Module sensitivity 

Sensitivity analysis for omitted confounders (Chernozhukov, Cinelli, Newey, Sharma & Syrgkanis, 2022, “Long Story Short”), as in DoubleML.

If an unobserved confounder A was omitted, the estimate is biased by at most

|bias| ≤ |ρ| · √(c_y · c_d / (1 − c_d)) · √(σ² · ν²)

where σ² is the residual variance of the outcome, ν² the variance of the Riesz representer, c_y the share of the outcome’s residual variation explained by A, c_d the relative gain in the Riesz representer’s variation from A, and ρ the correlation between the two (adversarial at 1). DmlFit::sensitivity_analysis turns a scenario (c_y, c_d, ρ) into bounds on θ with confidence bounds, and computes robustness values: how strong confounding (c_y = c_d) must be to move θ (RV) or its confidence bound (RVa) to the null hypothesis.

Structs

SensitivityBenchmark

Benchmark values for (c_y, c_d, ρ) from omitting observed covariates.

SensitivityParams

A confounding scenario for DmlFit::sensitivity_analysis.

SensitivityParts

The per-observation ingredients of a model’s sensitivity elements, produced while estimating the score.

SensitivityResult

Result of DmlFit::sensitivity_analysis.

Enums

Nu2Parts

How a model’s ν² is formed from per-observation elements.