analysis in this paper requires only one observation per project or program, so multiple observations
had to be reduced to only one. Since there is not yet a solid theoretical basis for preferring one method
to the other, an average rate was calculated for each research project. All in all, the more than 1,800
observations in the IFPRI study, boiled down to only 201 useful rate-of-return observations, of which
78 pertain to developed countries (mainly the USA) and 123 to developing countries. It is assumed
that these 201 rate-of-return observations represent a reasonable sample of all the agricultural R&D
projects that have been undertaken worldwide. Although R&D projects with low or negative rates of
return are probably under-represented in the sample, this has fairly little effect on the two parameters
that we want to estimate: the implicit cutoff rate and the slope coefficient. Each observation is given
the same weight, which implies that each R&D project is assumed to have cost the same.
Figure 4 plots the distribution of ex post rates of return for both developed and developing
countries. As can be seen, the mode for developed countries (estimated at a rate of return of 20%)
stands out more clearly than the mode for developing countries (estimated at roughly 40%).
Apparently, the assumption that the selection of R&D projects took place under more or less the same
budget constraints (i.e., the same cutoff rate) and with more or less similar innovation opportunities
holds less well for the developing countries as a group. This is confirmed by the quite different
region-specific distributions as reported in tables 2, 3 and 4.
Table 2 provides more detailed and differentiated information about the characteristics of the
various rate-of-return distributions. The differentiation between developing regions is somewhat
speculative because of the rather small number of observations. Still, the distributions follow the
expected pattern quite well. In all cases, the distribution is lopsided to the left with a median lower
than the mean, and an estimated mode lower than the median. The share of observations lower than
the mode ranges between 20% and 40%. When the distribution is perfectly representative (which
unfortunately is not the case), this share would give an objective indication of selection performance.
The current sample includes a few very high rates of return (the highest is 855%), which in
some instances causes rather high standard deviations and means that are biased upwards. Although
one does not want to exclude very high rates of return a priori (although some healthy skepticism is
warranted), statistically one expects them to be rare. However, the rate-of-return sample may not only
13