Abstract
A new simulation-based approach to sample size determination for hierarchical linear models (HLM) which uses preliminary data is proposed. This approach acknowledges uncertainty in parameter values associated with the estimates obtained from a preliminary data. Taking samples repeatedly, via computer simulation, from the posterior predictive distribution for future observations, we can estimate statistical power and the mean range of confidence interval (CI) numerically, and find sample size with which a desired level of the power or the mean range of CI is achieved. The proposed approach is applicable to any specific model in HLM, as long as a computer program is appropriately adapted to the model.
