Bias and Precision of Estimates from an Age‐Structured Stock Assessment Program in Relation to Stock and Data Characteristics

Abstract

Assessments for many U.S. Pacific coast groundfish stocks have been developed using the statistical catch‐at‐age method known as Stock Synthesis. This study used Monte Carlo simulation and a fractional factorial experiment to evaluate the effects of input data errors and stock characteristics on bias and precision in estimates of ending exploitable biomass, rate of fishing mortality, depletion, and other output variables. Nine factors were examined: length of the data series, rate of natural mortality, shape of the fishery selectivity curve, trend in fishing mortality, recruitment pattern, and level of sampling error in the data for catch, fishing effort, a survey biomass index, and sample size for fishery and survey age compositions. Length of the data series, age composition sample size, survey biomass variability, and fishing effort variability were the most influential factors for most of the output variables. The estimates of depletion had the least bias and the estimates of starting biomass the smallest variability; the estimates of ending recruitment had the greatest bias and largest variability. For all the output variables examined the estimates appeared to be median‐unbiased. For the conditions considered in the experiment, it appears that the accuracy of assessment estimates for ending exploitable biomass and projected catch would be more readily improved by increased age composition sampling than by comparable (but much more expensive) improvements in survey estimates of stock biomass.

Evaluating alternative estimators of fishery management reference points

“…Accordingly, either the number of years with survey data or the length of the catch time-series was important in determining each stock assessment method's ability to estimate h. Recruitment variability is important for nearly all of the estimators (Table 8 and Fig. 8), which is consistent with the findings of Yin and Sampson (2004).…”

“…Contrary to the results of Yin and Sampson (2004) and Punt et al (2002), which suggest that B 0 and B last share similar levels of error which factor out in the calculation of stock depletion, B 0 was estimated either better (rockfish and flatfish) or similarly (the hake) to stock depletion (Tables 4 and 5). The enhanced ability to estimate B 0 suggests that the biomass at the end of the model time period is not well determined, most likely due to incomplete information about incoming recruitments, subjecting B last to a higher level of error which does not factor out in the calculation of stock depletion.…”

“…Following Dichmont et al (2006) and Yin and Sampson (2004), the importance of the factors in Table 3b, and their first-order interactions, in determining estimator performance was explored by fitting a series of General Linear Models to the logarithms of the MAREs for each trial and life history type, separately for B 0 , B MSY , stock depletion, and h. The GLM models included: (a) models for Table 5 As for Table 4, except the results are for estimators of depletion Table 6 As for Table 4, except the results are for estimators of BMSY each individual factor in Table 3b on its own, (b) a model with all of the main factors in Table 3b, and (c) a model with main factors and first-order interactions selected using AIC (Burnham and Anderson, 2000). The purpose of the last analysis was to assess the importance (rather than statistical significance) of interactions among factors.…”

Performance of deviance information criterion model selection in statistical catch-at-age analysis

“…These assumptions were pragmatic choices and are common to many simulation studies evaluating assessment methods (e.g., Yin and Sampson, 2004). In real assessments these quantities are estimates, and their true values are unlikely to be truly constant.…”

Evaluating the estimation of fishery management reference points in a variable environment

“…The importance of the factors in Table 1(b), and their first-order interactions, in determining estimator performance was explored by fitting a series of general linear models (GLMs) to the logs of the MAREs for each trial and life history type, separately for B 0 , B MSY , current biomass relative to B 0 , and h (Dichmont et al, 2006;Yin and Sampson, 2004;Haltuch et al, 2008). The GLM included: (a) models for each individual factor in Table 1(b), (b) a model with all of the main factors in Table 1(b), and (c) a model with main factors and first-order interactions selected based on AIC comparisons (Burnham and Anderson, 2000).…”