The Pseudo-Population Interpretation of Weighting

What Does Weighting Create?

Inverse probability weighting changes how much each observed person contributes to an analysis. The pseudo-population interpretation gives a useful way to understand that change: under ideal conditions, weighting makes treatment independent of the measured covariates. The weighted treated and untreated groups then have the same covariate distribution, even if the original groups did not.

“Pseudo-population” does not mean that new people have been observed or that the original treatment assignments have changed. It is a population distribution defined by weighting observed records. The interpretation is exact for the population when the true propensity score is used and positivity holds. In a finite data set, especially with estimated propensity scores, the resulting balance is generally approximate and must be checked.

The balance identity below is a property of the treatment-assignment probabilities and the weights. It does not, by itself, require assumptions about potential outcomes. To interpret a weighted outcome contrast causally as an ATE, additional assumptions are needed: consistency, no interference between units, conditional exchangeability \(\{Y(1),Y(0)\}\perp A\mid X\), and positivity \(0<e(X)<1\) for covariate values in the target population. Correctly specified or otherwise consistently estimated propensity scores are also needed for the corresponding estimation procedure to work as intended.

The Balance Identity for ATE Weights

A distribution is characterized by the probabilities it assigns to measurable sets. Thus, to establish balance for general covariates—including continuous covariates—we show equality for every measurable set \(B\) of possible covariate values. We do not rely on the probability of any single covariate value being positive.

For treated people, the expected weighted contribution from people whose covariates fall in \(B\) is

The first equality conditions on \(X\). Because \(E[I(A=1)\mid X]=P(A=1\mid X)=e(X)\), the ratio in the middle is one wherever the weight is defined. Positivity ensures \(e(X)>0\) for treated weighting on the target support. For untreated people, the same argument gives

since \(P(A=0\mid X)=1-e(X)\). Each treatment group’s weighted covariate distribution therefore reproduces the source distribution of \(X\): after normalizing its total weighted mass, the probability of \(X\in B\) is \(P(X\in B)\). This holds for every measurable \(B\), so it establishes equality of the full distributions, not merely equality of selected means.

In an ideal population calculation, the total expected weight in each group is one: set \(B\) to the whole covariate space in the two identities. In an empirical sample, random variation and estimated scores mean neither exact total nor exact balance is guaranteed. The identity describes the target property of the weighting scheme, not a promise that every finite weighted data set will balance perfectly.

Stabilized Weights and Treatment Prevalence

Stabilized ATE weights multiply each unstabilized weight by the marginal probability of the person’s observed treatment. Write \(p=P(A=1)\). The treated weight is \(p/e(X)\), and the untreated weight is \((1-p)/\{1-e(X)\}\). These weights preserve the marginal treatment proportions in the ideal pseudo-population while still making treatment independent of \(X\).

To see both properties together, define the weighted joint probability of a covariate set and treatment group by taking the expected stabilized weight times the corresponding indicators. For any measurable \(B\),

Adding the two contributions gives weighted probability \(P(X\in B)\). In particular, the weighted marginal probability of treatment is \(p\), and the weighted joint probabilities factor into the treatment and covariate marginals. That factorization is precisely independence in the pseudo-population. Multiplying all weights by the same positive constant does not change the normalized distribution; what matters here is the relative weight assigned across people.

Continuous Covariates and the Role of the Set Argument

The measurable-set proof matters whenever \(X\) is continuous. If \(X\) has a continuous distribution, then typically \(P(X=x)=0\) for every individual value \(x\). A statement about the probability at one exact value would not establish balance. Probabilities over sets—such as intervals or regions—do.

What Balance Does—and Does Not—Establish

The pseudo-population interpretation is useful because it translates a formula into a distributional claim: the weighted treatment groups represent the same measured-covariate population. This is the intended design effect of ATE weighting. It can also help identify a mismatch between an intended weighting scheme and its implementation.

In practice, weights are usually calculated from an estimated propensity score \(\hat e(X)\), not the true \(e(X)\). The exact identity relies on the conditional probability \(P(A=1\mid X)\), so substituting an inaccurate estimate need not produce exact balance. Even a correctly estimated model need not yield exact empirical balance, particularly in a small sample. Balance should be assessed in the observed weighted data using covariate summaries and distributional comparisons relevant to the analysis.

A second condition is positivity. If some people have propensity scores near zero or one, the corresponding inverse probability weights become large. A few records can then dominate the weighted distribution, making estimates unstable and sample balance difficult to achieve. If a treatment group is absent for a covariate pattern, no finite inverse probability weight can create observed representation of that group in that pattern. Trimming, changing the target population, or choosing another method changes the analysis and requires substantive justification; it is not a way to make the original positivity assumption true.

Finally, balance on measured \(X\) is not proof of exchangeability. A weighted outcome contrast has a causal interpretation only when the relevant identification assumptions—including conditional exchangeability given the measured covariates—are credible. Weighting cannot balance an unmeasured confounder merely by balancing the observed covariates.

Comparison and Common Pitfalls

Weighting targets balance by reweighting all eligible observations, whereas matching constructs or selects comparison sets and may leave some observations unused. Both approaches depend on adequate overlap and measured covariates; they can yield different target populations when observations are excluded or weights are highly variable. Outcome regression instead models outcomes conditional on covariates and treatment. These approaches are not interchangeable guarantees: the choice depends on the estimand, design, assumptions, and diagnostics. The introductory material on inverse probability weighting establishes the weighting estimands and estimators; the focus here is the distributional interpretation behind their balance goal.

• Pitfall: treating exact population balance as a finite-sample guarantee. The identity uses the true propensity score in expectation. Estimated scores and random samples can leave residual imbalance.
• Pitfall: checking only covariate means. Matching means does not establish equality of distributions. Inspect distributions and important nonlinearities or interactions as well as means.
• Pitfall: interpreting balance as proof of no confounding. Balance refers to measured covariates. Unmeasured confounding and other identification failures remain possible.
• Pitfall: overlooking near-zero or near-one scores. Extreme weights indicate weak overlap and can make the weighted result depend heavily on a few observations.