How localised are knowledge spillovers? Evidence from microgeographic data on UK patent citations

ABSTRACT

We model the spatial characteristics of technological knowledge flows in the UK. Using a novel and highly accurate dataset of inventor locations, we test for localisation of knowledge spillovers in citations between UK patent applications from 1982 to 2015. We apply continuous distance localisation tests separately to patent citations in 313 technologies and find that spillovers are localised in far fewer technologies and at shorter distances than previous studies have suggested. Only 30% of technologies in the UK display localisation, knowledge spillovers decay rapidly at distances between 30 and 80 km, and spillovers within technologies are twice as frequently localised as spillovers between technologies. Our results suggest that technological and geographical proximity are important determinants of knowledge spillovers in the UK and that close physical proximity is particularly relevant for industrial sectors that are more reliant on tacit knowledge.

KEYWORDS:

• Knowledge spillovers
• localisation
• patent citations
• distance

JEL CLASSIFICATIONS:

• O31
• O33
• R12

Acknowledgements

We thank Dominique Guellec, Dietmar Harhoff, Imko Meyenburg and two anonymous referees for helpful comments. We are grateful to the UK Intellectual Property Office for providing inventor address data under the Open Government Licence 3.0.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are openly available in figshare at http://doi.org/10.6084/m9.figshare.16771195.

Notes

1 Other approaches not exploiting patent citations include R&D production functions (e.g. Bottazzi and Peri 2003) and Mincerian wage regressions (e.g. Rauch 1993; Rosenthal and Strange 2008).

2 Murata et al. (2014) find such downward bias to be large relative to the upward bias induced by imperfect matching of controls (Thompson and Fox-Kean 2005).

3 Patent citations may also be localised if inventive activity is co-located for reasons unrelated to knowledge spillovers, such as geographically concentrated specialised labour markets (Almeida and Kogut 1999; Breschi and Lissoni 2009) or firms and inventors co-locating to exploit specialised inputs (Porter 1990; Helmers 2017). We control for this by testing for localisation relative to existing patterns of concentration of patenting activity by using (Jaffe, Trajtenberg, and Henderson 1993) style control patents (see Section 3).

4 We choose the range [0, 170] because the median distance between originating and citing patents is 161 km.

5 For both local likelihood and local linear estimates we use a tricube weight function and select bandwidth by least squares cross validation. We use a local quadratic polynomial for our local likelihood estimates.

6 In contrast, the majority of inventor addresses in USPTO patent data are recorded as only an inventor's city and state.

7 In the US applicants are required to supply a complete list of material relevant to the patentability of the invention. Failure to do so risks revocation of any subsequently granted patent. Consequently, US applicants, and their patent attorneys, tend to cite every reference that might conceivably be related to the application, irrespective of whether they accurately represent knowledge spillovers (Michel and Bettels 2001).

8 The more common approach of using citations to originating patents filed in a fixed period (Jaffe, Trajtenberg, and Henderson 1993; Murata et al. 2014; Figueiredo, Guimarães, and Woodward 2015) results in too few observations to test for localisation separately in each technology class. For example, restricting our sample to the 29,103 applications filed between 1982 and 1989 yields a sample of only 8238 originating patents which attracted 18,402 citations.

9 UK postcodes identify either a single address receiving large volumes of mail or a small number (up to 100, though 15 is more typical) of usually adjacent addresses receiving smaller volumes of mail (ONS 2019).

10 Street names and zip codes are available for fewer than 20% of inventors (Breschi and Lissoni 2009; Murata et al. 2014).

11 Whilst the USPTO data should consistently record the location of each inventor's residence rather than their workplace, Kerr and Kominers (2015) discuss facing the same issue in their US data.

12 In some cases, this may even more than offset the advantages of using point locations for an inventor's reported address. For example, for an inventor who lives and works on opposite sides of a city, the city centre would be closer to their workplace than their home address.

13 To check the sensitivity of our results to the choice of a 6-month window between the filing dates of citing patents and potential controls, we also reran our main specification using both 3-month and 12-month time windows. Our results were essentially unchanged.

14 There are 639 IPC sub-classes in the 2006 version of the IPC that we use compared to the 417 USPC classes in the NBER Patent Citations Data (Hall, Jaffe, and Trajtenberg 2001).

15 The equivalent distances for our local likelihood and local linear based tests are 74 and 93 km respectively.

16 Since the boundary robust tests detect localisation more frequently, this may suggest that the K-density based tests under-detect localisation (Murata, Nakajima, and Tamura 2017). We discuss this in more detail in Section A.2 of the online appendix.

17 The distances within which 95% of localised classes are localised, Murata et al.'s (2014) widest extent of localisation, are 75 km for our K-density tests, 92 km for our local likelihood tests and 94 km for our local linear tests.

18 Although not directly comparable, our results are very similar to those of several papers applying continuous distance tests to identify clusters of innovative firms in the US which find that patent citations are localised within distances between 8–32 km (Buzard et al. 2017, 2020) and 80–96 km (Kerr and Kominers 2015).

19 This difference is statistically significant. McNemar tests reject a null hypothesis of equal proportions with test statistics of 13 ($p-\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{u}\mathrm{e}=0.000$) for K-density tests, 17 ($p-\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{u}\mathrm{e}=0.000$) for local likelihood, and 21 ($p-\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{u}\mathrm{e}=0.000$) for local linear.

20 This does mean, however, that our inter-class test for each technology will, on average, be based on fewer observed citations than the intra-class test for that technology. If the power of our tests to detect localisation is reduced in small sample sizes, this would cause us to under-detect localisation in inter-class citations. We present a robustness check against this concern in Section A.6 of the online appendix.

21 These regression-based approaches typically find negative coefficient(s) for geographic distance variable(s) and positive coefficient(s) for technological relatedness variable(s) added independently. We are not aware of any work explicitly including interaction terms between the two.

22 The frequency with which automotive technologies are localised is not apparent in Panel A of Table 4 as they are split between transport-related sub-classes in IPC section B (such as B60–Vehicles in general and B62D–Motor vehicles; trailers) and more general engineering sub-classes in IPC section F (such as F16–Engineering elements).

Additional information

Funding

We also thank Essex County Council for funding an earlier project that led to this paper.