Despite various inefficiencies described above the code is fairly fast. A test data set with 72 groups of 5 observations and 5 variables using the autoregressive correlation structure (one of the slowest) took about 5 seconds per iteration on a (shared) Sparcstation 10 or 6 seconds per iteration on a 75mHz Pentium-based PC running Linux. This data set had the same number of observations in each group; unbalanced data require an extra matrix inversion for each group that is smaller than the largest one and so would be slightly slower. The speed is only improved about 20% after compiling the code, presumably because it is mostly vectorised and so none of the inefficiency is easy to optimise away.