A generalized additive model (GAM) of blue shark, Prionace glauca, catch rates (catch per set) was fitted to data gathered by National Marine Fisheries Service (NMFS) observers stationed aboard Hawaii-based commercial longline vessels (N = 2010 longline sets) from March 1994 to December 1997. Its coefficients were then applied to the values of predictor variables, which were also contained in logbook records that described the remainder of fishery-wide effort during the study period (N = 41319 longline sets). The objective was to determine whether predictions generated by such a GAM could serve in lieu of observers on the large fraction of longline trips that do not carry an observer (approximately 95%). After deleting data considered false or inaccurate, much of which was associated with a small number of vessels, the relationship between catch rates as reported in logbooks and GAM predictions was expressed by log(e) (Y + 1) = 0.7952 log(e) (X + 1) - 0.0586 where Y is the catch rate (i.e., the number of blue shark caught per set) and X the GAM predictions (R-2 = 0.307, N = 40 243). Patterns of correspondence between logbook trends and GAM predictions were further refined by plotting the trends according to the type of fishing effort (e.g., tuna- or swordfish-directed). The highest mean catch rates reported in logbooks, the highest mean GAM predictions, and the greatest differences between the two occurred consistently in mid-year on swordfish trips. In contrast, mean values from logbooks and mean GAM predictions were closest for tuna-directed effort, but this reflected an order of magnitude reduction in the scale of catch rates rather than closely similar trends. A bootstrapping algorithm developed for the GAM yielded an estimate of 23.9% under-reporting for the study period, with approximate 95% prediction limits of 15.4-28.9%. We conclude that prediction with a GAM fitted to fishery observer data is a useful monitoring technique for the Hawaii-based commercial longline fishery. It allowed us: to gain insight into fleet-wide and individual logbook reporting practices, to estimate the relationship between logbook data and predicted values, to characterize bias in this relationship, and to identify patterns specific to each major sector of the fishery. (C) 2002 Elsevier Science B.V. All rights reserved.
