Innovations for Products, Innovations for Licensing:

Patents and Downstream Assets in the Software Security Industry

Alfonso Gambardella Università Bocconi, Milan, Italy

agambardella@unibocconi.it

Marco S. Giarratana Universidad Carlos III de Madrid

marco.giarratana@uc3m.es

Abstract

This paper develops a theoretical framework in which patents and control of downstream co-specialized assets affect the way firms profit from innovation according to the stage of industry evolution. We argue that in early stages of industry life cycle the effectiveness of patents in protecting innovations or in enhancing their visibility encourages technology licensing, while ownership of downstream assets helps launch new products. We test these ideas in the Security Software Industry, which arose about 20 years ago, and exhibits the typical features of an early stage industry. Moreover, SSI patents are an effective means for technological protection and for visibility. We employ panel data on 87 firms with patents in key SSI technologies during 1993-2000. After controlling for firm size, age, and R&D intensity, we find evidence consistent with our predictions.

Keyword: Technology Licensing, Patents, Downstream assets, Software

JEL: L8; O32

1. Introduction

Firms that want to profit from innovation face two strategies: selling the technology to another

party or embedding it in final products. While integration in products is the classical form of

commercial exploitation of innovation, technology trade has gained importance (Arora and

Gambardella 1994; Gans and Stern 2003; Cassiman and Veugelers 2006). In his seminal work,

Teece (1986) highlights the pivotal factors in this decision: Control of downstream co-

specialized assets and degree of Intellectual Propriety Right (IPR) protection (see also McGahan

and Silverman, 2006). The underlying theoretical argument is that if a firm owns the

downstream assets, it will embed innovation in final products. If not, and these assets are costly

to acquire, it may find it more profitable to sell the innovation. In this case, IPRs are important,

as technology deals are affected by transaction costs that are lowered by well-defined IPRs

(Arora et al. 2001). Recent case studies call for a fine-tuning of this theory. Sometimes

established firms with downstream assets sell their technologies (Arora et al. 2001); young start-

ups with scarce marketing and production assets exploit their innovations with final products

(Giarratana 2004); quite a few firms both sell and use their technologies internally (Ammon and

Laursen 2006).

This paper argues that what matters is not only the firm control of downstream assets,

but also how fragmented these assets are in the industry, where more fragmentation means more

sub-market niches and less direct firm competition. Since industry structure typically changes

over the industry life cycle, this means that downstream assets can play a different role

according to the stage of industry evolution (Klepper 2002; Klepper and Thomspon 2007). With

respect to IPRs, the literature has stressed the importance of patents in protecting technology

suppliers (e.g. Arora et al. 2001, Gans et al. 2002, Arora and Merges 2004). We also highlight

their role in lowering the search costs of potential technology buyers. Thus, in our view, well-

defined IPRs reduce transaction costs in technology trade for both the buyer and the supplier.

We summarize our theoretical framework in four industry scenarios, according to the high-low

importance of IPRs and fragmentation of downstream assets. Each scenario predicts the impacts

1

of the two factors on the firm propensity to exploit innovation through final products or

technology licensing.

Specifically, we are interested in testing the theoretical implications of the scenario in

which IPRs are relatively well-defined and downstream assets are fragmented. This is because,

according to our theory, this is the scenario in which a firm holding a new technology faces a

choice between licensing and developing the product. As we shall see, when IPRs are weak

firms integrate, and when IPRs are strong but downstream assets are homogenous licensing

either occurs because of pre-existing vertical specialization (i.e. not explained by our strategic

choices) or it does not occur. Also, this scenario depicts the case of technology-based nascent

industries. In these industries downstream assets are fragmented because a dominant design has

not yet emerged, and we study the implications of well-defined IPRs.

We look at the Security Software Industry (SSI), which is an ideal test-bed for the

scenario that we want to study: i) it is a recent technology-based industry with many sub-market

niches in which innovation plays a major role; ii) IPRs are well-defined, and firms usually

patent their new software algorithms, which are the core of a software security system; iii) the

industry exhibits significant entry and exit, with little sign of consolidation around a few large

players; and iv) firms tend to profit both through product releases and technology licensing

(Fosfuri and Giarratana 2007). By drawing on a sample of 87 firms with patents in key SSI

technologies, we predict the probabilities that a firm sells its technology or launches a new SSI

product using 1993-2000 panel data. To our knowledge, only Gans et al. (2002) provide

empirical evidence of the simultaneous determinants of firm choice to sell the technology or to

go downstream in the final market. They employ survey data on 118 start-ups that have received

a public or private grant. From an empirical perspective, our study makes two improvements.

First, we do not pre-determine the firm types that license or develop products, viz. we use a

sample of both small and large firms. Second, the industry exhibits a clear distinction between

cryptography technology and final product market. This reduces the ambiguity between

products and technologies, and therefore between licensing and product strategies.

2

One of the most relevant empirical results of our analysis is that patents encourage

technology trade. They favor the probability of selling the firm technology, and we discuss why

they capture something more than innovation in our regressions, particularly lower transaction

costs from both the demand and the supply-side of technology. We also find that downstream

assets are the main drivers of the firm propensity to release an SSI product. Finally, patents do

not produce a clear-cut effect on the propensity to release a product, and downstream assets do

not affect the probability of selling the technology.

The next section discusses our theoretical background. Section 3 describes the major

features of SSI. Section 4 presents the empirical evidence, and Section 5 discusses our results.

The final section concludes.

2. Theoretical background

Firms that control downstream assets tend to embody their innovations in final products. There

are many reasons. First, such assets, once in place, need to be fed with production and

commercialization activities to avoid under-utilization. Second, patents, secrecy, and other

means can be used to block or delay imitation, thereby reinforcing the ability of companies to

exploit the innovation commercially. Third, as shown by Arora and Fosfuri (2003) in a

theoretical model, and by Fosfuri (2007) empirically, the threat that technology buyers could

compete with the licensor discourages technology producers with downstream assets to trade

their technologies. All this suggests that firms with downstream assets rarely sell their

innovation rights. Typically they license if this does not erode their market share, as for example

when the potential technology buyer operates in a distant market, whether geographically (e.g.

international licensing) or technologically (e.g. the technology is used in distant product

applications). By a symmetric argument, the lack of downstream assets, and the cost of

acquiring them, favor technology licensing, as shown by the fact that this is the way in which

many innovative young entrepreneurial firms profit from their innovations in markets controlled

by established incumbents (Gans and Stern 2003).

