Overlap Weights and the ATO Population

Why Overlap Weights Define a Distinct Target

When treatment is much more likely for some covariate profiles than others, a population-wide treatment effect can depend heavily on comparisons that the observed data barely support. Overlap weights address this by emphasizing profiles for which either treatment could plausibly have been received. They do not simply “fix” poor overlap: they define a particular target population and an effect for that population.

Let $A\in\{0,1\}$ denote treatment, $X$ baseline covariates, and $Y(1)$ and $Y(0)$ the potential outcomes under treatment and control. The propensity score is $e(X)=P(A=1\mid X)$. The individual treatment effect is $Y(1)-Y(0)$. The target here is an average of this effect over a covariate distribution tilted toward treatment equipoise.

Definition: The overlap tilting function is $h(X)=e(X)\{1-e(X)\}$. The average treatment effect in the overlap population (ATO) is $$ \tau_{\mathrm{ATO}}= \frac{E[h(X)\{Y(1)-Y(0)\}]}{E[h(X)]}. $$ The target covariate distribution is the original distribution of $X$, reweighted in proportion to $h(X)$. Since $e(1-e)$ is largest at $e=0.5$ and approaches zero as $e$ approaches zero or one, the ATO most directly describes units for whom treatment assignment is relatively uncertain given $X$.

The ATO is a population effect, but not generally the average treatment effect (ATE) in the original population. Nor is it generally the average treatment effect among treated units (ATT). Its interpretation depends on the propensity-score distribution: changing the population or treatment-assignment process can change which covariate profiles receive the most target weight.

How the Treatment-Specific Weights Produce the Target

The overlap tilting function gives the target distribution; the weights applied to observed treated and control units are slightly different. This difference compensates for the fact that the treated and control groups have different covariate distributions before weighting.

Definition: For a treated unit, the overlap weight is $1-e(X)$. For a control unit, it is $e(X)$. Equivalently, both groups use the target tilt $h(X)=e(X)\{1-e(X)\}$ divided by their probability of being in their observed treatment group.

To see why, let $f(x)$ denote the population covariate density. Before weighting, the covariate density among treated units is proportional to $e(x)f(x)$. Multiplying it by the treated-unit weight $1-e(x)$ gives a density proportional to $e(x)\{1-e(x)\}f(x)$. Among controls, the covariate density is proportional to $\{1-e(x)\}f(x)$; multiplying by the control weight $e(x)$ gives the same result. Normalization makes the two weighted distributions identical.

\begin{aligned} f(x\mid A=1)\{1-e(x)\} &\ \propto\ e(x)f(x)\{1-e(x)\} = h(x)f(x),\\ f(x\mid A=0)e(x) &\ \propto\ \{1-e(x)\}f(x)e(x) = h(x)f(x). \end{aligned}

This is a population identity when the propensity score is the true treatment probability. In a finite sample, estimated propensity scores and random variation mean that the weighted covariate distributions need not match perfectly.

Worked Example: Deriving the Overlap Target Across Three Strata

Consider a hypothetical population divided into three baseline-covariate strata. Their population proportions are 0.50, 0.30, and 0.20, and their treatment probabilities are 0.10, 0.50, and 0.90, respectively. Find each stratum’s share of the overlap target.

For each stratum, multiply its population proportion by $h=e(1-e)$. The unnormalized target shares are:

\begin{aligned} \text{Stratum 1: }&0.50(0.10)(0.90)=0.045,\\ \text{Stratum 2: }&0.30(0.50)(0.50)=0.075,\\ \text{Stratum 3: }&0.20(0.90)(0.10)=0.018. \end{aligned}

The total is $0.045+0.075+0.018=0.138$. Divide each unnormalized share by 0.138. The target proportions are $0.045/0.138=0.3261$, $0.075/0.138=0.5435$, and $0.018/0.138=0.1304$. They sum to 1, up to rounding.

The middle stratum has a 30% share of the original population but about a 54.35% share of the overlap target. The first and third strata have treatment probabilities near zero and one, so each receives less target weight than its original population share. The target does not exclude them by a hard cutoff; it smoothly reduces their influence.

