New papers

Estimating causal effects from observational network data requires accounting for both individual treatment and neighborhood exposure. Existing outcome models commonly incorporate these intervention variables into a joint predictor and estimate causal effects through prediction contrasts, leaving the corresponding response components implicit. We propose the $\underline{N}$etwork $\underline{E}$ffect $\underline{D}$ecomposition $\underline{Net}$work (NEDNet), which parameterizes the conditional outcome response as a query-invariant pre-treatment baseline plus separate treatment- and exposure-response pathways. Path-specific neighborhood contextualization supplies local pre-treatment information to the two pathways. We characterize the fitted response through a unique reference-normalized decomposition, which is equivalent to a zero mixed treatment--exposure contrast. Experiments on semi-synthetic benchmarks derived from BlogCatalog and Flickr show that NEDNet achieves the lowest PEHE among the compared methods for individual, peer, and total effects across homogeneous and covariate-heterogeneous response settings.

Randomized experiments have long been the gold standard for scientists seeking to learn about cause and effect. When randomized experiments are infeasible, scientists often resort to observational studies, which are widely available and often large but rely on untestable assumptions that, when violated, may result in biased estimates. Uncertainty about bias leads to a phenomenon known as the illusion of learning from observational research (Gerber, Green and Kaplan, 2004a): absent prior information about bias, observational results cannot meaningfully contribute to the estimation of a causal parameter. To shatter the illusion, we take an empirical Bayes perspective. We show that the distribution of observational biases can be learned from calibration studies-experiments that target a causal effect that is known a priori to be zero. Calibration identifies the distribution of observational bias and allows observational studies to inform the estimation of causal parameters via empirical Bayes shrinkage. We formalize the illusion phenomenon in an empirical Bayes setting and show that, with an increasing number of calibration and observation studies, both the bias distribution and the causal effect can be consistently recovered. We illustrate our method through a simulation study and a semi-synthetic application based on Ferraro and Miranda (2013)'s water-usage experiment.

Pocock and Simon's minimization method is a widely used approach for covariate-adaptive randomization in clinical trials. However, it can exhibit the shift problem under unequal allocation and variance inflation of the imbalances in additional covariates. In this paper, we propose a new minimization method that preserves the conventional biased-coin allocation probability while ensuring marginal imbalances bounded in probability. Meanwhile, the proposed procedure guarantees that neither the shift problem nor variance inflation occurs. Simulation and empirical studies demonstrate the excellent performance of the proposed procedure.

We propose a design-based covariate adjustment that leverages information from auxiliary strata to improve the efficiency of average treatment effect estimation when target samples are sparse. The estimator borrows strength from related strata that may be correlated with the target, yielding gains for studies with limited observations or rare subpopulations. We present the estimator, develop a Neyman-style conservative variance estimator for hypothesis testing and confidence interval construction, and establish its large-sample properties. The procedure is straightforward to implement within standard randomization-based workflows and does not require a correctly specified outcome model or stochastic error assumptions. In simulations, the method delivers lower variance and narrower confidence intervals than conventional design-based or regression-adjusted estimators across a range of sparsity and correlation patterns. Applications to a schooling experiment and the ACTG 175 randomized trial illustrate practical gains in precision.

Statistical software makes methods widely available. Good statistical work, however, requires understanding the question, how the data were collected, and what conclusions the evidence can support. As AI agents make data analysis accessible to people without statistical training, this reasoning should be built into the systems themselves. Our review of selected AI agents for data analysis and study design points to a gap in documented support for scientific inquiry with users at runtime. In this paper, we advocate designing domain-specific interactive harnesses in which statistical principles guide proactive inquiry: agents recognize missing information that could change an analysis, ask questions users may not know to raise, and use the answers to guide or revise analytical decisions. Two practical entry points illustrate how statisticians can begin building: an implemented causal-consulting skill within an existing agent environment and Model Forge, a custom harness for interactive model and algorithm development. These design examples connect statistical principles to choices about context, memory, tools, and interfaces, and motivate evaluation of both computational correctness and the scientific consequences of interaction. We conclude with a broader interdisciplinary research agenda on adaptive human-agent interaction, including selective information acquisition, evolving user preferences, and evaluation under uncertainty.

