Research map / Probabilistic and scientific modeling / Treatment effects
It’s Hard to Be Normal: The Impact of Noise on Structure-agnostic Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows DML is minimax optimal for Gaussian or binary treatment in the partially linear model with black-box nuisances, but for non-Gaussian independent treatment noise constructs ACE, a cumulant-based estimator with r-th order insensitivity to nuisance error when the (r+1)-th treatment cumulant is nonzero.
Problem
Structure-agnostic estimation of the treatment effect theta_0 in the partially linear model given black-box nuisance estimates; whether DML is optimal depends on the treatment noise distribution.
Approach
Builds higher-order Neyman-orthogonal moment functions recursively from treatment cumulants estimated structure-agnostically from residuals, so that the error depends on the r-th power of nuisance error when the (r+1)-th cumulant of the treatment noise is nonzero; proves a Gaussian lower bound showing DML cannot be improved for Gaussian treatment.
Key results
- DML is shown minimax rate-optimal for Gaussian treatment noise, even with known variance, via a Gaussian treatment barrier lower bound.
- ACE achieves O(eps^r + n^{-1/2}) error when the (r+1)-th treatment cumulant is nonzero; in synthetic demand experiments, fifth-order ACE has the smallest bias and near-lowest RMSE.
- For binary treatment in the partially linear model, a new lower bound matches DML's rate up to constants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DML under Gaussian treatment noise | Partially linear treatment effect estimation · Gaussian treatment noise | structure-agnostic estimation error rate: DML is minimax rate-optimal; matches the minimax error rate whenever ${\epsilon}_1 = \gO(n^{-c}) for some positive const | No clear effect |
| Use DML for binary treatment | Partially linear treatment effect estimation · Binary treatment | minimax estimation error: matches the upper bound achieved by DML up to constant factors | No clear effect |
| Use ACE with order $r$ instead of DML | Partially linear treatment effect estimation · Independent non-Gaussian treatment noise | structure-agnostic estimation error rate: $\gO({\epsilon}^r+n^{-1/2})$ | Helped |
| Use the cumulant-based ACE estimator when $|\kappa_{r+1}|\geq\delta_{\mathsf{id},r}$ | ACE · Independent treatment noise with finite moments | structure-agnostic estimation error: ${\epsilon}_1^{r}{\epsilon}_2 + C_{\uptheta}{\epsilon}_1^{r+1} + 64C_{\mathsf{T}}^r\big(r^2C_{\mathsf{T}}+C_{\mathsf{Y}} | Helped |
| Use ACE with sub-Gaussian treatment noise | ACE · Independent sub-Gaussian treatment noise | structure-agnostic estimation error: ${\epsilon}_1^{r}{\epsilon}_2 + C_{\uptheta}{\epsilon}_1^{r+1} + 64(C_{\mathsf{g}}+\psi_{\upeta})^r\big(r^2(C_{\mathsf{g | Helped |
| Estimate noise cumulants from empirical residuals rather than noise moments | Structure-agnostic noise estimation · Cumulant estimation | minimax estimation error rate: $\gO(n^{-1/2}+{\epsilon}^r)$ versus $\Omega(n^{-1/2}+{\epsilon}^3)$ for $r=3$ and $\Omega(n^{-1/2}+{\epsilon}^2)$ for $r | Helped |
| Use ACE with Lasso-estimated high-dimensional linear nuisances instead of DML | Treatment effect estimation with sparse linear nuisances · High-dimensional linear nuisance model | maximum sparsity for order $n^{-1/2}$ estimation error: $o(n^{r/(r+1)}/\log p)$ versus $o(n^{1/2}/\log p)$ | Helped |
| Use fifth-order ACE rather than lower-order ACE estimators (ablation) | ACE · Synthetic demand estimation | total RMSE | Helped |
| Use fifth-order ACE rather than lower-order ACE estimators (ablation) | ACE · Synthetic demand estimation | bias | Helped |
| Use first-order ACE rather than higher-order ACE estimators (ablation) | ACE · Synthetic demand estimation | standard deviation | Helped |
| Use fifth-order ACE rather than first-order ACE (ablation) | ACE · Synthetic demand estimation | bias | Helped |
| Use fifth-order ACE rather than fourth-order ACE (ablation) | ACE · Synthetic demand estimation | estimation error | Helped |
Surprising findings
- DML is minimax optimal for Gaussian treatment noise, so distributional information about the treatment cannot improve structure-agnostic rates in that case.
- First-order ACE (which is DML) has the lowest standard deviation, while higher-order ACE reduces bias.
Limitations
- Main limitation is the assumed independence of treatment noise from covariates.
- Higher-order ACE needs the (r+1)-th treatment cumulant bounded away from zero; constants may grow with r.
- Experiments are small synthetic simulations only.
Experimental setup
Metrics: RMSE bias standard deviation confidence interval coverage
Baselines: double machine learning second-order orthogonal estimator
Builds on: double machine learning partially linear model structure-agnostic estimation
Closest papers by meaning
- Quantifying Aleatoric Uncertainty of the Treatment Effect: A Novel Orthogonal Learner (NeurIPS 2024 · similarity 0.87)
- Doubly robust identification of treatment effects from multiple environments (ICLR 2025 · similarity 0.87)
- Structural Causal Bottleneck Models (NeurIPS 2026 · similarity 0.86)
- Generalized Bayes for Causal Inference (NeurIPS 2026 · similarity 0.86)
- Impossibility of Distribution-Free Predictive Inference for Individual Treatment Effects (NeurIPS 2026 · similarity 0.86)
- An Analysis of Causal Effect Estimation using Outcome Invariant Data Augmentation (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.