Technological Innovation, Resource
Allocation and Growth∗
Leonid Kogan, Dimitris Papanikolaou,
Amit Seru and Noah Stoffman
Abstract
We construct novel economic measures of innovation at the firm and aggregate level. We
combine patent data for US firms from 1926-2007 with the stock market response to news
about patents to estimate the economic magnitude of firm-level technological innovations. The
stock market reaction to patent grants strongly predicts future citations the patent receives,
thus providing validation to our procedure. According to our measure, higher technological
innovation is followed by (i) an increase in total-factor productivity and output at the aggregate
level; (ii) a temporary decline in aggregate consumption, with a subsequent reversal; and
(iii) a decline in the aggregate Tobin’s Q. We find strong patterns of reallocation of capital
towards innovating firms within an industry, and a weaker pattern across industries. Finally, we
show that differences in the rate of technological innovation across industries are an important
determinant of cross-industry differences in output growth.
∗
Acknowledgements: We thank Hal Varian for his help in enabling us to extract information on Patents
from Google Patents database. We thank Nick Bloom, John Cochrane and seminar participants at Northwestern Brownbag lunch and SITE conference for helpful discussions. The authors thank the Fama-Miller
Center at University of Chicago, the Zell Center and the Jerome Kenney Fund for financial assistance.
A
Introduction
Economists since Schumpeter have argued that technological innovation is the engine that
sustains long-term economic growth. However, the exact role played by technical change in
economic growth and business cycle fluctuations remains unclear. Similarly, while technology shocks play a central role in macroeconomic real business cycle (RBC) models (Kydland
and Prescott, 1982; Long and Plosser, 1983), there is little consensus on what these shocks
are (Cochrane, 1994). A primary reason for these ambiguities is the difficulty in measuring
technological innovation – both at firm and aggregate level – in the data. This paper attempts to fill this gap. We construct a novel economic measure of innovation at the firm and
aggregate level for US firms from 1926-2007. We use this measure to assess the reallocation
and growth dynamics within industry and across industry after bursts of innovative activity. Using our measure, we find a stronger impact of technological innovation on resource
allocation and growth relative to what existing indicators of technological change suggest.
There have been several attempts to measure technological innovation. A large strand
of literature in industrial organization has measured innovation using research and development intensity (e.g., Griliches (1990)). This measure, however, does not capture innovation
directly, but rather attempts to quantify the resources devoted to innovative activities. Similarly, while patents are often used as a proxy for innovative activity, not all patents measure
technological change, and there is wide heterogeneity in economic value of patents (e.g.,
Kortum and Lerner (1998)). This shortcoming can be circumvented by focusing on future
citations each patent generates, since citations per patent have been shown to be a good
indicator of the patent’s value (e.g., Hall, Jaffe, and Trajtenberg (2005)). However, using citation information suffers from the look-ahead bias that could be especially problematic when
examining reallocation and growth dynamics subsequent to innovative activity. There have
been also a variety of indirect measures of technical change employed by macro-economists,
including Solow residuals (purified or unpurified), but the limitations of these measures are
well known (e.g., Basu, Fernald, and Kimball (2006)). Finally, while many of these shortcomings are resolved in the measure of Alexopoulos (2011) based on books published in
the field of technology, it is hard to evaluate reallocation and growth dynamics across firms
within industry and across industries using this measure.
We propose a new measure that circumvents several of the limitations of previous measures. We construct our measure both at the firm and the aggregate level over an sample
spanning more than eighty years. Our main novelty is that we combine patent data with the
stock market response to news about patent application/approval. Weighing each patent
based on the firm’s idiosyncratic stock return on the day the patent was applied or issued
1
serves as a signal of the economic value of each patent. The benefit of this value adjustment is that it does not use forward-looking information such as citations by future patents.
We show that, among other things, our measure of patent value strongly predicts future
citations.
We evaluate the validity of our measure of innovation in several ways. First, we show
that it captures known periods of high technological progress as well as firms driving these
waves (e.g., technologically progressive 1930s, see Field (2003)). Second, we show that
the empirical distribution of firm-level innovation measure is extremely fat-tailed and the
few large firms contributing disproportionately to the aggregate rate of innovation in the
economy vary by decade. This is consistent with past research which describes the nature
of radical innovations (Harhoff, Scherer, and Vopel (1997)). Finally, we find that firms that
innovate as per our measure have higher productivity subsequent to innovative activity and
the characteristics of these firms match those of innovators as described by Baumol (2002),
Griliches (1990) and Scherer (1983).
We find that higher technological innovation at the aggregate level is followed by an
increase in total-factor productivity in the economy. Output increases as a response to innovation, but only after a delay of 4-5 years. We find that technological innovation has a
negative effect on hours worked, though the effect is not always statistically significant. A
positive shock to innovation generates a U-shaped response on consumption growth: consumption is lower in the short-run but increases in the long-run. This is consistent with a
reallocation of resources away from consumption in the short-run towards the implementation
of innovation. One interpretation of these patterns is that, following an innovation shock,
the economy initially diverts some resources to reallocation, but eventually the aggregate
output rises.
To better understand the evidence at the aggregate level, we exploit the fact that our
innovation measure is available at the micro-level and directly estimate reallocation and
growth dynamics within and across industries consequent to innovative activity. We find
that, within an industry, inputs (labor and capital) get reallocated from the firms that
innovate less to those that innovate more. The evidence for between-industry reallocation
is weaker than the evidence for within-industry reallocation, consistent with the view that
some inputs are specific to the industry. Innovative activity predicts higher subsequent
output growth of innovating firms as well as their industry. More innovative activity leads
to a fall in the aggregate Tobin’s Q, suggesting that innovation erodes market valuations of
competing firms.
Our paper is related to several strands of the literature. First, it is related to literature
that explores the short-run impact of technology shocks on inputs (e.g., Gali (1999); Shea
2
(1999); Basu et al. (2006); Alexopoulos (2011)). While Gali and Basu et al. use different approaches to measure productivity shocks, both find that favorable technology shocks reduce
hours worked in the short run, which they attribute to sticky prices.1 Shea (1999) constructs
a measure of technology shocks using patent counts and R&D spending, and finds that labor
hours increase following a positive shock. Alexopoulos (2011) raises a number of concerns
with the measures of Shea, and proposes an alternative measure based on books published
in the field of technology. We contribute to the literature by constructing a measure of innovation using information from both firm’s patenting activity and the stock market reaction.
Thus, our measure sidesteps several of the criticisms levied against patent related measures,
since they do not use forward information. Furthermore, our measures are available at the
firm- or industry level, which allows us to trace the effect of innovation at its origin, and its
propagation through the economy.
Our work is also related to the literature that explores the micro-foundations of technology shocks. In particular, Gabaix (2011) proposes that if the distribution of firm size is
sufficiently fat-tailed, as is the case in the US and in most of the world, firm-specific shocks
can have substantial effects on aggregate quantities due to the failure of the law of large
numbers. The empirical distribution of firm-level innovation measure is fat-tailed, which is
consistent with the view that firm-specific shocks can affect aggregate quantities if the law
of large number fails. However, the fact that at the aggregate level, our equal- and valueweighted measures have similar impact on aggregate quantities suggests that firm-specific
shocks to a few large firms is not the only driver of our technological innovation measure.
In addition, our work is related to the literature on the economic impact of technological
innovation (Kortum and Lerner, 2000; Hall et al., 2005; Bloom, Schankerman, and Reenen,
2010). The work of Bloom et al. (2010), who aim to quantify the externalities of innovation
by disentangling the business-stealing effect from positive research spill overs, is particularly
close to our paper. We contribute to this literature by proposing a measure of patent quality
that is free from the look-ahead bias and assessing within- as well as between-industry
reallocation and growth dynamics after bursts of innovative activity.
Finally, our work is related to the endogenous growth literature that focuses on understanding the relationship between inputs and outputs in the production of R&D, and between
R&D output and productivity growth (e.g., Kortum (1993), Caballero and Jaffe (1993) and
Acemoglu (2009)). Moreover, our findings on drop in market Tobin’Q around aggregate
1
Gali’s approach rests heavily on the assumption that demand shocks cannot affect productivity in the
long run. The approach of Basu et al. 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, after correcting for increasing returns, imperfect competition, and cyclical factor utilization.
3
technological innovations is similar to Laitner and Stolyarov (2003) and other papers such as
Greenwood and Jovanovic (1999) and Hobijn and Jovanovic (2001) that argue that arrival
of a new type of capital can lead to a large fall in the stock markets capitalization in an
economy due to obsolescence of older capital.
The remainder of the paper is organized as follows. We explain the data construction in
Section B. Section C describes the construction of our innovation measure using stock prices.
Section D explores the response of aggregate productivity business cycle quantities on our
measure of innovation. Section E studies the response of individual firms and industries
on our innovation measure and documents patterns of reallocation. Section F explores the
response of the market value of capital both at the aggregate as well as the firm level.
Section G discusses the connection of our findings with existing models and concludes.
B
Data
Our measure of innovation relies on using information on patents that a firm creates and
the stock market response to news about these patents. We now discuss our data pertaining
to patents and stock market reaction. We also elaborate on other data that we employ in
several of our analysis.
B.1
Patent Data
Patents in the United States are filed with, and granted by, the United States Patent and
Trademark Office (USPTO). Regulations allow only an individual, not a corporation, to be
an inventor. However, the inventor can assign the granted property rights to a corporation
or another person. Therefore, when patents are granted they always have an inventor, and
sometimes an “assignee” (one or more corporations or people). To implement our measure
of innovation, we need to match all granted patents to public corporations with returns
in the CRSP database.2 Previous studies have relied on the NBER patent database, which
provides such a match for patents granted between 1976 and 2006. Because our study focuses
on low-frequency macroeconomic variables, it is important to have as long a time-series as
possible. We therefore created a new database covering patents granted between 1926 and
2010. We explain in this section how this database was constructed.
2
We focus on grant dates because prior to 1999, patents were made public only when they were granted,
so unsuccessful applications remained secret. Applications are now generally publicized 18 months after
filing, although the law allows for certain exceptions, including national security grounds.
4
Matching patents to firms
We construct a new database of patent filings and issues by downloading the entire history
of U.S. patent documents from Google Patents.3 Each of about 7.8 million patent files was
downloaded using an automation script.4 Google provides the patent number and filing date
for all patents. For most patents, particularly for years after 1920, Google also provides a
text version of the patent document that was created using OCR software. We use this text
version of the document to extract the names of corporations to which patents are assigned.
Since we merge our patent data with data on stock returns, we are limited to the period
after 1926, when the CRSP database begins.
Unfortunately, the OCR technology that Google applied to the patent filings is imperfect,
and many of the downloaded documents include a great deal of garbled text. We therefore
make use of a number of text analysis algorithms to extract relevant information from the
documents. We are especially interested in matching patents to the corporations to which
they have been assigned. This requires extracting assignee names and then matching them to
company names in CRSP. We now briefly discuss the steps our matching procedure followed.
We search the document for the words “assignee” or “assigned” and extract the text that
immediately follows. This text is either a company name, or the name of an individual to
whom the patent was assigned. We then count the number of times each assignee name
appears across all patent documents. We compare each assignee name to more common
names, and if a given name is “close” (in the sense of the Levenshtein distance5 ) to a much
more common name, we substitute the common name for the uncommon name. For example,
one of the most common names is “General Electric Company”, which is associated with
over 42,000 patents. We substitute this name for the far less common, but quite similar,
names “General Electbic Oohpany”, “General Electbic Cqhpany”, and “Genebal Electbic
Compakt”.