The joint expectation of functions of potential outcomes $(Y_1,Y_0)$ is a fundamental quantity in causal inference, particularly when investigating the variation or heterogeneity of causal effects. This paper provides novel assumptions for their identification: comonotonicity and countercomonotonicity. These assumptions offer a unified framework for their identification across discrete, continuous, and mixed outcome variables. In the absence of these assumptions, we derive sharp bounds for two broad classes of functions, yielding new results for bounding moments of individual causal effects, which are central to measuring effect heterogeneity. We present corresponding estimation methods and illustrate them in simulation studies and real-world datasets.

We consider inference on parameters and hypotheses selected using p-values. We provide a simple procedure to decompose a p-value into two components, one for inference and one for selection, that are independent under the null hypothesis. To control the Type 1 error rate conditional on selection of the null hypothesis from randomized p-values, we develop tests motivated by the selective inference approaches of data thinning, fission, and carving. Unlike the typical application of these approaches, however, our tests only make use of the p-values rather than the full data. Furthermore, we construct confidence intervals that attain the nominal conditional coverage and corresponding estimators. In the context of the file drawer problem, the winner's curse, and sequential testing, we show that our proposals offer both theoretical and empirical advantages over existing randomized and non-randomized approaches. Lastly, we apply our methods for selective inference on genetic associations using publicly-available p-values from the UK Biobank.

When we seek to uncover the causal effect of X on Y using observational data, we often include control variables in the hope that it will allow us to uncover our parameter of interest. However, there is a lack of guidance on how to pick and implement controls. The inclusion of what is known as ”bad controls” and the failure to include ”good controls” might result in bias. This article presents accessible guidelines for choosing and using control variables to identify causal effects.

Abstract Rerandomization is an experimental design technique that repeatedly randomizes treatment assignments until covariates are balanced between treatment groups. Rerandomization in the design stage of an experiment can lead to many asymptotic benefits in the analysis stage, such as increased precision, increased statistical power for hypothesis testing, reduced sensitivity to model specification, and mitigation of p-hacking. However, the standard implementation of rerandomization via rejection sampling faces a severe computational bottleneck in high-dimensional settings, where the probability of finding an acceptable randomization vanishes. Although alternatives based on Metropolis-Hastings and constrained optimization techniques have been proposed, these alternatives rely on discrete procedures that lack information from the gradient of the covariate balance metric, limiting their efficiency in high-dimensional spaces. We propose Langevin-Gradient Rerandomization (LGR), a new sampling method that mitigates this dimensionality challenge by navigating a continuous relaxation of the treatment assignment space with a Langevin diffusion. We discuss the trade-offs of this approach, specifically that LGR samples from a non-uniform distribution over the set of balanced randomizations. We demonstrate how to retain valid inference under this design using randomization tests and empirically show that LGR generates acceptable randomizations orders of magnitude faster than current rerandomization methods in high dimensions.

The Randomized Discontinuation Design (RDD) is a two-stage adaptive enrichment design that is uniquely suited to answer causal questions about the benefit of continuing treatment in patients who have already demonstrated response to an induction therapy. The estimand framework from the International Council for Harmonisation (ICH) E9(R1) addendum provides the necessary language to precisely define this scientific question. A major challenge in formulating estimands for RDD oncology trials is that the target population for the primary scientific question is itself defined by a post-treatment biological outcome in Stage 1, which makes it difficult to separate the process of selecting eligible participants from the causal inference problem in Stage 2. In this paper, we address this challenge using the Phase III JAVELIN Gastric 100 trial as an example. We argue that the "conventional" estimand can be formulated as an estimand targeting the response-defined subgroup of participants with evidence of response to the induction therapy in Stage 1, and we clarify the precise sense in which this subgroup is, and is not, related to the concept of a principal stratum. We also formulate a "full-path" estimand-capturing the patients' entire two-stage treatment journey-and connect it to the dynamic treatment regime literature. We also provide Directed Acyclic Graphs (DAGs) for the two estimands to clarify their respective causal structures. This paper aims to strengthen the dialogue among clinical research stakeholders tasked with the design and analysis of RDD trials and support more efficient regulatory decision-making. Trial Registration: ClinicalTrials.gov identifier: NCT02625610.

