WSP12351 - Laserfiche WebLink

Home Browse Search

, 1. tv ~... o ~ Indicators of salt pickup: 1, < Irrigated area (annual value), 2. Consumptive use in the preceding month. 3. Consumptive use in the second preceding month. 4. Consumptive use in the third preceding month. 5, Consumptive use in the fourth preceding month. For downstream sites (table 1), these variables were summed for all tributary upstream reaches. The models then were calibrated using variables that indicated the total water-resources development upstream from the site, Selection of independent variables for individual sites was made using stepwise-regression procedures. Four separate methods were used: (1) Back- ward elimination, (2) forward selection, (3) stepwise selection, and (4) maximum R2 improvement (SAS Institute Inc., 1985). Also, the diversion indicators were divided into three groups to evaluate the effects of col- linearity among the diversion indicators, One group included only total adjustments; the second included only the components of total adjustments-- consumptive use, net other diversions, and net release; and the third included all four of the diversion indicators. This combination of 4 methods and 3 groups of variables produced 12 separate selection options for a particular model calibration, Several secondary criteria were applied to distinguish among options and determine the overall best variable set. The first of these secondary cri- teria was to minimize the mean-square error. The second criterion was that the Mallows' Cp statistic approximately equal the number of parameters in the model. This minimizes the bias in the standard error caused by an over- parameterized model (Montgomery and Peck, 1982). The third criterion was that the coefficients for the development variables have the expected sign. The fourth criterion was that the model contain no strongly correlated variables or combinations of variables. This criterion was intended to decrease the collinearity among the independent variables. As a check for possible col- linearity, the variance inflation factors (VIF's) were evaluated for the selected variable sets, The VIF for a coefficient is a measure of the increase in the variance of that coefficient caused by correlation among the independent variables in the model. If VIF's are large, the associated coefficients may be poorly estimated (Montgomery and Peck, 1982). In practice, the variable sets selected by the four stepwise regression procedures usually were identical; therefore, the choice was only among the three variable groups. Normally, one of these groups was clearly superior in regard to all the secondary selection criteria, 15