Papers
You can find the complete list of papers here. It is possible to filter the papers by subtopics.
Each participant will present one paper. To express a preference, send us three numbers in order of preference. We will try to give everyone their first or second choice.
Before session 1. Please read the review papers #1 and #2 to have a better perspective about the background of the course. Papers tagged as Intro cannot be chosen for presentation.
1 Bayesian Causal Inference: A Critical Review
A critical review of Bayesian causal inference in the potential-outcomes framework, covering causal estimands, identification assumptions, assignment mechanisms, propensity scores, priors, covariate overlap, and sensitivity analysis. It also discusses instrumental variables and time-varying treatments.
A bridge between causal estimands and Bayesian analysis. Discusses how priors induce shrinkage and sparsity in parametric models and support sensitivity analysis around causal assumptions. It also surveys nonparametric Bayesian methods for point and time-varying treatments, with implementation examples.
3 The How and Why of Bayesian Nonparametric Causal Inference
A conceptual overview of the Bayesian nonparametric toolkit for causal inference. Discusses how to choose an appropriate method, how to avoid common pitfalls in high-dimensional settings, and why both the treatment selection and outcome processes often need to be considered.
4 Bayesian Inference for Causal Effects: The Role of Randomization
A foundational paper that defines causal effects as comparisons between potential outcomes and frames Bayesian causal inference as prediction of the missing potential outcomes. Explains the role of sampling, treatment assignment, and data-recording mechanisms, and when these mechanisms can be treated as ignorable.
5 Bayesian Nonparametric Modeling for Causal Inference
A core paper that uses Bayesian Additive Regression Trees (BART) to flexibly model the response surface in observational studies. In nonlinear simulation settings, BART accurately estimates average treatment effects and naturally accommodates treatment-effect heterogeneity.
6 Bayesian Nonparametric Generative Models for Causal Inference with Missing at Random Covariates
Models the joint distribution of the outcome, treatment, and confounders using an enriched Dirichlet process. The resulting generative model supports estimation of marginal and subgroup causal effects and imputes missing covariates within the same model under an assumption of ignorable missingness. A good contrast to BART and Bayesian Causal Forests.
Introduces Bayesian causal forests for estimating heterogeneous treatment effects in observational studies and addresses the phenomenon of regularization-induced confounding. Incorporates an estimated propensity function into the response model and separates the prognostic and treatment-effect components, allowing them to be regularized independently and the treatment effects to be shrunk towards homogeneity.
Proposes a semiparametric Bayesian tree-ensemble model for heterogeneous effects of a continuous exposure, moving beyond binary treatments. The model assumes that the exposure effect is linear while allowing its slope to vary flexibly across prespecified moderators.
A strong applied extension of Bayesian causal forests. Introduces targeted smooth Bayesian causal forests for treatment effects that vary smoothly over one target covariate. It induces smoothness through the terminal-node models and regularises the treatment and prognostic effects separately, with an application involving gestational age.
10 Shrinkage Bayesian Causal Forests for Heterogeneous Treatment Effects Estimation
The paper extends the idea of Bayesian Causal Forests to a sparsity inducing version which identifies covariates relevant to identify treatment-effect heterogeneity.
11 Bayesian Propensity Score Analysis for Observational Data
Develops a joint Bayesian propensity-score analysis that treats the propensity score as a latent quantity and propagates uncertainty about its estimation into the causal-effect analysis. This paper provides the starting point for the model-feedback problem considered in Paper 12.
12 Model Feedback in Bayesian Propensity Score Estimation
Shows that joint Bayesian estimation can allow information from the outcome model to feed back into the propensity-score model, thus biasing causal-effect estimates. It also examines how additional covariate adjustment can mitigate this problem.
13 Priors and Propensity Scores in Bayesian Causal Inference
Examines whether and how propensity scores should enter Bayesian causal inference. Under commonly used prior-factorization assumptions, the propensity-score model has no role in a Bayesian analysis. The paper explains why these assumptions can be problematic in high-dimensional settings and discusses Bayesian motivations for relaxing them.
14 Bayesian Doubly Robust Causal Inference via Posterior Coupling
Proposes a fully Bayesian version of doubly robust causal inference using posterior coupling. It incorporates propensity-score information through moment conditions, avoids conventional two-stage estimation, and is designed to prevent the model-feedback problem highlighted in Paper 12.
15 Bayesian Inference for the Causal Effect of Mediation
Bayesian nonparametric estimation of direct and indirect effects through a mediator, in the setting of a continuous mediator and a binary response.
16 Mediation Analysis Using Bayesian Tree Ensembles
Uses Bayesian tree ensembles to flexibly estimate mediation effects in the potential outcomes framework.
17 Estimating Heterogeneous Causal Mediation Effects with Bayesian Decision Tree Ensembles
Introduces a varying-coefficient BART model for estimating and regularising heterogeneous direct and indirect effects. The method shrinks effect estimates towards homogeneity and uses posterior summaries to identify and interpret subgroups with different mediation effects.
18 Bayesian Inference for Causal Effects in Randomized Experiments with Noncompliance
Foundational Bayesian noncompliance paper. Develops Bayesian inference for causal effects when treatment assignment is random but the treatment actually received is nonignorable because of noncompliance. Models latent compliance types and examines the roles of the exclusion restriction and monotonicity assumptions, with comparisons to intention-to-treat and instrumental-variable analyses.
19 Principal Stratification for Causal Inference with Extended Partial Compliance
Extends principal stratification to double-blind randomized trials with partial compliance and imperfect blinding, allowing compliance with the active treatment and placebo to differ. It states the additional assumptions required to estimate principal causal effects and causal dose–response relationships.
20 Assessing Causal Effects in the Presence of Treatment Switching Through Principal Stratification
Develops a Bayesian principal-stratification approach to treatment switching in randomized clinical trials. Defines latent strata according to patients’ potential switching behavior and estimates causal effects within those strata while accounting for censoring of switching and survival times.
21 Bayesian Regression Discontinuity Design with Unknown Cutoff
Introduces a Bayesian regression-discontinuity method for settings in which the treatment-assignment cut-off is unknown or uncertain. The method incorporates prior information about the cut-off, estimates the local treatment effect while propagating cut-off uncertainty, and can also be used to assess the validity of a suspected cut-off.
22 Inferring Causal Impact Using Bayesian Structural Time-Series Models
Uses a Bayesian diffusion-regression state-space model to predict the counterfactual trajectory that would have occurred without an intervention. Comparing this prediction with the observed post-intervention trajectory produces a posterior distribution for the causal impact over time, illustrated using an online advertising campaign.