PDF to HTML - Convert PDF files to HTML files

Europaudvalget 2003-04
EUU Alm.del Bilag 111
Offentligt

WORKING PAPER NO. 03-17/R

AN EMPIRICAL LOOK AT SOFTWARE PATENTS

James Bessen*

Research on Innovation and

Boston University School of Law (Visiting Researcher)

Robert M. Hunt**

Federal Reserve Bank of Philadelphia

First Draft:

This Draft:

August 2003

March 2004

[email protected]

** Ten Independence Mall, Philadelphia, PA 19106. Phone: (215) 574-3806. Email:

[email protected]

Thanks to Peter Bessen of May8Software for providing a software agent to acquire our patent database and

Annette Fratantaro for her work with the Compustat data set. Also thanks to John Allison, Tony Breitzman

and CHI Research, Iain Cockburn, Mary Daly, Dan Elfenbein, Terry Fisher, Bronwyn Hall, Joachim

Henkel, Brian Kahin, David Mowery, Leonard Nakamura, Cecil Quillen, Eric von Hippel, Rosemarie

Ziedonis and seminar participants at APPAM, Berkeley, EPIP Munich, Federal Reserve Banks of

Philadelphia and San Francisco, the Federal Reserve System Applied Micro meetings, Harvard, IDEI,

MIT, NBER, and OECD.

The views expressed here are those of the authors and do not necessarily represent the views of the Federal

Reserve Bank of Philadelphia or the Federal Reserve System.

2004,

Verbatim copying and distribution of this entire article for noncommercial use is permitted in any

medium provided this notice is preserved.

PDF to HTML - Convert PDF files to HTML files

AN EMPIRICAL LOOK AT SOFTWARE PATENTS

James Bessen

Research on Innovation and

Boston University School of Law (Visiting Researcher)

Robert M. Hunt

Federal Reserve Bank of Philadelphia

March 2004

Abstract:

U.S. legal changes have made it easier to obtain patents on inventions that use software.

Software patents have grown rapidly and now comprise 15 percent of all patents. They

are acquired primarily by large manufacturing firms in industries known for strategic

patenting; only 5 percent belong to software publishers. The very large increase in

software patent propensity over time is not adequately explained by changes in R&D

investments, employment of computer programmers, or productivity growth. The residual

increase in patent propensity is consistent with a sizeable rise in the cost effectiveness of

software patents during the 1990s. We find evidence that software patents

substitute

for

R&D at the firm level; they are associated with

lower

R&D intensity. This result occurs

primarily in industries known for strategic patenting and is difficult to reconcile with the

traditional incentive theory of patents.

Keywords: Software, Patents, Innovation, Technological Change

JEL classification: O34, D23, L86

PDF to HTML - Convert PDF files to HTML files

Introduction

Federal courts, and to a lesser extent the U.S. Patent and Trademark Office (USPTO),

dramatically changed standards for patenting software-related inventions over the last

three decades. During the 1970s, federal court decisions typically described computer

programs as mathematical algorithms, which are unpatentable subject matter under U.S.

law.

Systems using software could be patented, but only if the novel aspects of the

invention did not reside entirely in the software.

At this time, the U.S. Congress

considered the question of patenting software and instead opted to protect computer

programs under copyright law.

But after the Supreme Court decision in

Diamond v. Diehr

in 1981,

a series of

court and administrative decisions gradually relaxed the subject matter exception that

restricted the patenting of software-related inventions. The 1994 decision

In re Alapat

eliminated much of the remaining uncertainty over the patentability of computer

programs.

During this same period, new legislation and other court decisions lowered

standards for obtaining patents in general, while strengthening aspects of patent

enforcement.

This paper explores the general characteristics of software patenting over the last

two decades, paying particular attention to the rapid growth in software patenting and the

effect of this growth on R&D. We construct our own definition of a

software patent

(there is no official definition) and assemble a comprehensive database of all such

patents. In Section I we describe this process, and the process of matching these patents

to firm data in the Compustat database. In Section II we summarize the general

characteristics of this data. We find that over 20,000 software patents are now granted

each year, comprising about 15 percent of all patents. Compared with other patents,

See, for example, the Supreme Court decision in

Gottschalk v. Benson,

409 U.S. 63 (1972).

Parker v. Flook

437 U.S. 584 (1978).

U.S. Copyright law was amended in 1976, and more explicitly in 1980, to include computer

programs. See H. Rpt. No. 94-1476 (1976) and P.L. 96-517 (94 Stat 3028). There is a voluminous

literature on the merits of different forms of intellectual property protection for computer programs. See,

for example, Dam (1995), Graham and Zerbe (1996), and Samuelson et al. (1994).

450 U.S. 175 (1981).

33 F.3d 1526 (Fed. Cir. 1994).

PDF to HTML - Convert PDF files to HTML files

software patents are more likely to be assigned to firms, especially larger U.S. firms, than

to individuals. They are also more likely to have U.S. inventors. Surprisingly, most

software patents are assigned to manufacturing firms and relatively few are actually

assigned to firms in the software publishing industry (SIC 7372). Most software patents

are acquired by firms in industries that are known to accumulate large patent portfolios

and to pursue patents for strategic reasons (computers, electrical equipment, and

instruments). These large inter-industry differences remain even after we control for

R&D, software development effort, and other factors.

In Section III we perform regressions that explore the “propensity to patent”

software inventions. This builds on the model of Hall and Ziedonis (2001) which, in turn,

builds on the empirical literature of “patent production functions” (including Scherer

1965, Bound et al. 1984, Pakes and Griliches 1984, and Griliches, Hall, and Hausman

1986). We find a dramatic growth in software patent propensity even after controlling for

R&D, employment of computer programmers, and other factors. This growth is quite

similar to the remarkable growth in patent propensity that Hall and Ziedonis (2001)

found in the semiconductor industry. Productivity-based explanations are unlikely to

account for even half of the rise in software patent propensity. The pattern of the residual

increase is consistent with the explanation that changes in patent law made software

patents significantly more cost effective. We also find that industries known for strategic

patenting have much higher patent propensities.

Section IV explores the effect of software patenting on R&D. According to the

traditional incentive theory, making software patents more cost effective should increase

the profitability of firms that make software inventions. This, in turn, should induce them

to increase their R&D spending. To test whether this in fact happened, we use a well-

established empirical technique for estimating elasticities of factor substitution. We show

that the incentive hypothesis can be re-stated as the hypothesis that R&D and patents are

complements. This means that increases in the appropriability of software should lead to

greater R&D intensity. We find that this is not the case, however. Firms that increased

their software patenting relative to their overall level of patenting tended to

decrease

their R&D intensity relative to other firms. This result is robust to a variety of

PDF to HTML - Convert PDF files to HTML files

econometric and other considerations. But, again, this effect is concentrated in the

industries known for strategic patenting.

In Section V, we note that while our empirical analysis does not identify the

specific causal mechanisms at work, these results are difficult to reconcile with the

traditional incentive hypothesis. Strategic patenting provides an explanation for a rise in

patent propensity together with an apparent substitution away from R&D: firms may be

engaged in a patent “arms race.” The prominence of certain industries—known for

strategic patent behavior in other contexts—in our empirical results may not be

coincidental. Section VI concludes.

Background and Data

A. Changing Legal Treatment and Strategic Patenting Industries

The erosion of the subject matter exception for computer programs occurred against a

backdrop of broader changes in patent law following the creation of a unified appeals

court for patents suits in 1982 (see Hall and Ziedonis 2001, for a nice summary). The

court raised the evidentiary standards required to challenge patent validity and tended to

broaden the interpretation of patent scope (Rai 2003, Merges 1997). The court relaxed

the standards for evaluating whether or not an invention is obvious to practitioners skilled

in the art (Cooley 1994, Dunner et al. 1995, Hunt 1999, Lunney 2001). The court was

also more willing to grant preliminary injunctions to patentees (Cunningham 1995,

Lanjouw and Lerner 2001) and to sustain large damage awards (Merges 1997, Kortum

and Lerner 1999). Finally, plaintiff success rates in patent infringement suits have

increased substantially (Lerner 1995).

Some of these changes made patents “stronger,” in the sense that patents became

more likely to be upheld in court or more effectively enforced. Others made patents

“cheaper,” in the sense that lower patentability standards reduced the effort required to

obtain a patent on a given invention. Together, these changes made patents more

cost

effective

than before, generating more appropriability per dollar invested in obtaining and

asserting them.

PDF to HTML - Convert PDF files to HTML files

If patents did become more cost effective, the change is likely greater for software

than for other inventions, for two reasons. First, the presumption that computer programs

could not be patented was largely reversed by the mid-1990s. Second, a number of other

legal decisions relaxed the “enablement” requirement for software patents. Under U.S.

patent law, filers are required to provide detailed instructions explaining how the

invention works. It is supposed to be a “best mode” example of all of the patent claims.

For software patents and business methods, it seems the courts have largely eliminated

this requirement (Burk 2002, Burk and Lemley 2002).

In the words of an IBM patent

attorney, “[the patent standard] currently being applied in the U.S. invites the patenting of

ideas that may have been visualized as desirable but have no foundation in terms of the

research or development that may be required to enable their implementation” (Flynn,

2001). The combined effect of these regulatory changes is that software patents appear to

have gained greater appropriability and became less costly to obtain in absolute terms

over time and also possibly relative to other patents.

Yet the software industry was highly innovative and growing rapidly well before

software patents became commonplace. Nominal investment in software grew 16 percent

per annum during the 80s (and 11 percent per annum during the 90s, Grimm and Parker

2000). This innovativeness is important for two reasons. First, in interpreting results

below, the growing use of software is an important factor. Second, given this history, it is

not at all clear that patent protection was essential for innovation in this industry.

This is not unusual. When surveyed, American firms in a number of other

innovative industries (including semiconductors and precision instruments) rate patents

as a

relatively

less effective form of appropriability (Levin et al. 1987, Cohen, Nelson,

and Walsh 2000). Instead, they cite lead time advantages, learning curves,

complementary sales and service and secrecy as generally more important sources of

appropriability.

Reviewing case law, Burk and Lemley (2002, p. 1162) write: “For software patents, however, a series

of recent Federal Circuit decisions has all but eliminated the enablement and best mode requirements. In

recent years, the Federal Circuit has held that software patents need not disclose source or object code,

flow charts, or detailed descriptions of the patented program. Rather, the court has found high-level

functional description sufficient to satisfy both the enablement and best mode doctrines.” See also Cohen

and Lemley (2001) on the different treatment of software patents.

PDF to HTML - Convert PDF files to HTML files

Yet even in industries where patents are rated as ineffective, we sometimes

observe that firms sometimes acquire large patent portfolios. These industries, including

the computer, electrical equipment and instruments industries, are also found to account

for a major share of the growth in patenting in recent years (Hall 2003). Some researchers

have suggested that firms in these industries may patent heavily in order to obtain

strategic advantages, including advantages in negotiations, cross-licensing, blocking

competitors, and preventing suits (Levin et al. 1987, fn 29, and Cohen, Nelson, and

Walsh 2000). In principle, strategic patenting can arise whenever individual products

involve many patentable inventions and the cost of obtaining patents is sufficiently low

(see Bessen 2003 for a theoretical model). Firms may acquire large numbers of patents so

that even if they have an unsuccessful product, they can hold up rivals, threatening

litigation. Innovative firms may acquire “defensive” patent portfolios to make a credible

counter-threat. The outcome may involve the cross-licensing of whole portfolios, where

firms agree not to sue each other and those firms with weaker portfolios pay royalties

(Grindley and Teece 1997).

So, in what follows, it is natural for us to be alert to the possibility that software

patents may also be acquired for strategic purposes and we will find distinctive behavior

in the industries known for strategic patenting. An important implication of strategic

patenting is that policy changes that “strengthen” patents (or make them cheaper to

acquire) can lead to a kind of “Prisoner’s Dilemma” game that actually

decreases

the

private incentive to engage in R&D.

B. What Is a Software Patent?

How many software patents are being granted? Although the patent office maintains a

system for classifying patents, this system does not distinguish whether the underlying

technology is software or something else. Researchers must construct their own

definitions.

Some observers have sought to identify “pure” software patents where the

invention is completely embodied in software (e.g. Allison and Lemley 2000). We prefer

not to use such a definition for two reasons. First, beginning with

Diamond v. Diehr,

PDF to HTML - Convert PDF files to HTML files

attorneys have drafted software patents in such a way that they did not necessarily

appear

to be patents on software.