At this point, we have an assignee name for each patent. These names must be matched
to a company identifier such as the CRSP permno. This was accomplished in two steps. We
begin by looking only at patents that are also in the NBER database. For each assignee
name identified in the steps above, we count how many different permnos are matched to
3
http://www.google.com/patents
4
Google also makes available for downloading bulk patent data files from the USPTO. The bulk data
does not have all of the additional “meta” information including classification codes and citation information
that Google includes in the individual patent files. Moreover, the quality of the text generated from Optical
Character Recognition (OCR) procedures implemented by Google is better in the individual files than in the
bulk files provided by the USPTO. As explained below, this is crucial for identifying patent assignees.
5
The Levenshtein distance is the number of edits required to make one string match another string, where
an edit is inserting, deleting, or substituting one character.
5
patents in the NBER database. For example, all of the patents with an assignee name
“General Electric Company” are matched to one permno in the NBER database. We can
therefore safely assume that all of the patents assigned to the General Electric Company can
be matched to that permno, even for patents not included in the NBER data. Remaining
assignee names are matched to CRSP firm names using a name matching algorithm.6 The
algorithm uses a score based on the inverse word frequency to match assignee names to
possible company names. For example, the word “American” is quite common in company
names, and so contributes little to name matching; the word “Bausch” is quite uncommon,
so it is given much more weight. Visual inspection of the matched names confirms very few
mistakes in the matching.
Summary statistics
We now provide some statistics that lends credence to our method for extracting patent
information. Table 1 shows how many patents we match to companies. Of the 6.0 million
patents granted in or after 1926, we find the existence of an assignee in 2.8 million. The
matching procedure provides us with a database of 1.8 million matched patents, of which
435,814 (24%) are not included in the NBER data. Figure 1 graphs the total number of
patents matched by the year the patent was granted. Patents included in the NBER data,
which is the most comprehensive database previously available, are shown in light shading.
Patents unique to our database are presented in dark shading. Note that the two sets of
data appear to fit together fairly smoothly, and that even during the period covered by the
NBER data, our database adds an average of 2,187 patents to the NBER data.7
Table 2 provides additional summary statistics. Overall, our data provides a matched
permno for 66% of all patents with an assignee, or 31% of all granted patents. By comparison,
the NBER patent project provides a match for 32% of all patents from 1976-2006, so our
matching technique works quite well, even using only data extracted from OCR documents
for the period before the NBER data. Another point of comparison is Nicholas (2008), who
uses hand-collected patent data covering 1910 to 1939. From 1926-1929, he matches 9,707
patents, while our database includes 8,858 patents; from 1930-1939 he has 32,778 patents
while our database includes 47,039 matches during this period.
In our analysis we will restrict our attention to data till 2007 (i.e., we exclude years 20086
The algorithm is based on code written by Jim Bessen, available at http://goo.gl/m4AdZ.
7
Recall that we use information on the patent-assignee match in the NBER data to assist with our
matching, so the match during the overlapping period is generally the same, by construction. An exception
is for a few cases where there is apparently a mistake in the NBER match and our patent-assignee frequencybased matching system corrects an error.
6
2010). This allows us to circumvent making inferences with patent data that is censored,
since many patents towards the end of the sample may have been applied for but are not
granted (see Hall, Jaffe, and Trajtenberg (2001) for more details).
B.2
Stock Market and Financial Data
The return data used to assess the stock market response to news about patents are from
CRSP over the period 1926-2007. In several of our analysis we use financial and accounting
data that are from Compustat. The sample in these cases is determined by the availability
of Compustat data (available from 1951 onwards). As is standard, we omit financial firms
and utilities from our analysis.8 See appendix A for variable definitions.
C
Measuring innovation
In this section we explain how we construct our firm, industry and aggregate level measures
of innovation. It is well known in industrial organization literature that while patents can
be used as a proxy of innovative activity, many patents do not measure technical change.
This shortcoming has be circumvented by inferring the value of a innovation by assessing
the future citations a patent generates (e.g., Harhoff, Narin, Scherer, and Vopel (1999), Hall
et al. (2005)). However, we depart from this line of research since citations based measures
use information from the future and as a result the analysis on reallocation and growth
dynamics after a valuable innovation – as measured by instances when patents generate
higher ex post citations – could suffer from a look ahead bias.
Our insight is to instead construct measures that combine patent data with short term
stock market response to news about these patents as a way of assessing value of patents.
While a number of papers have examined the relation between firm-level patenting activity
and firm value (e.g., Hall et al. (2005)), researchers have not used short term market reactions
to patenting activity to gauge innovation quality, as we do. An exception to this is ?? (Aus)
who conducts an event study around grant dates of 600 patents belonging to biotechnology
firms.9 As he notes, application dates of patents are not widely publicized. In contrast, the
USPTO grants patents every Tuesday. The USPTO’s weekly publication Official Gazette,
8
There is an additional reason for leaving these firms from our analysis that follows from our model:
capital-producing firms do not make any investment decisions in the model.
9
More broadly, while there is a large literature in financial economics that shows that stock market
movements reflect the market’s expectation of future development in the economy, the use of stock price
movements in macro work has been limited with the exception of Beaudry and Portier (2006) who incorporate
aggregate stock price changes in the estimation of aggregate TFP.
7
published every Tuesday, mentions which patents are granted in each date. As a result,
similar to his study, we focus our analysis on grant dates.
We now explore the impact on firm’s stock price once a patent is granted. Our central
idea is that the stock price reaction around the day that the patent is granted to the firm
should contain some information about the value of the patent. We show that this is the case
by providing two pieces of evidence. First, we document that the volume of trading on the
firm’s shares increases around the time that the patent is issued. This finding suggests that
the fact that the patent is granted carries some information to the market. Second, since
citations generated by patents have been proven to capture value of a patent, we explore
whether the stock price reaction on the day the patent is granted is informative about the
number of future citations the patent receives.
Our analysis uses stock market data, so one worry is that we do not include private
companies several of whom might be responsible for large and more important productivity
shocks. However, as Baumol (2002) notes, that while several independent and private firms
might provide initial innovation, subsequent refinement through R&D that actually leads to
really large improvements in welfare is conducted by large publicly traded firms.
C.1
Stock prices response around patent grant window
To explore the response of trading volume on the announcement window, we regress a firm’s
trading volume x on a dummy variable G taking the value one if a patent was granted to
the firm f in day t:
xf t+k = b Gf t + ρx̄t−1 + γt + ζf m + ef t ,
(1)
for the days k = −1 to k = 5. Depending on the specification, we control for the average
P
trading volume in firm f in the days prior the event (x̄t−1 ≡ 15 5l=1 xt−1−l , firm-month (ζf m )
and calendar-day fixed effects (γt ). We restrict attention to the sample of firms that have
been granted at least one patent in the 1926-2008 period.
Table 3 shows the results of estimating equation (1). We find that there is a statistically
significant increase in share turnover around the day that the firm is granted a patent.
Volume increases in the day of the announcement, and remains temporarily higher for the
next couple of days. Depending on the specification, the total turnover in the first three
days after the announcement increases by 0.35-0.60%. Given that the daily median turnover
rate is 1.29%, this is an economically significant increase in trading volume, consistent with
the view that patent issuance conveys important information to the market.
The next step is to explore whether the stock price reaction on (or after) the day of the
announcement carries information about the likelihood of the patent receiving citations in
8
the future. To isolate market movements we focus on the firm’s idiosyncratic return rf t ,
constructed in two ways. Our first idiosyncratic return measure equals the firm’s return
minus the return on the market portfolio. This measure effectively assumes that all firms
have a beta of one with the market. Campbell, Lo, and MacKinlay (1997) refer to this
measure as the “market-adjusted-return model”, and it has the advantage that one does not
need to estimate the firm’s stock market beta therefore removing one source of measurement
error. Our second measure of idiosyncratic return is the firm’s stock return minus the return
on the beta-matched portfolio (CRSP bxret). This has the advantage that it relaxes the
assumption that all firms have the same amount of systematic risk, but is only available for
a smaller sample of firms. Moreover, we explore the role of the firm’s percentage, as well
as dollar return, where the latter is defined as the firm’s idiosyncratic rf t times the market
capitalization of the firm (in 1982 dollars) the previous day Sf t−1 .
Given our measure of the firm’s idiosyncratic price movements, we examine whether
the return of the firm on or around the patent grant day predicts future citations. A firm
can always choose not to implement a patent, therefore the value of the patent cannot be
negative. Hence, we also restrict attention to the positive stock price responses only. We
estimate:
Nj = b Rj + γy + γ log σj + ej ,
(2)
where Nj is the number of future citations received by patent j and Rj contains the total
stock price reaction of the firm around the day that patent j is granted, truncated at the
1% level within grant years. Older patents are more likely to receive more citations, so
we include grant-year fixed effects (γy ). Since we are focusing on the positive stock market
reaction only, we control for the firm’s idiosyncratic volatility σj , estimated over the previous
year the patent is granted.
We show the estimation results in Panel A of Table 4. We find a strong positive relation
between a firm’s idiosyncratic price movements on the day the patent is granted and the
number of future citations it receives, provided we truncate the idiosyncratic return at zero.
The relation between the firm’s idiosyncratic stock return ([rf t ]+ or [rf t ]+ ×Sf t−1 ) and future
citations is economically significant. Given that the idiosyncratic returns are highly skewed
it is more informative to look at percentile movements. An increase from the median to the
90% percentile in terms of stock price reaction is associated with 0.16 to 0.39 more citations,
depending on the specification. Given that the median number of citations a patent receives
is 4, this increase in citations is economically significant. As a robustness check, we also
estimate equation (2) with the poisson and the negative binomial model. Our estimation
results are similar using these alternative estimation procedures. Finally, the results using
9
our second idiosyncratic return measure (the firm’s return minus the beta-matched portfolio)
are similar, though quantitatively a bit weaker (the point estimates are one-third smaller in
magnitude). Since the market-adjusted returns appear more informative, on the remainder
of the paper we focus on that measure. Nevertheless, results using the beta-matched portfolio
returns are quantitatively similar and are available on the supplementary appendix.
Moreover, we perform two more robustness tests. First, our stock-price response measure
contains a large number of zeros (roughly 50%). Hence, we explore whether the positive effect
on citations comes from the transition from zero to positive, or does it also exist if we focus
on the positive responses alone. To this end, we estimate equation 2 restricting the sample
to those patents with a positive stock price response rf t > 0. As we see in Panel B of Table 4,
the effects are economically stronger: our OLS regressions imply that an increase from the
median to the 90-th percentile in terms of stock price response (within the set of positive
responses) is associated with 0.38 to 0.99 more citations.
As a second robustness test, we explore whether expanding the announcement day window is more informative about future citations. As we saw previously, trading volume is
higher on the two days following the day the patent is granted, suggesting that the stock
price movements following the announcement are also informative. The downside is that
increasing the announcement window adds more noise. We explore this tradeoff by expanding the range of our announcement window to [t, t + 2] and also to [t − 1, t + 4]. As we
see in Panel B of Table 5, the longer window is informative about future citations, though
coefficients are smaller in magnitude, possibly due to more measurement error. An increase
from the median to the 90-th percentile in terms of stock price reaction during the [t, t + 2]
([t − 1, t + 4]) window is associated with 0.15 to 0.47 (0.14 to 0.49) higher number of patent
citations.