Observational studies of institutional decision-making often rely on explanatory variables that are themselves partly constructed through the process under investigation. In sentencing research, many legal factors are defined by the judge and may therefore act simultaneously as confounders and mediators. Controlling for them may introduce post-treatment bias, whereas omitting them may introduce confounding bias. We propose overcoming this impasse by treating the influence of judicial prejudice on these legal factors as a form of differential measurement error. Using a typical regression model of racial disparities in the U.S. Federal Courts, we demonstrate how Monte Carlo simulations can be used to examine changes in estimates of racial disparity under different assumptions about the extent to which judicial prejudice is present in the construction of offence seriousness and pre-trial detention. We show that minor differential errors in some of these legal factors can substantially mask estimates of racial disparities in sentence length, and why the convention of reporting models with and without discretionary factors fails to reflect this form of model uncertainty.

Drawing causal conclusions from non-experimental (observational) data is a central aim in science but is extremely challenging. Without control over treatment assignment (as in an experiment), treated and control groups may not be comparable on unmeasured factors affecting treatment and outcome. To address such concerns, sensitivity analyses can be performed; sensitivity analyses acknowledge the existence of unmeasured factors but ask how strong they must be to overturn causal conclusions. We (re)introduce a particularly simple form of sensitivity analysis which can be performed by any reader of an observational study with access to only an estimate and confidence interval for a causal effect. This simple sensitivity analysis may serve as an effective tool for readers wishing to gauge the robustness of conclusions from an observational study.

Rerandomization is a modern experimental design technique that repeatedly randomizes treatment assignments until covariates are deemed balanced between treatment groups. This enhances the precision and coherence of causal effect estimators, mitigates false discoveries from p-hacking, and increases statistical power. Recent work suggests that balancing covariates via rerandomization does not alter the asymptotic precision of covariate-adjusted estimators, thereby making it unclear whether rerandomization is worthwhile if adjusted estimators are used. However, these results have two key caveats. First, these results are asymptotic, leaving finite sample performance unknown. Second, these results focus on precision, while other potential benefits, such as increased coherence among flexible estimators, remain understudied. Hence, in this paper we provide three main contributions: (i) a comprehensive review of the rerandomization literature, covering historical foundations, theoretical developments, and recent methodological advancements, (ii) an extensive simulation study examining finite-sample performance, and (iii) a practical guide for practitioners. Our study compares precision, coherence, power, and coverage of various estimators under rerandomization versus complete randomization. We find rerandomization to be a complementary design strategy that enhances the precision, robustness, and reliability of causal effect estimators, especially for smaller sample sizes.

Distributed multiple testing asks $N$ sites, each holding p-values for its own hypotheses, to control the false discovery rate (FDR) of the discoveries made across the whole network while communicating only a small number of bits. Greedy interval aggregation (Pournaderi and Xiang, IEEE TSIPN, 2023) meets the communication budget but controls FDR only asymptotically, and its FDR can exceed the target at finite sample sizes; the same interval counts both select the rejection regions and calibrate the stopping rule, which biases the selected interval densities upward. We propose \emph{budgeted BONuS-GA}: each site mixes synthetic uniform p-values into its data, the center ranks candidate p-value intervals across sites using the pooled counts, and each site's false discoveries are estimated from its own synthetic counts under a per-site share of the level. The procedure needs no knowledge of the null proportions, uses every p-value for both selection and inference, and controls $FDR\le\alpha$ at every sample size for any fixed assignment of hypotheses to sites, assuming only that the null p-values are independent uniforms, independent of the non-null p-values. It keeps the original $O(\sqrt m\log m)$ communication, $m$ being the total number of p-values, when $N=O(\sqrt m)$. We also give a sample-splitting alternative, cross-fit greedy aggregation, with finite-sample control under a random-site model, and we quantify the selection bias behind the original procedure's failure. In a monitoring-network simulation the budgeted procedure retains most of the power of its heuristic counterpart at moderate-to-large site sizes.