This makes the determination of a “pure” software patent

somewhat arbitrary and impractical for a comprehensive database. Second, we do not

necessarily assume that the subject matter exclusion is the

only

difference between

software patents and other patents. For example, we noted above that software patents are

subject to a different “enablement” requirement. So it is useful to study a somewhat

broader range of patents in any case.

Our concept of software patent involves a logic algorithm for processing data that

is implemented via stored instructions; that is, the logic is not “hard-wired.” These

instructions could reside on a disk or other storage medium or they could be stored in

“firmware,” that is, a read-only memory, as is typical of embedded software. But we want

to exclude inventions that involve only off-the-shelf software—that is, the software must

be at least novel in the sense of needing to be custom-coded, if not actually meeting the

patent office standard for novelty.

1. Identifying software patents

How can we identify patents that fit this description? Griliches (1990) reviews the two

main techniques that researchers have used to assign patents to an industry or technology

field: 1) using the patent classification system developed by the patent office; and 2)

reading and classifying individual patents. In this paper, we use a modification of the

second technique.

We began by reading a random sample of patents, classifying them according to

our definition of software, and identified some common features of these patents. We

used these to construct a search algorithm to identify patents that met our criteria. We

used this algorithm to perform a keyword search of the U.S. Patent Office database,

which identified 130,650 software patents granted in the years 1976 to 1999. Next, to

For instance, Cohen and Lemley (2001, p. 9) argue “The Diehr decision and its appellate progeny

created what might be termed ‘the doctrine of the magic words.’ Under this approach, software was

patentable subject matter, but only if the applicant recited the magic words and pretended that she was

patenting something else entirely.” In

Diamond v. Diehr,

the Supreme Court ruled that an invention using

temperature sensors and a computer program to calculate the correct curing time in an otherwise

PDF to HTML - Convert PDF files to HTML files

validate the accuracy of this algorithm, we compared the results of our search against a

random sample of 400 patents which we had read and classified into software patents and

other patents. We also compared our results to samples and statistics generated by other

researchers. The details of the algorithm and characteristics of the sample are described

in the Appendix.

Compared to our random sample of 400 patents, this algorithm had a false

positive rate of 16 percent (that is, 16 percent of the patents the algorithm said were

software patents, were not) and a false negative rate of 22 percent (that is, it failed to

identify 22 percent of the patents we categorized as software patents).

We performed a number of other checks. First, we compared our list of software

patents to a set of 330 software and Internet patents identified in research conducted for

the papers by Allison and Lemley (2000) and Allison and Tiller (2003).

These patents

were identified by reading a larger number of patents, but applying a more narrow

definition of software inventions (again, where the invention is completely embodied in

software).

Virtually all (92 percent) of the software patents identified by Allison were

categorized as software patents by our algorithm. Thus our false negative rate for “pure”

software patents appears to be quite small (8 percent). Second, using statistics generated

by the research in Allison and Lemley (2000), we calculated an upper bound on the

comparable false positive rate of 26 percent.

Given that we are using a broader

definition of software patent, it is reassuring to find that this number is not much larger

conventional process of molding rubber goods could be patented. We treat this as a software patent;

Allison and Lemley would not.

Thanks to John Allison for sharing his data with us. The data used in Allison and Lemley (2000) are

based on reading 1,000 randomly selected patents issued between mid 1996 and mid 1998. That data was

augmented in Allison and Tiller (2003) by examining 2,800 patents issued between 1990 and 1999

identified via a keyword search (for the terms

Internet

world wide web)

restricted to patents included in

classes 705, 707, or 709. Note that in the Allison and Tiller taxonomy, internet business method patents are

a subset of software patents.

Both papers state the following: “Another researcher might include within the Software classification

those inventions in which the algorithms are embodied in chips, but we have chosen to include within our

definition of Software only those inventions that consist purely of software that is not embodied in

hardware.”

Allison reports identifying 92 pure software patents in the sample of 1,000 evaluated in that paper

(private communication). This is somewhat higher than the 76 reported in the published version of the

paper. For the same years, the ratio of software patents to total patents calculated using our algorithm was

11.5%. Thus an upper bound on the false positive rate is (.115-.092*.92)/.115 = 26%.

PDF to HTML - Convert PDF files to HTML files

than the false positive rate generated when using our own random sample. Although the

algorithm does make errors, it performs reasonably well, and it seems unlikely that it

introduces significant biases to our patent counts or regression coefficients.

2. Using Patent Classes to Identify Software Inventions

Why did we use this rather laborious method rather than simply counting patents in

certain patent classifications? First, in a longitudinal study patent classes are problematic

because the classification system changes over time and the patent office continually re-

classifies issued patents. Moreover, lawyers are known to draft patents so that they avoid

falling into certain classes in order to influence the examiner’s prior art search or some

other aspects of the examination (Lerner 2004, p. 19).

Second, economists have long recognized the poor correspondence between

patent classes and economic concepts of industry or technology (see, for example,

Schmookler 1966, Scherer 1982a, Scherer 1984, Soete 1983, and Griliches 1990). Patent

classes are not designed with social scientists in mind but are used primarily to aid prior

art search. Although patent classes can be used effectively when they are confined to

select sets of well-defined subclasses (e.g., Schmookler 1966, Lerner 2004), or when

classifications are statistically distributed over industries (e.g., Silverman 1999), or when

taken as loosely representative (e.g., Graham and Mowery 2003), a definition based on

patent classes is likely to introduce significantly more inaccuracies than the approach

chosen in this paper.

In the case of the U.S. classification system, there are no patent classes for

software

per se.

Instead, software inventions are included in functional categories along

with hardware inventions. For instance, one class includes “arrangements for producing a

permanent visual representation of output data.” This is a functional description that

Allison and Lemley (2000) and Allison and Tiller (2003) also reject the idea of using patent

classifications to identify software patents. According to Griliches (1990), a patent class is “based

primarily on technological and functional principles and is only rarely related to economists’ notions of

products or well-defined industries (which may be a mirage anyway). A subclass dealing with the

dispensing of liquids contains both a patent for a water pistol and for a holy water dispenser. Another

subclass relating to the dispensing of solids contains patents on both manure spreaders and toothpaste

tubes” (Griliches 1990, p. 1666).

PDF to HTML - Convert PDF files to HTML files

includes software programs, hardware computer displays, and even electric and

mechanical signs that pre-date electronic computers.

Still, we did examine the efficacy of using the patent classification approach for

identifying software inventions, as proposed in Graham and Mowery (2003). In that

paper, the authors identified a number of subdivisions of the International Patent

Classification system (IPC) where many patents assigned to large U.S. software

companies may be found.

We compared the list of software patents identified by this

approach with our random sample of 400 patents. The results were significantly worse

than our own algorithm: a 30 percent false positive rate and a 74 percent false negative

rate. We also found that this definition would exclude half of the patents obtained by the

top 200 publicly-traded software firms during the 1990s, and a majority of the pure

software and Internet patents identified in Allison and Tiller (2003). In contrast, our

definition accounts for about 4/5 of all patents obtained by the top 200 software firms in

the 1990s.

Nevertheless, to check whether our main results were robust to our choice of

algorithm, we ran most of our tabular analyses and regressions below using the Graham-

Mowery definition of software patent. Although standard errors were predictably higher,

the results were broadly similar. For example, the distribution across industries was

similar and patents and R&D were found to be statistically significant substitutes.

Although Graham and Mowery’s approach is useful for obtaining a rough impression, it

is not the best technique for a more comprehensive study such as ours, and attempts to

improve its coverage by adding more classes are likely to increase the false positive rate.

To summarize, our algorithm has a positive error rate. For this reason, one can

find patents that our algorithm incorrectly classifies as software patents (see Hahn and

Wallsten 2003). But the rate of false positives is reasonably low and the algorithm

appears to be substantially more accurate than at least one alternative based on patent

classes. Moreover, the evidence suggests that there is no

systematic

bias in these errors:

our main results hold even when we use a very different definition of software patent.

The subdivisions include G06F 3/, 5/, 7/, 9/, 11/, 13/, and 15/; G06K 9/, and 15/; and H04L 9/.

PDF to HTML - Convert PDF files to HTML files

C. The Matched Sample

We also explore the characteristics of the firms that obtained software patents and, in

particular, the relationship between software patenting behavior and firm R&D

performance. To do this, we matched a large portion of both software patents and other

patents to firms in the 1999 vintage of the Compustat database.

Our main population of interest consists of U.S.-owned public firms that perform

R&D. This group performs a large share of domestic R&D and it should provide a

relatively stable group for comparison over time. But it does limit the relevance of our

conclusions to this group, however, and so our analysis has little to say about start-up

firms, individuals, universities, etc.

We begin by matching our patents to firms (i.e. the assignees) using the NBER

Patent Citations Data File (Hall, Jaffe and Trajtenberg 2001a).

This data set matches

patents to the 1989 vintage of firms contained in Compustat, so we do a variety of things

to supplement those matches:

1. We added the largest 25 publicly traded software firms ranked by sales (only

one of which is included in the NBER file).

2. We merged the data set with a set of firm-patent matches provided to us by

CHI Research.

That data encompasses most of the significant patenting

firms (public or private) over the last 25 years.

3. Using data contained in Compustat, we identified 100 of the largest R&D

performers in 1999 that were not already included in our data set. We matched

these firms and their subsidiaries to their patents using a keyword search on

the USPTO web site.

To be precise, we match patent numbers in our data set with those found in the NBER data set.

Where available, we use the firm CUSIP assigned by NBER to obtain financial data from Compustat.

We again match by patent number and use the firm CUSIP assigned by CHI. Details on CHI’s

proprietary data is described at http://www.chiresearch.com/information/customdata/patdata.php3. We are

grateful to Tony Breitzman for sharing this data with us.

PDF to HTML - Convert PDF files to HTML files

The final file held 4,792 distinct subsidiaries and 2,043 parent firms from 1980 to 1999,

an improvement of 1,230 subsidiaries and 305 firms over the NBER database for the

same period.

To test the coverage of this matched sample, we compared it to the target

population in Compustat (that is, all U.S. firms that publicly report financials with

positive R&D). The matched sample performed 91 percent of the deflated R&D in the

Compustat file over this period and accounted for 89 percent of the deflated sales by

R&D-performing firms. Moreover, the coverage ratios are roughly constant over the

entire sample period, varying only a few percentage points in each direction over two

decades. Over this period, the matched sample also accounts for 68 percent of all

successful U.S. patent applications by domestic non-government organizations (mostly

corporations) and 73 percent of software patents granted to these organizations. These

coverage ratios were also quite stable over the two decades.

However, only 37 percent of the R&D-performing firms contained in Compustat

are matched to their patents in our data set and this coverage declined over the sample

period as an increasing number of small firms have gone public since 1980. Thus this

matched sample is broadly representative of the firms that perform most of the R&D and

obtain the majority of patents, but it is not representative of entrants and very small firms.

Nevertheless, given the extent of our coverage of R&D and patents, the results we obtain

in our sample regarding patents and R&D will represent the overall interaction between

patents and R&D.

It is possible that sample selection may bias regression coefficients within the

matched sample. To check for this, we implement Heckman two-stage sample selection

models (below) and find, in general, that sample selection has little effect.

D. Other data

We used the NBER Patent Citations Data File (Hall, Jaffe and Trajtenberg 2001a) to

obtain data on citations received and numbers of claims. To obtain data on employment

The original NBER sample accounted for 47% of the successful patent applications to U.S. non-

government organizations; our sample accounts for 68% of these patents.

PDF to HTML - Convert PDF files to HTML files

of programmers and engineers, we used the Occupational Employment Survey conducted

by the BLS. This source provides detailed occupational employment of 3 digit SIC

industries. Because not all industries were covered in all years of the survey during the

early years, we linearly interpolated employment shares.

We also use a number of input

price indices from the BLS Multifactor Productivity series.

II.

Summary Statistics

Table 1 reports the number of software patents and other patents granted per year and

also the numbers of applications per year, conditional on the applications successfully

resulting in a grant by the end of 1999. As can be seen, their numbers have grown

dramatically in absolute terms and also relative to other patents. Today almost 15 percent

of all patents granted are software patents.

Table 1 also shows estimates of the number of software patents published by Greg

Aharonian.

The overall trends are quite similar and the numbers in recent years are also

quite close. Clearly, our definition of software patents is more inclusive, especially

during the early years.