3

In spite of these theoretical predictions, Gans et al. (2002) do not find any robust effect

of downstream assets on the propensity of start-ups to sell their technology, and Arora and

Ceccagnoli (2006) show that the firm downstream assets do not produce an effect per se, but

only mediated through the effectiveness of patents. By contrast, Fosfuri (2007) finds a

significant negative effect of downstream assets on the probability of licensing chemical

compounds for a sample of large petrochemical firms. Similarly, McGahan and Silverman

(2006) find the in industries in which downstream complementary assets are important, relevant

innovations (measured by patent citations) by “outsiders” – i.e. firms or institutions that do not

compete with the focal firm – have a positive impact on its financial value. This suggests that

the control of complementary assets forces the outsiders to develop alliances or licensing

agreements with the incumbents. This mixed evidence can be explained. The fragmentation of

downstream assets changes across industries and along the stages of an industry evolution

(Teece 1986). In the computer industry, Bresnahan and Greenstein (1997) provide an

illuminating example of what happens to industry structures and firm strategies when the

fragmentation of downstream assets changes. When they are not fragmented, firms invest in

endogenous sunk costs securing the control of downstream co-specialized assets. When

downstream assets partition for exogenous reasons, new market niches arise and new

competitors become leading suppliers, threatening the positions of the consolidated leaders.

According to the “industry life cycle” tradition (Jovanovic and MacDonald 1994;

Klepper 2002), fragmentation of downstream assets is typical of the initial period of an

industry's development, which also features intensive entry and exit of firms. Then, industries

undergo a “shake-out” phase, which leads to product standardization, fewer sub-market niches

and higher concentration (see also McGahan and Silverman 2001). Klepper and Thompson

(2007) explain the pre-shake-out pattern with the tendency of some industries to generate

continuous proliferation of new sub-market niches that make the standardization of downstream

co-specialized assets difficult. A common interpretation is that the proliferation of sub-market

niches emerges when industries become markets of general purpose technology and products

(Bresnahan and Trajtenberg 1995). These markets feature several potential applications that

4

give rise to a continuous life and death of sub-markets with specialized downstream assets.

Lasers, semiconductors, or software are typical examples. In sum, what matters is not only the

firm control of downstream assets, but also how partitioned these assets are in the industry. In

an industry, or an industry stage, in which downstream assets are fragmented, the control of

downstream assets could have a moderate impact on the decision to licence technologies.

At the same time, technology trade is bound by transaction costs. This stems largely

from the intangible nature of knowledge goods that makes the object of transaction ill-defined.

Asymmetric information or other contractual ambiguities then curb technology trade. As a

matter of fact, Arora and Gambardella (1994) argue that technology markets are more likely to

arise when technologies are more codified, and take the form of well-defined objects like codes,

algorithms, genes; however, the codification of technology is not sufficient for technology trade.

On the supply side, we know since Arrow’s (1962) work that information can be reproduced

easily, and this can destroy the rents of its original producer if it is not protected by property

rights. A growing literature has then noted that patents enhance technology markets because

they confer a property right to the technology producer (e.g. Gans et al. 2002; Arora and Fosfuri

2003; Arora and Merges 2004). On the demand side, another well-known feature of patents is

that they disclose information about the innovation. Today, with on-line databases and the great

attention paid by market players to patents, there is a great diffusion of information about them,

which reduces the search costs of technology buyers. Interestingly, codification helps both

sides. The effectiveness of patents is higher when the technology is codified because there are

fewer ambiguities about what is protected. At the same time, codified technologies can be

understood fairly well just from the patent documents. Their diffusion can then be enough to

boost the visibility of the innovation. The recent empirical evidence confirms that patents favor

technology licensing, even though it does not establish whether patents reduce transaction costs

5

from the point of view of the buyer or the supplier (Arora and Ceccagnoli 2006; Gambardella et

al. 2006).1

To summarize this discussion, below we define four scenarios according to the

importance of patents in an industry (high-low) and the fragmentation of downstream co-

specialized assets (high-low). In the four scenarios, we predict the effect that patents and

downstream assets have on the propensity of firms to profit from its innovation through

technology licensing or final products. The four scenarios are also synthesized in Table 1.

[TABLE 1 ABOUT HERE]

Scenario 1. Patents are important and downstream assets are fragmented. The

fragmentation of downstream assets means that competition among firms is less tight because of

the differentiation across niches. In turn, this lowers concerns about the rise of new competitors

when technologies are sold. In this scenario, firms can profit from innovation through both

licensing and product releases. Patents encourage licensing because protection prevents others

from reproducing the technology without consent, and visibility raises the demand for

technology. At the same time, the very fact that markets for technology are well-functioning

implies that patents have a limited effect on the probability of releasing a product because firms

without innovations could buy them on the market. By contrast, downstream assets ought to

have a positive effect on the firm propensity to release a product. At the same time, they will

have a moderate or null effect on licensing since the fragmentation of the industry implies that

there are few concerns of business stealing or cannibalization.

Scenario 2. Patents are important and downstream assets are homogenous. In this case,

if a) downstream assets and innovations are owned by the same organization, firms will exploit

their inventions and use patents to block potential competitors (i.e. computer hardware before

1990s). If b) innovation and downstream assets are owned by different organizations, there will

be a division of innovative labor with classical technology trade between young entrepreneurial

firms and large incumbents (e.g. biotechnology). In the a) case, firm investments in patents

1 Since a primitive reason why patents are important is that the technologies are codified, in this context important patents and codified technologies will be largely equivalent concepts. However, some legal attitudes of the Patent Offices could also play a role, i.e. software patentability.

6

should have a moderate or null effect on the probability of licensing and a positive one on the

probability of releasing a product since firms with downstream assets need inventions and

protection to exploit business opportunities. Downstream assets should have a positive effect on

the firm propensity to release a product and a negative or null effect on licensing. In the b) case,

firm patents ought to have a positive effect on the probability of licensing only in the case of

firms without downstream assets, and a limited or null effect on the probability of releasing a

product since firms without downstream assets cannot enter the downstream market.

Downstream assets should affect negatively the probability of licensing because there are

concerns about market share cannibalization. However, they should positively affect the release

of products.

Scenario 3. Patents are not important and downstream assets are fragmented. In this

case, even if firms are willing to license technology, transactions costs are high. There could be

a less than optimal level of technology trade, and firms could not exploit all the potential market

niches of the industry. The software industry in the 1980s is an example. Patents ought to have

no effect on the probability of licensing or product release. Downstream assets should

encourage product releases, but should not affect greatly the propensity to license.

