Trade, imitative ability and intellectual property rights

Abstract

Economic theory suggests some ambiguity concerning the effects of strengthening intellectual property rights (IPRs) on international trade. Here we extend the empirical literature that attempts to resolve this ambiguity. We use panel data to estimate a gravity equation for manufacturing exports, in aggregate and by industry, from five advanced countries to 69 developed and developing countries over the period 1970–1999. In particular, we use threshold regression techniques to determine whether the impact of IPR protection on trade depends upon the level of development, imitative ability and market size of the importing country. We confirm the importance of the importers’ imitative ability, and also find some evidence of a role for market size in this relationship. The individual industries present different patterns of thresholds and coefficients, with Total Manufacturing closely reflecting that of Fabricated Metal Products.

Appendices

Appendix 1: threshold regressions

Threshold regression addresses the issue of whether regression functions are identical across observations or whether there exists evidence of non-linearity with observations split into discrete classes. Hansen (1996, 2000) developed a method of estimating such thresholds and testing for their significance as well as constructing confidence intervals. Hansen (1999) extended this approach to fixed effects panel data. In this appendix we briefly describe the methods used in this paper.

The method can be described using the following two variable panel regression model

$$ y_{it} = \mu_{i} + \delta_{1} x_{it} I(q_{it} \le \lambda ) + \delta_{2} x_{it} I(q_{it} > \lambda ) + \varepsilon_{it} $$

(A1)

where I(·) is the indicator function and q _it is the threshold variable.²³ Here the observations are divided into two regimes depending upon whether the threshold variable, q _it , is smaller or larger than the threshold, λ. The coefficient on x _it is given by δ ₁ for observations with q _it less than or equal to λ and by δ ₂ for q _it greater than λ. Chan (1993) and Hansen (1999, 2000) recommend estimation of λ by least squares, which involves finding the value of λ that minimises the concentrated sum of squared errors, S ₁ , that is $\hat{\lambda } = \arg \min_{\lambda } S_{1} (\lambda )$. In practice this involves searching over distinct values of q _it for the value of λ at which the concentrated sum of squares is smallest. Hansen (1999) notes that it is undesirable to have two few observations in a particular regime, a possibility that can be excluded by constraining the possible values of λ to those for which a minimum percentage of observations are in each regime. In our analysis we impose the restriction that at least 20% of observations must lie in each regime.

Hansen (1999) suggests the following steps to estimate the threshold value: first, sort the distinct values of the observations on the threshold variable, q _it . Second, eliminate the smallest and largest η percent. For each of the remaining values of q _it estimate Eq. A1 and save the sum of squared residuals. Third, choose the estimate of $\hat{\lambda }$ as the value of q _it with the minimum sum of squared errors.

Having found a threshold it is important to determine whether it is statistically significant or not, that is, it is necessary to test the null hypothesis H₀: δ ₁ = δ ₂. Given that the threshold is not identified under the null hypothesis this test has a non-standard distribution and critical values cannot be read off standard distribution tables. Hansen (1999) proposes a bootstrap procedure to simulate the asymptotic distribution of the likelihood ratio test. This involves the following steps: first, estimate the linear model where no threshold is assumed and save the sum of squared residuals, S ₀ . Second, calculate the likelihood ratio test of the null hypothesis, H ₀, given by $F_{1} - [S_{0} - S_{1} (\lambda )]/\hat{\sigma }^{2} ,$ where $\hat{\sigma }^{2} = \left\{ {{1 \mathord{\left/ {\vphantom {1 {n[t - 1}}} \right. \kern-\nulldelimiterspace} {n[t - 1}}]} \right\}S_{1} (\hat{\lambda }),$ with n the number of cross-sectional units and t the number of time periods. Third, construct a bootstrapped sample under the null hypothesis by drawing from the normal distribution of the residuals from the linear model.²⁴ Fourth, using the bootstrap sample estimate the model under the null (linearity) and the alternative (threshold at $\hat{\lambda }$) and calculate the bootstrap likelihood ratio statistic, F ₁. Finally, repeat this procedure a large number of times (in our case 1,000) and calculate the percentage of draws for which the simulated statistic exceeds the actual one. This is the bootstrap estimate of the p-value for F ₁ under H ₀.

