Shape-Preserving Covariate Adjustment via Empirical Likelihood in Randomized Experiment

What the math gives to ML

The paper's transferable asset is not empirical likelihood alone, but the construction of one covariate-balanced probability measure whose every downstream functional automatically preserves distributional shape. This suggests replacing independently fitted quantile or CDF targets in a neural distributional head with a single empirical-likelihood-weighted target, avoiding crossing quantiles and non-monotone survival predictions caused by separate adjustment at each threshold or quantile. The weights provide exact finite-sample balance of selected covariate moments, while the shared weighted measure supports CDFs, quantiles, tail probabilities, and survival-derived quantities without separate correction procedures. The most practical first test is an auxiliary distributional-learning target for treatment-conditional neural predictors under covariate shift or imbalanced treatment groups.