We have developed a method for marginal analysis of the driving data that motivated this research. The method can be used to deal with correlation within subjects effectively. Two sources of within-subject correlation are considered: variability between subjects and serial correlation. We propose to address the variability between subjects by explicitly adjusting for subject as a fixed effect. To deal with serial correlation, we propose to use separated blocks, with a block-diagonal correlation matrix to account for as much correlation as possible within blocks, and a separation of suitable size to control the correlation between blocks. This procedure may be repeated randomly and the results will be synthesized using the multiple outputation technique. Model-based inference may be used to enhance the performance of the method, especially when the robust variance estimate performs poorly.
