Doubly Robust Estimation: The Guarantee, and Where It Breaks Down

Doubly robust estimation is a way of estimating a causal effect from observational data that stays consistent if either of two auxiliary models is correctly specified — the treatment-assignment (propensity) model or the outcome model — without requiring both to be right. That single property is the whole point of the method and the source of its name: an analyst who is not certain which of two model forms is correct gets a second, independent chance at an unbiased estimate, instead of betting the entire analysis on one specification.

This is a precise statistical guarantee, not a vague claim of extra robustness. It does not mean the estimate is safe from unmeasured confounding, and it does not mean the estimate is protected if both models happen to be wrong. The rest of this guide states the guarantee exactly, shows a reproducible worked comparison where a doubly robust estimator succeeds precisely where a single-model approach fails, and shows — with the same simulation — where the guarantee stops helping.

The two models every adjustment method leans on

Estimating a causal effect from observational data almost always requires modeling one of two things about the confounders, X:

• The treatment model (propensity model) — the probability of receiving treatment given X, written e(X). Inverse probability weighting and propensity score matching both depend entirely on this model being correctly specified.
• The outcome model — the expected outcome given treatment and X, written μ(A, X). Standard covariate-adjusted regression, and g-computation in the time-varying case, depend entirely on this model being correctly specified.

Both are working models fit to data, not the true data-generating process, so both can be wrong — a linear term standing in for a curve, an omitted interaction, a link function that doesn’t match the real relationship. A method that only ever uses one of the two models has exactly one chance to get the functional form right. Doubly robust estimation combines both models in a single estimator so that either one being right is enough.

What the guarantee actually says

The most common doubly robust estimator is augmented inverse probability weighting (AIPW). For a sample of n units with covariates Xi, treatment Ai, and outcome Yi, with fitted outcome-model predictions μ̂1(Xi) / μ̂0(Xi) and fitted propensity ê(Xi), the AIPW estimate of the average treatment effect is:

τ̂AIPW = (1/n) Σi [ (μ̂1(Xi) − μ̂0(Xi)) + Ai(Yi − μ̂1(Xi)) / ê(Xi) − (1−Ai)(Yi − μ̂0(Xi)) / (1−ê(Xi)) ]

Read it as an outcome-model prediction with a correction term added on. If the outcome model is exactly right, the correction term has expectation zero regardless of what the propensity model looks like, and the estimator collapses to a consistent outcome-model estimate. If the propensity model is exactly right instead, the correction term exactly cancels the outcome model’s bias, and the estimator collapses to a consistent IPW-type estimate. Either path alone is enough — that algebraic cancellation, not a rhetorical claim of general robustness, is what “doubly robust” refers to. The term itself and the augmented-IPW construction were formalized in the causal-inference and missing-data literature in the mid-1990s to mid-2000s, most associated with Bang and Robins’ 2005 Biometrics paper on doubly robust estimation in missing-data and causal-inference models (recalled as well-established methodological history, not independently re-checked against the original volume/page this session).

What it does not say: it does not protect against an unmeasured confounder omitted from X entirely — both models are still functions of the same measured covariates, so an omitted confounder biases both simultaneously, exactly as it would bias matching or regression alone. And it does not say the estimator is unbiased if both models are wrong — the worked comparison below shows exactly how badly that case can fail.

A worked comparison: where IPW alone and regression alone fail, and AIPW doesn’t

The comparison below is a seeded simulation (mulberry32 PRNG, seed 20260829), run directly for this guide, not an illustrative or hypothetical numbers — every figure is reproducible from the generating process described here. A single confounder X ~ N(0,1) drives both treatment assignment and the outcome, and both true relationships are quadratic in X:

• True propensity: logit P(A=1|X) = −0.2 + 0.55X − 0.45X² (mean propensity 0.37 across 200,000 simulated units, no extreme values driving the result).
• True outcome: Y = 1.0 + 1.2X + 0.8X² + 2A + ε, ε ~ N(0,1) — a constant additive treatment effect of exactly 2, so the true average treatment effect is known by construction.

Two versions of each nuisance model were fit to the same simulated sample: a correct version that includes the X² term, and a misspecified version that only includes X linearly — the single most common real-world specification error, an analyst assuming a relationship is linear when it isn’t. Results at N = 200,000 (true ATE = 2.000):

Estimator	Propensity model	Outcome model	Estimate	Bias
IPW alone	Misspecified (linear)	—	1.562	−0.438 (understates by 22%)
Regression alone	—	Misspecified (linear)	1.565	−0.436 (understates by 22%)
AIPW	Misspecified (linear)	Correct (quadratic)	2.000	−0.0001
AIPW	Correct (quadratic)	Misspecified (linear)	2.015	+0.015
AIPW	Misspecified (linear)	Misspecified (linear)	1.497	−0.503 (understates by 25%)

Both singly-specified methods are biased by essentially the same amount — unsurprising, since both are failing for the same reason: the omitted X² term is a real confounder of the constant-effect relationship, and leaving it out of either the propensity or the outcome model leaves that confounding unaddressed. AIPW recovers the true effect to within simulation noise in both directions of single misspecification — wrong propensity paired with a correct outcome model, and correct propensity paired with a wrong outcome model — which is the guarantee stated precisely rather than asserted. The last row is the guarantee’s limit, not an exception to it: with both models wrong, AIPW is biased by more than either singly-specified method alone, because the correction term is now built from two wrong models compounding rather than one wrong model being offset by a correct one.

