WMOD00293 - Laserfiche WebLink

Home Browse Search

. . seeded units showing a treatment response, that unit would have triple the response ofa simple model, which had each seeded unit showing a response. Net treatment responses were taken from the most successful partitions found for the BRE. Experimental units were added until a 0.05 one-tailed P level was achieved, where P is the probability of incorrectly concluding that there is a positive seeding effect when none exists (type I error). Each simulation was repeated 1000 times to estimate the number of experimental units needed to reach a specified power level (I - /1), where f3 is the probability of a type II error-the probability of not detecting a treatment response when one exists. SUMMARY AND RECOMMENDATIONS The main purpose of this paper is to explore how variable responses to seeding might have affected the statistical power of the test of past experiments. Thfs was done by examining how the experimental duration might vary for winter orographic cloud seeding statistical designs that were common in the past, using different assumed responses to treatment. Most previous investigations assumed that each treated experimental unit would have the same percentage precipitation increase. Increasing evidence shows that this simple model is physically unrealistic. Seeding likely results in large percentage increases from a fraction of cases that are particularly amenable to treatment, but has little or not effect on the remaining cases. This result is likely true for other type of weather modification (rain augmentation, hail suppression), so the results in this paper may have a more general application beyond winter orographic cloud seeding. Monte Carlo techniques were first applied to simulated experiments using nonseeded 6-h data from the BRE. Seeding response models had a percentage of precipitation increase to all or a fraction of seeded experimental units. Only the one-tailed probability level ex = 0.05 was considered in all simulations. Following the results of the exploratory statistical analysis of S for the cold partition, a net precipitation increase of 66% was approximated in each experiment. These simulations were calculated with all available non seeded cases with main ridge temperatures of -9 oC and colder and a westerly wind component at 700 hPa. The number of units required to achieve statistical significance varied considerably with the different assumed treatment responses. For the 6-h BRE dataset, only a single winter of randomized seeding would be required to achieve a = 0.05 with a power of 0.9 if all treated experimental units responded with a 66% increase (model I). However, if only one-sixth of the units responded, though each with a 396% increase, almost six winters of experimentation would be requiredto achieve the same a and power levels. Similar results were obtained with the BRE and Idaho 24-h datasets. Simulations that limited seeding responses to the smallest or null accumulations indicated that there was little likelihood of deteqted with an effect with confidence. Simulations were also done in which a constant increase in precipitation amount was added to all or a fraction of seeded units partitioned by natural precipitation accumulation during the experimental period. These simulations also demonstrated a sensitivity to the character of simulated precipitation increases for all the datasets. These simulations reinforced the experimental design issues highlighted by the models I, II and III presentations; more experimental units are needed if only a fraction oftreated units respond to seeding. The simulations illustrate the importance of considering power as well as a in experimental design. Conducing a randomized seeding experiment without some physically reasonable estimate of power is very much a "crap shoot". An acceptable a level mayor may not result in the time allowed for the experiment (often determined by the sponsor's patience and resources). If the desired a level is achieved, 72