Testing patent value indicators on directly observed patent value—An empirical analysis of Ocean Tomo patent auctions

Testing patent value indicators on directly observed patent value—An empirical analysis of Ocean Tomo patent auctions

Research Policy 43 (2014) 519–529 Contents lists available at ScienceDirect Research Policy journal homepage: www.elsevier.com/locate/respol Testin...

745KB Sizes 0 Downloads 0 Views

Research Policy 43 (2014) 519–529

Contents lists available at ScienceDirect

Research Policy journal homepage: www.elsevier.com/locate/respol

Testing patent value indicators on directly observed patent value—An empirical analysis of Ocean Tomo patent auctions Timo Fischer a,∗ , Jan Leidinger b a b

Schöller Chair in Technology and Innovation Management, Technische Universität München, Arcisstr. 21, D-80333 Munich, Germany Technische Universität München, Germany

a r t i c l e

i n f o

Article history: Received 7 October 2012 Received in revised form 4 July 2013 Accepted 24 July 2013 Available online 28 August 2013 Keywords: Patent value Patent value indicators Patent auctions Patents

a b s t r a c t The valuation of patents is an important, albeit challenging task. Extant research to identify patent value indicators has so far relied on expert estimates of patent value, exploited patent renewal data, or depended on more indirect measures of patent value. Recently, specialized market places for patent transactions have emerged that allow us for the first time to directly observe patent’s private value. One of the most prominent market places for patents is Ocean Tomo, a platform that offers periodical patent auctions. We make use of this auction data to empirically test predictions on patent value identifiers on real-world auction prices. We find empirical support for forward citations and the patent’s family size; however, both indicators explain only a small variance in patent value. In contrast, our full model explains a large share of variance, making us optimistic that with increased directly observed patent value, such models can be useful tools in patent valuation. © 2013 Elsevier B.V. All rights reserved.

1. Introduction Many studies have tried to identify indicators for patents’ private value (see Bessen, 2008 for a recent overview). Unfortunately, we do not observe a transparent liquid market for patents to obtain value estimates. Lanjouw and colleagues (1998, p. 407) summarize that “patent rights are seldom marketed,” and confirm the statement by Schankerman and Pakes (1986, p. 1052), that “their private value is in general unobserved.” Furthermore, to date, no precise or commonly agreed upon approach on monetary patent valuation exist. This difficulty in assessing a patent’s private value has complicated the attempts to identify indicators for patent’s value. Thus, researchers mostly relied on indirect measures to approximate patent value in order to explore patent value indicators, for example, renewal decisions (Bessen, 2008), the value of firms holding patents (Deng et al., 1999; Hall et al., 2005; Lerner, 1994), or the probability of infringement and challenging suits (Harhoff and Reitzig, 2004; Lanjouw and Schankerman, 2001). Others, like Albert et al. (1991), Harhoff et al. (1999, 2002), Reitzig (2003), and Gambardella et al. (2008) chose to obtain subjective patent value estimates from patentees, a set of experts, or the inventors in order to test patent value indicators. However, recently specialized platforms for patent transactions have emerged that facilitate observing patent’s private value. Ocean

∗ Corresponding author. Tel.: +49 89 289 25760. E-mail address: fi[email protected] (T. Fischer). 0048-7333/$ – see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.respol.2013.07.013

Tomo, a platform that offers periodical patent auctions, is one of the most prominent market places. We make use of the resulting auction data to test patent value indicators. Thus, while all of the prior studies were forced to employ only estimates of patent’s private monetary value, our study uses real-world patent auction data of more than 1800 patents. Since 2006, Ocean Tomo has hosted periodical auctions of intellectual property (IP). In the majority of cases, patents are auctioned, but sometimes trademarks, copyrights, and domain names are also offered. Selling entities range from individual inventors or investors, academic institutions, midsized companies to large corporations, and government agencies. We collected data from the auction catalogs and matched it to the outcome of each auction published by Ocean Tomo. We complement this dataset with patent-level data from PATSTAT and INPADOC patent databases. Eventually, our dataset included 1784 U.S. patents, 617 of which had been successfully sold at Ocean Tomo’s auctions. The patents sold on Ocean Tomo between 2006 and 2009 have a high share of information technology patents and are potentially not representative for all patent sales. This makes it difficult to compare our descriptive results on observed patent value to patent value estimates of previous studies that looked at different time frames and used less selective samples in detail. Nonetheless, we can report that the patent values observed in our study are roughly consistent with previously obtained patent value estimates. For the testing of patent value indicators, we use Heckman models to control as best as possible for any sample selection effects. However, if patents of some industries where not offered or sold at Ocean

520

T. Fischer, J. Leidinger / Research Policy 43 (2014) 519–529

Tomo, we cannot generalize our results based on them. To test for the predictive power of various patent value indicators discussed in the extant literature, we estimate regressions on the auction price of sold patents, taking into account the sample selection of sold over unsold patents and the sample selection of patents that were offered at Ocean Tomo over average technology and cohort matched patents. We find that forward citations and family size are significant indicators for patent value but find no support for the relation between value and the number of International Patent Classification (IPC) classes a patent applies to. Most interestingly, our full model has high explanatory power for patent value while the explanatory power of single patent value indicators is limited. 2. Theory and hypotheses The definition of a patent’s private value followed in this paper is in line with Harhoff and colleagues’ (2002, p. 1345–1348) definition of the asset value of a patent. They define a patent’s value by the benefits that the winner of a patent race will perceive. When a firm acquires a patent, it gains all associated rights including the right to exclude competitors from using the underlying invention and the right to block other patent rights that depend on the one transferred. Firms that unsuccessfully compete for the patent right suffer the consequences of a competitor becoming the leader. The difference in profits between the two options constitutes the asset value of the patent right. Early estimates of patent value showed a highly skewed distribution (e.g., Schankerman and Pakes, 1986). This considerable variation in the value of patents spurred the research for patent value indicators. In the following, we focus on three of the most often examined patent value indicators, which we chose to test using patent auction data.1 Nearly every researcher who examined patent value with the help of indicators from patent databases included the number of forward citations in his or her studies (e.g., Gambardella et al., 2008; Harhoff et al., 1999, 2002; Lanjouw and Schankerman, 1999; Trajtenberg, 1990), and all of them have assessed a significant and positive relationship between them. The number of family members of a patent (Gambardella et al., 2008; Harhoff et al., 2002) and the breadth of a patent (Harhoff et al., 2002; Lanjouw and Schankerman, 2001; Lerner, 1994) have similarly sound theoretical foundations, but while the size of a patent’s family is consistently significant as a value indicator, the results on the scope of a patent are more ambiguous.2

