Patent value indicators and technological innovation

Abstract

I provide empirical evidence that patent value indicators can identify technological innovation among inventors and small- and medium-sized enterprises in Sweden. Survey data on the commercialization of patents are related to patent value indicators (patent renewal, patent family size and forward citations) from archival sources. Among the patent value indicators, both the length of patent renewal and the size of the patent family indicate that a patent has been commercialized. Patent renewal for at least 6 years is sufficient to predict an accurate probability of commercialization. Furthermore, patent renewal is the only indicator revealing whether commercialization is successful. Forward citations have a moderate relationship with both commercialization and successful innovation, which may reflect the fact that citations are outside the control of the patentees or that forward citations measure spillovers rather than the private value. Although the correlations of the patent value indicators with technological innovation are noisy, this study provides stronger empirical support for the true relative value of different indicators with respect to innovation.

Appendices

Appendix A

The ordered probit model can be described as (Greene 1997):

$$y_{i}^{*} = X_{i} \alpha + \varepsilon_{i}$$

(A1)

where X_i is a vector of patent quality indicators and technology dummies; α is a vector of coefficients that indicates the influence of the independent variables on the profit level; and ε_i is a residual vector that represents the combined effects of unobserved random variables and random disturbances. The residuals are assumed to have a normal distribution, and the mean and variance are normalized to 0 and 1. The vector with the latent variable, y_i^*, is unobserved. The model is based on the cumulative normal distribution function, F(Xα), and is estimated via maximum likelihood procedures. The difference between this model and the two-response probit model is that in this model a parameter (threshold value), ω, is estimated by α. The probabilities P_i(y = k) for the three outcomes are:

$$\begin{gathered} P_{i} (0) = F( - X\alpha ), \hfill \\ P_{i} (1) = F(\omega - X\alpha ) - F( - X\alpha ), \hfill \\ P_{i} (2) = 1 - F(\overline{\omega } - X\alpha ), \hfill \\ {\text{where}}\;\sum\nolimits_{k = 0}^{2} {P_{i} (k) = 1} \hfill \\ \end{gathered}$$

(A2)

The threshold value, ω, must be larger than 0 for all probabilities to be positive.

To take account of selectivity, in the first step, a probit model estimates how different factors influence the decision to commercialize a patent (Greene 2002):

$$\begin{gathered} d_{i}^{*} = X_{i} \theta + u_{i} \hfill \\ d_{i} = 1\quad {\text{if}}\;d_{i}^{*} > 0\;{\text{and}}\;0\quad {\text{otherwise}} \hfill \\ \end{gathered}$$

(A3)

where d_i* is a latent index; d_i is the selection variable, indicating whether the patent is commercialized; X_i is a vector of explanatory variables that influence the probability that the patent is commercialized; θ is a vector of parameters to be estimated; and u_i is a vector of normally distributed residuals with zero mean and a variance equal to 1.

From the probit estimates, the selection variable d_i is then used to estimate a full-information maximum likelihood model of the ordered probit model (Greene 2002). In addition, the first step probit model is re-estimated. The residuals [ε, u] are assumed to have a bivariate standard normal distribution and correlation ρ. There is selectivity if ρ is not equal to zero. Note that this specification is not a two-step Heckman model. No lambda is computed and used in the second step.

Appendix B

See Table 12

Table 12 Descriptive statistics of dependent and explanatory variables