Online Appendix to “Technological Innovation,

Resource Allocation and Growth”

Leonid Kogan, Dimitris Papanikolaou,
Amit Seru and Noah Stoffman

A Patent 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 the data that we employ in our
analysis.
Patents in the United States are granted by the United States Patent and Trademark Office
(USPTO). We download the entire history of U.S. patent documents from Google Patents.1 Each of
about 7.8 million patent files was downloaded using an automation script.2
To construct our measure of innovation, we match all patents in the Google data to corporations
whose returns are in the CRSP database. Patent regulations require that only an individual, not a
corporation, can 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”, that is, one or more corporations or persons.
For most patents, Google provides a text version of the patent document, created using OCR
software. We use this text version of the document to extract the names of corporations to
which patents are assigned. However, OCR technology 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.
Our sample covers patents granted between 1926 and 2010 matched to firms with returns in
CRSP database. Since we merge our patent data with data on stock returns, we are limited to the
period after 1926, when the CRSP database begins.
1
[Link]
2
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.
Matching patents to firms

Here, we briefly discuss the steps our matching procedure followed, and provide extensive details
Section B. 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 is 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 distance, to a much more common name, we substitute the common
name for the uncommon name.3 For example, one of the most common names is “General Electric
Company”, which is associated with over 43,000 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 permco. This is 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 permcos are matched to patents in the NBER
database. For example, all of the patents with an assignee name “General Electric Company” are
matched to one permco 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 permco, even for patents
not included in the NBER data. Remaining assignee names are matched to CRSP firm names using
a name matching algorithm.4 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.

Extracting patent citations

We extract patent citations from three sources. First, all citations for patents granted between 1976
and 2011 are contained in text files available for bulk downloading from Google. These citations
are simple to extract and likely to be free of errors, as they are official USPTO data. Second, for
patents granted before 1976, we extract citations from the OCR text generated from the patent files.
We search the text of each patent for any 6- or 7-digit numbers, which could be patent numbers.
We then check if these potential patent numbers are followed closely by the corresponding grant
date for that patent; if the correct date appears, then we can be certain that we have identified
a patent citation. Since we require the date to appear near any potential patent number, it is
unlikely that we would incorrectly record a patent citation – it is far more likely that we would fail
3
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.
4
The algorithm is based on code written by Jim Bessen, available at [Link]

2
to record a citation than record one that isn’t there. Third, we complement our citation data with
the hand-collected reference data of Nicholas (2008). See Section B of this Appendix for a detailed
explanation of this process.

Summary statistics

We now provide some statistics that lend credence to our method for extracting patent information.
Table A.1 shows the number of patents we match to companies. Of the 6.2 million patents granted
in or after 1926, we find the presence of an assignee in 4.4 million. The matching procedure provides
us with a database of 1.9 million matched patents, of which 523,301 (27%) are not included in the
NBER data. Figure A.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.5
Table A.2 provides additional summary statistics. Overall, our data provides a matched permco
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 47,036 matches
during this period.

5
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 mostly the same, by construction. An exception is for cases where there is apparently
a mistake in the NBER match and our patent-assignee frequency-based matching system corrects an error.

3
B Patent Data Construction – Details
In this section we explain in detail how we constructed our new patent data set. The raw data are
very large and not very well structured, and thus required a great deal of effort to clean. We used a
number of techniques to extract, clean, and match assignees from patents. As with any such project
there is a trade-off between type-I and type-II errors (in this case, failing to match an assignee to
CRSP or incorrectly matching an assignee to CRSP). Our approach was to be as conservative as
possible, attempting to minimize mismatches while at the same time extracting as many correct
matches as possible.

B.1 Data sources

We use three sources of data to construct the new patent database:

1. Details of patents granted from 1976–2010 is available in high-quality text files available for
bulk downloading from Google, through a special data-hosting arrangement with the United
States Patent and Trademark Office (USPTO). The text files use one of two data structures
that allows relatively straightforward data extraction: files for 2001–present use XML, while
files for 1976–2000 use a fixed-width data structure with labeled fields.
2. Patents granted prior to 1976 are also stored on Google, but only in individual web pages
(one per patent). Information during this period is drawn from Optical Character Recognition
(OCR) of original patent documents, and is of highly-variable quality. There is very limited, if
any, structure to these files.
3. We use the NBER patent data (Hall and Trajtenberg, 2001), which covers the period 1976–2006,
to help with the matching and to validate our other data extraction methods.

Due to varying data sources and quality over time, it worth stressing that from 1976–2010 we
use the official records of the USPTO. As we discuss below, we are able to provide some additions
and corrections to the NBER data during the period of overlap with our data. Prior to 1976 the
data are more difficult to work with, but we have implemented a number of sophisticated text
analysis algorithms to create a very high-quality database.

Downloading individual patent files

We downloaded individual patent data from Google. The URL for each patent’s summary page is of
the form [Link] where RD0y is a 4-character code
used by Google to identify each patent. The IDs use any of the characters {a, . . . , z, A, . . . , Z, 0,
. . . , 9, , -}. There are 644 = 16.8 million possible IDs, but only about 8 million patents. However,
all 16.8 million URLs must be checked, because there is no publicly-available mapping of patent
numbers to the Google ID.

4
 A screen shot of the summary page for the patent with id RD0y, which is patent 4,345,262, is
shown in Figure 1. The main page includes—when available—the title of the patent, the filing and
grant dates, the abstract, inventor(s), original assignee(s), current classifications, and a record of
citations (out-cites) and references (in-cites). The information reported on this page by Google was
gleaned from the OCR analysis of the original patent document, and consequently less information
is reported for older documents, especially patents granted before 1976.

