Advertising , the Matchmaker ¤

This study models advertising content as a noisy signal on product attributes. Contrary to previous empirical studies that modeled advertising only as part of the consumer’s utility function, we formulate advertising also as an element in her information set. This approach yields the following implications. First, in some cases, exposure to advertising decreases the consumer’s tendency to purchase the promoted product. Second, exposure to advertising improves the matching of consumers and products. These implications enable the researcher to distinguish between the e¤ect of advertising on utility and its e¤ect through the information set using individual-level data on both consumption and advertising exposures. Using a dataset that was designed and created to test this model and its implications, we show that the theory is supported empirically. The structural estimates imply that an exposure to a single advertisement decreases the consumer’s probability of not choosing her best alternative by at least 16%. ¤We are deeply grateful to Dmitri Byzalov for excellent research assistance, to Steve Berry, Dennis Carlton, Zvi Eckstein, Michael Keane, Paul Klemperer, Aviv Nevo, Ariel Pakes, Peter Reiss, John Rust, Manuel Trajtenberg, Bruce Weinberg, Ken Wolpin, and to participants in seminars at Econometrics in Tel-Aviv, Haifa University, Harvard University, Hebrew University, MIT, NBER IO Meetings, NBER Productivity Seminar, NYU, Oxford University, University of Pennsylvania, Rochester University, Stanford University, Tel-Aviv University, and Yale University, for comments, and to several colleagues for helpful discussions. Anand is grateful to the Division of Research at Harvard Business School for ...nancial support. Shachar is grateful to the Israel Institute of Business Research at Tel Aviv University Faculty of Management for ...nancial support. ySoldiers Field Road, Boston, MA 02163. Phone: (617) 495-5082; Fax: (617) 495-0355; email: banand@hbs.edu. zEitan Berglas School of Economics, Tel Aviv University, Tel Aviv, Israel, 69978. Phone: 972-3-640-6311; Fax: 972-3-640-5621; email: rroonn@post.tau.ac.il.


Introduction
The idea that the development of science has driven the rapid economic growth of the Western world enjoys wide acceptance today. Though identifying the precise point at which this notion entered the marketplace of ideas proves difficult, the fact that both de Toqueville (1848) and Marx (1844) pointed to science as a key progenitor of the technological progress that stimulates economic growth suggests that this claim has held sway in intellectual circles for at least 150 years. 1 The widespread belief in this relationship generates important social consequences. Early sociologists noted that this claim to material usefulness affords science its status in modern society (Merton, 1938(Merton, , 1942Weber, 1946). Economists have drawn on this widely accepted relationship to justify the public funding of scientific research (e.g., Bush, 1945;Kuznets, 1959;Malakoff, 2000). And international organizations, such as UNESCO and the World Bank, use this logic to support the promotion of these expenditures worldwide (Schofer, Ramirez and Meyer, 2000).
In the attempt to validate this claim, researchers usually examine one of three types of evidence. 2 The earliest research on this topic analyzed the impact of public and private research and development expenditures on productivity increases in the United States, showing that these expenditures do indeed appear to stimulate growth (Mansfield 1972;Rosenberg, 1976;Schmitz, 1989;Adams 1990). Nevertheless, cross-national studies fail to confirm this effect, often finding a negative relationship between indicators of scientific research and economic growth (Williams, 1964;Shenhav and Kamens, 1991;Schofer, Ramirez and Meyer, 2000). A more recent strain of research -arguing that scientific practices can stimulate productivity even at the firm level -demonstrates that companies that adopt a scientific orientation outperform those that do not (Henderson and Cockburn, 1994;Gambardella, 1995;Zucker and Darby, 1996;Zucker, Darby and 1 The precise dating eludes us, but an examination of the classics of political economy points to the early 19 th century. Although the idea appears absent from the works of Adam Smith (1776) and Ricardo (1821), one can find it in de Toqueville, Marx and the later work of Alfred Marshall (1920). Gieryn (1983) also mentions it as an element of a 1866 discourse involving John Tyndall. 2 Other approaches offered, though not followed, include Sveikauskas' (1981) correlation of scientific employment to economic growth and Jaffe's (1989) investigation of the relationship between university research expenditures and local patenting rates. Brewer, 1998). 3 However, Stern (1999) points out and provides strong evidence that firms adopting a scientific orientation might accrue these benefits from compensating differentials in wages (i.e. talented workers accept lower wages in exchange for the ability to publish their research), rather than the productivity enhancements from science.
The most direct line of investigation shows that patents that originate out of university research receive more citations in the future, a measure of patent importance, suggesting that university research yields more important inventions (Jaffe and Trajtenberg, 1996;Henderson, Jaffe, and Trajtenberg 1998).
