Prediction Powered Inference (PPI)

• Prediction Powered Inference (PPI) is a framework that integrates large-scale predictive estimates with gold-standard labels to correct bias in statistical inference.
• It decomposes estimators into a prediction term and a rectifier term, ensuring valid inference even when the prediction model is imperfect.
• The method extends to various applications including survey sampling, decentralized data, and federated learning, with tuning parameters adapting to prediction quality.

Prediction-powered inference (PPI) is a statistical framework for inference with partially labeled data in which a large set of machine-generated predictions is combined with a smaller gold-standard labeled subset through an explicit bias-correction step. In its canonical form, PPI estimates a prediction-based quantity on a large unlabeled sample and then rectifies it using labeled observations, yielding valid confidence intervals for targets such as means, quantiles, and regression coefficients without requiring the prediction model to be correct (Angelopoulos et al., 2023). Subsequent work has reformulated PPI as bias-corrected $M$-estimation, linked it to survey sampling and missing-data theory, and extended it to informative labeling, decentralized data, sequential inference, bootstrap procedures, and multi-expert aggregation (Lee et al., 7 Jun 2026).

1. Core formulation

The standard PPI setting has two data sources: a small labeled sample $(X_i,Y_i)$ and a much larger sample of covariates $X_i$ for which a predictor $f$ supplies $\hat Y_i=f(X_i)$. For mean estimation, the basic estimator averages predictions over the large sample and adds a labeled-sample estimate of prediction error: $$\hat\theta_{\mathrm{PPI}} = \frac{1}{N}\sum_{i=1}^{N}\hat Y_i + \frac{1}{n}\sum_{i=1}^{n}(Y_i-\hat Y_i).$$ The second term is the “rectifier”: it estimates the average prediction error on the labeled data and corrects the full-sample prediction mean (Mozer, 19 Mar 2026).

A closely related finite-population presentation writes the target as

$\theta^*=\frac{1}{N}\sum_{i=1}^N Y_i$

and decomposes the estimator into a prediction term and a rectifier term,

$$\hat{\theta}_{\mathrm{PPI}} = \frac{1}{N}\sum_{i=1}^N \hat Y_i - \frac{1}{n_{\mathrm{lab}}}\sum_{i:R_i=1}(\hat Y_i-Y_i),$$

which is the same construction up to sign convention for the residual (Datta et al., 13 Aug 2025). In both formulations, the substantive idea is identical: predict everywhere, then use labeled outcomes only to estimate and remove the bias induced by replacing $Y$ with $\hat Y$.

For general inference problems, PPI is expressed through an estimating equation. One formulation defines a full-data target by

$(X_i,Y_i)$0

and replaces the infeasible full-data score with

$(X_i,Y_i)$1

Under simple random sampling without replacement, this score is design-unbiased for the full-data score $(X_i,Y_i)$2, creating what the later theory calls an “oracle bridge” between the computable prediction-assisted score and the infeasible full-data estimating equation (Lee et al., 7 Jun 2026).

2. Estimands and inferential mechanics

The original PPI framework was developed for quantities such as means, quantiles, and linear and logistic regression coefficients, and more generally for parameters defined as minimizers of expected convex losses (Angelopoulos et al., 2023). In that setting, if

$(X_i,Y_i)$3

then PPI decomposes the corresponding subgradient or estimating function into an imputed term and a rectifier,

$(X_i,Y_i)$4

where $(X_i,Y_i)$5 is computed from predictions and $(X_i,Y_i)$6 is estimated from labeled data (Cortinovis et al., 4 Feb 2025).

The associated confidence set construction is modular. One builds a confidence set for the prediction-based term and another for the rectifier, and then combines them through a Minkowski sum: $(X_i,Y_i)$7 This yields valid inference because $(X_i,Y_i)$8 belongs to the sum at the true parameter when both component intervals are correct (Luo et al., 2024). The same logic underlies the original convex-analysis presentation, where the inference problem is reduced to testing whether the rectified estimating equation includes zero (Angelopoulos et al., 2023).

