Prediction-Powered Causal Inference (PPCI)

• PPCI is a family of semi-supervised, prediction-assisted methods for estimating causal and structural parameters by merging small labeled samples with larger unlabeled datasets.
• It enhances efficiency by reducing the variance associated with regressor averaging while maintaining causal validity via bias correction and orthogonalization.
• The framework leverages techniques like debiased machine learning and targeted learning (DML/TMLE) to attain semiparametric efficiency in causal effect estimation.

Prediction-Powered Causal Inference (PPCI) denotes a family of semi-supervised and prediction-assisted methods for estimating causal or structural parameters when a relatively small labeled sample with outcomes is supplemented by a larger sample in which outcomes are missing or unavailable but regressors or covariates are observed. In the most explicit formulation, PPCI studies semiparametric efficient estimation of causal and structural parameters in a semi-supervised setting and asks when auxiliary unlabeled regressors can reduce asymptotic variance relative to estimators based only on labeled data (Kato, 11 Jun 2026). Closely related work places PPCI within the broader prediction-powered inference (PPI) program, where predictions are treated as auxiliary, variance-reducing ingredients rather than as substitutes for gold-standard outcomes, and where validity is recovered by explicit bias correction on labeled data (Song et al., 28 Jan 2026). A complementary causal perspective argues that causal inference is “a structured instance of prediction under distribution shift,” which gives PPCI a conceptual interpretation as prediction with selective labels plus correction for selection and target-domain mismatch (Fernández-Loría, 6 Apr 2025).

1. Conceptual foundations

PPCI is rooted in two ideas that are distinct but compatible. The first is the PPI principle: a large prediction-only or unlabeled sample can improve efficiency if prediction error is corrected using a smaller labeled subset, rather than treating predictions as truth (Song et al., 28 Jan 2026). The second is the causal view that the target of inference is defined by a shift between observed and unobserved outcome distributions, with identification resting on assumptions that justify transfer from observed source-domain data to the target causal distribution (Fernández-Loría, 6 Apr 2025).

In the semiparametric formulation of PPCI, the data consist of a labeled sample with outcomes and regressors and an unlabeled sample with only regressors. The unlabeled sample is informative because many causal targets are regression functionals averaged over a regressor distribution. The central efficiency insight is that outcomes affect the residual or noise component, while unlabeled regressors sharpen the averaging over $X$ component; consequently, unlabeled data can reduce the asymptotic variance attached to estimating the covariate-law part of the estimand, but they do not reduce outcome noise itself (Kato, 11 Jun 2026).

A broader interpretation of PPCI is also supported by work on trial generalization. There, an observational study is used only to train a predictor, while the randomized trial remains the source of causal identification. The observational study need not satisfy causal assumptions; it contributes predictive structure that may make the nuisance-learning problem easier, after which trial data correct any predictive bias (Demirel et al., 2024). This suggests that PPCI is not a single estimator but a design pattern: combine predictive structure with gold-standard causal information while preserving causal validity through correction, orthogonalization, or augmentation.

2. Statistical setup and target parameters

The most explicit PPCI framework considers two data-generating schemes (Kato, 11 Jun 2026). In the two-sample scenario one observes

$$\{W_i=(X_i,Y_i)\}_{i=1}^n \sim P_0,\qquad \{\widetilde X_j\}_{j=1}^m \sim Q_{0X},$$

with $N=n+m$ and labeled fraction $\rho=n/N\in(0,1)$. In the one-sample scenario one observes $(X,S,Y)$ with