These accounts typically assume -whether implicitly or explicitly -that science benefits society because its methods offer a superior process for generating high-quality knowledge that can enable new technologies. 4 Nevertheless, this highly favorable view of the scientific method conflicts with a substantial literature in the sociology of science that points to the difficulty the scientific community faces in defining its own boundaries (e.g., Collins, 1982;Gieryn, 1983). It also contrasts markedly with descriptions of the highly contentious and often subjective debates that frequently surround the interpretation of new evidence. 5 An alternative, perhaps somewhat more sociological, explanation exists. In addition to its methods, science also includes a set of social conventions (Merton, 1942). One in particular strikes us as important: the norm of publication (Bernal, 1939). Absent publication, much of the knowledge gained from scientific research must pass through interpersonal networks. To the extent that these social networks localize in both physical and social space, transmission will most frequently occur to those in close proximity to the inventors. Though these interpersonal networks allow for the diffusion of knowledge through space, publication elevates the diffusion process beyond the limited contacts available in social networks, thereby accelerating the 3 Interestingly, all of these studies investigate firms within a single industry, biotechnology. Thus, the more general impact of commercial enterprises adopting scientific practices remains unknown. 4 Some researchers, most notably Henderson, Jaffe and Trajtenberg (1998), suggest that universities generate a different type of patent -one with application across a broader range of technologies. We believe that diffusion might also drive these results and address this issue later in the paper. 5 A host of examples exist. Linus Pauling's contentious claim that vitamin C can control cancer and the ensuing debate (Richards, 1988). Martin (1989) reviews the controversy surrounding the fluoridation of water. McCrea and Markle (1984) examine the debate over estrogen replacement therapy. In addition to biology and medicine, a host of researchers focus on debates within the world of physics (e.g., Collins, 1975Collins, , 1998Pickering, 1984;Traweek, 1988). diffusion of knowledge. This diffusion might prove functional in the sense that it reduces the duplication of research effort, but it suggests that the norm of publication rather than the scientific method of developing knowledge provides this benefit.
To gain purchase on this issue, this study compares the citation patterns of inventions that draw on the scientific literature to those that do not. We attempt to identify the importance of communication in two ways: First, we compare the future patent citation rates across three groups of patents. The first group consists of patents that cite the scientific literature, while the second includes patents that cite published sources that do not meet the standards of scientific research (e.g., marketing reports and popular journals). The third group comprises all patents that do not cite published sources. If science generates higher quality knowledge, then the first group of patents should receive more citations than the other two groups. On the other hand, if communication accelerates the diffusion of information, then both groups of patents that cite published sources should receive more citations than the group that does not. Second, we compare the spatial diffusion of future patents that cite these various published sources to those that do not reference such literature. If publication expands the rate at which information diffuses through space, then patents that reference published sources should experience an expansion in the geographic range of the patents that cite them.
The results suggest that communication plays a large role in explaining the citation premium attributed to patents based on scientific research. Patents derived from sciencebased research do not show a significantly higher likelihood of being cited by all future patents. Nevertheless, patents that reference published materials -whether scholarly or not -receive citations at a much faster rate from distant inventors, suggesting that publication accelerates the flow of information through space. Thus, the citation premium accorded to science-based research appears to stem almost entirely from the fact that the faster dissemination of knowledge increases the number of inventors aware of these inventions. Though this rapid diffusion of information might benefit society by reducing the degree of inefficient duplicated effort, the scientific method, per se, does not appear to lead to more important inventions.
In addition to the theoretical issues, this question also addresses important policy issues.
For example, if much of the benefit of science accrues from its norms of openness, then public funding of research should include stipulations of quick and public disclosure of results to take advantage of the potential societal benefits. Moreover, if the technological benefits of science to society derive mainly from communication, then faster review and publication might increase the pace of technological change. For example, faster publishing and Internet-based dissemination of information may increase the rate of invention. We address these and other policy issues in detail in the discussion.

The Role of Science in Technological Advance
In considering the role that science plays in technological advance, one must first distinguish how science differs from non-science. Two types of criteria separate scientific and non-scientific activity in the literature: the method of knowledge generation and the professional norms. Philosophy tends to focus on the logic of the scientific method and how that method may engender the efficient production of knowledge.
Meanwhile, sociological accounts typically focus on the norms of the scientific community. Let us review each of these accounts. 6

Science as a Method
Much of the work in the philosophy of science focuses on the logic of theory development in scientific research. By focusing on these methods, science seeks both to differentiate itself from other pursuits, such as metaphysics, and to legitimate its activities. Thus, Compte (1853) distinguishes science from other activities by its reliance on observation and logic to generate theory. Nevertheless, this description fails to separate science from other forms of experiential learning. A more precise modern statement of the value of the scientific method comes from Popper (1965) and the 6 Though both portray scientists as belonging to a nearly homogenous population, scientists must admittedly contend with overlapping and conflicting roles in multiple communities (Moore, 1996). Likewise, the factors that demarcate science can vary across disciplines and often appear in the practices of 'non-scientists' (Gieryn, 1983). Nevertheless, while scientists likely differ in interesting ways from one another, we focus on those factors that typically differentiate those labeled as scientists from non-scientists. positivist school, who claim that the generation of falsifiable statements forms a more logically sound basis for building knowledge. Popper (1965) argues that science progresses by systematically testing the empirical implications of theories. Though scientists cannot test the theories themselves, they can deduce empirical observations implied by those theories. One sees this logic throughout work in the sciences. 7 If the empirical evidence, either from experiments or observation, fails to match the expectations of the theory, then the theory does not hold. Scientists must search for a new theory to explain both the new findings and those patterns that the previous theory explained.
Though more recent research raises questions regarding the validity of the positivist model of science, even the modifications to this theory typically argue that science seeks a close match to empirical observation. Thus, Kuhn (1958) notes that theories progress in two stages: normal science and scientific revolution. During normal science, research progresses along the path described in the positivist model. Through the course of this research, anomalies that fail to fit the dominant theory inevitably arise. Instead of simply declaring the theory wrong, scientists instead attempt to extend the theory incrementally until these theory fixes become unwieldy. At that point, the field reaches a stage of crisis and seeks out a new dominant paradigm. Nevertheless, even in Kuhn's account, science offers a superior means of generating knowledge.