Scenario 4. Patents are not important and downstream assets are homogenous. These

industries are the least intensive in technological trade. Examples can be found in the traditional

scale intensive industries, like petrochemical products. Patents should have no effect on either

the probability of licensing or of exploiting the invention downstream. Downstream assets

should favor product releases and reduce the propensity to license.

3. The Security Software Industry

3.1 Background

SSI is one of the newest segments of the software industry. So far, it is a good example of a no-

shake-out industry (Giarratana 2004). SSI has experienced an unprecedented growth in recent

years. While it started around the end of the 1980s, the SSI world market reached USD8.9 billon

in 2001, up from USD6.3 billion in 2000 and USD4.4 billion in 1999 (IDC 2002).

7

The technological foundation of the industry dates back to the late 1970s, when the US

government channelled important investments in military projects linked to the security of data

transmission. However, it was not until the late 1980s that commercial versions of software

security products were released into the civilian market. The inception of the industry therefore

coincided with the growing market for home personal computers and the development of the

Internet, with web financial transactions and data transmissions. This, in turn, created a growing

commercial demand with different requirements that broadened the spectrum and complexity of

the products and services offered. Specifically, the first software security product that we are

aware of was the first anti-virus released by McAfee in 1989.

Given the steady increase in the rates of entry and exit throughout its existence, SSI

features high industry turbulence (Fusfuri and Giarratana 2007). In part this can be explained by

the relatively small sunk costs needed to start an SSI venture. The initial amount invested to set

up Check Point Software in 1993 was only USD 300,000. Check Point Software, an Israeli

company, is currently one of the world leaders in SSI with sales of 580 million USD in 2005.

SSI has also experienced an intense proliferation of sub-market niches. They range from

basic security software, such as Virtual Private Networks, Firewall and Virus Scanning, to

advanced security services like Public Key Infrastructures, Security Certification, and

Penetration Testing. Table 2 shows the SSI major product niches that have arisen in fewer than

15 years of industry history using a six digit SIC code classification.

[TABLE 2 ABOUT HERE]

3.2 Algorithms, patents and licensing

The design of security software products is a complex undertaking. The technological core of

the product is the crypto-algorithm, which specifies the mathematical transformations that are

performed on the data. Speed of mathematical calculations and security level are the two main

features by which SSI products are evaluated. This is because the time consumed by the

encryption and decryption processes depends on the length of mathematical algorithms and the

8

power of computing machines (Giarratana 2004). A good algorithm aims at minimizing the

computer time needed to perform data transformation under a security level threshold.

The crypto-algorithm is the principal object of a firm patent. Patenting of the crypto-

algorithm has been encouraged by the loosening restrictions of the US Patent Office on software

patenting. Usually, SSI patents include flow charts reporting step-by-step the encryption-

decryption routines and the detailed description of all the mathematical procedures that perform

the encryption. Moreover, there are tests on the level of security and the speed of the process.

The crypto-algorithm is a classical example of a codified technology. For example, one of the

most important crypto-algorithms is the Elliptic Curve Cryptography patented by Certicom (US

Patent number 6,141,420). This algorithm needs only 160 computer bits to perform all the

procedures while the standard string needs about 1,024 bits. This is because while the standard

systems are based on integer calculus, the Certicom algorithm is based on elliptic curves that

can be calculated more easily and faster while providing the same level of security. Yet, this

also means that if unprotected, the technology is not hard to imitate, as it is clearly defined in

mathematical and software language, i.e. it is expressed in a well-defined code. For the very

same reasons, the patent defines clearly the object of what is protected, which reduces potential

ambiguities about it, and then raises the effectiveness of protection. Finally, the codified nature

of the algorithm and of its functions implies that the disclosure of information provided by the

patent raises its visibility, as potential users can rapidly verify its structure and utilization.

We confirmed the importance of the protection and visibility role of patents in SSI from

interviews with managers in four firms that have sold algorithms. We found the following: First,

patents are a good way of protecting the algorithms. Patents are necessary, because algorithms

tend to devaluate quickly. The interviewees told us that patents produce a lead time advantage

of no more than two years. Second, firms patent without a well-planned ex-ante strategy of

selling the technology. Our four firms were approached by future technology buyers who had

studied their technology and were interested in utilizing it in their product developments. This is

suggestive of the fact that the technology markets in SSI is in good part demand-driven, which

in turn raises the importance of the visibility function of patents. Our interviewees also

9

suggested that firms in SSI sell technology to gain reputation or to open up new markets, even if

the licensing revenues are quite sizeable.

The foregoing remarks make SSI an ideal example of a market for both products

(Firewall, Anti-virus, etc.) and technology. This is confirmed by Hoovers data

(www.hoovers.com), which classified the 2002 SSI revenues in sales of software products (52.3

%), services (30.3 %) and revenues from licensing the technological algorithms (17.4 %).

Moreover, our understanding of the industry is that, after a licensing deal, the licensee uses the

acquired technology either to create and improve a new algorithm or to embed it in its software

products. This highlights the high codification of the underlying technologies, as the algorithms

can be readily embedded in new systems, whether new algorithms or products, without any

major modifications of the technology itself or the system in which it is embedded. Two

examples of these two uses are the following. In 1999 IBM bought from Security Dynamic

Technology the right to use the SecurID patented algorithm (US Patent number 6,411,715) that

generates a one-time, pseudo-random code that changes every 60 seconds. IBM planned to

improve this technology in order to create a new system of data protection particularly suited for

banks and financial institutions.2 In 2000 Entex, a consulting company, signed a licensing

agreement with the software development company M-Tech Mercury Information Technology

to use M-Tech's P-Synch password management system as a user self-authentication tool in

conjunction with self-service IT applications deployed at customer locations by Entex.3

4. Empirical Evidence

4.1 The sample

Since we aim at studying the behaviour of firms holding a SSI innovation, our sample is

composed of all the firms with at least one software patent granted in technological classes that

are strategic for SSI (SSI Patents, hereafter). As previously discussed, in SSI patents are

effective, and therefore companies patent their algorithms frequently. Our discussion and 2 PR Newswire, 7/1/1999 p.38 3 Telecomworldwire, 17/4/2001 p.67

10

interviews with managers of SSI companies confirmed this presumption. It is unlikely that a

firm that has developed a new software security algorithm does not patent it. As a result, this

sampling procedure certainly captures all the firms with at least one algorithm in this business.

