Spatially Robust Inference with Predicted and Missing at Random Labels

A team of researchers has published a new paper addressing a critical, real-world problem in applied AI and statistics. When scientists use predictions from machine learning models (like GPT-4 or Llama 3) to fill in missing or expensive-to-collect data—a common practice in fields from epidemiology to economics—standard statistical methods for calculating uncertainty (confidence intervals) break down. This is especially true when the data has spatial patterns or when labels are Missing at Random (MAR), a common scenario in observational studies. The paper, 'Spatially Robust Inference with Predicted and Missing at Random Labels,' identifies that a popular technique called cross-fitting, used to estimate model parameters, introduces artificial correlations that distort variance estimators, leading to unreliable results.

To solve this, the authors propose a novel 'jackknife spatial heteroscedasticity and autocorrelation consistent (HAC) variance correction.' This method cleverly separates the genuine spatial dependence in the data from the noise induced by the cross-fitting procedure itself. The result is a 'doubly robust' estimator that produces confidence intervals that are asymptotically valid under standard conditions. In practical tests on benchmark datasets and simulations, their method showed 'substantial improvement in finite-sample calibration,' meaning the reported 95% confidence intervals actually contain the true parameter 95% of the time, a crucial requirement for trustworthy science. This work bridges a significant gap between modern AI prediction tools and rigorous statistical inference, providing a robust framework for decision-making when relying on imputed data.