Think outside the grids: An objective approach to define spatial strata for catch and effort analysis

Abstract

A method is developed to objectively stratify a study area for standardizing catch per unit effort (CPUE) data for use in a generalized linear modeling framework. The method involves two steps: (1) subdividing the study area into small grids and applying a boosted regression tree method based on the full model to reconstruct a rough population pattern among grids and (2) applying a k-medoids algorithm to partition the grids into discrete strata based on the trend and magnitude of expected CPUE in each grid (time series clustering) or the spatial proximity and average CPUE in each grid (spatial clustering). The main advantage of this method compared to other area stratification approaches is that the shape of the resulting strata best matches the population structure as inferred from CPUE. This approach reduced bias in indices of abundance compared to ad-hoc methods. The variant of the area stratification method based on time series clustering performed best. However, this approach often performed as poorly as ad-hoc methods when there was a unidirectional shift in fishing locations over time because strata did not necessarily capture the shifts in fishing grounds over time. However, when the strata were selected to capture the shift in fishing grounds, both arbitrary and cluster-based method produced an index of abundance that was close to unbiased. Finally, the study showed that goodness of fit criteria (AICc or BIC) can be used to select the optimal number of strata for CPUE standardization, but this does not guarantee improved accuracy of the derived index of abundance compared to a pre-specified number of strata. The importance of conducting sensitivity analysis and the use of an ensemble approach is highlighted.