This represents the best proxy, and to our judgment not noisy, of an ideal sample of all the firms

that own an SSI algorithm – the core technology in this business. Unfortunately, data on the

algorithms owned by SSI firms are not available on a large scale. At any rate, in our robustness-

check section below, we will also show our results after adding a control sample of firms

without SSI patents.

We built our sample from the LECG Corptech Patent database (www.lecg.com), which

covers about 80,000 software patents granted by USPTO from 1976 to 2000. We selected all the

patents in the US technological classes 380 (“Cryptology”) and 705, subclasses 50-79

(“Business Processing Using Cryptography”). We then obtained a sample of 87 firm-assignees

with at least one of these patents. We combined this information with the SSI Database

constructed by Giarratana (2004) which includes all the SSI product introductions and licensing

deals from the beginning of the industry in 1989 to 2000. The original data source is the

database Gale Group’s Infotrac Promt (www.gale.com), which is the new version of Predicast

database, used in several research articles (e.g. Pennings and Harianto 1992). Promt contains

comprehensive and reliable coverage of companies, products, markets, alliances, and deals from

a vast collection of journals, newsletters, news releases and newspapers. We constructed our

database from firm news announcements in Promt classified under SIC code 73726, which

corresponds exactly to Software Security, and under the firm events “New Product

Annoucement” or “Technology Licensing.”4 Since before 1993 our sample firms do not release

products or licensing contracts, we restricted our analysis to the period 1993-2000. We ended up

with a panel dataset of 696 observations, or 87 firms for 8 years. Our 87 sample firms released

21.3% of the total number of products and 72.4% of technology licensing registered in the SSI

Database.

4 For details about the construction of the database and the data cleaning process, see Giarratana (2004) and Fosfuri and Giarratana (2006).

11

4.2 Methodology and dependent variables

We run two logit regressions to estimate the probability that, in each year, the firm releases a

new SSI product or signs a licensing contract as a technology seller. The first dependent

variable is LICENSE which takes the value 1 if in the corresponding year the firm announces at

least one licensing deal as an algorithm seller, and 0 otherwise. The second dependent variable

is PRODUCT which takes the value 1 if in the corresponding year the firm announces the release

of at least one new SSI product, and 0 otherwise. Of our 87 firms, 44 launched at least one new

SSI product, and 36 signed at least one licensing contract as a seller. We run a logit for whether

products were released or licenses were issued rather than the number of products or licenses

because we are not interested in the scale of licensing or new product introduction but in the

licensing vs. product strategy. Our dummy for licensing or product is a better measure of the

intentions of the firms.

4.3 Independent variables

Patents. For each firm-year, we calculate the stock of SSI patents (PATENT), that is the

cumulative sum of all the SSI patents up to that year. As a robustness check, we introduce two

alternative patent stock variables using an annual discount factor of 10% and 15% respectively

(PATENT10 and PATENT15). These covariates proxy for the firm propensity to invest in

patenting the cryptology invention. We use stock measures to better capture potential learning

and knowledge base effects.

There are two important remarks about the use of patents in our regressions. The first

one is that our goal is not to study patents as a measure of innovation, but as factors that

encourage licensing or the development of new products. Since data on the number of firm

innovations are difficult to obtain, we have to resort to other controls (i.e. firm-level R&D, size

and age). However, since they are imperfect measures of the number of innovations, patents

could still proxy for them. Our claim is that patents will be also correlated with other aspects as

well, particularly protection and visibility.

12

This leads to our second point. We do not make any claim on the origin of different

patent propensities. Because we have firm-level data, we need patents to capture differences

across firms, or at least over time for the same firm. Firms could show differences in the

propensity to patent, which could stem from different sources. For example, some firms may

have internal legal offices that reduce the marginal cost of patenting, or simply greater

experience. Another important source of inter-firm differences is the IPRs managerial ability.

Some firms simply perceive the advantages of patents better than others, and therefore patent

more, e.g. to better protect around an innovation. Finally, good algorithms span more

applications, which lead to more patents. While this may also reflect differences in quality, a

larger number of patents produce greater protection and visibility in any case, and therefore

capture some of the aspects that we want to account for.

Downstream Assets. Previous studies on licensing measure downstream assets with

direct survey responses on the importance of downstream complementary assets (Gans et al.

2002), with a dummy that captures if the R&D and production personnel interact (Arora and

Ceccagnoli 2006), and with the firm market share in a particular market segment (Fosfuri 2007).

These indirect measures suggest that the use of real measures of downstream assets has proven

to be surprisingly difficult. We try to improve on this matter by jointly employing two standard

accounting measures like sales and fixed assets along with a natural proxy of downstream assets

like the firm trademarks. We construct two time-variant variables that we use alternatively to

proxy for downstream assets in software: the share of the live software trademarks on the total

firm trademark multiplied by firm fixed assets (ASSETS1), and the share of the live software

trademarks on the total firm trademarks multiplied by firm size in sales (ASSETS2). We are

basically using the share of software trademarks to denote the proportion of firm total sales or

fixed assets associated to software. We use both sales and firm fixed assets for robustness

purposes.

Trademarks are combinations of “words, phrases, symbols or designs that identify and

distinguish the source of the goods or services” (USPTO Documentation, http://tess.uspto.gov).

US trademark owners pay different types of fees for each class of goods/services for which a

13

trademark is registered, and they have to prove periodically that they are using the trademark in

the US market. Otherwise – even if the owner is willing to pay the fees – the trademark is

cancelled. A trademark protects a new logo on a product or a new advertising and branding

campaign. Moreover, software trademarks are not a recent phenomenon like software patents.

For example, the first software trademark was granted in 1967 to the firm Medelco, and

Microsoft has registered up to date more than 1,700 software trademarks, the first one in 1982.

Therefore, the trademarks are an excellent proxy for the “stock” of firm downstream assets in a

particular market segment, which is difficult to obtain from more conventional accounting

measures, also because if the trademarks are not commercialized they cannot be renewed. The

empirical literature on trademarks has emerged only recently. Previous studies have shown that

trademarks are a good proxy of the products/markets in which a firm operates and their level of

product advertising. For example, Seethamraju (2003) and Smith and Parr (2000) show that

trademarks are highly correlated with firm sales and stock market value.

