Tunable Overlap for Positivity Violations in Retrospective Cohorts

Observational studies of medical records often hit the same wall. For some patients, doctors almost always choose one treatment, so there are few or no comparable patients who got the other one. Estimators that reweight patients by the inverse of their chance of treatment then blow up, because a handful of patients receive enormous weights. Practitioners have two main fixes. They can trim weights at a percentile they pick by hand, or they can switch to overlap weighting, which smoothly downweights patients whose treatment was nearly certain. Overlap weighting behaves well, but it quietly changes the question: it estimates the effect in a "clinical equipoise" subpopulation where either treatment was plausible, not in the population the study set out to describe. Neither fix tells the analyst how much variance they are buying and how much of the original question they are giving up.

Wireless engineering faced a structurally similar binary choice and resolved it. Uplink and downlink traffic could either use separate frequency bands, which is safe but wastes spectrum, or share one band fully, which doubles bandwidth but lets strong base-station signals drown weak phone signals. The resolution was the $\alpha$-duplex design, which makes the shared fraction a continuous knob. When that knob is tuned separately for each cell rather than once for the whole network, it gives large gains. We propose to carry that design lesson, and only that lesson, into causal inference. Treat the degree of tilting toward the overlap region as a continuous, per-subgroup parameter. Choose it to minimize variance subject to an explicit, covariate-only budget on how far the implied target population may drift from the intended one. Report the whole variance-versus-drift frontier rather than one ad hoc point.

Core Claim

The wireless literature supplies the motivating evidence and the design pattern, not the mathematics. In-Band α-Duplex Scheme for Cellular Networks: A Stochastic Geometry Approach shows that full band sharing gives up to 97% downlink gain but up to 94% uplink loss. A tuned intermediate overlap instead improves both links, by 28% and 56%, over separate bands. Interference Management in Full-Duplex Cellular Networks With Partial Spectrum Overlap goes further and optimizes a separate overlap factor $\alpha_i$ per cell by interior-point methods, reporting a 53% network-rate gain over a single static $\alpha$. The causal analogue of "static $\alpha$" is a single global rule, whether a trimming percentile or the fixed overlap tilt, applied to every region of covariate space. That wastes information wherever overlap is already good and under-protects wherever it is poor.

Concretely, let $e(x)$ be the propensity score, meaning the probability of treatment given covariates $x$. We use the tilting family

$$h_\alpha(x) = \{e(x)(1-e(x))\}^{\alpha}, \qquad \alpha \in [0,1],$$

where treated units receive weight $h_\alpha/e$ and controls receive $h_\alpha/(1-e)$. Setting $\alpha=0$ recovers ordinary inverse-probability weighting for the average treatment effect. Setting $\alpha=1$ recovers overlap weighting, which Average treatment effect on the treated, under lack of positivity adapts to the treated population as OWATT. Intermediate values cap the largest weight at roughly $e^{\alpha-1}$ while moving the target less.

We allow a separate $\alpha_k$ for each of $K$ prespecified strata, defined either by clinical subgroup or by propensity quantile. We choose them by solving

$$\min_{\alpha_1,\dots,\alpha_K} \widehat{V}(\alpha) \quad \text{subject to} \quad D(\alpha) \le \delta .$$

Here $\widehat{V}$ is the sandwich or bootstrap variance of the weighted estimator. $D(\alpha)$ is the largest standardized mean difference between the tilted covariate distribution and the intended target population's covariate distribution. $\delta$ is a drift budget the analyst fixes in advance.

The key point is that $D$ depends only on covariates. It answers the objection of Weight Trimming and Propensity Score Weighting, which argued that tuning a trimming threshold is intractable because the true effect is unknown. We never tune against the outcome. We tune against a measurable statement about who the estimate describes. Drift in the covariates translates into bias for the original target only through how strongly the treatment effect varies with covariates. We bound that translation using an outcome model and report the bound separately, stating the assumption behind it openly.

Research Program

The work has three parts. The first is the estimator and its inference: the per-stratum tilted estimator, an augmented (doubly robust) version that adds an outcome regression, and variance estimates that account for $\alpha$ being selected from the data. For the last of these we use sample splitting, choosing $\alpha$ on one fold and estimating on the other, and we compare against a naive plug-in to show whether selection matters.

The second part is a frontier diagnostic. For each study we trace the curve of estimated variance against drift $D$ as $\delta$ varies, mirroring the uplink/downlink rate region in the $\alpha$-duplex papers. Global overlap weighting and each trimming percentile are marked as points on the same plot. The goal is to show that the per-stratum curve dominates them, reaching lower variance at equal drift, or to show where it does not.

The third part extends the method to censored time-to-event outcomes. There we estimate differences in restricted mean survival time with censoring weights. This matters because SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis finds that classical imputation degrades above 70% censoring. Multiplying censoring weights by propensity weights makes extreme weights worse, which is exactly the regime where tunable tilting should matter most.

