interrupted time series analysis

This study addresses the bias in estimating intervention effects within single-unit interrupted time series analyses that arises from ignoring concurrent or preceding disruptive interruptions—such as the COVID-19 pandemic. To mitigate this issue, the paper introduces Bayesian stacking into this domain for the first time, constructing a robust counterfactual prediction model by integrating multiple functional forms representing potential confounding interruptions. By combining count data modeling with flexible functional ensembles, the proposed approach substantially enhances the reliability of counterfactual inference under co-occurring disruptions. The method is empirically applied to assess the impact of Texas’s six-week abortion ban in 2021 on reported pregnancy counts, effectively adjusting for pandemic-related confounding and accurately isolating the true policy effect.

Modeling nonlinear and heterogeneous causal effects in interrupted time series with multiple subpopulations remains challenging. This paper proposes a Bayesian hierarchical generalized additive model (GAM) that integrates partial-pooling priors with a hierarchical model selection mechanism, enabling information sharing across groups while preserving subgroup-specificity. The method supports nonlinear intervention responses, multilevel structural modeling, and post-stratified causal inference, with robust estimation via MCMC. We evaluate it on three real-world applications: the impact of PSA screening introduction on prostate cancer diagnosis rates; changes in rural stroke/TIA hospitalization rates during early COVID-19; and heterogeneous effects of Missouri’s Medicaid expansion on payment methods across age and sex subgroups. Results demonstrate substantially improved accuracy in identifying heterogeneous effects and enhanced cross-group comparability, yielding an interpretable and generalizable causal framework for policy evaluation.

研究了在所有单元同时受到干预的情况下，如何通过高斯过程回归估计时间序列中的因果效应，以解决缺乏同期对照组的问题。

This study addresses the limitations of conventional difference-in-differences (DiD) methods in discrete-outcome panel data, where violations of the parallel trends assumption—such as mean reversion, counterfactual extrapolation beyond support, and ill-defined trends across multiple outcome categories—can induce substantial bias. The authors propose a novel identification strategy grounded in transition independence: absent treatment, the state transition dynamics of the treated and control groups are identical conditional on prior outcomes. Integrating this assumption with a latent class Markov model, the approach identifies latent individual types from short panels and consistently estimates the average treatment effect on the treated (ATT), effectively accounting for unobserved heterogeneity. Empirical results demonstrate that the estimated ATT under the proposed method differs markedly from conventional DiD estimates, confirming its validity and robustness.

This study addresses the limitation of traditional regression adjustment methods, which can only estimate average treatment effects and fail to capture the dynamic evolution of treatment effects over time. The authors propose a novel longitudinal treatment effect estimation framework that, for the first time, incorporates time-varying covariate transition mechanisms into regression adjustment. By modeling post-treatment covariate trajectories through transition kernels, the method enables a fine-grained characterization of effect heterogeneity across time. The proposed estimator is theoretically shown to achieve the semiparametric efficiency bound and possesses asymptotic normality. Both simulation studies and empirical analysis using A/B test data from a Japanese streaming platform demonstrate that the approach substantially reduces estimation variance and enhances statistical inference efficiency.

该文介绍R包tteICE，通过实施五种策略解决临床试验中因并发事件导致的时间到事件结果的治疗效果评估问题。

This study addresses the failure of conventional two-way fixed effects (TWFE) models in estimating treatment effects for repeated events, where disentangling the independent dynamic impacts of individual occurrences remains challenging. Focusing on recurrent shocks such as natural disasters, this work proposes a linear parametric framework grounded in a conditional parallel trends assumption based on effect accumulation. Inference is conducted using a sequential imputation estimator combined with Monte Carlo simulations. The proposed approach achieves robust and consistent estimation of the dynamic effects associated with single events, effectively recovering the aggregate trajectory while precisely isolating the independent impact of each shock. Simulation experiments further validate the superiority of this methodology over existing alternatives.

This study addresses the challenge of estimating the causal effect of longitudinal treatment strategies on survival outcomes using electronic health records, where monitoring frequency of covariates varies across patients, variable types, and time, and may itself carry information about underlying health status—potentially biasing conventional causal inference methods. For the first time, monitoring indicators are formally treated as time-varying confounders. The authors integrate inverse probability weighting, G-computation, and longitudinal targeted maximum likelihood estimation (TMLE) to develop a unified framework for causal effect estimation under informative monitoring. This approach substantially reduces bias arising from ignoring the monitoring mechanism and extends the applicability of both static and dynamic treatment strategies. Simulations confirm its validity, and an application to real-world ICU data demonstrates its ability to accurately assess the impact of different mechanical ventilation initiation strategies on mortality.

This study addresses the bias in causal inference caused by missing not at random (MNAR) mechanisms in longitudinal observational data. Building upon the self-censoring model, this work introduces shadow variable moment conditions and machine learning techniques, proposing a recursive identification formula and a pooled smoothing strategy to overcome information borrowing challenges under sparse measurements. Furthermore, nuisance function estimation is achieved by integrating path derivatives, cross-fitting, and ensemble learning. The primary contribution lies in constructing an asymptotically linear estimator that attains root-n consistent estimation of longitudinal modified treatment policy effects while enabling simultaneous inference across multiple time points.

Conventional regression discontinuity design (RDD) cannot estimate causal effects under dynamic threshold interventions—such as initiating diabetes prevention based on time-varying fasting glucose levels—because it ignores the temporal evolution of measured variables. Method: We propose the “dynamic marginal policy effect” as a causally identifiable target at the threshold and develop a nonparametric estimation framework integrating local linear regression with dynamic causal modeling. We rigorously establish its consistency and asymptotic normality. Contribution: This work is the first to extend RDD to dynamic threshold settings, overcoming the restrictive static-threshold assumption. Simulation studies demonstrate that our method substantially improves estimation accuracy for causal effects compared to classical RDD and existing static extensions. It provides a rigorous causal inference tool for time-varying decision-making contexts, particularly in chronic disease prevention and other adaptive clinical or policy interventions.