Causal excursion effect (CEE) characterizes the effect of an intervention under policies that deviate from the experimental policy. It is widely used to study the effect of time-varying interventions that have the potential to be frequently adaptive, such as those delivered through smartphones. We study the semiparametric efficient estimation of CEE and we derive a semiparametric efficiency bound for CEE under working assumptions, in the context of micro-randomized trials. We propose a class of two-stage estimators that achieve the efficiency bound and are robust to misspecified nuisance models. In deriving the asymptotic property of the estimators, we establish a collection of easy-to-use theorems for globally robust Z-estimators with either cross-fitted or non-cross-fitted nuisance parameters that are applicable beyond the proposed estimators for CEE. We demonstrate efficiency gain of the proposed estimator compared to existing ones through simulations and a real data application using the Drink Less micro-randomized trial.

Covariate-adaptive randomization (CAR) is widely adopted in clinical trials to balance baseline covariates across treatment arms.When outcomes are binary or other types of discrete data, generalized linear models (GLMs) are typically used as working models to improve the efficiency of treatment effect estimation.However, in high-dimensional settings, limited sample sizes may lead to unstable estimation and invalid inference.Moreover, when the underlying data-generating mechanism is misspecified relative to the GLM working model, efficiency gains over the unadjusted estimator are no longer guaranteed.To address these challenges, we first propose a high-dimensional GLM adjustment approach using ℓ1 regularization, resulting in a GLM-adjusted treatment effect estimator under CAR.When the target trial has a limited sample size, we further develop a transfer learning estimator that incorporates information from a large external trial regarded as source data.As long as the source and target populations are sufficiently similar, the transfer learning estimator requires weaker sparsity conditions and a smaller target sample size than the GLM-adjusted estimator to achieve asymptotic normality.To guarantee efficiency improve-Statistica Sinica: Newly accepted Paper ment relative to the unadjusted estimator, we additionally introduce a projection method.With this step, both the GLM-adjusted and transfer learning estimators achieve efficiency gains.Simulation studies and a semi-synthetic data application demonstrate the effectiveness of our transfer learning strategy and the projection method.

Qualitative comparative analysis (QCA) requires thresholds to construct sets from continuous scales, but a standard analysis is fixed to a single set of thresholds and does not directly reveal how the sufficiency structure changes when the target level of the outcome or the judgment criterion for a condition changes. This article introduces the R package ThSQCA, which uses the QCA package as its computational engine and provides a common workflow to run, aggregate, visualize, and report analyses that systematically vary the outcome threshold, the condition thresholds, or both. Using simulated data constructed for methodological illustration, we show that raising the target level of the outcome narrows the result from multiple sufficient paths to a single path (with lower consistency and higher coverage) and eventually yields no solution. We also show that varying the judgment criterion for a condition can make a solution disappear or produce a solution supported by very few cases. ThSQCA does not select an appropriate threshold automatically; it is a diagnostic workflow that clarifies the range over which a conclusion holds and the points at which it changes.

Background: Causal discovery algorithms aim to infer causal structure from observational data under explicit assumptions, yet their performance under realistic social science conditions remains insufficiently understood. Existing evaluations typically rely on large samples, high-dimensional settings, and clean data—conditions that rarely hold in applied research. Objectives: This study evaluates four base algorithms (PC, GES, FCI, GFCI_approx) and three bootstrap-enhanced variants (FGES_boot, GFCI_approx_boot, FCI_boot) under conditions intended to reflect constraints commonly encountered in applied social science research, with the aim of understanding how different causal discovery methods behave under realistic small-sample, low-dimensional, and imperfect-data conditions. Methodology: Algorithms were evaluated through a factorial simulation design spanning small sample sizes (N = 25–200), low dimensionality (p = 4–6), moderate outlier contamination, and moderate departures from linearity (20% of causal relationships generated using quadratic functions), complemented by an illustrative application to a real educational dataset (35 countries × 7 indicators). GFCI_approx is a custom hybrid causal discovery method inspired by selected components of the GFCI framework (Ogarrio et al., 2016); because it replaces the FGES search phase with hill-climbing optimization, it should be interpreted as a distinct hybrid method rather than an implementation of the original GFCI algorithm. Findings: Five main findings emerged. First, all algorithms exhibited a pronounced precision–recall asymmetry, with consistently higher precision (0.92–1.00) than recall (0.43–0.91) across the examined conditions. Second, recovery performance improved substantially between N = 50 and N = 100, while estimation uncertainty decreased and performance became more stable at larger sample sizes. Third, increasing dimensionality (p = 4 to 6) reduced recovery accuracy unless compensated by larger sample sizes. Fourth, moderate outlier contamination reduced performance by approximately 5–22%, whereas the moderate departures from linearity examined in this study had only a negligible effect (approximately 1.6%), indicating that outlier contamination was the dominant source of performance degradation. Fifth, an illustrative application to a real educational dataset showed that, after removal of a composite variable, PC and FCI converged on two undirected edges; bootstrap stability analysis provided stronger support for the Effectiveness–Equity edge (stability = 0.800, 95% CI [0.657, 0.943]) than for the Parent engagement–Effectiveness edge (stability = 0.733, 95% CI [0.575, 0.892]). Because no ground-truth causal structure is available, this application should be interpreted as an illustration of the proposed workflow rather than as a validation of recovered causal relationships. Conclusion & Implications: We provide practical recommendations regarding sample size, preprocessing, and robustness assessment, and argue that causal discovery in small-sample social science settings is best viewed as a conservative hypothesis-generation and filtering tool rather than a comprehensive structure recovery method. By aligning the simulation design with realistic applied constraints, this study helps bridge the gap between methodological research and empirical practice.

