WAHLEW: INCIDENTAL UUL^HIIN MUKiALlTK 

 40°N 



20 c - 



20 e 



40"S 



160»W 



140* 



120 c 



100' 



80« 



Figure 1.— The three areas of the eastern tropical Pacific used to stratify the data, bounded 

 by lat. 40°N, long. 160°W, lat. 40°S, and the western coastline of the North and South 

 American continents. 



portions of known dolphin sets which were cal- 

 culated from logbook data. 



I tested for significant between-strata differences 

 in mean total kill using analysis of variance (ANOVA) 

 methods of BMDP programs P7D and P2V (Dixon 

 1983). Violation of the ANOVA assumption of equal 

 cell variances may seriously distort significance 

 probabilities in unbalanced models such as in this 

 study (Glass et al. 1972). Because such distortion 

 could be great, test results were considered to be 

 inconclusive when significance probabilities were 

 close to 0.05. 



I was unable to test for the combined effect of all 

 four stratification factors using the whole data set 

 because data were sparse or unavailable in many of 

 the 144 cells of the proposed four-factor stratifica- 

 tion. Thus, ANOVA results for restricted subsets of 

 the data containing adequate sample sizes were 

 assumed to hold for subsets with inadequate sam- 

 ple sizes. To eliminate significant between-factor 

 interactions, I tried logarithmic and power trans- 

 formations of the dependent variable, total number 

 of dolphins killed. When these transformations failed 

 to eliminate the interactions, I partitioned the 



analysis into individual levels of the interacting 

 variables. 



When it was necessary to determine where the 

 within-factor differences occurred, £-tests for differ- 

 ences between all pairs of cell means were made. 

 Since I tested for differences between all pairs 

 rather than between a few preselected pairs, a dif- 

 ference was considered significant if its significance 

 probability was less than the quotient of 0.05 and 

 the number of pairs. This Bonferroni adjustment to 

 the significance level of each test assured a level of 

 0.05 simultaneously across all tests (Snedecor and 

 Cochran 1980). 



I computed lvalues for pairwise differences using 

 separate rather than pooled variance estimates 

 because the cell variances were unequal according 

 to Levene's test; this test was selected because it 

 is more robust under nonnormality than either the 

 common F-ratio or Bartlett's test (Brown and For- 

 sythe 1974). Degrees of freedom were calculated 

 with Satterthwaite's approximation, so that sig- 

 nificance probabilities could be obtained from an 

 ordinary ^-distribution (Snedecor and Cochran 

 1980). ' 



561 



