The number and size of larval fish samples are usually determined arbitrarily, despite the influence of these decisions on the precision of abundance estimates and the ability to detect differences among population estimates. Review of the literature suggests that most surveys of larval fish are based on few (median, 4), large (median, 300 m3) samples. To evaluate current sampling designs, we developed a model, based on published data, to predict the variance in larval fish abundance among replicate samples. Inter-replicate variance (s2) is strongly related to mean abundance (x) as: log10s2= 0.19 + 1.74·log10x;r2= 0.93,P< 0.0001. This relationship was tested with an extensive data set collected in the Hudson River, New York (weekly samples over 14 years in 12 regions of the 250-km-long river), and was found to be general across environments, life history stages, and species. The model was not affected by sample volume, Our analysis shows that half of published studies estimated larval abundance with a coefficient of variation of the mean (SE/x) of 0.5 or more and could only detect order-of-magnitude differences among sites or time periods. Thes2/xequation provides guidelines to select the number and size of samples that should be taken to achieve a required level of precision and to detect a given difference among population means.