Full text of Working Papers (Federal Reserve Bank of Richmond) : On the Employment Effect of Technology : Evidence from U.S. Manufacturing for 1958-1996, Working Paper 03-06

View original document

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

Working Paper Series

On the Employment Effect of
Technology: Evidence from U.S.
Manufacturing for 1958-1996

WP 03-06

Yongsung Chang
Federal Reserve Bank of Richmond
Jay H. Hong
University of Pennsylvania

This paper can be downloaded without charge from:
http://www.richmondfed.org/publications/

On the Employment Effect of Technology:
Evidence from U.S. Manufacturing for 1958-1996 *
Federal Reserve Bank of Richmond Working Paper 03-06
July 2003

Yongsung Chang
Research Department, Federal Reserve Bank of Richmond
and
Jay H. Hong
Department of Economics, University of Pennsylvania

Abstract:
Recently, Galí and others have found that technological progress may be contractionary:
a favorable technology shock reduces hours worked in the short run. We ask whether this
observation is robust in disaggregate data. According to our VAR analysis of 458 fourdigit U.S. manufacturing industries for 1958-1996, some industries do exhibit temporary
reduction in hours in response to a permanent increase in TFP. However, there are far
more industries in which technological progress significantly increases hours. Using
micro data on average price duration, we ask whether the difference across industries is
related to the stickiness of industry-output prices. Among 87 manufacturing goods, we do
not find such a relation.

Keywords: Technology Shocks, Hours Fluctuations, Sticky Prices
JEL Classification: E24, E32

* We thank Mark Bils for providing us monthly price-change frequency data used in Bils and Klenow
(2002). We thank Mark Bils, Larry Christiano, Martin Eichenbaum, Jordi Galí, Valerie Ramey, and John
Shea for helpful comments and suggestions on an earlier version. The views expressed herein are those of
the authors and do not necessarily reflect those of the Federal Reserve Bank of Richmond or the Federal
Reserve System. Chang: Yongsung.Chang@rich.frb.org; Hong: jayhwa@econ.upenn.edu.

Introduction

Despite controversies regarding its quantitative importance as a source of business-cycle ﬂuctua
tions, the employment eﬀect of technology is conventionally viewed as expansionary; technological
progress not only expands the production frontier but also creates jobs.
Recently, however, a number of studies—initiated by Galı́ (1999), Basu, Fernald, and Kimball
(1998), and Kiley (1998) and reinvestigated by Francis and Ramey (2002)—report that favorable
technology shocks may reduce total hours worked in the short run. This is an important ﬁnding
because, if conﬁrmed, the ﬂuctuation induced by technological progress may violate a simple fact of
the business cycle—the co-movement of output and employment, documented at least since Burns
and Mitchell (1946).1
In this paper, we investigate whether this observation is robust at a more disaggregate level.
According to our VAR analysis of 458 four-digit U.S. manufacturing industries for 1958-1996, some
industries exhibit a decrease in hours worked in response to a favorable technology shock, identiﬁed
by a stochastic trend component of total factor productivity (TFP). However, there are far more
industries in which a permanent increase in TFP leads to a signiﬁcant increase in both employment
and hours per worker in the short run. Among 458 four-digit industries, 148 industries exhibit a
statistically signiﬁcant increased hours of work in response to a favorable technology shock, whereas
only 13 industries exhibit signiﬁcant decreases in hours worked in the short run.
1

In Galı́ as well as in Kiley, and Francis and Ramey, a technology shock is identiﬁed by a stochastic trend of labor

productivity from a structural VAR. Basu et al. construct a measure of technology change from production functions,
controlling for imperfect competition, utilization, and aggregation eﬀects. In contrast, Shea (1998), distinctive for
his use of a direct measure of technology, ﬁnds that an increase in the orthogonal components of R&D and patents
tends to increase input use, especially labor, in the short run, but to reduce inputs in the long run.

Our results diﬀer from Kiley’s, which shows that employment decreases in response to a perma
nent increase in labor productivity in most two-digit manufacturing industries for 1968:II-1995:IV.
However, we do not view these ﬁndings necessarily conﬂicting. We ﬁnd that the stochastic trends
of TFP and labor productivity capture diﬀerent types of changes in production because labor pro
ductivity reﬂects changes in input mix as well as improved eﬃciency. For instance, disturbances
aﬀecting material-labor or capital-labor ratios (e.g., persistent movement of relative input price
changes or trends in sectoral labor supply) generate a negative correlation between labor produc
tivity and hours along the downward sloping marginal product of labor whereas such changes alone
do not aﬀect the TFP.
As the contractionary eﬀect of technology is incompatible with a baseline equilibrium model,
alternative models have been proposed. Our analysis sheds light on two alternative hypotheses:
sticky prices (Galı́, Kiley, and Basu et al.) and labor-saving technological progress (Francis and
Ramey). Intuitively speaking, if prices do not fall, demand remains unchanged and ﬁrms need
less input due to the improved technology.2 We test this hypothesis by asking whether the crosssectional diﬀerence in an industry’s hours response (to a technology shock from the VAR) can be
accounted for by the stickiness of industry-output prices (average duration of prices) constructed
by Bils and Klenow (2002). For 87 manufacturing goods, which we are able to match with the
employment and TFP in the NBER Database, we do not ﬁnd a systematic correlation between the
industry’s hours response and the average duration of product price.
2

Dotsey (1999) and Galı́ made it clear that the hours response depends on the nature of monetary policy as well

as on the price stickiness. For example, hours can increase even under the sticky price if monetary authority is very
´ Lop´ez-Salido and Vall´es (2002), the employment eﬀect of
accommodative to technology shocks. According to Gali,
technology varies across monetary policy regimes in the U.S.; the negative correlation between hours and technology
has weakened since the Volcker-Greenspan era.