a patent receives (e.g., Hall et al., 2005; Harhoff et al., 1999; Trajtenberg, 1990) in the patent examination process, prior art that would limit the patent’s novelty is identified. Patents representing prior art are cited and they receive forward citations. The more forward citations a patent receives, the higher is its contribution to the prior art, making it a good proxy for patent’s technological quality. Thus, we posit: Hypothesis 1. The number of forward citations a patent receives is positively related to its private value. 2.1.2. Patent’s economic relevance measured by family size The value of the patent should also depend on both the technological quality and the economical relevance of the underlying invention. Even if inventions are comparable regarding technological quality, market sizes or industry characteristics may differ, giving them diverging economic qualities. To proxy the economic relevance, we make use of the patent’s family size (Harhoff et al., 2003; Lanjouw et al., 1998; Putnam, 1996). A patent’s family size captures the number of jurisdictions in which patent protection for a single invention has been sought. The expansion of patent protection involves additional costs—e.g., translation, patent attorneys’ filing fees, examination fees—for every jurisdiction. If the applicant chooses to spend additional money, the exclusion right should be worth the extra costs. Hence, we posit: Hypothesis 2. The number of family members of a patent is positively related to its private value. 2.1.3. Patent scope measured by distinct IPC classes Furthermore, the scope of a patent should be related to its value. Broad patents read on many products or processes and hence increase the attractiveness of the right of exclusion (Merges and Nelson, 1990; van Zeebroeck et al., 2009). Furthermore, competitors will find it more difficult to “invent-around” a broader patent, adding value to the exclusion right. To proxy the scope of a patent, we used the number of distinct four-digit IPC classes to which the patent is assigned (Lerner, 1994). Hypothesis 3. The number of distinct IPC classes to which a patent is assigned is positively related to its value. 3. Empirical approach 3.1. Empirical setting

2.1. Indicators of patent value 2.1.1. Patent’s technological quality measured by forward citations From early on, the technological quality of a patent has been related to its value (Albert et al., 1991; Green and Scotchmer, 1995; Nordhaus, 1967). The higher a patent’s technological quality, the higher the patent’s legal robustness should be (e.g., Bessen, 2008; Reitzig, 2003). Furthermore, the higher the patent’s technological quality, the more inventions should build upon the underlying invention of the focal patent, thus increasing the value of its exclusion right. A widely accepted patent value indicator that captures patent’s technological quality is the number of forward citations

1 Previous studies also examined indicators such as the outcome of opposition cases for European patents (Harhoff et al., 2002), the number of backward references (e.g., Gambardella et al., 2008; Harhoff et al., 2002) or the number of claims (Bessen, 2008; Gambardella et al., 2008). As we analyze U.S. patents, we only include the latter two as control variables in our calculations. 2 Insignificant in Harhoff et al. (2002) and Harhoff and Reitzig (2004); significant and negative in Lanjouw and Schankerman (2001); significant and positive in Lerner (1994).

We make use of Ocean Tomo’s patent auctions data, which allows us to observe the private value of patents (cf. Schankerman and Pakes, 1986). Ocean Tomo claims to have held the first public auction of IP rights, such as patents, trademarks, and copyrights, in 2006 (cf. Tietze, 2012 for a detailed presentation of Ocean Tomo auctions). Between 2006 and 2010, Ocean Tomo held 10 different auctions. Every patent auction follows the same structure. First, auction date and location are announced by Ocean Tomo, followed by the registration of sellers and the patents they have on sale. These patents are evaluated by Ocean Tomo Patent Ratings, a specialized patent rating agency. Patents that meet certain quality standards set by Ocean Tomo (which are not disclosed) are accepted and published in the auction catalog. The next phase consists of the registration of the potential bidders and due diligence procedures that include private meetings or conference calls between seller and potential buyers. Finally, the auction itself takes place. Usually the auction (in fact, a series of auctions of many “lots”) is embedded into a two-day program of conferences and get-togethers at an exclusive and varying location. Bidder anonymity is secured by Ocean Tomo by identifying bidders by paddle number only. Bidders can even choose “double-blind bidding” by requesting Ocean

T. Fischer, J. Leidinger / Research Policy 43 (2014) 519–529

Tomo to provide a representative to execute the bidding on his or her behalf using written instructions. Bidding is also possible by telephone or via the Internet. As a further measure of bidder protection, Ocean Tomo states in its Bidding Procedures and Conditions of Sale brochures that it requires every seller to sign a pledge agreeing not to use the bidder’s participation as evidence in any action seeking to enforce a patent for any purpose. Patents that have not been sold remain available in the post-auction sales period. Several lots were auctioned in each of the 10 auctions we studied. A ‘lot’ consists of one to several patents offered by the same seller and contains at least one U.S. patent. Every patent in the lot is sold along with its complete family. While in most cases the patents offered within the same lot belong to the same product or at least the same industry, the grouping is done by the seller and in some cases a lot may contain non-related patents. Interestingly, sellers have the opportunity to set the minimum auction price they are willing to accept for the lot. In case the seller does not specify any minimum amount, Ocean Tomo applies a required reserve amount of $10,000. If the highest bid does not meet the reserve price, the transaction will not be realized. Ocean Tomo employs a first-price English auction with ascending prices (Tietze, 2008). Sellers pay a fixed registration fee of $1000 per lot, in cases where the minimum reserve price is set at $10,000. The fee rises to $3000 per single-patent lot or $6000 per multipatent lot if the seller decides to set a higher reserve price. Bidders have to pay a fixed registration fee of $1500. The final price to be paid by the buyer comprises the final bid price plus the buyer’s premium of 10% of the final bid price and any applicable taxes. The 10% buyer’s premium, plus the 15% seller’s premium fee on the final bidding price for successful sales, plus the registration fees constitute Ocean Tomo’s revenue at the auction. In 2009, Ocean Tomo sold its transaction division, including the auction business, to ICAP for the sum of $10,000,000. Unfortunately, ICAP has not continued Ocean Tomo’s policy of publishing detailed auction results. Thus, we were not able to include the results of the eleventh and subsequent auctions in this study. 3.2. Dataset The auction catalogs, published before each auction event, provide detailed information about each lot, the patents to be offered, and the name of the seller. Even a sample forward citation analysis is provided along with the names of possible licensees. Some of the catalogs feature—if provided by the seller—the expected value of the lot, while the reserve price established by the seller remains unknown in all cases. After the auction, Ocean Tomo reports the prices of the successfully sold lots. Unfortunately, Ocean Tomo does not report information on every bid made, only the final bid. Beginning with the first live patent auction in the spring of 2006 and ending with the last Ocean Tomo auction in July 2009, the dataset incorporates the catalogs of 10 live IP auctions with a total of 678 lots containing 1784 patents. In order to be able to focus solely on U.S. patents, we excluded 33 non-U.S. patents from the analysis. Furthermore, we dropped 76 U.S. patents that had not reached five years since the filing date, as it is standard practice to be able to account for truncation effects on forward citations by observing the same five-year time span of forward citations for all patents in the dataset. Of the 678 lots in our dataset, 262 have been successfully sold. Of these 262 sold lots, 167 contain only one patent while 95 lots contain multiple patents (on average 4.28). We complemented the dataset with information provided by the European Patent Office (EPO) from the INPADOC database as of April 2009 (EPO, 2009a) and the patent statistical database (PATSTAT) as of April 2009 (EPO, 2009b). The INPADOC database contains data on the legal events

