Causal Statistical Estimators

• Causal statistical estimators are methods that translate causal assumptions and observed data into average treatment effects, bridging potential outcomes and graphical models.
• They incorporate techniques such as g-computation, inverse probability weighting, and doubly robust estimators like AIPW and TMLE to balance bias, variance, and efficiency.
• Recent advances integrate cross-fitting, high-dimensional adjustments, and generative models to handle confounding, missing data, and partial identification for complex outcomes.

Searching arXiv for recent and foundational papers on causal statistical estimators, including doubly robust, cross-fit, graphical, high-dimensional, and weighting approaches. Causal statistical estimators are procedures that translate causal identification assumptions and observed-data structure into estimates of quantities such as the average causal effect, treatment-specific means, total effects in structural equation models, or partially identified bounds. In the potential-outcomes formulation, a standard target is

$\psi=\mathbb{E}[Y(1)-Y(0)],$

with observed data $O=(X,A,Y)$, binary treatment $A\in\{0,1\}$, and baseline covariates $X$. In the graphical formulation, the same target is obtained from interventional distributions such as $P(Y\mid do(A=a))$. Across these formulations, estimation is distinct from identification: assumptions such as consistency, conditional exchangeability, positivity, back-door admissibility, or front-door structure determine whether a parameter is identified, whereas the estimator determines bias, variance, efficiency, and inferential validity (Wijayatunga, 2014, Schomaker, 2020).

1. Identification, estimands, and causal models

Under consistency, conditional exchangeability, and positivity, the average treatment effect is identified by the g-formula,

$$\psi=\mathbb{E}\big[\mathbb{E}[Y\mid A=1,X]-\mathbb{E}[Y\mid A=0,X]\big].$$

The same result appears in the graphical setting through back-door adjustment. For a binary treatment and discrete confounders,

$$P(Y=y\mid do(A=a))=\sum_x P(Y=y\mid A=a,X=x)P(X=x),$$

with the continuous-$X$ analogue obtained by replacing sums with integrals. The graphical factorization $p(x,a,y)=p(x)p(a\mid x)p(y\mid a,x)$ makes explicit that adjustment for an admissible back-door set $X$ identifies the intervention distribution (Wijayatunga, 2014).

This equivalence between potential outcomes and directed acyclic graphs is central to the modern theory of causal estimators. In one direction, the potential-outcome assumptions $O=(X,A,Y)$0 and $O=(X,A,Y)$1 imply the adjustment formula. In the other, a back-door set in the DAG supplies the conditioning set required for ignorability. The literature therefore treats graphical structure as a mechanism for selecting adjustment variables and potential-outcome notation as a parameterization of counterfactual targets rather than as competing frameworks (Wijayatunga, 2014).

A recurrent point of clarification concerns regression. Regression parameters are conditional associational parameters, whereas causal parameters are typically marginal contrasts of potential outcomes. A coefficient on treatment in a regression model is not generally equal to a marginal causal effect. Equality requires additional conditions, including a valid back-door set, correct functional form, and, when a single coefficient is interpreted as a marginal effect, absence of effect modification together with a collapsible effect measure. The non-collapsibility of the odds ratio is especially consequential: conditional logistic-regression coefficients are generally not marginal causal log-odds ratios, even when there is no confounding (Schomaker, 2020).

Identification can also proceed through front-door structure. With treatment $O=(X,A,Y)$2, mediator $O=(X,A,Y)$3, outcome $O=(X,A,Y)$4, and an unobserved confounder between $O=(X,A,Y)$5 and $O=(X,A,Y)$6 but not between $O=(X,A,Y)$7 and $O=(X,A,Y)$8, the front-door formula yields

$O=(X,A,Y)$9

This permits estimation of complex nonlinear causal effects from observational data when the causal structure is known, even with missing data handled through multiple imputation under a missing-at-random design implied by the study DAG (Karvanen, 2014).

