Modern experiments increasingly collect rich post-treatment objects—such as text, images, and video—that may provide early surrogate signals for outcomes that are delayed, expensive, or observed only for a subset of units. A new paper on arXiv introduces SIRNET, a method designed to exploit these signals for treatment-effect estimation.
SIRNET performs causal adaptation of pretrained embeddings, integrating them into a semiparametric framework under the assumption of surrogacy. This allows the model to use high-dimensional surrogate information while maintaining interpretable treatment-effect estimates, even when the true outcome is not fully observed.
The approach is particularly relevant for settings where traditional outcome measurement is costly or time-consuming, and where off-the-shelf embeddings can be repurposed for causal inference. The paper is available as arXiv:2610.14011.