A. Who Owns Software Patents?

Table 2 shows characteristics of software patents compared with other patents, using data

from the NBER patent database. Software patents are more likely to be owned by firms

than by individuals or government. They are also more likely to be owned by U.S.

assignees and to have U.S. inventors.

After the U.S., the top countries ranked by

Thanks to Joseph Bush of BLS for providing the data. The employment categories we used for

programmers were occupation codes 25102, systems analysts, and 25105, computer programmers; for

engineers, we used 22100 (a group code). Because computer support occupations (25103 and 25104) were

lumped in with the other codes during early years of the survey, we make a proportional adjustment in

those years, reducing the employment counts of programmers by their relative share in the first year in

which all categories were reported.

Aharonian used the “I know one when I see one” criterion (private communication). For many years

Aharonian has written articles on software patents in his

Internet Patent News Service

(www.bustpatents.com). The numbers cited in table 1 are reprinted from p. 319 of Lessig (2001).

Allison and Lemley (2000) find that their sample of software patents has about the average share of

U.S. inventors, although they use a somewhat different method to classify inventors’ national origins.

PDF to HTML - Convert PDF files to HTML files

inventors are Japan (18 percent), Germany (3 percent), Great Britain (2 percent), and

Canada (2 percent).

Consistent with the findings of Allison and Lemley (2000), software patents tend

to receive a larger number of subsequent citations, and they have substantially more

claims per patent. Other research finds that these statistics may indicate greater

private

value, although not necessarily greater social value (Hall, Jaffe and Trajtenberg 2001b,

Lanjouw and Schankerman 2001, Allison et al. 2003). In other aspects, software patents

are similar to other patents.

To obtain more information about the firms that obtain software patents, Table 3

shows firm characteristics by patent type (software vs. non-software) for our matched

sample. Relative to other patents, software patents tend to be obtained by firms with

larger market value, sales, and R&D budgets. They are slightly more likely to be

obtained by newly public firms. Allison and Lemley (2000) also find that software

patents are more likely to be obtained by larger entities as classified by the patent office.

B. The Distribution of Software Patents Across Industries

Table 4 shows the industries of the firms obtaining software patents in the sample

matched to Compustat. Most of the software patents are obtained by manufacturing

firms, especially in the electronics and machinery industries, which include computers.

Software publishers (SIC 7372) acquire only 5 percent of the patents in this sample, and

other software service firms, excluding IBM, account for 2 percent.

The “usual

suspects” for strategic patenting—SIC 35, 36, 38 and IBM—account for 68 percent of

software patents.

The distribution of software patents across industries appears to reflect something

other than the

creation

of software. Columns 2 and 3 include two measures that reflect

software creation: the share of programmers and systems analysts employed in the

industry and the share of programmers, systems analysts and engineers. These are the

occupations most directly involved in the creation of software and so these shares

IBM accounts for 13 percent of the software patents in our sample and it is consistently the largest

software patentee. We break it out separately, as it is not representative of the software services industry.

PDF to HTML - Convert PDF files to HTML files

represent the relative software development effort; the first measure more narrowly than

the second.

The manufacturing sector acquires 75 percent of software patents but employs

only 11 percent of programmers and analysts (32 percent of software writers if engineers

are included). Software publishing and services (SIC 737, including IBM) acquires only

13 percent of software patents but employs 33 percent of programmers and analysts (18

percent if engineers are included). There is little reason to expect software developers

employed in software companies or finance or retailing to be far less productive than

software developers in manufacturing. These large differences suggest that industries

differ dramatically in the degree to which they seek patent protection for their software.

This disparity also appears in the fifth column, which reports the differences in

the simplest measure of patent propensity, the ratio of patents to R&D. Software

publishing firms get only a quarter of the number of patents per dollar of R&D that other

firms obtain. This corresponds to the views expressed by software publishing executives

that software patents are of little value to them (USPTO 1994).

The sixth column

displays a measure of software patent propensity derived from the regression analysis

described below. Overall, software patents are more likely to be obtained by larger firms,

established firms, U.S. firms, and firms in manufacturing (and IBM); they are less likely

to be obtained by individuals, small firms, foreign firms, and software publishers.

III.

The Rising Propensity to Patent Software

From 1987 to 1996 the number of successful software patent applications (granted by

1999) increased 16.0 percent per annum. This growth was greater than any of a number

of yardsticks one might measure it against: during roughly the same period, real

industrial R&D grew 4.4 percent per annum, employment in computer programming

related occupations grew at a 7.1 percent rate, and real business spending on own-

account and outsourced programming grew 7.4 percent per year. This growth occurred

Also, BEA analysis of software investment (Parker and Grimm 2000) implies that about 30 percent

of software is produced as packaged software, the primary product of firms in SIC 7372. Yet this industry

acquires only 5 percent of software patents.

PDF to HTML - Convert PDF files to HTML files

against a general background of rising patent propensity—the ratio of all domestic patent

applications to real R&D—grew about 2.1 percent per year.

We can identify several possible factors that might contribute to increased

software patenting: growth in R&D generally, growth in that portion of R&D which uses

software, greater productivity in software development, and changes in the cost-

effectiveness of software patents from regulatory or other sources. To help sort out the

roles played by these different factors, we use a “patent production function” model of

Hall and Ziedonis (2001).

This production function relates the number of successful

patent applications made by a firm each year to its size, relative R&D spending and other

characteristics, plus a time dummy, which serves to capture residual changes in patent

propensity.

A. Specification

In the preferred specification of this model, the expected number of patents for firm

year

conditional on firm characteristics for that firm and year, is

(1)





Capital

[

]

exp



Employees





Employee





where

is a dummy variable that equals one if the firm is a new entrant and zero

otherwise. The right hand side variables capture the effects of scale, R&D intensity,

capital intensity and new entrant status. The time dummy then captures changes in the

propensity to patent. Differences in the time dummies between two different years

correspond to log differences in the expected number of patents. This is our initial

specification to which we add additional controls, including firm effects and a measure of

software development intensity.

Total industrial R&D increased from $121 billion in 1988 to $164 billion in 1998 in 1996 dollars

(NSF 2003). The BLS Occupational Employment Survey estimates 904,430 employees in “computer

scientists and related occupations” in 1987-89 and 1,839,760 in 1998. The BEA estimates $38billion in

1996 dollars for business spending on own-account and custom software in 1989 and $64billion in 1998

(Parker and Grimm 2000). The general increase in patent propensity has been explored by Kortum and

Lerner (1999) and Hall and Ziedonis (2001).

The literature on patent production functions also includes Scherer 1965, Bound et al. 1984, Pakes

and Griliches 1984, and Griliches, Hall, and Hausman 1986.

PDF to HTML - Convert PDF files to HTML files

Since the left hand side variable is a count and many observations are zero, the

equation can be estimated with a Poisson regression, as is frequently done in the

literature. The Poisson regression assumes that the variance equals the expected value of

the left hand side variable, but often Poisson specifications fail to meet this assumption

and show “over-dispersion.” Tests using a negative binomial specification reveal over-

dispersion in our data too, and, following Hall and Ziedonis (p. 113), we take this as an

indication to use heteroscedastic-consistent standard errors.

Our analysis differs from Hall and Ziedonis in two important ways, however.

First, our dependent variable is not all of the patent applications of the firm, just software

patent applications. In principle, this should cause no problem—one still expects size,

R&D intensity, etc. to affect the level of software patents. But we may also want to

control for the degree to which the firm directs resources to software development, which

we do below. Second, Hall and Ziedonis study a narrowly defined industry whereas we

study a broad range of industries. For this reason, we will want to control for industry or

firm effects in some of our regressions.

In our base specification

is the number of software patents applications by firm

in year

that resulted in a patent granted by 1999.

Because patent prosecution

typically takes two years or so, we conduct our regressions through 1997. The other

variables are as follows: employees is the number of employees listed in Compustat in

thousands, R&D is deflated by the GDP deflator, and capital is property, plant and

equipment deflated by the NIPA capital goods deflator. The “new firm” dummy is equal

to one for the first five years a firm appears in Compustat.

B. Results

Column 1 shows the base regression. Column 2 adds seven industry dummies and

variables to capture the relative use of programming personnel. The first of these

variables is the ratio of “computer scientists and related occupations” to total industry

employment in the BLS Occupational Employment Survey (OES) for the firm’s SIC 3

PDF to HTML - Convert PDF files to HTML files

digit industry (or 2 digit if Compustat assigns the firm only two digits). The second

variable is the comparable share of engineers in total employment. The OES was

conducted on a three year cycle until 1995, so values for this variable in intervening years

are linearly interpolated. And for comparability, we only used OES data from 1987

onwards, so the second column covers a shorter interval.

The elasticity of the scale variable, employment, is similar to estimates in

previous studies. Hausman, Hall and Griliches (1984) obtained an elasticity of .87 on

R&D. Hall and Ziedonis obtained an elasticity of employment of .85 when they included

some firm controls. The coefficient of the R&D intensity variable is, however, much

higher than in Hall and Ziedonis, although when we include firm fixed effects below, this

difference largely disappears, suggesting that it is picking up unmeasured firm/industry

heterogeneity. Capital intensity (in column 2) is also significant and similar in magnitude

to one estimate by Hall and Ziedonis, but smaller than another. Hall and Ziedonis

interpret this coefficient as evidence that capital intensive firms may patent more because

they are subject to holdup by rivals who patent strategically. That is, this is evidence of

“defensive patenting.” Our result suggests this hypothesis might apply more generally.

1. New vs. Old Firms

We find that the “new firm” dummy variable is not significant. In contrast, Hall and

Ziedonis find that their dummy variable is highly significant and has a relatively large

coefficient.

They suggest that their result arises from new semiconductor design firms

that need patents to secure financing (see also Hall 2003). These smaller firms do not

have complementary manufacturing facilities that may provide another means of

appropriability.

In our data, we only observe patent applications that are successful. After November 2000, patent

applications are generally published 18 months after the filing date. But publication is not required if the

inventor does not seek patents abroad.

There is also a difference between the definition of our new firm variable and theirs. Hall and

Ziedonis include all firms that entered Compustat after 1982. We include firms that entered in 1982 or

later, but our dummy variable equals one only for the first five years of entry. Their initial interest was to

compare whether incumbent firms in particular benefited from the creation of the Court of Appeals for the

Federal Circuit. Our interest is whether entrants appear to find additional benefits from software patents,

something Hall and Ziedonis explored specifically for entrant design firms with an additional dummy

variable.

PDF to HTML - Convert PDF files to HTML files

Although our sample may not be representative of the entire population of new

firms, this regression does include a reasonably large sample of new firms (1,308

observations during the first five years of 314 firms). Hence our different result suggests

that either software patents are not comparably useful for obtaining financing in general

or this “vertical disintegration” strategy is not broadly relevant outside the semiconductor

industry, or both.

We explored this issue further by interacting the new firm dummy variable with

industry dummy variables. Only one of these interaction terms was statistically

significant, the one for SIC 73, business services, which in our sample largely consists of

software services and pre-packaged software firms. Column 3 shows a regression with

just this interaction term added. The coefficient is negative and statistically significant,

suggesting that new software firms obtain

fewer

software patents than established firms.

This suggests that new software firms may not obtain the same benefit from patents as do

new semiconductor firms.

2. Cross Industry Variation in Software Patent Propensity

Returning to Column 2, the coefficients on the industry dummies indicate large inter-

industry differences in software patent propensity even after controlling for industry

employment of programmers and engineers. Given the exponential specification, (1),

differences between industry coefficients correspond to differences in the log of software

patent propensity. Column 6 of Table 4 shows corresponding differences in software

patent propensity itself (that is, the exponential of the coefficients), normalized so that

the software patent propensity of SIC 73 equals one (this includes IBM). As can be seen,

the differences are quite large, with SIC 36 and SIC 38 obtaining an order of magnitude

more software patents, all else equal, and machinery, SIC 35, not too far behind. In

general, the industries that have a high propensity to patent software also have a high

patent propensity in general (see column 5 of Table 4), although the differences in

We also ran these regressions using the total number of patents, not just software patents, as the

dependent variable and obtained similar results (not shown). Graham and Mowery (2003) report that the

software patent propensities of established software publishers rose over the 1990s, but for entrant firms

(those founded after 1984) there was no discernable trend.

PDF to HTML - Convert PDF files to HTML files

software patent propensity are larger. In the literature, these are the industries known for

strategic patenting (see section IV).

