Laserfiche WebLink
<br />5. The reflectivity of the prospective cell must be greater on the second radar scan than <br />on the first, <br /> <br />6. The prospective control cells must reside in the 60-kilometer-wide annulus that is <br />centered on the mean range of the treated cells for which the controls are being <br />selected. <br /> <br />The last criterion is best understood by reference to figure 2.3, in which the treated cells, the <br />environmental cells, and the control cells are plotted. Note that the three treated cells (either <br />S or NS) are defined to have a 25-kilometer region of effect around them, and that the <br />environmental cells are defined as those cells which did not receive either S or NS treatmEmt <br />that live in this region of effect. The cells which are to be used as controls are depicted <br />schematically in the 60-kilometer-wide annulus. The center of this annulus is at the mean <br />range of the treated cells from the radar. To be consistent with criterion No.3, the annular <br />region in which the control cells can be selected must end at least 35 kilometers from the <br />treated cells. In essence, therefore, at least a lO-kilometer buffer exists between the area of <br />effect and the region that contains the control cells. <br /> <br />Before the control cells can be used for evaluation of the Thai cell experiments, the control <br />values corresponding to each of the experimental units must be weighted as a function of the <br />number of treated cells in that unit. If this weighting is not done, the overall control value <br />for the Sand NS units will be dominated by the unit that has the most control cells. Thus, <br />the control cells have the same weight and influence as the treated cells (either S or NS) that <br />they are meant to represent. <br /> <br />This approach worked quite well initially in Texas (Rosenfeld and Woodley, 1989), where the <br />control cells were positively correlated on a unit-by-unit basis with the corresponding <br />properties ofthe cells that were randomly selected for treatment but did not receive AgI (i.e., <br />the NS cells). The correlation was found to be strongest for Rvol and Amax and weakest for <br />Zmax' The Rvol correlations indicate that between about 30 and 60 percent of the rainfall <br />variability of the NS cells in Texas can be accounted for by the control cells. In cloud seeding <br />studies, this level of performance is rather good for one control variable. This approach is <br />expected to work as well in the AARRP if it proves possible to identify many control cells lor <br />each experimental unit. <br /> <br />2.11.3.2 Linear modeling of the unit rainfall <br /> <br />A number of additional ways exist to address the natural rainfall variability in a randomized <br />cloud seeding experiment. One possibility is to follow the example of Flueck et al. (1986) and <br />use a guided exploratory approach to linearly modeling the unit rainfall in the Thai program. <br />This process relies upon an initial crude conceptual model of Thai rainfall and repeated <br />interaction between the scientist and the computer-generated summary statistics to <br />systematically build a predictive statistical linear model of the unit rainfall and the potential <br />treatment effects. <br /> <br />The linear modeling will attempt to account for the effects of natural atmospheric processes <br />on the rainfall response variable (y), as well as the effects of the treatment. In statistical <br />terms, a statistical linear predictive model will be defined and fit to the unit area rainfall: <br /> <br />* <br />y = bo + b1xl + bZX2 + .000 + bkxk <br /> <br />18 <br />