Why Inverse Probability Weighting Identifies the ATE

Why Reweighting Can Recover a Marginal Effect

Treated and untreated people can differ in their covariates before treatment. As a result, their observed average outcomes need not compare like with like: part of the difference may reflect who received treatment rather than treatment itself. Inverse probability weighting (IPW) addresses this by giving more weight to observations whose treatment assignment was less likely given their covariates. The central question here is why that weighting targets the population means of the potential outcomes.

This is a population identification result. It says that under specified assumptions, a particular expectation involving the observed data equals a causal quantity defined using potential outcomes. It does not say that a finite-sample weighted average equals that quantity exactly, nor does it by itself guarantee that estimated weights will be reliable.

Definition: Let $A\in\{0,1\}$ indicate treatment, let $X$ denote pretreatment covariates, and let $Y$ be the observed outcome. For each unit, $Y(1)$ and $Y(0)$ are the potential outcomes under treatment and control, respectively. The average treatment effect (ATE) in the target population is $E[Y(1)]-E[Y(0)]$. The propensity score is $e(X)=P(A=1\mid X)$, so $1-e(X)=P(A=0\mid X)$.

Assumptions for the Identification Result

The argument requires that treatment and outcomes be well defined, that the potential outcomes have finite means, and that the following conditions hold in the target population.

Consistency: the observed outcome is the potential outcome corresponding to the treatment actually received: $Y=AY(1)+(1-A)Y(0)$. This presumes a well-defined treatment and no interference between units.
Conditional exchangeability: $(Y(1),Y(0))\perp A\mid X$. Once the measured pretreatment covariates $X$ are held fixed, treatment assignment is independent of both potential outcomes.
Positivity: $0<e(X)<1$ for covariate values occurring in the target population. Thus both treatment states are possible at every such covariate value.

Conditional exchangeability is the no-unmeasured-confounding condition relative to $X$; it is not established merely by calculating a propensity score. Positivity is also a population condition, not just a rule that all estimated scores must look comfortably away from zero and one. In applications, estimated probabilities near the boundaries can still make weights highly variable.

Theorem: Under consistency, conditional exchangeability, positivity, and finite outcome means, $$ E\left[\frac{AY}{e(X)}\right]=E[Y(1)] \qquad\text{and}\qquad E\left[\frac{(1-A)Y}{1-e(X)}\right]=E[Y(0)]. $$ Therefore, $$ E\left[\frac{AY}{e(X)}-\frac{(1-A)Y}{1-e(X)}\right] =E[Y(1)]-E[Y(0)], $$ which is the ATE.

Deriving the Treated-Outcome Identity

The first expression gives zero contribution to units who were not treated because $A=0$ for them. Among treated units, consistency makes their observed outcome equal to $Y(1)$. Conditional exchangeability then lets the treated units’ conditional mean represent the potential-outcome mean at the same $X$. The denominator $e(X)$ cancels the conditional probability of being treated.

\begin{aligned} E\left[\frac{AY}{e(X)}\right] &=E_X\left[E\left\{\left.\frac{AY}{e(X)}\right|X\right\}\right]\\ &=E_X\left[\frac{E[AY\mid X]}{e(X)}\right]\\ &=E_X\left[\frac{E[AY(1)\mid X]}{e(X)}\right]\\ &=E_X\left[\frac{P(A=1\mid X)E[Y(1)\mid A=1,X]}{e(X)}\right]\\ &=E_X\left[E[Y(1)\mid X]\right]\\ &=E[Y(1)]. \end{aligned}

The third line uses consistency for treated units: when $A=1$, $Y=Y(1)$, so $AY=AY(1)$. In the fourth line, $E[AY(1)\mid X]$ equals $P(A=1\mid X)E[Y(1)\mid A=1,X]$. Conditional exchangeability gives $E[Y(1)\mid A=1,X]=E[Y(1)\mid X]$; by definition, $P(A=1\mid X)=e(X)$. Positivity ensures the division by $e(X)$ is defined for the target population. The last line averages the conditional mean over the population distribution of $X$.

For the control outcome, consistency gives $(1-A)Y=(1-A)Y(0)$. Conditional exchangeability gives $E[Y(0)\mid A=0,X]=E[Y(0)\mid X]$, while $P(A=0\mid X)=1-e(X)$. Applying the same conditional-expectation steps yields $$ E\left[\frac{(1-A)Y}{1-e(X)}\right] =E_X\left[\frac{(1-e(X))E[Y(0)\mid A=0,X]}{1-e(X)}\right] =E[Y(0)]. $$ Subtracting the control identity from the treated identity gives the ATE identity in the theorem.

