Under the current kidney allocation system in the United States, kidneys are allocated to patients primarily through a combination of tissue matching, sensitization level, and waiting time. However, due to recent trends in medicine and the shortfall of kidney supply, the current system fails to match donors and recipients well. In an eort to improve the allocation system, the United Network of Organ Sharing (UNOS) defined principle factors that would determine a new allocation policy. The most prominent factor is patients' potential remaining lifetime.

Estimating "potential remaining lifetime" is complicated for several reasons. First, the characteristics of candidates in the waitlist are different from kidney recipients, implying that the mortality of candidates does not represent the mortality that would have occurred among recipients, had they not received a transplant. Second, treatment methods of patients without a transplant have changed over the last two decades, making lifetime predictions less certain. Lastly, the lifetime model should extrapolate future survival beyond the duration of the data.

In this paper, we use a parametric Weibull Accelerated Failure Time (AFT) model to predict survival rates and show its advantage over common in terms of prediction accuracy. We also use data mining methods, and in particular, classication and regression trees, to tackle recipients' selection bias and bias caused by changes in medical treatment.