WMOD00293 - Laserfiche WebLink

Home Browse Search

. . ~' -. the experimenters may have been very fortunate in the specific population of units that nature provided. But a replication of the experiment may take several times as long to complete. If the P value does not reach an acceptable a level, it may simply be that insufficient time was allowed to have a reasonable chance of achieving a significant result. That is, seeding may have been effective but was undetected because the number of experimental units was too limited (type II statistical error). The only correct conclusion is that seeding effectiveness was neither proved nor disproved. When a serious effort is made to consider both a and power in the experimental design, the treatment response model should be carefully selected. The particular model chosen has been shown to have a large -impact on estimated experimental duration. Unfortunately, present knowledge of treatment response is limited. However, model I, the most commonly used model, is likely unrealistic. Its use results in overly optimistic estimates of experimental duration. This presents the experimenter with a dilemma. It is imprudent to design a statistical cloud seeding experiment without reasonable a and power levels. However, accurate power-level estimates may be beyond the scope of present physical understanding. The best course for future investigations would therefore be to improve physical understanding before designing more statistical experiments. The implications of these results are sobering. They suggest that previous statistical experiments may have been handicapped by the assumption of a simplistic treatment response model,. which led to overestimates of experimental power. The number of experimental units required to achieve given a and power levels may be far larger than estimated for most past experiments, which could partially explain the frequent finding of inconclusive results. A number of points must be considered in designing future statistical experiments. First, strong predictor- covariate relationships with target area precipitation are necessary to reduce the number of experimental units needed to detect a seeding signal. Second, partitioning based on a solid physical understanding is needed to reduce the number of experimental units that have minimal or no response to seeding. Third, improved statistical techniques could be useful, but their development may be difficult in view of. . uncertainties of how treatment responses vary among the population of treated units. Fourth, and most important, a much improved physical understanding is needed prior to the development of any future statistical design. Over a decade ago, Braham (1981) made a strong case for improving our physical understanding before conducting any further "black box" experiments that had a major emphasis on demonstrating precipitation changes at the ground. Schaefer (1990) also argued for a better physical comprehension of the effects of cloud seeding. The results of this paper strengthen the case for improved physical understanding prior to further statistical testing of cloud seeding. 8.22. Huggins, A. W.,1996: Use of radiometry in orographic cloud studies and the evaluation of ground-based cloud seeding plumes. 13th Conference on Planned and Inadvertent Weather Modification, Atlanta, GA, American Meteorological Society, 142-149. , . INTRODUCTION AND BACKGROUND Since 1981 a program to evaluate the c~oud seeding operations of the state of Utah has been conducted. in southwestern and central Utah. The research is part of the Atmospheric Modification Program (AMP), a cooperative venture between six states, the Navajo Nation and the National Oceanic and Atmospheric Administration (NOAA). The Utah project is concerned with the modification of winter orographic cloud 73