Figure 1: Google summary page for U.S. patent 4,345,262

Using a Perl automation script, we sequentially navigated to each of the 16.8 million patent
summary pages.6 From this page, we stored all available information. The script then loaded the
“Read this patent” link, which loads a PDF version of the patent document. From here, we loaded
the “plain text” version of the document, which is simply the text derived from OCR of the PDF
document. Examples of these pages are shown in Figures 2 and 3. We saved the complete text
of the plain text version of each patent. After compression, the complete archive of text requires
approximately 56 gigabytes of disk space.
6
Google generally blocks users from downloading so many web pages. We are grateful to Hal Varian for his
assistance with arranging permission to access these pages.

5
 Figure 2: PDF view

Figure 3: Plain text view

Download bulk patent files

As part of a special arrangement with the USPTO, Google also makes available for downloading bulk
patent data files. 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 OCR procedures implemented by Google is better
in the individual files than in the bulk files provided by the USPTO. We therefore do not use the
bulk download files for data in the pre-NBER period.
For the post-NBER period, however, the bulk data files are of extremely high quality because

6
they are based on digital patent records as opposed to OCR data drawn from images of patent
documents. These data files are provided either in XML format or in a fixed-width record format.
In both cases, all fields (inventor name, grant date, etc.) are clearly identified. We rely on these
files to construct the database during the post-NBER period (2006–2009) and to make additions
and corrections to the NBER data.

B.2 Identifying assignees

Extracting assignee names

For data during the post-1976 period, we can use the XML files available for bulk download to
identify the assignee with virtually no errors.
During the pre-1976 period, we cannot rely solely on Google’s extraction of the filing and grant
dates or the assignee name because the OCR for patents frequently has errors. As an example,
consider patent 1,131,249, shown in Figure 4.

Figure 4: Title page of patent 1,131,249

It is clear to a human reader that this patent was assigned to the Allis-Chalmers Manufacturing
Company, but the OCR for this patent reads

EASLS B. KNIGHT, OF NORWOOD, OHIO, ASSIGNOR,, BY MESH’S ASSIGN1IBNTS, TO ALUSCHALME&S

MANOTAC/rURING- COMPANY, A COBPOBAT’lOH OF DELAY/ABE.

Consequently, Google records the assignee as “BY MESH S ASSIGNIIBNTS”, which is clearly not
accurate.
We therefore rely on a number of textual analysis algorithms to extract the assignee name from
the full text files we saved for each patent. In general, our approach to performing a “fuzzy” match
on a text string is to use the maximum likelihood n-gram approach described by Norvig (2009).

7
 We begin by identifying the text where the assignee, if there is one, will be named. We do this
by searching for words that appear similar to “assign”, “assignor”, or “assignee”. When found near
the beginning of the patent document, this word is typically followed closely by the name of the
assignee, so we extract a text string of 200 characters for further processing. The assignee may be a
person, or a corporation, in which case the name will include a word like “company”, “corporation”
or “incorporated”. If the word “assign” and its variants are not found, we assume the inventor did
not assign the patent to another entity.

Cleaning assignee names

After extracting the string that is likely to contain the assignee name, additional cleaning is necessary.
Because of OCR errors, company names may be garbled. For example, the General Electric Company,
which has more than 43,000 patents in our data, appears as “General Electbic Oohpany”, “General
Electbic Cqhpany”, and “Genebal Electbic Compakt”, among hundreds of other misspellings. To
fix these, we first count how many patents have been granted to each assignee name, regardless of
how the assignee name is spelled. In this example, General Electric Company appears in 42,693
patents, while each of the misspelled variants appears fewer than 5 times.
We then calculate the Levenshtein edit distance7 between each assignee name and all other
names that have more patents. If any assignee name is close to another assignee name that is
associated with many more patents, then the more common assignee name is substituted for the
less common name. This algorithm correctly identifies all of the misspellings noted above as being
General Electric.
After cleaning assignee names, we manually checked which misspelled names were matched to
the 500 assignees with the most patents to confirm that no significant errors were introduced in this
step.

Matching to CRSP

Having extract a list of assignee names, the next step is to match company names to the CRSP
permco identifier. This is accomplished in three steps.
We begin by looking only at those patents that are included in the NBER patent database.
7
The Levenshtein distance is the number of edits required to transform one string into another string, where
allowed edits are inserting, deleting, or substituting one character. For example, the Levenshtein distance between
“patent” and “parent” is 1, while the distance between “patent” and “apparent” is 3.

8
For each assignee name identified in the steps above, we count how many different permcos are
matched to patents in the NBER database. For example, all of the patents with an assignee name
“General Electric Company” are matched to one permco 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 permco, even for patents not included in the NBER data. This step allows us to draw on the
extensive data cleaning and matching project undertaken by Hall and Trajtenberg (2001) while at
the same time identifying some errors in the NBER database. For example, patent 4,994,660 was
assigned to General Electric but is identified in the NBER data as being assigned to Hitachi, Ltd.
Because our algorithm relies on name matching, and the assignee name in that patent is General
Electric, the patent is correctly identified in our data.
The first step only helps us match assignees with patenting activity during the period covered by
the NBER database. We therefore proceed with a second step to match remaining assignee names.
We do this with a name matching algorithm based on code written by Jim Bessen, available at
[Link] 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.
Finally, we identify the top 250 assignees (by patents) with no CRSP matches. We manually
matched these to CRSP whenever possible. Examples of firms requiring manual matching include
research subsidiaries such as 3M Innovative Properties Company, which was not successfully matched
to CRSP because its name differs substantially from its parent. Although we only checked 250
assignees, this manual check allowed us to match an additional 64,000 patents. Firms with high
patenting activity but not matched to CRSP are either private companies or foreign firms that are
not listed on U.S. exchanges, an example of which is Hoffmann-La Roche, the large Swiss drug
company.

