CausalPFN
A causal foundation model based on a Prior-Fitted Network (PFN) for individual treatment effect estimation. Its interval coverage is insufficient when aggregating at the unit level.
Timeline (last 90 days)
On the ACIC 2016 dataset, temperature-calibrated CausalPFN 95% intervals have coverage below the nominal level when the group size exceeds 30 units (out of 4802 total units).
The CausalPFN interval for the full-sample average effect contains the true value in only 20% of the datasets, far below the expected 95%.
On all six benchmarks, the jointly predictive model, used as a head on top of a frozen CausalPFN, greatly improves coverage performance.
Using the jointly predictive model as a head on top of a frozen CausalPFN is a modular repair strategy that preserves the original unit-level point estimation capability and only replaces the way intervals are constructed.
All Facts (4)
On the ACIC 2016 dataset, temperature-calibrated CausalPFN 95% intervals have coverage below the nominal level when the group size exceeds 30 units (out of 4802 total units).
50%UnverifiedThe CausalPFN interval for the full-sample average effect contains the true value in only 20% of the datasets, far below the expected 95%.
50%UnverifiedOn all six benchmarks, the jointly predictive model, used as a head on top of a frozen CausalPFN, greatly improves coverage performance.
50%UnverifiedUsing the jointly predictive model as a head on top of a frozen CausalPFN is a modular repair strategy that preserves the original unit-level point estimation capability and only replaces the way intervals are constructed.
50%