Evaluation of various soil erosion models with large data sets have co
nsistently shown that these models tend to over-predict soil erosion f
or small measured values, and under-predict soil erosion for larger me
asured values. This trend appears to be consistent regardless of wheth
er the soil erosion value of interest is for individual storms, annual
totals, or average annual soil losses, and regardless of whether the
model is empirical or physically based. The hypothesis presented herei
n is that this phenomenon is not necessarily associated with bias in m
odel predictions as a function of treatment, but rather with limitatio
ns in representing the random component of the measured data within tr
eatments (i.e., between replicates) with a deterministic model. A simp
le example is presented, showing how even a 'perfect' deterministic so
il erosion model exhibits bias relative to small and large measured er
osion rates. The concept is further tested and verified on a set of 30
07 measured soil erosion data pairs from storms on natural rainfall an
d run-off plots using the best possible, unbiased, real-world model, i
.e., the physical model represented by replicated plots. The results o
f this study indicate that the commonly observed bias, in erosion pred
iction models relative to over-prediction of small and under-predictio
n of large measured erosion rates on individual data points, is normal
and expected if the model is accurately predicting erosion rates as a
function of environmental conditions, i.e., treatments. (C) 1998 Else
vier Science B.V.