53
the sensitivity analysis. In some situations, for instance, when censoring percentages
are small and/or balanced across covariate groups, the estimated regression param-
eters and their confidence intervals may not change much. As a result, the typical
assumption of independent censoring may still be valid and traditional approach in
survival analysis can be used. In such situations, a sensitivity analysis can help us
decide how much confidence can be put in the results of analysis.
3.4 Simulation study of dependent censoring
In order to evaluate the method proposed in Section 3.3, we conducted a simulation
study, through which we compared the results from correctly assuming dependent
censoring with the ones from the false assumptions (traditional Coxph method).
We run 300 simulations with sample size 200. And for 10 simulations, we run 30
times of bootstrapping each to get the covariance matrices. The statistical software
package R is used to implement the simulation throughout this dissertation.
To generate data, we assume two covariates: B = 0 or 1 with equal probability and
centered Age which follows Uniform (-5,5). Assuming Weibull distributions, which
is a popular distribution to model the time to event data, we specified the marginal
distributions for events and informative censoring times T and C, respectively, by the
following survival functions.