Ensemble models for data-poor assessment: accounting for uncertainty in life-history information

Abstract

Abstract Length measurements from fishery catch can be used in data-limited assessments to estimate important population parameters to guide management, but results are highly sensitive to assumptions about biological information. Ideally, local life history studies inform biological parameters. In the absence of reliable local estimates, scientists and managers face the difficult task of agreeing on fixed values for life-history parameters, often leading to additional uncertainty unquantified in the assessment or indecision defaulting to status-quo management. We propose an ensemble approach for incorporating life history uncertainty into data-limited stock assessments. We develop multivariate distributions of growth, mortality, and maturity parameter values, then use bivariate interpolation and stacking as an ensemble learning algorithm to propagate uncertainty into length-based, data-limited stock assessment models. Simulation testing demonstrated that stacking across life history parameter values leads to improved interval coverage over simple model averaging or assuming the parameter distribution means when the true life-history parameter values are unknown. We then applied the stacking approach for a U.S. Caribbean stock where the Scientific and Statistical Committee did not accept the assessment due to uncertainty in life history parameters. Stacking can better characterize uncertainty in stock status whenever life-history parameters are unknown but likely parameter distributions are available.