Tracking the technological composition of industries with algorithmic patent concordances

ABSTRACT

Patents are a useful proxy for innovation, technological change, and diffusion. However, fully exploiting patent data for economic analyses requires linking patents to measures of economic activity, which has proven to be difficult. We construct probabilistic linkages between the U.S. Patent Classification (USPC) system and Cooperative Patent Classification (CPC) system and industry and product classifications including the North American Industrial Classification System (NAICS), International Standard Industrial Classification (ISIC), Harmonized System (HS) and Standard International Trade Classification (SITC). We use these concordances to evaluate the persistence of technology-industry relationships over time by generating linkages over different years of patent data. We find strong persistence in technology usage within industries and, until recently, relatively little change in the technology composition of industries over time. As the technology composition of industries becomes more stable, we find evidence of increased specialization. Finally, we show that industries that exhibit changing technology composition also show shifting occupational composition.

KEYWORDS:

• Patents
• concordance
• classifications
• USPC
• CPC
• technological change
• occupational composition

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 See Griliches (1990) for an early description of how patents can be used as measure of innovation. See Kortum and Lerner (1998) and Hall, Jaffe, and Trajtenberg (2001) for discussions of the heterogeneity of patent value and the use of citation measures to characterize the impacts of individual patents.

2 For example, memory storage devices have undergone incremental changes each year in the speed and amount of data one can store. However, there has also been significant long-term change in the underlying technology as the industry moves away from hard disk drives (HDD) to solid state drives (SSD).

3 These findings are consistent with Colombelli and Quatraro (2014), who find persistence within firm-level knowledge production functions.

4 Layering concordances is akin to translating Japanese into English by first translating it into French.

5 This same benefit may also be seen as a drawback as information about the original technological intent of the patent is lost.

6 The fact that technology classification systems change over time, and that those changes are applied to the entire patent corpus, also underscores the importance of a generalizable concordance that can be continuously updated as new technology classes emerge and are incorporated into the existing patent database. Our methodology allows us to do just that, re-calculating the concordance weights as new data become available.

7 We include updated search terms and key words for: 6-digit North American Industrial Classification System (NAICS), versions 1997, 2002 and 2007, 4-digit International Standard Industrial Classification (ISIC), versions 2.0, 3.0, 3.1 and 4.0, 6-digit Harmonized System (HS), versions 1997, 2002 and 2007, 4 and 5-digit Standard International Trade Classification (SITC), versions 2, 3 and 4.

8 See http://www.patentsview.org (accessed 2/2/2016).

9 PATSTAT Global bulk download was purchased and accessed in the second-half of 2016.

10 We elect to construct the yearly weights using the CPC classification because we have a larger patent corpus to match against (PATSTAT). We have performed an identical analysis using USPC codes and find similar levels of persistence and change. Also, the choice of industry classification is not of consequence as similar findings were made using alternative classification schemes and years.

11 This is similar to what is known as measuring the ‘persistence parameter’ in the macro literature. A study looking at the persistence of innovative activities using a similar methodology was done by Malerba, Orsenigo, and Peretto (1997). In our case, the larger the coefficient value of β₁ and the more variation that can be explained by the prior weights, the higher the degree of persistence.

12 We restrict our analysis to industries that have patent coverage in both (t − δ) and t. This allows us to disregard technology-industry entry and exit, which have the potential to bias our distance weighting.

13 We limit the analysis to this simple, reduced form equation to keep the association as general as possible and focus primarily on within-industry changes to account for unobserved year heterogeneity.

14 These NAICS-based tabulations rely heavily on the earlier USPC to SIC concordances originally developed in 1974.

15 A valid comparison with the concordance developed by the USPTO requires us to first translate our ALP concordances into the NAICS-based categories (OTAF codes) used by USPTO. We next ensure that our sample of patents are as close to the sample as that used by USPTO, meaning that we primarily rely on utility patents granted between 1976 and 2012 (the range of years covered by the USPTO report). In agreement with the USPTO methodology, we equally weight all associated industry codes, which yields us our comparison.

16 The new version of the crosswalk incorporates additional technology categories (IPC) that have been introduced since 2003, and converts the IPCs to NACE Rev. 2 (the original converted IPCs to NACE Rev. 1.1). There are also a handful of 4-digit IPCs that receive a proportional distribution into different NACE categories (for instance, B65D ‘Containers’ is allocated into NACE 22.22, 23.13, 17.21, 25.91, 13.9 and 16.24).

17 For this comparison, we extract the patents that can be concorded using both methodologies and assign equal weights to all 4-digit CPCs. We then concord them to the 2-digit NACE using the EUROSTAT concordance. For the ALP counts, we must first assign the CPCs to ISIC Rev. 4, which the NACE Rev. 2 is based upon. We can then easily convert them to 2-digit NACE Rev. 2 using the correspondence table found in UN Statistics Division website.

18 The set of patents used for the CPC concordance is significantly larger and based on worldwide patents (as opposed to only U.S. patents). The results also hold for USPC-based concordances.

19 The correlation between these two measures is −0.88.

20 For example, see Acemoglu and Autor (2011), Autor, Levy, and Murnane (2003), Bresnahan, Brynjolfsson, and Hitt (2002), Berman, Bound, and Griliches (1994).

21 A full description of the program can be found here: https://pypi.python.org/pypi/topia.termextract/ (accessed 2/2/2016).

22 See http://www.patentsview.org (accessed 2/2/2016) for more details about the PatentsView database. In order to maintain consistency, we limit the patents used in the CPC crosswalk to those applied for between 1976 and 2014, along with the patent actually being granted.

23 For PATSTAT patents, we utilize the first application date as the date for the patent and only included granted patents. As a result, patents in the later years of the PATSTAT (2011 and later) will be limited due to the average time between application and granting (typically 3–5 years).

24 The filter consistently removes between 20% and 25% of matches by industry group, reducing noise in our final weights.

25 For further discussion, see Lybbert and Zolas (2014).

26 The latest versions can be found here: https://sites.google.com/site/nikolaszolas/PatentCrosswalk.

27 There are over 19.4 million granted patents in the PATSTAT database with English language abstracts.