Where the guarantee stops helping: finite samples

The result above uses 200,000 simulated units specifically to show the guarantee at population scale, with sampling noise minimized. Real studies rarely have that much data. Running the identical setup 300 times at a more realistic N = 1,000 per sample shows what the guarantee does and doesn’t protect against once finite-sample variance is back in the picture:

Estimator (propensity / outcome)	Mean estimate	Bias	SD across runs	RMSE
IPW alone (misspecified)	1.562	−0.438	0.095	0.448
IPW alone (correct)	1.985	−0.016	0.101	0.102
Regression alone (misspecified)	1.565	−0.435	0.099	0.446
Regression alone (correct)	2.003	+0.003	0.068	0.068
AIPW (wrong PS / correct OM)	2.003	+0.003	0.069	0.069
AIPW (correct PS / wrong OM)	1.971	−0.029	0.157	0.160
AIPW (both wrong)	1.498	−0.502	0.099	0.512

Two things worth reading carefully in this table. First, the bias pattern holds up at realistic sample sizes, not just at population scale — both AIPW rows built from one correct model stay within a few thousandths of the truth on average across 300 independent samples, while both singly-specified misspecified estimators stay biased by roughly the same 0.44 they showed at 200,000 units. Second, and easy to miss: consistency is not the same guarantee as efficiency. AIPW built on a correct propensity model and a wrong outcome model has noticeably higher variance (SD 0.157) than a plain IPW estimator using that same correct propensity model alone (SD 0.101) — the wrong outcome model doesn’t introduce bias, but it does add noise the correction term has to carry. A doubly robust estimator being consistent under one correct model is not a reason to stop trying to get both models right; getting both right still buys real efficiency the guarantee alone doesn’t.

AIPW, TMLE, and where doubly robust methods fit in practice

AIPW is the estimator worked through above; targeted maximum likelihood estimation (TMLE) is the other commonly used doubly robust estimator, built on the same either-model-correct guarantee but through a different mechanism — it updates the outcome model itself using a targeting step informed by the propensity model, rather than adding a separate correction term after the fact. Both are legitimate members of the same doubly robust family; the choice between them is usually about estimating-equation properties and machine-learning integration (TMLE composes more naturally with flexible/ML-based nuisance models under cross-fitting) rather than about the core guarantee, which is shared.

In practice, doubly robust estimation is worth reaching for whenever an analyst genuinely isn’t confident in the functional form of either the treatment assignment mechanism or the outcome relationship — which describes most observational research questions with continuous or nonlinearly-related confounders. It is not a substitute for the design-stage work that any causal estimate depends on: identifying every plausible confounder, checking for colliders that adjustment can make worse rather than better, and being explicit about the assumption that there is no important confounder left unmeasured. Doubly robust estimation buys a second chance at getting the functional form right; it does not buy a second chance at having measured the right variables in the first place.

Frequently asked questions

Is doubly robust estimation the same thing as AIPW?

AIPW is the most common doubly robust estimator, but “doubly robust” describes a property — consistency if either nuisance model is correct — that more than one estimator can have. TMLE has the same property through a different construction, and both are correctly described as doubly robust.

Does doubly robust estimation fix unmeasured confounding?

No. Both the propensity model and the outcome model are functions of the same measured covariates. An important confounder that was never measured biases both models the same way it would bias propensity score matching or ordinary regression adjustment — the “either model can be wrong” guarantee only covers getting the functional form of the measured-covariate relationship wrong, not omitting a variable entirely.

What happens if both the propensity model and the outcome model are wrong?

The guarantee no longer applies, and the estimate can be biased — in the worked comparison above, more biased than either singly-specified method alone, because the correction term is then built from two incorrect models rather than one incorrect model being offset by a correct one.

Does a doubly robust estimator always have lower variance than a singly-specified one?

No. The repeated-sampling comparison above shows AIPW built on a correct propensity model and a wrong outcome model carrying visibly higher variance than a plain IPW estimator using that same correct propensity model alone. The guarantee is about consistency (correct on average, given enough data), not about matching or beating a correctly-specified single model on efficiency.

How is this different from g-computation or a marginal structural model?

G-computation uses only an outcome model; an IPTW-weighted marginal structural model uses only a treatment model. Each is consistent if its one model is correctly specified, and each has no fallback if it isn’t. Doubly robust estimation is the version of this same causal-inference toolkit that combines both models specifically so that either one being correct is enough.