To answer the hypotheses formulated in Section 6.1, I first compare the patenting values of Fabless and IP Core companies through univariate analysis comparing their variances and the mean of the distribution for the overall time period and individual years.
To control for the influences of other explanatory variables than being a technology provider I next use a panel dataset that combines patenting information (see Section 3.3.3) with publicly available company information (see Section 3.3.4) for both IP Core and Fabless companies. I analyze this using two different sets of regressions, the first set utilizes clustered standard errors on the firm level for the combined IP Core and Fabless company datasets, the second set uses firm-level fixed effects for the IP Core and Fabless company datasets separately. The unit of observation is a firm-year combination. I was not able to investigate Hypothesis 1 through a random effects panel regression since the Hausman test for correlation between the error terms and the independent variables rejected the Null (i.e., there was correlation between the error terms and the independent variables) requiring a fixed effects regression. This, however, would have eliminated my dummy variable ‘IP Core’ that is perfectly collinear with the firm-specific dummy variables since no firm changed from being IP Core to Fabless or vice versa. Therefore, I utilize an ordinary least squares (OLS) model with clustered standard errors since the panel character of the data is not essential to answering the first hypothesis (patenting characteristics are not likely to be dependent on the patenting characteristics of the previous year but rather to be determined by firm characteristics).
I compute for the regressors, in accordance with Hall and Ziedonis (2001), the decadic logarithm of R&D spending, number of employees, plant and property equipment (PPE), the firm age (Year – Year of Inception), and the year (to capture time trends). I use logarithms because they capture the decreasing changes in firm behavior for a given absolute change of a variable as the starting position of the variable increases (i.e., the patenting behavior of a firm is likely to change more as it grows from 100 to 200 employees than from 1100 to 1200 – a difference better captured by the logarithm) and also reduce the influence of outliers (Cohen et al., 2014). Herein, number of employees serves as a proxy for firm size and PPE as a proxy for capital intensity. Additionally, all monetary values are discounted to 2005 USD using the US gross domestic product
deflator (US Bureau of Economic Analysis, 2016). See Table 22 in Appendix A 4 for detailed per-company information of the explanatory variables.
As the dependent variable I use the decadic logarithm of the patenting intensity, which equals the number of patents filed in a given year divided by million dollars of R&D investment for that year in 2005 USD, plus one to enable computation of the log for firms that have not filed for a single patent in the respective year ( 𝑦𝑖,𝑡 =
𝑃𝑎𝑡𝑒𝑛𝑡𝑠𝑓𝑖𝑙𝑒𝑑𝑖,𝑡
𝑅&𝐷 𝑠𝑝𝑒𝑛𝑑𝑖𝑛𝑔𝑖,𝑡+ 1). This variable has also been used by Hall and Ziedonis (2001), termed as propensity to patent in their analysis of the patent paradox. I deviate from this terminology since I follow past scholars (e.g., Brouwer and Kleinknecht, 1999; Jell et al., 2016) in their description of two distinct, subsequent process parameters. The first process parameter, termed R&D efficiency, describes the number of patentable innovations generated per million dollars R&D budget. The second process parameter, termed propensity to patent, captures how many patents a company files given a certain set of patentable innovations. The described measure of number of patent applications divided by million dollars R&D budget (i.e. the product of the two process parameters introduced before) is called patenting intensity.
I further include a dummy for the three EDA companies due to their different business models (see Section 2.2.2) and create additional robustness tests excluding Qualcomm from the regressions due to its size and corresponding high patenting activity (see Section 3.3.3 and Section 4.5). A Qualcomm dummy with a coding of 1 for Qualcomm is not possible for the models using clustered standard errors on the firm level since this would have led to only one firm cluster exhibiting a ‘1’ while all the rest would have exhibited ‘0’ for this independent variable, which in turn leads to Stata no longer computing an F statistic and p-value for the entire distribution (see last entry of STATA help file ‘help j_robustsingular’ for details). Instead, I recalculate the regressions without the Qualcomm data as robustness tests and receive similar results (see Appendix A 11 for the corresponding figure).