The norms of science enforce this methodology. Students being trained in the sciences receive extensive indoctrination into the importance of using the 'scientific method'. As students, they take entire courses devoted to the proper methods for developing and testing theories. Following those courses, most programs place them in apprentice-like positions through research assistantships and post-doctoral appointments in which faculty can closely monitor the activities of these pupils and ensure that they conform to accepted practices. Even after leaving the training stage of their careers, the peer review process affords the community a mechanism for continuing to enforce a set of research standards. 7 The logic also dominates the social sciences. This very paper proceeds with the same type of reasoning.
To the extent that these methods support the more effective accretion of knowledge, one might expect the adoption of scientific methods to stimulate technological development.
If one views the process of invention as a search of a large and multi-dimensional technological space, then the development of theories provides an efficient mechanism for reducing the proportion of the space that one must sample (Fleming and Sorenson, 2000). When sampling technological space proves costly, either in terms of time or money, techniques that reduce the need for direct observation will yield economic benefits. Thus, one might expect science to offer a more efficient search process if its methods offer a superior mechanism for developing knowledge.
Work in the sociology of science, however, raises questions regarding the extent to which science actually does rely on accurate descriptions of reality as an objective function.
The debates over the interpretation of data in biology (McCrea and Markle, 1984;Martin, 1988) and physics (Collins, 1975(Collins, , 1998Pickering, 1984) suggest that social and political influences play a strong role in shaping the interpretation of experimental data. This methodologically relative view of science also resonates with a wide range of sociological studies that highlight the role of social processes in science (e.g., Merton, 1968). Thus, we also consider a less instrumental perspective on the role of science.

Science as an Institution
Robert K. Merton encouraged sociologists to consider science as an institution. Through their training and both positive and negative career incentives, scientists internalize a set of values that guide their activities. Though Merton (1942) identified several central norms in the scientific community -including universalism, communism, disinterestedness and organized skepticism -'communism' seems most important to the question of knowledge diffusion.
The norm of 'communism' strongly influences the practice of scientific research.
'Communism' refers to the idea that individual scientists believe that their property rights over their ideas extend only to the credit associated with finding it first (Merton, 1942).
Other scientists may freely use the ideas generated by their predecessors as long as they give credit to the originator of the idea. At first glance, this norm might appear to reduce the incentives for innovation. Nevertheless, scientists receive powerful, though somewhat indirect, incentives to innovate because rewards in the field invariably go to those that discover things first (Merton, 1957). These rewards take the form of recognition -through citations, prizes and the naming of species, theories and elementsand resources through research grants, endowed chairs, university-funded laboratories and graduate students.
Together, the norm of communism and the incentives surrounding primacy create an intense desire to publish research (Merton, 1942). Publication provides the means for conforming to the norm of communism, making one's ideas available to the scientific community. To allow these journals to transmit information effectively, scientists have developed highly specialized vocabularies and grammars to codify complex information.
Thus, journals can broadcast information that once may have required interpersonal communication for effective transmission. In addition to relaying new ideas to peers, publicly disseminating research establishes and allows one to defend a claim to primacy. This dual role thus reinforces the importance of publication.
Publishing might also, functionally, promote the accretion of knowledge. Bernal (1939) notes that the growth of science coincided with the rejection of the idea of secrecy. The rapid diffusion of knowledge through publication and other media allows researchers to avoid the duplication of effort and advance their research beyond the already known.
Scientists themselves implicitly recognize the importance of this cumulative research effort, as one observes in statements such as Newton's "If I have seen farther, it is standing on the shoulders of giants." 8 In addition to published journals, the scientific community has also generated a wide range of organizations that facilitate the flow of communication: conferences, departments, academies, etc. These associations influence the interaction patterns among researchers because they form conduits that shape the daily activities of researchers. For example, sociologists meet every year at a central location to encourage interaction with one another. In the physical and biological sciences, scholars tend to focus on meetings that congregate researchers with much narrower ranges of interests. By forming networks that bridge geographic space, these organizations expand the range of information diffusion (Sorenson and Stuart, 2001).

Empirical Strategy
To understand better the effect of science on technological development, this study focuses on the differences between patents that cite non-patent materials and those that do not. Patents provide a window on the generation and diffusion of knowledge through their references to other materials. In particular, patents reference two types of 'prior art'. First, they list those prior patents on which the new invention builds. Second, patents often reference other materials that led the inventor to come up with their new ideas. Similar to studies of citation patterns across academic papers, these references offer a paper trail for tracking the diffusion of information.
Patent citations offer a more objective measure of linkages, however, than citation patterns in publications. Unlike authors, inventors have an incentive to minimize their references to prior art, though the law requires them to reveal any directly relevant existing work (Carr, 1995). Patents represent a property right over the commercial developments based on some parcel of intellectual property. As such, patent applicants proceed with strong incentives to avoid actions that could delimit the range of their claim.
Both citations to other patents and citations to prior publications can potentially reduce the scope of their claims and limit the effectiveness of any future legal action defending it. Thus, patent applicants will seek to minimize these citations.
The patent review process places a check, however, on the failure to cite prior art. Based upon personal expertise and automated searches, patents examiners add missing citations to applications in the review process (Carr, 1995). Nevertheless, references to non-patent materials more likely come from the patent applicants themselves (Tijssen, 2001).
