= ...

B. Simulation Results We have undertaken a simulation study to examine the MLE of (J, and the sample size reestimation. We studied the following combination of situations: (J = 1, 3,5, 10,20; nl = 50 and 100. In each of the situations, J-tl = 0 and J-t2 varies so that the inverse of the coefficient of variation, A = (J-t2 - J-tl )/(J = fJ/(J, ranges from 0.20 to 0.50. In our experience, these wide ranges of (J and A cover most of the conditions in clinical trials involving normal approximation for sample size estimation. In addition, we also add A = 0 for the case when the null hypothesis is true. The choice of the constant c in (12) was 5.71, as discussed previously, representing situations of good estimation (for A = 0.35), overestimation (for A = 0.20 and A = 0), and underestimation (for A = 0.50) in the setting of initial values, respectively. The normal variates were generated by the NORMAL substitute of SAS (1990). One hundred replicates were run in each of the situations, which provided standard errors of the means (SEM) within the range of 0.010 (for (J = 1) to 0.221 (for (J = 20) in the estimation of a; or SEM/a = 0.01 for all the cases (see Table 1).