Speaker: Bijit Roy, Department of Statistics, George Washington University
Time: Friday, January 22, 2010, 3:30-4:30pm
Location: Conference room in the Department of Statistics
Abstract:
Traditionally problems involving failure times have been dealt with in two ways: using survival analysis or using hitting times of stochastic processes. The first uses past lifetime data and uses it to predict the current lifetime. Whereas the second method claims that the failure occurs when some danger indicator reaches (marker) a threshold, and tries to predict when that will happen. The most important drawback of traditional survival analysis is that it ignores available marker information. Whereas the criticism about the second method is that though we observe the danger indicators, they often are just indicators and not the cause, e.g. rising body temp may be an indicator of danger, but it is not the cause of death.
In practice the actual cause/mechanism of failure is often not well-known. So we cannot afford to ignore the mechanism and focus only on the indicators. We hope to address such issues by recognizing that the cause and the indicators of failure are related but different : the cumulative hazard and the marker. We present a formulation of the problem using these two quantities, and give a brief overview of a few different approaches to deal with the problem.