To estimate the optimal relative sample size of scientific observer data collected on Antarctic krill commercial fishing vessels, the relationship between statistical precision and sample size was estimated by using variance component analysis. Observer datasets from the Japanese krill fishery from 1995 to 2008 were analysed using a hierarchical Bayesian model. The models were composed of multistage cluster units (i.e. year, subarea, vessel, cruise and haul) based on a state–space model, separating biological process error in the population dynamics from fishery process as observation error. In both krill length and by-catch fish number, the parameters estimated by the Markov chain Monte Carlo (MCMC) method hardly show difference among years, subareas and vessels. The potent interaction effect between year and subarea suggests large spatio–temporal variability in the size structure of the krill population, which is presumably derived from large variability of recruitment causing difficulty in predicting krill population dynamics. Variances of observer datasets were calculated by the multistage sampling formula with the variance terms derived from the Bayesian model. For both krill length and by-catch fish number, vessel sample size shows marked effects on the coefficient of variation (CV), although haul sample size affects CV for only krill length data up to 10% haul coverage. These results suggest that data collection by scientific observers on board commercial vessels provides important information for the management of krill resources and the Antarctic ecosystem, while further discussion is needed about the optimal relative sample size to ensure the statistical precision required for the specific objective of a study that includes consideration of the cost of observer deployment.