114 
Fishery Bulletin 11 6(2) 
enabled us to predict across a much larger area, and 
because the research vessel survey data are zero in¬ 
flated (absences are less certain than presences among 
survey data), we were able to take advantage of the 
presence-only modeling capabilities of maximum en¬ 
tropy (Sundblad et al., 2014). 
Data 
We modeled habitat suitability by using research sur¬ 
vey data from 2001 through 2013—a timeframe that 
captures the current period of high Atlantic halibut re¬ 
cruitment. The U.S. National Marine Fisheries Service, 
NOAA, and Fisheries and Oceans Canada (DFO) have 
conducted seasonal or annual research surveys (or 
both) to foster an improved understanding of ground- 
fish abundance and distribution since the early 1960s 
(NOAA) and 1970s (DFO), (Azarovitz, 1981; Simon and 
Comeau, 1994). The Canadian surveys sample NAFO 
divisions off Newfoundland and Labrador (NF), Nova 
Scotia (NS), and within the Northern and Southern 
Gulf of St. Lawrence (GSL) (Fig. 1). The U.S. sur¬ 
veys sample Georges Banks, the Gulf of Maine, and 
the Bay of Fundy (here, we collectively refer to these 
regions as “U.S.” (Fig. 1). All surveys were conducted 
with bottom-trawl gear and fish abundance, biomass, 
water depth, and bottom temperature were recorded 
in a comprehensive database, which can be publically 
accessed from the Ocean Biogeographic Information 
System (available from website; Shackell et ah, 2005; 
DFO 6 ). Since the start of these surveys (-1963), they 
have been performed during all seasons and 7 differ¬ 
ent gear types have been used. A wide variety of fish 
species are captured over the course of each cruise; At¬ 
lantic halibut are usually caught as juveniles (annual 
median fish length 40-50 cm TL), and larger halibut 
are thought to outswim the trawls (Trzcinski and Bow¬ 
en, 2016). Samples were collected from all the NAFO 
divisions of interest, although there are variations in 
survey seasonality, intensity, and gear type, and there 
is difficulty in sampling over rough terrain. Because 
of these limitations and variations between surveys 
there are unsampled regions throughout the study area 
(Suppl. Tables 1 and 2) (online only), all of which can pro¬ 
duce an uneven and biased data set because some re¬ 
gions have more representative (larger in quantity and 
more evenly distributed) sample sets than others. 
If bias is not accounted for, the predictive power 
of the model is lessened (Philips and Dudik, 2008). 
Fourcade et al. (2013) explored 5 parameterizations 
available to maximum entropy modeling that are com¬ 
monly used to correct for bias, thus enabling models 
to predict more accurately in under-sampled locations. 
These methods include 1) systematic sampling (data 
are placed on a grid and one random sample per cell 
6 DFO (Department of Fisheries and Oceans Canada). 2016. 
2015 Maritimes research vessel survey trends on the Sco¬ 
tian Shelf and Bay of Fundy. Can. Sci. Advis. Seer. Sci. 
Resp. 2016/011, 66 p. [Available from website.] 
is used in analysis), 2) clustering samples (data are 
subset by using a principal component analysis that 
identifies spatial autocorrelation), 3) restricting the 
background (selecting only background points that fall 
within the extent of the survey [comparable to using 
locations where surveys yield true absence of fish]), 4) 
incorporating a bias file (values are weighted by using 
a raster that reflects the sampling effort or sampling 
probability), and 5) geographically splitting the data 
(the model is computed separately for each area and 
results are combined) (Fourcade et al., 2013). While 
designing our model, we tested several of these data 
correction techniques and compared their ability to im¬ 
prove output diagnostics; however we did not attempt 
to systematically sample or cluster data because this 
approach would have caused too many data to have 
been lost. 
We incorporated 5 environmental raster layers in 
our model: 1) bathymetry (created with the GEBCO 30 
arc-second grid obtained from the General Bathymetric 
Chart of the Oceans (available from website), 2) slope 
(created by calculating percent rise from the GEBCO 
30 arc-second grid), 3 and 4) seasonal mean bottom 
temperatures for summer and winter in degrees Cel¬ 
sius, and 5) the range in mean temperature between 
summer and winter, at a 0.1° resolution. Temperature 
data were obtained from the Global Ocean Reanalyses 
and Simulations (GLORYS; Mercator Ocean, available 
from website], which describe monthly mean ocean 
climate conditions at a 1/4° lat.xlong. resolution. We 
chose to work with GLORYS temperature data instead 
of the data that were collected from the trawl censors 
because GLORYS provided a complete and uniform 
coverage of the area that is more conducive to inter¬ 
polation. We limited our model to 5 predictor variables 
because the use of excessive variables can lead to over¬ 
fitting the data (Philips and Dudik, 2008). We selected 
variables that describe groundfish habitat (the bottom), 
and variables for which there were data for the entire 
study area. We interpolated the temperature layers by 
using ordinary kriging and 2001 through 2011 data. 
To prepare data for interpolation, we grouped monthly 
measures by season (summer: July-September, winter: 
January-March), calculated the mean annual tempera¬ 
tures for each season, and then assigned the 10-year 
average to each datum point. Finally, we incorporated 
regional variability in temperature by calculating the 
annual mean range in temperature by taking the dif¬ 
ference of the seasonal means, and averaging these 
values across the entire sampling period. 
We used 3 shapefiles (NAFO, available from website) 
to spatially classify the study area for spatial compari¬ 
sons: NAFO boundaries, the EEZ, and the Hague line 
(Fig. 1). The waters stretching from the northern limits 
of Baffin Island to Cape Hatteras are known as the 
NAFO Convention Area, which is divided into subar¬ 
eas, divisions, and subdivisions (Halliday and Pinhorn, 
1990). Although some of the areas of interest are for¬ 
mally referred to as “subdivisions,” we refer to all ar¬ 
eas as “NAFO divisions” throughout this analysis. We 