Although patent examiners often add citations to prior patents, their mandate does not extend their search to the vast array of published materials on which one could potentially draw. Therefore, inventors almost certainly have awareness of the non-patent references that appear on their patents.
Patents cite a wide range of sources besides previous patents; most notably, they make reference to various scientific, technical, and corporate literatures. Although other researchers have investigated the relationship between patents and non-patent references, they have typically explored small samples of references to the scientific literature using case study methods (e.g., Tijssen, 2001). Little work investigates the effects of scientific citations on large samples and none that we know of examines the importance of references to non-scientific sources.
Thus, we began our study by selecting two months of patents: May and June of 1990 (17,264 patents). Most of our data comes from the MicroPatent database, which comprises information from the U.S. Patent and Trademark Office (USPTO). To enable an analysis of the non-patent citations, a trained researcher categorized each of the 16,728 sources referenced on our sample of 17,264 patents. 9 She sorted the references into seven mutually exclusive and exhaustive categories (see Table 1 for the distribution): Scientific Index Journal: These publications appear in the Scientific Index. The journals in this category include both the familiar high prestige journals, such as Science and Physica, and a multitude of more obscure or non-English references -examples in the 9 One individual coded the entire sample. She began by comparing every reference to journals listed in the Scientific Index. She then proceeded based on the descriptions below. However, to assess the clarity and reliability of this classification schema, a second coder independently assigned a random sample of 100 references to these categories. The second coder agreed with the first on 96 of these cases. Given the consistency across independent coders, we feel confident that the classification scheme identifies distinct and cognitively meaningful categories. sample include Chermetinformatsia and Cryogenic Engineering. Although the quality of the journals indexed here undoubtedly varies, inclusion in the index denotes some level of acceptance in the scientific community. Therefore, we focus on the patents citing these journals as representing the fruits of scientific research.
Conference Proceedings: Though most of these conference proceedings refer to meetings for the presentation of scientific research, the standards at these meetings may not meet those necessary to receive publication in a peer-reviewed journal. One example of this type of citation appears in patent 4922432, which cites a paper entitled "The CMU Design Automation System -An Example of Automated Data Path Design," in the Proceedings of the 16 th Design Automation Conference. Automated data path design appears many years earlier in an engineering textbook (Mead and Conway, 1980); hence, the paper might fail the peer review process of a journal.
Technical Report: Most of the items in this category refer to research institute publications; for example, the Battelle Institute, "Final Report on High-Performance Fibers II, An International Evaluation to Group Member Companies" by Donald C.
Slivka et al., published in 1987, appears on one patent. Though many of these reports may describe the results of scientific research, the classification of them remains problematic since institutions self-publish these reports, thereby avoiding peer review.
Corporate Publication -Technical: Corporate technical publications typically refer to product specifications or schematics. Technical Bulletin BH183 (Series) by Howell Instruments Inc. of Fort Worth, TX, U.S.A., "3″-Dia. Digital Indicators", provides an example of one that appears in our data.
Book: Although the book once represented the primary mechanism for disseminating scientific knowledge, academics -especially those in the physical and biological sciences -have increasingly turned to journal articles as the preferred outlet for publishing research (Bazerman, 1988 To discriminate between the effects of scientific methods and publication, we focus on comparing citation patterns across patents that reference these various sources. Patents that cite journals in the Scientific Index most clearly draw from scientific research. Thus, we concentrate on these patents as representing science-based inventions. 10 Because these patents confound scientific methods and publication, we compare the distribution of their citations to those of patents that reference publications that clearly do not arise from scientific research. Of the categories described above, non-index journals and corporate non-technical publications most clearly meet this criterion. Therefore, our analysis will focus on patents that reference those types of publications. We examine the effects of these references on two dependent variables: the number of citations received and the spatial distribution of those citations.

Citation Rates
The most basic evaluation of the impact of science involves looking simply at the citation rates of patents that draw on science relative to those that do not. A substantial body of research based on patents uses the number of citations a focal patent receives from future patents as an indicator of the focal patent's economic and technical importance (e.g., Jaffe, Trajtenberg and Henderson, 1993). Though this measure undoubtedly offers a noisy proxy for these outcomes, it does allow the comparison of patents across classes and several studies have confirmed this general relationship (Trajtenberg, 1990;Albert, et al., 1991;Harhoff, et al. 1999;Hall, Jaffe and Trajtenberg, 2000). Transporting this measure to our question leads to the following expectation: If science offers a better method for developing knowledge, then patents based on its methods should receive more citations than those that do not employ the scientific method.
We begin by looking at simple mean citation rates. For each patent in the sample, we track all future patents through November of 1996 to determine how many citations it receives in its first six years. Two variables allow us to determine whether the patent draws on scientific research methods. If either the assignee (patent holder) is a university or the patent references a journal in the Scientific Index, then its development likely followed the methods of scientific research. Table 1 shows the average number of citations received depending on the type of materials referenced by the patent. As in earlier research (Jaffe and Trajtenberg, 1996;Henderson, Jaffe, and Trajtenberg 1998), universities receive more citations to their patents than other assignees (4.99 citations versus 3.80 for the entire sample), however, they account for only a small number of the patents citing the scientific literature (university patents also do not uniformly cite scientific articles; see Table 2). Whether from a university or not, all patents referencing the scientific literature receive more citations than patents that do not cite articles in the Scientific Index. On average, these patents receive 4.79 citations in the six years following their granting versus the 3.54 citations received by patents that do not reference published sources, a premium of 35%.