In the dual-threshold model we have

$$ y_{it} = \mu_{i} + \delta_{1} x_{it} I(q_{it} \le \lambda_{1} ) + \delta_{2} x_{it} I(\lambda_{1} < q_{it} \le \lambda_{2} ) + \delta_{3} x_{it} I(q_{it} > \lambda_{2} ) + \varepsilon_{it} $$

(A2)

where the two thresholds are ordered such that λ₁ < λ₂. It is a straightforward extension to search for the values of λ₁ and λ₂ that minimise the sum of squared errors. At the same time it can be expensive in terms of computation time to search for both thresholds simultaneously. Chong (1994), Bai (1997) and Bai and Perron (1998) have shown however that sequential estimation is consistent. This involves fixing the first threshold at $\hat{\lambda }_{1}$ and searching for a second threshold assuming that the first is fixed. The search takes place for values of q _it both above and below the first threshold, though the method can be adapted to ensure that a minimum of η per cent of observations are in each of the three regimes. It can be shown that the estimate of the second threshold, $\hat{\lambda }_{2}$, is asymptotically efficient using this method. This is not the case for $\hat{\lambda }_{1}$ however, because it was estimated from a sum of squared errors function that was contaminated by the presence of a neglected regime. To overcome this problem Bai (1997) recommends a refinement estimator for $\hat{\lambda }_{1}$ that involves fixing the second threshold at $\hat{\lambda }_{2}$ and searching for the first threshold again, now including the second threshold. This approach can be extended to consider the possibility of more than two thresholds.

In Sect. 4.3 we investigate the possibility of thresholds on more than one variable. Specifically, we examine interactions between the level of imitative ability and both the level of IPRs and market size. To allow such a possibility we adopt an approach similar to that described in the previous paragraph. We begin by fixing the threshold on our measure of imitative ability. In particular, we fix the threshold at the highest significant threshold on imitative ability and then search for a threshold on the other variable in the high imitative ability regime (i.e. for observations for which the measure of imitative ability is above the estimated threshold). Finally we consider the possibility of a threshold on these variables in the low imitative ability regime. Where the second threshold, that is the threshold in the high imitative ability regime, is significant we include it when searching for a third threshold. But then the estimated thresholds in the high imitative ability regime are not asymptotically efficient, since they were estimated from a sum of squared errors function contaminated by the presence of a neglected regime. To deal with this we follow Bai (1997) and re-estimate the high imitative ability threshold now including the estimated threshold in the low imitative ability regime.

Appendix 2: data

Our data is averaged over six 5-year periods, 1970–1974, 1975–1979, 1980–1984, 1985–1989, 1990–1994 and 1995–1999. Due to missing data for various variables the maximum number of observations is 2021. The data for population, GDP and GDP per capita came from the World Bank’s World Development Indicators (2001) database. Data on distance, common language and borders and landlockedness came from a website maintained by Jon Haveman. Trade data came from the OECD’s International Trade by Commodity Statistic (Historical Series, 1961–1990) and International Trade by Commodity Statistic (1990–1999). The trade data from 1961 to 1990 was in SITC rev. 2 and was converted to ISIC rev. 2 using a concordance supplied by the OECD. The data for the period 1990–1999 was in SITC rev. 3 and was converted to SITC rev. 2 and then ISIC rev. 2 again using a concordance supplied by the OECD. The education data was taken from the Barro and Lee (2001) database. The index of IPR protection is provided in Ginarte and Park (1997) and is the most commonly used indicator of IPR protection. This index was constructed for 110 countries quinquennially for the period 1960–1990. Five characteristics of patent laws are included: extent of coverage; membership in international patent agreements; provisions for loss of protection; enforcement mechanisms and duration of protection. Each was assigned a value ranging from zero to one and their unweighted sums formed the index, with a higher number signalling stronger IPR protection. This data has been updated to 1995 by Park who kindly supplied us with the full set of data. Table 10 provides summary statistics for all the variables.

Table 10 Summary statistics