B.3 Correcting grant dates

The filing and grant dates of the patents are subject to the same sort of OCR errors as the assignee
information. The grant dates are particularly important for our purposes because we use them to
calculate the return around the grant date. Since patent numbers are sequential by grant dates, it

9
is easy to infer missing or incorrect grant dates by comparing patent dates to the grant dates of
adjacent patents. The same is not true of filing dates, but do not use filing dates in our current
work.
To populate missing patent dates and correct mistakes we identify the 3 non-missing grant dates
immediately preceding and following each patent. For example, if patent k’s grant date is missing
but patents (k − 3,. . . ,k − 1,k + 1,. . . ,k + 3) have grant date D, then we set patent k’s grant date
to D. By applying this procedure iteratively we are able to correct most grant dates, with the
exception of patents whose grant dates are missing and lie at a boundary between two grant dates.
We fill in these missing boundary dates by manually checking their grant dates on the USPTO’s
web site.
While we don’t rely on filing dates in the paper, it is possible to correct large errors in filing
dates by identifying cases where filing dates occur after the grant date, or much earlier than the
filing dates of adjacent patents. These errors often occur only in the year, so we can keep the
recorded month and day the same while setting the year of the patent filing to the median filing
year of a 20-patent window centered on a patent with an apparent error.

B.4 Extracting citations

Extracting patent citations from the patent text documents presents another challenge. The format
of a patent document has changed several times, as has the location and formatting of citations
within the document. For example, Figure 5 shows the references section of patent 2,423,030, granted
in 1947. The format seen here is the first format used after patent citation began in February, 1947.

Figure 5: A patent citation section

10
 A human reader has no problem identifying the citations in this patent. But to understand the
considerable challenge faced in automating this identification, consider the OCR for this part of the
patent:

other side. 35 REFERENCES CITED

By this invention I am able satisfactorily and The following references are of record in the
conveniently to effect the drying of’Shaped pot- jjle of tllis patent:
tery or other ceramic articles either in their
moulds or otherwise, in a manner which min- UNITED STATES PATENTS
imises risk of injury by excessively rapid heating 40 Number Name Date
or moisture extraction. The invention is not, 1,767,872 Pox June 24, 1930
however, restricted to the example described as 1^934,904 Barnett et al Nov. 14’, 1933
subordinate details may be modified to suit dif- 2,257,180 Mayer Sept. 30, 1941
ferent requirements. 1,893,963 Russ Jan. 10,1933
Having thus described my invention what I 45
claim as new and desire to secure by Letters Pat- * ("uu-^ f A 1 Jun 11>
entis: Number Country Date
1. Means for drying ceramic ware, comprising 439,577 Great Britain Dec. 10,1935

Our approach is to identify any text that could be a patent number (a 6- or 7-digit number,
perhaps separated by commas, spaces, or other “noise” characters) and is closely followed by the
correct grant date for the cited patent. In particular, for every potential patent number we identify,
we determine its grant date and then search near the possible citation for that date. If the date
appears, we can be very confident that we have correctly identified a citation. For example, for the
patent shown in Figure 5 we extract the patent number 1,767,872 and then confirm that its grant
date—June 24, 1930—appears somewhere nearby in the text. By using this two-step process to
identify citations, our citation extraction is of very high quality—the probability that some random
7-digit number will be followed closely by the correct date is clearly extremely small.
Our citation extraction method provides more citations than what is available on the Google
summary page. For example, the Google summary page for the patent shown in the previous
example provides no citations at all, while our algorithm correctly extracted all four citations. (We
exclude citations to foreign patents, as these patents are not in our database.) In general, Google
does not currently report out-cites from patents granted before 1976, so we use this extraction
method on all patents granted between 1926 and 1975.

B.5 Data validation

As previously mentioned, any data extraction project such as this can lead to two types of errors:
matching a patent to a firm that is not the assignee, or failing to match a patent to a any firm when
it does have an assignee. Our strategy makes the first error very unlikely, as a match occurs only

11
when a name closely resembling a CRSP company name appears around the word “assignee” at the
beginning of patent document. We cannot be sure how many errors of the second type we made,
but we have taken care to ensure that our algorithms allow as flexible matching as possible.
We also did two final checks to check the quality of our matching strategy. First, we visually
inspected a random sample of 500 patents granted between 1926 and 1975 and confirmed that
assignees had been correctly extracted, and correctly matched if the assignee appeared in CRSP.
This is obviously a very small sample of patents, but this careful check confirmed that no serious
errors existed.
Second, we applied the extraction and matching algorithms we used in the pre-1976 period to a
random sample of 25,000 patents granted between 1976 and 1999. We then compared our matches
to the matches in the NBER data. None of our matches was incorrect, and only 3 patents were
incorrectly not matched to an assignee. In other words, applying the techniques we used on pre-1976
data to data from the NBER period yields results that are virtually identical to those in the NBER
database.

12
C Additional Results and Descriptive Statistics

C.1 Alternative Distributional Assumptions

Here, we briefly describe how we allowed for different distributional assumptions to construct the
filtered value of a patent.

Allowing for a non-zero mean

We now assume that patent value is drawn from a non-zero mean distribution, i.e., vj ∼ N (µ, σvj 2 ),
truncated at zero. In this case, the filtered value of vj as a function of the stock return is equal to

φ(Rj )
E[vj |rjl ] = (1 − δj )µj + δj rjl +
p
δj σξj ,
1 − Φ(Rj )

where φ and Φ are the standard normal pdf and cdf, respectively, and R and δ are the normalized
return and the signal-to-noise ratio respectively,

(1 − δj )µj + δj rjl
Rj = − p ,
δj σξj

and δj is the signal-to-noise ratio as defined in the paper.

Relative to the paper, however, the additional issue we need to worry about is how to estimate
µj . Following the same procedure as the paper to allow the mean µ to vary by firm and year is not
possible, since at higher frequencies the standard deviation term dominates the mean. Hence, we
will need additional assumptions to recover µj . We will do so in two ways. In both methods it is
useful to note that the average stock return on patent announcement dates equals

E[rj ] = E[vj ] = µj + 0.7979 σvj

1. The first way assumes that µ is constant across firm-years. In this case, we estimate µj by
first subtracting 0.7979 σvf t from firm returns on patent announcements, and then estimate
µj as the difference in average returns on announcement versus non-announcement days using
the full sample. Our point estimates imply a µj = −0.6%. We use this estimate when forming
the conditional value of a patent as outlined above.