The magnitude of these differences and the known importance of strategic

patenting in electronics and computers, suggests that these results may be the outcome of

strategic patenting. To test this idea further, column 4 drops the industry dummies but

includes instead a measure of the degree to which other firms in the industry patent. This

variable is the number of all patents obtained by other firms in the same 2 digit industry

as the observed firm, divided by the employment of those firms. The positive and

significant coefficient suggests that firms obtain more software patents in industries that

patent more overall, all else equal. This result is consistent with “defensive” patenting,

although it could also arise from other industry differences, such as large differences in

alternative means of appropriability.

3. Unobserved Heterogeneity

Our seven industry dummy variables do not capture the full extent of inter-industry

heterogeneity or other firm heterogeneity. This was confirmed by tests using the random

effects and fixed effects Poisson models in Hausman, Hall, and Griliches (1984). We then

face the choice of better controlling for this unobserved heterogeneity via a random

effects or a fixed effects model. A Hausman test rejects the null hypothesis that the

random effects estimates are consistent; that is, the firm effects appear to be correlated

with the coefficients (P = 0.000), indicating that fixed effects are preferred.

Column 5 presents the fixed effects Poisson regression. This is a conditional

maximum likelihood regression where the likelihood of each observation is conditioned

on the likelihood of the observed sum of software patents for each firm over all years in

the panel. This is only calculated for firms that have software patents in at least one year.

Since many firms obtain no software patents, the sample size is reduced considerably for

this regression. With fixed effects, the scale elasticity is somewhat smaller, and is

consistent with earlier research (Hausman, Hall, and Griliches 1984) and the R&D

intensity coefficient is now much more in line with the estimates in Hall and Ziedonis

(2001).

PDF to HTML - Convert PDF files to HTML files

In Figure 1, we plot the year dummies from this regression (normalized to equal 1

in 1987), those from the Column 1 regression, which spans a longer time period, and the

corresponding year dummies from Hall and Ziedonis’s preferred specification. As can be

seen, after the mid-80s, all three series grow rapidly, persistently and at about the same

rate. Compared to the rate for 1987, and holding all other factors constant, firms were

successfully applying for nearly 50 percent more software patents in 1991, and 164

percent more by 1996.

4. The Ebb and Flow of Copyright Protection

One factor that may explain the large increase in software patent propensity is

diminishing protection for computer programs provided by copyright. Lemley and

O’Brien (1997) describe the rise and fall of the “nonliteral infringement” doctrine.

Established in the 1986 decision in

Whelan Associates v. Jaslow Dental Laboratories,

gave copyright holders some protection over the features of their software programs. But

the doctrine was rejected in the 1992 decision in

Computer Associates International v.

Altai.

If the alternative of copyright protection affected firms’ propensity to obtain

patents, then one might expect a decrease in patent propensity after 1986, and then an

increase after 1992. No significant fluctuation is observed in Figure 1 and regression tests

for a break find only a very small change in trend.

Moreover, the industries that patent

most heavily also tend to use embedded software, so copyright protection was less

relevant to them.

C. Interpretation

We use the fixed effects specification to interpret the relative influence of various right

hand variables. Because (1) is exponential in form, changes in variables times their

coefficients affect the log of software patents and can be interpreted as growth rates. To

calculate annual growth rates, we evaluate the mean of each variable in 1996 and subtract

797 F.2d 1222 (3

Cir 1986) and 982 F.2d 693 (2d Cir 1992), respectively.

We used a regression similar to column 5 of Table 5 with a time trend instead of year dummies. We

interacted the time trend with an interval dummy for the years 1986-92 and found a very small (-.002), but

statistically significant reduction in the rate of growth in software patent propensity during this interval.

This effect is on the order of 5% of the overall growth rate in software patent propensity.

PDF to HTML - Convert PDF files to HTML files

the mean in 1987. We then multiply this difference by the estimated coefficient from

column 5 and divide by 9 (the number of years elapsed) to obtain an annual contribution

of the variable to the growth of software patents for our sample. The resulting estimates

are shown to the right of column 5. A similar calculation for the year dummies is shown

for all columns in a separate row.

The annual growth of software patenting among the firms in this sample (which

excludes firms without any software patents over this entire period) is 16.4 percent,

slightly larger than the aggregate 16.0 percent reported above. The majority of this

increase is captured by the contribution of the year dummies (10.8 percent). The next

largest contribution is from greater capital intensity (1.8 percent) followed by the growth

in programmer employment (1.2 percent), engineering employment (1.2 percent) and

R&D intensity (1.1 percent).

Thus the majority of the growth in software patenting is not attributed to any of

these explicit controls and can be attributed, instead, to rapidly rising patent propensity.

Note that this increase in patent propensity is quite close to the result obtained by Hall

and Ziedonis (2001) for all patents in just the semiconductor industry. A comparable

calculation on their preferred specification also results in a 10.8 percent annual growth in

patent propensity from 1987 – 95.

But such growth rates are far from typical for total

patenting in most industries.

1. Taking into Account Productivity Growth

In principal, software patent propensity can be decomposed into two factors: an increase

in the productivity of software developers, and change in the cost-effectiveness of

patenting.

The first factor means that software developers may produce more inventions

using the same level of inputs. The second means that firms find it more attractive to

patent a higher proportion of inventions, perhaps because the cost of doing so has fallen,

or because such patents provide larger benefits than they did in years past, or possibly

both.

Thanks to Rosemarie Ziedonis for graciously sharing data.

See Hall and Ziedonis (2001) for a similar discussion.

PDF to HTML - Convert PDF files to HTML files

A thumbnail calculation suggests that changing cost-effectiveness is the major

factor accounting for the rise in patent propensity. The productivity of software

developers is notoriously difficult to estimate. Although there has been strong market for

software development tools, the actual contribution of these tools to productivity is

widely debated and studies find that computer aided software engineering tools often go

unused (Kemerer 1992). Nevertheless, we can calculate an upper bound for programmer

productivity growth using price indices for the pre-packaged software industry. As we

discuss below, these estimates are likely to exaggerate productivity growth for software

developers involved in the R&D process.

From 1992 to 1997 receipts of the pre-packaged software industry (SIC 7372)

grew 18.5 percent per annum while employment grew 14.2 percent per annum (Census

2003). From the OES survey, we know that within SIC 737, employment of computer

science related occupations grew 2.4 percent faster than total employment, so we

estimate that programmer employment in pre-packaged software grew 16.6 percent per

annum. Thus nominal growth of software revenue per programmer was about 1.9 percent

per annum, less than a fifth of the estimated increase in software patent propensity.

But we must also take into account improvements in the quality of software

produced. A recent literature has developed price indices for pre-packaged software using

both matched-model methods and hedonic methods (see Grimm and Parker 2000, and

Abel, Berndt, and White 2003). The BEA has a matched model estimate of a -3.5 percent

price change per annum over this period and this corresponds well to other matched

model estimates. Hedonic price estimates, based on a small number of applications,

usually show more rapid price declines, although some economists have questioned

whether these estimates are reliable (Oliner and Sichel 1994). The BEA makes a seat-of-

the-pants quality adjustment to come up with a pre-packaged software price index that

has a growth rate of -6.8 percent over this same period. Combining these estimates, real

pre-packaged software per programmer grew either 5.4 percent per annum (using the

matched model index) or 8.7 percent per annum (using the quality adjusted index).

These figures are clearly less than the 10.8 percent annual rate of patent

propensity growth. But they very likely substantially overstate the growth in programmer

PDF to HTML - Convert PDF files to HTML files

productivity

per se.

The BEA, in fact, attributes

most

of estimated price decline to growth

in the market for software driven by rapidly falling prices for complementary computer

hardware (Grimm and Parker 2000, p. 6). With the total market growing rapidly each

year and small marginal costs of production, much of the estimated price decline is

attributable to economies of scale. But what matters for patent propensity is the

productivity of software developers at actually creating new software (e.g., new titles of

pre-packaged software) rather than duplicating software that already exists. Subtracting

the large scale economies, this productivity growth rate must be substantially less than

5.4 percent or 8.7 percent per annum. At the very least, growing cost-effectiveness of

patents accounts for some substantial portion of the rise in software patent propensity and

may well account for most of it.

D. The Cost-Effectiveness of Software Patents

Software patents may have become more cost-effective because of the regulatory changes

described in Section I. Eliminating the subject matter exclusion and reducing the non-

obviousness and enablement requirements may have made software patents much easier

(less costly) to obtain. Stronger enforcement and a greater presumption of validity may

have increased the appropriability each patent delivered. Both served to decrease the cost

of appropriability.

Strategic patenting may have amplified the effect of these changes. Reductions in

the cost of appropriability may encourage firms to pursue more aggressive strategic

behavior (Bessen 2003). This may, in turn, induce other firms to engage in defensive

patenting, further increasing the patent propensity. Both the importance of capital

intensity and the large industry effects found above suggest that strategic patenting may

play an important role and that this role may have increased over time.

In summary, the outsized growth in software patenting is not adequately

explained by research inputs, productivity growth, or other observable factors. There is a

possibility is that alternatives to patents became less effective, increasing the relative cost

effectiveness of software patents. But survey evidence suggests that trade secrecy did not decrease in

importance and may have increased (Cohen, Nelson, and Walsh 2000). We address changes in the efficacy

of copyright protection in section III.

Another

PDF to HTML - Convert PDF files to HTML files

significant residual trend in software patent propensity on the order of 5 percent to 10

percent a year, roughly 2.5 to 4 times larger than the trend for patents in general. Such

increases would be consistent with firm responses to regulatory changes that have

increased the appropriability of software patents, made them easier to obtain, or both.

Moreover, the pattern of large inter-industry differences in patent propensity (and the

importance of capital intensity) are consistent with models of strategic patenting

IV.

R&D and Software Patents

The traditional incentive theory suggests that more cost-effective patents for software

should lead to increased R&D spending, all else equal. With “cheaper” and/or “stronger”

patent protection, firms should patent more software inventions and/or realize greater

profits from the software inventions they patent. Expecting greater profits, firms should

find it profitable to spend more on R&D, all else equal. If the incentive hypothesis is

correct, the increase in the cost effectiveness of software patents identified in the

preceding section should be positively correlated with increases in firms’ R&D

investments. In this section, we test this hypothesis.

To do so, we must control for other factors so that, to the extent possible, we hold

“all else equal.” In a dynamic world with shifting demand and supply and changing

prices, firms may increase or decrease their R&D spending for reasons unrelated to

changes in the appropriability of software patents. For example, a firm facing growing

demand may increase R&D because increases in demand may make heretofore

marginally unprofitable R&D projects profitable. This renders any simple correlation

between software patenting and R&D spending unpersuasive.

Of course empirical economists have developed techniques for analyzing the

relationship between two quantities against a backdrop of shifting prices, namely, by

estimating elasticities of factor substitution (see Berndt 1991 for a general review). These

techniques allow economists to identify whether two factors are complements or

substitutes. The problem here is quite similar if one considers R&D and patents to be

factors of production (for the production of profits, rather than physical output). In other

words, the incentive theory is akin to saying that patents and R&D are

complements.

PDF to HTML - Convert PDF files to HTML files

Greater use of patents should be associated with greater use of R&D. And if patents

become more cost effective—that is, if their quality-adjusted price falls—then the share

of resources allocated to R&D relative to other factors of production should increase.

A. A Simple Model

To see that the incentive argument is equivalent to complementarity of R&D and patents,

it is helpful to consider a simple generalization of the formal patent race models that have

been used as the basis of the incentive hypothesis. In these models, the more R&D a firm

does on a project, the greater the probability that the firm will make a successful product

innovation (first). Abstracting away from issues of timing and taking the actions of other

firms as given, let

P(r)

be this probability of success. Since R&D activities on this project

will have diminishing returns, we assume that

is increasing, concave and that

(0)

0 .

Let

be the quantity of R&D performed at price

If the firm successfully innovates and gets a patent, then it earns a discounted

stream of rents during the term of the patent (and perhaps some profits afterwards). If the

discounted stream of production costs associated with the product is

then the profits

equal

⋅

, where

is the

markup above cost.

The expected profits on the project are

(2)

⋅

(

)

−

w r

In the patent literature, policy features such as patent term and scope affect

expected profits, and these features correspond to different levels of

In other words,

can be interpreted as a measure of appropriability.

To the extent that a project can be

protected by software patents, stronger, more cost-effective patents should increase

The firm’s optimal level of

can be found by examining the first order condition of (2).

Then straightforward calculation shows that increases in

generate increases in the

equilibrium level of

Moreover, if the scale of production remains relatively constant,

then the ratio of R&D expenses,

⋅

, to production cost,

, will also increase.

