中介分析笔记

High-dimensional mediation analysis in survival models (Chengwen Luo, 2020, PLOS COMPUTATIONAL BIOLOGY)

high-dimensional mediators，time-to-event outcome data.

Mediation models:
$$\begin{gathered} {\lambda }_i\left(t\mid X_i,M_i\right)={\lambda }_0(t)\exp \left\{\gamma X_i+{\theta }^T{Z}_i+{\beta }_1{M}_{1i}+\cdots {\beta }_p{M}_{pi}\right\} \\ {M}_{ki}=c_k+{\alpha }_k{X}_i+{\vartheta }^T{Z}_i+e_{ki},k=1,2,\cdots ,p \end{gathered}$$Main assumptions:

approximate the counterfactual outcome defined as log hazard as follows

$\begin{array}{rl}& \log \lambda \left(T\left(x,M_1^{x^{\ast }},\cdots ,M_p^{x^{\ast }}\right);t\mid Z\right)\\ & =\log {\lambda }_0(t)+{\theta }^{T}Z+\frac{1}{2}{\sigma }_{W_{\beta }}^2+{\beta }_1\left(c_1+{\vartheta }^{T}Z\right)\cdots +{\beta }_p\left(c_p+{\vartheta }^{T}Z\right)\\ & +\gamma x+\left({\alpha }_p{\beta }_p+\cdots +{\alpha }_p{\beta }_p\right)x^{\ast }\end{array}$
where ${\sigma }_{W_B}^2=\sum {\beta }_k^2{\sigma }_{M_k}^2$.

direct effect and total indirect effect on og hazard ratio as
Δ X → Y = log ⁡ λ ( T ( x ∗ , M 1 ( x ) , ⋯   , M p ( x ) ) ; t ∣ Z ) − log ⁡ λ ( T ( x , M 1 ( x ) , ⋯   , M p ( x ) ) ; t ∣ Z ) = ( x ∗ − x ) γ Δ X → M → Y = log ⁡ λ ( T ( x ∗ , M 1 ( x ∗ ) , ⋯   , M p ( x ∗ ) ) ; t ∣ Z ) − log ⁡ λ ( T ( x ∗ , M 1 ( x ) , ⋯   , M p ( x ) ) ; t ∣ Z ) = ( x ∗ − x ) ( α 1 β 1 + ⋯ + α p β p ) , \begin{aligned} \Delta_{X \rightarrow Y} &=\log \lambda\left(T\left(x^{*}, M_{1}(x), \cdots, M_{p}(x)\right) ; \mathrm{t} \mid \mathrm{Z}\right)-\log \lambda\left(T\left(x, M_{1}(x), \cdots, M_{p}(x)\right) ; t \mid Z\right) \\&=\left(x^{*}-x\right) \gamma \\ \Delta_{X \rightarrow M \rightarrow Y}&=\log \lambda\left(T\left(x^{*}, M_{1}\left(x^{*}\right), \cdots, M_{p}\left(x^{*}\right)\right) ; \mathrm{t} \mid \mathrm{Z}\right)-\log \lambda\left(T\left(x^{*}, M_{1}(x), \cdots, M_{p}(x)\right) ; \mathrm{t} \mid \mathrm{Z}\right) \\ &=\left(x^{*}-x\right)\left(\alpha_{1} \beta_{1}+\cdots+\alpha_{p} \beta_{p}\right), \end{aligned} ΔX→Y​ΔX→M→Y​​=logλ(T(x∗,M1​(x),⋯,Mp​(x));t∣Z)−logλ(T(x,M1​(x),⋯,Mp​(x));t∣Z)=(x∗−x)γ=logλ(T(x∗,M1​(x∗),⋯,Mp​(x∗));t∣Z)−logλ(T(x∗,M1​(x),⋯,Mp​(x));t∣Z)=(x∗−x)(α