This technical note studies whether one assignment-weighted linear estimator of a full-population exposure mean can remain design-unbiased under every graph in a finite candidate set. Exposure records own treatment and the presence of a treated neighbor; the true candidate graph is fixed throughout the experiment and response tables may be heterogeneous. Within the existing design-based Riesz framework, the note presents the finite affine-support condition and its nonnegative common-support corollary, then computes the support structure of candidate perfect matchings. Under uniform complete randomization with exactly two treated units, signed linear unbiased estimation of the (1,1) population mean is feasible exactly when every unit has a person who is its partner in none of the candidate matchings. For a specified four-unit, two-candidate matching experiment with positive treated-pair probabilities p_e and response values in [-M,M], the note establishes uniqueness in the common deterministic linear unbiased class and exact worst design variance R(p)=(M^2/4) sum_e 1/p_e. Uniform treated-pair assignment minimizes this risk at 9M^2; the known-graph linear unbiased oracle value under the same uniform experiment is 3M^2. These costs refer to the stated experiment and estimator classes. Version v1 is the first public technical-note release.

Experiments with spillovers create incentives for strategic behavior: when one household member's treatment can affect another, take-up decisions become interdependent. We propose an instrumental-variables framework grounded in game theory, in which take-up is a Nash equilibrium among a small number of agents, and all variables are discrete. Under minimal assumptions, without specifying how equilibria are selected, we derive sharp non-parametric bounds for direct, indirect, and offer-mediated effects. Game-theoretic restrictions such as supermodularity, symmetry, and dominance can be layered on via a tractable linear program, and we characterize when each does or does not tighten the identified set. We illustrate the usefulness of our approach to study household experiments by reanalyzing a banking intervention in Kenya.

The test-negative design has become popular for evaluating the effectiveness of post-licensure vaccines using observational data. In addition to its logistical convenience for data collection, the design is also believed to control for the differential health-care-seeking behavior between vaccinated and unvaccinated individuals, an important while often unmeasured confounder of the vaccination–infection relationship. Hence, the design has been employed routinely to monitor seasonal flu vaccines and more recently to measure COVID-19 vaccine effectiveness. Despite its popularity, the design has been questioned, particularly about its ability to fully control for unmeasured confounding. In this paper, we explore deviations from a perfect test-negative design, and propose various sensitivity analysis methods for estimating the effect of vaccination measured by the causal odds ratio in the subpopulation of individuals with good health-care-seeking behavior. We start with point identification of the causal odds ratio, comparing different forms of identification assumptions and their corresponding estimands. We then propose two approaches for conducting sensitivity analysis, addressing the influence of unmeasured confounding in two different ways. Specifically, one approach investigates partial control for unmeasured confounding, while the other examines the impact of unmeasured confounding on both vaccination and infection. Furthermore, we combine them to provide narrower bounds on the causal odds ratio, and further sharpen the bounds by restricting the treatment effect heterogeneity. Finally, we apply the proposed methods to evaluate the effectiveness of COVID-19 vaccines using observational data from test-negative designs.