$Y = S Y^* + (1-S)\mathrm{NA},$

under missing-at-random

$Y^* \perp S \mid X,$

and overlap

$\pi_0(X)=P(S=1\mid X)\ge c_\pi>0.$

The main estimand is a regression-functional target

$$\theta_0 = \mathbb E_{V_{0X}}\!\left[m(X,\gamma_0)\right] = \int m(x,\gamma_0)\,V_{0X}(dx),$$

where $\gamma_0(x)=\mathbb E_0[Y\mid X=x]$, $$\{W_i=(X_i,Y_i)\}_{i=1}^n \sim P_0,\qquad \{\widetilde X_j\}_{j=1}^m \sim Q_{0X},$$0 is a known linear functional map in $$\{W_i=(X_i,Y_i)\}_{i=1}^n \sim P_0,\qquad \{\widetilde X_j\}_{j=1}^m \sim Q_{0X},$$1, and the evaluation distribution is

$$\{W_i=(X_i,Y_i)\}_{i=1}^n \sim P_0,\qquad \{\widetilde X_j\}_{j=1}^m \sim Q_{0X},$$2

The presence of $$\{W_i=(X_i,Y_i)\}_{i=1}^n \sim P_0,\qquad \{\widetilde X_j\}_{j=1}^m \sim Q_{0X},$$3 is central: the target need not average over the labeled covariate law alone, and the unlabeled sample can therefore change the efficiency bound by improving estimation of the evaluation distribution (Kato, 11 Jun 2026).

The framework encompasses several canonical causal and structural targets (Kato, 11 Jun 2026). For the average treatment effect,

$$\{W_i=(X_i,Y_i)\}_{i=1}^n \sim P_0,\qquad \{\widetilde X_j\}_{j=1}^m \sim Q_{0X},$$4

For the average marginal effect,

$$\{W_i=(X_i,Y_i)\}_{i=1}^n \sim P_0,\qquad \{\widetilde X_j\}_{j=1}^m \sim Q_{0X},$$5

For the average policy effect,

$$\{W_i=(X_i,Y_i)\}_{i=1}^n \sim P_0,\qquad \{\widetilde X_j\}_{j=1}^m \sim Q_{0X},$$6

For covariate-shift mean or policy evaluation,

$$\{W_i=(X_i,Y_i)\}_{i=1}^n \sim P_0,\qquad \{\widetilde X_j\}_{j=1}^m \sim Q_{0X},$$7

A different but compatible causal setup appears in prediction-powered generalization from randomized trials to a target population (Demirel et al., 2024). There the causal target is

$$\{W_i=(X_i,Y_i)\}_{i=1}^n \sim P_0,\qquad \{\widetilde X_j\}_{j=1}^m \sim Q_{0X},$$8

identified under consistency, trial ignorability, treatment positivity, mean ignorability of trial participation, and positivity of trial participation via

$$\{W_i=(X_i,Y_i)\}_{i=1}^n \sim P_0,\qquad \{\widetilde X_j\}_{j=1}^m \sim Q_{0X},$$9

This formulation emphasizes target-population transport rather than semi-supervised missing outcomes, but it shares the same PPCI logic: prediction is used to simplify nuisance estimation, not to replace causal identification (Demirel et al., 2024).

3. Efficient influence functions, efficiency bounds, and orthogonality

The semiparametric core of PPCI is the derivation of the efficient influence function (EIF) and the associated efficiency bound (Kato, 11 Jun 2026). Because the target is linear in $N=n+m$0, the framework introduces the Riesz representer $N=n+m$1, defined by

$N=n+m$2

This object converts the target functional into an inner product under the labeled covariate law and is the key nuisance in PPCI.

The paper gives explicit representers for major causal functionals (Kato, 11 Jun 2026). For ATE,

$N=n+m$3

where $N=n+m$4. For AME,

$N=n+m$5

and if $N=n+m$6, this reduces to $N=n+m$7. For APE,

$N=n+m$8

For the covariate-shift mean,

$N=n+m$9

In the two-sample case, the labeled and unlabeled EIF components are

$\rho=n/N\in(0,1)$0

$\rho=n/N\in(0,1)$1

with

$\rho=n/N\in(0,1)$2

The efficiency bound is

$\rho=n/N\in(0,1)$3

Expanding,

$\rho=n/N\in(0,1)$4

The interpretation is explicit: the first term is an outcome-noise term, while the latter terms are regressor-averaging terms (Kato, 11 Jun 2026).