The results of this section suggest that the stock price reaction within a few days after
the patent is granted contains information about the value of the patent. We will use this
information as a way to appropriately weigh the number of patents when we construct
measures of innovation at the firm, industry or market level. Furthermore, since longer
windows appear to be a little bit more informative than simply focusing on grant date, we
explore alternative constructions of our measure using longer windows. In this paper, we
present results using the [t, t + 2] window. Results using longer windows are quantitatively
very similar.
10
C.2
Measuring innovation
Here, we use the stock market reaction of innovating firms to construct measures of innovation
at the firm, industry and aggregate level.
Firm-level measures of innovation
Based on the analysis above, we construct the following firm-level measure of innovation:
Af t =
P
d∈t [∆Sf d ]
Sf t
+
× If d
,
(3)
where, If d is an indicator variable that takes the value of one if a patent filed or granted by
firm f in day d and zero otherwise; ∆Sf d is the change in the firm’s market capitalization
due to it’s idiosyncratic return (Sf d−1 × rf d ); the firm’s idiosyncratic return around patent
granted on day d is measured the firm’s return minus the return on the market portfolio;
rf+d indicates that we only take those returns which are positive. Subtracting returns to a
beta-matched portfolio, rather than the market portfolio, leads to similar results.
We scale the firm-level innovation measure by firm f ’s end of year market capitalization
in year t, Sf t . Hence, our firm-level innovation measure can be interpreted as the fraction
of firm f that can be attributed to innovation in year t. Replacing market values in the
denominator by lagged market values gives very similar results.
There are two aspects related to the idiosyncratic returns that need discussion. First, the
news about a patent can get reflected in the stock returns when the patents are filed with
USPTO as well as when they are granted. In construction of our measure, we have examined
both grant and filing dates together as well as grant and filing dates separately. For brevity,
we present results using only the grant date in the paper. We note that this choice is also
appropriate since the news about granting of important patents is more likely to be easily
available to agents in the economy (e.g, financiers). This is important since we want to use
our measure to study reallocation dynamics – such as movement of inputs, such as financing,
from existing firms to innovating firms. Nevertheless, the results using both grant and filing
dates together produce results that are qualitatively similar to those in the paper. Second,
the days subsequent to the innovation over which the stock returns are calculated can also
be varied. For brevity, we report results for our measure constructed using announcement
day returns only. Extending the window around the announcement produces quantitatively
similar results.
11
Distribution of innovation
Our firm-level measure of innovation has fat tails. To see this, in Figure 3 we plot the
kernel density plot for the unconditional cross-sectional distribution of the dollar value of
P
innovation (log(Mit / i Mit )), where the dollar value of innovation for firm i in year t is
P
+
defined as Mit =
d∈t [∆Sf d ] × If d . To clearly see the property of this distribution we
P
also plot the distribution of log firm market capitalization (log(Vit / i Vit )) which is well
known to be fat-tailed. As we see, the distribution of dollar value of innovation is relatively
more fat tailed than the market capitalization one. Thus, the distribution of firm size implies
that the distribution of firm size is not solely driving our results.
Our findings highlight that a few large firms are very important for the aggregate rate
of innovation in the economy. Our findings are consistent with Harhoff et al. (1997), who
argue that the distribution of patent citations has fat tails. The fact the large firms are
important for innovation also supports the ‘granularity’ hypothesis of Gabaix (2011). The
innovative output of individual firms need not average out in aggregate, and thus offers a
micro-foundation for the time-variation in the aggregate amount of innovative activity.
We find that the largest firms responsible for innovation vary by decade. In the 1930s
and 1940s, AT&T and GM are responsible for a large share of innovative activity. In the
1950s and 1960s, du Pont and Kodak take a leading role. In 1970s and 1980s, a large share
of innovation takes place in Exxon, GE, 3M and IBM. Finally, in the 1990s and 2000s,
“new economy” firms are responsible for a large share of innovation, namely Sun, Oracle,
Microsoft, Intel, Cisco, Dell and Apple.
Aggregate measures of innovation
To construct aggregate measures of innovation, we need to aggregate our firm-level measures
across firms. Hence, we construct a dollar measure of innovation, at the industry or aggregate
level:
Ât =
X
(
X
f ∈Nt
d∈t
)
[∆Sf d ]+ × If d
.
(4)
Our aggregate measure is based on the level of stock prices. Hence, it will be mechanically
affected by the same economic forces that affect the level of the market, that might have
little to do with innovation (for example, changes in discount rates). Hence, we scale our
P
aggregate measure by the total level of the market At = Ât /Vt , where Vt =
f ∈Nt Sf t .
Thus, our aggregate measure is a value-weighted average of our value-weighted firm-level
innovation measure.
Our results in the previous section suggest that, in terms of dollar values, a few large
12
firms dominate. Indeed, our value-weighted measure is dominated by a few large firms. The
top 1% (5%) of firms in terms of innovation account for an average of 40% (70%) of the total
value of innovation across the 1926-2007 period.
An important question is whether the observed time-series variation in aggregate innovation is the result of disproportionately large idiosyncratic shocks that fail to be diversified
away, or the result of correlation in the underlying firm-level propensity to innovate. To
shed some light on this, we split the numerator in each of the two measures of innovation
(5)-(4) into two components: A1 , the part that is contributed by the top 1% of firms; and
A99 , the part that is contributed by the remaining firms. Hence, A1t + A99
t = At . If aggregate
innovation was determined by large idiosyncratic shocks, we would expect innovations in A1t
1
to be uncorrelated with A99
t . Instead, we find that the sample correlation between At and
A99
t . This result suggests that common movements in the propensity of firms to innovate are
a significant component of aggregate innovation shocks.
We also construct an equal-weighted measure of aggregate innovation as
Aew
t
1 X
=
Nt f ∈N
t
(
X
)
[rf d ]+ × If d
,
(5)
d∈t
where Nt is set of firms in CRSP at year t. Our equal-weighted measure is more diversified
across firms, since they all receive the same weight. As a comparison, the top 1% (5%)
of firms in terms of innovation account for an average of 6% (26%) of our equal-weighted
measure (5). We focus on the value-weighted measure At , but we find that using the equalweighted measure leads to quantitatively similar findings.
Finally, our measures are scaled by either the total number of firms in (5) or by the total
market capitalization of firms in (4). Consequently, they should not have a time-trend. One
source of concern is that our two aggregate measures are mechanically affected by either the
total number of firms in CRSP or by the level of the stock market. As a robustness test,
we construct alternative versions of (5)-(4) by instead using only the firms with positive
innovation amount in year t to define the scaling factor. Results using these two alternative
measures are quantitatively similar.
C.3
Validation of Measures
In this section we provide evidence that suggests that our measures are capturing technological innovation. We start with aggregate evidence. Figure 2 plots log of the two measures
over the period 1926-2006. As can be seen there is substantial variability in the two measures
over time. This is reasonable since there has been a wide variation in innovation that has
13
impacted technical change in the US over the last 80 years or so. It is also worth noting that
the wide variability in our measures is in stark contrast to number of patents or R&D stock
which has seen secular increase over time.
Our measures of innovative activity line up well with what are considered periods of
high technological progress in the U.S. For instance, our measures suggest high values of
technological innovation in the 1930s, a period referred to by many as the ‘most technologically progressive decade of the century’ (see Field (2003)). Moreover, when we dissect our
measures we find that firms that primarily contribute to technological developments during
the thirties are in the automobiles (such as General Motors) and telecommunication (such
as AT&T) sectors. This description fits well with studies that have examined what sectors
and firms led to technological developments and progress in the 1930s (see Smiley (1994)).
Similarly, our measures suggest higher innovative activity during 1960s and early 1970s –
again a period commonly recognized as a period of high innovation in the U.S (see Laitner and
Stolyarov (2003)). As has been noted, this was a period that saw development in chemicals,
oil and computing/electronics – the same sectors we find to be contributing the most to our
measure with major innovators being firms such as IBM, GE, 3M, Exxon, Eastman Kodak,
du Pont and Xerox.
Finally, developments in computing and telecommunication has brought about the latest wave of technological progress in the 1990s and 2000s, which coincides with the high
values of our measure. In particular, it is argued that this is a period when innovations
in telecommunications and computer networking spawned a vast computer hardware and
software industry and revolutionized the way many industries operate. Again, we find that
firms that are main contributors to our measure belong to these sectors with firms such as
Sun Microsystems, Oracle, EMC, Dell, Intel, IBM, AT&T, Cisco, Microsoft and Apple being
the leaders of the pack. We next turn to providing firm level evidence that lends additional
support to validity of our measures.
Characteristics of Innovating Firms
We start by exploring the characteristics of innovating firms in the data. This will allow us
to compare innovating firms identified by our measure with those found in earlier work.
In addition, the innovation of other firms can have an effect on a firm’s probability of
innovation, depending on whether innovations generate spill overs. To explore the effect
of innovation of a firm’s competitors on the firms likelihood of innovation, we construct
a measure of innovation of the firm’s competitors. We define a firm’s competitors as all
the firms in the same industry, defined at the 3-digit SIC level. We construct AIf t as an
14
value-weighted average of innovation of the firms in the same industry, excluding the current
firm.
We estimate the following regression:
rd
Af t = a0 + a1 log Qt−1 + a2 log Kt−1 + a3 log Kt−1
+ a4 AIf t−1 + uit ,
(6)
where the firm level variables in the regression include firm’s capital stock scaled by average
capital in the economy (K), its accumulated R&D expenditure stock normalized by average
assets in the economy (K rd ) and its Tobin’s Q (Q). The regression also includes industry
fixed effects to account for industry-level time invariant characteristics. In addition, we
also include time fixed effects to account for changing state of the business cycle as well as
changes in patent law and changes in the efficiency and resources of the USPTO (see Griliches
(1989)) during our sample period. Given that our innovation measures are censored at zero,
we estimate equation (6) using a Tobit regression and cluster the standard errors by firm
and year. Note that information on R&D expenditure is reliably reported in Compustat
only from 1975 onwards. As a result our sample period for regressions that use R&D stock
is restricted to 1975-2006.
As the results in the first four columns of Table 6 suggest, firms with high Tobin’s Q,
high-capital and high-R&D stock are more likely to innovate (i.e., coefficients a1 , a2 and
a3 are positive and significant). In addition, the estimated coefficient a4 is positive and
significant as well. This suggests that innovation by competitors in the past makes a firm
more likely to innovate. In other words, there is a persistent industry-specific component
of innovation. The results are robust across several specifications as shown in the table. In
the last four columns we use the sample from 1950 onwards by excluding R&D stock from
the specification. As can be observed the results are again quantitatively and qualitatively
similar. These findings are very comparable to those discussed in Baumol (2002), Griliches
(1990), Scherer (1965) and Scherer (1983) on the characteristics of firms that have conducted
radical innovation and have been responsible for technical change in the U.S.
Firm-level productivity
As a last test that lends support to validity of our measures, we examine if firms that innovate
have higher productivity subsequent to innovative activity. We measure productivity of a
firm in a given year by Yt /Kt where Y refers to sales (Compustat item sale), and capital (K)
to gross PPE (Compustat item ppegt).
