Patents, family, and size: evidence from Italian manufacturing firms

ABSTRACT

This study explores whether the probability to patent differs between family and non-family firms, and whether any potential difference between firm-type is moderated by size. The analysis is based on a large archive of patenting activities (Orbis–PATSTAT dataset) carried out by around 3700 Italian manufacturing firms over the 2010–2017 period. The results from a random effect probit model show that family firms patent less than non-family firms (the estimated average marginal effect of family ownership ranges from −0.055 to −0.032 according to model specification). Furthermore, the size effect is positive in every model, suggesting that the probability of patenting increases with size. While it is demonstrated that family firms remain less likely to patent than non-family firms, we also show that their disadvantages increase as they grow in size: in large firms, the probability of patenting is 0.22 for family firms and 0.39 for non-family firms. Importantly, the results hold when considering patent counts, citations and a number of additional sensitivity tests.

KEYWORDS:

• Family firms
• Patents
• Firm size
• Innovation

JEL CLASSIFICATION:

• D22
• L25
• L60
• O30

Disclosure statement

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

Notes

1 This study is based on patent applications filed at the EPO rather than at the Italian Patent Office. Given that national filings are likely to have lower average quality (Boeing and Mueller 2016; Deng 2007) and have lower costs associated with the European patenting route, the measure adopted in this study could be biased in favour of large companies compared to small and medium enterprises.

2 Only priority patents were included, while equivalent patent filings were excluded. A priority patent is the first patent filing made by applicants to protect the invention in a given country, and the equivalent patents are the subsequent filings made in other patent offices where protection is sought.

3 For a discussion of the advantages and disadvantages of using patents as a measure of technological change, see Archibugi and Pianta (1996) and Aiello et al. (2021a). Here, it is worth summarising a few points. While patents have some drawbacks as indicators of technological activity (not all inventions are patented, and the incentives to patent differ according to the sector and market), they present a number of advantages over alternative measures of innovation. Notably, patents are commensurable as they are based on an objective standard; that is, the type of invention that can be patented is clearly defined, meaning that patents are probably the most definite measure of innovation (Wang 2007). Indeed, compared with other innovation measures, usually gauged through surveys, patents are less exposed to personal views. They also reflect the quality of an innovation, as any patentable innovation is examined by experts who evaluate its novelty and utility. By contrast, reliable information on the quality of an innovation can rarely be gathered from other sources, especially if they are based on subjective judgements. Moreover, and differently from R&D expenditures, patents measure the outputs of the inventive process, thereby gauging the market value of an R&D project better than investments. Finally, patent data are quantitative and widely available. For these reasons, their use as a measure of the output of the inventive process has become widespread in the literature (Griliches 1990; Hall, Griliches, and Hausman 1986).

4 Firm size is measured in terms of annual turnover, which allows the sample to be split on the basis of the threshold values reported in the Commission Recommendation 96/280/EC (updated in 2003/361/EC of May 6, 2003).

5 Detailed sectoral and geographical distributions of the sample employed when performing the econometric analysis (i.e. NUTS-2 regions and NACE 2-digit code, respectively) are given in Tables A2 and A3 in the appendix.

6 It may be possible to estimate fixed effects with a Probit including firm-specific dummy variables, however, the estimates would not be consistent. See Cameron and Trivedi (2005) for an exhaustive discussion on nonlinear panel data models. Another potential method to be used is the fixed effects logit estimator. However, it impedes to assess the role of Family as it is time-invariant. Alternatively, after splitting the sample in FFs and non-FFs, the logit estimator might be applied to estimate the size-effect in the two groups. In such a case, results are difficult to be interpreted because the estimator drops all firms that never patent or always patent during the period under scrutiny, thereby significantly altering the sample size. The logit estimator drops 1187 firms, that is, almost one third of the sample.

7 Here, it is useful to say that we model the probability to patent because we are not able to distinguish between the propensity to patent, leading from inventions to patents, and research productivity, leading from research to inventions. Indeed, it is worth noting that both dimensions could determine a given number of patents, and were found to be affected by the design of policy tools, such as education, and policies on intellectual property and science and technologies (De Rassenfosse and de la Potterie 2009).

8 In the related literate, a threshold of at least 50% of the company’s shares is commonly used for privately held companies (e.g. Arregle et al. 2012; Arzubiaga et al. 2018; Broekaert, Andries, and Debackere 2016; Classen et al. 2014; Memili et al. 2015b; Meroño-Cerdán, López-Nicolás, and Molina-Castillo 2018; Miller, Minichilli, and Corbetta 2013; Vandekerkhof et al. 2018). This common practice seems to be appropriate for our setting because the structure of firm ownership in Italy is characterised by a limited number of shareholders with very large block holdings, thereby implying that a 50% stake is enough to achieve control of the company (Miller, Minichilli, and Corbetta 2013). Some robustness checks, however, are performed considering the thresholds of 40% and 60% and a continuous variable of family ownership.

9 The intangible nature of a stock of knowledge makes it difficult to determine the depreciation rate of knowledge. Some scholars use a rate of 15% (Aiello and Cardamone 2012; Goto and Suzuki 1989; Griliches and Mairesse 1983; Hall, Jaffe, and Trajtenberg 2005; Laurens et al. 2017), while others consider 10% (Bitzer and Stephan 2007; Montobbio and Solito 2018; Zawalińska, Tran, and Płoszaj 2018). While we applied a rate of 10%, a sensitivity analysis has been conducted using 15%. The results are robust, and available upon request.

10 A typical argument in the neo-Schumpeterian literature is that the characteristics of a particular sector or industry with which a firm is affiliated may affect its innovation activity. Different sectors have different technology and innovation opportunities and are thus characterised by different technological regimes (Malerba and Orsenigo 1993).

11 The Nomenclature of Territorial Units for Statistics (NUTS) is a geographical nomenclature subdividing the economic territory of the European Union into regions at three different levels (NUTS 1, 2 and 3 respectively, moving from larger to smaller territorial units).

12 The variables of the stock of patents and size are included with a one-year lag to take in account the likelihood that these factors will affect the probability to patent in technologies with a period lag.

13 Compared to Table 3, the results displayed in Tables 4 and 5 refer only to the model with the product term between Family and Size. However, the robustness checks have also been performed for the other three models of Table 3, and the estimates are provided as online supplementary material.

14 Following Chirico et al. (2020) we augment the model by including the quadratic term of family ownership. However, the likelihood-ratio test to determine whether adding Family Ownership² improves model fit yields evidence in favour of the model without the squared term. This conclusion is supported by the Bayesian information criterion (BIC) (Raftery 1995). This does not seem to be unexpected evidence, as the non-linearity of the estimator we use still captures the non-linear effect of Family Ownership (non-linearity is, for instance, depicted in Figure 3). In any case, the results of the model with the squared term of Family Ownership² are provided as online supplementary material.

15 The Poisson regression is considered appropriate for the analysis of discrete data with many zeros and small values (Greene, 2011), however, in our sample, there is overdispersion of patent data, as the variance is higher than the mean. In order to relax the assumption of equal conditional mean and variance functions (Greene, 2011), we thus also employ a negative binomial, which, in our case, turns out to be the most suitable method to model patent counts due to overdispersion (Poisson model results are in the appendix, Table A5). Another methodological choice made in this part of the study involves the use of random effect models. This is due to the fact that for every variable used in regressions, the total variation is prevalently due to between-variation rather than within-variation. In such a case, applying the fixed-effects estimator implies that the coefficients of the time-invariant regressors are not identified, and many observations are dropped because they are time invariant (Greene, 2011).