We deliberately do not borrow the stochastic-geometry interference mathematics to model spillover between patients (Gap 2 in the source file). That analogy is loose: patient spillover lacks the path-loss physics that makes those formulas tractable. Positivity is where the design lesson transfers cleanly.

Baselines

The naive reference point is the unadjusted difference in means and untrimmed inverse-probability weighting with a logistic propensity model. These show how bad things get when positivity fails. On the Setoguchi medical-claims simulation, Weight Trimming and Propensity Score Weighting reports 30.3% absolute bias for misspecified logistic IPW and 64.3% confidence-interval coverage.

Percentile trimming is the standard practitioner fix, run exactly as in that paper, sweeping the 99th down to the 50th percentile. Its reported best cases are bias falling to 6.5% at the 87th percentile and coverage reaching 100% at the 92nd. We report both the best percentile chosen with oracle access to the truth and a fixed 99th-percentile rule. Beating the oracle-chosen percentile without oracle access would be a strong result.

Untrimmed boosted-CART propensity weighting is the "better propensity model" competitor. The same paper reports it keeps the maximum weight at 20.3 versus 98.3 for logistic regression, and that trimming it hurts. We use identical boosted-CART propensities for every weighting method, so any gain comes from the weights and not from the propensity model.

Global overlap weighting and OWATT are the strongest existing method on this question. The OWATT paper reports the smallest RMSE and near-nominal 95% coverage in its simulations, and a MEPS White-versus-Asian expenditure disparity estimate of $2511.91 with reduced standard error. No numeric RMSE value was found on Lacuna for OWATT's simulations, so on MEPS we compare point estimates, standard errors, and drift. The per-stratum method must match or beat overlap weighting's variance at a strictly smaller drift, or beat it in RMSE against the original target estimand.

Augmented IPW with an outcome model is the doubly robust competitor. Each weighting scheme gets the same outcome model, the same cross-fitting, and the same tuning budget.

Two ablations isolate the key ingredient. The first is a single global $\alpha$ chosen under the same drift budget, which tests whether the gain comes from per-stratum tuning or merely from having a continuous knob. The second chooses $\alpha$ by cross-validated outcome-prediction error instead of the drift constraint, which tests whether the covariate-only criterion is doing real work.

Why not just…?

Why not just use overlap weights and accept the equipoise estimand? Often that is the right call, and our method reduces to it whenever the budget $\delta$ is loose. The problem arises when the clinical question concerns the whole population, for example the treated elderly and multimorbid patients whom trials exclude. Silently shifting the target then answers a different question. The MEPS and IHDP experiments measure how much the original-target RMSE differs between global overlap weighting and budgeted per-stratum tilting. If the gap is negligible across positivity severities, we will say that overlap weighting suffices.

Why not just fit a better propensity model, such as boosted CART? Lee et al. show this removes much of the need for trimming. However, flexible models reduce spurious extreme weights; they cannot create comparable patients where none exist. We therefore include simulation arms with genuine structural non-overlap, where some covariate regions are almost never treated, as well as arms with only model-induced extremes. We expect boosted CART to close the gap in the second setting but not the first.

Why not just use a doubly robust estimator and let the outcome model extrapolate? This is the real threat. With a well-specified outcome model, augmented IPW can tolerate poor overlap because the regression fills in the missing comparisons. Our claim is narrower: when the outcome model is misspecified in exactly the regions without overlap, tilting limits how much the estimate depends on extrapolation. The misspecified-outcome arms of the Setoguchi simulation settle this. If augmented IPW with untrimmed boosted weights matches our RMSE there, the per-stratum machinery is not worth its complexity, and we would recast the contribution as the drift diagnostic alone.

Proposed Experiments

The first test uses the Setoguchi claims simulation (1,000 runs, $N=500$, three nonlinearity tiers), with added arms of increasing structural non-overlap. We measure bias, RMSE, and coverage against the true average treatment effect, reproducing Lee et al.'s trimming numbers as a calibration check.

The second test uses IHDP, following the setup B surfaces used in Really Doing Great at Estimating CATE? A Critical Look at ML Benchmarking Practices in Treatment Effect Estimation. We induce positivity violations by deleting treated units within covariate regions, so that the true original-population effect is known while overlap degrades.

The third test uses MEPS as the real-data case. We report the variance–drift frontier and the per-stratum $\alpha_k$, and ask whether the chosen tilt concentrates on the few strata with extreme propensities while leaving well-overlapped strata near the untilted weights.

The fourth test uses the SurvHTE-Bench semi-synthetic MIMIC-IV datasets at low and heavy (above 70%) censoring. The target is the restricted mean survival time difference, and the comparisons are against the survival meta-learners the benchmark ranks highest.

The decisive check has two parts. At a matched drift budget, per-stratum tilting must reduce original-target RMSE relative to global overlap weighting, and relative to the best trimming percentile chosen without oracle access, in the structural non-overlap arms, while keeping coverage near 95%. It must also beat the global-$\alpha$ ablation there; otherwise the per-cell lesson from Interference Management in Full-Duplex Cellular Networks With Partial Spectrum Overlap does not transfer. Failing either part would falsify the proposal.