Final Technical Report: - Southwest Fisheries Science Center - NOAA
Final Technical Report: - Southwest Fisheries Science Center - NOAA
Final Technical Report: - Southwest Fisheries Science Center - NOAA
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
were built separately for each of these ecosystems with a model that was built using pooled data<br />
from both ecosystems. Modeling methodology followed Redfern et al. (2008), but only the<br />
60km resolution was used.<br />
3.5 Model Selection<br />
Model validation using an independent data set is an integral part of building robust<br />
cetacean-habitat models (Forney 1997 and 2000, Becker 2007). In this analysis, final models for<br />
the CCE and the ETP were selected using a two-part process in which models were initially built<br />
using stepwise variable selection based on the available SWFSC survey data through 2003.<br />
Candidate models were then evaluated in terms of their predictive capabilities when applied to<br />
data from the novel 2005 (CCE) and 2006 (ETP) SWFSC cetacean surveys (see 3.1.1 Marine<br />
Mammal Surveys). Predictions and overall model performance were compared to identify the<br />
best models.<br />
A collection of quantitative and qualitative methods were used to compare models.<br />
Average squared prediction error (ASPE) was used to assess each model’s prediction accuracy<br />
across all segments (n) within the entire study area, where<br />
Prediction accuracy was addressed in a spatial context using ratios of observed to<br />
predicted number of sightings (for the encounter rate models) or group size within each<br />
geographic stratum. These geographic strata were defined to be large enough to encompass a<br />
sufficient number of observations for a meaningful comparison of model predictions, yet<br />
environmentally distinct in terms of the biological and physical processes that determine habitat.<br />
In addition to examining the observed-to-predicted ratios themselves, we computed the sum of<br />
absolute deviations of the observed-to-predicted ratios, defined as<br />
observed<br />
1 ,<br />
predicted<br />
where the sum is taken over all geographic strata used in model evaluation. For both the ASPE<br />
and observed-to-predicted ratio computations, the Beaufort sea state variable was set to the<br />
observed value to generate encounter rate and group size predictions. Explained deviance, the<br />
likelihood analogue of explained variance, was used to assess each model’s fit to the assumed<br />
distribution for the data. Model complexity was evaluated by examining the number of predictor<br />
variables selected and their associated degrees of freedom, in conjunction with visual inspection<br />
20