Berkson Collider Bias Toy

Population independence, selected-sample association, and the collider path \(X \to S \leftarrow Y\).

HCGM toy

同一个总体,选中后换了答案

总体中 \(X\) 与 \(Y\) 近似无关;只看 \(S=1\) 的单位后, 二者出现负相关。这个 toy 保存的是选择条件改变了可见样本,而不是新增了 \(X \to Y\) 因果边。

1800
331

SCM Sketch

X Y S conditioning on S opens the collider
\(X := U_X\)
\(Y := U_Y\)
\(S := \mathbf{1}\{X + Y + U_S \ge \tau\}\)

Regression Summary

+0.029 near null
-0.528 opened collider
interactive projection

Population view

Before conditioning, \(X\) and \(Y\) are generated independently.

总体查询

在未选择的总体中,\(X\) 与 \(Y\) 的线性关系接近 0;黑色点只是满足 \(S=1\) 的那一部分。

Berkson collider bias scatterplot A generated scatterplot showing near-null population association and negative selected-sample association.

Paradox Diagnosis

总体关系接近 0;选择后关系为负。差异来自条件化 collider \(S\),而不是 \(X\) 和 \(Y\) 之间新增了因果边。

Population answer 未条件化 \(S\)
\(Y = a + \beta X\)
+0.029
+0.031
1800

这个查询保留所有单位,因此读到的是原始总体关系。

Selected answer 只看 \(S = 1\)
\(Y = a' + \beta' X \mid S=1\)
-0.528
-0.571
18.4%

选择条件让高 \(X\) 与高 \(Y\) 变成互相替代的入选方式,负相关由此出现。

Collider Path

\(X \to S \leftarrow Y\) 在总体中是关闭的路径;条件化 \(S\) 后, \(X\) 与 \(Y\) 在观测样本里相互牵连。HCGM 需要记录查询到底选择了哪些单位。

Canonical Figure

Generated two-panel Berkson collider bias figure

HCGM Robustness

Berkson HCGM selection robustness figure

Source Discipline

  • Source: scripts/generate.py
  • HCGM evaluator: scripts/generate_hcgm_selection_robustness.py
  • Data: data/berkson_synthetic.csv
  • Robustness: data/hcgm_selection_robustness.csv
  • Paper assets: tex/ and figures/
  • Projection: page/