2. Alternatively, we assume that µ + 0.7979 σv is constant across firm-years. That is, we allow

13
 µ to vary exactly with σv . In this case, we estimate the difference in average returns on
announcement versus non-announcement days using the full sample. The point estimate is
0.02%. We then recover our estimate µj = 0.0002 − 0.7979 σvj .

We find that in both of the cases above, our results are very similar to the benchmark case
where we assume µ = 0. Specifically, the correlation between the filtered value of the patent E[vj |rjl ]
constructed using different assumptions on µ in (a) and (b) above with our benchmark values are
in excess of 99%. Indeed, allowing the pre-truncation mean to vary has mostly a scaling effect on
our estimates. Since the results are essentially extremely similar to the paper, we do not reproduce
them here.

Exponential

As before, we assume that the stock return of the firm on the patent grant date is given by

r = v + ε.

We now assume that v is exponentially distributed with parameter 1/σv . As in the paper, we still
assume the noise term is normally distributed, ε ∼ N (0, σε2 ). Under these assumptions, we can solve
for the conditional expectation – equivalent of equation (4) in the paper – in closed form,

r !
2 exp(−R̃2 /2) σε
E[v|R] = R + σε √ −
π Gc (R̃/ 2) σv

where Gc is the complementary error function and

σε r
R̃ = − .
σv σε

As in the paper, we assume that the ratio σv2 /σε2 is constant across firms. We use our estimates
from section 1.4, which imply σv2 /σε2 = 0.0142, so we use that as our baseline case.

Cauchy

As before, we assume that the stock return of the firm on the patent grant date is given by

r = v + ε.

14
where now v is distributed according to the positive part of a Cauchy distribution centered at zero
with scale parameter γv and ε follows a Cauchy distribution centered at zero with scale parameter
γε . In this case, r on announcement date is also Cauchy with scale γv + γε . Under the assumption
that both ε and v are Cauchy distributed, the conditional value of the patent is now given by:

  

γε c(r) − γv 2 ln (c(r)) + 2 r c(r) + γv 2 arctan γrε − 2 γε c(r) − γv 2 ln (γv ) + rπ r2 + (γε − γv )2 γv




E(v|r) =    
2 γε γv r ln (c(r)) + 2 γv (γv 2 − γε2 + r2 ) arctan γrε − 4 γε γv r ln (γv ) + π (γv + γε ) (r2 + (γε − γv )2 )

where
c(r) = r2 + γε2 .

Since the second moments of the Cauchy distribution do not exist, we need alternative ways of
estimating its parameters than what we used in the text. We estimate the scale of the noise term,
γε , using one-half the interquartile range of firm-year idiosyncratic returns, with an adjustment
similar to the equation in footnote 14 in the paper. Regarding the estimation of the noise-to-signal
ratio δ̃ = γv /(γv + γε ), we can no longer estimate it using equation (17) in the paper. Absent a
different alternative, we use the same estimates as in the paper implying a δ̃ = 0.014.

15
 Table A.1: Number of patents

Data step Number of patents

Total downloaded patents 7,797,506

Granted in 1926 or later 6,272,428
Identified as having an assignee 4,374,524
Matched to CRSP 1,928,123
Of which:
Present in NBER data 1,404,822
New to this paper 523,301

The table provides details on patents in our sample discussed in Section II.A of the paper. 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 1,928,123 patents to CRSP firms, of which 523,301
(27%) are new to this study. Further details are reported in Table 2 and Figure 1. In the paper, we restrict attention
to the patents that have a unique assignee, patents for which we have non-missing data on three day announcement
return, market capitalization and return volatilities needed to compute our Θ̂ measure. The sample contains 1,801,879
patents.

16
 Table A.2: Assignee matching by Decade

Number of patents Number of unique

Years Total With assignee Matched to CRSP Matched firms CRSP firms

1926–1929 174,022 48,433 8,858 182 786

1930–1939 442,700 172,925 47,029 355 951
1940–1949 307,499 141,345 60,616 451 1,042
1950–1959 425,953 171,157 82,255 587 1,246
1960–1969 567,599 265,524 165,409 1,175 3,177
1970–1979 690,459 393,661 247,102 2,086 7,204
1980–1989 708,735 579,518 235,525 2,756 11,715
1990–1999 1,109,398 933,705 352,005 3,664 14,882
2000–2010 1,846,063 1,668,256 729,324 4,415 11,900

All years 6,272,428 4,374,524 1,928,123 7,864 26,660

The table shows summary statistics for patents in our sample by decade discussed in Section II.A of the paper.
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.

17
 Figure A.1: Number of Patents with Matched Assignees

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.

18
 Table A.3: Innovation and Firm Size

Size (book assets) 1 2 3 4 5

Patents, citation-weighted (Θcw ) 1.2 2.3 3.9 8.2 90.4

Citations to Patents 2.6 2.6 2.5 2.3 2.2
Patents, citation-weighted, scaled by assets (%) 6.6 4.4 3.0 2.5 2.4
Patents, citation-weighted, scaled by mkcap (%) 7.9 6.5 5.8 5.4 22.5
Patents, SM weighted (Θsm ) 0.3 1.2 3.5 15.5 603.8
Total Value to Number of Patents 0.6 1.1 2.0 4.3 18.1
Patents, SM weighted, scaled by assets (%) 3.5 3.3 3.5 4.7 10.6
Patents, SM weighted, scaled by mkcap (%) 1.8 2.4 2.8 3.9 12.3

