Evaluating the environmental and economic impacts oi agricultural poli
cies is not a simple task. A systematic approach to evaluation would i
nclude the effect of policy-dependent factors (such as tillage practic
es, crop rotations, and chemical use) as well as the effect of policy-
independent covariates (such as weather, topography, and soil attribut
es) on response variables (such as amount of soil eroded or chemical l
eached into the groundwater). For comparison purposes, the effects of
these input combinations on the response variable would have to be ass
essed under competing policy scenarios. Because the number of input co
mbinations is high in most problems, and because policies to be evalua
ted are often not in use at the time of the study, practitioners have
resorted to simulation experiments to generate data. However, generati
ng data from simulation models is often costly and time consuming; thu
s, the number of input combinations in a study may be limiting even in
simulation experiments. In this paper, we discuss the problem of desi
gning computer simulation experiments that require generating data for
just a fraction of the possible input combinations. We propose an app
roach that is based on subsampling the 1992 National Resources Invento
ry (NRI) points. We illustrate the procedure by assessing soil erosion
in a situation where there are ''observed'' data [reported by the Nat
ural Resources Conservation Service (NRCS)] for comparison. Estimates
for soil erosion obtained using the procedure we propose are in good a
greement with NRCS reported values.