Many real-world interventions are high-dimensional—think of a therapeutic message or an earnings call transcript. Classical causal estimators typically require that every possible intervention be observed, which is impossible when the intervention space is as large as all text strings. This paper adapts the idea of recasting causal inference as a supervised learning problem to handle such settings.

The authors show that the error of a learned causal effect estimator decomposes into moment-balancing errors of increasing order. They exploit the structure of the conditional outcome given confounders to design objectives that directly target these errors. They also introduce a way to project a high-dimensional treatment onto coarser attributes, so a single model can answer multiple causal questions without retraining for each attribute.

In experiments with continuous, discrete, and text treatments, the estimators generalize to held-out treatments. The results show that balancing higher-order moments helps when needed, and projected causal estimates are competitive with models trained specifically for each attribute.