Later work generalizes this scope beyond $(X_i,Y_i)$9-estimation. “Generalized Prediction-Powered Inference” extends the construction to any regular asymptotically linear estimator, allowing PPI analogues for functionals such as true positive rate, false positive rate, and area under the curve (Zou et al., 10 Feb 2026). This broadens PPI from a specific family of convex-loss procedures to a general recipe for prediction-assisted debiasing of regular estimators.

Computation has also been a major design axis. PPI++ replaces the original grid-based confidence-set inversion with a single optimization problem plus asymptotic covariance estimation, giving easy-to-compute confidence sets for parameters of any dimensionality and making generalized linear model applications substantially יותר tractable in practice (Angelopoulos et al., 2023).

3. Statistical properties, efficiency, and finite-sample limits

A central property of PPI is validity under imperfect prediction. The labeled sample is not used to validate a claim that the prediction model is correct; it is used to estimate the model’s error and propagate that uncertainty into the interval width (Angelopoulos et al., 2023). In the $X_i$0-estimation theory, consistency and asymptotic normality follow under simple random sampling without replacement, standard regularity conditions, and a nonsingular Jacobian for the full-data estimating equation (Lee et al., 7 Jun 2026).

The most favorable asymptotic regime occurs when the predictor is score-calibrated: $X_i$1 Under this condition, the PPI estimator attains the semiparametric efficiency lower bound. For mean estimation, score calibration reduces to the familiar regression condition $X_i$2 (Lee et al., 7 Jun 2026).

Efficiency, however, is not unconditional. PPI++ introduces a tuning parameter $X_i$3, with $X_i$4 giving standard PPI and $X_i$5 giving classical labeled-only inference. Asymptotically, the tuned procedure adapts to prediction quality and is designed to be no worse than classical inference using only the labeled data (Angelopoulos et al., 2023). In survey-sampling notation, this same tuning corresponds to the GREG-style coefficient in the estimator

$X_i$6

which interpolates between ignoring predictions and fully using them (Mozer, 19 Mar 2026).

Finite-sample analysis complicates the asymptotic picture. For mean estimation, “No Free Lunch: Non-Asymptotic Analysis of Prediction-Powered Inference” shows that PPI++ outperforms labeled-only estimation if and only if the correlation between pseudo- and gold-standard labels is above a sample-size-dependent threshold; in the Gaussian case, the threshold is

$X_i$7

The same paper shows that single-sample PPI++ is biased in finite samples and that its usual plug-in variance estimate can be overly optimistic (Mani et al., 26 May 2025).

A parallel critique appears in least-squares prediction-based inference. “Another look at inference after prediction” shows that standard PPI is valid but not always efficient, and proposes the Chen–Chen weighted augmentation, which guarantees efficiency improvement relative to labeled-only estimation in the least-squares setting. The paper explicitly notes that PPI can be less efficient than using only the labeled sample when predictions are weak (Gronsbell et al., 2024). Taken together, these results place PPI’s efficiency claims on a conditional basis: validity is robust, but efficiency depends on predictive quality, tuning, and the inferential regime.

4. Survey-sampling and missing-data roots

A major reappraisal of PPI identifies its core mean estimator as exactly the classical difference estimator from survey sampling: $X_i$8 The same paper shows that PPI++ is the generalized regression (GREG) estimator in survey-sampling notation (Mozer, 19 Mar 2026). This equivalence does not erase the distinct contemporary framing of PPI, but it does place its mean-estimation backbone squarely inside model-assisted survey estimation.

That comparison also clarifies the inferential distinction. In survey sampling, the finite population is fixed and randomness arises from the sampling design; in much of the PPI literature, the data are treated as i.i.d. draws from a superpopulation. The corresponding variance formulas differ in form but become asymptotically close when $X_i$9. In the survey-sampling presentation, the design-based variance of the difference estimator under simple random sampling without replacement is

$f$0

whereas a common superpopulation PPI variance expression is

$f$1

Both depend on the variability of the residuals rather than only on the raw variability of $f$2 (Mozer, 19 Mar 2026).

