Harris et al.: Otolith morphometric analysis for species discrimination of Sebastes melanostictus and S. aleutianus 
241 
Blackspotted Rougheye 
Figure 4 
Plot snowing how the logistic regression model assigns species identifications 
to each specimen in the testing data set collected in the Gulf of Alaska in 
2009 and 2013. Values on the y-axis are calculated probabilities of each speci¬ 
men being a rougheye rockfish ( Sebastes aleutianus). If P>0.5, the specimen is 
identified as a rougheye rockfish, and if P< 0.5, it is identified as a blackspotted 
rockfish (S. melanostictus). The left panel shows P-values for specimens that 
were genetically identified as blackspotted rockfish, and the right panel shows 
P-values for specimens that were genetically identified as rougheye rockfish. 
The error bars represent 95% prediction intervals for specimens, based on the 
standard error of the predictions. An identification is assigned only when the 
prediction interval does not cross the dashed line for probability of 0.5. Spec¬ 
imens classified as uncertain had a prediction interval that was not confined 
to a single species. 
(decreased accuracy) caused by discarding a number of 
otoliths (approximately 5% overall) that are ruled 
uncertain on the basis of the estimated error around 
each prediction. Likewise, if the uncertain specimens 
had been classified, overall accuracy would have 
increased but the number of misidentifications would 
have increased as well (Table 6). The logistic regression 
method developed in this study allows the use of an 
algorithm to determine when specimens are too diffi¬ 
cult or atypical to identify. Removal of the uncertain 
otoliths led to high specificity, with less than 5% of 
specimens misidentified. With both high accuracy and 
high specificity, this method ensures the highest qual¬ 
ity data are available for use in managing the popula¬ 
tions of these species. In contrast, if a simpler method 
that attempts to identify all specimens is used, some 
additional specimens will be correctly identified if they 
fall on the correct side of the threshold of P=0.5 (Fig. 4), 
but misidentifications double to as much as 9-10% in 
rougheye rockfish. In this study, no apparent pattern 
was observed with uncertain specimens with respect to 
age or location, meaning that the exclusion of uncertain 
specimens is unlikely to have caused a 
bias in this data set. 
A key finding in this study is the 
importance of age in distinguishing the 
2 species. A common practice in mor¬ 
phometric studies is to standardize all 
fish to a set size, usually determined 
by standard or fork length (Campana, 
1999; Lleonart et al., 2000). For this 
study, age was included as an interac¬ 
tion effect with the other predictors, 
allowing the model to account for it 
directly. In this study, we observed that 
almost no difference existed between 
the species when their otolith measure¬ 
ments were compared with fork length, 
but a clear separation in the growth 
patterns emerged when age was con¬ 
sidered (Figs. 2 and 3). Our findings are 
consistent with a growing understand¬ 
ing that, despite their similarities, the 
rougheye and blackspotted rockfish 
species have different rates of growth 
(Conrath, 2017; Shotwell et al., 2017) 
and specifically that rougheye rockfish 
grow faster and attain a larger size at 
age. The von Bertalanffy growth param¬ 
eters obtained in this study indicate a 
strong divergence between the 2 species 
(Table 4). Given our findings, it is pos¬ 
sible that using age alongside length or 
body size may help resolve other prob¬ 
lems of difficult stock or species dis¬ 
crimination. Including measurements 
such as fork length and age into a 
model as an interaction carries the few¬ 
est overall assumptions, makes full use 
of all available information, and allows the difference in 
growth rate to inform the results and improve the predic¬ 
tions of the model. 
The most significant disadvantage of this method is 
that reading the ages of otoliths is time consuming, and 
each specimen would not receive a final species identifi¬ 
cation for several weeks or months. Although measure¬ 
ments of age have a higher error than other data, the 
results of our simulations indicate that moderate age 
determination errors, consistent with commonly reported 
estimates, are unlikely to degrade the accuracy of this 
method. The AFSC reports an age determination CV of 
0.08 for rougheye rockfish, meaning that otolith age 
readers tend to disagree by 8%. In the simulation con¬ 
ducted in our study, the logistic regression model main¬ 
tains high overall accuracy (=90%) in this simulation 
(Fig. 5). This method is somewhat more vulnerable to 
systematic bias in age estimates. The drop in classifica¬ 
tion accuracy from increasing directional bias is not 
greater than that from random noise, but it happens 
unevenly. The core observation on which this model is 
based is that otoliths of rougheye rockfish are larger and 
