Confidence intervals on catch estimates from a recreational fishing survey: a comparison of bootstrap methods

Bootstrap methods are often used for confidence intervals on recreational fish catch estimates, because they are relatively robust and straightforward to implement. Such data are typically highly skewed and zero-inflated, presenting difficulties for many estimation methods. However, bootstrap performance in many situations is not well understood. Inaccurate confidence intervals can cause management errors, and biased intervals can promote errors in one direction. Although the analyses originate from recreational fisheries data, the conclusions are generally applicable to similarly distributed data from other sources. Using simulation, non-parametric bootstrap confidence intervals (bootstrap normal, bootstrap percentile, hybrid, bootstrap-t, BC, and BCa) on catch rate and total catch estimates from a recreational fishing survey were compared. The intervals' coverage (proportion of times the 'true' mean fell within the confidence intervals) and relative bias were also compared. The bootstrap-t, using a resample size of slightly less than n/2, provided confidence intervals with the most correct coverage for both parameters. Intervals were biased, usually substantially, for all other methods, with the commonly used bootstrap percentile among the more biased methods