The survey connection becomes even more explicit under informative labeling. When the labeled sample is not simple random and

$f$3

the unweighted rectifier is generally biased. “Prediction-Powered Inference with Inverse Probability Weighting” replaces it with Horvitz–Thompson or Hájek residual corrections,

$f$4

where $f$5. Standard PPI is recovered as the special case of equal inclusion probabilities (Datta et al., 13 Aug 2025). This extension places PPI inside the classical inverse-probability-weighting and unequal-probability sampling framework rather than outside it.

A complementary missing-data interpretation appears in the generalized ALE framework, where PPI is described as a computationally simple alternative to semiparametric missing-data estimators. That literature shows that PPI is generally not semiparametrically efficient outside restrictive oracle cases, but remains attractive because of its simpler construction (Zou et al., 10 Feb 2026).

5. Major variants and extensions

The PPI literature has developed by modifying either the prediction term, the rectifier, the sampling structure, or the inferential regime. Representative variants are summarized below.

Variant	Core modification	Paper
PPI++	Power tuning $f$6; optimization-based confidence sets	(Angelopoulos et al., 2023)
IPW-adjusted PPI	Horvitz–Thompson/Hájek residual correction under informative labeling	(Datta et al., 13 Aug 2025)
Fed-PPI	Federated model training and aggregation of local PPI statistics	(Luo et al., 2024)
StratPPI	Stratum-specific PPI with tailored allocation and tuning	(Fisch et al., 2024)
Bayesian/FAB/anytime-valid PPI	Posterior proxy estimands, Bayes-assisted rectifier intervals, confidence sequences	(Hofer et al., 2024, Cortinovis et al., 4 Feb 2025, Kilian et al., 23 May 2025)
Bootstrap/generalized PPI	GLM-calibrated bootstrap inference; extension to any regular ALE	(Efron, 26 Jun 2026, Zou et al., 10 Feb 2026)

Fed-PPI adapts PPI to private data silos. Each client trains a local model, federated learning aggregates those models, and local prediction terms, rectifiers, and variance estimates are then aggregated into a global confidence interval. The paper evaluates galaxy classification, Amazon deforestation, protein PTM odds ratios, gene-expression quantiles, health insurance logistic regression, and income linear regression, and reports that federated intervals are close to centralized PPI intervals even under Non-IID settings, though heterogeneity can widen intervals (Luo et al., 2024).

Stratified PPI exploits heterogeneity in predictor quality across strata. Under squared loss and natural allocation, the paper proves that StratPPI is asymptotically no worse than unstratified PPI++, with strict improvement when stratum-specific conditional means differ across strata. Empirically, it yields substantially tighter intervals in language-model evaluation settings such as Seahorse, AttributedQA, and Galaxy (Fisch et al., 2024).

Bayesian variants use posterior modeling of proxy quantities rather than purely analytic rectifier intervals. “Bayesian Prediction-Powered Inference” introduces difference, stratified, and chain-rule estimators tailored to autoraters with discrete outputs or nonlinear bias patterns (Hofer et al., 2024). FAB-PPI adds Bayes-assisted rectifier shrinkage with frequentist coverage guarantees and emphasizes the horseshoe prior because it shrinks aggressively near zero rectification while reverting toward standard PPI in low-prior-probability regions (Cortinovis et al., 4 Feb 2025). Anytime-valid PPI extends the framework to confidence sequences using Ville’s inequality and the method of mixtures, thereby providing time-uniform validity as labeled and unlabeled data accumulate (Kilian et al., 23 May 2025).