It is theoretically possible, however, that the scale of production for a successful

project may increase or decrease with appropriability—for example, an increase in

See Arora, Ceccagnoli, and Cohen (2003) for a similar decomposition taken to the data.

PDF to HTML - Convert PDF files to HTML files

might be associated with a change in market structure from a duopoly to a monopoly.

Diverse empirical evidence suggests that substantial changes in production scale are

unlikely in the broad range of industries studied here. Cohen and Klepper (1996) argue

that “the level of output over which rents from R&D are realized is closely related to the

firm’s output at the time it conducts its R&D.” More generally, in many industries, firms

achieve appropriability mainly through lead time, learning-by-doing, complementary

services, etc. and patents only play a secondary role (Levin et al. 1987, Cohen, Nelson,

and Walsh 2000). Any improvement in appropriability conveyed by patents is unlikely to

have a drastic effect on market structure in these industries because firms

already

have

significant appropriability. In Table 6 below, we find that software patents did not appear

to influence profit margins significantly, also suggesting that industry structure is

unlikely to change dramatically in response to the availability of software patents.

A positive association between appropriability and R&D cost share is a necessary

condition for patents and R&D to be Hicks-Allen complements. It also fits nicely with a

standard technique for evaluating factor demand by using cost share equations. This

technique uses relative cost shares (the cost of an input factor relative to total production

costs) regressed against log relative prices of input factors. For example, the ratio of

energy costs to total production costs is regressed against log relative prices of labor,

capital, energy, etc. These equations are derived from a flexible form translog cost

function (see Greene, 1997, Chapter 15.6). For a regression that uses one input factor as a

dependent variable (say, energy), if the relative price of another factor (say, labor) has a

positive coefficient, then these two goods cannot be complements.

Moreover, this test

takes into account both the direct effect of price changes on consumption of the left hand

side good (the substitution effect) and also the overall effect of the price change on total

costs (the income effect).

B. Estimation

To apply this technique to our problem, we can treat both R&D and patents as input

factors in addition to conventional input factors. Then, following the standard analysis, a

To be complements, the coefficient must be less than

(

−

∗

the product of the two cost shares.

PDF to HTML - Convert PDF files to HTML files

regression equation can be derived from a procedure of minimizing costs while holding

revenues constant. This equation would be

(3)

⋅

∑

⋅

where

is total cost (redefined here to include R&D costs as well as production costs),

is the “cost of appropriability” (the quality-adjusted cost of patents, not observed),

are prices of other input factors (taking one as numeraire), and the Greek symbols are

coefficients to be estimated.

This technique is preferred to estimating quantities when

analyzing firm-level data under the assumption that prices are likely to be exogenous to

the firm (Berndt 1991).

Since we do not observe the cost of appropriability, we do the following: We

include time dummies in the regression which will soak up, among other things, overall

changes in the cost of patenting and the “strength” of patents. We believe that a

substantial part of the rise in software patenting is a response to increases in the

“strength” of patents for software inventions and/or decreases in the cost of patenting

these inventions (see section III). In that case, the share of software patents in a firm’s

total patenting, denoted

serves as a proxy that captures variations in the relative cost of

appropriability across firms and over time.

Taking differences to sweep out firm fixed effects, our regression equation is

(4)

∆

⋅ ∆

∑

⋅ ∆

itj

We use five-year differences to reduce noise from measurement error and any biases

associated with partial factor adjustment.

We calculate the R&D cost share as the ratio of R&D spending to total costs

measured as the sum of cost of goods sold and selling, general and administrative

Typically, a system of equations is jointly estimated for each input factor (but one). Here, we

estimate only a single equation, in differences, because we do not have full information on other factor

inputs. Our estimates are still consistent, but not as efficient as they would be if we could estimate multiple

equations.

PDF to HTML - Convert PDF files to HTML files

expenses. Since not all firms report these cost items, we also use an alternative measure,

R&D/sales (see below).

The software share of patents,

is measured as the number of software patents

applied for in a given year divided by the total number of patents applied for that year,

conditional on the patents having been issued by 1999. Since patents take on average

about two years to issue, we only use patent applications through 1997. Also, this

software patent share measure is subject to sampling error—this measure will have large

variance for firms with small numbers of patents. To reduce heteroscedasticity, we use

Weighted Least Squares and we also report heteroscedastic consistent standard errors.

Below, we check that our results hold without this weighting. We also check for possible

correlation between

∆s

and the error term.

The price variables are annual industry price indices from the BLS Multifactor

Productivity report at the two-digit level for manufacturing and for the private non-farm

business sector for non-manufacturing firms. In addition to capital, labor, materials,

energy, and purchased services, we include an index for the price of IT capital, to control

for the influence of IT.

To check that our results did not depend on any possible errors

in these prices, we also ran our main regressions using a full set of interacted year and

two-digit industry dummies. The coefficients of interest changed only slightly.

C. Results

Column 1 in Table 6 shows a regression of equation (4) for the years 1991-97. The

coefficient on software patent share is negative and significant at the 5 percent level. A

negative value indicates a negative conditional correlation between our measure of the

declining relative cost of appropriability and the R&D cost share. That suggests these two

inputs are substitutes and not complements.

The weight used is

(

−

)

since the sampling variance for

is proportional to

−

We use data from the April 2001 release. Thanks to Bill Gullickson and Steve Rosenthal of BLS for

providing the data. We also used separate indices for the components of IT capital, but these did not

change the non-price coefficients and they were highly multi-collinear.

PDF to HTML - Convert PDF files to HTML files

As noted above, the cost variable is missing for many observations. Firms

reporting cost of goods sold tend to fall disproportionately in manufacturing and this

might bias results. Column 2 tests the significance of this sample selection and also of

firms not matched to the patent file by running a Heckman sample selection model. This

model consists of a regression equation and a probit equation (not shown) estimated on a

binary variable indicating whether the observation is included in the regression reported

in column 1. The regression coefficients are adjusted to reflect biases arising from the

sample selection process. For the probit, we used log market value, log deflated sales, a

new firm flag and year dummies on the right hand side. The probit coefficients are highly

significant and a likelihood ratio test indicates the probit regression as a whole is highly

significant. While a Wald test rejects the null hypothesis that the two equations are

independent, any resulting bias in the coefficient of software patent share seems small.

Estimated with the sample selection correction, the coefficient remains negative and is

highly significant.

Another way to test the incentive hypothesis is to see whether software share

changed the markup over cost. To do this, column 3 regresses the change in log deflated

sales against the change in software share and the change in log deflated cost. The

coefficient of the change in log cost is not significantly different from 1, consistent with

constant returns to scale. If software patents increased profit margins, then in this

regression the software share should have a positive coefficient. The estimated

coefficient is positive but quite small and not statistically significant. It appears the

software patents did not have any substantial effect on markups.

Although sample selection does not seem to indicate a substantial bias, our

regressions will be more representative, and estimated more precisely, if we expand the

sample size. We can do this by using the ratio of R&D to sales—R&D intensity—as the

dependent variable instead of R&D cost share. So long as firm profit margins do not

change dramatically, the ratio of R&D to sales will remain roughly proportional to the

ratio of R&D to cost. We verified that the ratio of cost to sales does not change much on

average (there is a variance of about .001 per year) where cost data is not missing. Also,

the previous regression shows that changes in profit margin are uncorrelated with

PDF to HTML - Convert PDF files to HTML files

changes in software share. This means that using R&D intensity instead of R&D cost

share will not bias the coefficient of software share. Finally, to eliminate cases where

profit margins were likely to change dramatically, we trimmed a small number of

observations where R&D spending exceeded half of annual sales.

Column 4 of Table 6 repeats the basic regression equation using the change in

R&D intensity as the dependent variable. The sample size is almost four times larger and

standard errors are about one third the size of those in column 1. The coefficient on

software share is quite similar. Further, in column 5 we repeat the Heckman sample

selection model regression. Here the null hypothesis that the sample selection equation is

independent of the regression equation cannot be rejected. Thus using the change in R&D

intensity as the dependent variable appears to provide a more representative and larger

sample. For this reason, we use this dependent variable in the subsequent analysis.

Our historical and legal research suggests that the substitution of patents for R&D

might be explained by changes in the cost/benefit of software patents—that these patents

were made “stronger” and/or cheaper during the 90s. Consistent with this view, we

conducted the previous regressions from 1991 to 1997. To check that this time period is

not arbitrary, column 6 repeats the regression of column 3 over a longer period with the

change in software share interacted with three time dummies. These results show that the

substitution effect follows the same time pattern as was observed for patent propensity in

the last section. There was a positive, but insignificant effect during the late 80s, and

negative and highly significant coefficients during the 90s. The magnitude of the

substitution effect appears to increase in the late 90s, but the difference from the early

90s is not statistically significant.

The observed substitution effect is economically significant. Thumbnail

calculations suggest that by the end of the 1990s, R&D would have been about 10

This screen eliminated 121 observations in total of some 39 firms. These firms were mainly startups

(34), mostly biotech (25) and each of these firms had observations for other years that were included in the

sample. To verify that this trimming did not introduce a bias, we performed a Heckman sample selection

analysis on just the selection created by this trimming. The null hypothesis that the sample selection

equation (with log market value, log sales, new firm flag, R&D intensity, lagged R&D intensity, and year

dummies as regressors) was independent of the regression equation could not be rejected (P = .859).

PDF to HTML - Convert PDF files to HTML files

percent higher without the substitution. This is equivalent to about $16 billion in private

R&D, or roughly eight years of the long run average increase in R&D intensity among

American firms.

1. Robustness Checks

To verify that our results were not unduly influenced by the use of weighted least

squares, we repeated the regression in column 4 using ordinary least squares but

eliminated observations for years where a firm filed fewer than five patents. This left

1,933 observations and a coefficient estimate for

∆s

of -.010 (.005), suggesting that the

substitution effect exists without the weights, but WLS estimates are more efficient.

It is possible that

∆s

might be correlated with the error term because R&D and

patenting decisions are made simultaneously. To test this, we instrumented

∆s

with

lagged values and re-ran the regression in column 4 as an IV regression. When we

instrumented with one and two year lags, the resulting coefficient was -.103 (.033),

suggesting that any bias might actually be downwards. Comparing the instrumented and

non-instrumented regressions, a Hausman test could not reject the null hypothesis that the

uninstrumented version is consistent with a P value of .999 (

(13) = 1.74). So

endogeneity of

∆s

does not appear to be a problem. Similar results were obtained using

just a two-year lag.

To make sure that the substitution result was not influenced by possible error in

the BLS price series we used, we replaced the changes in log prices with terms for the

year dummies interacted with 2 digit industry dummies.

Thus, we picked up all industry

trend variables in addition to sweeping out individual firm effects through differencing.

Given the large industry effects shown in Table 5, undoubtedly some of the variation in

∆s

was picked up in these interaction terms, reducing the coefficient on

∆s.

As expected,

We also performed all of the following analysis using the change in R&D cost share as the

dependent variable. Results were generally quite similar, but with larger standard errors

A 10 percent increase in software share over the decade times a coefficient of -.037 divided by mean

R&D intensity of about 3.5 percent. The dollar estimate is based on the level of industrial R&D in 1998

(NSF 2003). The average annual increase in the ratio of industrial R&D to net sales, as reported by NSF, is

.04 percentage points.

Several industries had very few firms, so we combined seven 2 digit industries into other industries.

PDF to HTML - Convert PDF files to HTML files

the estimated coefficient was a bit less, but still significant at the 1 percent level,

-.025 (.010). Thus the substitution effect appears robust to a variety of estimation

concerns.

2. Industry

Column 1 of Table 7 considers whether our results may be limited to one or two

industries by interacting the change in software share of patents with industry dummies.

We find a significant negative substitution effect for SIC 35 (which includes the

computer industry), SIC 36 (electronics, including semiconductors), and SIC 73

(business services, including software and IBM).

For other industries, the coefficient is

not significantly different from zero.

The substitution effect thus seems to occur mainly in those industries known for

strategic patenting. Column 2 divides the industries into strategic ones (SIC 357, 36, 38

and IBM) and the rest. Firms in the strategic patenting industries exhibit a large and

highly significant substitution effect; for firms in other industries, the coefficient is not

significantly different from zero, indicating neither a substantial substitution nor

complementary effect.

