What Do Technology Shocks Do?

Download What Do Technology Shocks Do?

Preview text

This PDF is a selection from an out-of-print volume from the National Bureau of Economic Research
Volume Title: NBER Macroeconomics Annual 1998, volume 13 Volume Author/Editor: Ben S. Bernanke and Julio Rotemberg, editors Volume Publisher: MIT Press Volume ISBN: 0-262-52271-3 Volume URL: http://www.nber.org/books/bern99-1 Publication Date: January 1999
Chapter Title: What Do Technology Shocks Do? Chapter Author: John Shea Chapter URL: http://www.nber.org/chapters/c11249 Chapter pages in book: (p. 275 - 322)

John Shea
What Do Technology Shocks Do?
1. Introduction
The real business cycle (RBC) approach to short-run fluctuations, pioneered by Kydland and Prescott (KP) (1982) and Long and Plosser (LP) (1983), has dominated the academic business-cycle literature over the last decade and a half. KP and LP were seminal in several respects. First, they reintroduced the Schumpeterian idea that stochastic technological progress could generate business cycles. Second, they argued that one could explain fluctuations using a frictionless neoclassical framework in which business cycles are optimal and therefore require no smoothing by policymakers. Third, they argued that business cycles could and should be explained using dynamic stochastic general equilibrium models in which preferences and production are explicitly spelled out in a way consistent with microeconomic first principles, such as optimizing behavior.
The RBC literature has broadened considerably since KP and LP. Recent research has introduced frictions such as imperfect competition (e.g., Rotemberg and Woodford, 1995), increasing returns to scale (e.g., Farmer and Guo, 1994), and price stickiness (e.g., Kimball, 1995), as well as alternative sources of shocks, such as government spending (e.g., Christiano and Eichenbaum, 1992), monetary policy (e.g., Christiano and Eichenbaum, 1995), and animal spirits (e.g., Schmitt-Grohe, 1997). The idea that business cycles should be analyzed using explicit dynamic stochastic general equilibrium models seems destined to be the main lasting contribution of KP and LP's work.
Meanwhile, the profession has largely ignored the empirical question of what role technology shocks actually play in business cycles. I believe that this is unfortunate, for four reasons. First, the idea that new prod-
I thank Kortum and Susanto Basu for providing their data, and the editors and participants, as well as seminar participants at Brown, for helpful comments.

276 • SHEA
ucts and processes are introduced at a time-varying rate is inherently plausible, at least at the disaggregated industry level. Second, much recent research exploring the effects of frictions on business-cycle propagation still assumes that cycles are driven by technology shocks (e.g., Cogley and Nason, 1995; Horvath, 1997; Carlstrom and Fuerst, 1997). It would be useful to know if this modeling strategy has any empirical foundation. Third, while few would argue any more that technology shocks are the only source of business cycles, it would still be useful to know if technology shocks can explain some part of fluctuations, particularly given that monetary, oil price, and other observable shocks seem unable to account for a large fraction of observed cyclical variation in output (Cochrane, 1994). Finally, even if technology shocks are not responsible for a large share of volatility, the response of the economy to technology could help distinguish between competing views of the economy's propagation mechanisms. In the baseline one-sector flexible-price RBC model, technology shocks shift out the production possibilities frontier, inducing short-run increases in investment, labor, and materials. In multisector models, industry technology shocks reduce input prices to downstream sectors, inducing increases in downstream input and output. Meanwhile, Gali (1996) and Basu, Fernald, and Kimball (1997) demonstrate that favorable technology shocks may reduce input use in the short run if prices are sticky; intuitively, if prices do not fall, output will be unchanged and inputs must fall to accommodate improved total factor productivity (TFP). Thus, one can potentially distinguish between sticky and flexible price models by examining whether technology shocks increase or decrease input use.
To date, the empirical case for technology has largely been made indirectly, by showing that plausibly calibrated models driven by technology shocks can produce realistic patterns of volatility and comovement. Of course, these quantitative exercises, while informative, do not tell us what technology shocks actually do. Two pieces of more direct evidence are that measured TFP is procyclical and that aggregate output potentially has a unit root, suggesting that at least some output shocks are permanent. However, it is now well known that neither of these facts proves that technology is important to business cycles. Observable nontechnology shocks cause procyclical movements in TFP, consistent with imperfect competition, increasing returns to scale, procyclical factor utilization, or procyclical reallocation of factors to high productivity sectors (e.g., Hall, 1988; Evans, 1992; Burnside, Eichenbaum, and Rebelo, 1995; Basu and Fernald, 1997). Meanwhile, demand shocks can have permanent effects on output in endogenous growth models (e.g., Stadler, 1990); and in any case, a unit root is consistent with transitory