In our study, for each firm-year, we calculated the share of software trademarks

registered and “live” at USPTO by our sample firms over the total live firm trademarks. In so

doing, for our sample firms, we downloaded trademarks data from the USPTO database

available at http://tess.uspto.gov. In order to distinguish the software trademarks, we apply a

search algorithm ([“computer software” or “operating system” or “computer program” or

“software algorithm” or “data processing” or “software application”]), to the front page of the

trademark in the description of Good and Service trademarked. To validate this proxy, we

confront our trademark values with the firm sales in software over total revenues that are

published by Software Magazine in the Software 500 List (survey data,

www.softwaremag.com). For four sample firms that are present in the Software 500 List, the

trademark and revenues values are very similar: H&P 0.28 from the 500 List, 0.30 from our

trademark proxy; IBM 0.52 vs. 0.46; Sun Microsystems 0.32 vs. 0.35; Apple 0.22 vs. 0.20 (data

at 2000). Financial data sources utilized are Bureau Van Dijk’s Osiris and Orbis, Compustat,

SEC filings and Hoovers.

14

Controls. First, we introduced time-variant controls. We have the firm age (AGE) in

years, its size (SIZE) in sales (USD million), and its R&D intensity (R&D) calculated as the

R&D expenditures of the firm divided by its sales. While size and age are standard controls, and

can be interpreted as proxies for a firm scale and experience, R&D plays a more important role.

As noted, since we want to check the effect of the firm patents on the probability of licensing,

and not of the patents as pure measures of the firm innovations, we need R&D to control for the

latter. To our knowledge, the only empirical evidence available that predicts the licensing

propensity using both patent and R&D is Fosfuri (2007) who finds that in the chemical industry

R&D intensity shadows the effect of patents on firm technological licensing. All the time-

dependent independent variables are lagged one year and enter the regression in a logarithmic

scale – results are similar without logarithms.

Second, we employ three sets of time-invariant dummies. We have three dummies to

control for the firm core business. They take value 1 if the firm core business is in software, SIC

code 737 (SOFTWARE), hardware, SIC code 357 (HARDWARE) or electronics, SIC code 359-370

(ELECTRONICS), and 0 otherwise (Source: Bureau Van Dijk’s Osiris). We then introduced two

geographical dummies that take value 1 if the firm headquarters are located in North America

(NORTH AMERICA), or in Europe (EU) and 0 otherwise. Finally, we introduced 7 year dummies.

4.4 Estimation results

Table 3 reports descriptive statistics for all the variables used in our analysis. Table 4 and 5

show the results for the probability of licensing and product releases. We present 6 different

models (Model 1-6) using the three patent specifications (PATENT, PATENT10, and PATENT15)

and the two proxies for the downstream assets (ASSET1 and ASSET2). All the controls are

always included as covariates.

[TABLES 3, 4 AND 5 ABOUT HERE]

The three patent stock specifications lead to similar results in all models. Patents have

an important positive effect on the probability of licensing an algorithm, but no effect on the

probability of releasing an SSI product. In a standard experiment, in Model 1, a standard

15

deviation increase from the mean in PATENT increases the probability of selling an algorithm by

34.1%. Our trademark-based proxies for downstream assets exhibit a symmetric pattern. They

have no effect on the probability of licensing but a positive and significant effect on the

probability of releasing an SSI product. A standard deviation increase from the mean in ASSET1

increases the probability of launching a new product by 12.4% (Model 1).

As far as our controls are concerned, size and age have a negative effect on the

probability of licensing. Old and large firms are less likely to be technology sellers of this

industry, which strengthens the classical framework whereby it is the small and new innovative

start-ups that typically act as technology suppliers. As expected, R&D intensity has a positive

and significant impact on the probability of licensing. We then obtain an effect of patents on the

probability of licensing after controlling for R&D.

In the product equation, age is significant with a negative sign. SSI is then a young

industry dominated by young firms on the product side as well. Note that size does not affect

the probability of releasing products, a result that might have been expected in the software

industry where scale economies play a rather marginal role. R&D-intensity is again important.

Apart from being more likely to license, more R&D-intensive firms are also more likely to

introduce a new SSI product. In this respect, R&D probably captures a good deal of the

innovation effect. Our interpretation is that in order to release an SSI product a firm needs some

downstream assets and an innovation, but patents per sè do not play a pivotal role for launching

a new product.

4.5 Robustness checks

We first check that the licensing and product decisions are not correlated. This implies no loss

in estimation efficiency by estimating (1) and (2) separately. Following Menard (1995: 80) we

compute the Pearson residuals for logit regressions, i.e. the residuals standardized by their

standard errors, and then the correlations among the residuals of each of the six pairs of product

and licensing regressions. For Model 1, residual correlation is 0.062; for Model 2 and 3 it is

0.060; for Model 4 and 5 it is 0.058; and for Model 6 it is 0.054. All correlation are not

16

statistically different from zero, except the first one at 10% level. We also perform a

multinomial logit with a dependent variable taking three values (null, product, licensing). Not

only are the estimation results (not shown here) similar to the independent regressions that we

have shown, but the Hausman and Small-Hsiao tests do not reject the hypothesis of

independence of irrelevant alternatives for every choice eliminated (licensing, product). We also

performed a no-time variant Logit regression estimating the firm probabilities to licence or to

launch a product during the whole time span. With independent variables set at mean values, or

at the pre-sample period (1992), results, which are available on request, did not change.

Our most important check, however, is that our results are robust to the use of a

different sampling method. We built a control sample of firms without SSI patents. From the

LECG-Corptech database we selected all the firms with business activity in SSI that are not

included in our initial sample.5 This produced a control sample of 79 firms that have some

business activity – product or services – in SSI. We do not know whether these firms have

produced an innovation, even though we know that they have not produced a patented

innovation. After collecting all the required variables, we performed the same regressions by

using a sample of 166 firms (87+79) over 8 years, which resulted in 1,328 observations. The

control sample average size and age are equal to USD 2,469,300 and 14 years respectively, very

similar to our initial sample data. We show the results of these estimations in Table 6, Models 1-

4, using only PATENT for simplicity (but the results are similar for all the patent specifications).

Table 6 confirms our earlier findings. The main exception is that patents are now significant and

positive in the product equations as well. This only confirms that our initial sampling method is

correct: with no-patent firms in the sample, patents still capture an innovation effect. We will

discuss this point further in the next section.

[TABLES 6 ABOUT HERE]

5 The LECG software database was obtained from CorpTech (Corporate Technology Information Services), and it provides information on more than 15,000 software companies active in the US. The LECG dataset reports whether a firm has ever been active in a certain business segment of the software industry, in our case Security Software. For more details on the LECG-CorpTech dataset see Lerner and Zhu (2005).

