A fourth check on the robustness of the results was performed by the inclusion of interaction terms with the R&D capital variable as well as with the productivity gap. None of the former interactions has any significant effect. Fifth, although the estimates of the productivity gap use the industry-specific Purchasing Power Parities of Kuroda (1996), the results are not sensitive to the use of the Unit Value Ratios of van Ark (1996). Lastly, within the dynamic panel data framework used here, the estimated coefficients may be subject to a finite-sample bias (see Baltagi, 1995, chapter 3). This bias disappears asymptotically at the number of time periods rises, but not as the number of units of observation rises. In the present context, with 27 annual observations, the bias is likely to be small.