Francis and Ramey illustrate that under a strong complementarity between capital and labor,
demand for labor input may decrease in the short run in the face of labor-saving technological
progress. On one hand, we ﬁnd that the labor share has indeed decreased signiﬁcantly in man
ufacturing in the past four decades (from .57 in 1958 to .38 in 1996) suggesting a presence of
labor-saving technological change in manufacturing. On the other hand, our data suggests that
under a strong complementarity (e.g., substitution elasticity between capital and labor as low as
.5), the decline of labor shares may be accounted for by the relatively slow accumulation of capital
in manufacturing as the measured capital-eﬀective labor ratios had decreased by almost 50 percent
during the sample period.
The paper is organized as follows. In Section 2, we brieﬂy describe our empirical method,
including the VAR and data, and report the estimates on the technology eﬀect on hours. In Section
3, we compare the stochastic trends of TFP and labor productivity, providing a reconciliation with
previous studies. In Section 4, we relate our results to the two hypotheses that allow for a negative
response of hours to technology. Section 5 provides caveats on our analysis. Section 6 is the
conclusion.

2
2.1

Evidence from Industry TFP and Hours
Identifying Technology Shocks

Following the tradition of Blanchard and Quah (1989) and earlier works (e.g., Galı́, Kiley), tech
nology shocks are identiﬁed by a structural VAR of productivity, xt , and hours, nt . Fluctuations
of industry productivity and hours are driven by two fundamental disturbances—technology and
non-technology shocks—which are orthogonal to each other. Only technology shocks can have a
permanent eﬀect on the level of industry productivity. Both technology and non-technology shocks
3

can have a permanent eﬀect on industry hours. We do not attempt to provide an interpretation
of non-technology shocks, which can be either aggregate (e.g., monetary shocks) or sectoral (e.g.,
reallocation shocks).3
We assume that the vector [Δxt , Δnt ]� can be expressed as a (possibly inﬁnite) distributed lag
of both types of disturbances:
�

Δxt
Δnt

�

�
=

C 11 (L) C 12 (L)
C 21 (L) C 22 (L)

��

εxt
εnt

�
= C(L)εt ,

(1)

where �xt and �nt denote, respectively, the sequence of technology and non-technology shocks. The
orthogonality assumption (combined with a standard normalization) implies E�t ��t = I. Our iden
tifying restriction corresponds to C 12 (1) = 0. The speciﬁcation (1) is based on the assumption that
industry hours and productivity are integrated of order one, which holds in most manufacturing
industries. Note that we do not impose a stationarity of hours (often adopted at the aggregate level
based on the balanced-growth path assumption). We also consider an alternative measure which is
stationary: the de-trended average workweek of production workers of the industry. In this case,

�t ]� , where n
�t denotes deviations of weekly hours from a
we estimate an analogous model for [Δxt , n
linear time trend. The main conclusion of the paper is not aﬀected by the choice of hours.4

2.2

Data

Industry productivity and hours from the NBER Manufacturing Productivity Database (Bartlesman
and Gray, 1996) are used to estimate (1). They include 459 four-digit manufacturing industry
data for 1958-1996 and largely reﬂect information in the Annual Surveys of Manufacturing. For
3

One class of models that is potentially inconsistent with our identifying restriction is endogenous growth models

in which non-technology shocks aﬀect the level of technology in the long run.
4

Altig, Christiano, Eichenbaum and Linde (2002) and Vigfusson (2002) suggest that a strong negative response

of hours might be due to omitted variables in a VAR and/or over-diﬀerencing of hours.

productivity, we use the measure of TFP growth contained in the Database (again see Bartlesman
and Gray), which is based on measuring separate factor inputs for non-energy materials, energy,
labor, and capital. For TFP higher than four-digit industries, we aggregate four-digit data weighting
by the industry’s value-added.
Industry output reﬂects the value of shipment divided by the price deﬂator of industry output.5 For hours worked, we use total hours employed in the industry, measured by the sum of hours
of production and non-production workers. There are no data on workweeks for non-production
workers. We follow the NBER Database’s convention of setting the workweek for non-production
workers equal to 40. We obtain a similar result when we assume that hours of non-production
workers are perfectly correlated with those of production workers. The NBER Database only in
cludes the wage and salary costs of labor. In calculating the industry labor share, we magnify wages
and salary payments to reﬂect the importance of fringe payments and employer FICA payments
in its corresponding two-digit manufacturing industry. The ratio of these other labor payments to
wages and salaries in two-digit industries, in turn, is based on information in the National Income
Product Accounts. Material expenditure includes expenditure on energy as well as on non-energy
materials. The capital’s share is calculated as a residual from labor and material share following the
Database’s convention.6 Finally, we use 458 industry data excluding “Asbestos Product” industry
(SIC 3292) due to termination of time series in 1993.
All VARs have a lag of one year. While our data are annual, in many industries, we maintain
nearly as many observations in quarterly data by constructing a panel. For example, in the estima
tion of the three-digit “men’s and boy’s furnishing and clothing” industry (SIC 232), we construct
5

Including or excluding inventory changes in output does not aﬀect the estimates in any signiﬁcant way.

This implicitly assumes a constant returns to scale production technology, a reasonable approximation of U.S.

manufacturing, according to Basu and Fernald (1997) and Burnside, Eichenbaum, and Rebelo (1995).

a panel by stacking six four-digit sub industry data (SIC 2321, 2322, 2323, 2325, 2326, 2329).
Likewise, for the panel estimation of a two-digit industry VAR, three-digit industry data are used;
for the panel estimation of aggregate manufacturing, durables, and nondurables, two-digit data are
used. When a VAR is estimated by panel data, industry dummies are included to allow for diﬀerent
average growth rates of TFP and hours across sub-industries. We also report the estimates based
on the aggregated time-series data. The results do not change signiﬁcantly. Standard errors are
computed by bootstrapping 500 draws.

2.3

Results

Figure 1 displays the responses of TFP and hours for the aggregate manufacturing industry. In
response to a one-standard-deviation technology shock (which eventually increases the manufac
turing TFP by 2.8 percent), hours worked increases by .7 percent at impact. Hours continues to
rise for two years until it reaches the new steady state, 1.7 percent higher than before. In response
to a non-technology shock, the manufacturing industry experiences a temporary increase of TFP,
suggesting a pro-cyclical factor utilization. Hours worked increases and remains high. We obtain
similar results with stationary hours, the (linearly de-trended) average workweek of production
workers. The average workweek increases in the short run in response to both technology and
non-technology shocks.
While we ﬁnd similar responses for durables and nondurables, the hours response varies vastly
across industries. For example, in “Transportation Equipment,” hours increases almost by 7 percent in response to a technology shock which increases the TFP by 4 percent in the long run;
whereas hours falls and persistently remains low in response to a technology shock in “Agricultural
Chemicals” (Figure 2).