In addition, the innovation of other firms can have a positive or a negative effect on
a firm’s profitability, depending on whether the innovation spill-over or business-stealing
15
effect dominates. In general, the presence of unobserved variables hampers the estimation
of externalities, as highlighted in Bloom et al. (2010). For example, common productivity
shocks could impact many firms in the same industry – thereby creating a positive correlation
between innovative activity of a firm’s competitors and its own sales. To explore the effect of
innovation of a firm’s competitors on the firms level of profitability, we include our measure
of innovation of a firm’s competitors.
We evaluate the relationship between subsequent productivity and innovation by a firm
or its competitors by estimating:
Yf t
Yf t−1
= a0 + a1 Zf t−1 + a2 Af t−1 + a3 AIf t−1 + ρ
+ uf t ,
Kf t
Kf t−1
(7)
where Af is our firm-level innovation measure, and AIt−1 is the amount of innovation by
competing firms. To isolate the effect of innovation from firm characteristics, we include a
vector of controls Z that includes firm’s Tobin’s Q, firm capital stock normalized by average
capital in the economy, firm sales normalized by capital, firm stock returns and volatility
In addition we include industry (I) and/or time (T) fixed effects and report standard errors
that are clustered by firm and year. The main estimates of interest in this specification are
a2 and a3 which capture the change in productivity of firm’s capital in response to changes
in innovation by the firm and its competitors, respectively.
The results are reported in Table 7 and show that there is a substantial increase in its
productivity of a firm’s capital subsequent to an innovation. The magnitudes are large as
well, especially when dynamic effects are taken into account.10 Our estimates imply that an
increase in innovation by the firm from the 50th to the 90th percentile leads to increase in the
firm’s productivity by 0.12 to 0.31, which is economically significant relative to the median
level of 2.4. The fact that the productivity of capital increases lends strong support to the
notion that our measure are meaningfully related to technological innovations. Furthermore,
we find some evidence of business-stealing using our value-weighted measure. In particular,
a one-standard deviation increase in the amount of innovation by the firm’s competitors
lowers firm-level productivity by 0.23 to 1.14 units.
D
Innovation and Aggregate Dynamics
We explore the response of aggregate quantities and prices to our innovation measure constructed at a more aggregate level. We focus on both economy wide and industry-level
10
We compute the cumulative dynamic response of productivity as ak /(1−ρ) using the coefficient estimates
of 7.
16
responses.
D.1
Aggregate variables
In this section we evaluate growth dynamics using our measure of innovation constructed
at the aggregate level. We present benchmark results and comparisons of these results
with other those obtained from other specifications as well as using alternative measures of
innovation.
Benchmark results
We start by exploring the relation between measures of innovation and quantities of interest
using VARs. We focus on utilization-adjusted total factor productivity from Basu et al.
(2006); per capita gross domestic product; hours worked; consumption of services and nondurables; investment in equipment and software; and the relative price of equipment and
software to the consumption deflator. We deflate quantity variables by population growth,
and nominal variables by the consumption deflator.
We estimate tri-variate VARs of the form Zt = [log(Xt ), log σt , log(At )]0 , where X is our
variable of interest and A represents our dollar measure of aggregate innovation, scaled by the
market capitalization of the market. We include the cross-sectional average of idiosyncratic
volatility σ to ensure that our innovation measure does not pick up movements in firmlevel volatility due to our truncation at zero. We include a deterministic trend, following
Alexopoulos (2011). We compute impulse responses by ordering the innovation shock A last,
hence the technology shock affects only the variables of interest with a lag. We select the
number of lags using the Akaike-Information criterion, which advocates a lag length of one
to two years for all of the systems. We compute standard errors by a bootstrap simulation
of 500 samples.
We plot the impulse-response functions in Figure 5, along with 1.65 standard error bands.
Total-factor productivity increases following an increase in innovation activity. The forecast
error variance attributed to our innovation measures 14% at the 8-year horizon. This is to
results in Alexopoulos (2011), but in contrast to Shea (1999) who uses only information on
patents and finds a negative relation.
Aggregate output displays a hump-shaped response. In the first two years, the response
of output is negative but not statistically significant. However, in the long-run output
increases by a substantial amount: a one-standard deviation innovation shock results to a
1.5% increase in aggregate output after 8 years. This finding is similar to the pattern found
by Basu et al. (2006) and Alexopoulos (2011). The share of forecast-error variance attributed
17
to our innovation measures is 9.6%.
Hours worked display a transitory drop in the short-run following a one-standard positive innovation shock. At the trough, hours worked drop by 1.1%, which is economically
significant given the fact that the annual standard deviation of percentage changes in hours
is 2.5%. This negative response of aggregate hours is somewhat puzzling, but is in line with
the evidence of Basu et al. (2006) and others.
Consumption of non-durable goods and services displays a U-shaped response. In the
first two years it shows a statistically significant drop of 0.45%. Subsequently, consumption
increases, leading to a 0.3% net increase after 8 years. The innovation shock accounts for
8.2% of the forecast-error variance of consumption growth after 8 years.
Investment in equipment and software displays a delayed positive response. In the first
two years it shows a negative, but not statistically significant response. However, it increases
subsequently. At the peak, investment increases by 3.6% following a one-standard deviation
innovation shock.
The price of equipment and software shows a positive permanent response. A onestandard deviation innovation shock increases the relative price of equipment by 0.9% in the
next two years. This response is economically significant, given that the annual standard
deviation of the price of equipment is 3%. However, at longer horizons, standard errors
increase so we cannot reject the null that the response is zero.
Alternative specifications
First, we explore the robustness of our findings to using a vector-error-correction model. The
Johansen test suggests the presence of one cointegrating relation in all systems. Thus, we
estimate tri-variate VECMs as above, and explore the response of aggregate variables to a
one-standard deviation innovation shock. We show the results in Figure 6. Allowing for a
cointegrating vector leads to similar qualitative results, though the quantitative magnitude is
stronger. A one-standard deviation innovation shock now increases total-factor-productivity
by 1.75% and aggregate output by 5% at the 8-year horizon. Hours worked drops by 1%,
though the drop is not statistically significant at longer horizons. Consumption drops by
0.5% in the short-run, but displays a net 0.5% increase after 8 years. Both the price and the
quantity of investment increase. At the peak, investment rises by 7.1%, while the price of
equipment displays a transitory increase of 0.9% over the next two years.
Second, we explore whether our measure of innovation contains incremental information
to stock prices. Following Beaudry and Portier (2006), we include the level of the stock market in our VAR, scaled by the consumption deflator and population. We depart from Beaudry
18
and Portier (2006) in that we include the level of the CRSP value-weighted rather the level of
the S&P 500 index, since the former includes all stocks traded on the three major exchanges.
We order the level of stock prices second, so now Zt = [log(Xt ), log(Mt ), log σt , log(At )]0 . We
find that controlling for the level of the stock market improves the ability of our technology
shock measure to forecast quantity variables, but worsens its ability to forecast productivity,
hours worked, and the relative price of equipment. Output, investment and consumption
increase by 2.6%, 0.9% and 4% at the peak following a one-standard deviation innovation
shock.
Third, one might worry that our measure of innovation uses information only from publicly traded firms and omits information on private firms. In particular, one concern may
be that the ratio of innovation that takes place in public versus private firms fluctuates as a
function of unobservable variables that have an effect of output, for instance credit market
conditions. We address this by including the log ratio of public-to-private Pt patents granted
per year into our VAR. We aggregate the annual patents produced by private firms by identifying patents produced by private firms following the same process as outlined in the data
section. In other words, Zt = [log(Xt ), log(Pt ), log σt , log(At )]0 . We find that doing so leads
to quantitatively similar results, and the response of productivity, output and consumption
has a similar shape as before. Our innovation shock accounts for 12-27% of the forecast
error variance of productivity, while the corresponding shares are 5.5-15.7% for output and
6.1-9.3% for consumption.
Fourth, we explore the sensitivity of our findings to the presence of the deterministic
time-trend; the inclusion of the average firm-level volatility measure; or using a equal- rather
than value-weighted innovation measure. Our results are qualitatively similar across different
specifications, see online appendix for details.
Comparison to Patents or R&D expenses
We explore whether our measure of innovation contributes relative to other commonly employed measures of technological innovation: the stock of R&D capital, and the log number
of patents. We estimate two bi-variate VARs for productivity, output and consumption, with
the log number of patents or R&D capital series ordered last. Note that the plots of these
two measures in Panels (c) and (d) of Figure 2 showed a strong time-trend in both series.
Hence, following Alexopoulos (2011) we include a deterministic time trend in our VARs. We
show the results in Figure 7. Using a simple patent count, or measuring the stock of R&D,
has essentially no ability to predict business cycle variables. The impulse responses of output
and consumption are not statistically significant, and either innovation measure accounts for
19
1.7-3.1% of the variance forecast error decompositions of output or consumption. In terms
of productivity, both measures fare better: they predict a statistically significant increase
in productivity at the 8-yr horizon, but only accounts for 4.7-13.1% of the forecast error
variance.
Granger causality
If our measure of innovation indeed represents a fundamental shock, then it should not be
predictable by output or productivity. The top panel of Table 8 shows that output and TFP
do not Granger-cause either of our measures of innovation.
In addition, we explore whether our measure of innovation is predictable by other measures of technological growth in the literature, for instance the book-based measures of Alexopoulos (2011) and the stock of R&D capital. Contemporaneously, our innovation measure
is uncorrelated with the number of books in Alexopoulos (2011); the correlation of growth
rates is less than 10% using annual data. In addition, the middle panel of Table 8 shows that
our measures of innovation are distinct from the measures of Alexopoulos (2011), in that
neither causes the other. Our measure is somewhat correlated with the number of patents
at 32-43%, but not with R&D spending (less than 10%). The bottom panel of Table 8 shows
that our measure is not Granger-caused by either the number of patents or R&D spending.
D.2
Innovation and long-run growth
In this section we evaluate the relationship between long-run effects and our measures of
innovation. It is difficult to make long-run statements using aggregate time series since the
power of such tests is likely to be low. To circumvent this issue we examine growth response
to innovation across different industries which allows us to make more reliable inferences. In
particular, we use the KLEMS industry-level output data provided by Dale Jorgenson and
investigate the relationship between our aggregate industry level innovation measures and
long-run output growth in these industries.
Before estimating regressions we present the relationship between our measures of innovation and industry growth in Figure 4. In the scatter graph, we plot the industry Aew
and Avw averaged over the first half of the sample (1960-1982) on the X axis and the corresponding output growth of the industry in the second half of the sample (1983-2006) on
the Y axis. As can be observed, on average, industries which experienced high technological
innovation in the first half of the sample were also the ones whose growth rate was subsequently higher in the second half of the sample. For example, industries such as Electrical
Machinery, Automotive and Communication which are in the highest quartile of innovation
20
during the first half of the sample had an annualized growth rate of more than 4% over the
second part of the sample. Similar correlation is found for low innovative industries such as
Textile and Utilities.
We formally test for the relationship shown in Figure 4. In particular, our specification
regresses industry level output growth in a period on last period’s industry level aggregate
innovation – with each period representing 5 years. As can be observed from impulse responses we showed earlier, a period of five years should allow us enough time to capture the
impact of innovation on output. These regressions control for lagged output growth, time
effects, industry stock return and industry volatility across specifications.
