지금까지 우리는 \(\hat{S}(t)\)가 Brownian motion로 converge in distribution하는 것에 대해 배웠고 이를 증명하기 위해 Martingale central limit theorem을 이용했다. 지금까지 우리의 관심사는 어떠한 하나의 population에서 온 sample data의 survival curve가 어떻게 수렴하는지 분포를 알고자 하는 것이었다면, 이번 챕터부터는 \(k\)개의 각기 다른 population이 존재할 때 그것들을 어떻게 비교하는지, 또 그러기 위해 어떠한 test를 사용하는지를 살펴볼 것이다.



Remark (Common Framework)

We have random variables

\(X_1^{(1)}, \ldots X_{n_1}^{(1)} \stackrel{\text{iid}}{\sim} F^{(1)}, \mbox{ }\mbox{ }\mbox{ }\mbox{ }Y_1^{(1)}, \ldots Y_{n_1}^{(1)} \stackrel{\text{iid}}{\sim} G^{(1)}\) (population 1)

\(X_1^{(2)}, \ldots X_{n_2}^{(2)} \stackrel{\text{iid}}{\sim} F^{(2)}, \mbox{ }\mbox{ }\mbox{ }\mbox{ }Y_1^{(2)}, \ldots Y_{n_2}^{(2)} \stackrel{\text{iid}}{\sim} G^{(2)}\) (population 2)

\(\mbox{ }\mbox{ }\mbox{ }\mbox{ }\mbox{ }\mbox{ }\mbox{ }\mbox{ }\mbox{ }\mbox{ }\mbox{ }\mbox{ }\mbox{ }\mbox{ }\mbox{ }\vdots\)

\(X_1^{(k)}, \ldots X_{n_k}^{(k)} \stackrel{\text{iid}}{\sim} F^{(k)}, \mbox{ }\mbox{ }\mbox{ }\mbox{ }Y_1^{(k)}, \ldots Y_{n_k}^{(k)} \stackrel{\text{iid}}{\sim} G^{(k)}\) (population \(k\))



Remark

We assume that we have \(k\) functions \(A_1,\ldots,A_k\) and for each \(h\) \((h=1,\ldots,k)\) we have a counting process \(N_h\) satisfying the Aalen model, i.e., \[ N_h(t)-\int_0^t Y_h(s)dA_h(s)\mbox{ }\mbox{ }\mbox{ }\mbox{ is a martingale}, \] where \(Y_h\) is predictable with respect to some filtration.

We want to test the null hypothesis \[ H_0: A_1=\cdots =A_k. \] We assume that the vector \((N_1\ldots,N_k)\) forms a \(k\)-dimensional counting process. That is, we assume that the counting processes never jump at the same time, and that the counting processes and predictable processes are all defined with respect to the same filtration.


For each \(h\), let \(\hat{A}_h\) be the Nelson-Aalen estimator of \(A_h\), and let \(\hat A\) be the Nelson-Aalen estimator based on the pooled sample, i.e., \[\begin{equation} \hat A(t)= \int_0^t \frac{I(Y_\cdot (s)>0)}{Y_\cdot (s)}dN_\cdot (s) \end{equation}\] where \[ N_\cdot (s)= \sum_{h=1}^k N_h(s)\mbox{ }\mbox{ }\mbox{ and }\mbox{ }\mbox{ } Y_\cdot (s)= \sum_{h=1}^k Y_h(s) \]


Assume \(H_0\) is true. Let \(A\) be the common value of the \(A_h\)’s. Let \[ M_h(t)=N_h(t)-\int_0^t Y_h(s)dA(s) \] and \[ M_\cdot(t)=N_\cdot(t)-\int_0^tY_\cdot(s)dA(s) \mbox{ }\mbox{ }\mbox{ }\mbox{ }\left( =\sum_{h=1}^k M_h(t) \right). \] Then, \(M_\cdot(t)\) is a martingale, since the sum of martingale is again a martingale. Hence, the common value \(A\) may be estimated by \(\hat A(t)= \int_0^t \frac{I(Y_\cdot (s)>0)}{Y_\cdot (s)}dN_\cdot (s)\) (처음에 배웠던, 양변에 derivative를 하여 구하는, Nelson Aalen공식 추론 방법으로 바로 구한다).

The general approach is to compare \(\hat A_h\) with \(\hat A\) through a weight process, then combine the \(k\) comparisons via quadratic form.



back