521

tied to patents in their time of validity. The PATSTAT database contains patent bibliographic data. Both EPO databases contain data on U.S. patents, which we use in this study. 3.3. Estimation strategy We use the auction price as our main dependent variable, employing logarithmic properties since this variable is highly skewed. If the patents are part of a multi-patent lot, we calculate the average selling price for all patents of such lot. As independent variables, we use five-year truncated forward citations, the family size, and the number of distinct four-digit IPC classes. We use only forward citations received up to five years after patent application in order to obtain the same citation window for each patent in our dataset. In our analysis, we use the INPADOC extended family provided in the PATSTAT database, as is common practice. This measure counts patent applications that share the same priority application or are indirectly linked via priority applications (Martinez, 2010). As not all patents are offered by Ocean Tomo, and as some of the lots that are offered are unsold, selection effects could bias our results. To account for these two selection effects, we first estimate a Heckprobit model with the auction outcome sold/unsold as the dependent variable in the main equation, taking a group comparison between average patents and the patents that are offered as the selection equation. To enable this comparison, we construct a cohort and technology matched control group of random, active U.S. patents. For every patent offered by Ocean Tomo, we drew a random U.S. patent of the same four-digit IPC class that was active at the time of the auction and which was applied for in the same year as the Ocean Tomo match patent.3 As a second step, we compute the Mills ratio of the Heckprobit model and use it as the selection control variable (Heckman, 1979) in our ordinary least squares (OLS) models that employ the lot price as the dependent variable.4 Each Heckman selection model should include a variable in the selection stage that has no theoretical or empirical influence on the dependent variable in the later stage.5 The selection model in our Heckprobit regression (average patent vs. patent offered by Ocean Tomo) includes the number of patents applied for by the selling entity at the United States Patent and Trademark Office as the identification variable.6 While the patent portfolio size of the seller theoretically should neither (and also empirically does not) affect the probability that a patent is sold on Ocean Tomo nor affect the auction price, this identification variable should and does affect the decision to offer a patent on Ocean Tomo. The bigger the patent portfolio of the firm, the smaller the probability that a patent from this portfolio will be offered for sale on Ocean Tomo. Large patent portfolios increase the probability to capture value with patents (Fischer and Henkel, 2013; Hall and Ziedonis, 2001). On the one hand, firms with large patent portfolios that exploit complementarities between their patents have fewer incentives to sell single patents because it would decrease the value of the whole portfolio. On the other hand, these firms can make use of the complementar-

3 Using a cohort and technology matched control group has the unique advantage of controlling for any technology and time differences between the comparison groups in the empirical setup. 4 As the Heckman second stage OLS regressions cannot use standard errors from a simple OLS regression (Heckman, 1979, pp. 158–159) we use bootstrapping to determine the correct Heckman standard errors (Hill et al., 2003). 5 Even if no identification variable exists, identification relying on function form differences between probit and OLS models is feasible. In particular, with large datasets as in our case, functional form differences are sufficient for identification (Cameron and Trivedi, 2009: 543). 6 This variable also captures firm size effects, cf. Harhoff and Wagner (2009) or Serrano (2010). As our sample is not restricted to medium and large firms, we cannot rely on company databases to obtain more direct variables for firm size.

522

T. Fischer, J. Leidinger / Research Policy 43 (2014) 519–529

ities in their patent portfolio to extract more value from individual patents and thus have less incentive to capture value by selling a particular patent. As an identification variable in the second selection stage (sold vs. unsold patents on Ocean Tomo), we chose the number of reassignments. The number of reassignments should and empirically has an influence on the probability that a patent is sold on Ocean Tomo, while it should not and does not have an influence on the resulting auction price. Firms usually sell a patent only if it is no longer necessary to protect the products they offer. The more often a patent is reassigned, the fewer firms rely on the patent. If no firm relies on the patent, it is unlikely that its reserve price is reached and the patent is sold. However, if a firm still relies on the patent, it should be auctioned for its full value to prevent any holdup situations (Reitzig et al., 2007). 3.4. Control variables We control for several variables usually employed in patent valuation studies. 3.4.1. Self-citations Self-citations are those a patent receives from documents belonging to the same assignee. Following Bessen (2008), a large number of self-citations may signal “thicket building” behavior, in which a patentee strengthens its patents by also patenting related technologies. Analogous to the number of forward citations, we account only for citations received until the fifth year of the patent after the filing date. As self-citations are not stored in PATSTAT, we had to allocate many resources for manual screenings of all citations within the five-year-truncation window.7 3.4.2. Backward references We also control for the number of backward references of a patent. A patent makes several references to other publications considered to be relevant prior art. If patents are found that are held against the claims of the patent applied for, so-called backward references to prior patents are placed. To avoid distortion of the data by companies that consistently pursuit a strategy of citing own patents, these so-called self-citations are subtracted from the total number of backward citations to the patent literature. However, the interpretation of the number of backward references a patent receives is not clear. While it has been suggested that researchers measure the amount of extant technology in a technology field (Ziedonis, 2004), other scholars argue that it also measures the scope of the patent (Harhoff et al., 2003), while others argue that it is a measure for the patent’s technical novelty (Carpenter et al., 1980; Reitzig, 2003). Patents also reference non-patent literature, which for the most part refers to articles in scientific journals. The number of these references can be used as a proxy for the proximity of the patent to science (Meyer, 2000; Narin and Noma, 1985; Narin et al., 1987, 1997). 3.4.3. Claims Furthermore, we control for the number of claims a patent makes. The number of claims may serve as another indicator of a patent’s scope next to the number of assigned IPC classes. However, this patent characteristic is also ambiguous. Despite being used as a measure for a patent’s scope, some scholars argue

7 In the first selection stage, where we compare patents offered on Ocean Tomo with cohort matched patents, we have no truncation effects since each tuple of comparison patents have the same age. As the amount of citations to check is prohibitively high in this setting, we cannot differentiate between forward selections and self-selections in the first selection stage.

