Laserfiche WebLink
found in examining subdivisions, e.g. monthly values, of the data; and (2) the problem of <br />multiplicity in the analyses may lead to some cases having apparently small P values that <br />cannot be accorded the same level of importance as the same P value would have in the <br />primary analysis. <br />Table 3 summarizes the key results of these exploratory analyses. In brief, none of <br />the MRPP P-values for the various seasonal comparisons is small enough to be <br />considered significant. Among the P values for the month-by-month comparisons, ones <br />for June involving the target area approach the level of 0.10, but in view of the <br />aforementioned multiplicity concerns this cannot be considered statistically significant. <br />Interestingly, the smallest P value (0.069) occurs for the August data in the north control <br />versus south control comparison – suggesting a possible “seeding effect” in the <br />comparison between these two non-seeded areas. All this, of course, merely reflects some <br />combination of the multiplicity factor and the likelihood that the natural variations in the <br />rainfall in the region overwhelm any effect of the NDCMP seeding upon the rainfall as <br />measured by the climatic gage network. <br />Using Figure 2 and monthly analogs thereof, it is possible to derive point <br />estimates and confidence intervals for any potential seeding effect. The procedure, <br />summarized in the appendix of Smith et al. (1997), uses separate LAD regression lines <br />for the historical and NDCMP years – though in the present instance the large P values do <br />not justify separate lines. The ratio of the slopes of those lines provides the point <br />estimate. Table 3 includes these point estimates, all of which are quite close to 1.0 – a <br />value that would indicate no effect. Confidence intervals shown in the table are obtained <br />with random permutations of the assignment of years to the two groups, followed by <br />recalculation of the LAD regression lines, and indicate the range of the ratio resulting <br />from such permutations. For the seasonal values, the confidence interval is roughly 1.0 <br />0.1, while the monthly intervals bracket 1.0 with a somewhat wider range. <br />Table 3. Key Results of Exploratory Statistical Analysis <br />Months:SummerJuneJulyAug <br />(June-Aug) <br />MRPP P values <br /> Target vs. Control 0.322 0.162 0.828 0.960 <br /> Target vs. N. Control 0.451 0.116 0.793 0.591 <br /> Target vs. S. Control 0.706 0.103 0.695 0.385 <br /> N. Control vs S. Control 0.802 0.626 0.879 0.069 <br />Point estimate of seeding effect1.0080.9500.9971.020 <br />90% confidence interval 0.91-1.10 0.82-1.10 0.86-1.17 0.88-1.23 <br />9 <br />