INSERT TABLES 1 & 2 ABOUT HERE
Though patents citing scientific articles receive more cites, every other form of publication also appears to accord the patents that reference it an increase in future citations. Although some of these categories, such as conference proceedings also confound publication with scientific method, at least two types of publications -the corporate non-technical and non-index journals -make no claim to scientific methods.
Patents that reference non-technical corporate publications receive 4.61 citations on average, a 30% premium, while those that refer to non-index journals receive 5.4 citations -53% more than patents that do not cite any publications and 13% more than patents citing journals in the scientific index.
Although these results suggest that publication, rather than the application of scientific methods, raises future citation rates, these averages do not account for other factors that might vary across patents that arise from science-based research versus those that do not.
For example, technological fields vary in their degree of activity. Consequently, the number of potentially citing inventors differs from one field to the next. To control for these differences, we modeled the number of citations each of our 17,264 patents received using negative binomial regression. 11 Three important control variables enter these regressions. First, an activity control accounts for differences in citation patterns across patent classes. The USPTO classifies patents into one or more of roughly 400 technological classes. If all patents fell into a single class, we could simply use fixed effects for each class; however, most patents (92%) fall into more than one class. Therefore, following fixed effects logic, we estimate the number of citations a patent will likely receive based upon its class memberships. 12 Second, the models include a control for the number of classes in which the patent falls.
Patents that fall in multiple classes may apply to a broader range of technologies, thereby expanding the number of future patents that might cite them. Finally, technologies also vary in their closeness to the technological frontier. Recent technical area -the average of the patent numbers of the focal patent's prior art (higher numbers indicate more recent 11 Because counts, such as the number of future citations, cannot fall below zero, linear regression can yield inefficient and biased coefficient estimates. Although researchers often use Poisson regression to model count data, our sample violates the assumption of a one-to-one mean/variance ratio, making the negative binomial a more appropriate specification (Cameron and Trivedi, 1986). 12 This amounts to a weighted mixture of the expected number of citations for a particular class based on citation activity from 1985 to 1990. See Fleming (2001) for a detailed explanation. technology) -controls for this difference across technologies. 13 Table 3 describes these controls and the other variables used in the count models.

INSERT TABLES 3 ABOUT HERE
Tables 4 and 5 present the results of these regressions. Model 1, which estimates a model with only the dummy variable for university assignees, shows that these patents receive 32% more citations on average. 14 After including the number of articles in the Scientific Index referenced by the patent in model 2, however, university patents only exhibit an 18% premium over non-university patents that draw on the scientific literature. Each reference to an Index journal increases the number of citations a patent receives by about 3%. As models 3, 4, 5 and 7 show, differences across technical areas account for part of this effect. In model 7, which includes fixed effects for each class, university patents receive 11% more citations (statistically insignificant), whereas each reference to a scientific publication increases the expected number of citations by 1%. This decomposition of effects suggests that most of the university effect comes from references to the scientific literature and the technological fields of research.

INSERT TABLES 4 & 5 ABOUT HERE
The critical results with respect to our question appear in table 5. Model 6 includes counts of both the number of articles referenced in journals that do not appear in the Scientific Index and the number of non-technical corporate publications cited -the two types of publications that most clearly do not meet the standards of scientific research.
Not only do both of these publications increase the number of citations that a patent receives, but also they have a larger effect on the future citation count than publications in scientific journals. Each reference to a non-index journal article increases the expected number of citations by 12% and each reference to a non-technical corporate publication predicts an 8% boost in citations. 15 Model 7 demonstrates that these effects remain robust after the inclusion of fixed effects for technological class. 16 In model 8, the results also hold when specifying the publication measures as dummy variables, rather than counts of the number of references. Nonetheless, this specification does show that the difference between the magnitudes of the effects of the different types of publications comes primarily from variation in the average count of references cited of each type (in model 8, referencing the Scientific Index increases expected citations by 10.7%, while referring to non-index journals and corporate non-technical publications boosts counts by 19.5% and 16.2% respectively). Finally, model 9 re-estimates the model including only future patents that do not share an assignee with the focal patent (i.e. excluding selfcitations) in the dependent variable. Removing self-citations changes the results minimally. However, the university dummy variable becomes significant once again, suggesting that universities engage in less self-citation than other patent holders.
These results raise doubt that the methods used in scientific research can account for the larger number of citations these patents receive. If those methods did explain the citation premium, then patents drawing on scientific methods should receive more future citations than both patents that do not cite published sources and those that reference non-scientific publications. Nevertheless, patents that reference non-scientific publications, such as corporate advertisements and popular magazines, appear to enjoy citation counts that equal or exceed those received by patents based on scientific research. Though this finding seems more consistent with the communication explanation of the role of science, it does not provide direct evidence of the diffusion process. Therefore, we proceed by analyzing the spatial distribution of citation linkages. 15 Comparing model 7 to one that constrains the coefficients of the different types of publication to equality indicates that these coefficients differ with a probability of more than 99% (using Haberman's chi-squared test). In models 8 and 9, however, the likelihood that the magnitudes of the dummy coefficients differ falls to less than 80% based on the same test. 16 A Hausman (1978) test confirmed the preference of the fixed effects model over a random effects specification (χ² = 128.8, 7 degrees of freedom).

Citation Distribution
Citations tend to localize in space because the networks that tie inventors together connect most densely when these actors lie in close geographic proximity to one another.