Table 1 Ocean Tomo sales 2006–2009. Auction

Patents

Lots

Patents sold

Lots sold

Sales (mil. $)

Spring 2006 Fall 2006 Spring 2007 Summer 2007 Fall 2007 Spring 2008 Summer 2008 Fall 2008 Spring 2009 Summer 2009

421 251 163 155 133 166 101 158 169 67

77 71 65 44 76 83 60 99 77 26

99 36 90 35 56 115 57 58 23 4

27 21 35 13 37 51 26 43 6 3

3.25 5.74 12.53 7.22 11.26 19.24 9.74 12.15 2.76 1.22

that the number of claims is correlated with the patent’s legal sustainability (Lanjouw and Schankerman, 2001; Reitzig, 2003). The more claims a patent has, the higher the chance that at least one will survive an invalidation procedure. 3.4.4. Generality Furthermore, we control for the generality of the patent. If a patent receives citations from a large number of technology fields, the patent has a “general” impact. Trajtenberg et al. (1997) developed the generality measure that determines the concentration of citations from different technology fields. The higher the generality measure, the more general the impact of the patent. 3.4.5. Patent age We also control for the age of the patent. With increasing age, the underlying technology reaches a higher degree of market penetration. However, the older the patent, the higher the chance that the underlying technology has become obsolete. Furthermore, the lifetime of the patent is limited.8 3.4.6. Additional controls We use several dummy variables to control for technology effects on a patent’s value. Furthermore, we control for the patent’s application period and use a dummy variable to capture whether the patent was prior art of a multi- or a single-patent lot. Lastly, we control for the auction in which the patent was sold. 3.5. Descriptive statistics Table 1 provides an overview of the number of sold and unsold patents in each auction, as well as the volume of total sales. The sales of the auctions in our dataset total $85.11 million, including the 10% buyer’s premium charged by Ocean Tomo on the final bid price. An average auctioned patent values $148,535, which converted to 1992 U.S. dollars for comparison with existing literature is $104,781. This observed patent value, based on real-world patent auction data, is roughly consistent with previous patent value estimates summarized in Bessen (2008). Bessen obtained a mean value estimate of $78,178 for all patentees of patents issued in 1991 and a value estimate of $113,067 for patents of U.S. public manufacturing firms. Barney (2002) obtained an estimate of $61,896 for all patentees of the 1986 patent cohort. Serrano (2005) focused on small patentees and obtained a mean value estimate of $47,456. Putnam (1996) obtained a mean value estimate of $188,355 for patents also filed abroad, for which Bessen imputes a value estimate of $78,800 for all patents.

8

All patents are more than five years old.

T. Fischer, J. Leidinger / Research Policy 43 (2014) 519–529

523

1991 to 2003 to calculate patent value.9 In recent years, smartphone wars and rising patent troll activities should increase the value of information technology- and communications-related patents. If we convert the auction sale prices to 1992 dollars to compare them to Bessen’s study, we find the mean value in our study to be 3.4 times higher in information technology and 2.6 times higher in telecommunications. Table 3 illustrates the differences between patents offered on Ocean Tomo and random (while cohort and technology matched) control patents. Patents offered on Ocean Tomo obtained more forward citations,10 contained more backward references, had more family members, were assigned to more IPC classes, and were more general than random patents. Other than with respect to the number of claims, we find no significant difference.11 Table 4 compares sold and unsold patents using t-tests. We find that the truncated number of forward citations is significantly higher among sold patents. The family size of sold patents is significantly higher than that of unsold patents. On the other hand, the number of IPC classes is not significantly different between sold and unsold patents. 4. Results 4.1. Heckprobit selection models Fig. 1. Share in offered patents and sold patents at Ocean Tomo live auctions and share in total number of active U.S. patents in 2009 by OST-INIPI/FhG-ISI technology nomenclature.

Fig. 1 shows the shares in offered and sold patents at Ocean Tomo’s live auctions sorted by technology fields and a comparison to all active U.S. patents in 2009. More than 53% of the patents offered at Ocean Tomo auctions belong to only three technology fields defined by the OST-INIPI/FhG-ISI technology nomenclature (OECD, 1994, p. 77): “information technology,” “telecommunications,” and “analysis, measurement, control technology.” Optics and audiovisual technologies add another 11%. Evidently, live auctions especially attract sellers of IP related to communications and information technologies. The percentage of patents offered at auctions that belong to any one of the above three categories is substantially higher than their share in the total number of U.S. patents active in 2009 (Fig. 1). Table 2 shows how the total sales of the auction distribute over technology fields of patents sold. Information technology patents generated the highest sales volume with $53.18 million, followed by $15.75 million generated by telecommunication patents, $7.51 million generated by patents in the “analysis, measurement, control technology” field, and $4.28 million in “optics and audiovisual technology.” In all four major technology fields, median values are much lower than the mean, indicating the skewed patent value distribution usually observed in literature on patent value. As a large share of patents auctioned on Ocean Tomo is related to information technology, the observed patent values are hard to compare in detail to other studies that found the majority of patent value concentrated in chemical, drug, and medical technologies (Bessen, 2008). A finding that is similar to Bessen, who also analyzed U.S. patents, is the relationship between median and mean values for patents in computers and communications related fields. While Bessen finds the mean value to be 2.1 times higher than the median, we observe a factor of 1.5 in information technology and 2.7 in telecommunications, suggesting a roughly similar value distribution (at least compared to chemical, drugs, and medical where Bessen observes factor 10 and more). The absolute median and mean values for information technology-related patents are hard to compare to extant literature, since we look at completely different time frames. Bessen uses data from renewal decisions from

We first estimate a Heckprobit model, which estimates the first selection stage (average patents vs. patents offered on Ocean Tomo), and the second selection stage (sold vs. unsold patents) jointly. Model 1a in Table 5 shows the first selection stage, the selection equation in the Heckprobit model comparing 1784 patents offered on Ocean Tomo with 1784 random patents that are matched by cohort and technology. Interestingly, a higher number of forward citations, a higher number of backward references, and a higher generality increase the probability that patents are offered on Ocean Tomo. Model 1b in Table 5 shows the second selection stage, the main equation in the Heckprobit model, comparing 573 sold patents on Ocean Tomo with 1211 unsold patents. While a higher number of forward citations, backward references to non-patent literature, a larger family size, and being part of a single-patent lot increase the probability that a patent is sold on Ocean Tomo, patent age decreases this probability. Both selection stages have low explanatory power; McFadden’s Pseudo R2 is .085 in Model 1a and .062 in Model 1b. 4.2. OLS regressions on patent price Table 6 shows the results of the OLS regressions that we rely on for hypotheses testing. Model 2a shows results without taking the selection stages from the Heckprobit into account. Model 2b shows results including a selection control variable based on the Heckprobit selection stages. Model 2c shows estimation results including a selection control but using only patents that were part of singlepatent lots, while Model 2d shows the result of the same empirical exercise with patents from multi-patent lots. Interpreting Model 2a, we find that forward citations are significantly, positively related to the selling price as hypothesized

9 Studies for French patents (Schankerman, 1998) and German patents (Lanjouw, 1998), use even older data. The same holds for analyses of U.S. patents by Barney (2002), Serrano (2005), and Putnam (1996). 10 As we use a cohort matched control group, we did not have to use truncated citations in this analysis. 11 We also estimate a probit model comparing patents offered on Ocean Tomo with random (while cohort and technology matched) patents. The marginal effects of the indicators of interest obtained in this probit are comparable to the level effects observed in the t-tests. The results are available from the authors upon request.

