Predictive models are increasingly used in surveys for tasks such as model-based and model-assisted estimation, as well as handling nonresponse through imputation and reweighting. The rise of statistical learning has provided survey statisticians with flexible tools, but incorporating them into survey estimation strategies poses significant challenges for valid inference. In particular, variance estimation becomes delicate when using black-box machine learning methods.

I will show that naively plugging machine learning predictions into classical estimators can lead to invalid inference and misleading uncertainty quantification. Through theoretical and empirical results, I will illustrate why many standard approaches fail with nonparametric regression. To address this, I will present an extension of the Double Machine Learning framework with cross-fitting to survey sampling that supports valid inference with arbitrary black-box methods. Under realistic conditions, the resulting estimators are shown to be square-root n consistent and asymptotically normal. I will also introduce new variance estimators based on cross-fitting that remain consistent across methods, enabling the construction of asymptotically valid confidence intervals. Issues related to model selection and aggregation will also be discussed. Simulation studies demonstrating the strong performance of the proposed methods will be presented.