Table 1 lists both unconditional and conditional correlations between the growth rates of hours
and TFP for two-digit industries.7 For aggregate manufacturing, the unconditional correlation between TFP and hours is .39 (with standard error of .05). The correlation conditional on technology
shocks is .72 (.22); the manufacturing industry tends to hire more labor when there is technological
progress. The conditional correlation on non-technology shocks is also signiﬁcantly positive, .73
(.03); a temporary increase in TFP is associated with longer hours of work.
The correlation conditional on technology varies across two-digit industries. Among those
statistically signiﬁcant, it ranges from -.79 (.46) in “Lumber and Wood Products except Furniture”
to .99 (.02) in “Tobacco Products.” Yet the majority of two-digit manufacturing industries show
positive correlations between TFP and hours conditional on technology shocks; 12 industries exhibit
.7 or higher. Among those statistically signiﬁcant at 10 percent, 14 industries exhibit a positive
correlation conditional on technology whereas only two industries exhibit a negative conditional
correlation. This pattern is robust across the level of aggregation. Among four-digit industries,
233 exhibit statistically signiﬁcant positive correlations conditional on technology whereas only 30
industries show negative correlations conditional on technology shocks.
The conditional correlation reﬂects the overall response of hours to a technology shock. As our
primary interest is in the short-run response, in Table 2 we summarize the number of industries
with positive and negative contemporaneous response of hours to technology from the bi-variate
industry VARs. The numbers in parentheses represent the cases that are statistically signiﬁcant at
7

Following Galı́, we compute the conditional correlation on technology based on VAR estimates. Given an estimate

for C(L), estimates of conditional correlations are obtained as:
�∞

1i 2i
j=0 Cj Cj
cor(Δxt , Δnt | ε ) = �
var(Δxt | εi ) × var(Δnt | εi )
i

for i = x, n, where var(Δxt | εi ) =

�∞

1i 2
j=0 (Cj )

and var(Δnt | εi ) =

�∞

2i 2
j=0 (Cj ) .

10 percent. Looking at the ﬁrst row, the two-digit industry panel-data estimates, 4 industries exhibit
negative responses (only one of them is statistically signiﬁcant at 10 percent), whereas 16 industries
show positive responses (8 of them signiﬁcant). The result is similar when the aggregated (nonpanel) data are used. There are 14 positive and 6 negative responses. For three-digit industry paneldata estimates, 115 (53 signiﬁcant) industries show a positive response and 25 (3 signiﬁcant) show
a negative response. Within the full sample of the 458 four-digit industries, 343 (148 signiﬁcant)
industries show a positive response, whereas 115 industries (13 signiﬁcant) show a negative response.
Despite considerable heterogeneity across sectors, the employment eﬀect of technology does not
appear strongly inconsistent with the equilibrium view: technological progress tends to increase
the demand for labor. However, its quantitative importance for the cyclical variation of hours (in
terms of the forecast error variances from the VAR) is small: technology shocks account for less
than 20 percent of two-year volatility of hours growth in manufacturing.8

TFP vs. Labor Productivity

Our result—hours worked increases in response to a trend component of TFP—appears at odds with
Kiley’s, which shows that the technology-driven components of labor productivity and employment
are negatively correlated for 15 of the 17 two-digit manufacturing industries for 1968:II-1995:IV.
However, we do not see our results as necessarily in conﬂict with Kiley’s. In Kiley as well as in Galı́
technology shocks are identiﬁed by the stochastic trends of labor productivity. In fact, when we
use labor productivity (instead of TFP), we also ﬁnd a strong negative response in hours in most
industries, consistent with Kiley.
8

This is a common result in structural VARs with long-run identiﬁcation such as Blanchard and Quah based on

aggregate unemployment and output.

In this section, we provide an explanation for this diﬀerence by showing that the stochastic
trends of labor productivity and TFP reﬂect diﬀerent types of changes in production process over
time. Figure 3 exhibits TFP, labor productivity (value-added divided by total hours), and total
hours worked in manufacturing for 1958-1996. While both TFP and labor productivity show
positive trends, the magnitudes are somewhat diﬀerent. To illustrate, suppose the labor-augmenting
technology, denoted by Xt , grows at rate g. If the balanced growth property were held, output,
capital, and labor productivity would grow at rate g and the measured TFP, Zt , would grow at
rate αg, where α is labor share. Given the average labor share of .5 in manufacturing during
the sample period, the labor productivity should have grown twice as fast as the measured TFP
to be consistent with the balanced growth path. Yet in the last 40 years, TFP has doubled and
labor productivity has tripled. In other words, the balanced growth path predicts that the capitaleﬀective labor ratio ( NKt Xt t ) is stationary. However, this ratio has decreased signiﬁcantly (almost by
50 percent) in aggregate manufacturing from 1958 to 1996.
The diﬀerence between the two productivity measures is dramatic in some industries. In
“Leather and Leather Products” no trend appears in the TFP shown in Figure 4, whereas labor
productivity exhibits a strong trend caused by a continuous decrease in employed hours over time.
For aggregate manufacturing, we could not reject the null-hypothesis of no co-integration between
TFP and labor productivity at 10 percent signiﬁcance level. The null-hypothesis of no co-integration
is not rejected at 10 percent for 17 of the 20 two-digit manufacturing industries.
Consider a production function Yt = F (Nt , Kt , Mt ; Zt ) where Yt , Nt , Kt , Mt , and Zt denote
output, labor, capital, material, and TFP, respectively. With lowercase letters for logged values,
the growth rate of labor productivity Δ(yt − nt ) is:
Δ(yt − nt ) � Δzt + (αn,t − 1)Δnt + αk,t Δkt + αm,t Δmt ,