When $\rho=n/N\in(0,1)$5, the bound reduces to

$\rho=n/N\in(0,1)$6

with

$\rho=n/N\in(0,1)$7

This is minimized at

$\rho=n/N\in(0,1)$8

yielding

$\rho=n/N\in(0,1)$9

Relative to labeled-only estimation under the same $(X,S,Y)$0 normalization,

$(X,S,Y)$1

so the efficiency gain is

$(X,S,Y)$2

This gives the precise sense in which unlabeled regressors can strictly improve causal inference: only the covariate-averaging component is reduced (Kato, 11 Jun 2026).

The one-sample EIF under missing outcomes is

$(X,S,Y)$3

with efficiency bound

$(X,S,Y)$4

The factor $(X,S,Y)$5 reflects missingness of outcomes (Kato, 11 Jun 2026).

A notable feature is that the EIF is also a Neyman orthogonal score. The paper states the identity

$(X,S,Y)$6

This implies that the leading bias is second order in nuisance-estimation errors, with product-rate condition

$(X,S,Y)$7

which is precisely the DML condition used to obtain asymptotic linearity and efficiency (Kato, 11 Jun 2026).

4. Estimation strategies: DML-PPCI and related prediction-powered constructions

The paper “Prediction-Powered Causal Inference by Automatic Debiased Machine Learning and Semi-Supervised Riesz Regression” introduces two estimators collectively called DML-PPCI (Kato, 11 Jun 2026). If the estimator is based on an estimating equation, it is EE-DML-PPCI; if it is based on targeted learning, it is TMLE-DML-PPCI.

For EE-DML-PPCI, the estimator is

$(X,S,Y)$8

With cross-fitting, the fold-specific version is

$(X,S,Y)$9

The paper describes this as the semi-supervised analogue of AIPW or ARW (Kato, 11 Jun 2026).

For TMLE-DML-PPCI, the regression nuisance is updated along the EIF direction: $Y = S Y^* + (1-S)\mathrm{NA},$0 with fluctuation parameter

$Y = S Y^* + (1-S)\mathrm{NA},$1

The resulting estimator is

$Y = S Y^* + (1-S)\mathrm{NA},$2

This is described as the semi-supervised analogue of Auto-TMLE (Kato, 11 Jun 2026).

Under nuisance consistency, the product-rate condition, and either Donsker-type conditions or cross-fitting, both estimators satisfy

$Y = S Y^* + (1-S)\mathrm{NA},$3

and hence

$Y = S Y^* + (1-S)\mathrm{NA},$4

The paper states that both EE-DML-PPCI and TMLE-DML-PPCI are regular and semiparametrically efficient (Kato, 11 Jun 2026).

A different PPCI-style estimator arises in trial generalization with an auxiliary observational study (Demirel et al., 2024). There the predictor $Y = S Y^* + (1-S)\mathrm{NA},$5 is trained on the observational study and then incorporated in one of two ways. The additive bias correction (ABC) estimator uses the identity

$Y = S Y^* + (1-S)\mathrm{NA},$6

defines

$Y = S Y^* + (1-S)\mathrm{NA},$7

and estimates the bias function

$Y = S Y^* + (1-S)\mathrm{NA},$8

from the trial, yielding

$Y = S Y^* + (1-S)\mathrm{NA},$9

The augmented outcome modeling (AOM) estimator augments the covariates with the predictor,

$Y^* \perp S \mid X,$0

defines

$Y^* \perp S \mid X,$1

and estimates

$Y^* \perp S \mid X,$2

These constructions show that PPCI can be implemented either as orthogonal semiparametric estimation or as bias-corrected generalization with predictive augmentation (Demirel et al., 2024).