What Do Technology Shocks Do? • 277
shocks driving an arbitrarily large fraction of short-run variation (Quah, 1989).
This paper takes a more direct approach to assessing what technology shocks do, an approach inspired by the large literature estimating the impact of monetary policy shocks on the economy (e.g., Christiano, Eichenbaum, and Evans, 1998). Using annual panel data for 19 U.S. manufacturing industries from 1959 to 1991, I employ vector autoregressions (VARs) to document the dynamic impact of shocks to two observable indicators of technological change: research and development (R&D) spending, and patent applications. R&D measures the amount of input devoted to innovative activity, while patent applications measure inventive output. Previous studies (e.g., Griliches and Lichtenberg, 1984; Lichtenberg and Siegel, 1991; Scherer, 1993), as well as results reported below, suggest that variation in R&D and patenting is related to long-run variation in productivity growth across firms and industries. Moreover, industry-level R&D and patents display nontrivial short-run fluctuations, as can be seen in Figure 1, which plots log real R&D and log patent applications by industry of manufacture and use for the U.S. aerospace industry. If technological progress is truly stochastic, then fluctuations in R&D should in part reflect variation in the perceived marginal product of knowledge, while fluctuations in patents should in part reflect shocks to the success of research activity. I use these fluctuations to estimate how a typical industry's inputs and TFP respond over time to technology shocks, and to quantify the share of industry volatility due to technology shocks. I estimate the impact of both own technology shocks and technology shocks in upstream input-supplying industries.
To be sure, fluctuations in R&D and patent applications may not be due to technology shocks alone. Griliches (1989), for instance, argues that patenting fluctuations in the U.S. are in part responses to factors such as changes in patent law and changes in the efficiency and resources of the U.S. Patent Office. Both R&D and patent applications, meanwhile, are a type of investment, and as such they may respond endogenously to output shocks, either because of financial-market constraints or because current shocks are positively correlated with the future marginal product of capital. My preferred VAR specifications address these concerns by including time dummies in the regressions and by placing the technological indicators last in the Choleski ordering used to decompose the VAR innovations into orthogonal components. The time dummies remove fluctuations in R&D and patent applications due to aggregate factors unrelated to true technological progress, such as changes in the number of patent examiners, provided that these factors affect all industries equally. My impulse responses therefore measure the

278 • SHEA



LogR&D Spending





LogPatentApplications, Industry ofManufacture






LqgPatentApplications, Industry ofUse







What Do Technology Shocks Do? • 279
impact of industry-specific technology shocks on industry-specific fluctuations in inputs and TFP, while my variance decompositions estimate the contribution of technology shocks to idiosyncratic industry fluctuations. Placing technology last in the ordering, meanwhile, defines technology shocks as the component of R&D or patenting orthogonal to both lagged technology and lagged and contemporaneous inputs and TFP. Empirically, innovations to industry output are positively correlated with innovations to both R&D and patent applications; placing technology last assumes that this contemporaneous comovement reflects an accelerator mechanism running from industry activity to technology, rather than an instantaneous impact of technology shocks on output. This assumption seems inherently plausible given the likely lags between R&D spending, invention, and diffusion of a new technology (Gort and Klepper, 1982).
My main empirical findings are as follows. First, favorable technology shocks—increases in the orthogonal components of R&D and patents— tend to increase input use, especially labor, in the short run, but to reduce inputs in the long run. Second, technology improvements tend to encourage substitution towards capital relative to materials and labor, as well as substitution towards nonproduction labor relative to production labor. These results are consistent with recent cross-sectional studies establishing a complementary long-run relationship between technological change and equipment (Delong and Summers, 1991) and skilled labor (Berman, Bound, and Griliches, 1994). Third, favorable technology shocks do not significantly increase measured TFP at any horizon, and indeed in some cases reduce TFP. This suggests that procyclical movements in TFP have little to do with the introduction of new products and processes. Fourth, technology shocks explain only a small share of idiosyncratic industry volatility of inputs and TFP at business-cycle horizons. This result is bad news for technology-shock-driven models, particularly given that industry-specific technology shocks are likely to explain industry-specific volatility better than aggregate volatility (Horvath, 1997). However, my results could be consistent with models in which technology contributes to low-frequency fluctuations (e.g., Jovanovic and Lach, 1997); or with models in which the important "real" shocks come from strikes, weather, cartel behavior, and so on; or with models in which "technology shocks" are not due to stochastic scientific and engineering developments, but to stochastic movements in management techniques or industrial organization that cause a given set of inputs to be more or less efficient. Finally, I find that technology improvements are more likely to raise TFP and reduce prices in industries characterized by process innovations than in industries dominated by product innovations. This suggests that my fail-

