Efficiency and bootstrap in the promotion time cure model (Q2405162)

The authors study various aspects of a semiparametric promotion time cure model in a thorough and mathematically rigorous way. \textit{D. Zeng} et al. [J. Am. Stat. Assoc. 101, No. 474, 670--684 (2006; Zbl 1119.62381)] gave a serious attempt to provide a rigorous theoretic study, but with unnecessarily complicated calculations on the nonparametric maximum likelihood estimator (NPMLE) of the vector of regression coefficients. The estimation of the variance of the estimator lacks clarity, alternative methods to do inference are not studied, and efficiency is only shown to be true for the estimator of the parametric component of the model. The paper under review illustrates that the NPMLE can be computed by maximizing a certain criterion function over a set of dimension \(d+1\) (for the dimension \(d\) of the covariates), expresses weak consistency and weak convergence of the NPMLE and an explicit formula of the asymptotic variance of the parametric component, and shows the efficiency of the NPMLE of both the parametric and nonparametric components. Section 2 sets up the model of the promotion time cure model as \[ S_0(t|x) =\exp \left(- \eta (\beta_0^Tx)\Lambda_0(t)\right), \] for the conditional survival function \(T\) given \(X=x\) and the vector \(\beta_0\in\mathbb R^d\) of regression coefficients and an improper continuous cumulative hazard function \(\Lambda_0\) with \(\Lambda_0(0)=0\). By fixing arbitrarily that \(\Lambda_0(+\infty) = 1\), the paper defines the model \[ {\mathcal P} = \left \{(x, t)\to S_{\beta, \Lambda}(t|x) = \exp \left(- \eta (\beta^Tx)\Lambda(t)\right): (\beta, \Lambda )\in \tilde{\Theta}\right \}, \] where this model naturally allows one to handle the situation where the event \(T=+\infty\) with positive probability since \(S_{\beta, \Lambda}(+\infty |x) >0\). Define a random variable \(C\in \mathbb{R}^+\) as censoring time such that the only observables are \(Y=\min (T, C)\) and \(\delta = 1_{T\leq C}\). The survival time \(T\) and the censoring time \(C\) are conditionally independent. Under certain assumptions, the model \({\mathcal P}\) is identifiable by Proposition 1, i.e., the model \({\mathcal P}\) is necessarily consistent guaranteed by the MLE. Section 3 gives precisely the NPMLE and its logarithmic scale, where the score operator \(B_1(\beta, \Lambda)\) for \(\beta\) (the differential with respect to \(\beta\) to the logarithm of the NPMLE) is given by \[ B_1(\beta, \Lambda)[a] = (d_{\beta} X^T a)\bigg (\delta - \int_0^{+ \infty} \eta (\beta^T X) (\Delta 1_{Y\geq u} + (1-\Delta))d\Lambda (u)\bigg ). \] The existence of \((\beta, \Lambda)\) follows from the facts that the logarithm scale equals minus infinity if \(\Lambda\) does not jump at one of the uncensored observations and that an increase of \(\Lambda\) outside the set of finite observed survival times always reduces the value of the preceding or forthcoming jump and the value of the likelihood. By a Lagrange procedure, the authors obtain the solution for \(\hat{\Lambda}_{\beta}\) and then use \(\hat{\Lambda}_{\beta}(+\infty)=1\) in the likelihood to obtain \(\hat{\beta}\). Section 4 gives efficiency bounds. The efficient score function for \(\beta\) is \(\tilde{B}_1\) the residual of the score \(B_1\) after projecting it onto the nuisance tangent space generated by \(B_2(\beta_0, \Lambda_0)[h]\). By the uniqueness of the orthogonal projection on closed subsets of Hilbert spaces, there exists a unique \(h_0\) such that for every \(h\) in the closed subset of Hilbert spaces \[ E\big [(B_1-B_2)(\beta_0, \Lambda_0)[h_0] B_2(\beta_0, \Lambda_0)[h] \big ] =0. \] Section 5 presents the weak convergence of the asymptotic properties in Theorem 2, the weak convergence of \(n^{1/2}((\hat{\beta}, \hat{\Lambda}) - (\beta_0, \Lambda_0))\) to a tight Gaussian process in Theorem 3, and shows that \((\hat{\beta}, \hat{\Lambda})\) is efficient. Therefore the asymptotic distributions of the proposed estimators cannot be directly used because they depend on many unknown quantities. Section 6 is to consider the weighted bootstrap to approach the asymptotic distribution. Theorem 5 shows that the bootstrap estimator \(n^{1/2}((\hat{\beta}^*, \hat{\Lambda}^*) - (\hat{\beta}, \hat{\Lambda}))\) follows the same asymptotic law. Section 7 gives the performance of the estimators for the model \(S(t| Z_1, Z_2) = \exp ( - \exp (\beta_0 + \beta_1 Z_1 +\beta_2 Z_2)\Lambda (t))\) for a uniformly distributed random variable \(Z_1\) on \([\alpha, \alpha +1]\) and a Bernoulli random variable \(Z_2\) taking the values 0 and 1 with equal probability. The objective is to investigate the effect of the sample size \(n\), the distribution \(\Lambda\) and the average amount of censoring and cure. Compared to the Bayesian bootstrap, the multinomial weights lead systematically to slightly conservative and wider confidence intervals. The asymptotic normal approximation gives satisfactory results in Table 3. Section 8 is devoted to all technical proofs of Proposition 1 and Theorems 2--5. The proof of Theorem 3 is completed by applying the well-known weak convergence theorem for \(Z\)-estimators to give the weak convergence of the NPMLE. The efficiency is obtained by using a suitable characterization of the influence function. The appendix section gives several technical lemmas for the proofs used in Section 8.