sector Models,” Journal of Monetary Economics 43 (1999), 391–409. Foerster, Andrew T., Pierre-Daniel G. Sarte, and Mark W. Watson, “Sectoral vs. Aggregate Shocks: A Structural Factor Analysis of Industrial Production,” Journal of Political Economy 119 (2011), 1–38. Gilchrist, Simon, and Egon Zakrajšek, “Credit Spreads and Business Cycle Fluctuations,” American Economic Review 102 (2012), 1692– 1720. Hamilton, James D., “What Is an Oil Shock?” Journal of Econometrics 113 (2003), 363–398. ——— “Causes and Consequences of the Oil Shock of 2007–08,” Brookings Papers on Economic Activity 40 (2009), 215–259. Hamilton, James D., and Michael T. Owyang, “The Propagation of Regional Recessions,” this review 94 (2012), 935–947. Harding, Don, and Adrian Pagan, “Dissecting the Cycle: A Methodological Investigation,” Journal of Monetary Economics 49 (2002), 365–381. ——— “Synchronization of Cycles,” Journal of Econometrics 132 (2006), 59–79. Harvey, Andrew C., and A. Jaeger, “Detrending, Stylized Facts and the Business Cycle,” Journal of Applied Econometrics 8 (1993), 231– 247. Hatzius, Jan, Peter Hooper, Frederic S. Mishkin, Kermit L. Schoenholtz, and Mark W. Watson, “Financial Conditions Indexes: A Fresh Look after the Financial Crisis,” NBER working paper 16150 (2010). Higgins, James J., Introduction to Nonparametric Statistics (Pacific Grove, CA: Brooks/Cole, 2004). Holly, Sean, and Ivan Petrella, “Factor Demand Linkages, Technology Shocks, and the Business Cycle,” this review 94 (2012), 948–963. Horvath, Michael, “Sectoral Shocks and Aggregate Fluctuations,” Journal of Monetary Economics 45 (2000), 69–106. Keynes, John M., The General Theory of Employment, Interest and Money (London: Macmillan, 1936). Kilian, Lutz, and Logan T. Lewis, “Does the Fed Respond to Oil Price Shocks?” Economic Journal 121 (2011), 1047–1072. Kim, Chang-Jin, and Charles R. Nelson, “Has the U.S. Economy Become More Stable? A Bayesian Approach Based on a Markov-Switching Model of the Business Cycle,” this review 81 (1999), 608–616. King, Robert, and Charles I. Plosser, “Real Business Cycles and the Test of the Adelmans,” Journal of Monetary Economics 33 (1994), 405–438. Kiyotaki, Nobuhiro, and John Moore, “Credit Cycles,” Journal of Political Economy 105 (1997), 211–248. Lam, Pok-Sang, “A Markov-Switching Model of GNP Growth with Duration Dependence,” International Economic Review 45 (2004), 175–204. Lee, Lung-Fei, “Simulated Maximum Likelihood Estimation of Dynamic Discrete Choice Statistical Models: Some Monte Carlo Results,” Journal of Econometrics 82 (1997), 1–35. Long, John B., Jr., and Charles I. Plosser, “Real Business Cycles,” Journal of Political Economy 91 (1983), 39–69. ——— “Sectoral vs. Aggregate Shocks in the Business Cycle,” American Economic Review 77 (1987), 333–336. Lucas, Robert E., Jr., “Understanding Business Cycles,” CarnegieRochester Conference Series on Public Policy 5 (1977), 7–29. McConnell, Margaret M., and Gabriel Perez-Quiros, “Output Fluctuations in the United States: What Has Changed since the Early 1980’s?” American Economic Review 90 (2000), 1464–1476. McQueen, Grant, and Steven Thorley, “Asymmetric Business Cycle Turning Points,” Journal of Monetary Economics 31 (1993), 341–362. Morley, James, “Macroeconomics, Nonlinear Time Series” (pp. 5325– 5348), in Robert A. Meyers, ed., Encyclopedia of Complexity and Systems Science (Berlin: Springer, 2009). Owyang, Michael T., Jeremy Piger, and Howard J. Wall, “Business Cycle Phases in U.S. States,” this review 87 (2005), 604–616. Perotti, Roberto, “Estimating the Effects of Fiscal Policy in OECD Countries,” CEPR discussion paper 4842 (2005). Primiceri, Giorgio E., “Time Varying Structural Vector Autoregressions and Monetary Policy,” Review of Economic Studies 72 (2005), 821–852. Ramey, Valerie A., “Identifying Government Spending Shocks: It’s All in the Timing,” Quarterly Journal of Economics 126 (2011), 1–50. Ramey, Valerie A., and Daniel J. Vine, “Oil, Automobiles, and the U.S. Economy: How Much Have Things Really Changed?”
ASYMMETRIC PHASE SHIFTS IN U.S. INDUSTRIAL PRODUCTION CYCLES (pp. 333–367), in Daron Acemoglu and Michael Woodford, eds., NBER Macroeconomics Annual 2010 (Chicago: University of Chicago Press, 2011). Shea, John, “Complementarities and Comovements,” Journal of Money, Credit, and Banking 34 (2002), 412–433. Sichel, Daniel E., “Inventories and the Three Phases of the Business Cycle,” Journal of Business and Economic Statistics 12 (1994), 269–277. Stock, James H., and Mark W. Watson, “Has the Business Cycle Changed and Why?” (pp. 159–230), in Mark Gertler and Kenneth Rogoff, eds., NBER Macroeconomics Annual 2002 (Cambridge, MA: MIT Press, 2003).
133
——— “Indicators for Dating Business Cycles: Cross-History Selection and Comparisons,” American Economic Review 100 (2010), 16–19. ——— “Disentangling the Channels of the 2007–2009 Recession,” Brookings Papers on Economic Activity 44 (2012), 81–156. ——— “Estimating Turning Points Using Large Data Sets,” Journal of Econometrics 178 (2014), 368–381. Train, Kenneth E., Discrete Choice Methods with Simulation, 2nd ed. (Cambridge: Cambridge University Press, 2009). Watson, Mark W., “Business Cycle Durations and Postwar Stabilization of the U.S. Economy,” American Economic Review 84 (1994), 24–46. Zarnowitz, Victor, Business Cycles: Theory, History, Indicators, and Forecasting (Chicago: University of Chicago Press, 1992).