(2)

where αn,t denotes output elasticity of labor, and so forth. These elasticities, measured by the
average revenue shares at time t and t − 1, are allowed to vary over time to accommodate a factorbiased technological progress.
Given the technology, as hours increases the labor productivity decreases along the downward
sloping marginal product of labor (αn < 1). In general, labor productivity growth can reﬂect
improved technology, decreased hours of work, or increased use of other inputs. Thus, changes
in material-labor and capital-labor ratios due to shifts in input prices aﬀect labor productivity,
whereas such changes alone will not aﬀect the TFP.
In order to understand the extent to which input growth accounts for labor productivity growth,
we decompose the stochastic trend of labor productivity into the input growth and TFP growth
based on (2). We ﬁrst expand the VAR to include other inputs such as capital and material:
[Δxt , Δnt , Δkt , Δmt ]� = C(L)εt . We impose a similar identifying restriction in which only tech
nology shocks aﬀect productivity x in the long run: C 12 (1) = C 13 (1) = C 14 (1) = 0. Two sets
of estimates for C(L) are obtained, respectively, with TFP (denoted by Model A) and labor pro
ductivity (denoted by Model B) as a measure of productivity x. We do not attempt to identify
other shocks as it would require further (probably controversial) restrictions. The contribution of
input growth on labor productivity is calculated based on its long-run multiplier from the VAR
and output elasticity (measured by average revenue share); speciﬁcally, (αn − 1)C 12 (1), αk C 13 (1),
and αm C 14 (1), respectively, for labor, capital, and material.
According to Model A (in which technological progress is identiﬁed by the permanent compo
nents of TFP) in Table 3, a permanent TFP shock increases the labor productivity by a similar
magnitude in the long run, as the contributions of inputs on labor productivity tend to oﬀset each
other. For aggregate manufacturing, a 2.76 percent increase (caused by a one-standard-deviation

shock from the VAR) of labor productivity is decomposed into -.51 percentage points due to an
increased hours of work ((αn − 1)Δn), -.26 due to a decreased capital (αk Δk), .77 due to an increased material (αm Δm), and 2.76 due to an improvement in TFP (ΔT F P ). A similar pattern is
found across two-digit industries; the labor productivity and TFP grow in a similar magnitude in
the long run.
By contrast, when labor productivity is used (Model B) to identify technology, a signiﬁcant
portion of labor productivity is explained by an input growth. For aggregate manufacturing, a 4
percent increase in output per hours (caused by a one-standard deviation shock from the VAR)
consists of 2.44 percentage points due to an increase in TFP, .28 due to decreased hours, -.08 due
decreased capital, and 1.36 due to increased material. In nondurables, hours plays a more important
role. For a 3.25 percent increase of labor productivity, a 1.49 percentage point increase is due to
increased TFP and a .79 percentage point increase is due to decreased hours. Overall, about half
of the trend in labor-productivity is accounted for by the input growth in manufacturing.

Implications for Two Alternative Hypotheses

The industry VAR analysis reveals a considerable heterogeneity across sectors in the hours response
to technology. A negative response, in particular, is apparently inconsistent with the prediction
of the baseline (ﬂexible-price) equilibrium model. Alternative models have been proposed to allow
for a negative response of hours to technological progress. In this section we document some facts
that provide implications for two alternative hypotheses: sticky prices (Galı́, Kiley, and Basu et
al.) and labor-saving technological progress (Francis and Ramey).9
9

Also, Jermann (1998) shows that a combination of habit formation in consumption and adjustment cost in

investment can generate a negative response of hours to a favorable technology shock.

4.1

Sticky Prices

Intuitively speaking, when price is ﬁxed, the demand for goods remains unchanged and ﬁrms need
fewer inputs, including labor, to produce the same amount of output thanks to the improved TFP.
We test this hypothesis by asking whether the industry response of hours (to technology shocks)
from a VAR is systematically correlated with the stickiness of industry-output price.
We take advantage of the recent micro data constructed by Bils and Klenow (2002), who
compute the average monthly price-change frequency for 350 goods and services from unpublished
data on the price quotes collected by the BLS for 1995-1997. We exploit the cross-sectional variation
in these measures. For 87 manufacturing goods, we are able to match the SIC industry classiﬁcation
with the ELIs.10 In matching the two data sets, each ELI corresponds to a four-digit SIC industry
for 44 goods. For 11 goods, one ELI item corresponds to multiple four-digit SIC industries. In
this case, we aggregate TFP (weighted by value-added output) and hours of the industries. For
32 goods, multiple ELIs belong to one three- or four-digit SIC industry. In this case, the CPI
weights from the BLS are used to calculate the average price-change frequency of the goods. The
average duration of price (the inverse of average price-change frequency) for 87 goods is 3.4 months.
Gasoline is at the ﬂexible end of the spectrum with an average duration of 0.8 months; newspapers
are at the sticky end with an average duration of 29.9 months.
In Figure 5, we plot the short run response of hours to a technology shock against the (log)
average monthly duration of prices for 87 manufacturing goods. The short run response refers to
the contemporaneous eﬀect on hours of a technology shock that increases the industry TFP by one
10

To calculate the CPI, the BLS collects prices for about 71,000 non-housing goods and services per month. These

are collected from around 22,000 outlets across 44 geographic areas. The BLS divides non-housing consumption into
roughly 350 categories called “entry-level items” (ELIs).

percent in the long run. Under the sticky-price hypothesis we expect a negative correlation between
the hours response and average price duration. However, no systematic relationship appears; the
cross-sectional correlation between the hours responses and average duration of prices is .02. We
repeat the same plot (the right panel), now with y-axis representing the hours response to a shock
that increases the labor productivity by one percent in the long run (based on VARs of hours and
labor productivity). Again, we do not ﬁnd a systematic relation between the hours response and
average duration of price.
Our evidence—a near zero correlation between price duration and the VAR statistics—does not
reject a potential role of sticky price for the propagation mechanism of business-cycle ﬂuctuations.
However, it suggests that price stickiness may not be a primary reason for ﬁrms’ use of labor input
diﬀerently across industries in the face of permanent changes in technology.11

4.2

Labor-Saving Technological Progress

Technological changes often accompany substitution of inputs in production (e.g., factor-biased
technological progress). Francis and Ramey show that labor-saving technological progress may
decrease hours worked under a strong complementarity between capital and labor. Consider a CES
production function that nests Francis and Ramey:
�
� σ
1
1− 1 σ−1
Y�t = αt (Xt Nt )1− σ + (1 − αt )Kt σ
.

