A Causal Transformation Model for Interval-Censored Data with Unobserved Confounding
Principal Investigator
Name
Rosalba Radice
Institution
Bayes Business School, City St George’s, University of London
Position Title
Professor of Statistics
Email
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCO-1772
Initial CDAS Request Approval
Dec 26, 2024
Title
A Causal Transformation Model for Interval-Censored Data with Unobserved Confounding
Summary
Motivated by studies investigating causal effects in survival analysis, we propose a structural bivariate transformation model to evaluate the impact of a binary treatment on a time-to-event outcome subject to interval censoring. The model incorporates unobserved confounding through a bivariate Gaussian distribution, where the dependence parameter captures the underlying confounding effects. Additive predictors are used to flexibly account for the influence of observed confounders.
To estimate the baseline survival function, we employ monotonic P-splines, ensuring smooth and reliable estimation. Regularization techniques, such as ridge penalties, are applied to control the effects of binary or categorical instruments, while interactions between treatment and modifier variables are incorporated to accommodate potential heterogeneity in treatment effects across subgroups. A key causal measure provided by this approach is the survival average treatment effect, offering an intuitive and interpretable summary of treatment impact.
Parameter estimation is performed using a computationally efficient and stable penalized maximum likelihood estimation method, with associated inferential results enabling the construction of confidence intervals. To illustrate the practical utility of the proposed framework, we analyse data from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial.
The entire modelling framework is implemented in the R package GJRM, facilitating easy application for researchers and practitioners interested in causal inference for survival data with interval censoring and unobserved confounding.
To estimate the baseline survival function, we employ monotonic P-splines, ensuring smooth and reliable estimation. Regularization techniques, such as ridge penalties, are applied to control the effects of binary or categorical instruments, while interactions between treatment and modifier variables are incorporated to accommodate potential heterogeneity in treatment effects across subgroups. A key causal measure provided by this approach is the survival average treatment effect, offering an intuitive and interpretable summary of treatment impact.
Parameter estimation is performed using a computationally efficient and stable penalized maximum likelihood estimation method, with associated inferential results enabling the construction of confidence intervals. To illustrate the practical utility of the proposed framework, we analyse data from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial.
The entire modelling framework is implemented in the R package GJRM, facilitating easy application for researchers and practitioners interested in causal inference for survival data with interval censoring and unobserved confounding.
Aims
1. Propose a flexible transformation structural model that can estimate a cuasl effect of a binary treatment on on a time-to-event outcome subject to interval censoring.
2. Apply the novel approach to data from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial.
3. Estimate the causal effect of flexible sigmoidoscopy in reducing the risk of colorectal
cancer occurrence as compared to the usual care.
4. Implement the framework in the R package GJRM, facilitating easy application for researchers and practitioners interested in causal inference for survival data with interval censoring and unobserved confounding.
Collaborators
Giampiero Marra, Department of Statistical Science, University College London