moderation analysis

This study addresses the problem of testing whether a treatment effect operates entirely through observed mediators and identifying causal mechanisms under control for covariates. The authors propose a statistical test based on double machine learning, extending— for the first time—the joint evaluation of full mediation and causal mechanism identification to non-randomized treatment settings. By integrating conditional independence testing, the method achieves root-n consistent and asymptotically normal inference even in the presence of high-dimensional covariates. Simulation studies demonstrate favorable finite-sample performance, and the approach is successfully applied to two randomized experiments examining maternal mental health and social norms.

This paper addresses the testability of unmeasured confounding in observational studies, aiming to determine whether valid causal inference is feasible. We propose the first statistically rigorous method to test the “no unmeasured confounding” assumption, achieved by formally establishing a mathematical correspondence between the potential outcomes framework and causal graph models—thereby clarifying the fundamental distinction between causal identification and conventional association-based inference. Our approach operates within linear structural equation models and leverages joint analysis of randomized controlled trial (RCT) data and observational data to calibrate statistical power and rigorously control Type I error. The key contribution is the first empirically implementable, reproducible diagnostic test for unmeasured confounding, providing practitioners with a practical tool to assess the credibility and scope of causal conclusions drawn from observational studies.

This study evaluates the causal effects of large-scale deplatforming interventions, exemplified by Reddit’s “Great Ban.” Leveraging longitudinal log data from 34,000 users and 53 million comments, it applies a difference-in-differences (DID) framework—the first such use for quantifying heterogeneous treatment effects in deplatforming. Results reveal: (1) 15.6% of banned users permanently disengage; (2) among retained users, aggregate comment toxicity declines by 4.1%, yet a distinct subgroup exhibits a 70% *increase* in toxicity; (3) this high-toxicity subgroup shows no significant rise in activity or engagement, challenging the common hypothesis that deplatforming intensifies extremist behavior. The paper’s contributions are threefold: it pioneers DID-based estimation of differential deplatforming effects; uncovers non-monotonic toxicity responses; and demonstrates substantial individual-level heterogeneity—refuting uniform-effect assumptions. These findings provide granular, causal evidence to inform platform content governance policies.

This study addresses the inconsistency in causal effect estimates between observational studies and randomized controlled trials (RCTs) by proposing the first unified framework for decomposing causal effect heterogeneity. The framework systematically identifies and quantifies three sources of heterogeneity: differences in covariate distributions, variation in mediating pathways, and shifts in outcome-generating mechanisms. Methodologically, it formally defines effect decomposition across data types (observational vs. experimental), integrating causal inference, sensitivity analysis, and decomposition modeling, while enabling robust parameter estimation under multiple hypotheses. Evaluated through simulation studies and an empirical analysis of the “Moving to Opportunity” experiment, the framework demonstrates improved interpretability, robustness, and policy generalizability in synthesizing evidence from heterogeneous data sources.

This paper addresses causal inference in network experiments subject to interference. We propose a purely design-based, model-agnostic weighted least squares framework. Methodologically, we first establish the equivalence between the Hájek estimator and a specific inverse-probability-weighted regression coefficient. Second, we develop a bias-corrected network-robust covariance adjustment that ensures design-based validity of standard errors under arbitrary regression misspecification. Theoretically, our estimator is consistent and asymptotically normal. Simulations and empirical applications demonstrate stable confidence interval coverage exceeding 95%. Our approach balances practical implementability, flexible incorporation of covariates, and design-based robustness—offering a new paradigm for causal inference in network experiments that unifies theoretical rigor with empirical applicability.

Toxic content propagation on online social platforms demands governance mechanisms that balance theoretical rigor with practical feasibility. This paper proposes a toxicity propagation simulation framework based on an extended SEIZ (Susceptible–Exposed–Infected–Zombie) epidemic model. It introduces, for the first time in content moderation simulation, user-level modeling of the Dark Triad personality traits—narcissism, Machiavellianism, and psychopathy—as key determinants of susceptibility and transmission behavior. We design a threshold-driven, configurable, and interpretable personalized moderator that dynamically adjusts intervention intensity according to individual psychological profiles, thereby departing from conventional uniform-intervention paradigms. Experimental results demonstrate that the proposed intelligent moderator significantly suppresses toxicity diffusion: average propagation rate and duration decrease by 47% relative to a baseline moderator. These findings validate the critical advantages of personality-aware moderation strategies in enhancing both intervention efficacy and interpretability.

This study addresses a critical gap in content moderation research by moving beyond the prevailing view of moderation as a monolithic intervention and instead examining the joint effects of moderator type, violation context, and linguistic style on user compliance and self-censorship. Grounded in the HAII-TIME framework, the analysis leverages over ten million moderation events from Reddit, integrating probabilistic behavioral classification, ANOVA, OLS regression, and PCA-based linguistic feature analysis. The findings reveal, for the first time, that bot moderators are more effective than human moderators at enhancing compliance while reducing self-censorship. The study further demonstrates that violation severity moderates the efficacy of linguistic strategies and incorporates violation salience into the HAII-TIME model. Among 480 language-context interactions, 33 effects remain significant after FDR correction, offering empirical foundations for context-adaptive moderation systems.

This study addresses the external validity bias that arises when generalizing results from randomized controlled trials to a target population due to distributional differences in unobserved effect modifiers. Building on omitted variable bias theory, the authors propose a novel sensitivity analysis framework that decomposes external validity bias into the product of effect modification strength and covariate distributional imbalance. The framework introduces a dimensionless sensitivity parameter based on partial R², enabling closed-form bias bounds and facilitating benchmarking against observed covariates. Simulation studies demonstrate that the proposed bounds remain conservative and achieve nominal coverage even under model misspecification, offering improved interpretability and practical utility compared to existing approaches.

This study addresses the challenge of causal inference when multiple unobserved versions of a treatment violate the Stable Unit Treatment Value Assumption (SUTVA), leading to biased estimates under standard approaches. To tackle this issue, the authors introduce, for the first time, a Mixture-of-Experts framework into causal inference, integrating latent variable modeling with causal identification theory. The proposed method enables the identification and estimation of version-specific causal effects even when treatment versions are unobserved. Theoretical analysis establishes the identifiability of the approach under plausible assumptions, and extensive numerical experiments demonstrate its effectiveness and robustness in accurately estimating causal effects across a variety of data-generating settings.

This study addresses a critical yet often overlooked issue in observational research: when proxy variables are used to control for unmeasured confounding, covariates highly correlated with the exposure may inadvertently amplify sensitivity to residual confounding—an effect commonly neglected in conventional sensitivity analyses. Within a regression framework, this work formally characterizes this phenomenon and introduces a novel, observable metric based on the ratio of the exposure model coefficient to the residual variance, which quantifies how covariate structure exacerbates sensitivity to unmeasured confounding. By integrating multicollinearity into the interpretive framework of sensitivity analysis, the approach is validated through linear regression, proxy variable modeling, and sensitivity assessment in the context of smoking and lung cancer. Empirical results demonstrate that increasing socioeconomic stratification over time has heightened the sensitivity of recent data to unmeasured confounding.