PARRACK: FISHING EFFORT FROM AERIAL SEARCH DATA 



Wallis one-way analysis of variance (Siegel 1956: 

 184-194) was also applied and indicated the same 

 general results. The probability of obtaining the 

 calculated test statistic under the hypothesis of no 

 difference among these groups with respect to 

 means is 0.553, 0.410, and 0.872 for estimators I, II, 

 and III, respectively. Both parametric and non- 

 parametric tests, then, indicate that the error 

 rates of each estimator are of the same magnitude 

 regardless of the kind of category estimated. 



Analyses for possible differences in error 

 coefficients among estimators were carried out in 

 the same manner. All error coefficients of estima- 

 tor I were considered as one group, of estimator II 

 as another, and of estimator III as the third group. 

 Both parametric analysis of variance and nonpar- 

 ametric techniques indicated that the different 

 estimators probably produced different error 

 coefficients. The likelihood of obtaining the cal- 

 culated F statistic under the hypothesis of no 

 difference among the groups is low (0.006). The 

 Kruskal-Wallis analysis technique also indicated a 

 low probability of obtaining the calculated statis- 

 tic under that hypothesis (0.007). 



Cumulative frequency distributions of the c 

 from 1971-73 estimates were used to compare 

 estimator performances and to establish estima- 

 tor dependability. These frequency distributions 

 were established in the following way. Arbitrary 

 bounds or intervals (ju) were set up so that the first 

 bound included error coefficients from -0.049 to 

 -1-0.049, the second from -0.099 to -1-0.099, and so 

 on. The number of error coefficients from Table 4 

 falling in each interval was counted; these counts 

 were then divided by the total number of 

 coefficients calculated for that estimator to estab- 

 lish the percent of occurrences in each interval. 

 These proportions were then interpreted to be the 

 likelihood of the error coefficient occurring within 

 each bound (<I>jli, Table 5). Graphs of these 

 probabilities (Figure 3) indicate that estimator III 

 is the most desirable. Its error coefl^cient is most 

 likely to occur within set error bounds of *0.50 or 

 less. For error bounds greater than *0.50, estima- 

 tor I was superior. Estimator II was always infer- 

 ior to estimator III, but for very narrow error 

 bounds (-0.20 and less) estimator II was superior 

 to estimator I. 



Although estimator II produced the least desir- 

 able calculations of days fished, a like algorithm 

 also based upon P{G/N) estimated days on 

 grounds acceptably well: 



Table 5.-Fre(|uency of error coefficients of estimates of (iay.s 

 fished, 1971-7:1 



Frequency of occurence 



INTERVAL OF ERROR COEEFICIENT 



Figure 3.- Probabilities of error coefficients occurring within set 

 bounds for three estimators of days fished. 



g = P{G/N)iVM-g')+g' 



(5.1) 



where g is estimated days on grounds and other 

 symbols are as before. Comparisons of calculated 

 days on grounds to reported days on grounds were 

 made when reported values were available (Table 

 6). These comparisons indicate that the error was 

 less than *0.50 for 83% of such estimates. 



Approximations of estimation dependability 

 may be established from the calculations of 



511 