280 • SHEA
ure to find strong effects of technology on TFP may be due in part to the failure of available price data to capture productivity gains caused by quality improvements and new product introductions.
Two other recent papers (Gali, 1996; Basu, Fernald, and Kimball, 1997) also investigate the short-run impact of technology shocks, in both cases using aggregate data. Gali estimates a structural vector autoregression for labor productivity and labor input in the United States, identifying technology shocks by assuming that only technology affects long-run productivity. Basu, Fernald, and Kimball correct industry-level TFP for variations due to increasing returns to scale, imperfect competition, and cyclical factor utilization, and then measure aggregate technology as an appropriately weighted average of sectoral technology. Interestingly, Gali (1996) and Basu, Fernald, and Kimball (1997) both find that favorable technology shocks reduce input use in the short run, consistent with sticky prices but contrary to my results.
These two papers represent a distinct advance over existing literature. Nevertheless, one might disagree with their methodologies for measuring technological change. Gall's approach rests heavily on the assumption that demand shocks cannot affect productivity in the long run. This assumption is inconsistent both with endogenous growth models and with models in which recessions cleanse the economy by wiping out low-productivity firms (e.g., Caballero and Hammour, 1994, 1996). Cleansing models, in particular, predict that favorable demand shocks will reduce long-run productivity, and Gali himself has in the past argued for such an interpretation of the data (Gali and Hammour, 1992). Interestingly, my impulse response functions suggest that input innovations lead to short-run increases in TFP, consistent with increasing returns or procyclical utilization, but long-run decreases in TFP, consistent with cleansing models.
Basu, Fernald, and Kimball's approach does not rely on long-run restrictions. It does, however, rely on the idea that TFP fluctuations are valid measures of stochastic technological progress at the two-digit industry level, once one corrects for increasing returns, imperfect competition, and cyclical factor utilization. This idea seems plausible, but it is not necessarily true, given that fluctuations in "corrected" sectoral TFP could still be due to nontechnology sources such as measurement error, within-sector factor reallocations, or inadequate corrections for increasing returns or cyclical utilization. Basu, Fernald, and Kimball's methodology would be more convincing if their corrected measure of technology could be linked to some sort of outside measure of technological progress, such as anecdotal evidence on the timing of particular technical changes in particular industries.

What Do Technology Shocks Do? • 281
The remainder of the paper proceeds as follows. Section 2 describes the data. Section 3 examines long-run and contemporaneous relationships between technological progress and my measures of innovative activity, largely to connect my work to previous studies. Section 4 presents evidence from VARs, and Section 5 concludes.
2. Data Description
My goal is to examine the time-series interactions between measures of technological change, such as patents and R&D, and measures of economic activity. Ideally, I would estimate these interactions using aggregate data for a single country, following the empirical literature on monetary policy. However, this approach is not feasible in my case. The only readily available data for patents and R&D are annual rather than quarterly or monthly, implying short aggregate time series. Even if higherfrequency data could be constructed, it is not clear that they would be useful, since the impact of technological change on the economy is likely to operate at a somewhat lower frequency than the impact of monetary shocks. To obtain sufficient degrees of freedom to estimate the impact of technology shocks with reasonable precision, I use panel data for 19 manufacturing industries covering 1959-1991, exploiting the fact that technological developments are not perfectly synchronized across industries. An alternative, worth pursuing in future work, would be to use annual aggregate data for a panel of countries, or for panels of both countries and industries.
Data on R&D by industry are taken from the National Science Foundation's annual survey of U.S. firms. I examine only company-financed R&D. Previous research using cross sections of industries and firms (e.g., Terleckyj, 1975; Lichtenberg and Siegel, 1991) has shown that longrun productivity growth is related to company-financed R&D, but not to federally financed R&D, suggesting that public R&D dollars are spent inefficiently or that they are spent in areas, such as defense or space exploration, where productivity measurement is difficult. I convert nominal R&D to 1991 dollars using the GDP deflator, then convert real R&D flows to an R&D capital stock, following Griliches (1973) and most other subsequent research. I employ a linear capital accumulation equation, assuming a 15% annual depreciation rate and setting the 1959 stock equal to the 1959 flow divided by 0.15 plus the industry's average R&D growth over the sample period; these assumptions are standard in recent literature (e.g., Lach, 1995; Keller, 1997). The empirical results are similar if I use real R&D flows instead of R&D stocks. As a timing convention, I include R&D spending in year t in the R&D stock for year

282 • SHEA






R&D Manuf. Patent Use Patent

Growth Patents Growth Patents Growth

Food (SIC 20)

879.1 4.62 311.2 0.63 1085 0.06

Textiles (SIC 22-23)

185.5 3.57 620.9 1.17 994 -0.81

Lumber (SIC 24-25)

