arXiv subjects

Maria Kudela

Publications and source records attributed to Maria Kudela.

3 recordsLinked to original sources

arXiv2026

Characterizing Estimator-Regime Transitions in Adaptive Enrichment Trials

Adaptive enrichment trials that identify responsive subpopulations allocate a fixed sample across them and estimate each one's effect from the resulting data. For a synthetic control design with adaptive recruitment and borrowing of control outcomes, we ask whether a doubly robust family of targeted maximum likelihood estimators should replace its sample mean; the choice can be made largely at the design stage. A crossing condition weighs the bias removed and the efficiency gained by the doubly robust family against the finite-sample inflation of nuisance estimation; an equal-error boundary, located by simulation, separates the families in the plane of sample size and bias. A consistency principle prices recruiting with the wrong family's variance functional. With three and eight simulated subpopulations the boundary's sign prediction for the mean squared error difference holds in $99.6\%$ of $10{,}302$ comparisons decided at two standard errors. A pre-specified selector commits to the doubly robust family where the squared anticipated bias exceeds the net variance term. Recruiting along that family's path costs up to $0.035$ in the area under the receiver operating characteristic curve; recruiting with the sample mean functional and routing only the analysis removes this cost and about nine tenths of its error excess.

stat.ME↗

arXiv2026

Artificial Intelligence in Pharmaceutical Development: Building Rigorous, Decision-Ready Evidence

Artificial intelligence (AI) and machine learning (ML) are reshaping the pharmaceutical lifecycle, from discovery and clinical development through manufacturing and regulatory review to safety surveillance and health economics, creating evidentiary challenges that predictive performance alone cannot address. This review examines the inferential, causal, validation, and documentation requirements for AI-enabled evidence generation in high-stakes settings and argues that they must be built into workflows. We outline the evidentiary functions that AI-derived outputs must satisfy before supporting consequential decisions: estimand definition, fit-for-purpose data collection, bias control, uncertainty quantification, validation, interpretability, and decision-grade documentation. Because AI both depends on and accelerates large-scale, non-randomized data collection, the integration of randomized controlled trials (RCTs) with real-world data (RWD) is where its evidentiary risks concentrate; we therefore present a five-step causal roadmap for that integration, in which external data supplement but never replace a concurrent control, and characterize the safeguards required across development, pharmacovigilance, regulation, and health technology assessment. Priority frontiers answer recurring implementation problems: conformal prediction for overconfident predictions, randomization-based inference for scarce data, and federated causal inference and privacy-preserving generative modeling for data that cannot be pooled. Evidence generation depends not on competition between RCTs and RWD, or between statistics and AI/ML, but on their principled synthesis under causal and inferential frameworks, with statistical expertise central to decision-ready evidence.

stat.AP↗

arXiv2026

Doubly Robust Estimation and Inference for Longitudinal Win Probabilities with Ordinal Outcomes in Observational Studies

Repeated ordinal outcomes capture how treatment benefits and harms evolve over time. We develop doubly robust estimation and inference for visit-specific win probabilities and their prespecified weighted summaries, with Desirability of Outcome Ranking as a motivating example. Baseline confounding is addressed through augmented inverse-probability weighting (AIPW) of treatment-specific marginal category probabilities or joint multinomial targeted maximum likelihood estimation (TMLE). We consider fold-specific and pooled targeting and establish conditions for double-robust consistency and semiparametric efficiency under participant-level cross-fitting. When only one nuisance model is correctly specified, inference based only on the efficient influence function may be invalid. For smooth parametric nuisance models, we derive covariance corrections from the complete estimating procedure, including targeting. When a correctly specified propensity model is estimated by maximum likelihood, a simplified score correction accommodates convergent flexible outcome learners without requiring their derivatives. The corrected covariance estimators retain within-participant dependence and support visit-specific, weighted-summary, and joint inference across repeated assessments. Simulations supported double robustness, demonstrated the importance of cross-fitting with overfit learners, and showed improved marginal and joint confidence coverage with the proposed corrections. TMLE, particularly with pooled targeting, often had lower root mean squared error under adequate treatment overlap, whereas AIPW was more stable under weak overlap. An electronic health record study of hospitalized adults with an acute respiratory illness illustrates how the trajectory and weighted summary reveal complementary aspects of the treatment comparison.

stat.ME↗