Common Errors in Semiparametric Causal Estimation

Why Semiparametric Estimates Can Still Go Wrong

An estimator can have an attractive theoretical property and still be used incorrectly. In particular, double robustness is not a general guarantee of validity: it does not fix a lack of treatment overlap, remove every bias from flexible nuisance models, or make an inappropriate standard error trustworthy. Reliable inference requires checking the conditions that support identification, estimation, and uncertainty quantification separately.

This tutorial focuses on the average treatment effect (ATE) estimated with the augmented inverse probability weighted (AIPW) score. The same diagnostic habits apply to related semiparametric estimators. Let \(A\) be a binary treatment, \(Y\) an outcome, and \(X\) measured covariates. Write \(e_0(X)=P(A=1\mid X)\) for the true propensity score and \(m_{a,0}(X)=E(Y\mid A=a,X)\) for the true outcome regression under treatment level \(a\). Under consistency, conditional exchangeability given \(X\), and positivity, the ATE is identified by \(\psi_0=E[m_{1,0}(X)-m_{0,0}(X)]\).

The usual double-robustness claim is conditional: under the identification and regularity conditions, the AIPW estimator can be consistent if either the propensity model or the outcome regression is correctly specified (and the relevant fitted nuisance estimates behave appropriately). That claim is not the same as saying every finite-sample estimate is stable, or that a confidence interval is valid regardless of how the nuisance estimates were obtained.

Error 1: Treating Double Robustness as Protection Against Poor Overlap

Positivity requires that both treatment levels be possible at covariate values in the target population: \(0<e_0(X)<1\) almost surely. A stronger uniform-positivity condition requires some \(\epsilon>0\) such that \(\epsilon\le e_0(X)\le 1-\epsilon\) almost surely. When the true propensity is close to zero or one, the inverse-propensity factors in the AIPW score can be large. This can make estimates highly variable even when the propensity model is correct.

If a treatment level has probability zero in a covariate stratum, the observed data provide no treated-versus-untreated comparison there. Double robustness cannot supply the missing information or identify the ATE for that stratum without additional assumptions. If probabilities are positive but very small, identification may hold while estimation remains unstable.

Clipping estimated propensities to a chosen range may limit extreme weights, but it changes the nuisance estimates used in the score and does not create observations from unsupported treatment groups. Its effects should be treated as a sensitivity or modeling decision, not as proof that positivity holds.

Error 2: Checking Each Nuisance Rate but Ignoring Their Product

For root-\(n\) inference, it is not enough to say that each nuisance estimate is “reasonably accurate.” A standard sufficient condition, alongside appropriate consistency, moment, and positivity conditions, is that the product of the propensity and outcome-regression errors is small enough. In \(L_2\) notation, a common version is

The notation \(o_p(n^{-1/2})\) means that the product is negligible relative to \(n^{-1/2}\) in probability as the sample size grows. If both errors are of order \(o_p(n^{-1/4})\), their product is \(o_p(n^{-1/2})\). The condition also allows one nuisance to converge faster than the other. The relevant error norms and inverse-propensity factors must be controlled; a rate statement alone does not replace the other regularity conditions.

A Short Justification: Why the Product Appears

The product-rate condition is connected to the bias remaining in the expected AIPW score when nuisance functions are estimated. Conditional on \(X\), take the expectation of the score using the true data distribution but fixed fitted functions \(\hat e,\hat m_1,\hat m_0\). Subtract the target conditional contrast \(m_{1,0}(X)-m_{0,0}(X)\). Directly collecting the treated and untreated terms gives

This expression is zero if either the propensity equals its truth or both outcome regressions equal their truths. That cancellation explains the double-robust structure. When neither is exact, the remaining bias depends on products of nuisance errors, with denominators that can amplify errors near the boundaries. Under suitable bounds, taking expectations and applying standard product bounds leads to the product-rate requirement. This is a bias calculation, not a proof that all other sources of error or all inferential conditions are automatically controlled.

Error 3: Evaluating Training Predictions on Their Own Observations

Flexible nuisance learners can adapt closely to the observations used to fit them. If the AIPW score is then evaluated on those same observations, the fitted residuals and predictions may be unusually favorable in-sample. The resulting empirical score can behave differently from the score that would be obtained on new observations, complicating control of overfitting bias and the usual asymptotic argument.

Cross-fitting addresses this concern by splitting observations into folds, fitting nuisance models without a fold, and using those fitted models to predict for observations in the held-out fold. Each observation's score is computed from nuisance predictions made without using that observation to train the nuisance fits. Cross-fitting is not a remedy for positivity violations, inadequate nuisance rates, or incorrect data splitting for dependent observations.

Error 4: Reporting a Standard Error for the Wrong Quantity

For an asymptotically linear AIPW estimator under its required conditions, a standard influence-function standard error is the empirical standard deviation of estimated influence-function contributions divided by \(\sqrt n\). Equivalently, it is the square root of their sample variance divided by \(n\). The contributions should correspond to the actual estimator and, when cross-fitting is used, to the cross-fitted predictions. A standard error calculated from the raw outcome variation alone generally does not represent the sampling variability of an AIPW estimate, which also depends on treatment assignment, weighting, outcome predictions, and covariates.

A common invalid shortcut is to use the standard deviation of \(Y\) divided by \(\sqrt n\) as though the estimate were a simple sample mean of outcomes. Another is to calculate influence-function values from predictions that do not correspond to the predictions used in the reported estimate. Both can produce numbers, but neither automatically estimates the sampling variation of the stated causal estimator. If observations are clustered or otherwise dependent, an independent-observation variance calculation may also be inappropriate; the variance estimator must reflect the sampling and dependence structure.

A Practical Diagnostic Sequence

Before interpreting a semiparametric estimate or confidence interval, separate the questions rather than treating “doubly robust” as a single pass-fail label.

Common Mistake

Wrong reasoning: “The AIPW estimate is doubly robust, so it remains reliable even with extreme propensities, slow nuisance fits, in-sample predictions, and a standard error based only on outcome variation.” Each part of this claim confuses a specific theoretical property with a different requirement. Double robustness does not repair nonidentification or weak overlap; it does not eliminate product-rate and regularity conditions for root-\(n\) inference; it does not make training predictions equivalent to out-of-fold predictions; and it does not justify a variance formula unrelated to the estimator's influence function.

Correction: Treat identification, nuisance estimation, prediction construction, and variance estimation as separate checks. A defensible conclusion is only as strong as the conditions supporting each step. Sensitivity to extreme propensities, nuisance fitting choices, or variance procedures should be reported as uncertainty about the analysis, not hidden by the phrase “doubly robust.”