Treatment-Specific Support Restrictions

Why Treatment-Specific Support Sets Matter

A propensity-score overlap plot can show that treated and control units occupy different score ranges. A support restriction makes that diagnosis operational: it specifies which scores are considered usable in each treatment group, then retains units only where both groups have adequate support. This matters because a score can be familiar in one group yet poorly represented in the other. Keeping a unit on the basis of only one group’s score range does not establish that the comparison group is available at that score.

Let \(A\in\{0,1\}\) denote treatment status, let \(X\) be the measured covariates, and let \(e(X)=P(A=1\mid X)\) be the propensity score. The aim here is to define and apply a support restriction, not to select its numerical boundaries. The introductory material on positivity and common-support restrictions motivates why unsupported comparisons are problematic; the key task here is to keep the group-specific rules logically consistent.

The neighborhood condition is important for continuous scores: a particular score value can have probability zero even when it is a plausible value in the group’s distribution. For example, a continuously distributed score can have positive probability in every small interval around 0.40 even though \(P(e(X)=0.40\mid A=a)=0\). Defining support by positive probability at the individual value would therefore be wrong.

In practice, analysts usually work with estimated or operational sets, such as score intervals judged to have adequate representation in each arm. Denote these sets by \(G_1\) and \(G_0\). They need not be exact population supports, and an observed unit can fall outside the operational set assigned to its own group. The safe common restriction is therefore explicit:

This rule is applied to both treated and control units. It checks membership in both sets, rather than assuming that observed scores automatically lie in their own group’s operational set.

From Group-Specific Rules to a Common Restriction

A common way to describe the rule is in terms of the comparison group: retain treated units only where controls have support, and retain controls only where treated units have support. In symbols, these are \(e(X)\in G_0\) for treated units and \(e(X)\in G_1\) for controls. Those opposite-group checks alone are equivalent to the intersection rule only when membership in each unit’s own group set is already guaranteed.

For exact population supports, that own-group membership holds with probability one: a random score drawn from the distribution whose support is \(S_a\) lies in \(S_a\) almost surely. Thus the opposite-group check yields the common support, up to probability-zero exceptions. With estimated or deliberately defined adequate-support sets, that guarantee does not generally hold. In that setting, require both own- and opposite-group membership, or simply apply \(G_1\cap G_0\) to every unit.

The short justification for the exact-support statement is that the score distribution in arm \(a\) assigns probability zero to the complement of its support. For a real-valued random variable, the complement can be covered by a countable collection of open intervals, each having probability zero by the definition of support. Their countable union also has probability zero. Consequently, a treated score lies in \(S_1\) almost surely and a control score lies in \(S_0\) almost surely. Intersecting with the other group’s support then leaves scores in \(S_1\cap S_0\).

Asymmetric Sets and Unequal Retention

The two group-specific support sets can be asymmetric: they may have different lower or upper limits, or differ in shape when they are not intervals. Their intersection is determined by the region both sets share. The common score region applies to both treatment groups, but the number or proportion retained in each group can still differ because the two groups have different score distributions.

This distinction separates asymmetric support information from different retention rules by treatment arm. The first is compatible with a common restriction: construct \(G_1\cap G_0\), then apply it to both arms. The second can retain treated scores in one region and control scores in another. Unless the regions coincide, the retained treated and control units are no longer selected from one shared score-defined population. That may be a different design with a different estimand, and it should not be presented as the ordinary common-support restriction.

What Population Does the Restriction Describe?

A shared score restriction defines a retained population, not merely a convenient subset of the original data. If \(C\) is the selected score region, the target population consists of units with \(e(X)\in C\). For example, a causal estimand for that restricted population can be written

where \(Y(1)\) and \(Y(0)\) are the potential outcomes under treatment and control. Interpreting an estimate as this causal effect still requires the relevant causal assumptions, including consistency and conditional exchangeability given the measured covariates. A restriction aimed at positivity does not establish those assumptions, nor does it guarantee adequate support for every covariate pattern merely because score ranges overlap.

If treated and control units instead receive different score restrictions, the retained treated units may correspond to one target population and the retained controls to another. A contrast between their outcomes does not automatically estimate the effect in either group’s retained population. The analyst must define and justify the target estimand rather than treating different arm-specific samples as interchangeable.

Diagnostics, Alternatives, and Common Pitfalls

Before using a support restriction, inspect the score distributions by treatment group and report the actual sets applied. Useful checks include confirming the interval intersection, checking endpoint inclusion, counting retained and excluded units separately by arm, and verifying that each retained score belongs to both operational sets. A table of counts by arm can reveal a mistaken implementation, although unequal counts alone do not indicate an error.

The main alternatives answer related but distinct design questions. Symmetric trimming uses a shared score rule, often specified with the same cutoff logic for both groups. Treatment-specific support sets allow the evidence about adequate coverage to differ by arm, but their intersection still supplies a common retained score region. Weighting can retain more observations while giving some observations greater influence; it does not make unsupported comparisons reliable merely by assigning weights. Outcome-model extrapolation may produce predictions in regions with little or no comparison-group data, but those predictions depend on modeling assumptions rather than observed overlap.

• Wrong reasoning: “This control score lies in the treated support, so it must be retained.”
• Correction: With estimated or operational sets, also check that it lies in the control set. Retain it only if it belongs to their intersection.
• Wrong reasoning: “The intervals are asymmetric, so the restriction must retain different score ranges by arm.”
• Correction: Asymmetric group-specific sets can define a single shared range through their intersection. Apply that range to both groups; different retained counts can result from different group distributions.
• Wrong reasoning: “Every observed score must belong to the support range reported for its own arm.”
• Correction: This is true almost surely for exact population supports, not necessarily for estimated boundaries or operational adequate-support sets.

Support restrictions also have limitations. Their boundaries may be uncertain, and a restriction can reduce sample size or change the population to which findings apply. A shared score range is not proof of covariate balance, adequate sample size, or causal identification. Those questions require additional diagnostics and assumptions.