The first three columns of Table 9 present the results of OLS with different control
variables. The results across all specifications are very consistent and indicate a strong
positive relationship between aggregate industry-level innovation and subsequent output
growth at low frequencies. In addition, the magnitudes of these results are large: an onestandard deviation increase in the average industry-level innovation in a period leads to an
increase in average output growth by 0.5-1.1% in the next period.
One source of concern is that the relationship between innovation and output growth
is driven by spurious omitted variables – such as unobserved prospects in an industry –
which drive both more innovative activity and higher output. To alleviate these concerns we
generate exogenous changes in R&D activity across industries by employing the Bloom et al.
(2010) instrument for firm level R&D activity. As discussed in Bloom et al. (2010), the firm
level tax price of R&D can be decomposed into a firm level component that is based on solely
on Federal rules. Even though the Federal R&D tax rules may be at an aggregate level, their
effect varies in the cross-section of firms for three reasons. First, the amount of tax credit
that can be claimed is based on the difference between actual R&D and a firm-specific ‘base’
– with the base definition itself changing over time. In addition, if the credit exceeds the
taxable profits of the firm it cannot be fully claimed and must be carried forward. With
discounting this leads to a lower implicit value of the credit for tax exhausted firms. This
generates additional firm level variation. Finally, these firm-specific components all interact
with changes in the aggregate tax credit rate itself.
We use the Bloom et al. (2010) firm level R&D instrument and construct the industry
wide Federal tax component by taking the average of this tax price across firms in a given
industry. We then instrument the endogenous innovation variables (A) and report our results
in the last three columns of Table 9. As we see, the magnitude of the effects are higher when
we employ the IV strategy. A one-standard deviation increase in the average industry-level
innovation in a period leads to a subsequent increase in average output growth by 0.8-1.7%.
We conclude that innovation, as captured by our measures, leads to long-run economic
21
growth.
E
Reallocation
In the last decade, the literature has recognized the importance of resource allocation for
economic growth (Restuccia and Rogerson, 2008; Hsieh and Klenow, 2009; Jones, 2011;
Acemoglu, Akcigit, Bloom, and William, 2011). Given an economy’s stock of productive
inputs, to maximize the economy’s overall level of production, resources need to be allocated
to the most productive firms and industries. If innovation raises a firm’s productivity, as
we saw in Section C.3, then economic optimization implies that productive resources should
flow into the innovating firm away from firms that do not innovate. The same argument also
applies to industries, although specificity of capital or labor inputs can potentially dampen
the amount of reallocation. Since our measure is based on micro data, it places us uniquely
to examine these issues. Furthermore, it allows us to trace the origin of innovation and its
propagation through the economy
In this section we explore the reallocation dynamics within and across industries. This
allows us to shed light on some of the evidence at the aggregate level, such as the negative
response of aggregate hours worked to an innovation shock. In particular, examining firmlevel data allows us to isolate the direct effect of innovation on the marginal product of labor,
from the potential general equilibrium effect of innovation on prices and Hours worked.
E.1
Firm-level evidence
This section explores the reallocation dynamics subsequent to innovation by a firm. In
particular, we employ both the firm and industry level measures of innovation and examine
how these measures are related to reallocation of physical capital, labor and financial capital.
We provide results for both within and across industry reallocation of inputs.
To explore reallocation of inputs x at the firm-level, we estimate the following specification:
xf t = a0 + a1 Zf t−1 + a2 Af t−1 + a3 AIf t−1 + ρ xf t−1 + uf t ,
(8)
where Af t−1 is our measure of innovation at the firm-level, and AIt−1 is the average level of
innovation of firm’s competitors defined in Section C.3. To isolate the effect of innovation
from firm-level characteristics, we include a vector of controls Z that includes firm’s Tobin’s
Q, firm capital stock normalized by average capital in the economy, firm sales normalized
by capital, firm stock returns and volatility. In addition we include industry (I) and/or
time (T) fixed effects and report standard errors that are clustered by firm and year. The
22
main estimates of interest in this specification are a2 and a3 which capture the change in
physical capital of a firm in response to changes in innovation by the firm and its competitors,
respectively.
Physical Capital
We first examine how physical capital gets reallocated subsequent to innovation by a firm
or by its competitors. We measure the firm’s net investment rate It /Kt−1 as the ratio of
capital expenditures (capx) minus sales of property plant and equipment (sppe) to lagged
gross property, plant and equipment (ppegt).
We report results in Table 10. As is evident, we find that subsequent to an innovation
by a firm, there is a substantial increase in its investment, regardless of whether we use the
equal weighted or value weighted measure of innovation. Since the firm-level innovation rate
is substantially skewed, we focus on inter-decile movements in firm-level innovation to explore
economic magnitudes. In particular, our estimates imply that an increase in innovation by
the firm from the 50th to the 90th percentile leads to increase in the firm’s investment rate
by 0.4-0.9%, including dynamic effects. This increase is statistically but also economically
significant relative to median firm investment rate is 10% in our sample.
Furthermore, we find evidence that physical capital flows from firms that do not innovate
to firms that do. If the firm does not innovate but is competitors do, then its investment
rate is substantially lower. A one-standard deviation increase in the level of innovation by
the firm’s competitors leads to a drop in the firm’s investment rate by 1.2-1.6%.
Labor
We now examine reallocation of labor subsequent to innovation by a firm. We measure
change in labor in a firm as ∆ log Nt , where N is the number of employees. As before, the
main estimates of interest in this specification are a2 and a3 , which capture the change in
employment of a firm in response to changes in innovation by the firm and its competitors,
respectively.
We show the results in Table 11. We find that subsequent to an innovation by a firm,
there is a substantial increase in its employment using either innovation measure. As before,
the economic magnitudes are significant. Our estimates imply that an increase in innovation
by the firm from the 50th to the 90th percentile leads to increase in employment of the firm
by 0.1-0.7%, compared to the median firm-level hiring rate of 3%.
In addition, there is a reverse movement of labor when a firm does not innovate but
its competitors in the same industry do. A one-standard deviation increase in the average
23
innovation of the firm’s competitors leads to a reduction of 1 to 3% in the firm’s hiring
rate. This increased reallocation response of labor relative to firm capital within industries
is consistent with the view that capital is more firm-specific than labor.
Financial Capital
Finally, we examine the reallocation of financial capital subsequent to innovation by a firm.
Increasing the quantity of inputs requires resources, so innovating firms may have to seek
external funds. We measure financial capital inflows (Ft /At−1 ) as debt issuance plus equity
issuance minus payout (Compustat sstk + dltis - prstkc-dv-dltr) normalized by assets (at).
We follow a specification similar to (8) and use financial capital inflows as the dependent
variable.
The results are reported in Table 12. As is evident, we find that subsequent to an
innovation by a firm, there is a substantial increase in its financial capital inflows. The
results are obtained regardless of whether we use the equal weighted or value weighted
measure of innovation. Our estimates imply that an increase in innovation by the firm from
the 50th to the 90th percentile leads to an increase of capital inflows to book assets of
0.6-0.7%, compared to the median level of zero capital flows.
Our results also suggest that there is an opposite effect when the firm does not innovate
but is competitors do. In particular, a one-standard deviation increase in the average innovation of the firm’s competitors leads to a reduction of 0.1-0.6% in net capital flows to the
firm. This negative effect suggests that firms that fail to innovate in an industry where other
firms do, have few investment opportunities. Hence, these firms increase payout to investors.
E.2
Industry-level evidence
So far we have focused on reallocation of inputs subsequent to innovation by a firm or its
competitors within an industry. We now conduct a similar exercise examining reallocation of
inputs across different sectors. The KLEMS dataset of Dale Jorgenson contains information
on the quantity and price of capital and labor inputs at the industry level. We estimate
specifications similar to (8) at the industry level:
xIt = a0 + a1 Zt−1 + a2 AIt−1 + a3 AM It−1 + ρ xIt−1 + uIt ,
(9)
where AIt−1 is our measure of innovation at the firm-level, AM It−1 is the average level of
innovation in the economy, excluding industry I, and Z is a vector of controls that includes
industry returns and volatility. The left-hand-side variables are the growth in the quantity
24
[kIt , hIt ] and price [pIt , wIt ] of capital and labor respectively used in industry I. The main
estimates of interest in this specification are a2 and a3 which capture the change in inputs
(physical capital, labor or financial capital) of an industry in response to changes in innovation by the industry and its competitors (all other industries), respectively. As a robustness
test, we also estimate equation (9) using IV, with our instrument described in Section (D.2).
Our results are qualitatively similar using instrumental variables.
We show our results in Table 13. There is some evidence that an increase in the amount of
industry innovation increases the quantity of capital and labor services in the industry. The
response of the quantity of capital is statistically significant, while the results for the quantity
of labor are statistically significant at the 10% level, provided we include time-dummies or
control for lagged industry returns and volatility.
We find evidence suggesting that increases in economy-wide innovation lead to reallocation of labor but not capital. In particular, the quantity of labor employed in the industry
drops sharply following an increase in innovation in industries. In contrast the quantity of
capital shows no response once we control for its lag. In addition, we see that the price of
capital does not respond to the innovation of other industries, whereas the price of labor
increases. These findings echo our results in the previous section and consistent with the
view that capital has higher specificity than labor (Hamermesh and Pfann, 1996; Hall, 2002).
If capital is specific to the industry, the amount of innovating activity in the economy need
not affect its price or quantity employed. In contrast, labor is more mobile. An increase in
innovating activity in other industries increases the outside option of workers, leading to an
increase in the equilibrium wage. We find no statistically significant response of the price of
capital or labor services on the industry’s own innovation shock.
F
Innovation and Tobin’s Q
In the last part of our empirical analysis we analyze the relationship between our measures of
innovation and Tobin’s Q – both at the firm and aggregate level. At the firm level, a successful
innovation should lead to an increase in the firm’s Tobin’s Q as it generates investment
opportunities. However, at the aggregate level, the effect can be theoretically ambiguous
due to the general equilibrium response of the price of capital, or because innovation renders
old capital obsolete as in the model of Laitner and Stolyarov (2003).
In Panel A of Table 14, we explore the response of firm’s Tobin’s Q following a successful
innovation. We regress the change in Tobin’s Q on the firm-level measure of innovation,
controlling for the firm’s idiosyncratic volatility, lagged values of Q and time fixed effects
to account for aggregate movements in innovation (Q) over this period. We find that tech25
nological innovation by a firm leads to an increase in the Q of the innovating firm relative
to the market Q. These results are statistically significant for both the equal and value
weighted measures of innovation. The magnitudes are large: a firm which finds itself in the
90-th percentile in terms of technological innovation experiences an increase in Tobin’s Q by
1.5-4.5%. These magnitudes are in line with those reported in Hall et al. (2005).
Next, we analyze how changes in Tobin’Q at the aggregate level are related to changes
in aggregate innovation. We show the results in Panel B of Table 14. We find that in the
time-series, our aggregate measure of innovation is negatively correlated with Tobin’s Q. In
particular, a one-standard deviation increase in the aggregate amount of innovation leads
to a 4.7-12% drop in aggregate Tobin’s Q. Given that the annual volatility of log changes
in Tobin’s Q is 18%, our findings suggest that technological innovation plays an important
role for aggregate valuations. Our results are in line with the views expressed in Hobijn
and Jovanovic (2001) and Laitner and Stolyarov (2003), who both argue that technological
innovation leads to obsolescence of old capital, and thus may lead to a decrease in Tobin’s
Q.