524

T. Fischer, J. Leidinger / Research Policy 43 (2014) 519–529

Table 2 Patent values in 1000 dollars by technology field (OST-INIPI/FhG-ISI technology nomenclature). Technology field

Patents sold

Mean value

Median value

Information technology Telecommunications Control technology Audiovisual technology Consumer goods Transport Electrical engineering Optics Medical technology Handling printing Chemical engineering Others

247 95 64 21 5 39 10 34 6 7 2 43

215.30 (151.88) 165.77 (116.94) 117.41 (82.83) 203.86 (143.81) 182.60 (128.81) 19.67 (13.88) 73.43 (51.80) 20.94 (14.77) 112.78 (79.56) 52.11 (36.76) 55.79 (39.36) 2.55 (1.80)

143.48 (101.22) 60.50 (42.68) 46.15 (32.56) 112.08 (79.07) 220.00 (155.20) 1.10 (.78) 46.45 (32.77) 20.63 (14.55) 36.83 (25.98) 36.67 (25.87) 55.79 (39.36) 1.57 (1.11)

Total value 53,179.05 15,748.04 7514.00 4281.00 913.00 767.20 734.34 711.79 676.70 364.76 111.57 109.48

Patent values in 1992 $ in parentheses to allow comparison to existing literature.

Table 3 Descriptives and t-test for equality of means between patents offered on Ocean Tomo and IPC/year matched control group. Variable

Forward citations (untruncated) Backward references to patent literature Backward references to non-patent literature Claims Generality Family size Number of IPCs Applied U.S. patents by seller Reassignments

1784 patents offered on Ocean Tomo

1784 technology and cohort matched control patents

t-test

Mean

STD dev.

Mean

p-Value

15.51 20.64 4.10 20.27 .39 14.84 1.79 3520.56 .83

26.17 23.27 9.79 17.55 .31 37.12 1.05 8387.04 1.00

(Hypothesis 1) and that family size is positively related to the selling price as hypothesized (Hypothesis 2). However, we find no empirical evidence for Hypothesis 3. The number of distinct four-digit IPC classes has no significant effect on the patent value. Regarding controls, it is interesting to see that the single-patent lot dummy is significant and positive. Also, the number of claims a patent makes is positively related to the resulting auction price, while the age of the patent has a negative impact. Model 2b, which controls for the selection effects, shows qualitatively identical results. Model 2c and Model 2d break down the data into an analysis of single-patent lot data and multi-patent lot data, respectively. Regarding Hypothesis 1 (number of forward citations), we also find empirical support for Model 2c and Model 2d. Regarding Hypothesis 2 (family size), we do find a positive but not significant relationship in these two models. However, it has to be taken into account that we reduced the size of the dataset and hence the statistical power of the model by splitting data into two groups. The number of IPC classes (Hypothesis 3) is also not significant in

10.96 15.18 2.95 19.66 .32 10.56 1.72 11,627.48 .64

STD dev. 18.82 17.45 8.22 16.68 .30 42.87 1.00 17,905.74 .84

.000 .000 .000 .283 .000 .000 .028 .000 .000

Models 2c and 2d. The number of claims and patent age are also statistically significant in Model 2d, but not in Model 2c. In the following, we discuss marginal effects to study the impact of each patent value indicator on patent value in more detail. Table 7 shows the results of Model 3, which uses exactly the same specifications as Model 2b with the difference that all variables are deflated. The first two data columns in Table 7 show coefficient estimates and standard errors, the second two data columns show marginal effects in U.S. dollars and the proportional increase or decrease of the average patent value in our study if the variable of focus is increased by one unit. In the following, we only interpret variables that are statistically significant. Receiving one more forward citation within the first five years on average increases patent value by an economical significant $14,224. As the average patent in our study values $148,535, this marginal effect is equivalent to a proportional increase in patent value of 9.58%. Adding one more family member increases the patent value by $750. Adding one more claim increases the patent value by $1744. Interestingly, patents sold in

Table 4 Descriptives and t-test for equality of means between patents sold on Ocean Tomo and unsold patents. Variable

Forward citations (five year truncated) Self-citations (five year truncated) Backward references to patent literature Backward references to non-patent literature Claims Generality Family size Number of IPCs Applied U.S. patents by seller Reassignments Patent age Single-patent lot patent

573 patents sold on Ocean Tomo

1211 unsold patents

Mean

STD dev.

Mean

STD dev.

p-Value

2.68 .127 24.14 5.66 19.93 .325 24.67 1.76 1077.70 .625 3599.04 .291

5.22 .573 28.16 11.80 17.15 .311 60.34 1.02 5452.17 .851 1387.43 .455

1.98 .194 18.98 3.36 20.44 .313 10.18 1.81 4345.22 .920 3375.12 .194

3.70 .707 20.36 8.58 15.73 .328 15.56 1.06 9353.17 1.05 1380.27 .396

.001 .049 .000 .000 .569 .432 .000 .291 .000 .000 .001 .000

t-test

T. Fischer, J. Leidinger / Research Policy 43 (2014) 519–529

525

Table 5 Heckman probit comparing patents offered on Ocean Tomo and IPC/cohort matched control group as selection equation and sold patents on Ocean Tomo compared to unsold ones as main equation. Model 1a–Selection equation in Heckprobit – probit comparing patents offered on Ocean Tomo to technology and cohort matched control patents

Model 1b–Main equation in Heckprobit – probit comparing patents sold on Ocean Tomo compared to unsold patents

Observations

3568 (1784 patents on Ocean Tomo/1748 match patents)

1784 (573 sold patents/1211 unsold patents)

Dependent variable

Patent is offered on Ocean Tomo

Patent is sold on Ocean Tomo

Variable

Coef.

SE

AME

SE

ln forward citations (untruncated) ln forward citations (five year truncated) ln self-citations (five year truncated) ln backward references to non-pat. literature ln backward references to patent literature Family size Number of four-digit IPC classes Number of claims Generality Applied U.S. patents by seller Reassignments Patent age Single-patent lot patent dummy Patent application cohort controls Auction controls Four-digit IPC class controls Constant

.068***

(.020)

.257***

.007

.040* .133*** .00002 .010 −.001 .220** −.00003*** .111***

(.024) (.028) (.0001) (.022) (.001) (.078) (.0000) (.025)

−.469***

(.088)

LR Test

X2 (9)/Prob > X2

McFaddens R2

.085

.015* .050*** .000009 .004 −.0002 .083** −.00001*** .042***

(.009) (.010) .0002 .008 .0005 .029 (.0000) (.009)

Coef.

SE

AME

SE

.124* −.085 .108* −.027 .003** −.035 −.003 .168 −.000001 −.142** −.0002* .161* 4 of 19 significant 7 of 9 significant 2 of 14 significant .449

(.074) (.109) (.056) (.062) (.001) (.034) (.002) (.210) (.00002) (.042) (.0001) (.085)

.041* −.028 .036* −.009 .001*** −.011 −.001 .055 −.000001 −.047** −.00006* .053* 4 of 19 significant 7 of 9 significant 2 of 14 significant

(.021) (.036) (.015) (.021) (.0003) (.011) (.001) (.064) (.00001) (.016) (.00003) (.027)

(.775)

X2 (54)/Prob > X2

422.48/.000

138.28/.000

.062

Robust standard errors in parentheses. * p < .1. ** p < .01. *** p < .001. Table 6 OLS regressions on patent price. Model 2a – OLS on price of all sold patents

Model 2b – OLS on price of all sold patents with Heckman selection control

Model 2c – OLS on patent price of single-patent lots with Heckman selection control

Model 2d – OLS on patent price of multi-patent lots with Heckman selection control

Observations

573

573

167

406

Dependent variable

Logarithmic price

Variable

Coef.