2. Core estimators for average causal effects

The standard estimators for the average causal effect are organized around nuisance functions $A\in\{0,1\}$0 and $A\in\{0,1\}$1. G-computation, or outcome regression, plugs an estimate of $A\in\{0,1\}$2 into

$A\in\{0,1\}$3

Inverse probability weighting instead targets the same estimand through the propensity score,

$A\in\{0,1\}$4

These are singly robust estimators: consistency requires correct estimation of $A\in\{0,1\}$5 for g-computation and of $A\in\{0,1\}$6 for IPW. IPW is often less efficient and can be unstable under near-violations of positivity because extreme weights inflate variance (Zivich et al., 2020, Wijayatunga, 2014).

Augmented inverse probability weighting and targeted maximum likelihood estimation introduce the doubly robust class. The canonical AIPW estimator is

$A\in\{0,1\}$7

Its defining property is consistency if either the propensity score model or the outcome model is correctly specified. TMLE starts from initial estimates of $A\in\{0,1\}$8 and $A\in\{0,1\}$9, forms the clever covariate

$X$0

and updates the outcome regression through a targeted fluctuation so that the resulting plug-in estimator solves the efficient influence-function estimating equation in sample (Zivich et al., 2020, Benkeser et al., 2019).

The efficient influence function for the average treatment effect in the nonparametric model is

$X$1

or, in equivalent notation,

$X$2

AIPW is the empirical mean of an estimated efficient influence function, and TMLE is asymptotically equivalent to the doubly robust estimator under correct targeting. When both nuisance models are consistently estimated, these procedures are semiparametrically efficient (Wijayatunga, 2014, Benkeser et al., 2019).

A persistent misconception is that flexible outcome regression alone is sufficient for valid causal inference. The evidence summarized here does not support that view. Outcome-regression coefficients remain conditional quantities, and with flexible machine learning they may exhibit bias and under-coverage if they are used without orthogonalization, targeting, or cross-fitting. Likewise, entropy-balancing or other balancing procedures do not remove the need for identifying assumptions and are not generally assumption-free substitutes for outcome or propensity modeling (Schomaker, 2020, Källberg et al., 2022).

3. Cross-fitting, nuisance learning, and estimator choice

The contemporary semiparametric literature treats cross-fitting as a device for valid inference with flexible nuisance estimation. In $X$3-fold cross-fitting, nuisance functions are fitted on $X$4 folds and evaluated on the held-out fold, producing out-of-fold predictions $X$5 and $X$6. For the doubly robust estimator this yields

$X$7

Cross-fitting breaks “own observation” ties, weakens Donsker-type restrictions, and permits slower nuisance convergence rates because the product of nuisance estimation errors, rather than each rate separately, need only be $X$8 for $X$9-consistent asymptotically normal inference (Zivich et al., 2020).

Simulation evidence for the average causal effect under complex nonlinear data-generating mechanisms is unusually sharp on this point. With correctly specified parametric nuisance models, g-computation, IPW, AIPW, and TMLE were all unbiased with near-nominal coverage; g-computation was most efficient and IPW least efficient. With Super Learner nuisance estimation, singly robust estimators deteriorated, and non-cross-fitted doubly robust estimators had under-coverage of approximately $P(Y\mid do(A=a))$0 to $P(Y\mid do(A=a))$1. Double cross-fit AIPW and double cross-fit TMLE, by contrast, achieved near-nominal coverage of approximately $P(Y\mid do(A=a))$2 to $P(Y\mid do(A=a))$3, low bias, and RMSE around $P(Y\mid do(A=a))$4 at $P(Y\mid do(A=a))$5. Cross-fitting did not remove bias from misspecification, and repeated partitioning was needed for stability; at least $P(Y\mid do(A=a))$6 to $P(Y\mid do(A=a))$7 partitions improved stability in the paper’s implementation (Zivich et al., 2020).