Identification and Estimation

Overlap weighting defines a target, but a causal interpretation also requires assumptions linking observed outcomes to potential outcomes. In particular, consistency means the observed outcome is the potential outcome corresponding to the treatment actually received. Conditional exchangeability requires $Y(1),Y(0)\perp A\mid X$. The covariates $X$ must be measured before treatment and sufficiently control confounding for this assumption to be plausible.

For identification, there must also be treatment variation in the covariate profiles contributing to the target. In the usual positivity condition, $0<e(X)<1$ over the relevant covariate support. Profiles with $e=0$ or $e=1$ have $h=0$, so they contribute no overlap target weight. However, if nearly all profiles are close to those extremes, the ATO can represent a narrow subset of the original population.

Under these assumptions, the conditional mean difference among units with covariates $X=x$ identifies the conditional average treatment effect. Averaging that difference using the target covariate distribution identifies the ATO:

\tau_{\mathrm{ATO}} = \frac{E\left[h(X)\left\{E(Y\mid A=1,X)-E(Y\mid A=0,X)\right\}\right]} {E[h(X)]}.

A direct sample estimator uses the treated and control overlap weights to form two weighted outcome means, then takes their difference. With estimated propensity scores $\hat e_i$, the estimator is:

\widehat{\tau}_{\mathrm{ATO}} = \frac{\sum_{i:A_i=1}(1-\hat e_i)Y_i}{\sum_{i:A_i=1}(1-\hat e_i)} - \frac{\sum_{i:A_i=0}\hat e_iY_i}{\sum_{i:A_i=0}\hat e_i}.

The two weighted means target the same overlap covariate distribution in the population. Their difference therefore targets an average treatment effect for that distribution, provided the identification assumptions hold. The weighting formula alone does not ensure exchangeability, correct propensity-score estimation, or adequate finite-sample balance.

Worked Example: Calculating an ATO From Stratum Effects

Use the three-stratum hypothetical population from the previous example. Suppose the stratum-specific average treatment effects are 2, 5, and 8 outcome points. Calculate the ATO.

The normalized overlap-target proportions are 0.3261, 0.5435, and 0.1304. The ATO is their weighted average of the stratum-specific effects:

\begin{aligned} \tau_{\mathrm{ATO}} &=0.3261(2)+0.5435(5)+0.1304(8)\\ &=0.6522+2.7175+1.0432\\ &=4.4129\text{ outcome points}. \end{aligned}

Thus the average effect for this overlap target is about 4.413 outcome points. It is not the ATE for the original population: weighting the three effects by the original proportions would give $0.50(2)+0.30(5)+0.20(8)=4.1$ points. The difference arises because the target gives more weight to the middle stratum, where treatment is most balanced and the effect is 5 points.

Worked Example: Computing Weighted Outcome Means

In a small hypothetical data set, suppose the propensity scores are known. Two treated units have scores 0.20 and 0.80 and outcomes of 10 and 14 points. Two control units have the same respective scores and outcomes of 7 and 9 points. Calculate the overlap-weighted outcome contrast.

The treated units receive weights $1-e$, so their weights are 0.80 and 0.20. The treated weighted mean is:

\frac{0.80(10)+0.20(14)}{0.80+0.20} =\frac{8+2.8}{1}=10.8\text{ points}.

The control units receive weights $e$, or 0.20 and 0.80. Their weighted mean is:

\frac{0.20(7)+0.80(9)}{0.20+0.80} =\frac{1.4+7.2}{1}=8.6\text{ points}.

The estimated contrast is $10.8-8.6=2.2$ points. This arithmetic illustrates the estimator, not a reliable estimate from a large study: with only four units, sampling variation may be substantial, and the causal interpretation still depends on the identification assumptions.

What the ATO Does—and Does Not—Say

The ATO’s defining feature is its focus on treatment equipoise conditional on measured covariates. A propensity score near 0.5 indicates that, given those covariates, either treatment was reasonably plausible. The overlap tilt $e(1-e)$ reaches its maximum of 0.25 at 0.5 and is symmetric: $e=0.1$ and $e=0.9$ have the same tilt. This symmetry is about the propensity score, not a claim that the units or their treatment effects are otherwise alike.