(SE)

Coef.

(Heckman SE) Coef.

ln forward citations (five years) ln self-citations (five years) ln backward references non-pat. lit. ln backward references patent lit. Family size Number of four-digit IPC classes Generality Number of claims Patent age Single-patent lot patent dummy Reassignments Patent portfolio size Patent application cohort controls Auction controls Four-digit IPC class controls Heckman selection control Constant

.302*** .125 .014 −.021 .004*** −.045 −.075 .009*** −.0003** .764*** −.071 −.000008 10 of 19 significant 9 of 9 significant 4 of 12 significant

(.059) (.173) (.045) (.064) (.001) (.057) (.147) (.003) (.0001) (.108) (.056) (.00002)

(.076) (.146) (.050) (.069) (.001) (.066) (.145) (.003) (.0001) (.126) (.065) (.00002)

7.586***

(.296)

.226** .192 −.053 .003 .003** −.032 −.104 .010*** −.0002* .677*** .024 −.000006 4 of 19 significant 9 of 9 significant 6 of 12 significant −1.067* 8.681***

Prob > F 2

.000 2

R (adj. R )

.760

Logarithmic price

Logarithmic price

(.466) (.645)

.000 (.736)

.763

One-sided tests for (directed) hypotheses, two-sided tests for (undirected) controls. * p < .1. ** p < .01. *** p < .001.

Logarithmic price (Heckman SE) Coef. (.203) (.981) (.142) (.165) (.029) (.262) (.466) (.008) (.0003)

.098* .273* −.038 .065 .0003 −.046 −.141 .012*** −.0002*

(.067) (.149) (.053) (.074) (.001) (.057) (.152) (.003) (.0001)

−.201 .00002 0 of 15 significant 5 of 9 significant 2 of 9 significant −.466 10.417***

(.208) (.00003)

.084 −.00005* 4 of 19 significant 9 of 9 significant 5 of 12 significant −.919* 8.125***

(.076) (.00002)

(1.753) (1.996)

.000 (.739)

(Heckman SE)

.577** −.911 .061 −.049 .012 −.019 −.094 .0009 −.00005

.429

(.414) (.494)

.000 (.217)

.859

(.838)

526

T. Fischer, J. Leidinger / Research Policy 43 (2014) 519–529

a single-patent lot are, on average, $148,142 more valuable than patents sold in a bundle. Comparing proportional increases to Bessen’s (2008) study, we observe some interesting similarities and differences. The proportional increase of claims is comparable at the 1–2% level. However, we observe that the effect of an additional citation is much higher at 9.58% compared to 5%. This could be due to the fact that our sample is focused on information technology patents and that the marginal effect of citations is higher for information technology than for other fields. An alternative explanation is that we see a higher effect since we use five-year truncated citations compared to a longer exposure time frame, while citations have decreasing returns. Self-citations that show a 3% proportional increase effect in Bessen’s study have no significant effect in ours. The patent’s family size was, unfortunately, not included in Bessen’s study. 5. Discussion We find significant empirical support for two of our three hypotheses: the number of forward citations received by a patent is positively related to its value, confirming Hypothesis 1; and the number of family members is positively related to its value, confirming Hypothesis 2.12 Interestingly, we find no support for Hypothesis 3. The number of assigned IPC classes does not have an influence on a patent’s value. Our study comes with some limitations. Most notably, we encountered data limitations that we could not completely solve. We only observed the highest bidding price in successful transactions. The amounts that bidders were willing to pay in many cases remained undocumented because the secret reserve price of the seller was not met. We were able to remedy at least parts of this problem through utilizing a sample selection model following Heckman (1979). The most severe data problem we encountered was that many patents were offered in multi-patent lots that do not allow for precise hypothesis testing on patent value indicators as we do not know why some patents were bundled while others were not. To gain some understanding of the differences in patent characteristics, we compare the two groups in Model 4 (Table 7), using a probit model. Interestingly, we find no differences regarding forward citations between patents of single- and multipatent lots. Furthermore, we find the number of IPC classes and family members are different between patents of single- and multipatent lots. As we do not find any differences between the two groups with respect to forward citations (Hypothesis 1) or claims, we are confident that these findings are robust to any influence of patent bundling. For the other value indicators that show differences between single- and multi-patent lots, we have to trust that differences between the two groups can be controlled using a single-/multi-patent lot dummy. Furthermore, many patents that have not been sold may have been traded during some point in the post-auction phase, but appear as unsold items in our database since post-auction transactions are not made public. Lastly, large shares of patents sold on Ocean Tomo are information technology patents, and so potentially are not representative of all patent sales. The Ocean Tomo data we use is the first data available to study real patent sales. Due to the lack of existing studies, it is difficult to assess whether patents that were offered on Ocean Tomo are representative of all patent sales on the market. We are aware of studies on markets for patents that rely on patent reassignment data (e.g., Serrano, 2010). However, such data comes with the drawback that it not only records patent reassignment due to the

12 This is statistically significant in Models 2a and 2b. However, we do not find empirical support for Hypothesis 2 in Models 2c and 2d that break down the data into two groups and, hence, have weaker statistical power.

sale of a patent but also due to various legal reasons. For example, Fischer and Henkel (2012), who also used patent reassignment data, report that they had to manually clean their dataset extensively to rule out the reassignment of patents in corporate structures that did not reflect real patent sales. Such reassignments are done for tax optimization in international corporate structures, among other reasons. Hence, it is not surprising that we—relying on a more direct measurement of patent sales—find a different industry mix than Serrano (2010) based on patent reassignment data.13 Of course, because Ocean Tomo provides visibility of only the tip of the iceberg of patent sales, we do not know if that tip is representative to the whole iceberg. We have no reason to believe that some industries were not aware of Ocean Tomo and did not use it to sell their patents; nor do we believe that Ocean Tomo is, in general, more attractive for selling information technology patents than other patents. We just cannot rule out this possibility. Furthermore, Ocean Tomo auctions took place between 2006 and 2009. During this time frame, smartphone wars started and patent troll activity grew substantially; both phenomena are related to information technology patents. Therefore, the data we have is only representative for the time frame in which Ocean Tomo auctions took place. This makes it difficult for us to compare our descriptive results in detail to the findings of other studies that looked at different time frames and used less selective samples. On the other hand, it is somewhat comforting that the patent values we are able to observe for the first time are roughly consistent with the patent value estimates of previous literature. For the hypothesis testing, we use Heckman models to control as best we can for any sample selection effects. However, if patents of some industries were not offered or sold on Ocean Tomo, we cannot generalize our results to them. As our dataset does not cover patents of all technology fields, we cannot claim that our results hold for all patents. However, we are confident that they hold for the technology fields for the patents we were able to study using Ocean Tomo (see Fig. 1). Hence, an important avenue for further research would be to gather more data revealed by patent auctions to build a more representative database that will allow generalizing results on all technology fields. On the other hand, we can rule out that only information technology patents on Ocean Tomo drive the results we reported. To do this, we estimated our models on data without information technology patents and observed, qualitatively, the same results.14 Despite these limitations, our paper makes a strong contribution to the ongoing research on patent value indicators and patent valuation. We show for the first time that the hypothesized relation between patent value and received forward citations and patent value and family size holds when using a direct measure for patent’s private value and not only proxies, as has previously been done. As forward citations are the dominant indicator for patent value used in current research, this admittedly is not surprising; however, it is a very important contribution nonetheless. Furthermore, the increase in average patent value by $14,224 for every citation a patent receives in its first five years emphasizes the economic relevance of this indicator. Regarding the number of distinct IPC classes as a measure of a patent’s scope as was proposed and tested by Lerner (1994) on a sample of biotech patents, our results are in line with those of

