When Prediction Error Is Not Enough: Evaluating Nuisance-Function Prediction for Causal Estimation
The episode discusses a paper evaluating nuisance-function prediction error for causal estimation. Hosts discuss how joint measures like D joint are not strong enough to predict causal bias, emphasizing the need to consider sample-to-sample variation and data structure complexity in simulations. The key takeaway is that high predictive accuracy does not guarantee reliable causal inference, pushing for metrics that ensure statistical stability.