Cross-Fitted Survey-Weighted TMLE

An interactive companion to the results behind the paper

This site is the interactive companion to Cross-Fitted Survey-Weighted TMLE with Design-Based Variance for Causal Machine Learning (M. Ehsan Karim). It reproduces the paper’s simulation and NHANES results: every figure and table here is rebuilt from the manuscript’s frozen result files (multiple imputation, $m = 40$) and is labelled with its location in the paper — a main-text Table/Figure or a Web Table.

Important

The manuscript and its Web Appendix are the source of record. The numbers here are frozen snapshots that match the paper; if anything ever differs, the paper governs. Code and frozen data: https://github.com/ehsanx/Survey-TMLE-codes. Rendered 2026-06-25.

What’s reproduced here, and what’s in the paper

This site reproduces the paper’s results: the simulation suite (Web Appendix D), the four NHANES analyses with multiple imputation and sensitivity checks (Web Appendices E–F), the per-example survey-weighted Table 1s (Web Appendix E), and the variance-method benchmark and compute landscape (Web Appendix G). The full theory and proofs (Web Appendix A) are summarised on the Methods page rather than reproduced here — see the paper’s Web Appendix for the complete statements.

The three messages

⚖️ Weights → point estimate Ignoring the survey weights biases the point estimate — it targets a sample, not the population, quantity. (See NHANES.) 🧱 Clustered variance → honest CI Ignoring stratification and clustering understates uncertainty: the interval is too narrow even when the point estimate is right. (See Robustness.) 🔁 Cross-fitting → valid ML inference Flexible machine learning without cross-fitting under-covers once the learners cross a complexity boundary; PSU-level cross-fitting restores it. (See Simulation.)

Explore

The simulation — the full Web Appendix D: the headline coverage ladder, the complete results table, and every supporting experiment (correctly-specified control, the cross-fitting isolation, double robustness, the nuisance product rate, large-sample behaviour, type-I/power, the AIPW benchmark, the deployable CV-vs-CF comparison, and other estimands).
The NHANES applications — the population risk difference for four contemporary questions (E1–E4) under five arms, plus multiple imputation, the propensity-floor diagnostics, the single-fit-vs-cross-fit divergence, the covariate-set sensitivity checks, and an external AIPW cross-check.
Robustness & sensitivity — informative selection, the finite-population correction, a high-design-effect stress, and jackknife versus linearization variance.
Methods — a one-page, plain-language recap of the estimator and the design-based variance.

Note

The recommended estimator is Fully-Aware-CF (survey-weighted, cross-fitted at the primary-sampling-unit level). It is the arm the paper’s theory covers for machine-learning nuisances.

--- title: "Cross-Fitted Survey-Weighted TMLE" subtitle: "An interactive companion to the results behind the paper" --- This site is the interactive companion to **Cross-Fitted Survey-Weighted TMLE with Design-Based Variance for Causal Machine Learning** (M. Ehsan Karim). It reproduces the paper's **simulation** and **NHANES** results: every figure and table here is rebuilt from the manuscript's frozen result files (multiple imputation, $m = 40$) and is labelled with its location in the paper — a main-text Table/Figure or a Web Table. ::: {.callout-important} The manuscript and its Web Appendix are the **source of record**. The numbers here are frozen snapshots that match the paper; if anything ever differs, **the paper governs**. Code and frozen data: <https://github.com/ehsanx/Survey-TMLE-codes>. Rendered 2026-06-25. ::: ::: {.callout-tip collapse="true"} ## What's reproduced here, and what's in the paper This site reproduces the paper's **results**: the simulation suite (Web Appendix D), the four NHANES analyses with multiple imputation and sensitivity checks (Web Appendices E–F), the per-example survey-weighted Table 1s (Web Appendix E), and the variance-method benchmark and compute landscape (Web Appendix G). The full **theory and proofs** (Web Appendix A) are summarised on the [Methods](methods.qmd) page rather than reproduced here — see the paper's Web Appendix for the complete statements. ::: ## The three messages ```{=html} <div class="card-grid"> <a class="msg-card" href="nhanes.html"> ⚖️ Weights → point estimate Ignoring the survey weights biases the point estimate — it targets a sample, not the population, quantity. (See NHANES.) </a> <a class="msg-card" href="robustness.html"> 🧱 Clustered variance → honest CI Ignoring stratification and clustering understates uncertainty: the interval is too narrow even when the point estimate is right. (See Robustness.) </a> <a class="msg-card" href="simulation.html"> 🔁 Cross-fitting → valid ML inference Flexible machine learning without cross-fitting under-covers once the learners cross a complexity boundary; PSU-level cross-fitting restores it. (See Simulation.) </a> </div> ``` ## Explore - [**The simulation**](simulation.qmd) — the full Web Appendix D: the headline coverage ladder, the complete results table, and every supporting experiment (correctly-specified control, the cross-fitting isolation, double robustness, the nuisance product rate, large-sample behaviour, type-I/power, the AIPW benchmark, the deployable CV-vs-CF comparison, and other estimands). - [**The NHANES applications**](nhanes.qmd) — the population risk difference for four contemporary questions (E1–E4) under five arms, plus multiple imputation, the propensity-floor diagnostics, the single-fit-vs-cross-fit divergence, the covariate-set sensitivity checks, and an external AIPW cross-check. - [**Robustness & sensitivity**](robustness.qmd) — informative selection, the finite-population correction, a high-design-effect stress, and jackknife versus linearization variance. - [**Methods**](methods.qmd) — a one-page, plain-language recap of the estimator and the design-based variance. ::: {.callout-note} The recommended estimator is **Fully-Aware-CF** (survey-weighted, cross-fitted at the primary-sampling-unit level). It is the arm the paper's theory covers for machine-learning nuisances. :::