Bias reduction in the presence of informative censoring: application of the Cox model to multidrug-resistant tuberculosis cohort analyses.