17

Finally, we performed the same regressions with the original sample of 87 firms but

without the five largest firms (first quintile of size distribution). This checks whether our results

are driven by some outlier observations. The results, shown in Table 6, Models 5-8, are very

similar to those for the 87 firm sample.

5. Discussion

5.1 Patents

This is one of the first studies that show how patents support technology trade. Our claim is that

patent stocks capture something more than a pure innovation effect. In our view more patents

means fewer transaction costs both for the supplier – innovation protection – and for the buyer –

fewer search costs. Algorithm technology is a complex innovation, and most firms, particularly

the start-ups, have just one or few algorithms. Therefore, the discontinuity of the innovation

event means that patents tend to increase more than proportionally with the number of

algorithms. The first patent accounts for the development of a new algorithm, while additional

patents reflect factors other than the sheer number of innovations.

Our initial sampling method eliminates part of the innovation effect associated with

patents. We cannot be sure that the firms with no-patents have not developed an algorithm;

however, we know that the firms with patents are quite likely to have one. So, with a sample of

firms with patented technologies, the number of patents is more likely to reflect other aspects

than the sheer number of innovations. In this respect, beyond the control of firm R&D, age and

size, patents affect only licensing, and not product launching. Our interpretation is

straightforward. If patents are measuring IPR protection or innovation visibility, they do not

raise significantly the odds of a product innovation – beyond our controls. On the contrary, they

increase the probability of licensing an algorithm. When we employ a firm sample with or

without patents, patent variables gain importance as a measure of innovation. As a result, not

only are patents important for licensing, but they also raise the probability of launching a new

product. In sum, SSI firms need an innovation to launch a product or to sell a technology.

18

However, for the product strategy they could buy innovations from third parties and use other

tools rather than patents to block rivalry, while for the licensing strategy they need patents to

lower transaction costs.6 Also, we cannot exclude that patents help firms extract more rents from

their products.

5.2 Industry scenario

Our data confirm the predictions of Scenario 1. The industry is young with no-shake-out, at

least so far. Because of its fragmented downstream co-specialized assets, market niches

proliferate and patents encourage technology sale. Downstream co-specialized assets are

important for operating in final markets, but because of their fragmentation there are weak

threats of business cannibalization that could influence the propensity to license.

In addition, not only is SSI segmented in at least six major product niches dominated by

different specialized firms (mostly young start-ups), but the demand structure in SSI also

reflects two distinct types of customers. On the one hand, there are large ICT firms that

represent technology-skilled, highly-selective customers that choose the best products on the

market (hardware, telecommunication, and semiconductor producers). On the other hand, we

have technologically less sophisticated customers like banks, financial institutions and credit

cards that demand global security packages (Giarratana 2004; Fosfuri and Giarratana 2006).

Thus, we can identify two types of producers and sellers of SSI products. The first type

is firms specialized in niches that release innovative products in only one niche and that are

typically sellers of algorithms (RSA Data Security, Verisign, Certicom, Check Point Software).

The second type is more service-oriented firms that create diversified portfolios of security

products serving the less technologically advanced customers with a global security demand.

Thus, the route of success for a start-up is to enter and specialize in a particular sub-market

niche trying to remain at the innovation frontier in that niche and investing in downstream assets

6 Interestingly, none of the 79 firms with no patents that we added to our sample produced a technology license in the period that we considered. This suggests either one of the following situations. First, firms with an algorithm also hold patents. Since firms can license an algorithm only if they have it, only firms with patents license. Second, even if firms without patents hold an algorithm, patents encourage their licensing.

19

that, being niche-specific, are not very costly and can be defended easily. Start-ups could also

increase their revenues and reputation by selling their technology to firms specialized in other

niches or to large incumbents, which own diversified product portfolios and can sell to

customers that the start-ups will rarely be able to reach (e.g. banks and financial institutions). In

unreported regressions, if we introduce the number of licensing contracts signed as a technology

buyer to explain the probability of releasing a product, the sign of this covariate is positive and

significant, while all other variable signs and standard errors remain largely similar.

Finally, for firms with a patented innovation, patents increase the probability of

licensing the inventions, but they do not have a clear effect on entry in downstream markets.

This is because, when markets for technology are well-functioning, like in SSI, firms with no

innovations can buy them from third parties and embed them in proprietary products. Therefore,

patent effectiveness, which allows for a good functioning of this market for technology, is also

the source of segmentation of the downstream product market. This is because the firms without

innovations can buy them and serve market niches that would otherwise remain unexploited.

6. Conclusions

This paper tried to develop the theoretical framework that supports the firm decision of

commercially exploiting innovations. In this respect, we try to improve on the current literature

(Teece 1986; Arora et al. 2001; Gans et al. 2002; Arora and Ceccagnoli 2006) by developing a

more comprehensive framework in which the choice to launch a product or to license a

technology are predicted from different combinations of patent importance and nature of

downstream co-specialized assets. In our theory section, we merge two literature traditions: the

studies on the role of IPRs and downstream co-specialized assets (Gans and Stern 2003; Arora

and Ceccagnoli 2006; Fosfuri 2007), and the industrial organization contributions on industry

evolution (Thomspon and Klepper 2007). Given that downstream co-specialized assets can be

either homogeneous or fragmented in several sub-market niches, typically according to the stage

of industry evolution, we defined four scenarios in which the control of downstream assets and

20

the firm propensity to patent could produce alternative effects on the probability of selling a

disembodied technology or releasing a product in the final market.

To our knowledge, together with Gans et al. (2002), this is one of the first attempts to

collect data on both technology licensing and products from existing databases in order to

jointly test hypotheses about company behavior on these matters. We tested our predictions

using data on the Security Software Industry. This is a new industry that has not featured an

industrial shake-out, and it is characterized by proliferation of market niches and important

functions of patents. Our findings confirm that downstream assets are pivotal in the final market

penetration, but have no effect on the licensing probability. Conversely, patents increase the

firm probability of selling technology, but do not have a clear effect on entry in final market.

Our results are useful to both managers – especially technology entrepreneurs – and

policy makers. For managers, we show that firms should seriously consider the alternative

options of selling technology and not just final products. This could help in particular the young

ventures that may increase their revenues and chance of survival. As a matter of fact, most

Security Software firms patent cryptography inventions without considering the business

potential of markets for technology. From the interviews that we conducted we found that the

market for technology was a demand-pull event, initiated by large incumbents who seek for

algorithms owned by small ventures. A more structured strategic thinking about how firms

could profit from innovation in the market for products and technologies could not only

encourage more technology sales, but also increase substantially the effectiveness of firm

investments in R&D and patents. For example, firms may not invest in innovation because they

do not have the downstream assets to incorporate the innovation in final products. But if they

consider technology licensing, they may do so even if they do not own such assets.