5. Relation to general PPI, robustness mechanisms, and failure modes

PPCI inherits several methodological lessons from the broader PPI literature. The paper “Demystifying Prediction Powered Inference” stresses that PPI is a bias-corrected inferential framework, not “use machine learning predictions as if they were outcomes” (Song et al., 28 Jan 2026). In mean estimation,

$Y^* \perp S \mid X,$3

and in generic M-estimation,

$Y^* \perp S \mid X,$4

This formulation is directly relevant to PPCI because it frames prediction as a variance-reduction device plus a labeled-data correction term (Song et al., 28 Jan 2026).

The same paper isolates three assumptions in its baseline PPI setup: distribution comparability or MCAR,

$Y^* \perp S \mid X,$5

independence between the pre-trained model and the internal inference data, and complete covariate information (Song et al., 28 Jan 2026). A plausible implication for PPCI is that these conditions are replaced or augmented by causal identification assumptions, but the underlying warning remains: validity depends not only on predictive accuracy but also on how labels are missing or selectively observed.

The failure mode most emphasized in the PPI literature is double-dipping. If the prediction model is trained using data that also enter the inference stage, especially the same labeled observations used for bias correction, then predictions become too optimistic, residual correction is biased, variability is understated, and confidence intervals become anti-conservative (Song et al., 28 Jan 2026). The paper reports that in the Mosaiks housing example, reusing training data can reduce nominal $Y^* \perp S \mid X,$6 coverage dramatically, down to about $Y^* \perp S \mid X,$7 in some small-labeled-sample settings (Song et al., 28 Jan 2026). For PPCI, this strongly suggests that sample splitting or cross-fitting is structural rather than optional when nuisance learners are trained internally.

Another lesson concerns efficiency. Basic PPI is valid but does not uniformly improve precision. For mean estimation under MCAR, the variance difference relative to complete-case analysis is

$Y^* \perp S \mid X,$8

so PPI improves on complete-case analysis roughly when

$Y^* \perp S \mid X,$9

If predictions are weak, PPI can be worse than complete-case analysis (Song et al., 28 Jan 2026). Related linear-regression work revisits this issue and proposes a weighted augmentation, the Chen–Chen estimator,

$\pi_0(X)=P(S=1\mid X)\ge c_\pi>0.$0

showing that it is at least as efficient asymptotically as labeled-only inference, unlike the unweighted PPI augmentation (Gronsbell et al., 2024). This suggests that PPCI designs may benefit from control-variate or covariance-weighted augmentation, especially when predictive signal is uneven.

A further robustness refinement appears in “FAB-PPI: Frequentist, Assisted by Bayes, Prediction-Powered Inference” (Cortinovis et al., 4 Feb 2025). Standard PPI decomposes the estimating equation as

$\pi_0(X)=P(S=1\mid X)\ge c_\pi>0.$1

with

$\pi_0(X)=P(S=1\mid X)\ge c_\pi>0.$2

FAB-PPI modifies only the rectifier estimate: $\pi_0(X)=P(S=1\mid X)\ge c_\pi>0.$3 then solves

$\pi_0(X)=P(S=1\mid X)\ge c_\pi>0.$4

With a horseshoe prior,

$\pi_0(X)=P(S=1\mid X)\ge c_\pi>0.$5

the method has an infinite spike at zero and Cauchy-like or power-law tails, which makes it shrink aggressively when predictions are good and asymptotically revert to standard PPI when $\pi_0(X)=P(S=1\mid X)\ge c_\pi>0.$6 is large (Cortinovis et al., 4 Feb 2025). The paper itself is not a causal paper, but it is directly relevant to PPCI because it provides a prior-assisted mechanism for improving prediction-powered estimators while preserving asymptotic frequentist coverage.

