Has New Zealand benefited from its investments in research & development?

Abstract

We use panel data for nine industries to evaluate research and development (R&D) investments in New Zealand over the past forty years. We estimate the impact of R&D stocks in a particular industry on output per person in that industry and on output per person in the rest of the economy. We examine both public and private R&D investments. Privately provided R&D has a statistically significant positive impact on own-industry output per person, suggesting it increases productivity. However, publicly provided R&D has no impact on own-industry output per person. There is also evidence that private R&D in certain industries positively affects output per person in the rest of the economy, i.e. it generates positive spillovers. There is no evidence of positive spillovers from publicly provided R&D.

^a The views presented in this article do not necessarily reflect those of the Department of Labour and Motu Economic and Public Policy Research.

Acknowledgements

We are very thankful to Arthur Grimes, Hans-Jürgen Engelbrecht, Francisco Nadal De Simone, and David Mayes for comprehensive comments on an early version.

Notes

^a The views presented in this article do not necessarily reflect those of the Department of Labour and Motu Economic and Public Policy Research.

¹ Nonrival means that the marginal cost for an additional user is negligible.

² It is quite possible that there was an adjustment period that lasted several years after the reform, as growth rates have improved substantially from 1993 onwards.

³ We are aware of two papers only that examine R&D spillovers in New Zealand. Eveleens and Scobie (1986a, b) examined R&D spillovers in agriculture using different data and estimation method and found a positive relationship between R&D and agricultural productivity. Johnson (1999) used a data set similar to ours, but a different method to examine economy-wide R&D spillovers and found pretty results similar to ours.

⁴ Mansfield et al. (1977) is an example of case studies for manufacturing innovations.

⁵ Wieser (2004) provides a long list of references using this method. Also see Engelbrecht (1997), Engelbrecht (2002) and Coe and Helpman (1995) for OECD. Another approach is to estimate cost functions as systems of equations that include factor demand equations. For example see Mohnen et al. (1986), Bernstein and Ishaq (1991), and Mohnen and Lepine (1991). An additional approach is to regress R&D stocks or expenditures on indirect measures of total factor productivity. For example see, Bönte (2004), Engelbrecht (2002, 2003), Engelbrecht and McLellan (2001) and Johnson (1999).

⁶ Capacity utilization could be a proxy for capital utilisation as an additional regressor. However, the New Zealand Institute for Economic Research (NZIER) survey of capacity utilisation does not match the industry specifications that we have followed in this article, e.g. the industries are very different. For this reason we could not use their data.

⁷ Romer (1986) used physical capital to test for spillovers, Griliches (1979) used R&D and Lucas (1988) used human capital. Nelson and Phelps (1966) argue that the stock of human capital affects output growth (or per capita output growth) because it affects the adoption and the absorption of new technologies so one can have a product term. Engelbrecht (2002, 2003) tested this hypothesis for many countries and found some evidence for it.

⁸ The capital stock measure is subject to common criticisms that arise from guessing the depreciation rate and the initial stock, and the assumption of a constant rate of depreciation. According to Philpot (1994, 1995 and 1996), a base year was established in the past from known statistical collections of asset inventories and then added to each year by new investment, and subtracted from by estimated depreciation. In the gross stock definition, depreciation is based on the expected life of assets and is formally run down in the last eight years of each asset class. By and large, the rate of depreciation on capital allowed by the IRD is higher and hence the size of the net stock (as Philpott called it) is lower. These data are NZSIC-consistent (New Zealand Standard Industrial Classification) and based on existing capital formation data from Statistics New Zealand at that time. For the period since 1989–90, capital formation is taken from NZSIC-consistent capital formation data. Stocks were extrapolated using deflated capital formation series and the average implicit depreciation rate in the Philpott series for 1986–87, 1987–88, 1988–89 and 1989–90.