3. Outsourcing Software Production

The use of software throughout the economy grew rapidly during the 1990s and so it is

important to examine the possibility that the increasing ubiquity of software may

somehow spuriously cause the correlations observed in Table 6. There are two potential

mechanisms to consider: the use of purchased software within the firm generally, and the

use of software in the R&D process itself.

Investment in software also grew rapidly during the 90s, especially in the form of

purchases of pre-packaged software (see Grimm and Parker 2000). It is possible that as

firms purchased more software from other companies, they reduced the internal

If IBM is excluded, the coefficient on SIC 73 is no longer significantly different from zero.

PDF to HTML - Convert PDF files to HTML files

production of software, some of which may have been classified in the R&D budget.

This would have reduced R&D spending and R&D intensity.

Two arguments militate against this explanation. First, one would expect that if

firms are developing less software internally, they are less likely to obtain software

patents. In that case, software patent share would not rise as much, or might even fall.

Then the correlation between changes in software patenting share and R&D intensity

would likely be

positive.

Second, our regressions control for externally purchased IT,

which includes software (both outsourced and pre-packaged).

To test this further, we

also performed the regressions with separate price indices for computers and software

(these are components of the BLS IT price index) and found essentially the same results.

So the use of purchased or outsourced software does not seem to explain the negative

association between software patent share and R&D intensity reported in Table 6.

4. The Use of Software in R&D

The second possible mechanism for a spurious correlation is an increase in the

importance of software in the R&D process itself. Firms employ software to design

products, both by using tools, like computer-aided engineering software, and by using

embedded software in hardware products that were formerly hard-wired. In other words,

software may reduce the cost of performing R&D. So in principle, it is possible that firms

that use software to design new products might obtain patents on this software and, under

certain conditions, they might also spend less on R&D.

There are several reasons to suggest that this does not explain the negative

correlation we observe. First, this explanation only works if the demand for R&D is

relatively inelastic. If software use reduces the cost of performing R&D, then R&D

projects that were marginally unprofitable before become profitable. If there are a

sufficient number of marginal projects, then the lower effective cost of each project will

In Compustat, software development spending is included in R&D, unless it is paid for by a

customer.

The coefficient on the change in the log of the IT price index is consistently positive, indicating that

IT does substitute for R&D. Dropping this variable from the regression only changes the coefficient of

slightly. That suggests the relationship between utilization of IT and R&D is different from the interaction

between software patents and R&D.

PDF to HTML - Convert PDF files to HTML files

be offset by the increase in the number of projects performed. That is, if R&D demand is

sufficiently elastic, any reductions in the effective cost of conducting R&D would cause

firms to

increase

R&D intensity. We know from studies of the R&D tax credit that the

demand for R&D is rather elastic with respect to its tax price. Berger (1993) finds that

reductions in the tax price of R&D

increased

R&D intensity. In a survey of the literature,

Hall and van Reenen (2000) find that the tax price elasticity of R&D is around -1; at this

level, changes in the price of R&D should have little effect on R&D intensity. Thus, this

evidence suggests that the demand for R&D is sufficiently elastic so that software use

cannot explain the negative correlation.

Second, we can also control for software use. If increased use of software in the

R&D process is causing the negative coefficient on the change in software share, then

including the change in the relative employment of programmers and engineers in the

regression should reduce or eliminate this association. Column 3 of Table 7 repeats the

R&D intensity regression, but includes the change in the share of programmers and

systems analysts in industry employment and the change in the share of employment of

engineers. The coefficient of

∆s

hardly changes, indicating that the substitution effect is

orthogonal to the employment of software writers.

Third, it is hard to reconcile this explanation with the cross-industry pattern of the

substitution effect. Software is used in the design and development of new products and

processes in a wide range of industries, some of them strategic, such as computers and

electronic equipment, and some in the non-strategic group, such as automotive, chemical

processes and non-computer machinery. But although both groups of industries use

software in products and in the design of products, the substitution effect is observed

only in the strategic group. It seems unlikely this pattern is simply coincidence.

The coefficient for programmers is negative and significant. One interpretation of this is that firms

shifting from manufacturing to services may reduce their R&D and increase their IT resources, resulting in

an increase in the employment share of programmers. The mean change in programmer employment share

is .001, so the net effect is not very large. Moreover, using interaction regressions, we found that this effect

is largely limited to non-manufacturing industries.

PDF to HTML - Convert PDF files to HTML files

5. Variation by Firm Size, Incumbency, and Other Factors

Some researchers suggest that stronger patent protection for software may facilitate

vertical dis-integration. In particular, small firms may have increased their software R&D

in order to acquire more patents to license or sell to large firms. If this were so, software

patents might complement R&D at small firms. On the other hand, if software patents are

associated with strategic patenting, then small firms might be forced to acquire

“defensive” patents and they might also face lower appropriability. Then R&D would

substitute for software patents at small firms.

Column 4 explores the effect of firm size by interacting the change in software

share with dummies for firms of small and medium employment size (less than or equal

to 1,000 employees and greater than 1,000 but less than or equal to 10,000).

The

coefficients on these interaction terms thus indicate differences between the smaller size

classes and large firms. Large firms exhibit the strongest substitution effect, but the

effects for the smaller classes are not significantly different statistically. And the smallest

size class hardly differs from the largest class. These results do not support the view that

the substitution is purely a large firm effect or that software patents have encouraged

greater R&D among small firms.

Column 5 looks at the related question of differences between incumbent and

entrant firms. It adds an interaction of the change in software share with a dummy

variable that is 1 if the firm entered Compustat after 1985. There were 297 observations

where this dummy was set in this regression. Here again, the substitution effect for

entrants differs only slightly from the effect for incumbents and this difference is not

statistically significant. We repeated the regressions in Columns 3 to 5 for just the group

of firms in the strategic patenting industries and obtained similar results.

We also conducted other regressions (not shown), which included additional

controls for firm size (lagged log deflated sales), new firms, firm risk (standard deviation

of annual stock price growth), and liquidity (change in long-term debt divided by

capital). Including these control variables had little effect on the coefficient on software

PDF to HTML - Convert PDF files to HTML files

share. Finally, we considered the role of accounting changes but concluded this could not

explain the observed relationship between software patents and R&D intensity.

Interpreting Our Results

While our empirical technique cannot identify the causal relationships, it does reveal a

number of patterns we can compare to a variety of theories about the effects of extending

patent protection to computer programs. For example, the facts seem difficult to reconcile

with the traditional incentive hypothesis.

Our results make quite clear that software patenting by and large has little to do

with the pre-packaged software industry. That industry acquires a very small fraction of

all software patents. Although it has been argued that software patents may encourage

entry into an industry dominated by powerful firms, software entrants have a relatively

low software patent propensity. However, we do not have a comprehensive picture of all

entrant firms.

Two-thirds of all software patents are obtained by a small group of industries

known for strategic patenting (SIC 35, 36, 38 and IBM). Moreover, this group appears to

play a unique role throughout our analysis. This group has a much higher propensity to

obtain software patents and, in the 1990s, appears to have substituted these patents for

R&D. For firms in other industries, there does not appear to be any relationship between

an increasing focus on software patenting and R&D investments.

We think the central role of this group of industries is not accidental—that is, the

phenomenon we observe with software patents may be the

same phenomenon

other

researchers have observed for patenting in general for these industries:

Of the 2,991 observations, 432 have employment less than 1,000 and 1,177 have employment greater

than 10,000.

Beginning in 1985, the FASB required firms to capitalize software development expenses (but not

research or maintenance expense, which usually account for most software costs). Based on some simple

calculations (details available from authors) we find that this change can explain neither the timing nor the

sign of our results.

In this discussion it is important to recall that the estimated effects are based on patentability

standards as they existed after the early 1980s. It is possible the effects might have been different under

different (say higher) patentability standards, but that question cannot be addressed here.

PDF to HTML - Convert PDF files to HTML files

1. Firms in these industries report that patents are relatively less effective compared to

firms in other industries, yet they obtain patents to aid negotiations, to cross-

license, to block competitors, and to prevent suits (Levin et al. 1987, Cohen,

Nelson, and Walsh 2000). Firms in some of these industries engage in strategic

cross-licensing of whole portfolios (Grindley and Teece 1997, Hall and Ziedonis

2001). These industries have rapidly increased their rates of patenting in general

(Hall 2003).

Our results are consistent with the explanation that these industries acquired large

numbers of software patents because they became a cost-effective means of

building strategic portfolios.

2. The growth in patenting overall exhibits a sharp break in 1984 (Hall 2003). Patent

propensity in the semiconductor industry also began a sharp and steady rise

beginning in 1984 (Hall and Ziedonis 2001). We find a sharp rise in software

patent propensity beginning in about 1984 (see Figure 1) that continues to increase

at just about the same rate as Hall and Ziedonis found for all patents in the

semiconductor industry.

A consistent explanation is that the formation of the unified court of appeals for

patents in the early 80s initiated changes in patent law that increased the cost

effectiveness of patents, especially software patents. In industries prone to strategic

patenting, this might have set off an “arms race” to build large portfolios, and

obtaining software patents became a convenient way to accomplish this goal.

3. Among firms in the strategic patenting industries, we find evidence that software

patents substituted for R&D during the 90s. This, too, is consistent with strategic

patenting of various sorts (Bessen 2003). Maturing firms with diminished

competitive advantage from technology might choose to harvest patent royalties

from their past research in lieu of further R&D, especially if legal changes make

patents more cost effective.

In turn, other firms, facing increased payment of

For example, when Louis Gerstner Jr. became CEO of IBM in 1993, he slashed R&D by over $1

billion and also shifted the business much more toward services, so that the R&D-to-sales ratio has

declined steadily. At the same time he initiated a much more aggressive patent-licensing strategy.

PDF to HTML - Convert PDF files to HTML files

royalties, may choose to reduce R&D, possibly diverting resources to build

“defensive” patent portfolios. In these cases, greater software patenting would be

associated with decreased R&D intensity.

Strategic patenting provides a parsimonious explanation for our main results.

Note that “strategic patenting” may involve several different forms of related behavior

(see Cohen, Nelson, and Walsh 2000) and different causal mechanisms for the observed

negative correlation between an increased focus on software patents and declining

relative R&D intensity (e.g., diminished technological opportunity for mature firms or

defensive patenting for small firms).

The

combination

of our results is difficult to reconcile with the hypothesis that

software patents increased R&D incentives. It would require several coincidences: Rising

patent propensity must result from a very large increase in the productivity of R&D

(beginning about 1984) that occurs in only a handful of industries (but not the software

industry) and yet without regard to the hardware/software distinction. It must also be the

case that demand for R&D is price inelastic in those same industries, but not in the rest of

the economy. This seems a rather unlikely combination.

Our interpretation might seem to conflict with the findings of Arora, Ceccagnoli,

and Cohen (2003). They find that firms’ ratings of “patent effectiveness,” as reported in

the Carnegie Mellon survey (Cohen, Nelson, and Walsh 2000) are positively correlated

with a measure of patent propensity. They also find that patent effectiveness is often

positively correlated with R&D spending. But Arora et al. are estimating a cross-

sectional effect, while our results may reflect a shift in an industry

equilibrium

that

cannot be observed in cross-sectional data. This is particularly true because strategic

patenting may have characteristics of a “Prisoner’s Dilemma.” Policy changes may

encourage firms to acquire more patents, yet, because other firms are also acquiring

larger portfolios, the equilibrium result may be lower appropriability (Bessen 2003). In

other words, Arora et al. show how R&D varies with patent effectiveness while we

explore how changing patent policy might have changed the relationship between patent

effectiveness and R&D.

PDF to HTML - Convert PDF files to HTML files

VI.

Conclusion

The rapid growth in software patenting has not been driven by the software publishing

industry, but instead by a group of industries in computers, electronics and instruments.

Other researchers have established the role of strategic patenting in these industries and

strategic patenting provides a parsimonious explanation for our results. According to this

theory, software patents are significant because they provide a cost-effective way for

firms to build strategic patent portfolios.

Legal scholars sometimes argue that patent law should treat computer programs

no differently than any other invention. This paper does not address arguments about

legal consistency, but instead explores the economic effects of granting software patents

in the U.S. during the 1990s. Our results are difficult to reconcile with the traditional

incentive theory—that granting more patents will increase R&D investments. Rather, if

legal changes have encouraged strategic patenting, the result might well be less

innovation.