13 Serrano (2010) finds a comparable proportion of patent reassignments in the fields of chemistry, drugs, and medical on the one hand, and information technology on the other. This suggests that we under sampled the technology fields of chemistry, drugs, and medical. 14 The results were qualitatively the same with the exception of the number of claims a patent makes. This patent value indicator has the same sign but is no longer statistically significant. Of course, this model has less statistical power due to fewer observations.

T. Fischer, J. Leidinger / Research Policy 43 (2014) 519–529

527

Table 7 Marginal effects of patent value indicators. Model 3 (Model 2b with deflated variables) Observations

573

Dependent variable

Price

Variable

Coef.

(Heckman SE)

Marginal effect in $

Proportional increase (%)

Forward citations (first five years) Self-citations (first five years) Backward references non-patent literature Backward references patent literature Family size Number of four-digit IPC classes Generality Number of claims Patent age Single-patent lot patent dummy Reassignments Patent portfolio size Patent application cohort controls Auction controls Four-digit IPC class controls Heckman selection control Constant Prob > F R2 (adj. R2 )

14.224** −30.577 −.785 −1.030* .750* 19.728 16.673 1.744* −.022 148.142** −9.401 −.008 1 of 19 significant 1 of 9 significant 3 of 12 significant −110.294 41.272 .000 .245

(6.097) (30.014) (1.033) (.620) (.012) (18.221) (36.209) (.772) (.022) (52.433) (18.290) (.005)

14,224 $ −30,577 $ −785 $ −1030 $ 750 $ 19,728 $ 16,673 $ 1744 $ −22 $ 148,142 $ −9401 $ −8 $

9.58 −20.59 −.53 −.07 .50 13.28 11.22 1.17 −.01 99.73 −6.33 −.01

(86.642) (103.283) (.171)

One-sided tests for (directed) hypotheses, two-sided tests for (undirected)controls. * p < .1. ** p < .01. *** p < .001.

earlier quantitative studies that have found no significant relationship between the indicators and their value proxies (Harhoff et al., 2002). For the case of number of claims, our results are in line with Gambardella et al. (2008) that found a significant relationship of this variable to patent value. Hence, the number of claims a patent makes should indeed be considered an important patent value indicator, while the number of assigned IPC classes shows no general predictive power. Regarding the challenge of patent valuation in practice, this study shows that forward citations, family size, and the number of claims a patent makes can serve as first indicators of patent value if at least five years of citation data is available. Interestingly however, adding forward citations increases the explained variance by only 0.5% points, from 75.8% to 76.3%, in Model 2b, and 1.2% points, from 74.8% to 76%, in Model 2a. Thus, the explanatory power of forward citations is quite limited despite its strong relation to patent value and its economically significant marginal effect. This is in line with previous research that recognized a noisy relationship between forward citations and patent value. The addition of five-year forward citations improves explained variance by only 1.4% points in Gambardella et al. (2008), who use expert estimates as a proxy for patent value. Bessen (2008) reports that no citationbased value indicator explained more than six percent of variance in the patent value estimates based on renewal data. Similarly, adding family size increases explained variance by only 0.3% points from 76% to 76.3% in Model 2b; adding claims increases explained variance by only 0.6% points from 75.7% to 76.3% in Model 2b. Most interestingly for patent valuation in practice, despite the limited explanatory power of single patent value indicators, our full model explains a large share of variance in patent value. Model 2b explains 76% of the variance (adjusted R2 74%) in patents’ private value.15 In contrast, Gambardella et al. (2008) report that their

15 Even if we drop auction control dummies and the single- or multi-patent lot dummy, and focus only on patent-related data, we can still explain 58.3% of variance in Model 2b.

full model explains only 11.3% of their value measure’s variance.16 It seems that relying on less “noisy,” directly observed patent value as a dependent variable strongly improves model fit.17 While single patent value indicators still cannot explain much variance in patent’s directly observed value, the full model containing all patent value indicators as well as controls has strong explanatory power. Most likely, we simply have to accept that there is no single patent value indicator with strong explanatory power, even if we use a direct measure for patent’s private value. An interesting avenue for further research is, on the one hand, to gather more data on publicly revealed patent’s private value to increase generalizability of this type of research and, on the other hand, to think of patent value indicators that cover all relevant dimensions of patent value to further increase predictive validity. To sum up, this study makes us optimistic that models combining many variables from patent databases can reach a level of explanatory power of patent’s private values, making them useful tools in patent valuation.

Acknowledgement We thank Philipp Kröger for valuable research support. Errors and oversights are ours alone.

Appendix A. See Table 8.

16 Unfortunately, this was the only comparable study for which we were able to find a reported measure for explained variance. 17 The high model fit is also not explained by our information technology-focused sample. Taking out information technology patents even further increases explained variance.

528

T. Fischer, J. Leidinger / Research Policy 43 (2014) 519–529

Table 8 Probit model comparing patents from single-patent lots to patents from multi-patent lots. Model 4 Observations

1.784

Dependent variable

Patent in a single-patent lot

Variable

Coef.

SE

AME

SE

ln forward citations (five years) ln self-citations (five years) ln backward references to non-patent literature ln backward references to patent literature Family size Number of four-digit IPC classes Number of claims Generality Applied U.S. patents by seller Days since issue Patent cohort controls Auction controls Four-digit IPC class controls Constant

.080 −.901*** −.014 −.036 −.031* −.110* .003 .267* −.00002 −.0004** 8 of 15 significant 7 of 9 significant 7 of 11 significant −.188

(.058) (.228) (.042) (.052) (.012) (.057) (.002) (.130) (.00002) (.0001)

.017 −.196*** −.003 −.008 −.007* −.024* .001 .058* −.000005 −.0001** 8 of 15 significant 7 of 9 significant 5 of 11 significant

(.012) (.049) (.009) (.011) (.003) (.012) (.001) (.028) (.000003) (.00003)

(.227)

Robust standard errors in parentheses. * p < .1. ** p < .01. *** p < .001.

