Efficient Targeted Maximum Likelihood Estimators for Two-Phase Design Problems
| Authors | Sky Qiu et al. |
| Year | 2026 |
| Field | AI / ML |
| arXiv | 2602.24131 |
| Download | |
| Categories | stat.ME, stat.ML |
Abstract
In a typical two-phase design, a random sample is drawn from the target population in phase 1, during which only a subset of variables is collected. In phase 2, a subsample of the phase-1 cohort is selected, and additional variables are measured. This setting induces a coarsened data structure on the data from the second phase. We assume coarsening at random, that is, the phase-2 sampling mechanism depends only on variables fully observed. We review existing estimators, including the generalized raking estimator and the inverse probability of censoring weighted targeted maximum likelihood estimation (IPCW-TMLE) along with its extensions that also target the phase-2 sampling mechanism to improve efficiency. We further introduce a new class of estimators constructed within the TMLE framework that are asymptotically equivalent.
Engineering Breakdown
Plain English
This paper addresses the statistical and computational problem of efficient estimation in two-phase sampling designs, where a large cohort is screened in phase 1 with limited variables, and a smaller subsample is intensively measured in phase 2 with additional expensive variables. The authors review existing approaches like generalized raking and inverse probability of censoring weighted targeted maximum likelihood estimation (IPCW-TMLE), then introduce a new class of estimators built within the TMLE framework that jointly targets both the phase-2 sampling mechanism and the estimation target to improve efficiency. The key insight is that coarsened data structures from two-phase designs can be handled more efficiently by simultaneously modeling the sampling process and the parameter of interest, reducing variance compared to methods that ignore the sampling mechanism.
Core Technical Contribution
The paper's main novelty is a unified TMLE-based framework that extends beyond IPCW-TMLE by explicitly targeting the phase-2 sampling mechanism alongside the scientific parameter, yielding improved efficiency bounds. Rather than treating the two-phase design as a nuisance to be corrected post-hoc, the authors construct estimators that leverage information about why certain subjects were selected for phase 2, embedding that knowledge into the estimation procedure itself. This represents a shift from one-step correction methods to a principled multi-target optimization that exploits the full coarsening structure inherent in the design. The new class of estimators is shown to be semiparametric efficient under their assumed data-generating mechanism.
How It Works
The method operates on data with a specific hierarchical structure: phase 1 collects variables X (basic covariates) on all N subjects; phase 2 selects n < N subjects and measures additional variables Y, with selection indicators R depending only on phase-1 data. The TMLE framework starts by fitting an initial estimator of the treatment/outcome relationship using available data, then applies a targeted fluctuation that adjusts for both measurement coarsening and sampling bias. The key technical components are: (1) a nuisance parameter model for the phase-2 selection probability P(R=1|X), (2) a target parameter model for the scientific estimand (e.g., treatment effect), and (3) a clever covariate design that simultaneously optimizes both models through a single one-step update. The output is a double-robust estimator that converges to the true parameter at root-n rate even under misspecification of either the selection model or outcome model.
Production Impact
For engineers building observational analytics systems, this approach directly improves handling of stratified data collection pipelines where intensive measurement is limited to subgroups. In real-world settings like healthcare studies (where genetic testing or imaging is expensive) or market research (where follow-up surveys target subsets), this framework replaces ad-hoc weighting schemes with principled uncertainty quantification. Implementation requires building dual-path inference pipelines that maintain the sampling design metadata and fit selection propensity models alongside outcome models; the computational cost is modest (approximately 2-3x cost of single-model estimation) but yields substantially lower confidence interval widths. The trade-off is increased model complexity and strict requirement that phase-2 selection depends only on observed phase-1 variables (the coarsening-at-random assumption), which must be validated in production through sensitivity analyses.
Limitations and When Not to Use This
The framework critically assumes coarsening-at-random (CAR), meaning phase-2 selection can depend only on fully observed phase-1 variables, not on unmeasured confounders or the unobserved phase-2 outcome itself; violations of this assumption lead to inconsistent estimation. The method requires accurate specification of the selection mechanism P(R=1|X), and misspecification can degrade efficiency gains or introduce bias if the outcome model is also misspecified (though double-robustness protects against single-source misspecification). For very small phase-2 sample sizes (n < 50), finite-sample properties and cross-fitting procedures become critical and are not fully characterized. The paper does not address mixed coarsening patterns (e.g., some variables missing not-at-random, some subjects with partial phase-2 data) or adaptive phase-2 sampling designs where enrollment changes based on interim results.
Research Context
This work extends the TMLE framework, which has become a standard tool for semiparametric inference in causal inference and complex sampling designs, specifically to the two-phase design setting that has been studied classically in survey statistics and epidemiology. The paper builds on decades of survey methodology (particularly work by Rotnitzky and others on inverse-probability weighting in stratified designs) and brings modern semiparametric theory to bear on coarsened data problems. It contributes to the broader literature on efficient influence functions and targeted learning by showing how to optimally extract information when data collection is constrained by budget or feasibility. The results open directions for higher-order designs (three or more phases) and extensions to handle missing-data mechanisms beyond coarsening-at-random.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::