6. Applications, empirical behavior, and misconceptions

The clearest direct PPCI claim in the recent literature is that unlabeled covariates can attain a smaller asymptotic variance than the efficiency bound attainable from labeled observations alone, provided the target is a regression functional averaged over a regressor distribution (Kato, 11 Jun 2026). This is a semiparametric statement, not a claim that unlabeled data identify causal effects by themselves. The improvement arises only through better estimation of the covariate-law component of the estimand.

In trial generalization, prediction-powered methods are reported to facilitate better generalization when the auxiliary observational study is high-quality and remain robust when it is not, including when it has unmeasured confounding (Demirel et al., 2024). The paper’s empirical summary is specific: using the observational study alone is not reliable when confounding is high; trial-only outcome modeling struggles when the trial is small and $\pi_0(X)=P(S=1\mid X)\ge c_\pi>0.$7 is complex; ABC and AOM can substantially outperform trial-only outcome modeling when the trial is small and the predictor captures much of the structure in $\pi_0(X)=P(S=1\mid X)\ge c_\pi>0.$8; AOM is more robust than ABC because it can ignore an unhelpful predictor; and when the trial is large enough, the advantage disappears (Demirel et al., 2024).

The general PPI literature reports a parallel pattern. When predictions are sufficiently informative, PPI variants can produce tighter confidence intervals than complete-case analysis, but if assumptions are violated the gains may disappear or validity may fail (Song et al., 28 Jan 2026). Under missing-not-at-random mechanisms, the paper states that all methods, including classical inference using only labeled data, yield biased estimates (Song et al., 28 Jan 2026). This is particularly relevant to PPCI because selective outcome observation is inherent to causal problems; prediction assistance does not remove the need for ignorability, overlap, or transport assumptions.

Several recurrent misconceptions are clarified by these papers. One misconception is that PPCI means imputing missing outcomes with black-box predictions. The literature explicitly rejects this: predictions are auxiliary, and treating them as ground truth generally yields biased point estimates and anti-conservative confidence intervals (Song et al., 28 Jan 2026). A second misconception is that any additional unlabeled data must improve efficiency. The semiparametric theory shows improvement only in the regressor-averaging term, while the broader PPI theory shows that weak predictions can even worsen efficiency unless tuning or weighting is used (Kato, 11 Jun 2026, Song et al., 28 Jan 2026, Gronsbell et al., 2024). A third misconception is that auxiliary observational data must be causally valid to be useful. In trial generalization, the observational study is used purely as a source of predictive structure, and the method makes no causal assumptions on that study (Demirel et al., 2024).

More broadly, the argument that causal inference is “prediction under distribution shift” helps explain why PPCI is conceptually coherent without collapsing causal inference into ordinary supervised learning (Fernández-Loría, 6 Apr 2025). In that account, the source domain contains selectively observed labels, the target domain contains unobserved potential outcomes or counterfactual contrasts, and causal assumptions play the role of transfer assumptions that justify moving information across distributions.

7. Limitations, scope conditions, and future directions

The current PPCI literature is explicit that prediction-powered methods do not eliminate the causal layer. PPI infrastructure by itself does not identify causal effects; causal extensions require treatment assignment structure, potential outcomes, causal identification assumptions, nuisance functions for treatment propensity and outcome models, and possibly treatment-effect heterogeneity (Song et al., 28 Jan 2026). This is why the 2026 PPCI paper builds directly on semiparametric EIF and Riesz machinery rather than merely transplanting generic PPI estimators (Kato, 11 Jun 2026).

