164 
) 
ot 
(2) 
oO 
oS 
> 
Oo 
Oo 
— 
=) 
(S) 
1S) 
<x 
E 
Ww 
nS 
uw 
Fishery Bulletin 119(2—3) 
5 10 15 20 25 300 10 15 20 25 30 
YO S 10 Ww A 2s) co 
Year 
Figure 7 
Accuracy (%) of determining overfishing status over time for each of 4 estimation models under cases 0-12 
(CO-C12). Overfishing status is determined by dividing fishing mortality rate (F) by F\;,,:4, which was set to 
the F that corresponds to maximum sustainable yield. The models, evaluated in this study for use in stock 
assessments, include the Assessment Model for Alaska (AMAK), the Age Structured Assessment Program 
(ASAP), the Beaufort Assessment Model (BAM), and Stock Synthesis (SS). 
to identification of common features used (Table 2) in 
this study could be applied to multi-model comparisons 
in other studies. They might also prove useful in ensem- 
ble modeling, which is now gaining traction in fisheries 
science as a means to combine the estimates from multi- 
ple stock assessment methods (Brodziak and Legault, 
2005; Brodziak and Piner, 2010). Ensemble modeling 
loosens the assumptions associated with selecting a sin- 
gle “best” assessment model (Rosenberg et al., 2018). 
Stewart and Martell (2015) proved that ensemble 
modeling benefits from guidelines for developing sets of 
candidate models. The steps developed in this study to 
identify common features across alternative models may 
facilitate selection of plausible models and identification 
of the sources of differences among estimates before con- 
structing an ensemble. 
Examination of similarity in initial numbers at age 
The approach used in each EM to estimate the initial num- 
bers at age depends on 3 factors: the selected initial year; 
the level of fishing, if any, that typically occurred prior to 
that initial year; and the availability of age data beginning 
with the initial year, such that the initial non-equilibrium 
age composition can be estimated. In principle, if informa- 
tive age data are available across years, the EM results 
should not be dependent on the choice of the initial year 
because the estimated age composition in the initial year 
could have been alternatively estimated by starting the 
model at an earlier year and estimating the age composi- 
tion for that year as projected from earlier recruitments. 
Also, when age composition data from the fleet and survey 
are available in the first year, the numbers at age for the 