A useful contrast is the target induced by common weighting choices. ATE weighting targets the original covariate distribution, ATT weighting targets the treated covariate distribution, and overlap weighting targets a distribution proportional to $e(1-e)f(x)$. The ATO is useful when the scientific question concerns the population in which treatment choice is most ambiguous, or when very extreme propensity scores make ATE or ATT weighting highly sensitive to a few units. It should not be selected merely because its estimate is convenient.

Overlap weighting is also different from restricting the sample to a chosen common-support range. A restriction discards units according to a cutoff and changes the target to the retained population. Overlap weighting smoothly changes the contribution of units across propensity scores. Both approaches can shift the target away from the original population, and neither creates information about treatment effects in covariate regions where one treatment is absent.

Diagnostics, Failure Modes, and Limitations

Diagnostics should address both the propensity-score model and the weighted comparison. Inspect the estimated propensity-score distributions by treatment group, identify regions with few or no units in one group, and check covariate balance after weighting. Examine the actual weight distributions and whether a small number of observations dominate the weighted means. These checks do not prove exchangeability; they can reveal important failures of the design.

A practical advantage of overlap weights is that their raw unit weights are bounded between zero and one when propensity scores lie between zero and one. This avoids the unbounded inverse-probability weights that can occur near propensity-score extremes. Bounded weights do not guarantee precision or good identification. If overlap is poor, few units may carry appreciable weight, the effective information can still be limited, and the ATO may apply only to a small or unusual part of the source population.

Estimated propensity scores also matter. A misspecified treatment-assignment model can leave covariate imbalance after weighting. A correctly fitted score model does not address unmeasured confounding, and balance on observed covariates does not establish that all confounders were measured. If the ATO is reported, describe its target and the population to which it is intended to apply; do not label it simply “the treatment effect” without qualification.

Common mistake: “Overlap weights put more weight on people who resemble the whole study population, so their result estimates the population ATE.” This is not generally correct. The weighting function $e(1-e)$ changes the covariate distribution, emphasizing units near treatment equipoise. The resulting causal contrast is the ATO, unless the ATO happens to equal the ATE in a particular setting. Report the target explicitly and do not infer an effect for poorly represented profiles from the overlap-weighted result.

The choice among overlap weighting, ATE or ATT weighting, and sample restriction is ultimately a choice about the question and target population, as well as a statistical design decision. State the target first, check whether the data support comparisons for it, and interpret the result only for that target under the stated causal assumptions.

Check Your Understanding

Answer each question, then compare your reasoning with the brief solution.

A covariate profile has propensity score 0.25. What are its overlap tilt, treated-unit weight, and control-unit weight?
Solution: The tilt is $0.25(0.75)=0.1875$. The treated-unit weight is $1-0.25=0.75$, and the control-unit weight is $0.25$.
Why do treated and control units receive different overlap weights even though both weighted groups target the same covariate distribution?
Solution: Their unweighted covariate distributions differ by their treatment probabilities. The treated density is proportional to $ef(x)$, so weighting by $1-e$ yields a density proportional to $e(1-e)f(x)$. The control density is proportional to $(1-e)f(x)$, so weighting by $e$ yields that same target.
Suppose two strata have normalized overlap-target proportions 0.60 and 0.40, with treatment effects of 3 and 6 days. What is the ATO?
Solution: $0.60(3)+0.40(6)=1.8+2.4=4.2$ days.
Does a propensity score of 0 or 1 contribute positive overlap target weight? What is the implication?
Solution: No. At either extreme, $e(1-e)=0$. Such profiles do not contribute to the overlap target; if many profiles are near these extremes, the ATO may describe a limited subset of the original population.
After overlap weighting, observed covariates are balanced between treatment groups. Does that establish a causal interpretation?
Solution: No. Balance is a useful diagnostic for measured covariates, but it does not establish conditional exchangeability or rule out unmeasured confounding. Consistency and the other identification conditions must also be plausible.