Estimation of the subsurvival function for time-to-tumor in survival/sacrifice experiments (Q1186023)

The paper's aim is modelling, estimation and study of the asymptotic properties of the incidence rate of a tumor in animal carcinogenic bioassay when the time-to-onset of the tumor is unobservable or indetectable before death. The approach uses random sacrifices as a method for obtaining data that will then be used together with survival data, to estimate different functions in the model as e.g. disease resistence function \(Q(t)\), and survival function \(P(t)\). The essential data needed from a single animal are: its group, age at death, mode of death (natural or sacrifice) and a list of the tumors present at death. Let \(X_ i\) denote the time until the first tumor onset or until time- to-death without tumor, whichever occurs first for animal \(i\). Let \(K=1\), if \(X_ i= \)time-to-first-tumor and \(K=2\), if \(X_ 1= \)time-to-death- without-tumor. Let \(X_ 2\) denote the time-to-death with tumor if \(K=1\) and \(X_ 2=+\infty\) if \(K=2\). Let \(Y_ i\) be the random time of sacrifice. Then the subdistribution functions for time-to-tumor \(F_ 1\) and time-to-death-without-tumor \(F_ 2\) are defined as \(F_ k(t)= \operatorname{Prob} (X_ 1\leq1,\;K=k)\). The subsurvival function for time-to-tumor is \(S_ 1(t)=1-F_ 1(t)\) and it represents the probability that an animal is either alive and tumor free at time \(t\) or dead without tumor at \(t\). A new closed-form nonparametric estimator is proposed for the subsurvival function \(S_ 1\) combining sacrifice data and data from natural deaths of animals without assumptions concerning the lethality of the tumor and without specification of cause of death. This is based on a four-variate counting process (recording the number of deaths with and without tumor and summarizing the sacrifice experience with and without tumor) and on a process \(R(t)\) recording the number of animals at risk time \(t\). Approximate Kaplan-Meier estimates \(P\) and \(F_ 2\) are obtained by using stochastic integrals with respect to martingales for asymptotic results. Proofs of consistency and asymptotic normality of \(S_ 1\) are given.