Prediction-Powered Inference (PPI) was introduced in 2023 as a method for combining machine learning predictions with labeled data to estimate population quantities. The idea: use a large unlabeled dataset scored by a predictive model, then correct the model's systematic errors using a smaller labeled validation set.
This paper shows that PPI is algebraically equivalent to the difference estimator of Cassel, Särndal, and Swensson (1976). PPI++ — the improved variant with a tuning parameter — is equivalent to the generalized regression (GREG) estimator. Both were developed in survey sampling theory fifty years ago.
The correspondence is exact, not approximate. The same formula, the same variance reduction mechanism, the same optimality properties. The survey sampling literature derived these estimators for combining auxiliary information (census data, administrative records) with probability samples. PPI rediscovered them for combining model predictions with labeled data. The mathematical structure is identical because the statistical problem is identical: you have an imperfect proxy for the quantity of interest and a smaller dataset that lets you measure the proxy's error.
The paper's tone is constructive rather than dismissive. Both communities benefit from recognizing the connection. PPI researchers gain access to decades of theory on calibration, post-stratification, and non-response adjustment. Survey statisticians gain access to PPI's extensions to new problem types and its software ecosystem.
But the deeper lesson is about disciplinary amnesia. A foundational technique can be reinvented from scratch, published in top venues, and widely adopted — all without anyone noticing it already existed in an adjacent field. The barrier to recognizing prior work is not obscurity but vocabulary. The same idea dressed in different notation becomes invisible.
Note (2026-11-06): I checked this essay against the raw arXiv abstract of Mozer, 'PPI is the Difference Estimator: Recognizing the Survey Sampling Roots of Prediction-Powered Inference' (2603.19160, v1). The abstract names the 1976 difference estimator only as 'Cassel et al. (1976)'; from my own knowledge, not the abstract, that 1976 paper is by Cassel, Sarndal and Wretman, so 'Cassel, Sarndal, and Swensson' (a later book) looks wrong. It says 'PPI plus' corresponds to the generalized regression (GREG) estimator of Sarndal et al. (2003), so my 'both were developed fifty years ago' is loose for GREG. The author argues the equivalence and then considers what PPI inherits, what is genuinely new, and where inferential claims 'require care' (mode of inference, the role of the unlabeled pool, differential prediction error for subgroup estimands such as the average treatment effect); my text says the variance reduction mechanism and optimality properties are the same without those qualifications. I checked the abstract and author list only.