Can licensing induce productivity? Exploring the IPR effect

Abstract

Licensing is one of the main channels for technology transfer from foreign-owned multinational enterprises (MNEs) to domestic plants. This transfer, occurring within and across industries, results in technology spillovers that may affect both intra- and interindustry productivity. Intellectual property rights (IPR) legislation may increase (or reduce) this effect. Using Chilean plant-level data for the 2001–2007 period and an exogenous variation from an IPR reform in 2005, we explore whether or not IPR affects the spillover effects of licensing on productivity. We find that stronger IPR reduces backward spillovers but increases forward spillovers. Moreover, the IPR legislation effect is stronger on firms that are on average smaller, and have low productivity. Our results are robust not only to a series of definitions of IPR, licensing, and productivity, but also to a set of different specifications.

Appendices

A. Appendix I: robustness

See Tables 5, 6, 7, 8, 9 and 10.

Table 5 Placebo regressions

Table 6 Spillover effects under different IPR measures panel IV in levels

Table 7 Spillover effects under different IPR measures (panel IV and first differences, with lagged control variables)

Table 8 Robustness checks for specification issues

Table 9 Different IV following (Lu et al. 2017)

Table 10 Different IV following (Javorcik et al. 2018)

B. Appendix II: TFP estimation methods

We follow closely the excellent survey of Van Beveren (2012) regarding TFP estimation which departs from an standard Cobb–Douglas production function with capital, unskilled labour and skilled labour: $Y_{_{ijrt}}=A_{_{ijrt}}K^{^{\beta _{k}}}_{_{ijrt}}L^{s^{\beta _{ls}}}_{\ _{ijrt}}L^{u^{\beta _{lu}}}_{\ _{ijrt}}$; where $Y_{ijrt}$ is output of the firm i in sector j and region r at time t; $K_{ijrt}$ is the capital stock; while $L^s$ and $L^u$ are the number of skilled and unskilled workers, respectively. If we apply logarithms our production function becomes: $TFP_{ijrt}=y_{ijrt}-\beta _kk_{ijrt} -\beta _{ls}l^s_{ijrt}-\beta _{lu}l^u_{ijrt}$.

If we were to estimate this equation using OLS, the main assumption would be that all inputs in the production function are exogenous. However, inputs in the production function are not chosen independently, creating an endogeneity problem. Moreover, selection bias is introduced when, in an unbalanced panel of firms, the decision on allocation of inputs (especially variable inputs) is not independent of the decision to continue operating in the market. Therefore, the estimation technique has to take into account that the estimates are conditional on the survival of the firm. To overcome endogeneity and selection bias, Olley and Pakes (1996) assumes that capital is a fixed input that depends on an investment process. Therefore, capital in period t depends on capital in $t-1$ and investment $i_{t-1}$. This timing corrects the endogeneity issue concerning the capital since the decision of capital is taken before the knowledge of the productivity shock. To estimate this, Olley and Pakes (1996) propose a three-step estimation process where the investment of the firm is a monotonically increasing function of productivity and existing capital. If the relationship is monotonically increasing, as explained in Arnold (2005), the investment decision is a function of the productivity and the capital stock.²⁶ Thus for the first step, the estimating equation becomes: $y_{it}=\beta _0+\beta _{l}l_{it} +\phi (i_{it}, k_{it})+\eta _{it}$.

Since the functional form of $\phi (.)$ is unknown, it can be proxied non-parametrically by a third or fourth-degree polynomial. Thus, in the first stage, both $\widehat{\beta _l}$ and $\widehat{\phi }$ can be estimated using regular Ordinary Least Squares (OLS). In the second step, they introduce a correction for the selection bias (exit decision), and in the third step, they recover consistent estimates of $\beta _l$ and $\beta _k$. However, the main limitation of this methodology is that there could be a large number of “zero” investment observations (not all firms invest every single period). Thus a considerable amount of information is potentially lost. Since there is a monotonic relation between investment and productivity, all the observations with zero investment must be dropped from the sample, which could imply a significant loss of observations in the dataset (depending on how many firms do not invest). Moreover, since investment is not always positive, Levinsohn and Petrin (2003) proposes using intermediate inputs as a proxy for productivity shocks. As Arnold (2005) points out, typically, there are significantly less zero- observations in intermediate inputs than in investment.

More recently, Ackerberg et al. (2015) have re-examined the estimation methods for production functions. They shed some light on some issues that might hinder the methodology proposed by Olley and Pakes (1996) and Levinsohn and Petrin (2003). They argue that there may be significant problems with the estimation of production functions if the methods mentioned above are used. The most critical issue is collinearity problems that arise in both methods.²⁷ The estimation has to be performed using either non-linear least squares or an optimization routine. The OLS and the Olley and Pakes distributions are very asymmetric, while the Levinsohn and Petrin, and the ACF estimators behave more smoothly. Levinsohn and Petrin estimation overestimates TFP compared to the ACF estimator.

Using the Chilean Encuesta Nacional Industrial Anual (ENIA), which is a national representative survey of the manufacturing sector for 2001 – 2007, we follow (Ackerberg et al. 2015) employing the number of skilled and unskilled workers within each establishment to identify $\beta _{ls}$ and $\beta _{lu}$ from the production function and electricity consumption as a proxy for investment to identify the function $\phi (\cdot )$ and thus be able to recover $\beta _k$.

Table 11 Summary statistics of different TFP estimators

Fig. 4

Comparison of TFP densities from different methods

Table 11 shows summary statistics of different TFP estimators, while Fig. 4 shows the distribution for those estimators. Moreover, Table 13 illustrates the coefficients for the different methods.

In order to check the robustness of the main results of the paper, estimates of Table 2 are depicted in Table 12. As it is shown, the results are qualitatively similar to the ones presented in the main section of the paper (Table 13).

Table 12 Main specification under different TFP methodologies

Table 13 TFP estimation by different methodologies