496 



Fishery Bulletin 101(3) 



Table 7 



Significant parameters and interactions among parameters influencing bycatch estimates from simulated data. Significance was 

 tested by using ANOVAs. *** =P< 0.001, ** = 0.001 5 P < 0.02, * = 0.02 <.P<. 0.05. AvgF = the mean catch of fish; AvgS = the mean 

 catch of shrimp; FCV = fish CV; SCV = shrimp CV; ECV = Effort CV; r = the correlation coefficient between the catch offish and the 

 catch of shrimp, n = observer coverage; P(S) = probability of catching shrimp; and P(F) = probability of catching fish. The ranges 

 of values for all parameters are shown in Table 1. The CPUE-mean-per-unit estimator (CPUE=catch per unit of effort) uses unit 

 effort as a proxy for sample size. The basic F:S ratio estimator is the mean of the individual fish to shrimp ratios, and the gi-and 

 F:S ratio estimator is the ratio of the mean catches offish and shrimp. The basic CPE ratio estimator is the mean of the ratios of 

 catch per effort, where effort is a variable measure of effort, and the grand CPE estimator is the ratio of the mean catch offish to 

 the mean estimate of effort. 



Normal distribution 



Main effects 

 SCV 

 ECV 



Two-way interactions 

 AvgS X AvgF 

 SCVxn 

 FCVxn 

 ECV X 71 

 ECVxr 



Three-way interactions 

 AvgS x AvgF X FCV 

 AvgS X FCVxn 

 AvgS xnxr 

 AvgF X SCV X ECV 

 AvgF X FCVxn 

 AvgF X ECV xn 

 SCV X FCV X ECV 

 SCVxnxr 

 FCV X ECVxr 

 FCVxn xr 



*** 

 ** 



** 

 ** 



contuiued 



shrimp. In fact, only the mean-per-unit estimators gave 

 overall bycatch estimates that were statistically unbiased 

 for both the normally distributed and delta lognormally 

 distributed data. Three other estimators, the grand F;S and 

 the grand CPE ratio estimators for the normally distrib- 

 uted data and the grand CPE ratio estimator for the delta 

 lognormally distributed data, gave bycatch estimates that 

 differed by less than \7i on average from the actual simu- 

 lated bycatch, although these differences were statistically 

 significant. In these simulations, the reason that a bias of 

 less than 1*% was statistically significant was probably due 

 to the large number of fleets included in the paired-sample 

 t-tests for each distribution, which made the standard er- 

 rors and confidence intervals very small. Both the F;S and 

 CPE basic ratio estimators (mean of the ratios) performed 

 poorly in terms of bias in both normally distributed and 

 delta lognormally distributed data, overestimating bycatch 



by between 10% and 427%, regardless of whether shrimp 

 catch or hours fished was used as the auxiliary variable. 



The ANOVAs performed on the simulated data indicated 

 why the different estimates are so variable, and showed 

 the complexities of the interactions among the parameters. 

 The CPUE-mean-per-unit estimators displayed the fewest 

 main effects and showed the fewest higher-level interac- 

 tions among parameters. In the normally distributed data, 

 there were no significant main effects or 2-way interac- 

 tions for the CPUE-mean-per-unit estimator although 

 observer coverage, (the sample size for each fleet) was a 

 factor in both significant 3-way interactions. For the delta 

 lognormally distributed data, the CV of the catch offish 

 was a significant maui effect for the CPUE-mean-per-unit 

 estimator, and observer coverage was a factor in three of 

 the four significant 3-way interactions. The probability of 

 catching fish and the CV of the fish catch were also fac- 