The same concern with parameter-targeted evaluation motivates recent model-selection methods for causal estimators. One approach scores CATE and IV estimators out of sample using Normalized ERUPT and energy distance. In randomized CATE problems, corrected outcomes $P(Y\mid do(A=a))$8 should have similar joint distributions across treatment groups when the model is well calibrated; in the IV setting, corrected outcomes should similarly equalize across access groups. These scoring rules support automated hyperparameter optimization over causal learners rather than relying on generic predictive accuracy alone (Kraev et al., 2022).

A complementary selection procedure estimates the squared $P(Y\mid do(A=a))$9-deviation of each candidate estimator from a benchmark estimator known to be asymptotically unbiased for the target. Its criterion is

$$\psi=\mathbb{E}\big[\mathbb{E}[Y\mid A=1,X]-\mathbb{E}[Y\mid A=0,X]\big].$$0

with a modified positive-part version used in practice. Under regularity conditions, the selected estimator asymptotically has risk no larger than the benchmark, whereas standard cross-validation can be systematically biased toward low-variance but biased procedures because it targets prediction rather than the causal parameter itself (Rothenhäusler, 2020).

4. Weighting, overlap, and target-population design

Weighting estimators adjust for covariate imbalance by constructing a target population through a weight function $$\psi=\mathbb{E}\big[\mathbb{E}[Y\mid A=1,X]-\mathbb{E}[Y\mid A=0,X]\big].$$1. In the balancing-weight framework,

$$\psi=\mathbb{E}\big[\mathbb{E}[Y\mid A=1,X]-\mathbb{E}[Y\mid A=0,X]\big].$$2

This nests the average treatment effect, the average treatment effect on the treated, the average treatment effect on the controls, and the overlap estimand. A unified family proposed for estimand selection is

$$\psi=\mathbb{E}\big[\mathbb{E}[Y\mid A=1,X]-\mathbb{E}[Y\mid A=0,X]\big].$$3

with ATE given by $$\psi=\mathbb{E}\big[\mathbb{E}[Y\mid A=1,X]-\mathbb{E}[Y\mid A=0,X]\big].$$4, ATT by $$\psi=\mathbb{E}\big[\mathbb{E}[Y\mid A=1,X]-\mathbb{E}[Y\mid A=0,X]\big].$$5, ATC by $$\psi=\mathbb{E}\big[\mathbb{E}[Y\mid A=1,X]-\mathbb{E}[Y\mid A=0,X]\big].$$6, and ATO by $$\psi=\mathbb{E}\big[\mathbb{E}[Y\mid A=1,X]-\mathbb{E}[Y\mid A=0,X]\big].$$7 (Barnard et al., 2024).

This formulation makes explicit that overlap problems are simultaneously estimation problems and target-population problems. The same paper decomposes the difference between a weighted estimator and the original target into statistical bias and estimand mismatch. The first reflects finite-sample and modeling limitations of the estimator; the second reflects deviation from the original research population induced by the chosen weighting scheme. To operationalize this trade-off, the paper proposes two design-based energy-distance metrics, summarized by permutation p-values: one for population representativeness relative to the original target and one for residual weighted imbalance in fitted propensity-score distributions. The resulting selection rule searches over $$\psi=\mathbb{E}\big[\mathbb{E}[Y\mid A=1,X]-\mathbb{E}[Y\mid A=0,X]\big].$$8 to balance preservation of the original population against reduction in bias and variance (Barnard et al., 2024).

