826 



Fishery Bulletin 98(4) 



ing site were separated from samples collected farther 

 downstream, and the difference in large-scale coverage 

 was accommodated. 



Finally, statistical tests were conducted to examine 

 effects of year and region on age-0 density, which was not 

 straightfoi-ward owing to the occurrence of many zeros and 

 a few very large values in the catch data. Four different 

 procedures were employed and the results were compared. 

 All statistical tests were accomplished with SYSTAT for 

 Windows (Wilkinson, 1996). 



The first procedure was a two-way analysis of variance 

 (ANOVA) of two different transformations of all four years 

 of data. One ANOVA was based on a log-transformation 

 (Sparholt et al.. 1991) (log n,(density)+0. 00001, nonzero 

 catches ranged from 0.00001-0.6 fish per m ' I and the other 

 was based on rank-transformed data (Conover, 1980). Nei- 

 ther transformation, however, removed the spike of zeros 

 in the frequency histogram of densities and thus made 

 the requisite assumption of normality tenuous. To avoid 

 empty cells data from the Unimak region were omitted; 

 they were collected only during 1987-88. 



The second procedure was a distribution-free, nonpara- 

 metric test to again look for effects of year and region 

 (Methven and Bajdik, 1994). Unlike Methven and Bajdik 

 ( 1994), however, the age-0 data were not paired; therefore 

 the signed rank test was replaced by the Kruskal-Wallis 

 test. Because this is a one-way test, subsets of the data 

 were selected so that the effect of the first factor was 

 examined within a level of the second factor (e.g. year 

 within a region) and vice-veisa (e.g. region within a year) 

 (Valle et al., 1999). In all, seven tests wei'e run. Signifi- 

 cance of the outcome for each test was based on the Bon- 

 ferroni-corrected alpha level of 0.05. 



The third and fourth procedures were conducted primar- 

 ily as a means to circumvent the zero-catch problem but 

 also because of differences in the information conveyed 

 by zero and nonzero catches (Pennington, 1983; Randa, 

 1984). Zero catches indicate absence, and nonzero catches 

 indicate presence as well as some measure of abundance. 



The third procedure involved chi-square tests to test 

 whether the presence or absence of age-0 pollock was asso- 

 ciated with year or region. Annual or regional differences 

 in age-0 frequency of occurrence indicate a change in fish 

 patchiness, or a change in the ability to target them; both 

 may be associated with a change in abundance. It was 

 necessary to conduct one chi-square test for each region 

 to examine annual differences and then for each year to 

 examine regional differences. Thus, Bonferroni corrections 

 were again necessary. 



The fourth procedure was a two-way ANOVA test of 

 effects of region and year on nonzero densities. The non- 

 zero densities appeared normally distributed after being 

 log,||-transformed. Again, the Unimak data were omitted. 



A year-region interaction term was included in the 

 ANOVA tests of the first and fourth procedures. Signifi- 

 cance of the interaction term indicates that annual dif- 

 ferences vary by region, or that regional differences vary 

 by yean The former identifies regions with high annual 

 variability; whereas the latter implies annual variation 

 in fish distribution. Significance of the interaction term 



determined which multiple comparison test was used post 

 hoc. If it was not significant, a Bonferroni multiple compar- 

 ison test was used for comparing levels of significant main 

 effects. If it was significant. Fishers least significant dif- 

 ference (LSD) test was preferred. In SYSTAT, the LSD test 

 does not automatically correct for the number of compari- 

 sons being made (Wilkinson, 1996). This correction was 

 desirable because only a subset of all possible pair-wise 

 comparisons was of interest. Only comparisons among 

 regions within year and among years within region were 

 of interest, those involving different years and regions (i.e. 

 1985 Kodiak versus 1986 Shumagin) were excluded. A 

 Bonferroni correction was then applied to the remaining 

 comparisons to maintain an overall 0.05 alpha level. 



Examination of effects of year and region on age-0 length 

 was accomplished by using a two-way analysis of covari- 

 ance (ANCOVA). An ANCOVA was necessary to account 

 for variation in length due to differences in the collection 

 day-of-year (expressed as the number of days since 1 Jan- 

 uary). The first ANCOVA was used to examine data col- 

 lected from the Kodiak and Shumagin regions during all 

 four years. The second ANCOVA was conducted with just 

 the 1987 and 1988 data to compare lengths among all 

 three regions. The covariate was not allowed to interact 

 with the main effects (ie. slope homogeneity was assumed 

 rather than tested) owing to the short duration of collect- 

 ing in some regions and years in relation to the variation 

 in mean length per haul. As in Anderson et al. (1995), the 

 dependant variable was mean length per haul rather than 

 individual length measurements; this variable simplified 

 the model but allowed no within-haul variability. Mean 

 size was weighted by mortality-adjusted fish density. 



The relationship between annual estimates of age-0 den- 

 sity or size and recruitment was examined graphically 

 because there were insufficient degrees of freedom for a 

 global significance test of correlation. Pollock recruitment 

 was indicated by the Gulf-wide abundance of age-2 individu- 

 als (Bailey et al., 1996a). No attempt was made to account 

 for autocorrelation in this time series because pollock exhibit 

 low autocorrelation in recruitment (Hollowed et al.. 1998). 



Results 



Overall, 406 trawl samples were collected yielding 384 

 density estimates (zero and nonzero catches) and 335 esti- 

 mates of mean length (Table 2). Stations were occupied 

 from west to east, except in 1985 when sampling pro- 

 gressed from cast to west and then doubled back to end 

 near northeast Kodiak Island. Sampling around Kodiak 

 Island was mostly over the inner shelf and in bays ( Fig. 1). 

 Coverage of the outer shelf was best farther west, between 

 158° and 164°W longitude. Sample depth varied consider- 

 ably, and it appears that the daytime vertical position of 

 age-0 pollock was also quite variable (Fig. 2). 



Abundance 



Results from each of the four statistical procedures con- 

 sistently indicated that year and region interact in their 



