Laserfiche WebLink
<br />o [hll'j 3 <br /> <br />Predictor variables will also be important in the demonstration pro- <br />jects. In the context of CREST, they consist of any measurable parame- <br />ters that are sufficiently associated with target area precipitation to <br />provide a reasonably good prediction of that precipitation. While <br />information such as that available from weather maps or individual <br />rawinsondes may be useful, it has been found that precipitation obser- <br />vations near the target area (control) provide the best prediction of <br />actual target area precipitation. The predictor variable precipitation <br />observations need to be located sufficiently upwind or cross wind of <br />the target area to be unaffected by seeding. Use of predictor vari- <br />ables is particularly advantageous because several analyses of weather <br />modification projects have suggested that Type I statistical errors <br />often occur. That is, although the projects were randomized, nature <br />did not provide totally similar subpopulations of seeded and nonseeded <br />units, even with rather large total populations. In such cases, <br />statistical analysis could indicate significant differences between <br />seeded and nonseeded subpopulations. These differences may incorrectly <br />be attributed to the effects of seeding when in fact such indicated <br />differences are really due to the "luck of the draw." Clearly it is <br />desirable to guard against Type I statistical errors. The use of <br />predictor variables allows this to be done and also provides for more <br />sensitive statistical testing because some of the variance among <br />treatment units is explained. In general, the closer the correlation <br />between the target area precipitation and the predictor variables, the <br />more sensitive the statistical analysis and the fewer treatment units <br />required to indicate a given level of statistical significance. <br /> <br />Nonparametric statistical tests will be applied to CREST precipitation <br />observations from seeded and nonseeded treatment units. These will <br />consist of the well-known Wilcoxon test, rerandomization procedures, or <br />the more recently developed MRPP (multiresponse permutation procedures). <br />These tests will be applied to ranked residual data resulting from <br />application of a median regression line to the target and control <br />(predictor variable) precipitation data. <br /> <br />Although the primary evaluation of increased streamflow will be based <br />on conversions of statistically determined increases in precipitation <br />through precipitation-runoff relationships, a secondary evaluation will <br />also be made directly using seasonal streamflow data. This streamflow <br />analysis will be made using a target-control approach comparing the <br />streamflow in the target subbasins with historically highly correlated <br />streamflows in nearby subbasins not affected by seeding. Because of <br />the diluting effect of the randomization procedures, the increased <br />seasonal streamflow will not be very great and, therefore, cannot <br />be discerned at the same level of statistical significance as the <br />increased precipitation. A statistically significant streamflow <br />evaluation is possible if a yearly randomization period was used, but <br />this was ruled out as the primary evaluation because the demonstration <br />program would have to be conducted for at least 10 years to reach a <br />comparable level of significance. Nevertheless, the secondary target- <br />control evaluation of streamflow will complement and reinforce the <br />primary evaluation. In addition, the evaluation of the chain of <br /> <br />32 <br />