116 
Fishery Bulletin 11 6(2) 
layers for 5 environmental variables (covariates). To 
account for depth-related biases (because the surveys 
rarely exceed 400 m, although we know that adult 
halibut can inhabit depths of ~1000-m (Miller et al., 
1991; Cargnelli et al., 1999), the extent of the output 
prediction layer was limited to the area of the strata 
sampled. Because of this restriction, it is highly likely 
that some potentially suitable areas were not mapped; 
however we contend that we are capturing the major¬ 
ity of juvenile habitat because there is an association 
between depth and halibut size (Bigelow and Schro- 
eder, 1953; Collette and Klein-MacPhee, 2002) because 
juveniles typically occupy shallower depths and move 
deeper with age (e.g., Sigourney et al., 2006). 
Our model with the strongest diagnostics incor¬ 
porated a restricted background (with true absence 
points) and its data were divided by survey (Table 1). 
By default, a model will select a set of random “back¬ 
ground” locations (in unsampled locations) to represent 
pseudo-absence points. By using true absence points in 
place of the background data, valuable information is 
incorporated into the model. We ran 3 separate models 
for separate surveys: NS and U.S., GSL, and NF, and 
then combined the 3 output layers, keeping the largest 
habitat suitability value for each location. This model¬ 
ing approach made full use of the data available, and 
running separate models accounted for inherent differ¬ 
ences among the surveys (sampling effort, catchability 
of gear type, and seasonality (Suppl. Tables 1 and 2) 
(online only) and enabled a consideration of smaller scale 
trends and potential spatial variation in habitat pref¬ 
erences in different regions. 
To evaluate the models, we performed a cross-valida¬ 
tion and compared “area under omission curve” (AUC) 
values. For cross-validation, we withheld 20% of the in¬ 
put data (test data) when running the model, and then 
compared corresponding predicted values from the re¬ 
sulting habitat-suitability layer with the true values 
at the test-data locations (a good model can accurately 
predict the likelihood of species presence in test-data 
locations). Cross-validation also produces AUC values 
that indicate the proficiency of a model in differenti¬ 
ating between presence and absence sites and is the 
standard for maximum entropy assessment (Elith et 
al., 2006; Phillips and Dudik, 2008). On a scale from 
0 to 1, an AUC value greater than 0.9 is widely ac¬ 
cepted as “excellent” or “high” accuracy, and less than 
0.6 is generally considered a “fail” because 0.5 means 
probabilities are no better than random (Phillips and 
Dudik, 2008; Halvorsen, 2013). For the remainder of 
the analysis, we used the habitat-suitability layer from 
the strongest performing model, and we also excluded 
the GSL because tagging evidence supports the man¬ 
agement of this area as that of a separate stock (Mc¬ 
Cracken, 1958; Neilson et al. 5 ; Stobo et al., 1988; den 
Heyer et al., 2012; Le Bris et al., 2017). 
For regional comparisons, we partitioned the habi¬ 
tat-suitability layer using the NAFO and EEZ shape- 
files as boundaries and calculated statistics within 
each division. We quantified suitable habitat by using 
a classification rule: a pixel was classified as “suitable” 
if its predicted habitat-suitability value was greater 
than 0.54. Typically values greater than 0.5 are ac¬ 
cepted as the probability of presence of halibut at sites 
where environmental conditions are ‘typical’ of pres¬ 
ence (Phillips and Dudik, 2008; Elith et al., 2011), and 
by increasing this threshold, we increased this like¬ 
lihood. The proportion of the total available suitable 
habitat to fall within each NAFO division (p t ) and the 
catch shares per NAFO division allocated on the basis 
of abundance of Atlantic halibut (p ( ) were both calcu¬ 
lated as 
where the summed suitable habitat (km 2 ) or the 
summed survey locations with species presence with¬ 
in i NAFO division (sj), were divided by the totals 
summed across all j NAFO divisions. We will refer to 
these values as “SH-based shares” and “abundance- 
based shares,” respectively. 
Assuming a constant relationship between poten¬ 
tial abundance and availability of suitable habitat, 
we plotted the relationship between SH-based shares 
and abundance-based shares (survey data) against a 
1:1 baseline. We interpreted this baseline as “expected 
habitat productivity” where proximity to the baseline 
indicates whether the productivity of an area is above, 
meeting, or below expectations. The abundances deter¬ 
mined from research surveys were derived from the 
same data set used to quantify juvenile habitat. Any 
positive relationship between habitat and abundance 
determined from research surveys would suggest that 
the amount of habitat is related to juvenile production. 
To explore the empirical relationships between suit¬ 
able habitat availability and fishery productivity (e.g., 
Brosse et al., 1999; Holbrook et al., 2000; VanDerWal 
et al., 2009) we plotted the relationship between SH- 
based shares and commercial landings in recent and 
historical fisheries (abundance-based shares [land¬ 
ings]). This approach allowed us to test the feasibility 
of using the 1:1 baseline and to model suitable hab¬ 
itat shares as a proxy for production and to explore 
the potential for similarities between adult and ju¬ 
venile halibut distributions. We further explored this 
potential overlap in choice habitat by overlaying the 
habitat-suitability layer with the 2010-2014 Canadian 
commercial fisheries map of landings, by catch weight 
(Butler and Coffen-Smout, 2017), and by extracting 
habitat suitability values associated with recent land¬ 
ings data which have a minimum legal fish size of 81 
cm TL (DFO 2 ). 
There are several assumptions inherent in this ap¬ 
proach. First, is that good juvenile habitat leads to 
high adult abundance. Although we do not have direct 
evidence to support this assumption, it is supported 
by the nursery-size hypothesis [stated earlier], and in¬ 
ferred by density-dependent mortality during early life- 
history stages (lies and Sinclair, 1982; Rijnsdorp et al., 
1992; Gibson, 1994; Beverton, 1995; Sundblad et al., 
