g-computation

This paper addresses selection bias in epidemiology under complex causal structures—specifically, the coexistence of treatment-induced selection and multiple biases (e.g., absence of a joint adjustment set). We propose two novel g-computation estimators. Methodologically, we are the first to unify these two challenging forms of selection bias within the g-computation framework, implementing them via stacked estimating equations. Leveraging causal graph identification, Monte Carlo simulation, and finite-sample theoretical analysis, we rigorously establish consistency and asymptotic normality. Simulations demonstrate that the new estimators significantly outperform conventional adjustment methods in small samples. Our core contribution lies in bridging causal identification with implementable estimation: systematically translating causal graph–based inferential results into interpretable, computationally feasible, and statistically guaranteed estimation strategies.

In randomized clinical trials, g-computation often yields biased treatment effect estimates and underestimates variance after covariate adjustment—particularly in small samples or rare-event settings—where maximum likelihood estimation may fail. To address this, we introduce, for the first time, a systematic bias-correction framework into g-computation. Our method employs a generalized Oaxaca–Blinder estimator for debiasing, integrated with Firth’s penalized likelihood correction and asymptotic bias analysis to derive a bounded, robust variance adjustment. The resulting estimator improves finite-sample accuracy and inferential stability without compromising efficiency. Through extensive simulations and reanalyses of real clinical trials, we demonstrate that our approach effectively balances the bias–efficiency trade-off, yielding more reliable and practically applicable unconditional treatment effect estimates.

In experimental causal inference, design uncertainty—arising from the assignment mechanism—is often overlooked under small sample sizes and heterogeneous treatment effects; conventional causal bootstrap methods apply only to completely randomized designs and average treatment effect estimation. This paper addresses this limitation by introducing integer linear programming into the causal bootstrap framework for the first time, enabling computation of the worst-case copula under generalized assignment mechanisms (e.g., conditional unconfoundedness, bounded confounding) to uniformly calibrate design uncertainty. The method accommodates both linear and quadratic treatment effect estimators and is supported by asymptotic theory establishing its validity. Monte Carlo simulations demonstrate that, in small-scale geographic experiments, the proposed approach substantially improves confidence interval coverage and precision while delivering more robust control of Type I error.

This study addresses the bias introduced by external control data in hybrid controlled trials and the reliance of conventional methods on strong exchangeability assumptions. The authors propose a model-robust G-computation approach that achieves unbiased and efficient estimation under weaker assumptions by adjusting for baseline covariates. Notably, the method does not require exchangeability and retains consistency and asymptotic normality even when the outcome regression model is misspecified, offering robustness, simplicity, and efficiency. Integrating variable selection with covariate adjustment, theoretical analysis, simulation studies, and an empirical application to an HIV treatment trial demonstrate that the proposed method effectively controls bias and substantially improves estimation efficiency across diverse scenarios.

This study addresses the bias in causal effect estimation arising from the coexistence of unmeasured cluster-level confounding and treatment effect heterogeneity in observational clustered data. To tackle this dual challenge, the authors propose an intra-group g-computation approach: clusters are first stratified by observed treatment prevalence, within-stratum g-computation is implemented using random-effects models, and estimates across strata are then aggregated to correct for bias. This method innovatively embeds random-effects modeling within the g-computation framework, effectively mitigating both sources of bias. Simulation studies demonstrate that the proposed estimator achieves the lowest root mean squared error when unmeasured confounding and heterogeneity co-occur. Applied to data from Bangladesh, the method reveals that adolescent pregnancy is associated with an average reduction of 0.12 in child height-for-age Z-scores (95% CI: [–0.18, –0.06]).

This study addresses a critical limitation of the traditional synthetic control method: when treatment effects are weak, systematic bias can shift the center of confidence intervals, leading to misleading inferences. To remedy this, the authors propose a novel time placebo–guided approach that explicitly quantifies and corrects for such bias. By retrospectively assigning placebo intervention dates within the observed panel and refitting the synthetic control model at each, the method directly estimates the bias distribution under the null hypothesis. This enables the construction of nonparametric confidence intervals calibrated to maintain nominal coverage regardless of the true effect trajectory. The proposed procedure achieves stable, bias-corrected inference with fixed interval width, substantially enhancing the robustness of causal conclusions in synthetic control applications.

This study addresses the challenge of accurately estimating the causal effects of multiple interventions on average length of stay in hospital quality improvement, where data scarcity and complex underlying mechanisms hinder reliable inference. The authors propose expert-guided g-computation (egg-computation), a novel framework that integrates Gantt charts with causal directed acyclic graphs (DAGs) to unify expert knowledge and empirical evidence. Clinical expert judgment is selectively incorporated only where causal identification is otherwise impossible, and large language models (LLMs) are leveraged to scalably generate causal graphs and estimates of time savings. In simulations, the method outperforms conventional causal inference approaches; when applied to evaluate eleven real-world hospital interventions, LLM-assisted results show high concordance with human expert assessments, demonstrating an efficient and scalable solution for causal effect estimation.

This study identifies and quantifies a critical implementation flaw in the gsynth R package (prior to version 1.3.1) when combining interactive fixed effects–expectation maximization (IFE-EM) estimators with parametric bootstrap inference: the algorithm erroneously substitutes in-sample residuals for out-of-sample prediction errors, leading to systematically underestimated standard errors and compromised inferential validity. Through Monte Carlo simulations, placebo tests, recomputed standard errors, and robustness checks using the generalized synthetic control method (GSCM), we demonstrate that the original approach yields substantially inflated false positive rates. After correction, most estimated treatment effects lose statistical significance, and reanalysis of three APSR articles using GSCM invalidates their core conclusions, underscoring the essential role of methodological rigor in empirical political science research.

This study addresses the limited accessibility of the g-formula in causal inference due to its mathematically opaque formulation for those with modest statistical backgrounds. Under the standard assumptions of consistency, positivity, and conditional exchangeability, the authors systematically reformulate the g-formula into two nonparametrically equivalent representations: a non-iterative (NICE) and an iterative (ICE) form. This novel decomposition clarifies the g-formula’s intrinsic connections to the law of iterated expectations and conditional expectation operators. Through three progressively complex numerical examples—spanning settings with both fixed and time-varying confounders—the paper intuitively illustrates the computational mechanics and causal identification logic underlying the g-formula. The proposed framework substantially enhances interpretability and generalizability, offering practitioners a transparent and unified pathway for estimating causal effects.

This study addresses the challenges of estimating causal effects of time-varying interventions on rare survival outcomes in large-scale longitudinal observational studies, where high computational costs and severe class imbalance often hinder reliable inference. The authors propose a subsampling and inverse probability reweighting framework tailored for longitudinal survival data, which integrates seamlessly with existing causal estimators—such as g-formula–based ICE—while preserving estimator consistency and substantially reducing computational burden. This approach represents the first application of a subsampling strategy that jointly optimizes computational efficiency and statistical consistency in the context of causal inference for longitudinal rare events, effectively mitigating model instability induced by outcome imbalance. Simulations and an empirical analysis using electronic health records to assess the impact of social-behavioral factors on suicide risk demonstrate that the method markedly improves computational efficiency while enhancing both the stability and accuracy of causal estimates.