Thursday, Aug 7: 8:30 AM - 10:20 AM
4205
Contributed Papers
Music City Center
Room: CC-212
Main Sponsor
Biopharmaceutical Section
Presentations
Causal inference plays an important role in precision medicine, which estimates the treatment effect for the "right" patients within a heterogeneous disease population. Artificial intelligence (AI) has emerged as a powerful tool for advancing causal inference, enabling the identification of subgroups where treatment effects vary significantly. This paper explores the integration of AI-based causal effect modeling for subgroup identification.
Our approach employs causal machine learning methods, such as causal forests, to estimate Conditional Average Treatment Effects (CATE) for individual latent subgroups characterized by features such as demographics, baseline disease characteristics, genetic markers, and outcomes. These methods go beyond traditional regression-based inference by modeling complex interactions and nonlinear relationships between features and treatment effects.
Applications of the methodology are demonstrated through simulation study on disease progression endpoints in a clinical trial, where AI-based causal modeling will be compared to traditional latent model approach in identifying patient subgroups that benefit most from a novel treatment
Keywords
Causal inference
AI-based
Precision medicine
Subgroup Identification
latent model
disease progression
Clinical trials are critical for determining the efficacy of medical interventions, yet the specific mechanisms by which these interventions produce their effects often remain unclear. This abstract introduces a Bayesian Mediation Analysis framework designed to shed light on the complex pathways through which treatment influences clinical outcomes. This approach:
1. Integrates Prior Knowledge.
Permits flexible modeling-both hierarchical and non-hierarchical-while incorporating informative priors, enhancing interpretability and robustness,
2. Naturally addresses Challenging Data Scenarios frequently encountered in clinical trials like: limited sample sizes, missing data, and nonlinear models,
3. Facilitates effect size comparisons.
Provides a systematic way to assess and compare direct versus indirect effects, and
4. Elucidates Causal Pathways by employing Gaussian Processes and Spline models
to capture heterogeneous underlying causal structures.
We demonstrate the method's utility through simulated datasets, showing how Bayesian Mediation Analysis can reveal nuanced treatment mechanisms and support more precise inferences about causal pathways.
Keywords
Bayesian Mediation Analysis, Gaussian processes, Spline Models, Effect Size Comparisons, Stan, MCMCregress and brm R packages
We propose a novel framework for causal inference that utilizes transfer learning methodologies to estimate the average treatment effect (ATE) in a target population where the primary outcome and negative control outcomes (NCOs) are unobserved. Our approach leverages a source dataset, which contains both the primary outcome and NCOs, to predict individual-level NCOs in the target dataset and calibrate the treatment effect estimate. Specifically, we develop a model that transfers information from the source population to the target population, using NCOs as a bias correction tool to enhance causal validity. We establish the identifiability conditions for our approach and derive implications for the observed data distribution. Through simulation studies, we demonstrate that our method accurately recovers the true ATE and improves bias correction with pseudo NCOs-calibration. To illustrate its practical utility, we apply our method to evaluate the impact of GLP-1 receptor agonists on mental health disorders across multiple clinical sites. Our findings highlight the potential of transfer learning-based causal inference in addressing challenges posed by incomplete outcome data setting.
Keywords
Causal inference
Transfer learning
Digital twin
Negative control outcomes
Bias calibration
Average treatment effect
In a treat-through study evaluating two therapeutic doses, randomization occurred at Week 0, and then after the same induction treatment, patients initiated one of the two randomized doses. A key question arises when evaluating the dose response at study completion: How should we account for differences observed at intermediate milestones, when patients received identical treatment prior to initiating their randomized dose? Specifically, what if the high dose had an early advantage that it maintained through the long term, or if the low dose outperformed in the short term but was eventually overtaken by the high dose? This presentation explores the use of propensity score weighting to analyze such data, adjusting for potential treatment differences (delta) between the two doses and assess the dose response when there is, or is not, a detectable delta at the study's conclusion. Simulation analyses were carried out to evaluate its operating characteristics.
Keywords
Causal Inference
Propensity Score
Dose response
Clinical trial design
Population-adjusted indirect comparisons (PAICs) are essential in health technology assessments (HTAs) for evaluating treatment effectiveness when direct head-to-head randomized controlled trials are unavailable. Simulated treatment comparison (STC) is a key PAIC method to adjust for baseline differences between trial populations and enable valid comparisons. However, traditional STC methods regress baseline covariates at the individual level in the index trial but estimate conditional treatment effects at mean covariate values in the comparator trial. This mismatch introduces aggregation bias in marginal treatment effect estimates. This paper introduces STC with Direct Marginalization (STC-DM), a novel extension to improve marginal treatment effect estimation. By directly accounting for baseline covariate heterogeneity, STC-DM mitigates aggregation bias and enhances population comparability. Simulation studies demonstrate its superior accuracy and robustness over standard STC methods. The presentation concludes with a discussion on STC-DM's implications and applications, highlighting its potential to enhance the validity and reliability of comparative effectiveness evaluations.
Keywords
Indirect treatment comparisons
Marginal treatment effect
Population adjustment
Simulated treatment comparison
Non-randomized studies are increasingly used to support decisions on the comparative effectiveness of interventions. However, identifying valid causal effects relies on certain assumptions. The assumption of no unmeasured confounding cannot be statistically tested, but a range of methods for sensitivity analysis have been developed. This research aimed to guide the selection of suitable methods to assess the robustness of study results to unmeasured confounding. Current recommendations on sensitivity analysis from regulatory and health technology assessment (HTA) guidelines were summarized. Commonly used methods were evaluated, along with recent approaches designed to overcome the limitations of established techniques. Methods were further categorized based on confounder measurability in external data sources and ease of implementation. In this presentation, key findings will be summarized; practical considerations, such as the assessment goals, nature of the unmeasured confounders, and availability of information on confounders, will be discussed to facilitate the selection of methods and promote transparency in reporting; and methods will be illustrated through examples.
Keywords
Unmeasured confounding
Sensitivity analysis
Non-randomized studies
Mediation analysis aims to elucidate the intermediate variables or processes that mediate the relationship between an intervention and its ultimate impact on outcomes.This multifaceted approach enhances our understanding of not only whether an intervention works but also how it works, providing a comprehensive perspective for researchers, clinicians, and policymakers.By dissecting these pathways, researchers gain valuable insights into the causal chain of events, informing the development of more targeted and effective interventions.This presentation will cover both Bayesian informative and noninformative priors and frequentist approaches.Sizes of direct and indirect effects of treatment on dependent variable will be examined.Partial correlations between mediating paths will also be accounted for in modeling by including the correlation of their error term variances. Presenting a totality of evidence from using both approaches within a regulatory context can offer a more comprehensive and informed response, especially when navigating complex datasets and topics. We illustrate our findings by assessing the improvements in functioning scores in patients with MDD with simulations.
Keywords
Mediation Analysis
Bayesian
Frequentist
Informative and non-informative priors