Patent citations reexamined

• Date: 01 March 2020
• Published date: 01 March 2020
• DOI: http://doi.org/10.1111/1756-2171.12307
• Author: Jeffrey Kuhn,Kenneth Younge,Alan Marco

RAND Journal of Economics

Vol.51, No. 1, Spring 2020

pp. 109–132

Patent citations reexamined

Jeffrey Kuhn∗

Kenneth Younge∗∗

and

Alan Marco∗∗∗

Many studies rely on patent citations to measure intellectual heritageand impact. In this article,

we show that the nature of patent citations has changed dramatically in recent years. Today,

a small minority of patent applications are generating a large majority of patent citations,

and the mean technological similarity between citing and cited patents has fallen considerably.

We replicate several well-known studies in industrial organization and innovation economics

and demonstrate how generalized assumptions about the nature of patent citations have misled

the field.

1. Introduction

Scholars often use patenting as an empirical measure of innovation. Seminal work by

Griliches (1981) measured patent counts, and later work weighted such counts by the number

of forward citations received by each patent on the view that more important patents receive

more references (Trajtenberg, 1990a). Later studies used patent citations to measure private

patent value (Lanjouw and Schankerman, 2001; Harhoff et al., 1999), firm market value (Hall,

Jaffe, and Trajtenberg, 2005), cumulative innovation (Caballero and Jaffe, 1993; Trajtenberg,

Henderson, and Jaffe, 1997), geographic spillovers (Jaffe, Trajtenberg, and Henderson, 1993),

technology life-cycles (Mehta, Rysman, and Simcoe, 2010), social importance (Moser,Ohmstedt,

and Rhode, 2017), originality (Jung and Lee, 2016), and technological impact (Corredoira and

∗University of North Carolina at Chapel Hill; jeffrey_kuhn@kenan-flagler.unc.edu.

∗∗ ´

Ecole Polytechnique F´

ed´

erale de Lausanne; kenneth.younge@epfl.ch.

∗∗∗Georgia Tech; alan.marco@pubpolicy.gatech.edu.

Wethank Ian Cockburn, Rui de Figueiredo, Jeff Furman, Michelle Gittelman, Dietmar Harhoff, Robert Merges, Michael

Meurer, Atul Nerkar,Gaetan de Rassenfosse, Katja Seim, Tim Simcoe, Neil Thompson, Tony Tong,Andrew Toole, Brian

Wright, Noam Yuchtman, and anonymous reviewers for helpful comments. We also thank seminar participants at the

2016 Searle Center Conference on Innovation Economics, the 2017 Boston University LawSchools Technology Policy

and Research Initiative Conference, the 2016 Munich Summer Institute, the United States Patent and Trademark Office

(2015), Skema Business School (2015), the 2016 DRUID Conference, and the 2016 Academy of Management Annual

Meeting. The authors thank the United States Patent and Trademark Office for providingaccess to data. The authors also

thank Google for a generous research grant of computing time on the Google Cloud.

C2020, The RAND Corporation. 109

110 / THE RAND JOURNAL OF ECONOMICS

FIGURE 1

TOTAL CITATIONSMADE BY YEAR, DIVIDED BY NUMBER OF BACKWARD CITATIONS MADE BY

CITING PATENT

Banerjee, 2015). Overall, a search for “patent citations” on Google Scholar (conducted January

1, 2018) returns over 21,000 results.

Recent research, however, has begun to call into question a straightforward interpretation of

patent citations. Abrams, Akcigit, and Popadak (2018) find an inverted-U relationship between

patent citations and market value, suggesting that high citation counts may indicate strategic use

of the patent system (instead of impactful innovation). Evidence also has emerged that the search

for and disclosure of prior art varies between applicants in systematic ways (e.g., Sampat, 2010).

Citations may suffer from significant noise and measurement error (Gambardella, Harhoff, and

Verspagen,2008; Roach and Cohen, 2013), and the comparison of patents between cohorts can be

problematic because citation counts have inflated substantially over time (Marco, 2007). Failing

to correct for time period, technology, and geographic region can introduce significant bias into

an analysis (Lerner and Seru, 2017). Correcting citation counts to return to the original goal of

devising a measure of innovation activitythat is broadly comparable across contexts, however, is

problematic due to endogeneity in the pendency, citation lags, and filing years of a given sample

(Mehta, Rysman, and Simcoe, 2010).

In this article, we highlight an important change in the data generating process of patent

citations that changes the statistical nature of patent citation measures and the appropriate use

of patent citations in current and future research. Specifically, we observe a dramatic increase

in the number of citations generated per year and relate that change to a small proportion of

patents flooding the patent office with an overwhelming number of references. Figure 1 shows

the number of backward citations over time, segmented bythe number of citations made by each

patent. In recent years, 46.8% of references derive from less than 5% of patents with more than

100 backward citations.1

To examine how the aforementioned changes in citation patterns affect the information

content and quality of citation-based measures, we use a vector space model based on Younge

and Kuhn (2015) that compares the text of every patent to every other patent granted by the

1Conversations with examiners suggest that reviewing more than 100 citations is extremely difficult, given time

constraints.

C

The RAND Corporation 2020.