Get access

Bayesian belief network models for species assessments: An example with the Pacific walrus

Authors

  • James G. MacCracken,

    Corresponding author
    1. United States Fish and Wildlife Service, Alaska Regional Office, 1011 E Tudor Road, MS-341, Anchorage, AK 99503, USA
    • United States Fish and Wildlife Service, Alaska Regional Office, 1011 E Tudor Road, MS-341, Anchorage, AK 99503, USA.
    Search for more papers by this author
  • Joel Garlich-Miller,

    1. United States Fish and Wildlife Service, Alaska Regional Office, 1011 E Tudor Road, MS-341, Anchorage, AK 99503, USA
    Search for more papers by this author
  • Jonathan Snyder,

    1. United States Fish and Wildlife Service, Alaska Regional Office, 1011 E Tudor Road, MS-341, Anchorage, AK 99503, USA
    Search for more papers by this author
  • Rosa Meehan

    1. United States Fish and Wildlife Service, Alaska Regional Office, 1011 E Tudor Road, MS-341, Anchorage, AK 99503, USA
    Current affiliation:
    1. Alaska Ocean Observing System, 1007 W. 3rd Ave., Anchorage, AK 99501, USA.
    Search for more papers by this author

  • Associate Editor: Donaghy Cannon

Abstract

In 2008, the U.S. Fish and Wildlife Service was petitioned to list the Pacific walrus (Odobenus rosmarus divergens) under the U.S. Endangered Species Act (ESA). Research into stressors that may be negatively affecting walruses is incomplete. We developed a Bayesian belief network model structured around the ESA 5-factor analysis during a workshop attended by walrus and ESA experts to 1) elicit expert opinion on important stressors and their effects, 2) develop the model, and 3) develop and analyze plausible future scenarios. The listing factors and associated stressors were organized as sub-models, capturing the cumulative effects of the factors through model output, which was the probability of negative, neutral, or positive effects. We found that in a time-constrained workshop, the graphical display of Bayesian belief networks allowed for rapid development, assessment, and revision of model structure and parameters. We modeled up to 3 scenarios (most likely-, worst-, and best-case) for each of 4 time periods (recent past, contemporary, mid-century, and late-century). Model output for the recent past (reference condition) was consistent with observations and provided a baseline for comparison of the outcomes of other periods and scenarios; stressor effects became increasingly negative with time. However, scenario analyses indicated that mitigation of relatively few stressors could reduce the cumulative effects of the listing factors. Uncertainty in model output was lowest for the past but differed by only 7% among the other time periods. We used 4 types of sensitivity analyses to identify explanatory variables that had the greatest influence on model outcomes. Published 2012. This article is a U.S. Government work and is in the public domain in the USA.

Ancillary