Worked Examples

Worked Example: Two Covariate Groups and Exact Recovery

Consider a hypothetical population in which $X$ has two equally common groups. The treatment probabilities are $e(0)=0.20$ and $e(1)=0.80$. Suppose the conditional potential-outcome means are $E[Y(1)\mid X=0]=10$, $E[Y(1)\mid X=1]=14$, $E[Y(0)\mid X=0]=4$, and $E[Y(0)\mid X=1]=8$. Assume consistency, conditional exchangeability, and positivity hold.

The marginal treated potential-outcome mean is obtained by averaging over the population groups: $$ E[Y(1)]=0.50(10)+0.50(14)=12. $$ The treated IPW identity gives the same result group by group. In group $X=0$, the probability of treatment $0.20$ is canceled by the inverse weight $1/0.20=5$; in group $X=1$, $0.80$ is canceled by $1/0.80=1.25$: $$ E\left[\frac{AY}{e(X)}\right] =0.50\left(\frac{0.20(10)}{0.20}\right) +0.50\left(\frac{0.80(14)}{0.80}\right) =0.50(10)+0.50(14)=12. $$

For control, the treatment probabilities are $0.80$ in group 0 and $0.20$ in group 1. Thus $$ E\left[\frac{(1-A)Y}{1-e(X)}\right] =0.50\left(\frac{0.80(4)}{0.80}\right) +0.50\left(\frac{0.20(8)}{0.20}\right) =0.50(4)+0.50(8)=6. $$ The ATE is $12-6=6$ outcome units. All treatment probabilities are strictly between zero and one, and the example stipulates exchangeability and consistency, so the identity applies.

Worked Example: Unequal Covariate Prevalence

In a second hypothetical population, $P(X=1)=0.30$ and $P(X=0)=0.70$. Let $e(0)=0.25$ and $e(1)=0.75$. Suppose the conditional means are $E[Y(1)\mid X=0]=8$, $E[Y(1)\mid X=1]=16$, $E[Y(0)\mid X=0]=3$, and $E[Y(0)\mid X=1]=7$. Assume the three identification conditions hold.

First calculate the marginal potential-outcome means: $$ E[Y(1)]=0.70(8)+0.30(16)=5.60+4.80=10.40, $$ $$ E[Y(0)]=0.70(3)+0.30(7)=2.10+2.10=4.20. $$ Thus the ATE is $10.40-4.20=6.20$ outcome units.

The population treated-weighted expectation recovers the treated mean: $$ E\left[\frac{AY}{e(X)}\right] =0.70\left(\frac{0.25(8)}{0.25}\right) +0.30\left(\frac{0.75(16)}{0.75}\right) =0.70(8)+0.30(16)=10.40. $$ Similarly, the control-weighted expectation is $$ E\left[\frac{(1-A)Y}{1-e(X)}\right] =0.70\left(\frac{0.75(3)}{0.75}\right) +0.30\left(\frac{0.25(7)}{0.25}\right) =0.70(3)+0.30(7)=4.20. $$

The unweighted observed group means need not equal these marginal means. Here $P(A=1)=0.70(0.25)+0.30(0.75)=0.40$. Among treated units, the probability of $X=1$ is $0.30(0.75)/0.40=0.5625$, so their observed mean is $0.4375(8)+0.5625(16)=12.50$. Among controls, $P(A=0)=0.60$ and the probability of $X=1$ is $0.30(0.25)/0.60=0.125$, giving mean $0.875(3)+0.125(7)=3.50$. The crude difference is $12.50-3.50=9.00$, not the ATE of $6.20$. It compares groups with different covariate distributions.

Worked Example: What Fails When Positivity Fails

Suppose a hypothetical target population has two equally common groups. Everyone in group $X=0$ is untreated, so $e(0)=0$, and their observed outcome is 4. Everyone in group $X=1$ is treated, so $e(1)=1$, and their observed outcome is 10. The consistency condition can hold, but positivity fails in both groups.

The treated weighting expression divides by $e(0)=0$ for group 0, and the control expression divides by $1-e(1)=0$ for group 1. More fundamentally, the data do not reveal $Y(1)$ for group 0 or $Y(0)$ for group 1. To see why this matters, consider two possible underlying populations with the same observed outcomes. In one, the missing potential outcomes are $Y(1)=4$ for group 0 and $Y(0)=10$ for group 1. Then the ATE is $0.50(4+10)-0.50(4+10)=0$. In another, those missing outcomes are $Y(1)=8$ and $Y(0)=6$. The ATE is then $0.50(8+10)-0.50(4+6)=4$ outcome units. Both populations produce the stated observed data, but have different ATEs. Without further assumptions, the observed data cannot identify which ATE is correct.