Other extensions broaden either the inferential targets or the prediction architecture. A bootstrap-based alternative recalibrates $f$7 through a low-dimensional GLM and uses a two-level bootstrap to avoid asymptotic variance derivations; under that model, unlabeled data may help substantially for correlation-type functionals but may add little for estimating $f$8 (Efron, 26 Jun 2026). PAS combines within-task PPI debiasing with empirical-Bayes shrinkage across many parallel mean-estimation problems (Li et al., 20 Feb 2025). Local PPI adapts the rectifier idea to local multivariable regression for $f$9 and $\hat Y_i=f(X_i)$0 (Gu et al., 2024). MOE-powered inference replaces a single predictor with a mixture of experts chosen to minimize PPI variance and obtains a best-expert guarantee (Gu et al., 30 Apr 2026). The same prediction-powered correction template has even been extended from inference to semi-supervised model training via an unbiased gradient estimator with online power tuning (Shoham et al., 26 Oct 2025).

6. Assumptions, diagnostics, and empirical behavior

The original practical workflow emphasizes three core assumptions: labeled and unlabeled data are comparable, corresponding to MCAR in the baseline setup; the predictor is trained externally and does not overlap with the inference sample; and the covariates required for prediction are observed for all units (Song et al., 28 Jan 2026). The same synthesis recommends covariate-distribution diagnostics such as standardized mean differences, Kolmogorov–Smirnov tests, and energy distance, while stressing that the main assumptions are only partially testable from observed data (Song et al., 28 Jan 2026).

When the predictor is learned from the labeled data used for inference, label reuse becomes a central failure mode. The M-estimation optimality paper shows that vanilla PPI with a learned predictor can severely undercover, especially when labels are scarce, whereas cross-fitting and single-fit variance correction restore near-nominal 95% coverage in its simulations (Lee et al., 7 Jun 2026). The practical review makes the same point empirically: in the Mosaiks housing data, double-dipping produced anti-conservative confidence intervals and coverages, with some coefficient coverages falling to around 50% in the reported experiment, whereas holdout training maintained nominal behavior (Song et al., 28 Jan 2026).

Sampling assumptions also matter. Under informative labeling, IPW-adjusted PPI requires positivity or overlap, correct specification of the propensity model when probabilities are estimated, and the MAR-type condition that, conditional on $\hat Y_i=f(X_i)$1, label inclusion does not depend on the unobserved outcome (Datta et al., 13 Aug 2025). Under missing-not-at-random mechanisms, the practical review states that all methods, including classical inference using only labeled data, yield biased estimates (Song et al., 28 Jan 2026).

Subgroup structure introduces another important caveat. For treatment-effect estimation, a pooled rectifier can fail when prediction errors differ across treatment arms. If $\hat Y_i=f(X_i)$2, then the bias of a pooled-rectifier treatment-effect estimator is

$\hat Y_i=f(X_i)$3

which does not vanish simply by increasing the number of labels if differential error persists. The proposed remedy is arm-specific rectification, directly analogous to stratified survey estimation (Mozer, 19 Mar 2026).

Empirically, PPI has been demonstrated in proteomics, astronomy, genomics, remote sensing, census analysis, and ecology (Angelopoulos et al., 2023). Under informative labeling in the NHANES BMI example, the IPW paper reports empirical biases of about $\hat Y_i=f(X_i)$4 for HT, $\hat Y_i=f(X_i)$5 for Hájek, $\hat Y_i=f(X_i)$6 for unweighted PPI, and $\hat Y_i=f(X_i)$7 for weighted PPI, while the PPI variants yielded narrower intervals than pure design-based estimators (Datta et al., 13 Aug 2025). At the design stage, power analysis for PPI++ shows that when $\hat Y_i=f(X_i)$8,

$\hat Y_i=f(X_i)$9

so the labeled sample-size reduction scales roughly with the $$\hat\theta_{\mathrm{PPI}} = \frac{1}{N}\sum_{i=1}^{N}\hat Y_i + \frac{1}{n}\sum_{i=1}^{n}(Y_i-\hat Y_i).$$0 between predictions and ground truth (Chen et al., 17 Mar 2026).

Across these developments, PPI is best understood not as a single estimator but as a general bias-corrected, prediction-assisted inferential template whose validity depends on explicit sampling and training assumptions, and whose efficiency depends on prediction quality, residual structure, and the chosen variant (Song et al., 28 Jan 2026).