Size (Market cap of equity) 1 2 3 4 5

Patents, citation-weighted (Θcw ) 1.3 3.4 6.0 14.9 81.4

Citations to Patents 2.2 2.4 2.5 2.5 2.3
Patents, citation-weighted, scaled by assets (%) 4.0 4.4 4.0 3.4 3.1
Patents, citation-weighted, scaled by mkcap (%) 21.9 9.8 7.2 6.0 3.4
Patents, SM weighted (Θsm ) 0.1 0.8 2.2 9.3 618.4
Total Value to Number of Patents 0.3 0.6 1.2 2.7 19.1
Patents, SM weighted, scaled by assets (%) 1.2 2.4 3.5 4.7 13.9
Patents, SM weighted, scaled by mkcap (%) 2.4 2.9 3.1 4.3 10.6

Table reports mean value within each quintile. SM values are deflated by CPI (units are USDm
P in 1982). Quintiles
are computed using annual breakpoints. Citation-weighted patent counts are computed as j 1 + Cj /C̄j where Cj is
number of cites to patent j and C̄j is the mean number of cites to patents granted in the same year as patent j.

19
 Table A.4: Firm-level innovation measure: changes in distribution across decades

Decade Mean Sd p25 p50 p75 p90 p95 p99

1950 3.1 6.3 0.0 0.4 3.1 9.4 16.2 32.7

1960 4.7 10.0 0.0 0.0 4.7 14.4 23.8 51.6
1970 1.7 5.3 0.0 0.0 0.7 4.3 9.3 30.4
1980 1.2 4.0 0.0 0.0 0.0 3.1 8.1 21.7
1990 3.0 11.1 0.0 0.0 0.0 6.7 17.9 56.0
2000 5.6 19.2 0.0 0.0 1.5 14.6 32.6 86.5

Table reports the distribution of our baseline measure θfsm across decades. Units are in percentage terms.

20
 Table A.5: Mean innovation across industries

Ind Code Industry Name θcw θsm

1 Food Products 0.76 0.98
2 Beer & Liquor 0.16 2.10
3 Tobacco Products 0.32 1.59
4 Recreation 2.19 1.39
5 Printing and Publishing 0.66 0.18
6 Consumer Goods 4.02 3.48
7 Apparel 0.51 0.22
8 Healthcare, Medical Equipment, Pharmaceutical Products 9.09 9.13
9 Chemicals 6.97 5.93
10 Textiles 1.05 0.33
11 Construction and Construction Materials 2.50 1.20
12 Steel Works Etc 1.78 1.38
13 Fabricated Products and Machinery 6.67 3.66
14 Electrical Equipment 8.09 4.58
15 Automobiles and Trucks 4.67 2.72
16 Aircraft, ships, and railroad equipment 6.22 3.85
17 Precious Metals, Non-Metallic, and Industrial Metal Mining 0.52 0.32
18 Coal 0.23 0.09
19 Petroleum and Natural Gas 0.72 1.43
21 Communication 0.41 0.67
22 Personal and Business Services 2.15 2.25
23 Business Equipment 7.45 7.19
24 Business Supplies and Shipping Containers 2.78 2.39
25 Transportation 0.05 0.05
26 Wholesale 0.42 0.24
27 Retail 0.13 0.12
28 Restaurants, Hotels, Motels 0.05 0.03

Table reports mean value of normalized firm-level innovation θ (multiplied by 100) within each Fama-French industry
(using their 30 industry classification). We exclude financial firms and utilities.

21
 Table A.6: Estimates of Patent Value: Descriptive Statistics

Moment C C/C̄ Rf Baseline Exponential Cauchy

E[v|Rf ] ξ E[v|Rf ] ξ E[v|Rf ] ξ
(%) (%) USDm (%) USDm (%) USDm
Mean 10.26 1.18 0.07 0.32 10.36 0.40 12.79 0.15 5.13
Std. Dev 20.13 1.98 3.92 0.20 32.04 0.30 39.75 0.11 16.13
Percentiles
p1 0 0.00 -9.93 0.11 0.01 0.13 0.01 0.05 0.00
p5 0 0.00 -5.15 0.14 0.04 0.16 0.05 0.06 0.02
p10 0 0.00 -3.55 0.16 0.11 0.19 0.13 0.07 0.05
p25 1 0.20 -1.67 0.20 0.73 0.24 0.89 0.09 0.33
p50 5 0.62 -0.09 0.27 3.22 0.33 3.95 0.13 1.52
p75 11 1.38 1.62 0.37 9.09 0.46 11.23 0.18 4.45
p90 24 2.78 3.82 0.53 22.09 0.66 27.28 0.27 10.95
p95 38 4.06 5.73 0.68 38.20 0.85 47.27 0.34 19.13
p99 90 8.84 11.49 1.07 121.39 1.35 150.46 0.55 60.04

The table reports the distribution of the following variables across the patents in our sample: the number of future
citations till the end of our sample period C; the number of citations scaled by the mean number of cites to patents
issued in the same year C̄; the market-adjusted firm returns Rf on the 3-day window around patent grant dates;
the filtered component of returns E[v|Rf ] related to the value of innovation – using equation (4); and the filtered
dollar value of innovation ξ using equation (3) deflated to 1982 (million) dollars using the CPI. In addition to the
baseline case, we also report results using two alternative distributional assumptions. First, we assume that the
component of firm return due to the patent, v, is exponentially distributed with scale parameter 1/σv . As before, we
assume the signal-to-noise ratio is constant across firms; using our estimates from equation (6) in the paper, we obtain
σv /σε ≈ 0.014, so we use that. As before, we allow σε to vary by firm-year and follow the same exact procedure as
in the baseline case. Second, we assume that v is distributed according to a Cauchy truncated at zero with scale
γv , while ε is distributed according to a Cauchy with parameter γε . We estimate the scale of the noise term, γε ,
using one-half the interquartile range of firm-year idiosyncratic returns, with an adjustment similar to the equation in
footnote 13 in the paper. Regarding the estimation of the noise-to-signal ratio δ = γv /(γv + γε ), we can no longer
estimate it using equation (6) in the paper because the variance of the Cauchy distribution does not exist. Absent a
different alternative, we use the same estimate as in the paper. We restrict attention to the patents for which we have
non-missing data on three day announcement return, market capitalization and return volatilities needed to compute
our Θ̂ measure. The sample contains 1,801,879 patents.

