Finite-Population Inference with ML-Based Predictions

Predictions from machine learning can be used at multiple stages of a survey as "basic ingredients" for estimating finite-population parameters. This course explains how to use ML-based predictions to develop valid statistical inference for model-assisted estimation and imputation with item nonresponse. For unit-nonresponse weighting, we clarify the challenges ML introduces for inverse-probability weighting and what can—and cannot—currently be justified. We cover standard point estimators and doubly robust approaches, variance estimation via cross-fitting and influence functions, and the construction of asymptotically valid confidence intervals. Practical considerations such as estimator selection, hyperparameter tuning, and weight trimming, are emphasized. By the end, attendees will have a clear map of state-of-the-art methods for drawing valid inference when ML predictions are used in the survey pipeline, and an understanding of the open problems for unit nonresponse.