What the Result Does—and Does Not—Establish

The theorem explains why inverse-probability weighting targets a marginal effect: within each covariate value, the treatment-assignment probability in the numerator of the conditional expectation is canceled by its inverse in the weight. Averaging the resulting conditional potential-outcome mean over the target population distribution of $X$ then gives the marginal mean. The same reasoning applies to control.

This is not a guarantee that an analysis using estimated propensity scores will work well. In practice, the weights use estimated rather than known probabilities, and the sample is finite. Very small estimated $e(X)$ values for treated units, or very small $1-e(X)$ values for controls, can produce large weights. Such weights can make estimates unstable and sensitive to a few observations. Inspect the distributions of estimated propensities and weights, and assess whether treatment groups have covariate support in common. Covariate imbalance after weighting is a warning that the chosen model or available overlap may be inadequate; apparent balance alone does not prove conditional exchangeability.

The key causal assumption that cannot generally be verified from observed outcomes is conditional exchangeability. If important common causes of treatment and outcome are omitted from $X$, the weighted groups may remain confounded even if the propensity model is estimated accurately. Positivity also matters to the estimand: removing unsupported regions or changing the target population may be necessary in an application, but that changes the question being answered.

Comparison With Other Approaches

Approach	How it targets the ATE	Key caution
Inverse probability weighting	Weights observed outcomes by inverse treatment probabilities so each treatment group represents the target covariate distribution in expectation.	Extreme weights can make finite-sample results unstable; identification still requires the stated assumptions.
Outcome standardization (g-computation)	Models conditional outcome means and averages predicted outcomes over the target population’s covariates.	Relies on adequate outcome-model specification and the causal identification assumptions.
Unadjusted group contrast	Subtracts the observed mean outcome among controls from that among treated units.	Generally compares different covariate distributions, so it does not identify the ATE when those differences confound treatment and outcome.

IPW and outcome standardization use different statistical models, but neither removes the need for consistency, conditional exchangeability, and positivity. The unadjusted contrast is appropriate as a causal estimate only under additional conditions that make the treatment groups comparable without adjustment.

Common Mistakes

Confusing identification with estimation. The theorem is about population expectations. A finite-sample weighted mean will generally not equal the population potential-outcome mean exactly.
Assuming weights create exchangeability. Weighting cannot adjust for a confounder that is not adequately represented in the measured covariates and propensity model.
Ignoring positivity because no estimated probability is exactly zero. Probabilities close to zero or one can still lead to highly variable weights, and a true structural lack of support cannot be fixed by computation.
Interpreting the weighted treated group as the treated population only. For the ATE identity, treated and control weighted expectations each target the marginal population mean of their respective potential outcome.

Key takeaway: Under consistency, conditional exchangeability, and positivity, inverse treatment-probability weighting cancels the treatment-selection probability conditional on $X$. The resulting expectations identify $E[Y(1)]$ and $E[Y(0)]$; their difference identifies the ATE.

Check Your Understanding

Use the identification conditions and equations above to answer each question. Solutions follow the questions.

If $e(X)=0.40$ for a covariate group, what inverse weight appears in the treated identity for that group?
A population has two equally common covariate groups with conditional treated potential-outcome means 5 and 11. What is $E[Y(1)]$, and what does the treated weighting identity equal?
In a population, $P(X=0)=0.60$, $P(X=1)=0.40$, and the conditional control potential-outcome means are 2 and 7. Find $E[Y(0)]$.
Why can accurate estimation of $e(X)$ not repair unmeasured confounding?
What identification problem arises if $e(X)=1$ for a covariate value present in the target population?

Solutions. (1) The treated weight is $1/e(X)=1/0.40=2.5$. (2) The marginal mean is $0.50(5)+0.50(11)=8$; under the assumptions, $E[AY/e(X)]=8$ as well. (3) $E[Y(0)]=0.60(2)+0.40(7)=1.20+2.80=4.00$. (4) If a confounder is omitted, conditioning on the modeled covariates may not make treatment independent of the potential outcomes; correctly estimating treatment probabilities based on incomplete covariates does not supply the missing information. (5) No untreated units occur at that covariate value, so the data do not reveal $Y(0)$ there; the control weight would also divide by $1-e(X)=0$. The marginal control mean, and hence the ATE without further assumptions, is not identified by these data.