Finally, it is important to stress that the regressions I run are not able to generate statements with respect to causality; they merely allow for the identification of correlations between dependent and independent variables. One illustrative example of this issue would be a positive correlation between firm size and patenting intensity. On the one hand, a company could start pursuing a more aggressive R&D agenda of hiring new researchers, which would, in turn, lead to a higher patenting intensity; thereby, the
increase in size causes an increase in patenting intensity. On the other hand, this same company could have had a breakthrough innovation using the existing staff some two years earlier and then began both filing a higher number of patents to exclude competitors from exploiting this innovation and hiring more staff to prepare for internal exploitation of the new technology. In this case, I would also see a rise in number of employees and patenting intensity, yet they would not be causally linked one way or another; rather, both would be due to the independent event of having a breakthrough innovation. Since my data does not allow distinction between these various scenarios (and the plethora of alternative imaginable scenarios leading to the same outcomes), I refrain from making statements with regard to causality in most cases, except where one decision clearly precedes the other (like choice of business model).
To investigate the first hypothesis, whether being an IP Core provider leads to a significantly higher incentive to patent, I compute four regressions including different dependent variables using regressions and clustered standard errors on firm level since I believe that the patenting intensity of year t does not depend so much on the patenting intensity of year t-1 itself, but rather that the driver of patenting behavior is the underlying firm. I therefore decided to address this firm heterogeneity by clustering standard errors on the firm level and treating the individual years as independent observations (see Equation 2). To account for overall time trends, I include Year as a regressor.
Equation 2: OLS regression of R&D expenditure, capital intensity, size, year and age on patenting intensity using clustered standard errors on firm level
Model 1:
𝑦𝑖,𝑡 = 𝛽0+ 𝛽1∗ 𝐼𝑃𝐶𝑜𝑟𝑒𝑖 + 𝛽2∗ 𝐿𝑜𝑔𝑅𝑛𝐷𝑖,𝑡+ 𝜀𝑖,𝑡 Model 2:
𝑦𝑖,𝑡 = 𝛽0+ 𝛽1∗ 𝐼𝑃𝐶𝑜𝑟𝑒𝑖 + 𝛽2∗ 𝐿𝑜𝑔𝑅𝑛𝐷𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽3 ∗ 𝐿𝑜𝑔𝑃𝑃𝐸𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽4
∗ 𝐿𝑜𝑔𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽5∗ 𝑌𝑒𝑎𝑟𝑡+ 𝜀𝑖,𝑡 Model 3:
𝑦𝑖,𝑡 = 𝛽0+ 𝛽1∗ 𝐼𝑃𝐶𝑜𝑟𝑒𝑖 + 𝛽2∗ 𝐿𝑜𝑔𝑅𝑛𝐷𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽3∗ 𝐿𝑜𝑔𝑃𝑃𝐸𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡 + 𝛽4
∗ 𝐿𝑜𝑔𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽5∗ 𝑌𝑒𝑎𝑟𝑡+ 𝛽7∗ 𝐸𝐷𝐴𝑖 + 𝜀𝑖,𝑡 Model 4:
𝑦𝑖,𝑡 = 𝛽0+ 𝛽1∗ 𝐼𝑃𝐶𝑜𝑟𝑒𝑖 + 𝛽2∗ 𝐿𝑜𝑔𝑅𝑛𝐷𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽3∗ 𝐿𝑜𝑔𝑃𝑃𝐸𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡 + 𝛽4
∗ 𝐿𝑜𝑔𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽5∗ 𝑌𝑒𝑎𝑟𝑡+ 𝛽6∗ 𝐴𝑔𝑒𝑖,𝑡+ 𝛽7∗ 𝐸𝐷𝐴𝑖 + 𝜀𝑖,𝑡
To investigate the second hypothesis regarding determinants of patenting behavior, I split my dataset into IP Core and Fabless datasets and perform panel regressions using firm-level fixed effects for each separate dataset since I am not concerned about losing the IP Core dummy for this inquiry. As I did for the clustered standard errors, I build corresponding models to increase comparability but omit the EDA control variable since it is collinear with the firm fixed effects (see Equation 3).