Our main suggestion for policy makers is that the effectiveness of patent regimes should

be adapted to industrial downstream conditions. We have shown that patents could be really

important in markets that tend to be highly segmented. These are usually general purpose

industries that produce products with several end-use applications and characterized by a

proliferation of sub-market niches (i.e. lasers or semiconductors). In these industries, wherein

21

the control of downstream assets is spread across several organizations, technology deals can

facilitate the quick exploitation of many market niches and they can make it possible to reach

many customers. High transaction costs in technology deals, possibly produced by an imperfect

patent regime, could especially harm the development of these markets. We do not know if SSI

could have experienced the evolution trajectory that we have described if patentability in

software had not gone through its reforms of the 1980s. Needless to say, patent reinforcement

and effectiveness could play a different role in industries where downstream assets are

homogenous and oligopolistic forces prevail. In this respect, our key conclusion is that in order

to assess the implications of stronger patents one needs to assess carefully the nature of

downstream product markets and more generally industry structure and the stage of life cycle in

which the industry is in. Hopefully, our Scenarios have helped provide some keys to

understanding these different implications.

References

Arora A., Ceccagnoli M. 2006. Patent Protection, Complementary Assets, and Firms:

Incentives for Technology Licensing. Management Science 52: 293-308.

Arora A., Fosfuri A. 2003. Licensing the Market for Technology. Journal of Economic

Behaviour and Organization 52: 277-295.

Arora A., Fosfuri A., Gambardella A. 2001. Markets for technology: Economics of

innovation and corporate strategy. Cambridge, Mass.: MIT Press.

Arora A., Gambardella A. 1994. The changing technology of technological change:

general and abstract knowledge and the division of innovative labour. Research Policy 23 (5):

523-532.

Arora A., Merges R. 2004. Specialized supply firms, property right and firm

boundaries. Industrial and Corporate Change 13(3) 451–475.

22

Arrow, K.J. 1962. Economic Welfare and the Allocations Resources for Inventions.

R.R. Nelson, ed. The Rate and The Direction of Inventive Activities. Princeton University Press,

Princeton NJ.

Bresnahan T.F., Greenstein S. 1997. Technological Competition and the Structure of the

Computer Industry. Journal of Industrial Economics 47(1): 1-40

Bresnahan T.F., Trajtenberg M. 1995. General Purpose Technologies, Engines of

Growth. Journal of Econometrics 65 (1): 83-108

Cassiman B., Veugelers R. 2006. In Search of Complementarity in the Innovation

Strategy: Internal R&D and External Knowledge Acquisition’, Management Science 52(1): 68-

82

Cohen W., Levinthal D. 1990. Absorptive capacity: A new perspective on learning and

innovation. Administrative Science Quarterly 35(1) 128-152.

Fosfuri A. 2007. The Licensing Dilemma: Understanding the Determinants of the Rate

of Technology Licensing, Strategic Management Journal forthcoming.

Fosfuri A., Giarratana M.S. 2006. Product Strategies and Survival in Schumpeterian

Environments: Evidence from the Security Software Industry. MIMEO, Universidad Carlos III

de Madrid

Gambardella A., Giuri P. and Luzzi A. 2006. The Market for Patents in Europe.

MIMEO. Università Bocconi, Milan.

Gans J., Hsu D., Stern S. 2002. When does start-up innovation spur the gale of creative

destruction? RAND Journal of Economics 33(4): 571-586.

Gans J.S., Stern S. 2003. The Product Market and the Market for Ideas:

Commercialization Strategies for Technology Entrepreneurs. Research Policy 32 (4): 333-350.

Giarratana M.S. 2004. The birth of a new industry: entry by startups and the drivers of

firm growth. The case of encryption software. Research Policy 33(5): 787-806

IDC 2002. Worldwide software security report.

Jensen R., Thursby M. 2001. Proofs and Prototypes for Sales: The Licensing of

University Inventions. American Economic Review 91 (1): 240-259.

23

Jovanovic B., MacDonald G.M. 1994. The Life Cycle of a Competitive Industry.

Journal of Political Economy 102(2): 322-347

Klepper S. 2002. Firm survival and the evolution of oligopoly. RAND Journal of

Economics 33(1): 37-61

Klepper S., Thompson P. 2007. Submarkets and the Evolution of Market Structure, The

RAND Journal of Economics, forthcoming

Lerner, J., Zhu, F. (2005) What is the Impact of Software Patent Shift? Evidence from

Lotus vs Borland. NBER Working Paper # 11168

McGahan, A.M., B.S. Silverman. 2001. How does innovative activity change as

industries mature? International Journal of Industrial Organization 19(7): 1141-1160

McGahan, A.M., B.S. Silverman. 2006. Profiting from Technological Innovation: The

Effect of Competitor Patenting on Firm Value. Research Policy, forthcoming

Menard S. 1995. Applied Logistic Regression Analysis. Sage Publications: London

Pennings J.M., Harianto F. 1992. The Diffusion of Technological Innovation in the

Commercial Banking Industry, Strategic Management Journal 13(1): 29-46

Teece DJ. 1986. Profiting from Technological Innovation. Research Policy 15(6): 285-

305.

24

Table 1: Four theoretical scenarios according to patent effectiveness and fragmentation of downstream co-specialized assets

Fragmentation of downstream co-specialized assets

High Low

High

Scenario 1

Licensing and product

releases could be mutually

exploited

Examples: earlier phases of

industry evolution or no-

shake industries where

innovation and IPR are well

defined (Lasers, Semicond-

uctors)

Scenario 2

Case a): Limited licensing

Case b): Licensing with

classical division of

innovative labour between

small and large firms

Examples: a) later phases of

industry evolution or shake

industries where innovation

and patents are important

(consumer electronics,

computer hardware) and b)

innovation and downstream

assets are own by different

organizations (Biotech)

Patent

effectiveness

Low

Scenario 3

Hard to license even when

there is willingness to

Examples: earlier phases of

industry evolution or no-

shake industries in which

patents are not effective

(Software in the 1980s)

Scenario 4

Limited licensing. Mature

sectors.

