FISHERY BULLETIN: VOL. 74, NO. 3 



T.^BLK 4. -Error coefficients of three estimate.s of days fished. 



'GDR = German Democratic Republic. 

 ^FRG = Federal Republic of Germany. 



I 



.100 



.000 



.100 



.000 



.100 



Estpmotor I 



I |"l |"i |»-r 



T 



Estimotor II 



rA. 



I ' I ' I 



U^^l 



J^ 



^■ 



' ' ' ' 1.30 ' 1 60 ' Z.bC 



I ' I ' I ' I ' I ' I   ■..30.,eo'^..o 



-1.00 -JBO -.60- -fiO -.20 .20 .40 .60 .80 1.00 hAO l.TO 2 60 



Figure 2.-0bserved frequency of error coefficients (t) of three 

 estimators of days fished. 



less than 0.01. In the case of error coefficients of 

 estimator I, that probability is 0.05. 



An analysis of variance for a one-way layout 

 with unequal replication (Steel and Torrie 



1960:112-114) was employed to investigate possible 

 differences in the mean values of estimation error 

 in regards to the kinds of category estimated. 

 Here, a separate analysis was carried out for each 

 of the three estimators. All error coefficients for 

 estimates of total days fished by country, all gear 

 types combined, were considered as a single group; 

 error coefficients of estimates of days fished by 

 each gear type (i.e., estimates of days fished by all 

 side trawls and by all stern trawls) as another 

 group; and error coefficients of estimates of days 

 fished for gear-country categories as the final 

 group. F-tests indicated a high probability (>0.25) 

 that errors were the same among these groups in 

 the cases of estimators I and III. Estimation of 

 gear totals (i.e., total stern trawl and total side 

 trawl) was not possible via estimator II because 

 days on grounds were not reported for all coun- 

 tries. The analysis for estimator II therefore 

 included two groups, i.e., gear-country categories 

 and country totals. As before, the likelihood that 

 error rates were the same among these two groups 

 was high (>0.25). 



Although application of the F-test requires that 

 normality assumptions be made, the test is robust 

 in regard to violations of these assumptions 

 (Scheffe 1959:361-364) so that limited deviations 

 from normality are likely to be of limited con- 

 sequence. However, the nonparametric Kruskal- 



510 