This paper analyzes spillover effects in spatial (network) models when the neighborhood (adjacency) matrix is contaminated by measurement error from reporting, aggregation, or disclosure imperfections, leading to inconsistent estimation of network effects. We introduce a regularization framework for the latent network that allows for sparse and/or low-rank structure and accommodates potential correlation between measurement errors and outcomes. We propose two estimators: (i) a two-stage procedure that first denoises the adjacency matrix and then incorporates the purified network into a regression analysis, and (ii) a Generalized Method of Moments (GMM) estimator that jointly estimates regression parameters and refines the network structure. We then establish strictly improved consistency rates for the spillover effect estimator relative to naive estimation ignoring measurement error. Simulations demonstrate that, in the presence of noisy networks, our approach reduces the root mean squared error of spillover estimates relative to conventional methods by approximately $50-80\%$. We apply our framework to examine the international spillover of economic growth, and the tax competition across U.S. states, illustrating that denoising might restore Leontief stability and yields improved estimates of spillovers.

Abstract Artificial intelligence (AI) applied to multiomics data has expanded precision medicine's capacity to predict disease risk, prognosis, and treatment response. However, predictive accuracy alone does not answer the central question of precision therapeutics: which intervention will improve a given patient's outcome. Conventional multiomics models primarily capture statistical associations and remain vulnerable to confounding and distributional shift, limiting clinical actionability. Causal artificial intelligence (causal AI), defined here as the integration of causal inference with machine learning (ML) to estimate intervention effects rather than predict outcomes, offers a complementary framework. This review critically examines the conceptual foundations, terminology, methodological maturity, and translational status of causal AI in precision therapeutics, distinguishing causal inference, causal ML, causal discovery, and causal representation learning, and emphasizing multiomics integration and biological mechanisms. I review causal forests, targeted maximum likelihood estimation, Mendelian randomization, and target trial emulation, alongside layer‐specific opportunities across multiomics data. Emerging applications in oncology, cardiometabolic disease, and psychiatry illustrate the potential of causal AI to characterize treatment‐effect heterogeneity, though most remain at hypothesis‐generating or retrospective stages, and methodological, interpretability, regulatory, and equity challenges remain substantial. Mechanism‐anchored causal AI offers a promising but still immature path toward closing the gap between molecular prediction and clinical intervention.

Non-compliance in randomised controlled trials (RCTs) may distort the estimate of treatment effect in intention-to-treat (ITT) and per-protocol (PP) analyses. For time-to-event outcomes, Inverse Probability of Censoring Weighting (IPCW) method and Rank Preserving Structural Failure Time model (RPSFTM) have been proposed to address treatment switch across two arms, with treatment effects summarised in terms of hazard ratio (HR) or restricted mean survival time (RMST). In this paper, we adapt the IPCW and RPSFTM methods to consider one-sided non-compliance affecting the experimental arm. For IPCW, non-compliers in the experimental arm were censored at the time of non-compliance. For RPSFTM, the causal model was modified to reconstruct the counterfactual survival time for non-compliers in the experimental arm. We quantify the treatment effects using HR, difference (dRMST) or ratio (rRMST) of RMST and milestone survival difference (MSD). We compare the aforementioned methods via rigorous simulation and a real-world data analysis based on a randomised trial of nasopharyngeal cancer. The simulation results demonstrated that RPSFTM-related models produced smaller bias, whereas IPCW generally performed well under high non-compliance for all summary measures. However, the RMST-based estimates had lower power than the HR estimate under idealised proportional hazard setting. When illustrated using real-world data, RPSFTM-related models yielded larger estimate of treatment effect for summary measures of dRMST and rRMST, whereas IPCW produced larger estimates when quantified in terms of HR and MSD. For handling treatment non-compliance in RCTs, it is important to choose the summary measure and model carefully on a case-by-case basis.

