Wednesday, Aug 5: 8:30 AM - 10:20 AM
6417
Contributed Papers
Thomas M. Menino Convention & Exhibition Center
Room: CC-102B
Main Sponsor
Section on Statistics in Epidemiology
Presentations
Spillover effects arise when an intervention received by one unit affects the outcomes of units within a predefined group, referred to as an interference set. Such effects commonly occur in clusters/sociometric networks. HIV prevention programs are often delivered to communities as intervention packages that contain multiple components. Disentangling component-specific effects of intervention packages is essential for fully understanding the effectiveness of HIV interventions. However, existing causal methods for estimating spillover effects are typically limited to settings with a single intervention, or multiple interventions that are analyzed as a whole, thereby unable to provide insights into which components were driving (or hindering) the effectiveness. Here, we develop novel causal methods for time-varying exposure to intervention packages with interference. We expand partial interference assumption, use marginal structure models to estimate the potential outcome, with time-updated inverse probability weighting to adjust for confounders. Generalized estimating equations is employed to derive closed-form robust variance estimators.
Keywords
Spillover effect
intervention package
time-varying exposure
marginal structure model
Conflicting randomized controlled trial (RCT) results complicate evidence synthesis and regulatory decision making. For example, the Meis trial evaluating 17-α-hydroxyprogesterone caproate for preventing recurrent preterm birth (PTB) found a protective benefit, whereas the confirmatory PROLONG trial found no effect despite identical protocols. Differences in study populations are a hypothesized explanation, but existing transportability approaches to assess reconcilability are underpowered and do not account for unmeasured effect modifiers (EMs). We propose a regression-based proximal causal inference framework for reconciling RCTs using proxies for unmeasured EMs. We develop hypothesis tests for reconcilability of conditional causal effects on additive and multiplicative scales under generalized linear models. One test extends transportability methods to account for unmeasured EMs; another directly tests equality of model coefficients for improved power. We also introduce an equivalence test based on the mean squared distance between conditional causal effects. Simulations assess finite-sample performance, and the methods are applied to the PTB trials to evaluate reconcilability.
Keywords
Proximal causal inference
Reconcilability of randomized trials
Unmeasured effect modification
Structural causal models
Preterm birth
Equivalence testing
In a randomized trial with perfect compliance, unadjusted estimators of treatment efficacy are unbiased. However, when trial participants do not receive the treatment to which they were randomized (referred to as non-adherence), unadjusted estimators may incur substantial bias. Correction of this bias is made more complicated in a cluster-randomized trial (CRT) than in individually-randomized settings due to the hierarchical nature of the data. Building upon prior work for missing outcome data in CRTs, we propose a two-stage estimation procedure utilizing targeted maximum likelihood to obtain robust and asymptotically semi-parametric efficient estimators of treatment efficacy in the presence of non-adherence. We outline theoretical arguments under which our proposed procedure accounts for non-adherence at the cluster level, at the individual level, and at both levels simultaneously. In simulation experiments, we verify that the procedure corrects bias and enhances precision compared to unadjusted analyses. Our work provides a foundation for flexible and robust non-adherence adjustment in parallel-arm CRTs, leveraging causal machine learning to mitigate bias and improve efficiency.
Keywords
cluster randomized trial
non-adherence
causal inference
targeted maximum likelihood estimation
machine learning
two stage estimation
Negative controls--or proxies--are variables assumed or known not to be involved in certain causal pathways, and have historically been used as bias detection agents. Recently, proximal causal inference has emerged as a promising framework for using these variables to directly identify a causal relationship of interest in the face of unmeasured confounding. Proximal methods, however, rely on untestable, often opaque identifying assumptions involving so-called "bridge function" or "completeness" conditions. In this work, we relax these assumptions in a commonly adopted single outcome proxy setting, and show that the negative control non-trivially restricts the counterfactual outcome distribution. Moreover, we derive nonparametric, robust, efficient estimators of sharp bounds for mean counterfactuals. These bounds are non-smooth, non-closed-form solutions to linear programs involving potentially high-dimensional nuisance functions; our statistical approach has implications for a wide class of such challenging functionals. Practically, our proposal can be used to leverage proxies for causal and missing data problems, achieving sharp, valid inference under transparent assumptions.
Keywords
causal inference
unmeasured confounding
partial identification
negative controls
proximal causal inference
Chronic, inflammatory oral diseases have increasingly been suggested as a potential modifiable risk factor for dementia, yet causal relationships remain unclear. Prior literature has relied heavily on traditional case-control and proportional hazards models that inadequately control for time-varying confounders. In this paper, we employ a target trial (TTE) framework using Longitudinal Targeted Maximum Likelihood Estimation (LTMLE), a modern causal inference method which yields more robust estimates under time-varying confounding than traditional methods. Using data from the UK Biobank, including 164,282 participants aged ≥60, we estimate the causal effect of oral disease exposure timing on the 25-year incident dementia risk. We find that those exposed at baseline, 5 years, and 10 years demonstrate a consistent 22-24% increased 25-year risk of dementia compared to those not exposed (RRs: 1.22–1.24; 95% CIs 1.02–1.50). Overall, our findings are consistent with a causal relationship between long-term oral disease exposure and elevated dementia risk, underscoring the importance of early oral intervention and further research into oral disease-modifying treatment.
Keywords
causal inference
target trial emulation
Dementia
large-scale prospective cohort comprising extensive genomic, phenotypic, and longitudinal health data with decades of follow-up
Longitudinal Targeted Maximum Likelihood Estimation (LTMLE)
oral disease
Target trial emulation (TTE) applies the principles of randomized trials to observational studies, often to compare sustained treatment strategies. However, its utility for comparing two different schedules of the same repeated treatment, such as annual versus biennial mammography screening, is understudied. We present key obstacles and potential sources of bias in this setting and propose solutions. We outline two analytical approaches for a treatment schedule TTE: sequential trials with clone-censor-weighting and a marginal structural model. Through simulation, we compare the performance of these approaches for estimating marginal risk differences between annual and biennial treatment schedules over time for a survival outcome. We show that TTE with pooled sequential trials results in bias when eligibility criteria do not ensure sufficient washout, and that allowing eligibility after previous treatment changes the estimand regardless of washout. These biases can be mitigated by clearly defining baseline treatment history in the target population and standardizing estimates accordingly. This work highlights the pitfalls and implicit assumptions of methods common to TTE.
Keywords
Target trial emulation
Marginal structural model
Longitudinal data
Causal inference
Sequential trials
Survival analysis
Defining a start of follow-up for untreated individuals is often challenging in target trial emulations that compare treatment to no treatment. A common solution is matching treated and untreated individuals, but this approach can implicitly alter the target population and reduce statistical efficiency. We propose a causal estimand for treatment effects based on cumulative incidences that marginalize over treatment initiation time and baseline covariates. We develop a simple g-computation estimator using Cox models, along with a one-step estimator that accommodates flexible machine learning models for nuisance estimation. We apply our proposed estimators in simulations and in a study to assess the effectiveness of the Pfizer-BioNTech COVID-19 vaccine to prevent SARS-CoV-2 infections in children 5-11 years old. In both settings, we find that our proposed estimators yield similar scientific inferences while providing significant efficiency gains over commonly used matching estimators. These results suggest that our framework may provide a practical and more efficient alternative to matching for target trial emulation studies.
Keywords
Causal inference
Target trial emulation
Estimands
Vaccine effectiveness
Matching
Machine learning