152.0 4.81 605.9 0.08 597 -0.05

Paper (SIC26)

611.7 5.53 482.4 0.04 490 0.08

Industrial chemicals (SIC 281-282, 3211.7 2.80 3758.8 0.75 2518 0.47


Drugs (SIC 283)

2629.4 7.75 825.5 5.90 1100 2.73

Other chemicals (other SIC28)

1053.0 5.27 2517.4 0.54 1261 0.70

Petroleum (SIC29)

1812.4 2.82 1745.5 -1.30 1659 0.22

Rubber (SIC30)

693.2 4.01 1586.0 1.03 1348 1.45

Stone (SIC32)

574.5 3.05 506.8 1.60 557 0.76

Metals (SIC33)

835.9 1.51 373.1 0.30 795 0.06

Metal prods. (SIC34)

684.1 2.06 3737.6 0.18 1979 0.24

Computers (SIC357)

5172.6 6.66 1114.3 2.70 1333 3.09

Other nonelec. equip, (other SIC 2102.3 5.08 10966.1 -0.15 4084 -0.33


Electronics & commun. equip. (SIC 5018.4 6.65 5629.4 1.51 4456 1.76


Other electric equip. (Other SIC36) 2043.7 0.99 4154.1 0.41 2779 0.43

Aerospace (SIC 372, 376)

4022.4 4.81 276.9 -1.22 392 -0.77

Autos & other transp. equip. (SIC 5701.8 4.37 1972.1 -0.28 2787 -0.09


Instruments (SIC38)

3100.3 7.42 3626.7 2.33 1268 1.59

t, so that I can interpret the correlations between R&D and other variables as reflecting a contemporaneous response of R&D to industry activity. I use data for 19 manufacturing industries; these are listed in Table 1 along with sample means of real R&D flows in millions of dollars and the growth rate of the R&D stock. The largest flows of company R&D are found in automobiles, electronics, and computers; the fastest-growing R&D stocks are in drugs, electronics, computers, and instruments. Note that my baseline sample omits nonmanufacturing industries as well as some manufacturing industries (tobacco, printing and publishing, leather, and miscellaneous manufacturing) whose R&D data are lumped together by the NSF. The share of overall R&D accounted for by these sectors is trivial for most of my sample period.
I must mention two problems with these data. First, to avoid disclosure of individual firms' operations, the NSF suppresses some industryyear observations. In virtually all such cases, the NSF suppresses either company-financed or total (including federally financed) R&D, but not

What Do Technology Shocks Do? • 283
both, so that I can interpolate gaps in company R&D using growth of total R&D. Second, the NSF data are collected at the company level. All R&D spending performed by a company is assigned to the industry in which the company had the most sales, even if part of the R&D was conducted in establishments belonging to another industry. Given that R&D is typically performed in large conglomerated firms, the assignment of R&D to particular industries is presumably subject to error. Particularly troubling is the fact that a given firm's industry classification can change over time as its pattern of sales changes, creating the possibility of large movements in measured industry-level R&D spending unrelated to actual changes in spending at the establishment level. Griliches and Lichtenberg (1984) attempt to overcome this problem by using R&D data grouped by applied product field rather than by industry of origin. Unfortunately, the reporting requirements of the NSF's product field survey were burdensome on participating firms, leading to spotty coverage. The survey was reduced from annual to biannual beginning in 1978, and was discontinued in 1986.
Patent data for U.S. industries are not routinely available. The reason is that the U.S. Patent Office assigns new patents to technological fields, but not to industries. Estimating patents by industry for the U.S. thus requires a mapping from technological fields into industries. The most satisfactory mapping available is the Yale Technology Concordance (YTC), described by Kortum and Putnam (1997). This concordance uses the fact that the Canadian patent office assigns patents to technological fields, to industries of manufacture, and to industries of use; for instance, a new farm tractor invented in an aerospace establishment would be assigned to the agricultural machinery sector (industry of manufacture) and to agriculture itself (industry of use). The YTC estimates mappings between technological field and industries of manufacture and use using the Canadian data, then applies the Canadian mapping to U.S. patents by technological field. For this study, I use annual data on U.S. patent applications grouped both by industry of manufacture and by industry of use, generously provided by Sam Kortum. I convert the annual flows of patents to stocks using the same method as for R&D; the empirical results again are similar if I use flows instead of stocks. Note that patents grouped by date of application are superior to patents grouped by date of grant, both because application presumably coincides with the economic viability of an innovation, and because historically there have been long and variable lags between application and granting in the United States, caused in part by changes in the resources of the U.S. Patent Office (Griliches, 1989).
I must again acknowledge potential problems with these data. First,

Preparing to load PDF file. please wait...

0 of 0
What Do Technology Shocks Do?