5 months and three intervals with breaks at 5 and 10 months. The break at 5 months has an unbalanced expected number of events at the maximum likelihood estimates (4 and 14) and takes much longer to converge using the EM algorithm than with a single break at 10 months. However, conclusions for the three covariates are the same regardless of the choice of interval. Stage of disease and CD4 lymphocyte are important predictors while dose is not. More advanced patients develop resistance faster. Patients with higher CD4 lymphocyte counts have a lower risk of developing resistance. Alternatively, as described by Rosenberg,21 one can choose a large number of breakpoints and use penalized likelihood techniques to smooth the estimates and avoid numerical problems associated with overparameterization. Caution would be needed in interpreting the asymptotic results if the number of intervals approached the number of events. Although further work needs to be done in choosing the best number of intervals, it seems based on the results presented here that the method is fairly robust to the choice of interval.

0 0

Post a comment