G
Discussion and Conclusion
We construct novel economic firm and aggregate level measures of innovation exploiting
patent data for US firms from 1926-2007 in conjunction with the stock market response to
news about patents. Our measures capture known periods of high technological progress as
well as firms driving these waves.
We find that higher technological innovation in the economy is followed by an increase
in total-factor productivity. Output increases as a response to innovation, but only after
a delay of 4-5 years. We find that technological innovation has a negative effect on Hours
worked, though the effect is not always statistically significant. A negative response of hours
worked on a positive technology shock is consistent with RBC models with sticky prices (e.g.
Gali (1999) and Basu et al. (2006)). A positive shock to innovation has a U-shaped response
on consumption growth: consumption is lower in the short-run but increases in the long-run.
This is consistent with a reallocation of resources away from consumption in the short-run
towards the implementation of innovation.
The empirical distribution of firm-level innovation measure is extremely fat-tailed – even
more so than the distribution of firm size. Following the arguments in Gabaix (2011), this
is consistent with the view that firm-specific technology shocks in a few large firms could
have substantial effects on the aggregate rate of innovation in the economy. However, the
fact that our results for the effect of innovation on business-cycle quantities do not depend
26
on whether we construct aggregate innovation measures as equal- or value-weighted averages
of firm-level innovation suggests that the innovative activity of a few large firms cannot be
solely responsible for our results.
To better understand the evidence at the aggregate level, we exploit the fact that our
innovation measure is available at the micro-level and trace the effect of innovation at its
origin. We find evidence for within- as well as between-industry reallocation: capital, labor
and financing flows into firms or sectors that innovate, and flows away from parts of the
economy that lag behind in innovative activity. Our findings on movement of labor to
innovating firms are consistent with Lentz and Mortensen (2008) who find similar evidence
in data on Danish firms and rationalize it using a Klette and Kortum (2004) equilibrium
model of growth through innovation extended to allow for firm heterogeneity. The patterns
of reallocation are stronger within- rather than between-industries: we find no movement
of capital across and weak evidence for movement of labor into innovative industries. Both
these results are consistent with specificity of inputs – with capital more specific to an
industry than labor (Ramey and Shapiro, 2001) – dampening the amount of reallocation
across sectors.
In summary, our results suggest that technological innovation accounts for a significant
part of long-run growth, both at the aggregate as well as at the industry level. We find that
industries that innovated in the past subsequently grow faster, especially once we address
concerns that the innovative output is potentially endogenous to industry forces.
Finally, we briefly explore the effect of innovation on asset prices. We find that an
increase in innovative activity leads to a fall in aggregate Tobin’s Q. This is consistent with
Hobijn and Jovanovic (2001) and Laitner and Stolyarov (2003) who argue that technological
innovation leads to obsolescence of old capital, and hence to a decrease in aggregate Tobin’s
Q.
In a companion paper, we rationalize all these findings in a GE model that allows one to
study the effect of innovation on reallocation and growth dynamics in an unified framework.
In the model, technological innovation shocks manifest as a wedge between the productivity
of new versus old capital, and hence trigger a reallocation of resources away from consumption
and into producing new capital goods. As a result, innovation leads to increased capital
accumulation and hence growth in the long-run at the expense of current consumption. In
addition, an increase in the rate of innovation leads to a fall in aggregate Tobin’s Q, because
the market value of existing firms with old production units drops.
27
References
Acemoglu, D. (2009). Introduction to modern economic growth. Princeton University Press.
Acemoglu, D., U. Akcigit, N. Bloom, and K. William (2011). Innovation, reallocation and growth.
Working paper, Stanford University.
Alexopoulos, M. (2011). Read all about it!! what happens following a technology shock? American
Economic Review 101 (4), 1144–79.
Basu, S., J. G. Fernald, and M. S. Kimball (2006). Are technology improvements contractionary?
American Economic Review 96 (5), 1418–1448.
Baumol, W. (2002). The free-market innovation machine: analyzing the growth miracle of capitalism. Princeton University Press.
Beaudry, P. and F. Portier (2006). Stock prices, news, and economic fluctuations. American
Economic Review 96 (4), 1293–1307.
Bloom, N., M. Schankerman, and J. V. Reenen (2010). Identifying technology spillovers and product
market rivalry. Working paper, Stanford University.
Caballero, R. J. and A. B. Jaffe (1993). How high are the giants’ shoulders: An empirical assessment
of knowledge spillovers and creative destruction in a model of economic growth. In NBER
Macroeconomics Annual 1993, Volume 8. National Bureau of Economic Research, Inc.
Campbell, J., A. Lo, and A. MacKinlay (1997). The econometrics of financial markets. Princeton
University Press.
Cochrane, J. H. (1994, December). Shocks. Carnegie-Rochester Conference Series on Public Policy 41 (1), 295–364.
Field, A. J. (2003). The most technologically progressive decade of the century. American Economic
Review 93 (4), 1399–1413.
Gabaix, X. (2011). The granular origins of aggregate fluctuations. Econometrica 79.
Gali, J. (1999). Technology, employment, and the business cycle: Do technology shocks explain
aggregate fluctuations? American Economic Review 89 (1), 249–271.
Greenwood, J. and B. Jovanovic (1999). The IT revolution and the stock market.
Griliches, Z. (1989). Patents: Recent trends and puzzles. Technical report.
28
Griliches, Z. (1990). Patent statistics as economic indicators: A survey. Journal of Economic
Literature 28 (4), 1661–1707.
Hall, B. H., A. B. Jaffe, and M. Trajtenberg (2001). The nber patent citation data file: Lessons,
insights and methodological tools. Technical report.
Hall, B. H., A. B. Jaffe, and M. Trajtenberg (2005). Market value and patent citations. The RAND
Journal of Economics 36 (1), pp. 16–38.
Hall, R. E. (2002, March). Industry dynamics with adjustment costs. NBER Working Papers 8849,
National Bureau of Economic Research, Inc.
Hamermesh, D. S. and G. A. Pfann (1996). Adjustment costs in factor demand. Journal of
Economic Literature 34 (3), pp. 1264–1292.
Harhoff, D., F. Narin, F. M. Scherer, and K. Vopel (1999). Citation frequency and the value of
patented inventions. The Review of Economics and Statistics 81 (3), 511–515.
Harhoff, D., F. M. Scherer, and K. Vopel (1997). Exploring the tail of patented invention value
distributions. Technical report.
Hobijn, B. and B. Jovanovic (2001). The information-technology revolution and the stock market:
Evidence. American Economic Review 91 (5), 1203–1220.
Hsieh, C.-T. and P. J. Klenow (2009). Misallocation and manufacturing tfp in china and india.
The Quarterly Journal of Economics 124 (4), 1403–1448.
Jones, C. I. (2011). Misallocation, economic growth, and input-output economics. Working paper,
Stanford University.
Jorgenson, D. W. and K. J. Stiroh (2000). U.s. economic growth at the industry level. American
Economic Review 90 (2), 161–167.
Klette, T. J. and S. Kortum (2004). Innovating firms and aggregate innovation. Journal of Political
Economy 112 (5), 986–1018.
Kortum, S. (1993). Equilibrium r&d and the patent-r&d ratio: U.s. evidence. American Economic
Review 83 (2), 450–57.
Kortum, S. and J. Lerner (1998, June). Stronger protection or technological revolution: what is
behind the recent surge in patenting?
Carnegie-Rochester Conference Series on Public Pol-
icy 48 (1), 247–304.
Kortum, S. and J. Lerner (2000, Winter). Assessing the contribution of venture capital to innovation. RAND Journal of Economics 31 (4), 674–692.
29
Kydland, F. E. and E. C. Prescott (1982). Time to build and aggregate fluctuations. Econometrica 50 (6), pp. 1345–1370.
Laitner, J. and D. Stolyarov (2003). Technological change and the stock market. The American
Economic Review 93 (4), pp. 1240–1267.
Lentz, R. and D. T. Mortensen (2008, November). An empirical model of growth through product
innovation. Econometrica 76 (6), 1317–1373.
Long, John B., J. and C. I. Plosser (1983). Real business cycles. Journal of Political Economy 91 (1),
pp. 39–69.
Nicholas, T. (2008). Does innovation cause stock market runups? evidence from the great crash.
American Economic Review 98 (4), 1370–96.
Ramey, V. A. and M. D. Shapiro (2001, October). Displaced capital: A study of aerospace plant
closings. Journal of Political Economy 109 (5), 958–992.
Restuccia, D. and R. Rogerson (2008, October). Policy distortions and aggregate productivity with
heterogeneous plants. Review of Economic Dynamics 11 (4), 707–720.
Scherer, F. M. (1965). Firm size, market structure, opportunity, and the output of patented
inventions. The American Economic Review 55 (5), pp. 1097–1125.
Scherer, F. M. (1983, March). The propensity to patent. International Journal of Industrial
Organization 1 (1), 107–128.
Shea, J. (1999). What do technology shocks do? In NBER Macroeconomics Annual 1998, volume
13. National Bureau of Economic Research, Inc.
Smiley, G. (1994). The American economy in the twentieth century. College Division, SouthWestern Pub. Co.
30
Appendix A: Data
Business-cycle data
Productivity is utilization-adjusted TFP from Basu et al. (2006). Populations is from the
U.S. Census Bureau (http://www.census.gov/popest/national/national.html). Output and
consumption are from the Bureau of Economic Analysis. Output is gross domestic product (NIPA Table Table 1.1.5) divided by the consumption price index (St Louis Fed, CPIAUCNS). Consumption is consumption of non-durables plus services, deflated by the price
index of non-nudrables and services respectively (NIPA Tables 1.1.5, 2.3.4). To get hours
worked, we merge series CEU0500000007 and EEU00500005 from the BLS, times total private employment(BLS, CEU0500000001) divided by population.
Firm-level data
We define the investment rate as capital expenditures (Compustat: capx) minus sales of
property, plant and equipment (sppe) divided by lagged gross property, plant and equipment
(ppegt); the return on assets as operating income (ib) plus depreciation divided by lagged
lagged gross property, plant and equipment; Tobin’s Q as the sum of the market value of
common equity (CRSP December market capitalization), the book value of debt (dltt), the
book value of preferred stock (pstkrv), minus the book value of inventories (invt) and deferred
taxes (txdb), divided by gross property, plant and equipment (ppegt); and book-to-market
as the ratio of the market value of common equity divided by the book value of common
equity (ceq). All variables are winsorized by year at the 1% level.
Other Data
The industry-level data is from the KLEMS dataset of Dale Jorgenson. We use industry
value added (constant prices) as measure of industry output. The aggregate Tobin’s Q
information comes from Flow of Funds Z1 report, Table B.102. We construct Tobin’s Q as
the ratio of line 35 “Market Value of Equities Outstanding” to line 32 “Net Worth (Market
Value)”. Finally, the time series information on R&D expenditure spending in the US is
obtained from the NSF website.
31
Tables
Table 1: Number of patents
Data step
Number of patents
Total downloaded patents
7,797,506
Granted in 1926 or later
5,988,864
Identified as having an assignee
4,386,506
Matched to CRSP
1,840,636
Of which:
Present in NBER data
1,404,822
New to this paper
435,814
The table provides details on patents in our sample. We begin with all patents downloaded from Google
Patents, and restrict the sample to post-1926. Not all patents have assignees, and among those that do, not
all are companies in CRSP. We are able to match 1,840,636 patents to CRSP firms, of which 435,814 (24%)
are new to this study. Further details are reported in Table 2 and Figure 1.