A variety of studies, beginning with Park (1926) demonstrate that social actors more frequently form ties when they reside near to one another. Bossard (1932) found that the likelihood of marriage between two individuals decreased rapidly as the distance between them lengthened, a finding confirmed in both marital and friendship ties in a series of subsequent investigations. Studies of the workplace indicate that the frequency of interaction between co-workers depends on the proximity of their offices to one another (Allen, 1977). More recently, Sorenson and Stuart (2001) demonstrate that spatial proximity plays a strong role in determining which entrepreneurs venture capitalists decide to fund. The importance of propinquity arises from two factors: First, the likelihood that any two individuals will meet in the course of their day-to-day activitiesthereby having the opportunity to form a tie -declines rapidly with distance (Hawley, 1971). Second, even when a tie does occur, the costs of maintaining that tie likely increase, as actors must bridge increasingly wide expanses to interact (Zipf, 1949). Thus, individuals budgeting scarce resources will discontinue these distant ties more readily.
By removing the flow of information from the constraints of social networks, publication should accelerate the diffusion of ideas in space. When private information travels through interpersonal ties, it cannot escape the spatial limitations of the network. Hence, studies repeatedly show circumscribed transmission of tacit or confidential data. For example, Hedstrom (1994) finds that the geographic configuration of communication networks in the Swedish population can explain the spatial contagion in these organizations' founding. Similarly, Baker (1984) shows in his study of a commoditiesexchange trading floor that the structure of interaction influences the volatility of securities prices. Nevertheless, when information becomes publicly available, its transmission should transcend these network limitations, flowing much more freely to loosely-or un-connected individuals. If publication in scientific journals accelerates the spatial diffusion of information, these patents should receive citations from more geographically distant future patents.
To investigate the diffusion of ties, we model the probability that a future patent cites a given focal patent as a function of distance, publication and a variety of control variables; essentially, we estimate the likelihood of tie formation. Although many studies proceed by analyzing every possible dyad and using logistic regression to estimate the effects of a covariate vector on the likelihood of a tie (e.g., Podolny, 1994;Gulati, 1995), this strategy creates two problems. Methodologically, it fails to account correctly for nonindependence across cases, as each actor enters the analysis many, many times. 17 The large number of repeat occurrences of each actor can lead to systematic underestimation of the standard errors for actor attributes that do not change from dyad to dyad.
Pragmatically, this strategy presents a second obstacle; the observation of all possible dyads can prove burdensome computationally, especially for large networks. For example, consideration of all potential dyads in our sample would require us to create a matrix with more than eleven billion cells. Sampling randomly from the set of potential patent dyads offers one possible solution to these issues. Nevertheless, this approach falls short of the ideal because it ignores the fact that the realized ties provide most of the information for the estimation of the factors that affect tie likelihood (Coslett, 1981;Imbens, 1992;Lancaster and Imbens, 1996).
To address these issues, our study uses a matched sample design. Thus, we include all 70,271 citations that actually appear in our sample of 17,264 patents. For comparison, we then create a matched sample of potential cites that did not occur, pairing each of the 17,264 patents in the sample with four patents 18 chosen randomly from the patents assigned between July 1990 and June 1996. Though this generated a data set of 139,487 dyads, our analyses restricted the estimated sample to the 76,807 cases where the focal inventor resides in the U.S. To address the fact that focal patents enter the data more than once, we report robust standard errors estimated without the assumption of independence across observations on the same patent.
The use of a matched sample introduces one new problem. Logistic regression can yield biased estimates when the proportion of positive outcomes (ties) in the sample does not match the proportion of ties in the population. In particular, uncorrected logistic regression using a matched sample tends to produce underestimates of the factors that predict a positive outcome (King and Zeng, 2001). Large samples do not necessarily alleviate this problem. Following Sorenson and Stuart (2001), we adjust the coefficient estimates using the method proposed by King and Zeng (2001) for the logistic regression of rare events to correct for this potential bias. 19 In addition to the variables previously discussed, the dispersion analysis requires a measure of distance. All patents list the address of the inventor on the front page of the patent application.
To locate inventors, we match the inventor's 3-digit zip code 20 to the latitude and longitude of the center of the area in which the patent assignees reside based on information from the U.S. Postal Service. Using spherical geometry, the distance between the two points, i and j, is: where latitude (lat) and longitude (long) are measured in radians. C converts the result into linear units of measure, C = 3437 corresponds to miles. In the models, a logged term accounts for the fact that the relevance of a mile declines with distance. We assign each 19 The traditional logistic regression model considers the dichotomous outcome variable to be a Bernoulli probability function that takes a value 1 with the probability π: where X is a vector of covariates and β is a vector of parameters. Researchers typically use maximum likelihood methods to estimate β. King and Zeng (2001) prove that the following weighted least squares expression estimates the bias in β generated by oversampling rare events: , and w 1 represents the fraction of ones (events) in the sample relative to the fraction in the population. Essentially, one regresses the independent variables on the residuals using W as the weighting factor. Tomz (1999) implements this method in the relogit Stata procedure. 20 The USPTO includes 5-digit zip information, but we chose to reduce the measurement error by using cleaned data. CHI, an information provider, has called every patent holder to verify the inventor's location; however, the only record this information at the 3-digit level. dyad including one foreign patent a distance of 2500 miles separating the two patents and include a dummy variable, foreign assignee, to capture any systematic bias in this assignment. 21 Beyond the variables used in the negative binomial models, the tie estimation includes several additional controls. Same zip indicates that the inventors of both patents reside in the same 3-digit zip code. Same class denotes dyads where both patents belong to the same primary class. Same assignee essentially indicates self-citations. Foreign assignee takes a value of one if the potentially citing inventor resides outside the US, or zero otherwise. Time (grant to cite) measures the difference in patent numbers between the two patents, a measure of the time that has elapsed between them. Finally, both university indicates that both patents belong to universities.