Asymptotic validity is the dominant guarantee. In FAB-PPI, frequentist coverage is preserved asymptotically under a CLT-type condition for the estimators of $\pi_0(X)=P(S=1\mid X)\ge c_\pi>0.$9 and $$\theta_0 = \mathbb E_{V_{0X}}\!\left[m(X,\gamma_0)\right] = \int m(x,\gamma_0)\,V_{0X}(dx),$$0 (Cortinovis et al., 4 Feb 2025). In DML-PPCI, asymptotic linearity and semiparametric efficiency require nuisance consistency, the product-rate condition

$$\theta_0 = \mathbb E_{V_{0X}}\!\left[m(X,\gamma_0)\right] = \int m(x,\gamma_0)\,V_{0X}(dx),$$1

and either Donsker-type conditions or cross-fitting (Kato, 11 Jun 2026). This suggests that finite-sample behavior may depend materially on nuisance quality, overlap, and the stability of the Riesz estimation problem.

Nuisance estimation is itself a substantive problem. PPCI relies on estimating the Riesz representer, and the paper therefore develops semi-supervised generalized Riesz regression via the objective

$$\theta_0 = \mathbb E_{V_{0X}}\!\left[m(X,\gamma_0)\right] = \int m(x,\gamma_0)\,V_{0X}(dx),$$2

Its sample analogue is regularized as

$$\theta_0 = \mathbb E_{V_{0X}}\!\left[m(X,\gamma_0)\right] = \int m(x,\gamma_0)\,V_{0X}(dx),$$3

The population excess risk equals a Bregman divergence,

$$\theta_0 = \mathbb E_{V_{0X}}\!\left[m(X,\gamma_0)\right] = \int m(x,\gamma_0)\,V_{0X}(dx),$$4

and the KKT conditions imply balancing equations through the semi-supervised imbalance gap

$$\theta_0 = \mathbb E_{V_{0X}}\!\left[m(X,\gamma_0)\right] = \int m(x,\gamma_0)\,V_{0X}(dx),$$5

This makes clear that computational and statistical complexity are concentrated in nuisance learning and balance control (Kato, 11 Jun 2026).

The paper also provides rate guarantees. If the dual target $$\theta_0 = \mathbb E_{V_{0X}}\!\left[m(X,\gamma_0)\right] = \int m(x,\gamma_0)\,V_{0X}(dx),$$6 is $$\theta_0 = \mathbb E_{V_{0X}}\!\left[m(X,\gamma_0)\right] = \int m(x,\gamma_0)\,V_{0X}(dx),$$7-Hölder on $$\theta_0 = \mathbb E_{V_{0X}}\!\left[m(X,\gamma_0)\right] = \int m(x,\gamma_0)\,V_{0X}(dx),$$8, then

$$\theta_0 = \mathbb E_{V_{0X}}\!\left[m(X,\gamma_0)\right] = \int m(x,\gamma_0)\,V_{0X}(dx),$$9

where $\gamma_0(x)=\mathbb E_0[Y\mid X=x]$0. Moreover, if

$\gamma_0(x)=\mathbb E_0[Y\mid X=x]$1

then the DML product-rate condition holds (Kato, 11 Jun 2026). These are implementation-relevant guarantees, but they also underscore that PPCI remains sensitive to high-dimensional nuisance complexity.

A final boundary question concerns scope. Not all causality papers with “prediction” in their framing are PPCI papers. For example, “Probably Approximately Correct Causal Discovery” studies finite-sample causal hypothesis discrimination and sample complexity for propensity scores, IV, and SCCS, but it does not use a predictor-plus-correction architecture and is therefore conceptually adjacent rather than an instance of PPCI (Wei et al., 25 Jul 2025). By contrast, work on trial generalization with observational predictors and work on DML-PPCI both fit the PPCI pattern because they explicitly combine predictive models with gold-standard causal data and then debias or orthogonalize the result (Demirel et al., 2024, Kato, 11 Jun 2026).

Taken together, these papers define PPCI as a semiparametric, prediction-assisted approach to causal inference in which unlabeled covariates or auxiliary predictive data reduce the difficulty of estimating regression functionals, while causal validity is retained through correction, orthogonality, and identification assumptions rather than through trust in predictions themselves (Kato, 11 Jun 2026, Song et al., 28 Jan 2026, Demirel et al., 2024).