32
33
Total
173,809
442,199
306,827
425,066
561,300
636,097
708,744
1,109,399
1,625,423
5,988,864
Number of patents
With assignee Matched to CRSP
63,958
8,858
215,613
47,036
167,037
60,571
217,341
82,245
327,754
163,851
419,033
236,190
579,098
235,532
933,428
352,005
1,463,244
654,345
4,386,506
1,840,633
Number of unique
Matched firms CRSP firms
180
786
351
951
446
1,042
585
1,246
1,166
3,177
2,081
7,204
2,729
11,715
3,577
14,882
1,078
11,900
8,729
26,660
Backward citations
>1
Average
72,575
1.1
235,268
1.6
204,908
2.3
298,767
2.9
430,916
4.1
576,699
8.3
670,862
12.0
1,037,054
13.3
937,052
4.1
4,464,101
6.7
The shows summary statistics for patents in our sample by decade. Column 2 shows the total number of patents, and column 3 shows how many
patents are identified as having an assignee. Column 4 shows how many of those patents with assignees are matched to a company in CRSP. (The
remaining assignees are either individuals, private companies, or the matching process was unable to identify the correct company.) Columns 5 and
6 show how many unique firms there are matched to patents or in CRSP. The final two columns report the number of patents with more than one
backward citation, and the average number of backward citations.
Decade
1926–1929
1930s
1940s
1950s
1960s
1970s
1980s
1990s
2000s
All years
Table 2: Assignee Matching by Decade
Table 3: Stock turnover around patent grant days
Turnover
(i)
(ii)
(iii)
xt−1
xt
xt+1
xt+2
xt+3
xt+4
-0.333
0.059
0.198
0.097
-0.025
-0.350
(-8.21)
(4.37)
(8.92)
(5.50)
(-1.86)
(-9.38)
-0.232
0.137
0.273
0.186
0.083
-0.198
(-9.60)
(5.18)
(6.55)
(4.97)
(2.61)
(-7.22)
-0.022
0.119
0.196
0.184
0.161
0.051
(-2.39)
(5.95)
(6.18)
(5.76)
(5.87)
(3.05)
Controls
F ×M
x̄t−1
x̄t−1 , day
Table shows the output of the regression of stock return turnover (xt = volt /shroutt ) on a dummy variable
taking the value 1 if a patent was granted to firm i on day t. We estimate three specifications: in (i) we
include firm-month fixed effects; in (ii) we control for the lagged average stock turnover over the last 5 days,
P5
x̄t−1 ≡ 15 l=1 xt−1−l ; in (iii) we control for x̄t−1 and calendar date fixed effects. We cluster standard errors
by year and report t-statistics in parenthesis. We restrict the sample to firms that have been granted at
least one patent.
34
Table 4: Number of future citations and announcement day return
Number of
citations
rf d
B: Patents with positive
A: All patents
grant day returns only
0.174
(0.14)
rf d × Sf d−1
0.001
(1.21)
[rf d ]+
18.459
7.063
28.910
11.053
(4.33)
(3.10)
(4.54)
(3.45)
[rf d ]+ × Sf d−1
R2
N
Control (log σf t−1 )
0.113
0.113
0.114
1690017
1690017
1690017
-
-
-
1.212
1.581
(6.95)
(6.11)
0.114
0.117
0.114
1690017 1690017
1690017
825126
Y
-
0.116
Y
-
0.116
1.342
1.762
(7.15)
(6.17)
0.114
0.118
825126 825126 825126
Y
-
Table shows output of a regressions of number of future citations on the firm’s idiosyncratic return (rf d ),
defined as the return of the firm minus the return on the market portfolio, on the day the patent is issued.
We interact the firm’s idiosyncratic return with the firm’s market capitalization in the previous day Sf d−1
in 1982 dollars (USb) to obtain a measure of dollar idiosyncratic return. All specifications include year-fixed
effects. Depending on the specification we control for log firm idiosyncratic volatility (log σf t−1 ) in previous
year. Firm idiosyncratic volatility in year t is computed as the standard deviation of daily idiosyncratic
returns in year t. We cluster standard errors by year and report t-statistics in parenthesis.
35
Y
Table 5: Number of future citations and announcement day return, longer event window
[
P2
[
P2
+
l=0 rf d+l ]
+
l=0 rf d+l ]
[
P4
[
P4
l=−1 rf d+l ]
l=−1 rf d+l ]
12.608
5.642
(5.08)
(4.11)
× Sf d−1
0.643
0.839
(6.33)
(5.98)
+
+
9.584
4.355
(4.74)
(3.67)
× Sf d−2
N
R2
Control (log σf t−1 )
1690017
1690017 1690017
1690017 1690017
1690017
0.488
0.637
(5.50)
(5.48)
1690017 1690017
0.113
0.116
0.114
0.117
0.113
0.116
0.114
0.117
-
Y
-
Y
-
Y
-
Y
Table shows output of a regressions of number of future citations on the firm’s idiosyncratic return (rf d ),
defined as the return of the firm minus the return on the market portfolio, on the day the patent is issued.
We interact the firm’s idiosyncratic return with the firm’s market capitalization in the previous day (Sf d−1 )
in 1982 dollars (USb) to obtain a measure of dollar idiosyncratic price change. All specifications include
year-fixed effects. Depending on the specification we control for log firm idiosyncratic volatility (log σf t ).
We cluster standard errors by year and report t-statistics in parenthesis.
36
Table 6: Which firms innovate?
Sample
log Qt−1
log Kt−1 /K̄t−1
rd
/At−1
log Kt−1
(1975-2006)
0.0172
0.0208
0.0169
0.0208
0.0343
0.0218
0.0278
0.0220
(11.09)
(12.73)
(11.03)
(12.72)
(22.71)
(19.28)
(21.34)
(19.35)
0.0516
0.0534
0.0511
0.0534
0.0399
0.0487
0.0395
0.0490
(43.65)
(45.06)
(44.55)
(44.87)
(32.83)
(48.95)
(37.87)
(49.00)
0.0479
0.0423
0.0452
0.0423
(28.45)
(22.50)
(26.67)
(22.46)
0.1021
0.0355
0.5308
0.0595
(4.34)
(1.87)
(68.99)
(4.08)
AI,t−1
N
Fixed Effects
(1950-2006)
35706
35706
35621
35621
135282
135282
134026
134026
T
I, T
T
I, T
T
I, T
T
I, T
Tobit regressions, t-statistics in parenthesis. SE clustered by firm. Kf t is gross PPE, Af t is book assets,
Kfrdt is accumulated R&D expenditures. Specifications include year (T) and/or industry (I) fixed effects.
37
Table 7: Firm-level productivity
yf t
Sales-to-Capital
Sales-to-Employees
(1)
(2)
(3)
(4)
(5)
(6)
-0.215
-0.128
-0.115
-0.047
-0.009
-0.090
(-17.03)
(-9.61)
(-8.01)
(-3.87)
(-0.68)
(-6.68)
0.088
0.134
0.127
0.119
0.123
0.170
(5.54)
(8.05)
(7.65)
(9.15)
(9.13)
(11.51)
134987
134987
134987
126941
126941
126941
0.835
0.837
0.839
0.824
0.823
0.834
Fixed Effects
I
I
I,T
I
I
I,T
(Size )
Y
Y
Y
Y
Y
(R, σ, Q)
-
Y
Y
-
Y
Y
(yf t−1 )
Y
Y
Y
Y
Y
Y
AIt−1
Af t−1
Observations
R2
Table shows output of the regression:
mpkf t = a0 + a1 Xt + a2 Af t−1 + a3 AIt−1 + uit .
Here yf t refers to log sales (Compustat item sale) + change in inventories (invt) over capital (Columns 1-3)
or number of employees (Columns 4-6). Depending on the specification, we control for lagged values of firm
size (log capital or number of employees), firm stock return (R), firm volatility (σ), industry (I) or time
(T) fixed effects, and lagged values of the dependent variable. Standard errors are clustered by firm. All
variables are winsorized by year at the 1% level.
38
Table 8: Granger causality tests
Variable does
A does
not Granger
not Granger
cause A
cause variable
Productivity
0.150
0.043
Output
0.916
0.027
Bowkers technology books
0.217
0.141
Library of Congress new technology books
0.231
0.813
Computer software and hardware books
0.469
0.590
Computer software, hardware, and network books
0.473
0.630
Telecommunications books
0.547
0.286
0.655
0.136
Variable
Output and productivity
Technology measures of Alexopoulos (2011)
Other technology measures
R&D Spending
Table features p-values of Granger causality tests, based on a 3-variable VAR [Xt , σt , At ] with a deterministic
trend.
39
Table 9: Long-run Industry-level regressions
∆yt
At−1
5-year avg output growth
0.078
0.076
0.045
0.119
0.101
0.066
(2.68)
(2.55)
(2.63)
(2.42)
(2.31)
(2.10)
0.292
0.296
0.391
0.238
0.270
0.371
Observations
264
264
264
174
174
174
Estimation
OLS
OLS
OLS
IV
IV
IV
I , σI )
(Rt−1
t−1
-
Y
Y
-
Y
Y
(∆yt−1 )
-
-
Y
-
-
Y
R2
Table reports results from a regression of log output y on amount of innovation [AI ] at the industry level,
I
controlling for lagged output growth, industry stock return RI and industry volatility σt−1
. Data is from
Dale Jorgenson’s 35-sector KLEM, described in Jorgenson and Stiroh (2000). Output is value added in
constant prices. Sample is 1960-2007 and covers 33 industries after excluding the finance and government
enterprises sector. Here period refers to 5 years, variables are period averages. We report t-statistics in
parenthesis, with errors clustered by industry. All specifications include time-fixed effects.
40
Table 10: Firm-level reallocation: Investment
it
(1)
(2)
(3)
(4)
(5)
-0.094
-0.043
-0.050
-0.037
-0.037
(-16.38)
(-7.45)
(-10.88)
(-5.83)
(-7.02)
0.052
0.034
0.040
0.045
0.047
(8.41)
(5.63)
(8.30)
(7.43)
(9.69)
0.065
0.143
0.250
0.168
0.262
Observations
140943
140943
140943
140943
140943
Fixed Effects
I
I
I
I,T
I,T
(Size - K)
Y
Y
Y
Y
Y
(R, σ, Q, Y /K, E/K)
-
Y
Y
Y
Y
(it−1 )
-
-
Y
-
Y
AIt−1
Af t−1
R2
Table shows output of the regression:
it = a0 + a1 Zf t−1 + a2 Af t−1 + a3 AIt−1 + uit .
Here it refers to capital expenditures (Compustat item capx) minus sale of property, plant and equipment
(Compustat item sppe) over lagged capital stock (Compustat ppegt). Depending on the specification, we
control for lagged values of Tobin’s Q, firm size (log capital), sales-to-capital (Y /K), profitability (E/A),
firm stock return (R), firm volatility (σ), industry (I) or time (T) fixed effects, and lagged values of the
dependent variable . Standard errors are clustered by firm. All variables are winsorized by year at the 1%
level.