⁹ In international literature, part-time employment is usually defined as working 30 hours or less per week and given a weight of 0.50. Before 1986, the QES did not include agriculture so Philpott used employment from the agriculture Census.

¹⁰ Researchers are asked to tick the type of output class of their research, e.g., if research is about sheep the researcher would classify it as agricultural research and if it is about fish harvest then it is classified as fisheries. They are also asked whether they consider their research as private or government or university research. Although some of privately provided R&D projects are publicly funded the percentages are small. The Ministry of Research, Science and Technology Report (1997, p. 11) publishes the self-funding ratio for private R&D expenditures. We calculate the average over the period 1990 to 1997 to be 86.5%. The remaining money includes not only government funds, but also funds from universities, nonprofit organisations and from abroad.

¹¹ For example, the sub-groups for agriculture are sheep (meat), sheep (wool), sheep (general), beef production, dairy production, alternative animal species and generic animal research; for processing are meat, dairy and other processes; and for manufacturing are material and industrial processing, engineering, electronic and instruments, fibre, textile and skin processing, wood and paper processing. For classifications without subgroups, descriptions of related research are pretty vague. For example, energy research is defined as ‘information bases for prospecting, production and use of all energy sources’.

¹² The standard value of depreciation ranges from 0 to 10 (Griliches, 1995). For more details about these calculations see Johnson (1999). He also experimented with different rates of depreciation. He reported that as the rates of depreciation increase the estimated elasticities got smaller, but the rate of returns remained unchanged.

¹³ We tested the OLS regression residuals of Equation 4 ϵ_it for unit root. We used several different tests. Results are not reported, but they are available upon request. There is strong evidence that we can reject the hypothesis of unit root in the residuals ϵ_it . There is only one case out of 20, where the Levin et al. (2002) did not reject the hypothesis of a unit root in the residuals. The test seems to be very sensitive to the specification of the lag structure.

¹⁴ Dropping the subscripts for simplicity, the GMM estimator minimises with respect to β for a chosen pxp weighting matrix W, where and h is a Txp matrix of instruments that includes up to 8 lags of the RHS variables.

¹⁵ Mairesse and Hall (1996), Blundell and Bond (1999) and Arellano and Bover (1995) discuss this problem in more detail and introduce small sample adjustments to the weighting matrix to improve inference. As discussed subsequently, we implement one of these adjustments but it has little impact on our results.

¹⁶ We also report the Arellano–Bond (1991) GMM estimator. This is an estimator specifically designed for dynamic panel data with fixed effects and takes advantage of moment conditions beyond the first-order. This approach does not require the assumption of zero covariance across years or homogeneity across industries for efficiency. It involves transformation of instruments in first differences. However, it still suffers from the ‘weak instruments’ problem and, in theory, requires that there are more cross-sectional units than time series observations (i.e. N > T), which is not the case in our data. Therefore, the results of this estimator must be interpreted with care.

¹⁷ We also estimated the model in first-difference specifications and found that the results change only slightly.

¹⁸ We also used the Arellano–Bover (1995) orthogonalization method to adjust our Arellano–Bond GMM estimates for weak instrument bias, but this had no impact on the magnitude of α. We also tried using capital stock measured at the beginning of the period instead of the end, but this also had no impact on the results.

¹⁹ Griliches and Mairesse (1990) also found a similar evidence for the United States and even stronger evidence of diminishing returns for Japan. Their explanation was that the exclusion of raw materials and intermediate products from the production function might the reason for these results. Nickell et al. (1992) found similar results for the United Kingdom, which they attribute to measurement errors in k _it and l _it . We don’t have data for intermediate raw material by industry in New Zealand so we could not test the above hypothesis.

²⁰ Given the estimates of ρ, the long-run values of γ₁ are 0.30, 0.32, 0.64 and 0.41 for the four first regressions, which are quite large. Wang and Tsai (2003) reported large elasticities for high-tech firms in Taiwan. In the long run, the estimated returns can be as high as 24%.