Design of Partially Observable Survival Data

Authors: 
Giovanni Parmigiani
Duke University

Nov 30 1991

This paper considers life testing and survival analysis problem in which it is impractical or too expensive to obtain exact information about the failure time. Rather, information is obtained by observing the process at specified times. The design problem considered here is the choice of the sequence of inspection times that best balances costs with loss of information due to the inexact observation of failure times.

A general solution for the non-sequential case is derived by recursive dynamic methods. The case of exponential failures with unknown parameters is analyzed in detail and it is proved proof that the optimal inspection are robust with respect to the costs, that a control limit policy is optimal and that, if an unlimited number of inspection is optimal, times between inspection increase geometrically.

Keywords: 

Survival Data, Inspection schedule, kullback-leibler information, bayesian design of experiments

Manuscript: 

PDF icon 1992-34.pdf