Surrogate Marker Paradox Toy

Treatment improves a marker, the marker predicts the outcome, and the real target still worsens.

HCGM toy

替代指标变好,真实目标变坏

\(A\) 明显提高 surrogate marker \(M\),并且 \(M\) 在组内正向预测 primary outcome \(Y\)。但 \(A\) 对 \(Y\) 的总效果为负,说明优化对象和真正查询对象错开了。

1200
2

SCM Sketch

A M Y positive direct negative path
\(M := \alpha_0 + \alpha_A A + U_M\)
\(Y := \beta_0 + \beta_M M + \beta_A A + U_Y\)
\(\alpha_A > 0,\ \beta_M > 0,\ \beta_A + \beta_M\alpha_A < 0\)

Effect Summary

+1.218 surrogate improves
-0.738 primary worsens
interactive projection

\(A \to M\) view

The treatment arm has a visibly higher surrogate marker.

替代指标查询

如果只问 A 是否提升 M,答案是肯定的;但这个查询还没有回答 Y 是否变好。

Surrogate marker paradox scatterplot Generated plots showing treatment effects on surrogate marker, treatment effects on primary outcome, and marker-outcome association within arms.

Paradox Diagnosis

A improves M, M predicts Y within each arm, but A lowers Y. The surrogate query and the primary-outcome query are not interchangeable.

Surrogate query 问 M,会得到好消息
\(\mathbb{E}[M \mid A=1]-\mathbb{E}[M \mid A=0]\)
+0.304
+1.521
+1.218

这个结果只证明 marker 变好,没有证明主要结果已经变好。

Target query 问 Y,答案反过来
\(\mathbb{E}[Y \mid A=1]-\mathbb{E}[Y \mid A=0]\)
+2.246
+1.509
-0.738

这个查询直接对准 primary outcome,因此暴露了负的总效果。

Path accounting 正向中介路径不够大
\(\beta_A + \beta_M\alpha_A\)
+0.936
-1.650
-0.714

HCGM 要把“优化哪个变量”和“真正关心哪个变量”分开保存。

Target Mismatch

\(M\) 是有用信号,但不是目标本身。只要 \(A \to Y\) 的直接负路径足够强, "marker 提升" 就不能自动替代 "outcome 改善"。

Canonical Figure

Generated three-panel surrogate marker paradox figure

HCGM Robustness

HCGM path robustness grid for the surrogate marker paradox

Source Discipline

  • Source: scripts/generate.py
  • HCGM evaluator: scripts/generate_hcgm_path_robustness.py
  • Data: data/surrogate_marker_synthetic.csv
  • Robustness data: data/hcgm_path_robustness.csv
  • Paper assets: tex/ and figures/
  • Projection: page/