PDF to HTML - Convert PDF files to HTML files

APPENDIX

A. Search Algorithm

The search query used is:

((“software” in specification) OR (“computer” AND “program” in specification))

AND (utility patent excluding reissues)

ANDNOT (“chip” OR “semiconductor” OR “bus” OR “circuit” OR “circuitry” in title) ANDNOT

(“antigen” OR “antigenic” OR “chromatography” in specification)

Table A1. Summary Statistics

Table 5 Sample

Mean

Annual successful

patent applications

Software patent

applications

Firm sales (mill. $96)

Firm R&D (mill. $96)

Percent not reporting

R&D

Firm employees

(1000s)

Percent new firms

No. of observations

No. of firms

25.9

3.6

2,632.7

100.2

27.9%

13.3

10.0%

13,136

1,443

2.4

37.2%

28,581

4,559

1987 - 97

Median

375.1

11.7

1,032.0

36.8

41.6

2.8

Total R&D

Performing

Sample

Mean

Median

Table 7 Sample

Mean

70.2

13.4%

4,967.5

190.9

21.4

10.7%

2,991

489

6.0

Median

2.7%

1,100.1

29.8

1,579.0

59.6

7.0

31.6%

11,107

2,351

1991 - 97

0.6

83.0

4.0

Corresponding

R&D Performing

Sample

Mean

Median

Note: The total R&D sample corresponding to Table 5 consists of all Compustat observations for US firms

with non-missing R&D. The sample corresponding to Table 7 consists of all Compustat observations for

US firms with non-missing or non-zero R&D for the current year and the 5-year lag. Sales and R&D are

deflated using the GDP deflator. New firms are firms during the first five years in which they appeared in

Compustat.

PDF to HTML - Convert PDF files to HTML files

References

Abel, Jason R., Ernst R. Berndt, and Alan G. White, 2003 “Price Indexes for Microsoft’s

Personal Computer Software Products,” NBER Working Paper No. 9966.

Allison, John R. and Mark A. Lemley. 2000. “Who’s Patenting What? An Empirical Exploration

of Patent Prosecution,”

Vanderbilt Law Review,

Vol. 58, pp. 2099-2148.

Allison, John R., Mark A. Lemley, Kimberly A. Moore, and R. Derek Trunkey, 2003. “Valuable

Patents,” George Mason Law and Economics Research Paper No. 03-31.

Allison, John R. and Emerson H. Tiller, 2003, “Internet Business Method Patents,” in Wesley M.

Cohen and Stephen A. Merrill, eds.,

Patents in the Knowledge-Based Economy,

National

Research Council, Washington: National Academies Press, pp. 259-84.

Arora, Ashish, Marco Ceccagnoli, and Wesley M. Cohen. 2003. “R&D and the Patent

Premium,” NBER Working Paper No. 9431.

Berger, Philip. 1993. “Explicit and Implicit Effects of the R&D Tax Credit,”

Journal of

Accounting Research,

Vol. 31, pp. 131-71.

Berndt, Ernst R. 1991.

The Practice of Econometrics: Classic and Contemporary.

New York:

Addison-Wesley Publishing.

Bessen, James. 2003. “Patent Thickets: Strategic Patenting of Complex Technologies,” Research

on Innovation Working Paper.

Bound, John, Clint Cummins, Zvi Griliches, Bronwyn H. Hall, and Adam Jaffe. 1984. “Who

Does R&D and Who Patents?” in Griliches, Zvi, ed.,

R&D, Patents, and Productivity,

Chicago: University of Chicago Press.

Burk, Dan L. 2002. “Patent Disclosure Doctrines: Enablement and Written Description,” Public

Hearings on Competition and Intellectual Property Law and Policy in the Knowledge-Based

Economy. Washington: Federal Trade Commission.

Burk, Dan L., and Mark A. Lemley. 2002. “Is Patent Law Technology Specific?”

Berkeley

Technology Law Journal,

Vol. 17, pp. 1155-1206.

Cohen, Julie, and Mark A. Lemley. 2001. “Patent Scope and Innovation in the Software

Industry,”

California Law Review,

Vol. 89, pp. 1-57.

Cohen, Wesley M., and Steven Klepper. 1996. “A Reprise of Size and R&D,”

Economic

Journal,

Vol. 106, pp. 925-51.

Cohen, Wesley M., Richard R. Nelson, and John P. Walsh. 2000. “Protecting Their Intellectual

Assets: Appropriability Conditions and Why U.S. Manufacturing Firms Patent (or Not),”

NBER Working Paper No. 7552.

Coolley, Ronald B. 1994. "The Status of Obviousness and How to Assert it as a Defense,"

Journal of the Patent and Trademarks Office Society,

Vol. 76, pp. 625-44.

Cunningham, M.A. 1995. “Preliminary Injunctive Relief in Patent Litigation,”

IDEA,

Vol. 35,

pp. 213-59.

Dam, Kenneth W. 1995. “Some Economic Considerations in the Intellectual Property Protection

of Software,”

Journal of Legal Studies,

Vol. 24, pp. 321-77.

PDF to HTML - Convert PDF files to HTML files

References

Dunner, Donald R., J. Michael Jakes, and Jeffrey D. Karceski. 1995. “A Statistical Look at the

Federal Circuit’s Patent Decisions; 1982-1994,”

Federal Circuit Bar Journal,

Vol. 5, pp.

151-80.

Flynn, John D. 2001. “Comments on the International Effort to Harmonize the Substantive

Requirements of Patent Laws” IBM submission at

http://www.uspto.gov/web/offices/dcom/olia/harmonization/TAB42.pdf

Gallini, Nancy T. 2002. “The Economics of Patents: Lessons from Recent U.S. Patent Reform,”

Journal of Economic Perspectives,

Vol. 16, pp. 131-54.

Graham, Stuart J. H., and David C. Mowery, 2003. “Intellectual Property Protection in the U. S.

Software Industry,” in Wesley M. Cohen and Stephen A. Merrill, eds.,

Patents in the

Knowledge-Based Economy,

National Research Council, Washington: National Academies

Press, pp. 219-58.

Greene, William H. 1997.

Econometric Analysis,

edition. Upper Saddle River, N.J.: Prentice

Hall.

Griliches, Zvi. 1990. “Patent Statistics as Economic Indicators: A Survey,”

Journal of Economic

Literature,

Vol. 28, pp. 1661-1707.

Griliches, Zvi, Bronwyn H. Hall, and Jerry A. Hausman. 1986. “Patents and R&D: Is There a

Lag?”

International Economic Review,

Vol. 27, pp. 265-83.

Grimm, Bruce and Robert Parker. 2000. “Recognition of Business and Government

Expenditures for Software as Investment: Methodology and Quantitative Impacts, 1959-98,”

Bureau of Economic Analysis, mimeo.

Grindley, Peter C., and David J. Teece. 1997. “Managing Intellectual Capital: Licensing and

Cross-Licensing in Semiconductors and Electronics,”

California Management Review,

Vol.

39, pp. 8-41.

Hahn, Robert W., and Scott Wallsten. 2003. “A Review of Bessen and Hunt’s Analysis of

Software Patents,” AEI-Brookings Joint Center, mimeo.

Hall, Bronwyn H. 2003. “Exploring the Patent Explosion,” invited lecture at the ZEW Workshop

on Empirical Economics of Innovation and Patenting, Mannheim, Germany.

Hall, Bronwyn H., and John van Reenen. 2000. “How Effective Are Fiscal Incentives for R&D?

A Review of the Evidence,”

Research Policy,

v. 29, pp. 449-69.

Hall, Bronwyn H., Adam B. Jaffe, and Manuel Trajtenberg. 2001a. “The NBER Patent Citations

Data File: Lessons, Insights and Methodological Tools,” NBER Working Paper No. 8498.

Hall, Bronwyn H., Adam B. Jaffe, and Manuel Trajtenberg. 2001b. "Market Value and Patent

Citations: A First Look," University of California at Berkeley Economics Department

Working Paper No. E01-304.

Hall, Bronwyn H., and Rosemary Ham Ziedonis. 2001. “The Patent Paradox Revisited: An

Empirical Study of Patenting in the U.S. Semiconductor Industry, 1979-1995”

RAND

Journal of Economics.

Vol. 32, pp. 101-128.

PDF to HTML - Convert PDF files to HTML files

References

Hausman, Jerry A., Bronwyn H. Hall, and Zvi Griliches. 1984. “Econometric Models for Count

Data with an Application to the Patents-R&D Relationship,”

Econometrica,

Vol. 52, pp.

909-38.

Hunt, Robert M. 1999. “Patent Reform: A Mixed Blessing for the U.S. Economy?" Federal

Reserve Bank of Philadelphia

Business Review

(November/December), pp. 15-29.

Hunt, Robert M. 2001. "You Can Patent That? Are Patents on Computer Programs and Business

Methods Good for the New Economy?" Federal Reserve Bank of Philadelphia

Business

Review,

Quarter, pp. 5-15.

Jaffe, Adam B. 2000. “The U.S. Patent System in Transition: Policy Innovation and the

Innovation Process,”

Research Policy,

Vol. 29, pp. 531-557.

Kemerer, Chris F. 1992 “How the Learning Curve Affects CASE Tool Adoption,” IEEE

Software, v. 9, pp. 23-8.

Kortum, Samuel, and Josh Lerner. 1999. “What is Behind the Recent Surge in Patenting?”

Research Policy,

Vol. 28, pp.1-22.

Lanjouw, Jean O., and Josh Lerner. 2001. “Tilting the Table? The Use of Preliminary

Injunctions,”

Journal of Law and Economics,

Vol. 44, pp. 573-603.

Lanjouw, Jean O., and Mark Schankerman, 2001. “Characteristics of Patent Litigation: A

Window on Competition,”

RAND Journal of Economics

v32, n1 (Spring 2001): 129-51.

Lerner, Josh. 1995. “Patenting in the Shadow of Competitors,”

Journal of Law and Economics,

Vol. 38, pp. 463-95.

Lerner, Josh. 2004. “The New New Financial Thing: The Sources of Innovation Before and

After

State Street,”

NBER Working Paper No. 10223.

Lessig, Lawrence. 2001.

The Future of Ideas: The Fate of the Commons in a Connected World.

New York: Random House.

Levin, Richard C., Alvin K. Klevorick, Richard R. Nelson, and Sidney G. Winter. 1987.

“Appropriating the Returns from Industrial Research and Development,”

Brookings Papers

on Economic Activity,

Vol. 3, pp. 783-820.

Merges, Robert P. 1997.

Patent Law and Policy, 2

Ed.

Charlottesville, VA: Michie Law

Publishers.

Merges, Robert P. 1999. “As Many as Six Impossible Patents Before Breakfast: Property Rights

for Business Concepts and Patent System Reform,”

Berkeley Technology Law Journal,

Vol.

14, pp. 577-615.

National Science Foundation. 2003.

Research and Development in Industry: 2000.

Arlington,

VA: National Science Foundation, Division of Science Resources Statistics.

Oliner, Stephen D., and Daniel E. Sichel. 1994, “Computers and Output Growth Revisited: How

Big is the Puzzle?”

Brookings Papers on Economic Activity,

1994:2, pp. 273-334.

Pakes, Ariel, and Zvi Griliches. 1984. “Patents and R&D at the Firm Level: A First Look,”

Economic Letters,

Vol. 5, pp. 377-81.

PDF to HTML - Convert PDF files to HTML files

References

Rai, Arti K. 2003. “Engaging Facts and Policy: A Multi-Institutional Approach to Patent System

Reform,”

Columbia Law Review,

Vol. 106, pp. 1035-1135.

Samuelson, Pamela, Randall Davis, Mitchell D. Kapor, and J.H. Reichman. 1994. “A Manifesto

Concerning the Legal Protection of Computer Programs,”

Columbia Law Review,

Vol. 94,

pp. 2307-2431.

Scherer, Frederic M. 1965. “Firm Size, Market Structure, Opportunity, and the Output of

Patented Inventions,”

American Economic Review,

Vol. 55, pp. 1097-1125.

Scherer, Frederic M. 1982a. “The Office of Technology Assessment and Forecast Industry

Concordance as a Means of Identifying Industry Technology Origins,”

World Patent

Information,

pp. 12-17.

Scherer, Frederic M., 1982b. “Demand-Pull and Technological Invention: Schmookler

Revisited,”

Journal of Industrial Economics,

Vol. 30, pp. 226-37.

Scherer, Frederic M. 1984. “Using Linked Patent and R&D Data to Measure Interindustry

Technology Flows,” in Griliches, ed.,

R&D Patents and

Productivity, University of Chicago

Press, pp. 417-61.

Schmookler, Jacob, 1966.