(3)

Here Y�t denotes the value-added output, Xt , a labor-augmenting technological progress, and σ the
substitution elasticity between capital and labor. In Francis and Ramey, the production technol11

Carlsson (2000) and Marchetti and Nucci (2001) provide evidence supporting the sticky price hypothesis based

on, respectively, Swedish and Italian manufacturing data. Both studies use a method similar to Basu et al. to identify
the technology and ﬁnd that a negative response of hours to a technology shock is more pronounced in sectors with
stickier prices. We discuss the method of Basu et al. in Section 5.

ogy is a Leontief (σ = 0) and the labor-saving technological progress refers to an increase in Xt
associated with a decrease in αt . The ﬁrst order conditions of a ﬁrm’s cost minimization imply:

Wt Nt
αt
Kt 1 −1
=
(
)σ ,
Rt Kt
1 − αt Nt Xt

(4)

t Nt
where Wt and Rt denote wage rate and rental rate for capital. The factor-share ratio ( W
Rt Kt ) reﬂects

αt
the output-input elasticity ( 1−α
) and the capital-eﬀective labor ratios ( NKt Xt t ). If the production
t

function is a Cobb-Douglas (σ = 1), capital-labor ratio has no impact on factor shares; the labor
share simply reﬂects a technological change in α. However, when capital and labor are complements
(σ < 1), a decrease in capital-eﬀective labor ratio decreases the labor share relative to capital.
We note two observation in manufacturing in the last four decades. First, the labor share
(in value-added) has signiﬁcantly decreased in manufacturing (from .57 in 1958 to .39 in 1996).
Second, the capital-labor ratio has not kept up with the measured TFP. Given the time series of
factor shares, capital-labor ratios, and Xt (based on the TFP in Section 2), one can compute the
implied values of αt over time that satisfy the equation (4).12 Figure 6 shows the implied times
series of αt , respectively, with σ of 2/3, 1, and 1.5 for aggregate manufacturing. This range of σ
includes the empirically plausible values for U.S. manufacturing (Lucas, 1969; Berndt, 1976).
The implied values of αt has indeed decreased over time, supporting the mechanism proposed
by Francis and Ramey. Even when capital and labor exhibit a fairly low degree of substitutability
(σ = 2/3), almost half of the decline of labor shares are attributed to the decrease of αt . However,
a too strong complementarity (such as Leontief) may rule out the role of αt in the downward
trend of labor shares because the capital-labor ratio has not grown as fast as the labor-augmenting
12

The measured TFP is adjusted by the (two-period average) labor shares to be consistent with the equation (3):

αt Xt = T F Pt .

productivity in manufacturing. (Recall (3).) For example, when σ = .5, the observed decline of
labor share is consistent with a constant α because

Kt
Nt Xt

has decreased by almost 50 percent during

the sample period.

Some Caveats

Shea and Basu et al. investigate the employment eﬀect of technology with disaggregate data. Both
studies use industry TFP as we do, but draw somewhat diﬀerent results; Basu et al. ﬁnd a negative
correlation between technology and inputs, especially for labor; Shea ﬁnds hours increases in the
short run but decreases in the long run. We brieﬂy describe the methodological diﬀerences here.
We share the concern of Shea and Basu et al. that the measured TFP contains cyclical com
ponents such as factor utilization. In Basu et al., TFP is corrected for both capital utilization
and labor eﬀort. Despite their careful analysis, this method is potentially vulnerable to a possible
spurious negative correlation between the corrected TFP and hours, the variable used to approx
imate the utilization rate. (See Bils [1998] for detailed discussion on this.) Shea takes a unique
approach by making use of direct measures such as R&D and patents. However, confronted with
an identiﬁcation problem in a VAR, he imposes a restriction on the contemporaneous eﬀects. The
technology variable is placed last in a VAR: industry R&D is allowed to respond to the business
cycle but not vice versa. We rely on a long-run restriction on the times series of TFP assuming
that intensity of factor utilization has no trend during the sample period.
We provide three caveats on our empirical analysis. First, TFP in the NBER Database is
constructed under two assumptions: the price-cost markup of 1 and constant returns to scale
technology. While they are reasonable approximations of U.S. manufacturing (e.g., Basu and
Fernald, and Burnside et al.), input growth and TFP growth may be spuriously correlated if the

true markup is higher than 1 (Hall, 1987).13 Suppose the growth rate of the “true” TFP is Δzt∗
and the markup is µ. The measured TFP growth (incorrectly assuming a markup of 1) is:

Δzt = Δzt∗ + (µ − 1)(αm Δmt + αn Δnt + αk Δkt ).

(5)

If the actual markup is above 1, the measured TFP is spuriously correlated with inputs. In
Table 4, we re-estimate the bi-variate VAR of hours and TFP adjusted for the markup of 1.1 and
1.2 based on (5). With the makeup ratio of 1.1, the result remains the same: technology shocks
tend to increase hours worked in the short-run. When the markup is 1.2, fairly high given the small
proﬁt rates in manufacturing, a permanent increase in TFP now has a negative impact on hours
in the short run and virtually no eﬀect on hours in the long run. However, we are concerned that
too high a markup may generate a spurious negative correlation between the corrected TFP and
inputs.
Second, our analysis is based on the gross output. The contribution of material input does
not appear in the net output measure such as value added. When the value added measures are
used, the stochastic trends in TFP and labor productivity do not diverge so much as in the gross
output measure.14 Nevertheless, dissimilarity between the two measures persists in manufacturing.
In a bi-variate VAR of aggregate manufacturing, hours increases by .32 percent initially and by
.85 percent in the long run in response to a one-standard-deviation permanent TFP shock (which
eventually increases the TFP by 8.9 percent). The conditional correlation between hours and the
permanent component of TFP is .57 (with standard error of .43). By contrast, hours decreases
13

We thank Jordi Galı́ for suggesting this exercise.

