Choose the Propensity-Score Target Before Choosing Weights

State whose effect matters

The average treatment effect in all eligible people and the average effect among those actually treated are different targets. They may differ when effects vary across people. Write the target population and treatment contrast before selecting matching or weighting settings.

The WeightIt weighting-methods documentation distinguishes supported estimands and links their choice to the substantive question. Software labels do not decide which population is clinically or policy relevant, and a default ATE setting is not a neutral choice when the research question concerns treated people.

Map the target to its weighting rule

For a binary treatment with propensity score e(X), conventional ATE weights are 1/e(X) for treated observations and 1/(1−e(X)) for controls. Conventional ATT weights are 1 for treated observations and e(X)/(1−e(X)) for controls. These rules emphasize different parts of the observed population.

Record which treatment level is considered treated and confirm that coding is consistent across the score model, weights and outcome contrast. If using overlap weights or another target, describe the represented population explicitly rather than reporting a generic propensity-weighted effect.

Compare a hypothetical observation

Consider a control observation with estimated treatment probability 0.8. Its ATE weight is 1/(1−0.8) = 5, whereas its ATT weight is 0.8/(1−0.8) = 4. A treated observation with probability 0.8 receives ATE weight 1.25 but ATT weight 1.

These differences are not software discrepancies. They arise because the estimands require different reweighting. The example also shows why copying weights from a previous analysis without preserving its target definition can silently change the meaning of a new comparison.

Review balance and support for that target

Inspect covariate balance against the intended target distribution, not an arbitrary pooled distribution. Examine overlap, influential weights and effective information in both groups. A satisfactory balance summary does not rule out unmeasured confounding, selection bias or poor measurement of the covariates used.

Keep the estimand, eligibility rules, treatment coding and weighting method together in the analysis specification. If inadequate support leads to a revised target, label it as a substantive change and describe who is no longer represented. Report the treatment effect on a clearly named scale with an uncertainty method compatible with weight estimation. Readers should be able to tell whose outcomes the estimate averages, not merely which package produced it.

Source verification: 6 November 2026. Examples are hypothetical.