6  Results

In this section we present the results of estimating the relationship between R&D and productivity for the agricultural sector. We take each of the three formulations in turn. In addition we present the results of using a simplified model following Scobie and Eveleens (1987).

6.1  Using stocks of knowledge

The results from running the Philips-Loretan model using R&D stocks generated by the PIM are summarised in Table 5.[27] We ran a number of variants of the basic model including and excluding human capital, and with and without the dummy variable to represent structural shifts. In no case was the variable representing human capital significant. In both cases where the structural shifter for post 1984 was included the coefficient was highly significant. When human capital is removed from the model, the coefficient on the domestic R&D stock becomes significant at the 1% level.

The facts that 1) both human capital and domestic R&D are insignificant (or barely so) when both are included in the regression; and 2) each becomes individually significant when the other is omitted, point to the presence of multicollinearity. In fact, the simple bivariate correlation between these two variables is 0.99. Further evidence that multicollinearity is a problem is given by running the same regression in different (arbitrary) time periods. If two explanatory variables are correlated, a different sample will likely produce opposite results. We ran the regression with both variables included (and the dummy excluded) for the period 1927 to 1964 and the same regression from 1964 to 2001, and found that, while the coefficient on human capital was positive in both samples (although insignificant in the earlier period and significant in the latter period), the coefficient on the domestic R&D stock went from being negative and significant in the first period to being positive and significant in the period 1964-2001.[28] Such correlations mean that the corresponding regression coefficients cannot be interpreted because it is impossible to fix or control one variable while changing the other in the presence of this high correlation.

One solution to the problem is to remove those variables which are highly correlated with others (in this case we removed the human capital index) and therefore redundant. However, the drawback of this approach is that no information is obtained on the deleted variable while the importance of those in the equation may be overstated. Hence the significant coefficient on the domestic R&D stock in models 3 and 4 may be picking up some of the contribution of the omitted variable for human capital as well as the R&D effect itself.

If we ignore this and conclude that the elasticity of MFP with respect to domestic R&D is 0.148 (model 3 which is the preferred specification), then this implies a rate of return of 16.7% to the domestic R&D stock (in the long-run), assuming an R&D intensity (defined as the R&D stock divided by GDP) equal to the average over our sample period.[29] This return is lower than that estimated when we include human capital in the equation (model 1), indicating that the estimated coefficient on domestic R&D in Model 3 (our preferred specification) is not picking up the effect of human capital (the omitted variable) as well as the effect of domestic R&D.

The coefficient on cumulated patents (our proxy for the foreign stock of knowledge) is highly significant and positive in all four models. This indicates that foreign spill-ins to the agricultural sector are an important source of new knowledge and they are associated with the productivity performance of this sector. The estimated elasticity of MFP with respect to the foreign spill-in stock ranges from 0.25 to 0.35. That is, for a 10% increase in the number of patents granted in the US, MFP in the agricultural sector of New Zealand would increase from between 2.5% to 3.5%.

The coefficients on both weather and extension variables are never significant; except for extension in Model 4. In the case of extension this result is somewhat surprising, as with more extension workers we would expect new knowledge to be disseminated to users faster and therefore for more extension workers to have a positive impact on MFP.

Note: The asterisks indicate the degree of significance of the estimated coefficient; *** = 1%; ** = 5%; * = 10% and an absence of asterisk indicates the coefficient was only significant at more than 10%.