22
 Table A.7: Forward Citations and Patent Market Values – Alternative Distributions

(1) (2) (3) (4) (5)

A. Exponential
log(1 + Cj ) 0.174 0.099 0.055 0.013 0.004
(9.99) (9.44) (10.28) (13.84) (5.12)
B. Cauchy
log(1 + Cj ) 0.173 0.096 0.059 0.016 0.004
(10.25) (9.49) (10.35) (12.86) (4.84)
Controls
Firm Market Capitalization - Y Y Y Y
Volatility - - Y Y -
Fixed Effects TxC TxC TxC TxC TxC
F FxT

Table reports the equivalent of Table 2 in the paper under two alternative distributional assumptions. Panel A presents
results under the assumption that the component of firm return due to the patent, V , is exponentially distributed
with scale parameter 1/σv . Panel B presents results under the assumption that v is distributed according to a Cauchy
truncated at zero with scale γv , while ε is distributed according to a Cauchy with parameter γε . See notes to Table A.6
for more details.

Table A.8: Innovation and Firm Profit Growth – Alternative Distributions

A. Exponential
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.018 0.029 0.036 0.042 0.046 -0.015 -0.029 -0.032 -0.035 -0.038
[3.61] [4.49] [3.74] [3.81] [3.60] [-2.96] [-5.05] [-7.25] [-5.98] [-5.84]
B. Cauchy
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.018 0.027 0.035 0.040 0.045 -0.012 -0.024 -0.027 -0.031 -0.034
[5.74] [5.95] [5.18] [4.98] [4.97] [-2.19] [-3.40] [-4.95] [-5.05] [-4.81]

Table reports the equivalent of Table 4, Panel A in the paper under two alternative distributional assumptions. Panel
A presents results under the assumption that the component of firm return due to the patent, V , is exponentially
distributed with scale parameter 1/σv . Panel B presents results under the assumption that v is distributed according
to a Cauchy truncated at zero with scale γv , while ε is distributed according to a Cauchy with parameter γε . See
notes to Table A.6 for more details.

23
 Table A.9: Innovation and Firm Growth: Results using Alternative Scaling (Market Capitalization)

a. Profits
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.006 0.017 0.023 0.028 0.034 -0.027 -0.034 -0.038 -0.043 -0.043
[2.45] [5.53] [4.34] [4.21] [4.07] [-5.64] [-3.29] [-4.14] [-4.33] [-4.37]
b. Output
Firm Competitors
1 2 3 4 5 1 2 3 4 5
-0.003 0.001 0.003 0.012 0.021 -0.041 -0.051 -0.058 -0.056 -0.064
[-1.34] [0.27] [0.54] [1.60] [2.23] [-6.23] [-3.52] [-3.65] [-3.47] [-3.84]
c. Capital
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.003 0.007 0.011 0.015 0.021 -0.013 -0.027 -0.039 -0.050 -0.062
[1.74] [2.19] [2.39] [2.58] [2.74] [-3.33] [-4.94] [-5.63] [-6.22] [-6.71]
d. Labor
Firm Competitors
1 2 3 4 5 1 2 3 4 5
-0.001 0.002 0.006 0.011 0.014 -0.019 -0.026 -0.032 -0.033 -0.032
[-0.61] [0.79] [1.53] [1.92] [1.99] [-6.19] [-4.61] [-5.29] [-4.40] [-4.23]
e. TFPR
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.003 0.010 0.012 0.016 0.017 -0.005 -0.009 -0.013 -0.013 -0.013
[1.09] [3.17] [2.74] [4.22] [4.74] [-2.62] [-3.59] [-4.53] [-3.09] [-2.58]

Table repeats the analysis in Table 4 in the paper. Rather than book assets, we now scale the firm’s dollar value of
innovation by its end of year market capitalization. Similarly, innovation by competing firms is constructed as the
dollar value of innovation divided by their total market capitalization, in a manner analogous to equation (11) in the
paper. See notes to Table 4 in the paper for more details.

24
 Table A.10: Innovation and Firm Profit Growth – Patent citations measured within a fixed window

A. Citations within 3-years of patent grant

Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.007 0.012 0.017 0.021 0.025 -0.005 -0.006 -0.007 -0.006 -0.006
[4.41] [4.92] [5.08] [4.99] [5.31] [-1.85] [-1.56] [-1.76] [-1.13] [-1.06]
B. Citations within 5-years of patent grant
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.008 0.014 0.019 0.024 0.027 -0.005 -0.007 -0.009 -0.007 -0.007
[4.73] [5.58] [5.35] [5.35] [5.71] [-1.69] [-1.78] [-2.11] [-1.43] [-1.24]
C. Citations within 10-years of patent grant
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.008 0.015 0.022 0.027 0.032 -0.003 -0.005 -0.009 -0.007 -0.007
[4.86] [5.77] [5.62] [5.69] [6.59] [-0.89] [-1.29] [-2.11] [-1.28] [-1.23]

Table reports the equivalent of Table 5, Panel A in the paper under different ways of adjusting patent citations for
truncation lags. In each of the panels A, B, and C, we measure forward citations over the first N years after the
patent is issued, where N = 3, 5, 10. We then repeat the analysis in Table 5 by also excluding the last N years from
the sample. See notes to Table 5 in the paper for more details.

