Health services research often involves multilevel data, with participants nested in families, clinicians, case managers, community care programs, HMOs, or countries. Longitudinal studies are also common, and studies may be both multilevel and longitudinal. Flexible and efficient analytic methods for data that are multilevel, longitudinal, or both are increasingly accessible to researchers. However, the implications of these methods for research design have not yet been well explored. As a result, researchers remain quite uninformed about crucial design decisions: How many persons should be sampled per cluster and how many clusters? How long should a longitudinal last and how frequent should the observations be? What impact will covariates have on the precision and power for studying key relationships? The overriding goal of the proposed project is to develop, test, and disseminate systematic ways of thinking about the planning of multilevel and/or longitudinal research. Planning will be based on a family of hierarchical models with random coefficients and related population-average statistical models. We propose to derive standard errors and optimal allocation algorithms and produce interactive, user-friendly software, enabling researchers to a) assess power, minimum detectable effects, and needed sample sizes given assumptions regarding costs, variance components, and covariates; and b) gauge the sensitivity of such planning decisions to uncertainties about these assumptions. The findings will be disseminated in articles, a monograph, and in two workshops that will introduce researchers to well-documented and freely available software.
|Raudenbush, S W; Xiao-Feng, L (2001) Effects of study duration, frequency of observation, and sample size on power in studies of group differences in polynomial change. Psychol Methods 6:387-401|