The value-added-based TFP growth of the industry is obtained by Δz� =

Δz
1−αm

as suggested in Basu and Fernald

(1999), where Δz is TFP growth based on gross output. This implicitly assumes a Leontief technology between value
added output and material input.

by .89 percent initially but increases by .1 percent in the long run in response to a permanent
labor productivity shock (which eventually increases the labor productivity by 5.9 percent). The
conditional correlation of hours and permanent components of labor productivity is -.58 (with
standard error of .22).
Finally, our data are limited to manufacturing, no longer a major sector of the U.S. economy.
When the aggregate (nonfarm business sector) TFP and hours are used, hours worked slightly
decreases (statistically not signiﬁcant) in response to a permanent TFP shock (Table 4). However,
as Figure 7 shows, we still see a diﬀerence in the hours responses; a permanent labor productivity
shock generates a much more pronounced negative response of hours. Given the considerable
heterogeneity within manufacturing, it appears that more research on micro and historic data—
such as Gort and Klepper (1982), Grilliches and Lichtenberg (1984), Kortum (1993), Shea (1998),
and Basu et al. (1999)—are needed to better understand what technology shocks are and what
they do.

Conclusion

Based on aggregate time series of labor productivity and hours, Galı́ and many others report
that favorable technology shocks may reduce hours worked in the short run, which is apparently
incompatible with the baseline equilibrium model. We investigate whether this ﬁnding is robust in
more disaggregated data.
According to our analysis of 458 U.S. manufacturing industries for 1958-1996, hours response
varies vastly across industries. Many industries exhibit reduction in hours in response to a per
manent increase in TFP, consistent with earlier studies. However, there are far more industries in
which technological progress leads to a signiﬁcant increase in hours both in the short and long run.

We provide a reconciliation with earlier studies by showing that the stochastic trends of labor pro
ductivity and TFP reﬂect diﬀerent changes in production in manufacturing as labor productivity
reﬂects changes in input mix as well as improved eﬃciency.
Our analysis sheds some light on two hypotheses that allow for a negative response of hours
to technology: sticky prices and labor-saving technological progress. For 87 manufacturing goods,
the cross-sectional correlation between the hours response (to technology) and the measure of price
stickiness (average duration of output price) is close to zero. We ﬁnd that about half of the observed
downward trends in labor shares in manufacturing may be due to technological changes in the form
of labor saving. While considerable work remains to be done, the employment eﬀect of technology
in U.S. manufacturing does not seem strongly inconsistent with the prediction of the equilibrium
view.

References
[1] Altig, David, Lawrence Christiano, Matin Eichenbaum, and Jesper Linde (2002) “Technology
Shocks and Aggregate Flucutations” mimeo.
[2] Basu, Susanto, and John Fernald (1997) “Returns to Scale in U.S. Production: Estimates and
Implications” Journal of Political Economy 105 249-283.
[3] Basu, Susanto, Miles Kimball, and John Fernald (1998) “Are Technology Improvements Con
tractionary?” International Finance Discussion Paper No. 625, Board of Governors of the
Federal Reserve System.
[4] Berndt, Ernst R. (1976) “Reconciling Alternative Estimates of the Elasticity of Substitution”
Review of Economics and Statistics 58:1, 59-68.

[5] Bils, Mark (1998) “Discussion” Beyond Shocks: What Causes Business Cycles Conference
Series No. 42, Federal Reserve Bank of Boston, 256-263.
[6] Bils, Mark, and Peter Klenow (2002) “Some Evidence on the Importance of Sticky Prices”
mimeo.
[7] Blanchard, Olivier J., and Danny Quah (1989) “The Dynamic Eﬀects of Aggregate Demand
and Supply Disturbances” American Economic Review 79:1, 1146-1164.
[8] Burns, A. F., and W. C. Mitchell (1946) Measuring Business Cycles National Bureau of
Economic Research.
[9] Burnside, Craig, Martin Eichenbaum and Sergio Rebelo (1995) “Capital Utilization and Returns to Scale and Externalities?” NBER Macroeconomics Annual 67-109.
[10] Carlsson, Mikael (2000) “Measures of Technology and the Short-Run Responses to Technology
Shocks” Manuscript.
[11] Dotsey, Michael (1999) “Structure from Shocks” Federal Reserve Bank of Richmond Working
Paper No. 99-6.
[12] Francis, Neville, and Valerie Ramey (2002) “Is the Technology-Driven Real Business Cycle
Hypothesis Dead? Shocks and Aggregate Fluctuations Revisited” NBER Working Paper No.
8726.
[13] Galı́, Jordi (1999) “Technology, Employment, and the Business Cycle: Do Technology Shocks
Explain Aggregate Fluctuations?” American Economic Review 89, 249-271.
[14] Galı́, Jordi, J. David Lopéz-Salido and Javier Vallés (2002) “Technology Shocks and Monetary
Policy: Assessing the Fed’s Performance” Manuscript.

[15] Gort, M., and S. Klepper (1982) “Time Paths in the Diﬀusion of Production Process” Economic
Journal 92, 630-653.
[16] Grilliches, Zvi, and F. Lichtenberg (1984) “R&D and Productivity Growth at the Industry
Level: Is There Still a Relationship?” Patents and Productivity, Zvi Grilliches (ed.) Chicago,
University of Chicago Press 466-501.
[17] Hall, Robert E. (1987) “Productivity and Business Cycles” Carnegie-Rochester Conference
Series on Public Policy, 27, 421-444.
[18] Jermann, Urban (1998) “Asset Pricing in Production Economies” Journal of Monetary Economics, 257-275.
[19] Kiley, Michael (1998) “Labor Productivity in U.S. Manufacturing: Does Sectoral Comovement
Reﬂect Technology Shocks?” mimeo.
[20] Kortum, Samuel (1993) “Equilibrium R&D and Patent R&D Ratio: U.S. Evidence” American
Economic Review 83, 450-457.
[21] Lucas, Robert E. Jr. (1969) “Labor-Capital Substitution in U.S. Manufacturing” in A.C. Har
berger and M.J. Bailey eds., The Taxation of Income From Capital Washington, The Brookings
Institution, 223-274.
[22] Marchetti, Domenico and Francesco Nucci (2001) “Unobserved Factor Utilization, Technology
Shocks and Business Cycles” Working Paper No. 392, Bank of Italy.
[23] Shea, John (1998) “What Do Technology Shocks Do?” NBER Macroeconomics Annual 275310.
[24] Solow, Robert (1957) “Technical Change and the Aggregate Production Functions” Review of
Economics and Satistics 39, 312-329.
20

[25] Vigfusson, Robert (2002) “Why Does Employment Fall After a Positive Technology Shock?”
Board of Governors, Federal Reserve Bank, mimeo.

