DeepBlip: Estimating Conditional Average Treatment Effects Over Time

Haorui Ma, Dennis Frauen, Stefan Feuerriegel

Keywords: causal inference, treatment effect estimation, structural nested mean models, medical time-series data, personalized medicine

TL;DR: We are the first to build a neural framework using blip functions to estimate CATE over time, which provides a stable learning framework for efficient offline evaluation of personalized treatment strategies over long time horizons.

Abstract: Estimating the conditional average treatment effect (CATE) over time is crucial for making personalized decisions in medicine. Yet, existing neural methods for this task have limitations: they either (1) do not adjust for time-varying confounding and are thus biased (e.g., causal transformer), or (2) become unstable over long time horizons because the method has to learn the full counterfactual outcome trajectories (e.g., MSNs, G-computation). To address these limitations, we propose DeepBlip, the first neural framework that leverages the blip function from structural nested mean models to break the joint effect of treatment sequences over time into localized, time-specific ``blip effects''. As a result, we learn a simpler estimand that does not require learning full counterfactual outcome trajectories, which is thus more stable over long horizons. Further, our DeepBlip adjusts for time-varying confounding and is thus unbiased. Our DeepBlip seamlessly integrates sequential models like LSTMs or transformers to capture complex temporal dependencies. Our DeepBlip has two further strengths for medical practice: (i) The loss is Neyman-orthogonal, meaning it is robust against model misspecification. (ii) The blip effects can be used to predict treatment effects for new treatment sequences without re-computation, which allows to identify optimal treatment sequences through offline evaluation. Finally, we evaluate our DeepBlip across various clinical datasets, where it achieves state-of-the-art performance.

Primary Area: causal reasoning

Submission Number: 7633