Contente and Rossi-Wongtschowski: Fish assemblages in the southeastern Brazilian Bight 
225 
Table 2 
Means, ranges, and measurement sources of oceanographic and spatial factors sampled during ECO- 
SAR IV-VII cruises and used as explanatory variables in the analysis conducted with the distance- 
based nonparametric linear modeling (available in DistLM in the PERMANOVA+ add-on for the soft- 
ware PRIMER 6). Eventual transformations before regression also are provided. Standard deviations 
fSDs) of the means are provided in parentheses. x=variable; na=not applicable; Z=local depth; Z a =mean 
depth of the fish aggregation at the water column; CTD=conductivity, temperature, and depth. 
Variable 
Mean (SD) 
Range 
Source 
Transformation 
Latitude (°) 
na 
22.6-28.6 
GPS 
Longitude (°) 
na 
41.7-48.6 
GPS 
Distance from shore (km) 
31.7 (28.5) 
1.0-132.0 
Satellite 
ln(x+l) 
Z local (m) 
47 (20) 
20-102 
Echosounder 
Z bottom (m) ; 
10 (14) 
0-53 
Echosounder 
ln(x+l) 
Za (m) 
24 (14) 
11-84 
Echosounder 
ln(x+l) 
Chlorophyll-a (mg/m 3 ) 2 
2.06 (2.00) 
0.30-6.30 
Satellite 
Bottom salinity 
35.38 (0.60) 
33.80-36.00 
CTD profiler 
Vx 
Surface salinity 
34.58 (1.68) 
28.70-36.60 
CTD profiler 
ln(x+l) 
Bottom temperature (T°C) 
17.90 (3.06) 
12.90-26.00 
CTD profiler 
Surface temperature (T°C) 
23.47 (2.07) 
18.70-28.60 
CTD profiler 
^Distance from the seafloor to the bottom of the net. 
2 Satellite-derived data for surface levels of chlorophyll-a averaged over the period of the cruise were taken 
from the Moderate Resolution Imaging Spectroradiometer ( MODIS ) data set available from NASA. Avail- 
able from website, accessed April 2015). Data are available with 8-day temporal resolution and 4-km spa- 
tial resolution. Therefore, we used data averaged over the period of the cruise. 
was the same response matrix used in PERMANOVA, 
but it was reduced to 63 tows and 31 species because 
1) the data from the ECOSAR III cruise were not con- 
sidered because we did not have satellite chlorophyll- 
a data available for the period of this cruise, and 2) 
species occurring in <3% of the tows of the resulting 
matrix were excluded. The second matrix was one of 
environmental and spatial variables (matrix of predic- 
tors), containing 63 samples and 11 variables (Table 
2) . The abundance data of the response biomass matrix 
were previously transformed to scale reduction. 
This removal of data from the ECOSAR III cruise 
did not affect the relation between the conclusions 
drawn from the results of PERMANOVA and SIMPER 
and those drawn from the analyses described below. 
This relation was not affected because we obtained 
practically the same results (not shown) in repeating 
the PERMANOVA and SIMPER analyses without the 
data from the ECOSAR III cruise. 
The relationship between species and their environ- 
ment was assessed by using the distance-based, non- 
parametric linear modeling available in DistLM (An- 
derson et al., 2008) in the PERMANOVA+ add-on (vers. 
1.0.1) for the software PRIMER 6, vers. 6.1.11. This 
routine fits predictor variables (matrix of predictors) 
to a set of response variables (the response biomass 
similarity matrix) on the basis of any resemblance 
measure (Anderson et al., 2008). Unlike the multivari- 
ate regressions where multinormality is assumed, this 
routine tests the null hypothesis of a no species-envi- 
ronment relationship through permutation (Anderson 
et al., 2008), making it quite suitable for our data that 
were not normally distributed. Moreover, the flexibility 
of the routine to run on any resemblance measure al- 
lowed the use of the Bray-Curtis index (Anderson et 
al., 2008), which best described the distribution of spe- 
cies abundance in our data (Legendre and Legendre, 
1998; Clarke and Warwick, 2001). The significance of 
test results was based on 9999 permutations. 
To meet the assumption of linearity of model, (An- 
derson et al., 2008), before running DistLM, we checked 
for outliers and skewness in draftsman plots among the 
predictors, transforming them when necessary (Table 
1; Legendre and Legendre, 1998). We also removed 
one redundant predictor when correlated (Pearson’s 
correlation coefficient [r]) at r>0.70 (Dormman et al., 
2013) with another one. We used information criteria 
to choose the most parsimonious model, in other words, 
the simplest model (the one with the lowest number 
of predictors) with enough suitable explanatory power 
(Burnham and Anderson, 2002). We based our selec- 
tion on both the corrected Akaike information criterion 
(AIC c ) and the Bayesian information criterion (BIC) 
(Burnham and Anderson, 2002). The model selection 
procedure best, available in the software PRIMER 6, 
provided the 20 best models. Balancing the severity of 
the BIC with the flexibility of the AICc to include vari- 
ables in models has been shown to be a robust selection 
procedure (for more details, see Anderson et al., 2008). 
The following steps were conducted to select a model: 
the number of variables in the model was defined as 
the number of variables most frequently found among 