Entropy balancing provides a different route to the same general objective. For ATE estimation it enforces three-way balance, requiring weighted treated means, weighted control means, and full-sample means of selected balance functions $$\psi=\mathbb{E}\big[\mathbb{E}[Y\mid A=1,X]-\mathbb{E}[Y\mid A=0,X]\big].$$9 to coincide. The KL version minimizes $$P(Y=y\mid do(A=a))=\sum_x P(Y=y\mid A=a,X=x)P(X=x),$$0 subject to balance and yields exponential-tilting weights, while the quadratic Rényi version minimizes $$P(Y=y\mid do(A=a))=\sum_x P(Y=y\mid A=a,X=x)P(X=x),$$1 and yields weights linear in the balance functions. Large-sample theory shows that these estimators are not generally consistent unless implicit parametric assumptions for the propensity score or the conditional outcomes are satisfied. In particular, exact balance does not by itself imply nonparametric consistency, and the divergence choice induces different implicit links: log-link conditions for KL and inverse-linear conditions for quadratic Rényi (Källberg et al., 2022).

Weak overlap also motivates adaptive replacements for the classical propensity score. In one collaborative targeted-learning construction, the propensity score $$P(Y=y\mid do(A=a))=\sum_x P(Y=y\mid A=a,X=x)P(X=x),$$2 is replaced by

$$P(Y=y\mid do(A=a))=\sum_x P(Y=y\mid A=a,X=x)P(X=x),$$3

the conditional probability of treatment given the conditional mean outcome. This modification yields collaborative TMLE and one-step estimators that are asymptotically linear under suitable conditions and nonparametrically super-efficient at the true data-generating distribution. The variance improvement is local rather than uniform, so the gain comes with nonregularity and more delicate inference; in finite samples, influence-function standard errors can underestimate variability and bootstrap checks are recommended (Benkeser et al., 2019).

5. Structural, high-dimensional, and mixed-source estimators

Not all causal statistical estimators target marginal treatment effects under ignorability. In recursive linear structural equation models with causal sufficiency, total causal effects can be estimated directly from a maximally oriented partially directed acyclic graph. If the effect of $$P(Y=y\mid do(A=a))=\sum_x P(Y=y\mid A=a,X=x)P(X=x),$$4 on $$P(Y=y\mid do(A=a))=\sum_x P(Y=y\mid A=a,X=x)P(X=x),$$5 is identified from the MPDAG, then

$$P(Y=y\mid do(A=a))=\sum_x P(Y=y\mid A=a,X=x)P(X=x),$$6

and the corresponding recursive least-squares estimator is

$$P(Y=y\mid do(A=a))=\sum_x P(Y=y\mid A=a,X=x)P(X=x),$$7

This estimator is asymptotically most efficient among all regular covariance-based estimators for any identified total effect, including point and joint interventions, and its efficiency result does not require Gaussian errors (Guo et al., 2020).

In sparse linear cyclic causal models, interventions are needed for identifiability in general. Two estimators analyzed in this setting are the LLC estimator and a two-step penalized maximum-likelihood estimator. Under near-optimally chosen interventions, the maximum-likelihood estimator is asymptotically near minimax optimal over sparse causal graphs, whereas LLC rates depend unfavorably on conditioning constants. The minimax lower bound scales as

$$P(Y=y\mid do(A=a))=\sum_x P(Y=y\mid A=a,X=x)P(X=x),$$8

while the two-step penalized MLE attains an upper bound of order

$$P(Y=y\mid do(A=a))=\sum_x P(Y=y\mid A=a,X=x)P(X=x),$$9

up to logarithmic and stability-dependent factors (Hütter et al., 2019).

High-dimensional confounding has also produced estimator-aggregation strategies. One proposal averages $X$0 candidate ATE estimators,

$X$1

using equal weights $X$2. The conservative variance estimator

$X$3

avoids estimating pairwise covariances across heterogeneous high-dimensional procedures. Simulations with a library of ten estimators showed that the average and a trimmed average were typically close to the best candidate in MSE and always protected against the worst candidate in the sense formalized by the paper’s risk bounds (Antonelli et al., 2019).