References Albert, M.B., Avery, D., Narin, F., McAllister, P., 1991. Direct validation of citation counts as indicators of industrially important patents. Research Policy 20, 251–259. Barney, J.A., 2002. A study of patent mortality rates: using statistical survival analysis to rate and value patent assets. AIPLA Quarterly Journal 30, 317–352. Bessen, J., 2008. The value of US patents by owner and patent characteristics. Research Policy 37, 932–945. Cameron, A.C., Trivedi, P.K., 2009. Microeconometrics Using Stata. Stata Press, College Station, TX. Carpenter, M., Cooper, M., Narin, F., 1980. Linkage between basic research literature and patents. Research Management 23, 30–35. Deng, Z., Lev, B., Narin, F., 1999. Science and technology as predictors of stock performance. Financial Analysts Journal 55, 20–32. European Patent Office, 2009a. INPADOC database – legal status data. http://www.epo.org/patents/patent-information/raw-data/test/product-1411.html (accessed 21.04.2011). European Patent Office, 2009b. Worldwide patent statistical database (PATSTAT). http://www.epo.org/patents/patent-information/raw-data/test/product-1424.html (accessed 21.04.2011). Fischer, T., Henkel, J., 2012. Patent trolls on markets for technology–An empirical analysis of NPEs’ patent acquisitions. Research Policy 41, 1519–1533. Fischer, T., Henkel, J., 2013. Complements and substitutes in profiting from innovation—a choice experimental approach. Research Policy 42, 326–339. Gambardella, A., Harhoff, D., Verspagen, B., 2008. The value of European patents. European Management Review 5, 69–84. Green, J.R., Scotchmer, S., 1995. On the division of profit in sequential innovation. RAND Journal of Economics 26, 20–33. Hall, B.H., Jaffe, A., Trajtenberg, M., 2005. Market value and patent citations. RAND Journal of Economics 36, 16–38. Hall, B.H., Ziedonis, R., 2001. The patent paradox revisited: an empirical study of patenting in the US semiconductor industry, 1979–1995. RAND Journal of Economics 32, 101–128. Harhoff, D., Narin, F., Scherer, F., Vopel, K., 1999. Citation frequency and the value of patented innovation. Review of Economics and Statistics 81, 511– 515. Harhoff, D., Reitzig, M., 2004. Determinants of opposition against EPO patent grants – The case of biotechnology and pharmaceuticals. International Journal of Industrial Organization 22, 443–480. Harhoff, D., Scherer, F., Vopel, K., 2002. Citations, family size, opposition and the value of patent rights. Research Policy 32, 1343–1363. Harhoff, D., Scherer, F., Vopel, K., 2003. Exploring the tail of patented invention value distributions. In: Granstrand, O. (Ed.), Economics, Law and Intellectual Property: Seeking Strategies for Research and Teaching in a Developing Field. Kluwer Academic Publisher, Dordrecht, The Netherlands, pp. 279– 310. Harhoff, D., Wagner, S., 2009. The duration of patent examination at the European Patent Office. Management Science 55, 1969–1984. Heckman, J., 1979. Sample selection as a specification error. Econometrica 47, 153–162.

Hill, R.C., Adkins, L.C., Bender, K.A., 2003. Test statistics and critical values in selectivity models. Advances in Econometrics 17, 75–105. Lanjouw, J., 1998. Patent protection in the shadow of infringement: simulation estimations of patent value. Review of Economic Studies 65, 671–710. Lanjouw, J., Pakes, A., Putnam, J., 1998. How to count patents and value intellectual property: the uses of patent renewal and application data. Journal of Industrial Economics 46, 405–433. Lanjouw, J., Schankerman, M., 1999. The quality of ideas: measuring innovation with multiple indicators. NBER Working Paper No. 7345. National Bureau of Economic Research, Cambridge, MA. Lanjouw, J., Schankerman, M., 2001. Characteristics of patent litigation: a window on competition. RAND Journal of Economics 32, 129–151. Lerner, J., 1994. The importance of patent scope: an empirical analysis. RAND Journal of Economics 25, 319–333. Martinez, C., 2010. Insights into different types of patent families. OECD STI Working paper No. 2010/2. Organisation For Economic Co-Operation and Development, Paris. Merges, R.P., Nelson, R.R., 1990. On the complex economics of patent scope. Columbia Law Review 90, 839–916. Meyer, M., 2000. Does science push technology? Patents citing scientific literature. Research Policy 29, 409–434. Narin, F., Kimberly, H.S., Olivastro, D., 1997. The increasing linkage between U.S. technology and public science. Research Policy 26, 317–330. Narin, F., Noma, E., 1985. Is technology becoming science? Scientometrics 7, 369–381. Narin, F., Noma, E., Perry, R., 1987. Patents as indicators of corporate technological strength. Research Policy 16, 143–155. Nordhaus, W.D., 1967. The Optimal Life of a Patent. Cowles Foundation for Research in Economics at Yale University, New Haven, CT. OECD, 1994. The Measurement of Scientific and Technological Activities Using Patent Data as Science and Technology Indicators: Patent Manual 1994. Organisation For Economic Co-Operation and Development, Paris. Putnam, J., 1996. The value of international patent rights. Yale University, New Haven, CT, Ph.D. Thesis. Reitzig, M., 2003. What determines patent value? Insights from the semiconductor industry. Research Policy 32, 13–26. Reitzig, M., Henkel, J., Heath, C.H., 2007. On sharks, trolls, and their patent prey – Unrealistic damage awards and firms’ strategies of ‘being infringed. Research Policy 36, 134–154. Schankerman, M., 1998. How valuable is patent protection? Estimates by technology field. RAND Journal of Economics 29, 77–107. Schankerman, M., Pakes, A., 1986. Estimates of the value of patent rights in European countries during the post-1950 period. Economic Journal 96, 1052–1076. Serrano, C.J., 2005. The market for intellectual property: evidence from the transfer of patents. Job market working paper. University of Minnesota and Federal Reserve Bank of Minneapolis. Serrano, C.J., 2010. The dynamics of the transfer and renewal of patents. RAND Journal of Economics 41, 686–708. Tietze, F., 2008. Technology market intermediaries to facilitate external technology exploitation. The case of IP auctions. Working Paper 55. Hamburg University. Tietze, F., 2012. Technology market transactions: auctions, intermediaries and innovation. Edward Elgar, Cheltenham, UK.

T. Fischer, J. Leidinger / Research Policy 43 (2014) 519–529 Trajtenberg, M., 1990. A penny for your quotes: patent citations and the value of innovations. RAND Journal of Economics 21, 172–187. Trajtenberg, M., Henderson, R., Jaffe, A., 1997. University versus corporate patents: a window on the basicness of invention. Economics of Innovation and New Technology 5, 19–50.

529

van Zeebroeck, N., van Pottelsberghe de la Potterie, B., Guellec, D., 2009. Claiming more: the increased voluminosity of patent applications and its determinants. Research Policy 38, 1006–1020. Ziedonis, R., 2004. Don’t fence me in: fragmented markets for technology and the patent acquisition strategies of firms. Management Science 50, 804–820.