To learn more about the code powering our lab, follow us on GitHub or Hugging Face.
–
causalimages: This package enables causal inference with image and image-sequence data—spanning Earth observation, biomedical, and social-science imagery—by turning pixels into estimands. It provides tools for image-based treatment-effect heterogeneity and image-based deconfounding, plus utilities for producing embeddings and efficiently handling large image corpora. Designed for reproducible, model-agnostic workflows, causalimages lets researchers diagnose where effects are strongest and control confounding directly from images without bespoke deep-learning pipelines. [Preprint] [.bib]
unshrink: unshrink is a lightweight Python package for post-hoc debiasing of ML predictions so they can be used for downstream causal inference without collecting new labels. It implements two complementary corrections: Linear Calibration Correction (LCC), which rescales predictions via a fitted intercept and slope, and Tweedie’s correction, which applies a Tweedie-style score-based adjustment to undo shrinkage. The API treats models as black boxes and returns corrected predictions (and treatment-effect estimates) with uncertainty estimates. [Article] [.bib]