The results of these regressions appear in table 6. If science offers a superior method for generating knowledge, then patents that draw on these methods should enjoy higher citation rates from all future patents. On the other hand, if science increases the velocity of information diffusion through its norms of openness, then publication should expand the spatial range of citing patents. The first model simply shows that both patents from universities and those that reference articles from journals in the Scientific Index have a higher probability of receiving a citation from any future patent. Model 11 tests the dispersion of these cites across both geographic and technological space by interacting Scientific Index with both the distance between the two patents and same class. Including terms to account for the distance, both in geographic and social space (in terms of class and assignee), dramatically improves the model. All of the terms show that proximity increases the likelihood that a future patent cites the focal patent. As one would expect if publication diffuses knowledge more rapidly, references to the scientific literature reduce this localization. In model 12, a number of additional controls account for other factors that might differ across patents citing science and those that do not. The inclusion of 21 Dropping the foreign patents does not change the qualitative interpretation of the interaction terms. However, it does produce a negative bias in the main effects of variables that increase the dispersion of citations. To avoid this bias, we reported models including foreign patents. One could also specify distance by its inverse so that zero becomes a meaningful value to assign to patents at a large but unknown distance. Doing so generates an identical qualitative pattern of results.
these variables weakens the interaction between Scientific Index and technological class; however, the other results of theoretical interest remain unchanged.

INSERT TABLE 6 ABOUT HERE
The magnitude of these effects appears quite substantial. One cannot easily interpret coefficients in isolation in logistic regression because their effects change depending on the levels of the other variables. Our calculations here calculate the effect of changing publication and distance setting all other variables to their mean levels across the population analyzed. Based on the estimates from model 12, the relative risk that a future patent will cite the focal patent declines by 82% as one moves from a location just outside the focal patent's zip code to one 2000 miles away if a patent cites no publications. When a patent references a single Scientific Index publication, the change in relative risk over the same distance drops to 75%; with four references (the 75 th percentile of patents citing science) it goes to only 33%. With regard to social distance, the relative risk that a future patent from a different class will cite the focal patent increases by about 5% (not significant) if it references a Scientific Index journal.
Time plays an important role in diffusion processes. Over time one would expect social networks to substitute for the dissemination of knowledge through publication by passing this information through interpersonal communications. Thus, the most pronounced effects of publication should appear in the early years after a patent's granting. To check for this pattern, we estimate the model on just those future patents granted within two years of the focal patent's granting and compare it to estimates based on patents granted four to six years following the focal patent. 22 Splitting the data along these lines reveals some interesting patterns. Most notably, references to the scientific literature expand the dispersion of future citations in geographic and technical space substantially in the first two years. In the first two years, the relative risk of citation by a patent 2000 miles distant is 96% lower than for one in the adjacent zip code, a difference that declines to 75% with a single reference to an index journal. Meanwhile, the relative risk of a citation from a future patent in a different class increases by roughly 80% when the focal patent references a scientific article. By four years out, however, references to the scientific literature no longer appear to extend the spatial diffusion of citations. Table 7 shows that similar patterns of results hold for publications that do not ascribe to the norms of science. This comparison again uses the two types of publications that appear most distant from the scientific community: non-index journals and non-technical

Discussion
The results support the idea that science-based patents receive more citations because the act of publication allows their ideas to diffuse more rapidly. Although patents from universities and those that make reference to scientific publications receive more citations than those that do not reference non-patent material, patents that cite any type of publication -including items such as popular magazines and marketing materialsexhibit a similarly large number of citations from future patents. Therefore, the number of citations received by university patents, and those citing the scientific literature, appears to depend more on the fact that these references represent a broad dissemination of knowledge through publication than on the methods used in scientific research. The spatial distribution of these future citations further supports this pattern. Patents making reference to publications more frequently receive citations from geographically distant inventors and from those operating in different technological spaces. Together these results strongly support the idea that science generates higher citation rates because its norms of openness and publication accelerate the diffusion of knowledge.
In estimating the effects of publication, the models actually suggest that non-scientific publications produce a larger increase in both the number and dispersion of future citations than scientific references. These differences might arise for a variety of social factors. For example, specialty journals may actually impede the movement of ideas across academic departments 23 by narrowing the range of readership to those interested in a specific topic. Also, although the development of specialized languages may enable the codification of knowledge, it also excludes outsiders -a fact that Merton (1938) refers to as the "cult of unintelligibility". These specialized vocabularies likely impede the flow of ideas across community boundaries from scientists to engineers (Allen, 1977).
Nevertheless, these private vocabularies allow scientists to maintain their status by excluding outsiders, so even if common language could transmit the same ideas, they might have a vested interest in maintaining these specialized languages (Abbot, 1988).
Although we believe that the results strongly support communication -as opposed to quality differences -as the driving force differentiating science-based invention from other scientific research, some alternative explanations warrant consideration.
One difficulty in investigating the distribution of citations geographically comes from separating the flow of knowledge from the distribution of inventive activity. If citations appear highly localized, two factors could explain that distribution. On the one hand, interpersonal communication might distribute important information through space.
