When Sample-Splitting Choices Matter in TMLE and AIPW: Cross-Fitting, Targeting, and Repetition
Abstract
Variants of targeted maximum likelihood estimation (TMLE) and augmented inverse-probability weighting (AIPW) with machine-learning nuisance models differ in which observations train them, how bias is corrected, and whether splits are repeated. In known-truth simulations of a binary-outcome average treatment effect, eight headline estimators shared each setting's library, inner folds and truncation bound; coverage differences were judged against a prespecified two-percentage-point margin. With a low-complexity library and good overlap (n = 3000), all but a half-sample baseline covered equivalently (0.928-0.939). With one random forest on the primary data-generating mechanism, estimators fitting the propensity in-sample undercovered severely (0.12-0.48), whereas cross-fitting both nuisances raised coverage to 0.91-1.00, two overcovering. Under near-positivity, every AIPW estimator covered better than every TMLE estimator with the unweighted fluctuation (0.907-0.917 versus 0.656-0.806 at the primary setting, all undercovering); looser truncation widened the gap. At the primary settings, pooled-versus-fold-specific and single-versus-double cross-fit constructions (involving four additional estimators) were usually equivalent; under extreme weights, pooled-targeting estimates were 1.7-1.8 times as dispersed as fold-specific ones. Repetition reduced partition dependence but made several intervals more conservative. Two real-data analyses illustrate these sensitivities. Results support cross-fitting both nuisances when empirical-process conditions cannot be justified, and assessing truncation and AIPW-versus-TMLE sensitivity under poor overlap.
Related
- Finite-sample guarantees for data-driven operator splitting methods via martingale inequalities
- Parallel Repetition for Entangled Games via Conditional Relative Entropy
- Selecting Repetition Counts Across Model Scales in Data-Constrained Pretraining
- The Sombor Index of Connected Graphs with Given Repetition Number and Cyclomatic Number
- A COVID-19 transmission model with optimal control integrating BERT with LHS and least-squares fitting for real-world consistency
BibTeX
@misc{karim2026when,
title = {{When Sample-Splitting Choices Matter in TMLE and AIPW: Cross-Fitting, Targeting, and Repetition}},
author = {M. Ehsan Karim and Rainie Fu},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.16192},
eprint = {2610.16192},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.16192},
}
RIS
TY - UNPB
TI - When Sample-Splitting Choices Matter in TMLE and AIPW: Cross-Fitting, Targeting, and Repetition
AU - M. Ehsan Karim
AU - Rainie Fu
PY - 2026
DA - 2026/10/12
PB - arXiv
DO - 10.48550/arxiv.2610.16192
UR - https://arxiv.org/abs/2610.16192
AB - Variants of targeted maximum likelihood estimation (TMLE) and augmented inverse-probability weighting (AIPW) with machine-learning nuisance models differ in which observations train them, how bias is corrected, and whether splits are repeated. In known-truth simulations of a binary-outcome average treatment effect, eight headline estimators shared each setting's library, inner folds and truncation bound; coverage differences were judged against a prespecified two-percentage-point margin. With a low-complexity library and good overlap (n = 3000), all but a half-sample baseline covered equivalently (0.928-0.939). With one random forest on the primary data-generating mechanism, estimators fitting the propensity in-sample undercovered severely (0.12-0.48), whereas cross-fitting both nuisances raised coverage to 0.91-1.00, two overcovering. Under near-positivity, every AIPW estimator covered better than every TMLE estimator with the unweighted fluctuation (0.907-0.917 versus 0.656-0.806 at the primary setting, all undercovering); looser truncation widened the gap. At the primary settings, pooled-versus-fold-specific and single-versus-double cross-fit constructions (involving four additional estimators) were usually equivalent; under extreme weights, pooled-targeting estimates were 1.7-1.8 times as dispersed as fold-specific ones. Repetition reduced partition dependence but made several intervals more conservative. Two real-data analyses illustrate these sensitivities. Results support cross-fitting both nuisances when empirical-process conditions cannot be justified, and assessing truncation and AIPW-versus-TMLE sensitivity under poor overlap.
ER -