Research map / Probabilistic and scientific modeling / Treatment effects
Bayesian Neural Controlled Differential Equations for Treatment Effect Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BNCDE, a Bayesian neural controlled differential equation that couples neural CDEs with latent neural SDEs over their weights to give posterior predictive distributions of continuous-time potential outcomes under treatment.
Problem
Individualized treatment effect estimation over time in continuous time from observational data lacks principled Bayesian uncertainty quantification; existing continuous-time method TE-CDE gives only point estimates (with ad-hoc MC dropout).
Approach
Model the time-varying weights of neural CDE encoder and decoder as stochastic processes driven by latent neural SDEs trained with variational Bayes (ELBO with path-space KL against OU priors), yielding full posterior predictive distributions of potential outcomes that capture model and outcome uncertainty.
Key results
- BNCDE credible intervals are faithful (empirical coverage at least the nominal level), whereas TE-CDE with MC dropout is overconfident with under-coverage.
- BNCDE CrIs are significantly sharper (narrower) than TE-CDE's across all prediction windows.
- BNCDE achieves clearly lower MSE of point estimates than TE-CDE, including under varying noise levels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BNCDE instead of TE-CDE with MC dropout | Treatment effect estimation on the tumor simulation · Tumor simulation, prediction windows Δ=1, 2, 3 | empirical coverage of credible intervals | Helped |
| Use BNCDE instead of TE-CDE with MC dropout | Treatment effect estimation on the tumor simulation · Tumor simulation, prediction windows Δ=1, 2, 3 | median credible interval width | Helped |
| Use BNCDE instead of TE-CDE with MC dropout | Treatment effect estimation on the tumor simulation · Tumor simulation, prediction windows Δ=1, 2, 3 | MSE of point estimates | Helped |
| Use BNCDE instead of TE-CDE with MC dropout | Treatment effect estimation on the tumor simulation · Tumor simulation, model uncertainty comparison | informativeness of model uncertainty | Helped |
| Use BNCDE instead of TE-CDE with MC dropout | Treatment effect estimation on the tumor simulation · Tumor simulation with varying noise levels | MSE of point estimates | Helped |
| Use BNCDE instead of TE-CDE with MC dropout | Treatment effect estimation on the tumor simulation · Tumor simulation with informative sampling | overall performance | Helped |
Surprising findings
- MC dropout estimates of model uncertainty in TE-CDE are nearly uninformative, with MSE decreasing almost randomly under deferral.
Limitations
- Black-box, not explainable; ELBO depends on diffusion hyperparameter with noticeable effect on conservatism.
- Training requires solving high-dimensional SDEs, the computational bottleneck; does not account for confounding bias.
- Evaluated only on simulated tumor growth data; needs sufficient data for robust estimation.
Experimental setup
Datasets: simulated lung cancer tumor growth data (PK-PD model)
Benchmarks: simulated tumor growth data with prediction windows 1, 2, 3 days
Metrics: empirical coverage of credible intervals median CrI width MSE of Monte Carlo mean estimates normalized MSE vs deferral rate
Baselines: TE-CDE with MC dropout
Builds on: neural controlled differential equations neural stochastic differential equations TE-CDE Ornstein-Uhlenbeck priors
Closest papers by meaning
- ODE Discovery for Longitudinal Heterogeneous Treatment Effects Inference (ICLR 2024 · similarity 0.87)
- Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time (NeurIPS 2026 · similarity 0.86)
- Treatment Effect Estimation for Optimal Decision-Making (NeurIPS 2025 · similarity 0.86)
- Continuous Treatment Effect Estimation Using Gradient Interpolation and Kernel Smoothing (AAAI 2024 · similarity 0.85)
- Quantifying Aleatoric Uncertainty of the Treatment Effect: A Novel Orthogonal Learner (NeurIPS 2024 · similarity 0.85)
- Stabilized Neural Prediction of Potential Outcomes in Continuous Time (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.