9711 - Laserfiche WebLink

Home Browse Search

a decline in abundance from the early 1990s to the present (tig. 4). However, as is apparent in the data, these estimators are very imprecise with corresponding poor ability to detect significant trends. Additionally, preliminary analyses of data collected in this program suggest that the 2007 estimate may be up to twice as large as the 2006 estimate (R. Van Haver- beke, U.S. Fish and Wildlife Service, oral commun.; 2007). ASMR without Tag-Cohort Specific Data Overall, the three ASMR formulations generally agree that adult (age-4+) HBC abundance has gradually increased since about 2001 (lig. 5). Among these three models, the 2006 adult abundance estimate is 6,690 (95% credible interval (CI) 6,403-6.994), 6,768 (95% CI 6,397-7,131), and 6.648 (95% CI 6.222-7,1(2) for models ASMR 1, 2, and 3, respec- tively. These results suggest that this population has increased 35%-40% from an estimated low abundance of approximately 4,800-5.000 during 2000-01. Estimated recmitment (age-2) among models is also in agreement (fig. 6). Following low recmitment for brood years during the early I 990s, all the models suggest that reclUitment incTea<;ed through the latter part of the 1990s. The biggest discrepancy amOl;g the three models is that ASMR 1 suggests a decIine in recruitment following the 2000 brood year, while the other two models suggest stability. The stmctural assumptions of model ASMR 3 (see Coggins and others, 2006b) do not permit a reliable recmitment estimate for brood year 2003. An additional differ- ence in the model results is in the estimation of instantaneous adult mortality (M",), where adult mortality ranges from 0.119 (ASMR 1) to 00133 (ASMR 3). Model Evaluation and Selection With these results in hand, the question becomes which model is best? Stated another way, which model produces results most consistent with or best supported by the data? In this case. the discrepancies among model results related to adult abundance are not large; as a result, selecting the best model is probably not critical from a management or con- servation perspective. However, the models~ do reflect rather different hypotheses about recmitment trends. Model ASMR I supports the hypothesis that recmitment has declined follow- ing the 2000 brood year, while the other two models suggest relative stability. Therefore, selecting the model (hypothesis) that is the most consistent with the data is desirable. Model tit to the data was assessed by plotting the Pear- son residuals of the observed and predicted numbers of fish marked and recaptured for each year and age (figs. 7-9). The patterns in Pearson residuals for both ASMR I (fig. 7) and ASMR 2 (fig. 8) demonstrate systematic lack of tit for partiClI- lar sets of cohorts. This is best seen in the recapture residuals where it is apparent that there are more fish observed than predicted for about eight pre-1990 cohorts, particularly for observations after 2000. Additionally, there are fewer recap- tures associated wi th the 1992 cohort than are predicted. These 11 systematic trends are likely imposing bias in the model results for ASMR I and ASMR 2. In contrast, there is much less systematic lack of fit in the residual patterns for ASMR 3 (fig. 9). Additionally, among the three models tbe Pearson residual standard deviation is smallest for ASMR 3. The finding that ASMR 3 has the best fit among the three models is not surprising, since it has the largest parameter set. Although ASMR 3 only varies 13 parameters in the direct numeric search, the conditional maximum likelihood estimates are used for each age- and time-specific capture probability (Coggins and others, 2006b). Therefore, assuming a liberal maximum longevity of 50 years, ASMR 3 has 895 param- eters. The question then becomes whether these additional parameters are justified. To provide insight into this question, the relative K-L distance was estimated using AIC (table 1). These results strongly indicate that model ASMR 3 is superior to ASMR 1 and 2; a finding congruent with the evaluation of model fit using Pearson residuals. Although it is comforting to find agreement between these two evaluations, one should ask: why is the more complicated structure of ASMR 3 needed to adequately fit the data? Since the fundamental difference between ASMR I and 2 and ASMR 3 is the amount of f1exibility in age- and time-specific capture probabilities, the pattern in ASMR 3 estimated capture probabilities is of interest (fig. 10). The discrepancy in capture probabilities between sampling period 2 (i.e., 1991-95; heavy gray lines) and sampling period 4 (i.e., . 2000-06: heavy black lines) suggests a major shift in the gear selectivity. Sampling since 2000 appears to be much less effec- tive at capturing fish between ages 9-20 years old than was sampling during the second period. Since structural assump- tions in ASMR 1 and ASMR 2 require that vulnerability is asymptotically related to age, it is not surprising that these models are not able to account for this unexpected pattern, and thus display poor model fit. ASMR with Tag-Cohort Specific Data In addition to repeating the analyses by Coggins and others (2006a) above, the ASMR models were also fit to the tag-cohort specific data using the log-likelihood in equation (2). The trends in adult abundance and recruitment are quite similar to those found using the simpler log-likelihood (figs. II and 12). In general, adult abundance estimates are slightly higher at the beginning of the time-series and slightly lower at the end. Adult abundance estimates for 2006 are 6,057 (95% CI 5,797-6308), 6,138 (95% CI 5,842-6,458), and 5,893 (95% CI 5,554-6,242) for the ASMR 1, 2, and 3 models, respectively. Adult mortality (Moo) estimates from the models fit to the stratified data indicate slightly higher adult mortality (0.128,0.137, and 0.148 for ASMR 1-3, respectively) than when fit to the pooled data. This finding is consistent with the more rapid decay observed in the time-series of adult abun- dance. Another difference is a slight increase in the precision of the estimates using the stratified data (fig. 13). This obser- vation provides marginal confirmation of the assertion that it