Why Combine Two Models?
A conditional treatment-effect learner aims to estimate how the average causal effect varies with measured characteristics \(X\). One approach models the outcome under each treatment level; another uses treatment probabilities to adjust for which outcomes were observed. Each approach can be vulnerable when its model is wrong. A doubly robust learner combines both in a pseudo-outcome: a constructed response whose conditional expectation equals the treatment effect if either the outcome models are correct or the treatment-probability model is correct.
The robustness claim concerns a precise expectation identity, not a guarantee that every finite-sample estimate is accurate. It requires causal identification assumptions, valid probability estimates, and appropriate conditions for estimating the conditional expectation. In particular, it does not protect against both nuisance models being wrong.
Estimand, Nuisance Functions, and Assumptions
Let \(W\in\{0,1\}\) denote treatment, \(Y\) the observed outcome, and \(X\) the measured covariates. Let \(Y(1)\) and \(Y(0)\) denote the potential outcomes under treatment and control. The target is the conditional average treatment effect:
Define the propensity score \(e_0(x)=P(W=1\mid X=x)\), and the outcome regression functions \(\mu_w(x)=E[Y\mid W=w,X=x]\), for \(w\in\{0,1\}\). The subscript on \(e_0\) and the use of \(\mu_w\) distinguish the true nuisance functions from fitted or candidate functions used in a learner.
To identify \(\tau(x)\) from observed data, assume consistency, conditional exchangeability, and positivity on the covariate values of interest. Consistency means \(Y=WY(1)+(1-W)Y(0)\). Conditional exchangeability means \((Y(1),Y(0))\) is independent of \(W\) given \(X\). Pointwise positivity means \(0<e_0(x)<1\) on the target support. Also assume the relevant conditional means exist. Under these assumptions, \(\mu_w(x)=E[Y(w)\mid X=x]\), so \(\tau(x)=\mu_1(x)-\mu_0(x)\).
The first term is the outcome-model contrast. The remaining terms correct it using observed residuals, scaled by treatment probabilities. A treated unit contributes an outcome residual to the treated outcome model; a control unit contributes a residual to the control outcome model. The corrections are designed so that errors in one nuisance model can be offset when the other nuisance model is correct.
The Double-Robustness Identity
The central result can be checked by conditioning on \(X=x\). Write \(g=g(x)\), \(m_w=m_w(x)\), and use the true conditional means \(\mu_w(x)\) and propensity \(e_0(x)\). Because \(E[W(Y-m_1)\mid X=x]=e_0(x)\{\mu_1(x)-m_1(x)\}\), and \(E[(1-W)(Y-m_0)\mid X=x]=(1-e_0(x))\{\mu_0(x)-m_0(x)\}\), we obtain:
The second line follows by subtracting \(\tau(x)=\mu_1-\mu_0\) from the first line and collecting the coefficients of \(m_1-\mu_1\) and \(m_0-\mu_0\). It gives the exact conditional bias for fixed candidate functions. If \(g=e_0\), the factor \(g-e_0\) is zero, regardless of the outcome functions. If \(m_1=\mu_1\) and \(m_0=\mu_0\), both outcome-model errors are zero, regardless of \(g\). Thus either a correct propensity model or correct outcome models make the conditional expectation equal to \(\tau(x)\).
Worked Example: Correct Propensity, Misspecified Outcomes
Suppose, hypothetically, that within a covariate stratum the true treatment probability is \(e_0(x)=0.40\), and the true conditional means are \(\mu_1(x)=10\) and \(\mu_0(x)=6\) outcome units. Then \(\tau(x)=10-6=4\) outcome units. Consider candidate outcome predictions \(m_1(x)=8\) and \(m_0(x)=5\), both incorrect, but use the correct propensity \(g(x)=0.40\).
Substitute into the conditional expectation formula:
The result equals the true conditional effect despite both candidate outcome means being wrong. The corrected treated residual contributes \(2\) units, while the corrected control residual subtracts \(1\) unit. This verifies the identity for the stated hypothetical stratum; it does not imply that a fitted propensity will equal the true one in an applied analysis.
Worked Example: Correct Outcome Models, Misspecified Propensity
Now suppose, hypothetically, that \(\mu_1(x)=m_1(x)=12\) and \(\mu_0(x)=m_0(x)=8\) outcome units, while the true treatment probability is \(e_0(x)=0.25\). Use an incorrect candidate propensity \(g(x)=0.50\). The conditional effect is \(12-8=4\) outcome units.
The outcome-model errors are both zero, so the conditional expectation is:
Here the propensity model is wrong, but the residual corrections have expectation zero because each outcome model gives the true conditional mean. The identity requires both outcome means to be correct; getting only one of them right is not enough for this argument.
From Pseudo-Outcomes to a Conditional Effect Fit
In practice, nuisance functions are estimated from data. Write the fitted functions as \(\hat m_1,\hat m_0,\hat g\), substitute them into \(\phi\), and treat the resulting values as responses in a regression or other conditional-mean learner using \(X\) as predictors. For a simple subgroup fit, this can mean averaging pseudo-outcomes within each subgroup. More flexible regressions can represent smoothly varying effects, but their predictions still depend on adequate data and model behavior across the covariate range.