When observational and interventional data are both available, a distinct bias-variance trade-off appears. In a confounded linear model with multivariate treatments, one can combine the unbiased but high-variance interventional estimator $X$4 with the biased but low-variance observational estimator $X$5 through

$X$6

The optimal matrix weight is

$X$7

where $X$8 is the confounding bias of the observational estimator. This matrix-weighted class is asymptotically unbiased in the infinite-sample limit, whereas the pooled estimator formed from the union of interventional and observational data is asymptotically biased unless the observational-to-interventional sample-size ratio tends to zero (Kladny et al., 2023).

6. Complex outcomes, missing data, and partial identification

Causal statistical estimation extends beyond scalar outcomes with full support. For functional outcomes, one recent framework defines pointwise effects

$X$9

integrated functionals, and norm-based dynamic effects built from Fréchet means. Estimation combines operator-valued kernel ridge regression with empirical Fréchet means and, when phase variation is important, Fisher–Rao registration through square-root slope functions. The corresponding doubly robust functional estimator has the same structure as scalar AIPW, but now at each point $p(x,a,y)=p(x)p(a\mid x)p(y\mid a,x)$0: $p(x,a,y)=p(x)p(a\mid x)p(y\mid a,x)$1 The paper gives consistency for Fréchet-mean estimators and asymptotic normality for Euclidean dynamic effects in finite-dimensional grids (Raykov et al., 6 Mar 2025).

When outcomes are truncated by death, the primary difficulty is that the longitudinal outcome may be undefined at follow-up. Several estimands are therefore distinguished. The survivor average causal effect,

$p(x,a,y)=p(x)p(a\mid x)p(y\mid a,x)$2

targets always survivors. Restricted mean survival time,

$p(x,a,y)=p(x)p(a\mid x)p(y\mid a,x)$3

targets survival directly. Composite estimands such as pairwise comparison and the survival-incorporated median require an explicit ordering of death relative to longitudinal health states. The paper argues that, because there is no natural notion of ordering and distance for outcomes truncated by death, the stratified average causal effect combined with restricted mean survival time provides a more complete characterization of treatment effects (Ortholand et al., 29 Apr 2026).

Partial identification leads to yet another class of estimators. If the joint law of potential outcomes is not identified, sharp bounds for moments of $p(x,a,y)=p(x)p(a\mid x)p(y\mid a,x)$4 can be written as conditional optimal transport problems. With observed joint laws $p(x,a,y)=p(x)p(a\mid x)p(y\mid a,x)$5 and $p(x,a,y)=p(x)p(a\mid x)p(y\mid a,x)$6 for $p(x,a,y)=p(x)p(a\mid x)p(y\mid a,x)$7 and $p(x,a,y)=p(x)p(a\mid x)p(y\mid a,x)$8, the lower bound is

$p(x,a,y)=p(x)p(a\mid x)p(y\mid a,x)$9

A wavelet-based primal estimator for this conditional optimal transport problem exploits smoothness of the observed densities. For quadratic costs it yields asymptotic normality of the estimated lower bound when $X$0, enabling Wald inference for the partial-identification set rather than only point estimation (Lin et al., 24 Feb 2026).

Generative estimators based on structural causal models supply a different generalization. Neural autoregressive density estimators learn each mechanism $X$1 in a known DAG and then answer interventional queries by truncated factorization. This produces estimates of $X$2, ATEs, CATEs, and full interventional outcome distributions through Monte Carlo sampling. In the synthetic studies reported, the method recovered causal effects from nonlinear systems under back-door and front-door identification, while also showing clearly that hidden confounding biases both neural and linear estimators when the identifying graph is misspecified (Garrido et al., 2020).

Finally, incomplete observational data can be handled within the identification formula rather than outside it. In a front-door setting with nonlinear causal effects and missing-at-random covariates, multiple imputation by chained equations combined with generalized additive models can estimate

$X$3

from incomplete data. In the example studied, complete-case analysis was biased upward for the average causal effect, whereas multiple imputation combined with front-door g-computation recovered the causal curve without significant bias over the supported exposure region (Karvanen, 2014).