25
 Table A.11: Innovation and Firm Growth: Controlling for R&D spending of Firm and Competitors

Profits
Firm a. Competitors
1 2 3 4 5 1 2 3 4 5
0.017 0.027 0.034 0.039 0.043 -0.018 -0.034 -0.037 -0.040 -0.045
[3.33] [4.12] [3.39] [3.48] [3.31] [-3.24] [-5.42] [-8.23] [-6.36] [-6.38]
b. Output
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.008 0.013 0.018 0.022 0.028 -0.016 -0.034 -0.044 -0.048 -0.056
[2.85] [3.07] [2.85] [2.62] [3.15] [-3.64] [-6.33] [-8.43] [-7.91] [-7.98]
c. Capital
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.010 0.021 0.028 0.034 0.039 0.002 -0.005 -0.012 -0.020 -0.029
[8.40] [6.69] [5.76] [4.40] [4.13] [0.38] [-0.82] [-1.59] [-2.35] [-3.15]
d. Labor
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.007 0.014 0.019 0.023 0.025 -0.007 -0.017 -0.021 -0.021 -0.019
[5.65] [4.39] [4.11] [3.76] [3.30] [-1.64] [-3.86] [-4.34] [-3.67] [-3.00]
e. TFPR
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.012 0.015 0.017 0.021 0.022 -0.001 -0.006 -0.010 -0.015 -0.017
[2.18] [2.07] [2.56] [3.28] [3.98] [-0.52] [-2.20] [-3.20] [-4.59] [-4.09]

Table repeats the analysis of Table 4 in the paper including the firm’s R&D spending as an additional control. We
control for the firm’s ratio of R&D spending to sales, as well as the ratio of total R&D spending to total sales of
competing firms. See notes to Table 4 in the paper for additional details.

26
 Table A.12: Innovation and Firm Growth: Controlling for Measures of Investor Attention

Horizon 1 2 3 4 5
A. Control for number of WSJ articles
Firm 0.017 0.030 0.036 0.043 0.045
[2.60] [4.31] [2.89] [3.23] [3.12]
Competitor -0.033 -0.053 -0.047 -0.057 -0.074
[-13.84] [-5.67] [-5.53] [-6.45] [-6.78]
N 28966 26293 24077 19541 15529
B. Control for number of analysts
Firm 0.011 0.018 0.022 0.026 0.028
[2.66] [4.45] [3.50] [3.37] [3.07]
Competitor -0.020 -0.040 -0.042 -0.042 -0.046
[-2.47] [-5.36] [-7.12] [-5.38] [-4.99]
N 77483 69322 62209 55734 49858
control for institutional ownership
Firm 0.019 0.029 0.036 0.040 0.042
[3.18] [4.51] [3.70] [3.23] [3.00]
Competitor -0.015 -0.032 -0.034 -0.037 -0.040
[-2.26] [-4.33] [-5.68] [-4.53] [-4.61]
N 70874 61737 53941 47165 41338

Table repeats the analysis in Table 4 in the paper using additional controls for the degree of investor attention. In
panel A we control for the (log one plus the) number of articles that mention the firm in the Wall Street Journal. The
data is available over the 2000-2007 period. The data is from matching news articles in Factiva to firms following the
procedure in Butler and Gurun, 2012. In Panel B, we control for the (log of one plus the) number of analysts covering
the stock. The data is from I/B/E/S and covers the 1975-2010 period. In Panel C, we control for the fraction of
institutional ownership. The data is from Thomson Reuters Institutional (13f) Holdings - Stock Ownership Summary
and covers the 1980-2010 period. See the notes to Table 4 in the paper for additional details.

27
 Table A.13: Innovation and Firm Growth: IV using tax price of R&D

Profits
Firm a. Competitors
1 2 3 4 5 1 2 3 4 5
0.116 0.221 0.273 0.352 0.409 -0.116 -0.207 -0.266 -0.327 -0.373
[1.92] [2.09] [1.94] [1.93] [1.75] [-2.54] [-2.50] [-2.37] [-2.30] [-2.08]
b. Output
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.069 0.099 0.148 0.208 0.262 -0.070 -0.134 -0.187 -0.233 -0.273
[1.91] [1.60] [1.67] [1.76] [1.69] [-2.29] [-2.49] [-2.44] [-2.35] [-2.16]
c. Capital
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.076 0.128 0.176 0.233 0.296 -0.092 -0.171 -0.242 -0.319 -0.386
[2.12] [2.02] [2.00] [2.04] [1.97] [-2.99] [-3.06] [-3.08] [-3.12] [-2.94]
d. Labor
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.052 0.083 0.113 0.148 0.195 -0.073 -0.139 -0.181 -0.213 -0.233
[1.64] [1.43] [1.39] [1.39] [1.39] [-2.55] [-2.76] [-2.55] [-2.28] [-1.95]
e. TFPR
Firm Competitors
1 2 3 4 5 1 2 3 4 5
0.092 0.149 0.170 0.212 0.229 -0.063 -0.095 -0.117 -0.130 -0.142
[2.26] [2.44] [2.31] [2.38] [2.16] [-1.81] [-1.77] [-1.78] [-1.65] [-1.55]

Table repeats the analysis of Table 4 in the paper using instrumental variables. We use the R&D tax credit variation
as an instrument for our innovation measure, following Bloom, Schankerman, and Van Reenen (2013). This R&D
price is constructed at an annual level for each firm using state-level R&D tax credits. See Bloom et al. (2013) for
more details on the construction of this variable. We instrument for the firm’s own innovation θf t using the firm-level
tax price; we instrument for the innovation by competing firms θI\f,t using the average R&D price of competing firms.
The first-stage F statistics vary from 17.1 to 61 across specifications and horizons. We cluster standard errors by firm.
See notes to Table 4 in the paper for additional details.