I also perform robustness tests removing individual years and individual companies, the two crisis years (2008 and 2009), and the two largest Fabless companies (Qualcomm and AMD whose average R&D budgets over the years 2005-2013 eclipse those of the next smaller company, NVidia, by a factor of 3 and 1.7, respectively, and can be assumed to operate differently due to their scale).
Equation 3: Firm-level fixed effects regressions of R&D expenditure, capital intensity, size, year and age on patenting intensity
Model 1:
𝑦𝑖,𝑡 = 𝛽0+ 𝛽1∗ 𝐿𝑜𝑔𝑅𝑛𝐷𝑖,𝑡+ 𝜈𝑖+ 𝜀𝑖,𝑡 Model 2:
𝑦𝑖,𝑡 = 𝛽0+ 𝛽1∗ 𝐿𝑜𝑔𝑅𝑛𝐷𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽2∗ 𝐿𝑜𝑔𝑃𝑃𝐸𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽3∗ 𝐿𝑜𝑔𝐸𝑚𝑝𝑙𝑖,𝑡 + 𝛽4∗ 𝑌𝑒𝑎𝑟𝑡+ 𝜈𝑖 + 𝜀𝑖,𝑡
Model 3:
Omitted since additional variables collinear with firm-level fixed effects Model 4:
𝑦𝑖,𝑡 = 𝛽0+ 𝛽1∗ 𝐿𝑜𝑔𝑅𝑛𝐷𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡 + 𝛽2∗ 𝐿𝑜𝑔𝑃𝑃𝐸𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽3∗ 𝐿𝑜𝑔𝐸𝑚𝑝𝑙𝑖,𝑡 + 𝛽4∗ 𝑌𝑒𝑎𝑟𝑡+ 𝛽5∗ 𝐴𝑔𝑒𝑖,𝑡+ 𝜈𝑖 + 𝜀𝑖,𝑡
Finally, during separate robustness tests I find that, in contrast to the clustered standard error regressions where no quadratic terms were significant, the PPE per Employee exhibits a significant quadratic relationship in the fixed-effects regressions.
Therefore, I also compute a third set of regressions.
Equation 4: Firm-level fixed effects regressions of R&D expenditure, (quadratic) capital intensity, size, year and age on patenting intensity
Model 1:
𝑦𝑖,𝑡 = 𝛽0+ 𝛽1∗ 𝐿𝑜𝑔𝑅𝑛𝐷𝑖,𝑡+ 𝜈𝑖+ 𝜀𝑖,𝑡 (same as Model 1 in Equation 3) Model 2:
𝑦𝑖,𝑡 = 𝛽0+ 𝛽1∗ 𝐿𝑜𝑔𝑅𝑛𝐷𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽2∗ 𝐿𝑜𝑔𝑃𝑃𝐸𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽3
∗ 𝐿𝑜𝑔𝑃𝑃𝐸𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡∗ 𝐿𝑜𝑔𝑃𝑃𝐸𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽4∗ 𝐿𝑜𝑔𝐸𝑚𝑝𝑙𝑖,𝑡 + 𝛽5
∗ 𝑌𝑒𝑎𝑟𝑡+ 𝜈𝑖 + 𝜀𝑖,𝑡 Model 3:
Omitted since additional variables collinear with firm-level fixed effects Model 4:
𝑦𝑖,𝑡 = 𝛽0+ 𝛽1∗ 𝐿𝑜𝑔𝑅𝑛𝐷𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽2∗ 𝐿𝑜𝑔𝑃𝑃𝐸𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽3
∗ 𝐿𝑜𝑔𝑃𝑃𝐸𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡 ∗ 𝐿𝑜𝑔𝑃𝑃𝐸𝑝𝑒𝑟𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽4 ∗ 𝐿𝑜𝑔𝐸𝑚𝑝𝑙𝑖,𝑡+ 𝛽5
∗ 𝑌𝑒𝑎𝑟𝑡+ 𝛽6∗ 𝐴𝑔𝑒𝑖,𝑡+ 𝜈𝑖 + 𝜀𝑖, 𝑡
The next section details the regression results for the models described in this chapter and provides robustness tests.