Interference between participants in two-sided marketplace experiments can severely bias conventional A/B tests. While two-sided randomization (TSR), which randomizes both supply and demand, enables estimators that mitigate this bias, existing analyses offer limited guidance on which estimator to prefer for the balanced market conditions common in practice, as they largely focus on extreme supply- or demand-constrained regimes. We introduce the Double Diagonal (DD) estimator, a novel method for interference-robust treatment effect estimation. The DD estimator contrasts the outcomes of treated-treated and control-control pairs and uses the off-diagonal interactions (treated-control and control-treated) to correct for interference-induced bias. We provide the first theoretical comparison between the DD and the standard naive TSR (TSRN) estimator under a simplifying mean-field market model, proving two key results: (i) DD's bias is strictly smaller than the standard estimator's below a critical threshold of the outside-option parameter; in particular, DD's bias is smaller whenever the outside option is at most twice as attractive as platform options; and (ii) the two estimators exhibit biases in opposite directions, with DD underestimating and the standard estimator overestimating the true effect, respectively. Extensive numerical experiments in realistic market settings--spanning heterogeneous treatment effects, supply-demand imbalances, and varied market structures--corroborate our theoretical findings. Across a wide range of configurations, the DD estimator demonstrates the lowest bias, establishing it as a robust and often superior choice for reliable experimentation in two-sided marketplaces.

We consider causal effect estimation in observational studies with continuous treatment. We introduce a general class of causal estimands that, as a special case, includes the average causal effect of assigning treatment according to an arbitrary counterfactual distribution. Balancing estimators are particularly attractive tools for covariate adjustment in this setting because they avoid directly modeling the treatment assignment mechanism, thereby reducing the potential for instability and misspecification bias in subsequent causal effect estimation. In this paper, we develop a general framework of covariate balancing for continuous treatment with the proposed class of causal estimands, leveraging the assumption of continuity of treatment dose. As in the binary treatment setting, analysts specify the basis functions to be balanced, the allowable degree of covariate imbalance, and a criterion for selecting among admissible weights. Under a minimum variance criterion for choosing weights, we establish consistency, characterize the effect of basis misspecification, and prove asymptotic normality. We develop consistent variance estimation, which yields asymptotically valid confidence intervals. Under homoskedasticity, our estimator is provably more efficient than inverse probability weighting based on the \emph{true} propensity score. Simulation studies support our theoretical results, and we illustrate the proposed method by estimating the effect of vocational training hours on expected future earnings using data from the U.S. Job Corps study.

Randomized clinical trials remain the gold standard for evaluating treatment effects but are often infeasible for rapidly assessing drug repurposing opportunities, particularly in complex conditions such as Alzheimer's disease. Target trial emulation (TTE) provides a principled framework for causal inference with electronic health records and, when coupled with modern machine learning, enables scalable analyses across distributed health systems. We present FLAT-WHITE (Federated Lossless Algorithm for Target Trial Emulation with Overlap Weighting for Healthcare Data Integration via Trusted and Efficient Communication), a novel federated causal machine learning pipeline for high-throughput treatment effect estimation without sharing patient-level data. FLAT-WHITE is a one-shot, lossless algorithm, requiring a single communication round while achieving results equivalent to centralized pooled analyses. With overlap weighting on the propensity score, it robustly addresses heterogeneity in treatment assignment and ensures valid causal inference in decentralized settings. We demonstrate FLAT-WHITE in a Alzheimer's drug repurposing study, scanning 181 candidate drugs using data from four health learning systems encompassing approximately 150,000 patients with mild cognitive impairment. Findings were validated externally in 339,180 patients from the TriNetX global research network. FLAT-WHITE enables rapid, large-scale, and secure evaluation of treatment effects and establishes a foundation for collaborative clinical research across diverse healthcare systems.

Methods for causal inference are well developed for binary and continuous exposures, but in many settings, the exposure has a substantial mass at zero-such exposures are called semi-continuous. We propose a general causal framework for such semi-continuous exposures, together with a two-stage estimation strategy. A two-part propensity structure is introduced for the semi-continuous exposure, with one component for exposure status (exposed vs unexposed) and another for the exposure level among those exposed, and incorporates both into a marginal structural model that disentangles the effects of exposure status and dose. The two-stage procedure sequentially targets the causal dose-response among exposed individuals and the causal effect of exposure status at a reference dose, allowing flexibility in the choice of propensity score methods in the second stage. We establish consistency and asymptotic normality for the resulting estimators, and characterize their limiting values under misspecification of the propensity score models. Simulation studies evaluate finite sample performance and robustness, and an application to a study of prenatal alcohol exposure and child cognition demonstrates how the proposed methods can be used to address a range of scientific questions about both exposure status and exposure intensity.