Many studies have compared load estimation methods
[ Bennett and Sabol, 1973;
Johnson, 1979;
Verhoff et al., 1980;
Walling et al., 1981;
Dickinson, 1981;
Dolan et al., 1981;
Fenn et al., 1985;
Dann et al., 1986;
Somlyody, 1986;
Ferguson, 1987;
Richards et al., 1987;
Walling et al., 1988;
Preston et al., 1989;
Clarke, 1990a;
Crawford, 1991;
Cohn et al., 1992a].
Recent studies using both Monte Carlo
and bootstrap methods
[ Preston et al., 1989;
Crawford, 1991] have found
that rating curve methods, when unbiasing factors are used,
provide nearly-unbiased, low variance estimates of nutrient loads from large
basins (
100 square miles).
Preston et al. [1989]
considered estimation of total phosphorus loads on the Grand River (Michigan),
and
Crawford [1991]
considered suspended sediment loads
on the Big Blue and the Wabash Rivers (Indiana).
Crawford also found that a
bias-corrected, transformed-linear model produced more accurate estimates
than those obtained from a non-linear model.
Cohn et al. [1992a]
evaluated the performance of rating curve
estimators for six species of nutrients at four tributaries to the Chesapeake
Bay.
They found that there was always a statistically-significant
lack of fit between the rating curve model and the data, but split-sample
studies did not reveal a significant bias in load estimates when
was employed.
The rating curve, once retransformation bias was removed, worked well for
estimating nutrient loads from large basins.
Stratified ratio estimators, when properly implemented, have been found to be about as precise as rating curve estimators. Preston et al. [1989], Richards et al. [1987] and Dolan et al. [1981] found that stratified ratio estimators can be both robust and of low variance for estimating nutrient loadings from mid-sized watersheds (100-2000 square miles) to the Great Lakes.
However, in a study of small, mountainous drainages, Thomas [1985] found that all rating curve estimators were substantially biased, the smearing estimator being the least biased (17% upward bias). As expected, stratified random sampling showed no significant bias.