Skip to main content
. 2013 Oct 8;19(12):3592–3606. doi: 10.1111/gcb.12335

Table 1.

Modelling techniques used in the study, model names used in article, and the list of environmental variables considered. See Materials and methods, Data S1 and Couce et al. (2012) for more details

Model Description Model name Variables
Boosted regression trees (BRT; Friedman, 2001) is a decision-tree-based statistical technique. A single tree is built by repeatedly splitting the data finding a cutoff value for one of the environmental variables so that the homogeneity of the resulting two groups is maximized. For BRT models, a sequence of trees (typically >1000) is produced, each grown on reweighted versions of the data, assigning ever-increasing weight to the cases misclassified by previous trees. The final prediction is obtained by the weighted average of predicted values across the sequence of trees.
BRTOPT 27 environmental variables:

  • Sea surface temperatures (SST; annual average, monthly maximum and minimum, annual range, standard deviation of monthly means, weekly maximum and minimum, and standard deviation of January and July means)

  • Salinity (annual average, and monthly maximum and minimum)

  • Nutrients (annual average phosphate and nitrate concentrations)

  • Irradiance (annual average, and monthly maximum and minimum), attenuation coefficient at 490 nm wavelength (annual average), and depth of light penetration (annual average, and monthly maximum and minimum)

  • Aragonite saturation (annual average ΩArag)

  • Dust level (annual average)

  • Cyclone activity (30-year average)

  • Current speed (annual average, and monthly maximum and minimum). For data sources see Couce et al. (2012)
BRT models were fitted in R, with version 1.6–3.1 of the gbm library (Ridgeway, 2007). Different values for the Tree Complexity (TC), Learning Rate (LR) and Bagging Fraction (BF) were considered, and the combination that minimized the predictive deviance was chosen in each case (for BRTOPT: TC = 10, LR = 0.01 and BF = 75; for BRTSIM: TC = 7, LR = 0.01 and BF = 50). See Couce et al. (2012) for more details.
BRTSIM Five variables containing the most relevant information according to BRTOPT's output:

  • SST (annual average, monthly minimum, and weekly maximum)

  • Depth of light penetration (monthly minimum)

  • Irradiance (annual average)

  • Plus aragonite saturation (annual average ΩArag)
Maximum entropy modeling (MaxEnt; Phillips et al., 2006) is a presence-only technique for the prediction of species geographic distributions, based on the environmental conditions of sites of known occurrence. Assumes environmental factors act as constraints on the distribution of a species and that within those constraints, the species will occupy the available habitat in a way that maximizes entropy.
MaxEntOPT Same variable set as BRTOPT
Our models were developed with Maxent version 3.3.3e, with default values for the convergence threshold (10−5) and maximum number of iterations (500). MaxEntSIM Same variable set as BRTSIM