Multiplicative Bias Correction in Kernel Hazard Estimation (Q4255149)

One observes \(n\) individuals. Let \(N_{i}\), \(i=1,2,\ldots,n\), count the observed failures for the \(i\)-th individual in the time interval \([0,1].\) Here \(N_{i}\) is a one-dimensional counting process. The random intensity is defined as \(\lambda_{i}(t)=\alpha(t)y_{i}(t)\) with no restriction on the functional form of \(\alpha(\cdot)\), and \(y_{i}\) is a predictable process taking values in \(\{ 0,1\}\) indicating (by the value 1) when the \(i\)-th individual is under risk. The author considers the multiplicative bias corrected estimator \(\hat a(t)= \tilde a(t)\hat g(t)\), where \[ \tilde a(t)=\sum\limits_{i=1}^{n} \int [K_{b}(t-s)/y^{(n)}(s)]dN_{j}(s), \] \(K\) is a probability density function symmetric about zero and Lipschitz continuous on its support \([-1,1]\), \[ y^{(n)}(s)=\sum\limits_{i=1}^{n}y_{i}(s),\quad \hat g(t)=\sum\limits_{i=1}^{n} \int [K_{b}(t-s)/\tilde a(s)y^{(n)}(s)]dN_{i}(s). \] Under some assumptions the author proved that \(\hat a(t)-a(t) \Rightarrow N(b^4 {\bar m}_{t},(nb)^{-1}\bar{v}_{t}^2)\), where \[ m_{t}=k_1^2 a(t)/4(a''/a)''(t),\quad v_{t}=a(t)/\gamma(t)\int\limits_{-1}^1 \Gamma_{k}(v) dv, \] \[ \Gamma_{k}(u)=2K_{b}-K_{b}\ast K_{b}(u),\quad k_1=\int\limits_{-1}^1 v^2 K(v) dv,\quad k_2=\int\limits_{-1}^1 K^2(v) dv, \] where \(\gamma(t)\) is a certain function. The author presents a simulation study which indicates good performance of the method. An application to the development of Danish mortality is provided.