If the same observations are used both to fit highly flexible nuisance models and to construct pseudo-outcomes, overfitting can compromise the useful behavior of the estimating procedure. A common practical safeguard is sample splitting or cross-fitting: estimate nuisance functions on one part of the data and calculate pseudo-outcomes on observations not used to fit those functions, rotating the held-out part across folds. This can reduce overfitting-related dependence; it does not make misspecified models correct or replace causal identification assumptions.
Worked Example: A Two-Group Pseudo-Outcome Fit
Consider the following hypothetical observations. Suppose nuisance predictions were estimated on separate training data and are constant within each covariate group. For both groups, \(\hat g(x)=0.50\). In group A, \(\hat m_1=10\) and \(\hat m_0=8\); in group B, \(\hat m_1=14\) and \(\hat m_0=9\). Outcomes are measured in outcome units.
| Group | W | Observed Y | \(\hat m_1\) | \(\hat m_0\) |
|---|---|---|---|---|
| A | 1 | 12 | 10 | 8 |
| A | 0 | 7 | 10 | 8 |
| B | 1 | 15 | 14 | 9 |
| B | 0 | 8 | 14 | 9 |
For the first observation, \(W=1\), so only the treated residual correction is present:
For the second observation, \(W=0\), so the control correction is subtracted:
The estimated group-A conditional effect from a two-level regression—that is, the mean pseudo-outcome in group A—is \((6+4)/2=5\) outcome units. For group B, the treated observation gives \(14-9+(15-14)/0.50=5+2=7\). The control observation gives \(14-9-(8-9)/0.50=5-(-2)=7\). Thus the group-B mean is \((7+7)/2=7\) outcome units. This simple fit estimates a difference of \(7-5=2\) outcome units between the group-specific effects. The calculations illustrate fitting, not a claim that these four observations establish the true effects; the pseudo-outcomes vary with observed residuals and the nuisance predictions.
Positivity, Overlap, and Weight Stability
Pointwise positivity, \(0<e_0(x)<1\), is the identification condition: at each relevant covariate value, both treatment states must be possible. It does not by itself bound inverse-propensity weights across the covariate distribution. For example, propensities \(e_0(x_n)=1/n\) are positive and below one for every \(n>1\), yet the treated weight \(1/e_0(x_n)=n\) grows without bound.
A stronger condition, often called uniform overlap, requires some fixed \(\epsilon>0\) such that \(\epsilon\leq e_0(x)\leq1-\epsilon\) throughout the target covariate support. This bounds both inverse weights: \(1/e_0(x)\leq1/\epsilon\) and \(1/\{1-e_0(x)\}\leq1/\epsilon\). Such a condition can control extreme weights, but it is stronger than pointwise positivity and may not hold in a given population. Even with pointwise positivity, very small or very large estimated propensities can make pseudo-outcomes highly variable. Restricting the target population or modifying weights changes the analysis and may change the estimand; it should not be presented as a consequence of double robustness.
What Double Robustness Does—and Does Not—Mean
The identity is conditional and exact for fixed nuisance functions. With estimated functions, it motivates consistency when one nuisance component is correctly estimated under suitable regularity conditions, including adequate overlap and control of estimation error. It does not say that every pseudo-outcome equals an individual treatment effect, nor that an average of a small sample of pseudo-outcomes is exact. The pseudo-outcome is a constructed response whose conditional mean is the target under the stated conditions.
- Both nuisance components can fail. The bias expression contains the product of propensity error and outcome-model error. If both are wrong, the product need not vanish.
- Extreme weights can inflate variability. Pointwise positivity alone does not guarantee bounded weights; inspect overlap in the target support.
- Effect prediction needs its own care. A pseudo-outcome can be noisy, and a poor second-stage conditional-mean fit can still produce poor estimates of how effects vary with \(X\).
- Robustness is not causal identification. The pseudo-outcome cannot repair unmeasured confounding, inconsistency, or a mismatch between the target population and the data being analyzed.
Common mistake. “The estimator is doubly robust, so positivity is automatically adequate and the estimate remains reliable whenever either model is approximately plausible.” The correction is that double robustness is an expectation identity under identification assumptions, not a bound on inverse weights or a blanket finite-sample guarantee. Pointwise positivity supports identification; uniform overlap is a stronger condition that bounds weights. If both nuisance models are wrong, or if weights are unstable and the conditional-effect fit is inadequate, the resulting estimates can still be unreliable.
Check Your Understanding
Answer each question using the definitions and conditions above.
- State the three causal identification assumptions used to relate the observed conditional outcome means to the conditional treatment effect.
- Why does a correct propensity function make the conditional expectation of the pseudo-outcome equal \(\tau(x)\), even when the outcome functions are misspecified?
- If both outcome regressions are correct but the propensity function is wrong, what happens to the conditional expectation, and why?
- Explain why \(0<e_0(x)<1\) does not guarantee bounded inverse-propensity weights across the covariate support.
- What does a second-stage regression of estimated pseudo-outcomes on \(X\) estimate under the double-robustness conditions?