Invention and Economic Growth,

Cambridge, MA: Harvard

University Press.

Silverman, Brian S. 1999. “Technological Resources and the Direction of Corporate

Diversification,”

Management Science,

Vol. 45, pp. 1109-1124.

Soete, Luc, 1983, “Comments on the OTAF Concordance between the US SIC and the US Patent

Classification,” mimeo.

USPTO. 1994.

Hearings on Software Patent Protection,

Washington: United States Patent and

Trademark Office, January-February 1994.

USPTO. 1996. “Examination Guidelines for Computer Related Inventions,”

Federal Register,

Vol. 61, pp. 7478-92.

PDF to HTML - Convert PDF files to HTML files

Table 1. Number of Software Patents

Patents Issued

Software

Patents

1976

1977

1978

1979

1980

1981

1982

1983

1984

1985

1986

1987

1988

1989

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

765

884

897

795

1,080

1,275

1,402

1,443

1,939

2,453

2,657

3,530

3,495

4,974

4,704

5,347

5,862

6,756

8,031

9,000

11,359

12,262

19,355

20,385

21,065

23,406

Aharonian Total Utility

Estimates

Patents

100

150

200

250

300

350

400

500

600

800

1,600

1,300

1,500

1,624

2,400

4,569

6,142

9,000

13,000

17,500

21,000

70,226

65,269

66,102

48,854

61,819

65,771

57,888

56,860

67,200

71,661

70,860

82,952

77,924

95,537

90,364

96,513

97,444

98,342

101,676

101,419

109,645

111,983

147,519

153,486

157,595

166,158

Software/

Total

1.1%

1.4%

1.6%

1.7%

1.9%

2.4%

2.5%

2.9%

3.4%

3.7%

4.3%

4.5%

5.2%

5.5%

6.0%

6.9%

7.9%

8.9%

10.4%

10.9%

13.1%

13.3%

13.4%

14.1%

Successful Patent

Applications

Software Total Utility

Patents

853

1,094

1,170

1,439

1,633

1,821

2,233

2,297

2,641

2,924

3,482

4,055

4,841

5,755

6,471

7,091

8,149

9,459

12,251

16,617

17,085

13,087

65,804

65,978

65,601

65,726

66,491

63,910

65,009

61,563

67,071

71,442

75,088

81,458

90,134

96,077

99,254

100,016

103,307

106,848

120,380

137,661

131,450

114,881

2002

24,891

167,438

14.9%

Note: Utility patents excluding re-issues. Successful patent applications are the number of patent

applications that resulted in patent grants by the end of 1999.

PDF to HTML - Convert PDF files to HTML files

Table 2. Characteristics of Software Patents (1990-95)

Software

Patents

Assignee type

Non-gov’t. org. (firm)

Individual/ unassigned

Government

U.S. assignee (if assigned)

U.S. inventor

Mean citations received

Number of claims

Percent of self-citations

88%

11%

70%

69%

9.7

16.8

12%

Other Patents

80%

18%

51%

53%

4.6

12.6

13%

Percent of patents owned by

63%

top 5 percent of assignees

Note: Total patents: 39,700 software, 546,058 other. Self-citations is average of upper and lower

bounds (see Hall, Jaffe, and Trajtenberg, 2001a). Differences between the means in the first two

columns are all significant at the 1 percent level.

Table 3. Firm Characteristics by Patent Type (1990-99)

Software

Patents

24,485

13,382

956

5.8%

Non-software

Patents

11,554

8,940

376

5.1%

Median firm market value

(million $96)

Median firm sales

(million $96)

Median firm R&D

(million $96)

Newly public firm

Note: Table shows patent characteristics by type for years 1990-99 from the sample matched to

Compustat. Covers 48,072 software patents and 274,529 non-software patents. Newly public

firms first appeared in the Compustat file within the last 5 years.

PDF to HTML - Convert PDF files to HTML files

Table 4. Successful Software Patent Applications by Industry

(1994-97)

(1)

Software

patents

(2)

Share of all

Prog.

(3)

Prog. +

Engineers

(4)

All

patents

(5)

All

Patents/

R&D

(6)

Software

patent

propensity

(Table 5.2)

Manufacturing

Chemicals (SIC 28)

Machinery (SIC 35)

Electronics (SIC 36)

Instruments (SIC 38)

Other manu.

Non-manufacturing

Software publishers

(SIC 7372)

Other software (SIC

737 exc. 7372, IBM)

Other non-

manufacturing

Addendum: IBM

75%

24%

28%

25%

11%

89%





�½

33%





32%

13%

68%

88%

15%

17%

27%

11%

18%

12%

3.8

2.5

4.2

6.8

7.1

2.3

3.0

1.0

2.8

3.4

5.0

1.5

4.4

9.6

8.7

1.9

18%

49%

1.0

(all SIC 73)

55%

3.8

Note: covers 28,268 software patents and 148,552 total patents for firms in the sample of

software patents matched by CUSIP to firms in Compustat. The numbers are for patent

applications that were resulted in an issued patent by 1999. Programmer employment is the share

of total employment accounted for by computer programmers and systems analysts in 1997 from

the Occupational Employment Survey of the BLS (which includes government employees). The

third column adds employment of engineers. For SIC 737 the combined employment shares also

include IBM. The fifth column shows patents granted per $10 million of R&D in 1996 dollars.

The last column is the exponential of the coefficient for the industry given in Table 5, column 3,

normalized so that the patent propensity of SIC 73 equals one. The last row shows IBM’s patents

as a portion of all patents (not just those in our matched sample).

PDF to HTML - Convert PDF files to HTML files

Table 5. The Propensity to Patent Software

Dependent Variable: Annual Number of Successful Software Patents Applications

Ln employment

Ln R&D / Emp.

No R&D dummy

Ln Capital / Emp.

Programmers / Emp.

Engineers / Emp.

New firm dummy

New firm x SIC 73

Other patents / Emp.

Industry dummies

SIC 28

SIC 35

SIC 36

SIC 38

Other manu.

SIC 73

Other non-manu.

Firm fixed effects

Period estimated

Annual growth rate of

year dummies

No. observations

Log Likelihood

80 - 97

10.3%

23,108

-68,548

87 – 97

9.4%

13,136

-42,963

87 – 97

9.5%

13,136

-42,870

87 – 97

7.2%

13,136

-48,528

-5.58

-4.50

-3.73

-3.83

-5.33

-5.99

-4.65

(.32)*

(.25)*

(.24)*

(.25)*

(.63)*

(.28)*

(.03)*

(.05)*

(.17)*

(.04)

(.03)*

(.06)*

(.18)*

(.05)*

(3.13)*

(.55)*

(.13)

(.03)*

(.06)*

(.18)*

(.05)*

(3.09)*

(.55)*

(.15)

(.33)*

(.03)*

(.05)*

(.19)

(.05)*

(.84)*

(.44)*

(.13)

Annual

Contribution

(.02)*

(.12)*

(.03)*

(.75)*

(.58)*

.88

1.01

-1.00

-.08

.88

.77

-.47

.37

12.44

4.13

.88

.77

-.49

.37

12.75

4.05

.12

-.90

.88

.78

-.16

.26

7.21

6.57

.11

.16

.62

.22

.35

.27

13.32

6.06

0.5%

1.1%

-0.1%

1.8%

1.2%

.19

(.13)

.03

(.02)*

-5.52

-4.45

-3.68

-377

-5.26

-5.96

-4.58

(.31)*

(.25)*

(.23)*

(.25)*

(.62)*

(.28)*

Yes

87 – 97

10.8%

6,587

-9,659

Note: Heteroscedastic-consistent standard errors in parentheses. All regressions include year

dummies; columns 1 and 4 include an intercept (not shown). Asterisk indicates significance at

the 1 percent level. Regressions are Poisson regressions. Column 5 has firm fixed effects

(Hausman, Hall, and Griliches, 1984). R&D is deflated by the GDP deflator, capital is property,

plant and equipment deflated by the NIPA capital goods deflator, and employment is in

thousands. The “new firm” dummy is equal to one for the first five years a firm appears in

Compustat. Other patents / employees is the ratio of patents to employees for other firms in the

same 2-digit SIC industry for the given year. Annual growth rate of year dummies is year

dummy for 1996 minus the year dummy for the earliest year divided by the number of years

elapsed (no dummy is estimated for 1997). Annual contribution is the estimated coefficient times

the difference in sample means for the variable between 1996 and 1987 divided by 9 (the number

of years).

PDF to HTML - Convert PDF files to HTML files

Table 6. R&D Cost Share Estimation

∆

Cost

WLS

1991-97

∆

Cost

∆

Sales

WLS

1991-97

∆

Sales

WLS

∆

Sales

∆

Sales

WLS

Heckman

1991-97

Heckman

1991-97

1985 - 97

∆s

∆

-.039

(.017)

-.030

(.003)*

.007

1.018

(.071)

(.024)*

-.037

(.006)*

-.037

(.002)*

∆s

x (1985-89)

∆s

x (1990-93)

∆s

x (1994-97)

No. total

observations

(firm years)

No. selected

LR test of probit

Wald test of

independent

equations

Adjusted

.018

-.031

-.050

774

29,832

757

(9) = 1059.9

P=.000

(1) = 571.6

P=.000

.128

.974

.038

774

2,991

29,832

2,913

(9) = 3370.6

P=.000

(1) = .04

P=.839

.066

5,285

(.012)

(.012)*

(.015)*

Note: Observations are firm-years, but differences are taken over five years. Robust standard

errors are in parentheses. All regressions include year dummies and changes in log factor prices.

The prices are for capital, labor, energy, materials, services and IT for the two-digit industry

from the BLS. Asterisk indicates significance at the 1 percent level. For columns 1, 3, 4, and 6,

the sample includes firms matched to patent file with positive patents and non-missing data. All

regressions are weighted by

(

−

)

where

is the number of patents granted.

software share of patents applications. The Heckman sample selection equations consist of a

regression equation and a selection probit. The regression coefficients are shown, the probit

regressors are log market value, log sales, new firm flag and time dummies. To measure the

goodness of fit for the probit, we show a likelihood ratio test that all coefficients except the

constant term are zero. To determine whether sample selection biases the coefficients, we show a

Wald test that the selection and regression equations are independent. In columns 4-6,

observations are excluded where R&D > �½ sales (see text).

PDF to HTML - Convert PDF files to HTML files

Table 7. R&D Intensity Estimation, 1991-97

Dependent Variable:

∆

Sales

.001

(.010)

-034

-.506

.120

(.011)*

(.136)*

(.087)

-.051

(.018)*

-.037

(.011)*

∆s

∆

Programmers/Emp.

∆

Engineers / Emp.

∆s

x (Emp.

≤

1,000)

∆s

x (1,000<Emp.

≤

10,000)

.007

.039

(.030)

(.022)

∆s

x (Entrant)

∆s

x strategic industry

∆s

x SIC 28

∆s

x SIC 35

∆s

x SIC 36

∆s

x SIC 38

∆s

x Other manu.

∆s

x SIC 73

∆s

x other non-manu.

No. observations

Adjusted

.002

-.060

.029

-.029

-.090

.027

.014

-.077

-.053

2,991

.060

(.031)

(.010)*

(.026)*

(.028)

(.013)

(.031)*

(.044)

(.019)*

(.028)

2,991

.051

2,967

.058

2,991

.045

2,991

.043

Note: Robust standard errors in parentheses. All regressions include year dummies and changes

in annual log factor prices at the two-digit industry level. Asterisk indicates significance at the 1

percent level. Time differences are five years and observations are firm years. Sample includes

firms matched to patent file with positive patents and excludes observations where R&D > �½

sales.

∆programmers/total

employment and

∆engineers/total

employment are the changes in the

shares of the respective occupational employment for the three-digit industry. All regressions are

weighted by

(

−

)

where

is the number of patents granted.

is software share of

patents granted, sales are deflated, and “Entrant” is a dummy for firms that entered the

Compustat database after 1985. “Strategic industries” are SIC 357 (computers), SIC 36

(electrical), SIC 38 (instruments) and IBM.

PDF to HTML - Convert PDF files to HTML files

Figure 1. Year Dummies, Software Patent Propensity Regressions

2.5

1.5

Hall-Ziedonis (semiconductor)

Poisson (Col. 5.1)

Poisson fixed effects (Col. 5.5)

0.5

Year

Source: Hall and Ziedonis (2001) and Table 5. Year dummies from Poisson regressions

correspond to log relative patent propensity. Normalized to 1 in 1987.