OBJECTIVES: The objective of this analysis is to compare several covariance structures which are used in the modeling of longitudinal data. METHODS: A PUBMED search reveals is a steady increase in prospective observational studies over the past five years.  Repeated measures models are frequently used to analyze longitudinal data. For the purpose of these comparisons, a series of longitudinal datasets are simulated.  In order to facilitate comparisons with applications to longitudinal datasets involving utilities; the dependent variable in the simulation datasets is a continuous variable restricted to the support interval [0, 1].  The predictor variables include a set of categorical and continuous variables, including a time varying covariate.  Datasets with four different types of time dependence were compared (no time trend, log time trend, linear trend, exponential trend).  Models with the following covariance structures were evaluated: compound symmetry, unstructured, autoregressive, heterogeneous autoregressive, variance components and toeplitz.  Model comparisons were based upon Akaike information criteria (AIC) and the Bayesian information criteria (BIC). RESULTS: The preferred covariance structures for the dataset without a time trend were heterogeneous autoregressive (AIC) and unstructured (BIC).  The preferred covariance structure for the log trend dataset was unstructured (AIC and BIC).  The preferred covariance structures for the linear trend dataset were variance components (AIC) and heterogeneous autoregressive (BIC).  The preferred covariance structure for the exponential trend dataset was variance components (AIC and BIC). CONCLUSIONS: The unstructured covariance matrix is often the default choice for the covariance matrix for longitudinal models.  This model has the least number of assumptions and allows for the modeling of each patient individually.  However, the unstructured covariance structure requires the most degrees of freedom and in some cases the estimated covariance matrix does not converge.  In these cases, covariance structures such as variance components and heterogeneous autoregressive may present attractive options.