On the non-identifiability problem arising on the poly-Weibull model

The poly-Weibull model is a general family which arise on competing risk scenarios when there is no direct information about which risk was responsible for the failure. Its advantage over the single-Weibull model is to allow not only constant, increasing or decreasing hazard curves with zero or non zero asymptotes, but also nonmonotones ones, including bathtub-shaped. In this paper we consider the nonidentifiability problem, which arises when the shape parameters of a poly-Weibull model are close. Theoretical calculation of the information and correlation matrixes are used to assess when a poly-Weibull model is likely to be feasible. From the practical point of view, a graphical method, based on the total-time-on-test plot and its simulated envelop, is considered for detecting when a poly-Weibull model is likely to be identifiable. We also provide a general framework for constructing hypothesis tests for non-identifiability by using parametric bootstrap-based methods. We set up a simulation study and show that the bootstrap tests have desirable properties with respect to size and power. In some situations only a single Weibull model is enough for fitting the data. We determine the cost of estimating the parameters of a single-Weibull model if a bi-Weibull model is considered instead.