41
Table 11: Firm-level reallocation: Labor hiring
∆nt
(1)
(2)
(3)
(4)
(5)
AIt−1
-0.154
-0.093
-0.095
-0.050
-0.050
(-19.59)
(-11.61)
(-12.14)
(-5.58)
(-5.67)
0.012
0.039
0.048
0.050
0.057
(1.42)
(4.80)
(6.22)
(6.17)
(7.33)
0.028
0.064
0.073
0.079
0.086
Observations
133136
133136
133136
135500
133136
Fixed Effects
I
I
I
I,T
I,T
(Size - N )
Y
Y
Y
Y
Y
(R, σ, Y /N, E/A)
-
Y
Y
Y
Y
(∆nt−1 )
-
-
Y
-
Y
Af t−1
R2
Table shows output of the regression:
∆nt = a0 + a1 Zf t−1 + a2 Af t−1 + a3 AIt−1 + uit .
Here ∆nt refers to log employment growth (Compustat item emp). Depending on the specification, we
control for lagged values of firm size (log no. of employees), sales-to-employees (Y /N ), profitability (E/A),
firm stock return (R), firm volatility (σ), industry (I) or time (T) fixed effects, and lagged values of the
dependent variable. Standard errors are clustered by firm. All variables are winsorized by year at the 1%
level.
42
Table 12: Firm-level reallocation: Financial inflows
∆f inf t
(1)
(2)
(3)
(4)
-0.019
-0.023
-0.032
-0.010
-0.015
(-2.91)
(-3.74)
(-5.66)
(-1.45)
(-2.42)
0.062
0.053
0.050
0.063
0.059
(7.19)
(6.31)
(6.69)
(7.56)
(7.95)
0.097
0.153
0.173
0.166
0.184
Observations
148565
148565
148565
148565
148565
Fixed Effects
I
I
I
I,T
I,T
(Size - K)
Y
Y
Y
Y
Y
(R, σ, E/A)
-
Y
Y
Y
Y
Controls (∆f inf t−1 )
-
-
Y
-
Y
AIt−1
Af t−1
R2
Table shows output of the regression:
f inf t = a0 + a1 Zf t−1 + a2 Af t−1 + a3 AIt−1 + uit .
Here f inf t refers to the ratio of total financial inflows, defined as debt issuance plus equity issuance minus
payout (Compustat sstk + dltis - prstkc-dv-dltr), over book assets. Depending on the specification, we
control for lagged values of firm size (log capital), profitability (E/A), firm stock return (R), firm volatility
(σ), industry (I) or time (T) fixed effects, and lagged values of the dependent variable. Standard errors are
clustered by firm. All variables are winsorized by year at the 1% level.
43
Table 13: Industry reallocation
Quantity of capital services
AIt−1
AM −It−1
R2
Quantity of labor services
0.027
0.024
0.034
0.028
0.024
0.034
0.035
0.041
(2.25)
(2.05)
(3.26)
(2.47)
(1.19)
(1.69)
(1.80)
(2.19)
-0.088
-0.025
-0.327
-0.211
(-4.16)
(-1.19)
(-6.70)
(-3.73)
0.017
0.093
0.045
0.073
0.164
0.186
0.167
0.186
Price of capital services
AIt−1
AM −It−1
Price of labor services
-0.043
-0.042
-0.046
-0.044
0.009
0.009
0.009
0.007
(-1.41)
(-1.23)
(-1.62)
(-1.41)
(0.81)
(1.06)
(1.22)
(1.01)
-0.070
-0.016
0.102
0.099
(4.77)
(4.65)
(-1.02)
(-0.21)
R2
0.003
0.031
0.116
0.148
0.031
0.178
0.461
0.464
Observations
1395
1363
1395
1363
1395
1363
1395
1363
I , σ I , ∆y
(Rt−1
t−1 )
t−1
-
Y
-
Y
-
Y
-
Y
Time Effects
-
-
Y
Y
-
-
Y
Y
Table reports results from a regression of the quantity [k, h] and price [p, w] of capital and labor services on
the amount of innovation at the industry level [AI ] and on the amount of innovation of all other industries
I
[AM −I ]. We control for controlling for time effects (T), industry stock return RI , industry volatility σt−1
and
one lag of the dependent variable. Data is from Dale Jorgenson’s 35-sector KLEM, described in Jorgenson
and Stiroh (2000). Sample is 1960-2007 and covers 33 industries after excluding the finance and government
enterprises sector. We report t-statistics in parenthesis, with Newey-West errors.
44
Table 14: Innovation and Tobin’s Q
∆ log Qf t
Af t
R2
A. Firm level
∆ log Qt
0.010
0.410
0.338
(8.08)
(14.49)
(12.45)
∆ log (At )
R2
B. Aggregate level
-0.155
-0.223
-0.240
(-3.47)
(-4.68)
(-4.20)
0.182
0.345
0.348
0.000
0.088
0.098
(log Af t−1 )
-
Y
Y
(log At−1 )
-
Y
Y
(log Qf t−1 )
-
Y
Y
(log Qt−1 )
-
Y
Y
(∆ log σf t )
-
-
Y
(∆ log σt )
-
-
Y
Table shows output of a regression of changes in Tobin’s Q at the firm (Panel A) or aggregate level (B) on
our measure of innovation, controlling for changes in volatility and lagged. Sample is 1952-2008. Results in
Panel A use data from Compustat, include time-fixed effects and standard errors clustered at the firm level.
Results in Panel B use data from the flow of funds, and standard errors are computed using the Newey-West
estimator.
45
46
The figure shows the number of patents matched to CRSP firms by year of patent grant. Light shading denotes patents included in the NBER patent
data set, while dark shading denotes patents that are new in our paper.
Figure 1: Number of Patents with Matched Assignees
47
log At
−0.5
1920
−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
−1
1920
−0.5
0
0.5
1
1960
year
1960
year
c) R&D Stock
1940
a) Innovation At
1940
1980
1980
2000
2000
1960
year
1980
1940
1960
year
1980
b) Number of patents
1940
2000
2000
d) New technology books (Library of Congress)
−0.3
1920
−0.2
−0.1
0
0.1
0.2
−0.8
1920
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
Figure plots log values (minus a linear trend) of a) our measure of innovation At ; b) Total number of patents granted; c) R&D Capital stock (from
BEA); d) Number of new technology books in the Library of Congress (from Alexopoulos (2011)).
log R & Dt
Figure 2: Aggregate measures of innovation
log Nt
log Bt
Figure 3: Cross-sectional distribution of dollar value of innovation vs firm size
0.2
0.15
0.1
0.05
0
−25
−20
−15
−10
−5
0
Figure plots kernel density plots for the unconditional cross-sectional distribution of log firm market capP
P
italization log(Vit / i Vit ) (dotted line) and the dollar value of innovation log(Mit / i Mit ) (solid line).
P
+
The dollar value of innovation for firm i in year t is defined as Mit =
× If d . We use the
d∈t [∆Sf d ]
Epanechnikov kernel and 300 points to estimate the probability density.
48
Figure 4: Innovation and Industry Growth
0.08
Elec. Mach
Mach
∆ ln Y (1983−2006)
0.06
Auto
0.04
0.02
0
Finance
Furniture
Rubber
Trade
Trans
Mine
Stone
Const
MetalMine
Agric
Misc
Manuf
Fabr
FoodWoodMetal
Elec
CoalUtilities
Paper
Print
Bus. Services
Comm
Instr
Trans. Equip.
Chem
Refine
Txtile
Oil
−0.02
−0.04
0
Smoke
Clothes
Gas Utilities
Leather
0.05
A
vw
0.1
0.15
(1960−1982)
0.2
Figure plots the average output growth rate of 34 industries during the 1983-2006 period, versus the amount
of innovation in 1960-1982. Data is from Dale Jorgenson’s website. Output is measured as value added in
constant prices.
49
Figure 5: Impulse responses, Benchmark
−3
8
x 10
0.035
0.03
7
0.025
6
0.02
5
0.015
4
0.01
3
0.005
0
2
−0.005
1
−0.01
0
0
1
2
3
4
5
6
7
−0.015
0
8
1
2
3
Years
4
5
6
7
8
7
8
7
8
Years
a) Productivity, utilization adjusted
b) Output
−3
8
x 10
0
6
−0.002
4
−0.004
−0.006
2
−0.008
0
−0.01
−2
−0.012
−4
−0.014
−6
−0.016
−0.018
0
1
2
3
4
5
6
7
−8
0
8
1
2
3
4
5
6
Years
Years
c) Hours worked
d) Consumption
0.06
0.025
0.05
0.02
0.04
0.03
0.015
0.02
0.01
0.01
0
0.005
−0.01
−0.02
0
−0.03
−0.04
0
1
2
3
4
5
6
7
−0.005
0
8
Years
e) Investment in Equipment& Software
1
2
3
4
5
6
Years
f) Relative price, Equipment& Software
Figure shows impulse response of productivity and output from a 3-variable VAR [Xt , σit , At ] with a deterministic trend. We use our measure of innovation constructed using the grant-day to grant-day plus two days
stock price response. Dotted lines represent 90% confidence intervals using standard errors are computed
using 500 bootstrap simulations.
50
Figure 6: Impulse responses, Error Correction Model
0.1
0.035
0.08
0.03
0.025
0.06
0.02
0.04
0.015
0.01
0.02
0.005
0
0
−0.005
0
1
2
3
4
5
6
7
−0.02
0
8
1
2
3
a) Productivity, utilization adjusted
0.015
0
0.01
−0.005
0.005
−0.01
0
−0.015
−0.005
−0.02
−0.01
1
2
3
4
5
6
7
8
5
6
7
8
6
7
8
b) Output
0.005
−0.025
0
4
Years
Years
5
6
7
−0.015
0
8
1
2
3
Years
4
Years
c) Hours worked
d) Consumption
0.12
0.02
0.1
0.01
0.08
0
0.06
0.04
−0.01
0.02
−0.02
0
−0.03
−0.02
−0.04
0
1
2
3
4
5
6
7
−0.04
0
8
Years
e) Investment in Equipment& Software
1
2
3
4
5
Years
f) Relative price, Equipment& Software
Figure shows impulse response of productivity and output from a 3-variable VECM [Xt , σit , At ] with a
deterministic trend and one cointegrating relation. We use our measure of innovation constructed using the
grant-day to grant-day plus two days stock price response. Dotted lines represent 90% confidence intervals
using standard errors are computed using 500 bootstrap simulations.
51
Figure 7: Number of Patents and R&D Stock
−3
9
x 10
−3
8
x 10
8
6
7
6
4
5
2
4
3
0
2
−2
1
0
0
1
2
3
4
5
6
7
−4
0
8
1
2
3
Years
4
5
6
7
8
7
8
7
8
Years
No. patents - Productivity
R&D Stock - Productivity
−3
6
x 10
0.02
4
0.015
0.01
2
0.005
0
0
−2
−0.005
−4
−0.01
−6
−0.015
−0.02
0
1
2
3
4
5
6
7
−8
0
8
1
2
3
No. patents - Output
5
6
R&D Stock - Output
−3
4
4
Years
Years
−3
x 10
3
x 10
2
2
1
0
0
−2
−1
−2
−4
−3
−6
−8
0
−4
1
2
3
4
5
6
7
−5
0
8
Years
1
2
3
4
5
6
Years
No. patents - Consumption
R&D Stock - Consumption
Left panel shows impulse response of productivity, output and consumption from a bi-variate VAR [Xt , Nt ]
with a deterministic trend. Right panel shows impulse response of productivity, output and consumption
from a bi-variate VAR [Xt , RDt ] with a deterministic trend. Dotted lines represent 90% confidence intervals.
52