Examples: later phases of

industry evolution or shake-

out industries where patents

are not effective (Petro-

chemicals)

25

Table 2: Product niches in the SSI

Niche Description

Authentication-Digital Signature

Products for authentication of digital documents with a copyrighted mark

Antivirus Programs that detect and clean viruses from computers

Data and Hardware Protection

Products securing integrity of sensible data stored in hard drivers

Firewalls A kind of checking door between different networks

Utility Software Utility software programs that assure protection and execution of operating systems and applications, giving the possibility to recreate the content of some data packages lost

Network Security and Management

Network security management packages that guarantee the high performing functioning of different networks

Source: Fosfuri and Giarratana (2007) and Giarratana (2004)

Table 3 Descriptive statistics Mean Dev, stand. Min Max Dependent variables PRODUCT 0.176 0.381 0 1 LICENCE 0.116 0.320 0 1 Independent variables Core variables PATENT 71.33 235.26 1 1,941 PATENT10 57.96 185.69 1 1,487 PATENT15 52.40 166.48 1 1,361 ASSET1 507,490 2,097,629 0 18,480,400 ASSET2 1,024,377 3,752,077 0 36,451,211 Time variant controls AGE 17.99 22.31 0 124 SIZE 2,816,394 9,551,466 76 87,548,000 R&D 0.129 0.111 0 0.593 Time invariant controls SOFTWARE 0.563 0.496 0 1 HARDWARE 0.126 0.332 0 1 ELECTRONICS 0.137 0.345 0 1 NORTH AMERICA 0.850 0.356 0 1 EU 0.068 0.253 0 1

26

Table 4 Logit regressions predicting the probability of licensing an algorithm

Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Independent variables Core variables PATENT

0.227** (0.118)

0.225* (0.121)

PATENT10

0.248** (0.121)

0.242** (0.124)

PATENT15

0.258** (0.123)

0.225** (0.121)

ASSET1

0.038 (0.035)

0.038 (0.035)

0.038 (0.035)

ASSET2

0.014 (0.033)

0.014 (0.033)

0.014 (0.033)

Time variant controls AGE

-0.702** (0.171)

-0.705** (0.171)

-0.706** (0.171)

-1.267** (0.328)

-1.271** (0.328)

-1.267** (0.328)

SIZE

0.092 (0.109)

0.082 (0.109)

0.077 (0.109)

-0.350** (0.189)

-0.353** (0.188)

-0.350** (0.189)

R&D

2.911* (1.680)

2.853* (1.680)

2.820* (1.680)

0.472** (0.179)

0.466** (0.178)

0.472** (0.179)

Time invariant controls CONSTANT

-3.182** (1.478)

-3.042** (1.488)

-2.970** (1.492)

-3.976** (1.851)

-3.922** (1.850)

-3.976** (1.851)

Core sector dummies Geographical dummies Year dummies

YES YES YES

LogL -177.7 -177.5 -177.4 -178.0 -177.8 -178.0 Notes: * indicates p<0.10, ** indicates p<0.05. Heteroskedastic consistent standard errors in parenthesis. 696 observations. All covariates (but the dummies) are in logs.

27

Table 5 Logit regressions predicting the probability of launching a SSI product

Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Core variables

PATENT 0.067

(0.110) 0.065

(0.105)

PATENT10 0.069

(0.113) 0.065

(0.107)

PATENT15 0.072

(0.114) 0.068

(0.108)

ASSET1 0.097** (0.033)

0.097** (0.033)

0.097** (0.033)

ASSET2 0.076** (0.029)

0.076** (0.029)

0.076** (0.029)

Time variant controls

AGE -0.634** (0.155)

-0.634** (0.155)

-0.634** (0.155)

-0.761** (0.249)

-0.761** (0.249)

-0.761** (0.249)

SIZE 0.747

(0.473) 0.746

(0.473) 0.745

(0.473) -0.171 (0.296)

-0.171 (0.295)

-0.171 (0.295)

R&D 3.644** (1.663)

3.633** (1.666)

3.620** (1.667)

3.591** (1.669)

3.579** (1.671)

3.566** (1.672)

Time invariant controls CONSTANT

-5.598** (1.379)

-5.574** (1.395)

-5.546** (1.401)

-4.241** (1..438)

-4.248** (1.434)

-4.239** (1.429)

Core sector dummies Geographical dummies Year dummies

YES YES YES

LogL -218.8 -218.8 -218.9 -219.1 -219.1 -219.1 Notes: * indicates p<0.10, ** indicates p<0.05. Heteroskedastic consistent standard errors in parenthesis. 696 observations. All covariates (but the dummies) are in logs.

28

Table 6. Robustness checks

With control sample Sample Without the 5 largest firms Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7 Model 8 Licensing Product Licensing Product Core variables

PATENT 0.394** (0.115)

0.386** (0.114)

0.184** (0.075)

0.191** (0.074)

0.219* (0.118)

0.206* (0.119)

0.054 (0.111)

0.045 (0.111)

ASSET1 0.032

(0.041) 0.141** (0.032)

0.038 (0.035)

0.098** (0.033)

ASSET2 0.048

(0.046) 0.156** (0.035)

0.062 (0.040)

0.116**(0.037)

Time variant controls

AGE -1.197** (0.215)

-1.195** (0.215)

-0.589** (0.166)

-0.600** (0.166)

-0.712** (0.172)

-0.713** (0.173)

-0.649** (0.156)

-0.645**(0.156)

SIZE -0.200* (0.106)

-0.187* (0.107)

0.083 (0.075)

0.073 (0.075)

0.090 (0.109)

0.072 (0.109)

0.291 (0.207)

0.278 (0.207)

R&D 5.466** (1.797)

5.471** (1.805)

3.118** (1.613)

3.056** (1.613)

2.856* (1.675)

2.880* (1.686)

3.606** (1.658)

3.558** (1.663)

Time invariant controls CONSTANT

-12.09** (1.480)

-5.355 (1.489)

-6.638** (0.938)

-4.370** (0.935)

-3.160** (1.480)

-3.046** (1.482)

-5.520** (1.377)

-5.433**(1.374)

Core sector dummies Geographical dummies Year dummies

YES YES YES

LogL -138.2 -137.9 -242.1 -241.9 -175.1 -174.4 -213.8 -213.1 Notes: * indicates p<0.10, ** indicates p<0.05. Heteroskedastic consistent standard errors in parenthesis. Model 1-4 1,328 observations, Model 5-8 656 observations. All covariates (but the dummies) are in logs.

29