Identifiability and consistency in masking models for competing risks data, III.
Abstract
We consider the estimation of the marginal distributions with competing risks data and
masked failure cause (called MCR data). The problem arises naturally when the failure
time and the cause are independent, which is the case when the failure timess of different
components in the system are independent and follow a proportional hazards model. We investigate the
identifiability issue for two popular models in the literature on the MCR data. A finding of the paper
is that the parameters in the model proposed by Flehinger et al. (2001) is not identifiable unless additional
constraints are imposed on the parameters. Thus the likelihood-based statistical analyses such as that in
Mukhopadhyay (2006, p. 816) based on the latter model in the literature are erroneous. We also show that the marginal
MLE of the failure cause under the other model is
consistent and efficient under certain regularity assumptions. This is joint work with
Qiqing Yu.