Table 1: Correlations between TFP and Hours in Manufacturing for 1958-1996
SIC

Industry

Unconditional

Conditional
Technology
Nontechnology

Aggregate Manufacturing

.3953∗∗
(.0560)

.7262∗∗
(.2218)

0.7238∗∗
(0.0320)

Durables

.5026∗∗
(.0624)

.7098∗∗
(.2114)

.7555∗∗
(.0329)

.0327
(.0788)
.5578∗∗
(.0585)
.4763∗∗
(.0495)
.2266∗∗
(.0743)
.4474∗∗
(.0694)
.5473∗∗
(.0525)
.4415∗∗
(.0714)
.5198∗∗
(.0527)
.3599∗∗
(.0686)
.2182∗∗
(.0550)

−.7909∗
(.4696)
.8521∗∗
(.0948)
.8958∗∗
(.1193)
.4518
(.6231)
.9463∗∗
(.0538)
.9469∗∗
(.0613)
.7189∗
(.3728)
.9794∗∗
(.0186)
.8938∗∗
(.0899)
.7012∗∗
(.3276)

.5878∗∗
(.1289)
.7661∗∗
(.1214)
.7147∗∗
(.0316)
.7098∗∗
(.1288)
−.6360
(.6418)
.7462∗∗
(.2236)
.8238∗∗
(.0324)
−.5694∗∗
(.1681)
−.6201
(.5788)
.6603
(.4319)

.2698∗∗
(.0600)

.7094
(.4755)

.7121∗∗
(.0936)

−.0142
(.0848)
.4122∗∗
(.0711)
.1902∗∗
(.0635)
.2701∗∗
(.0793)
.1027
(.1115)
.2947∗∗
(.0749)
.0656
(.0554)
.2672∗
(.1364)
.3128∗∗
(.0819)
.1264
(.0792)

.1044
(.6851)
.9989∗∗
(.0262)
.6293∗
(.3318)
.9006∗∗
(.2193)
−.9997
(.7403)
.9330∗∗
(.3853)
−.5983∗∗
(.2634)
.9964∗∗
(.4957)
.8154∗∗
(.3047)
.3057
(.7011)

−.6404
(.6169)
−.6492
(.5474)
.6716
(.5362)
.6635
(.6507)
.6680∗∗
(.2343)
.7086∗∗
(.2018)
.6281∗∗
(.0404)
.7229
(.6635)
.6733
(.4344)
.7189∗∗
(.3385)

24
25
32

Lumber And Wood Products,
Except Furniture
Furniture And Fixtures

Stone, Clay, Glass,
And Concrete Products
Primary Metal Industries

Fabricated Metal Products

Industrial, Commercial Machinery
And Computer Equipment
Electronic Equipment,
Except Computer Equipment
Transportation Equipment

36
37
38
39

Measuring, Analyzing,
And Controlling Instruments
Miscellaneous
Manufacturing Industries

Nondurables
20

Food And Kindred Products

Tobacco Products

Textile Mill Products

Apparel And
Other Finished Products
Paper And Allied Products

26
27
28
29
30
31

Printing, Publishing,
And Allied Industries
Chemicals And Allied Products
Petroleum Reﬁning
And Related Industries
Rubber And
Miscellaneous Plastics Products
Leather And Leather Products

Note: The correlation conditional on technology and non-technology shocks are estimates
22 standard errors. Those with double asterisks
from the VAR. The numbers in parentheses are
are statistically signiﬁcant at 5 percent.

Table 2: Short-Run Response of Hours to Technology in Manufacturing for 1958-1996
Data

Number of Industry
Negative Positive

two-digit

panel
aggregated

4 (1)
6 (0)

16 (8)
14 (5)

three-digit

panel
aggregated

25 (3)
36 (7)

115 (53)
104 (42)

115 (13)

343 (148)

four-digit

Note: The number of industries with a positive or negative short run response of
hours to a technology from industry VARs of TFP and hours. Those in parenthe
ses are the number of industries whose estimates are statistically signiﬁcant at 10
percent.

Table 3: Decomposition of Stochastic Trends in Labor Productivity: Percentage Changes

Durables

Manufacturing
Aggregate
4.51
2.54
3.13
5.38
4.39
4.75
6.81
4.94
4.38
4.10

2.98

2.76

Δ(y − n)

−0.00

0.15
−2.73
−1.44
−0.27
−3.03
−1.90
−0.49
−5.15
−3.08
−1.08

−0.63

−0.51

(αn − 1)Δn

−0.26
−1.73
−0.15
−0.28
−0.09
−0.45
−1.13
−0.08
−0.91
−1.02

−0.45

0.06
0.02
−0.37
−0.35
−0.06
0.12
−0.00
−0.35
0.54
−0.08

−0.07

−0.26

αk Δk

−1.01
1.60
0.44
1.23
−0.52
1.11
−0.36
−0.46
0.79
0.55

0.06

0.48
2.34
1.50
0.52
3.20
2.41
1.63
5.91
1.91
1.00

1.14

0.77

αm Δm

2.87
5.46
3.21
3.94
3.86
3.90
4.65
4.46
3.43
7.36

2.93

3.81
2.91
3.44
5.48
4.28
4.12
5.67
4.53
4.02
4.26

2.55

2.76

ΔT F P

3.57
6.96
4.97
6.59
4.24
4.95
5.10
6.43
4.40
8.62

3.25

6.15
4.03
4.13
5.97
4.86
6.78
9.36
5.61
4.81
5.11

5.03

4.00

1.30
3.35
−0.08
1.96
1.66
1.11
1.26
1.26
1.66
2.49

0.79

2.03
−0.76
0.49
0.30
−1.81
−0.34
0.08
−2.43
−0.84
0.44

0.11

0.28

(αn − 1)Δn

0.10
−0.46
0.02
−0.23
0.06
0.02
−0.41
−0.17
−0.19
−0.40

−0.23

0.43
0.28
−0.06
−0.20
0.01
0.02
0.16
−0.39
0.64
0.11

0.08

−0.08

αk Δk

0.93
1.48
3.38
1.79
−0.84
0.95
0.99
4.20
0.91
1.07

1.20

0.66
2.81
1.63
2.19
3.42
2.48
1.91
5.51
1.98
1.13

1.30

1.36

αm Δm

1.25
2.59
1.65
3.06
3.36
2.88
3.27
1.14
2.03
5.46

1.49

3.04
1.71
2.06
3.66
3.24
4.61
7.21
2.93
3.03
3.42

3.53

2.44

ΔT F P

Model B
Δ(y − n)