28
 Table A.14: Industry Output and Innovation – Comparison with Citation-Weighed Patents

Horizon (years) 1 2 3 4 5 6 7 8
A. Industry output (quantity)
SM 0.005 0.009 0.013 0.018 0.026 0.042 0.055 0.065
[2.73] [2.49] [2.83] [3.02] [3.29] [3.49] [3.32] [3.23]
R-sq 0.054 0.085 0.119 0.150 0.183 0.217 0.241 0.261
CW 0.007 0.016 0.024 0.031 0.038 0.044 0.049 0.053
[3.07] [3.61] [4.05] [4.11] [3.96] [3.81] [3.61] [3.29]
R-sq 0.051 0.086 0.119 0.148 0.176 0.197 0.215 0.229
B. Industry output (value added)
SM 0.001 0.002 0.004 0.007 0.012 0.027 0.038 0.047
[0.61] [0.44] [0.64] [0.79] [1.15] [2.35] [2.78] [2.83]
R-sq 0.014 0.027 0.044 0.059 0.077 0.098 0.121 0.140
CW 0.004 0.009 0.013 0.017 0.021 0.025 0.028 0.027
[1.57] [1.75] [1.77] [1.76] [1.78] [1.75] [1.68] [1.45]
R-sq 0.015 0.029 0.045 0.060 0.077 0.092 0.109 0.122
N 1395 1364 1333 1302 1271 1240 1209 1178

Table reports the relation between innovation and output growth at the industry level. We construct industry-level
innovation indices as P i
i f ∈I Θf,t
θI,t = P ,
f Bf t

using both the market based measure (i = sm) as well as for cohort-adjusted, citation-weighted patent counts (i = cw).
We report the estimated coefficients aτ from a specification similar to equation (12) in the paper,
i
xt+τ − xt = a0 + aτ θI,t + ρxt + ZIt + ut+τ .

where x is log industry output (quantity in Panel A, value added in Panel B) and Z is a vector of controls that
includes log capital, log employment, mean industry id. volatility and time effects. We compute standard errors using
i
Newey-West. To compare across the two measures, we scale θI,t to unit standard deviation.

29
 Table A.15: Innovation and Aggregate Growth – Comparison with Citation-Weighted Patents

(1) (2) (3) (4) (5) (6) (7) (8)

A. Aggregate Output
χcw 0.005 0.009 0.013 0.015 0.017 0.022 0.023 0.022
[1.08] [1.10] [1.36] [1.40] [1.70] [2.31] [2.24] [2.13]
R-sq 0.078 0.113 0.19 0.254 0.289 0.347 0.373 0.423
B. Aggregate TFP
χcw 0.004 0.007 0.011 0.016 0.02 0.022 0.023 0.026
[2.30] [2.20] [2.27] [2.94] [3.12] [3.14] [3.33] [3.89]
R-sq 0.201 0.305 0.397 0.493 0.535 0.559 0.592 0.652

Table repeats the analysis of Figure 5 in the paper using an alternative index of innovation that is constructed using
patent citations. Specifically, in a direct analogy to equation (18) in the paper, the value of the index in year t is
given by
P
j∈Jt Ĉj
χcw
t = ,
Yt
where Ĉ is the number of citations to patent j in the first 10 years since its grant date, Jt is the set of patents issued
in year t (including both private and public firms) and Y is aggregate output. Due to truncation, the sample ends in
2000. We report the estimated coefficients aτ from the following specification
L
X
xt+τ − xt = a0 + aτ log χ̂cw + cl xt−l + ut+τ .
l=0

Here, x is log aggregate output (panel A) or log TFP (panel B). We scale log χ̂cw to unit standard deviation. We
examine horizons of one to five years. We select the number of lags L using the BIC criterion, which advocates a lag
length of one to two years depending on the specification. We compute standard errors using Newey-West.

30
 ·10−2
2

Filtered patent value, E[v|R]

1.5

0.5

−0.2 −0.1 0 0.1 0.2 0.3

Firm stock return, R

Figure A.2: Plot compares the filtered values, E[v|R] across three different assumptions: our baseline case
(black); the assumption that v is exponentially distributed (blue); the assumption that ε is Cauchy distributed
and v follows a truncated Cauchy (at zero). See notes to Table A.7 for more details on the estimation of the
parameters. We use the sample mean for the variance (or scale) of the error term to draw these graphs. We
use the implied estimates from equation (6) in the main text to calibrate δ across the three cases.

31
 Figure A.3: Innovation and Aggregate Growth – VAR results

(a) Output (b) TFP

.04
.01
(i) scaled by GDP

.02

.005
0

−.02 0
1 2 3 4 5 1 2 3 4 5
Years Years

.01
.04
(ii) scaled by MKCAP

.02
.005

−.02 0
1 2 3 4 5 1 2 3 4 5
Years Years

Figure shows impulse response of output per capita and productivity to innovation using bi-variate VARs. We
obtain impulse responses by ordering our innovation measure last. We select lag length based on the BIC criterion.
Dotted lines represent 90% confidence intervals using standard errors are computed using 500 bootstrap simulations.
Productivity is utilization-adjusted TFP from Basu, Fernald, and Kimball (2006). Output is gross domestic product
(NIPA Table 1.1.5) divided by the consumption price index (St Louis Fed, CPIAUCNS). Output per capita is computed
using population from the U.S. Census Bureau.

32
References
Basu, S., J. G. Fernald, and M. S. Kimball (2006). Are technology improvements contractionary?
American Economic Review 96 (5), 1418–1448.

Bloom, N., M. Schankerman, and J. Van Reenen (2013). Identifying technology spillovers and
product market rivalry. Econometrica 81 (4), 1347–1393.

Hall, B., A. J. and M. Trajtenberg (2001). The NBER patent citation data file: Lessons, insights
and methodological tools. Technical report, NBER Working Paper 8498.

Nicholas, T. (2008). Does innovation cause stock market runups? Evidence from the great crash.
American Economic Review 98 (4), 1370–96.

Norvig, P. (2009). Natural langugage corpus data. In T. Segaran and J. Hammerbacher (Eds.),
Beautiful Data. O’Reilly Media.

33