文章目录
这一章节主要介绍了如何在生存模型中进行因果分析.
17.1 Hazards and risks
Administrative censoring: 指开始追踪和停止追踪的时间有差异.
设生存时间为 T T T, 并让 k k k表示第 k k k个时间单位.
生存概率: P r [ T > k ] \mathrm{Pr}[T > k] Pr[T>k].
risk: 1 − P r [ T > k ] = P r [ T ≤ k ] 1 - \mathrm{Pr}[T > k] = \mathrm{Pr}[T \le k] 1−Pr[T>k]=Pr[T≤k].
hazard: P r [ T = k ∣ T > k − 1 ] \mathrm{Pr}[T=k|T>k-1] Pr[T=k∣T>k−1].
显然生存概率是单调递减的, 而risk是单调递增的, 但是hazard并不绝对.
17.2 From hazards to risks
用
D
k
=
1
D_k=1
Dk=1表示个体在第
k
+
1
k+1
k+1个时间单位状态为死亡(0则表示生存).
则
P
r
[
D
k
=
0
]
=
∏
m
=
1
k
P
r
[
D
m
=
0
∣
D
m
−
1
=
0
]
.
\mathrm{Pr}[D_k = 0] = \prod_{m=1}^k \mathrm{Pr}[D_m=0|D_{m-1}=0].
Pr[Dk=0]=m=1∏kPr[Dm=0∣Dm−1=0].
注:
P
r
[
D
m
=
0
∣
D
m
−
1
=
0
]
=
P
r
[
D
m
=
0
]
P
r
[
D
m
−
1
=
0
]
P
r
[
D
1
=
0
]
=
1.
\mathrm{Pr}[D_m=0|D_{m-1}=0] = \frac{\mathrm{Pr}[D_m=0]}{\mathrm{Pr}[D_{m-1}=0]} \\ \mathrm{Pr}[D_1 = 0] = 1.
Pr[Dm=0∣Dm−1=0]=Pr[Dm−1=0]Pr[Dm=0]Pr[D1=0]=1.
且容易证明:
P
r
[
D
k
=
0
]
=
P
r
[
T
>
k
]
P
r
[
D
k
=
1
]
=
P
r
[
T
≤
k
]
P
r
[
D
k
=
1
∣
D
k
−
1
=
0
]
=
P
r
[
T
=
k
∣
T
>
k
−
1
]
.
\mathrm{Pr}[D_{k} = 0] = \mathrm{Pr}[T > k] \\ \mathrm{Pr}[D_{k} = 1] = \mathrm{Pr}[T \le k] \\ \mathrm{Pr}[D_{k} = 1|D_{k-1} = 0] = \mathrm{Pr}[T = k|T > k-1].
Pr[Dk=0]=Pr[T>k]Pr[Dk=1]=Pr[T≤k]Pr[Dk=1∣Dk−1=0]=Pr[T=k∣T>k−1].
这相当于, 只要我们建模出hazards, 就能够根据上面的公式推导出risks.
P
r
[
D
k
=
0
∣
A
=
0
]
\mathrm{Pr}[D_k=0|A=0]
Pr[Dk=0∣A=0]
采用同样的方法.
17.3 Why censoring matters
如果考虑administrative censoring, 则我们实际上需要考虑
P
r
[
D
k
c
ˉ
=
0
ˉ
=
0
∣
A
=
a
]
,
c
ˉ
=
(
c
1
,
c
2
,
⋯
,
c
k
e
n
d
)
.
\mathrm{Pr}[D_k^{\bar{c}=\bar{0}}=0|A=a], \bar{c} = (c_1, c_2, \cdots, c_{k_{end}}).
Pr[Dkcˉ=0ˉ=0∣A=a],cˉ=(c1,c2,⋯,ckend).
其等价于
∏
m
=
1
k
P
r
[
D
m
=
0
∣
D
m
−
1
=
0
,
C
m
=
0
,
A
=
a
]
.
\prod_{m=1}^k \mathrm{Pr}[D_m=0|D_{m-1}=0, C_m=0, A=a].
m=1∏kPr[Dm=0∣Dm−1=0,Cm=0,A=a].
注意, 这是因为
C
m
=
0
C_m=0
Cm=0表示
C
m
′
=
0
,
∀
m
′
≤
m
.
C_{m'} = 0, \quad \forall m' \le m.
Cm′=0,∀m′≤m.
所以在这种情况下, 我们需要对
P
r
[
D
m
=
0
∣
D
m
−
1
=
0
,
C
m
=
0
,
A
=
a
]
.
\mathrm{Pr}[D_m=0|D_{m-1}=0, C_m=0, A=a].
Pr[Dm=0∣Dm−1=0,Cm=0,A=a].
17.4 IP weighting of marginal structural models
和普通的IP weighting一样我们需要估计
P
r
^
[
A
=
1
∣
L
]
.
\widehat{\mathrm{Pr}} [A=1|L].
Pr
[A=1∣L].
17.5 The parametric g-formula
需要估计
P
r
[
D
m
+
1
=
0
∣
D
m
=
0
,
L
=
l
,
A
=
a
]
.
\mathrm{Pr}[D_{m+1}=0|D_m=0, L=l, A=a].
Pr[Dm+1=0∣Dm=0,L=l,A=a].
17.6 G-estimation of structural nested models
Fine Point
Competing events
censoring 的一种特殊情况.