:
:
:
:
:
:
:
:
:
:
2.53

0.25
−2.58
−0.92
−0.48
0.14
−0.86
0.67
−1.13
−0.51
−0.76

Model A

Nondurables
1.84
2.75
2.58
4.42
3.39
3.70
3.83
2.79
2.80
6.13

24
25
32
33
34
35
36
37
38
39
:
:
:
:
:
:
:
:
:
:

20
21
22
23
26
27
28
29
30
31

Note: The decomposition is based on equation (2). Model A and Model B identify the technology shock, respectively, from
the stochastic trends of TFP and labor productivity.

Table 4: Imperfect Competition, Net Output, and Aggregate Economy
Productivity
Measure

Gross Output
Short Run Long Run

Net Output
Short Run Long Run

TFP
(µ = 1)

.0088
(.0065)

.0192∗∗
(.0065)

.0062
(.0075)

.0155∗∗
(.0071)

TFP
(µ = 1.1)

.0055
(.0076)

.0161∗∗
(.0078)

.0012
(.0074)

.0109
(.0077)

TFP
(µ = 1.2)

.0042
(.0073)

.0111
(.0082)

−.0059
(.0076)

.0034
(.0088)

Labor
Productivity

−.0177∗∗
(.0045)

−.0053
(.0075)

−.0158∗∗
(.0052)

.0082
(.0073)

Aggregate Economy
Short Run Long Run
TFP

−.0044
(.0043)

.0041
(.0048)

Labor
Productivity

−.0111∗∗
(.0038)

−.0024
(.0048)

Note: The numbers denote the short- and long-run response of hours to a permanent increase
in productivity. Those in parenthesis are standard errors. The aggregate economy reﬂects
the nonfarm business sector.

Figure 1: Impulse Responses of TFP and Hours – Aggregate Manufacturing

Response of Productivity to technology shock
0.035

Response of Hours to technology shock
0.04

0.03

0.03
percentage

percentage

0.025
0.02
0.015

0.02
0.01

0.01
0

0.005
0

−0.01

year

Response of Hours to non−technology shock
0.06

0.05

0.04

percentage

−3
Response
x 10 of Productivity to non−technology shock
10

4
2
0
−2

3
year

0.03
0.02
0.01

year

Note: Figure depicts the impulse response of TFP and hours to (one-standard-deviation)
technology and non-technology shocks. The dotted lines represent the 90 percent conﬁdence
intervals based on bootstrapping 500 draws.

Figure 2: Impulse Responses of TFP and Hours – Agricultural Chemicals

Response of Productivity to technology shock
0.1

Response of Hours to technology shock
0
−0.01
percentage

percentage

0.08
0.06
0.04
0.02
0

−0.02
−0.03
−0.04

−0.05

year
Response of Productivity to non−technology shock
0.04

0.06
percentage

percentage

Response of Hours to non−technology shock
0.08

0.03
0.02
0.01

0.04

0.02

0
−0.01

3
year

year

Figure 3: TFP, Labor Productivity, and Hours – Manufacturing

3.5
TFP
Labor Productivity
Hours Worked

2.5

1.5

1
1955

1960

1965

1970

1975

1980

1985

1990

1995

2000

Year

Note: All variables are relative to the 1958 value. Labor productivity is value added output
divided by total hours worked.

Figure 4: TFP, Labor Productivity, and Hours – Leather and Leather Products

SIC 31
2.5
TFP
Labor Productivity
Hours Worked

1.5

0.5

0
1955

1960

1965

1970

1975

1980

1985

1990

1995

2000

Year

Note: All variables are relative to the 1958 value. Labor productivity is value added output
divided by total hours worked.

Figure 5: Price Duration and Hours Response to TFP

TFP shock

Labor Productivity shock

2
cor(srr,duration) = 0.08

1.5

0.5

short−run response

cor(srr,duration) = 0.02

−0.5

−1

−1.5

−2
−1

log(duration)

Note: The x-axis denotes the (log of) average monthly duration of industry output price
based on Bils and Klenow (2002). The y-axis represents the short run response of hours
to a shock that increases industry TFP (or labor productivity in the right panel) by one
percent in the long run.

Figure 6: Output-Labor Elasticity: Manufacturing

0.65
σ=1
σ=2/3
σ=1.5
0.6

0.55

0.5

0.45

0.4

0.35

0.3
1955

1960

1965

1970

1975

1980

1985

1990

1995

2000

Year

Note: The ﬁgure depicts the output-labor elasticity (α) implied by the equation (4). Three
lines correspond to the substitution elasticity between capital and labor (σ) of, respectively,
2/3, 1, and 1.5.

Response of TFP

Response of Hours

0.025

0.015

0.01
0.02

0.005

percentage

0.015

0.01
−0.005

0.005
−0.01

−0.015

period

Response of Labor Productivity

Response of Hours

0.018

0.01

0.016
0.005
0.014

percentage

0.012

0.01

0.008

0.006

−0.005

−0.01

0.004
−0.015
0.002

−0.02

period

6
period

Note: The top panels show the responses of aggregate TFP and hours to a (one-standarddeviation) permanent TFP shock; the bottom panels show those of aggregate labor produc
tivity and hours to a (one-standard-deviation) permanent labor productivity shock. The
dotted lines represent the 90 percent conﬁdence intervals based on bootstrapping 500 draws.

Full text of Working Papers (Federal Reserve Bank of Richmond) : On the Employment Effect of Technology : Evidence from U.S. Manufacturing for 1958-1996, Working Paper 03-06

FRASER