Prediction-powered inference
New paper in Science on using ML-based methods to improve classical confidence intervals or p-values. (also on arXiv) The idea is to use a large collection of unlabeled data along with an ML method to estimate labels on the unlabeled data. Then we can use that data to improve the confidence interval given all the data. This reduces the amount of data you need to make a key “discovery” as shown in Table 2.
