Causal Foundation Model
A class of models that use foundation models for causal inference. They can estimate treatment effects at the unit level, but when aggregating at the group level they may ignore error correlations, leading to inaccurate intervals.
Timeline (last 90 days)
Causal foundation models (CFMs) perform impressively on unit-level treatment effect estimation, but their confidence intervals for group-level average effects are severely miscalibrated.
The coverage collapse is a systematic problem rather than random error, meaning the risk of decisions based on these intervals is significantly underestimated.
CFMs simply combine unit-level variances at inference time, implicitly assuming that errors across different units are independent.
When making decisions that rely on group-level average effects, one should not blindly trust the aggregated intervals output by unit-level models, and should consider a jointly predictive model or post-hoc calibration.
All Facts (4)
Causal foundation models (CFMs) perform impressively on unit-level treatment effect estimation, but their confidence intervals for group-level average effects are severely miscalibrated.
50%UnverifiedThe coverage collapse is a systematic problem rather than random error, meaning the risk of decisions based on these intervals is significantly underestimated.
50%UnverifiedCFMs simply combine unit-level variances at inference time, implicitly assuming that errors across different units are independent.
50%UnverifiedWhen making decisions that rely on group-level average effects, one should not blindly trust the aggregated intervals output by unit-level models, and should consider a jointly predictive model or post-hoc calibration.
50%