However, citations could also localize simply because everyone researching the problem resides in a concentrated geographic area. Jaffe, Trajtenberg, and Henderson (1993) controlled for this problem by restricting their analysis to patents within a particular technology class. They measured the likelihood of a local citation to a patent, compared to a temporally proximal patent in the same technology class. Though this sampling method can yield biased estimates of the predictors of citation likelihood, it does control for the distribution of activity. We addressed this issue in two ways. First, we created a 23 The strength of this delimitation appears to vary dramatically from field to field. For example, Friedkin (1978) finds substantial variation in the proportion of out of department ties in his study of six departments at an elite U.S. school. distance control by averaging the mean distance between all dyadic pairs within a class.
Technological fields with highly localized activity will receive lower scores on this control than those with more dispersed activity. The inclusion of this term does not affect the interaction terms between publication and spatial distance. Second, we estimated a set of models (unreported) predicting the distance between citing and cited patent including fixed-effects by class to capture differences in the distribution of activity.
These models also show that patents citing published sources receive cites from more distant inventors. 24 Though technological fields do vary in the geographic distribution of activity, these differences do not appear to explain our results.
Another explanation offered for the larger number of citations received by university patents argues that academics only patent their most valuable research. Presumably then, the quality of knowledge produced by both universities and the private sector does not differ substantially. Researchers working in academia simply maintain a higher threshold over which an idea must pass before they consider it worthy of patenting. In support of this idea, as the number of patents issued to universities has increased, the citation premium received by university patents appears to have declined (Henderson, Jaffe and Trajtenberg, 1998;Hicks, et al., 2001). Nonetheless, this explanation applies to the number of citations that one would expect a science-based patent to receive rather than the geographic distribution of those citations. Thus, it cannot explain the pattern of results seen here. Moreover, this explanation fails to account for the increased number of citations received by all patents referencing scientific and non-scientific publications whether assigned to a university or not.
Others contend that universities differ in the type of knowledge they produce, rather than on some hierarchical dimension of quality. Nevertheless, the diffusion effects may account for this claim as well. Henderson, Jaffe and Trajtenberg (1998) argue that universities produce patents that differ from industry-based research in the basic-ness of the knowledge generated. In other words, university research applies to a broader range of inventions. Their evidence for this difference comes from the number of future citations patents receive outside their own technological class. Nevertheless, differences in the diffusion of knowledge would generate the same pattern of effects. Networks localize in social space as well as geographic space; researchers more likely know other researchers within their own field. Therefore, one would expect information traveling through social networks to diffuse slowly across technological fields. As with geographic space, publication removes the dissemination of knowledge from the constraints of these networks. Thus, one would expect publication to increase the flow of information across technological fields much as it does across geographic space, an expectation supported by the dispersion analysis.
Our results suggest a range of policy implications from the firm to the national level. At the firm level, the fact that publication extends the flow of knowledge offers interesting insight into the burgeoning literature on the growing use of science in for-profit organizations. These studies frequently argue that firms that adopt the norms of science experience superior performance compared to those that do not (Cohen and Levinthal, 1990;Henderson and Cockburn, 1994;Zucker and Darby, 1996;Zucker, Darby and Brewer, 1998). Although many of these studies intimate that science improves research productivity, our results suggest instead that one of two alternative explanations accounts for these results. First, our results appear to complement recent work by Stern (1999), which suggests that firms benefit by promoting science not because its methods inherently offer value but rather because talented employees desire the ability to publish.
Hence, they will accept lower compensation from firms that allow them to stay active in the scientific community. A second possible explanation stems from the fact that publication reduces the constraints of location. Firms must typically locate near to critical resources because the information critical to mobilizing those resources flows through spatially constrained social networks (Sorenson and Audia, 2000;Sorenson and Stuart, 2001). However, by drawing information from published sources, firms that do science-based research may gain flexibility in location, allowing them to avoid some labor market competition by locating facilities distant from rivals. Like Stern's explanation, this possibility points to a reduction in labor costs, rather than an increase in the quality of research as the advantage of adopting a scientific orientation.
From a public policy point of view, our results imply a clearer course of action. The research on productivity suggests that research and development may stimulate GDP growth (e.g., Adams, 1990). The process of science itself appears to offer little benefit; rather, what benefits might accrue from science-based research stems primarily from the dissemination of information -sometimes called spillovers. The rapid diffusion of knowledge can improve the efficiency of investments in R&D by reducing the degree of effort duplication. To the extent that this reduction in overlapping research activities increases the efficiency of R&D investments in stimulating economic growth, public funding sources should require timely publication of the results of the research projects they support to maximize societal benefits.
The wisdom of this course of action, however, might vary across countries, or even regions within countries. Research finds an inconsistent relationship between research and economic growth. Whereas studies of the U.S. often find a positive relationship, cross-national investigations find a negative relationship between publications and economic growth (Schofer, Ramirez and Meyer, 2000 The belief that science promotes economic growth plays an important role both in the allocation of resources to science and to its status in society. Nevertheless, we lack a clear understanding of both the accuracy of this belief and the mechanisms underlying the relationship. This paper takes one step toward resolving this ambiguity by demonstrating that science appears to benefit technological innovation primarily by expanding the flow of information in geographic and social space. Nonetheless, completely understanding this issue will require substantial future research.     Models 8 and 9 enter a dummy variable (i.e. 0 or 1) for whether a patent references each type of publication, rather than a count of the number of publications referenced of each type. Model 9 also excludes self-citations from the dependent variable, the count of future patent citations.