Audit Stabilized Inverse-Probability Weights From Numerator to Outcome Model

Write both probability models

A stabilized weight contains a numerator as well as a denominator. Record exactly which variables enter each model and which observed treatment probability is used for each participant. For a simple binary baseline treatment, stabilization commonly multiplies an ATE weight by the marginal probability of the observed treatment group.

The WeightIt probability-to-weight reference documents this construction and its estimand options. Stabilization is not the same operation as truncating large weights, and it does not eliminate the need for adequate overlap or measured-confounder control.

Check coding with a small audit table

For several synthetic observations, retain treatment status, propensity score, numerator, denominator and final weight. Verify that controls use the probability of receiving control, not the probability of treatment. Check that every probability lies in its valid range and that no denominator is zero.

Then summarize weight quantiles, extreme observations and the weighted covariate distribution. Compare the results with an unstabilized implementation on the same target as a diagnostic. Do not judge success solely by whether the average weight is close to one; model misspecification and omitted confounding can remain.

Recompute a hypothetical pair

Suppose 30% of a fictional sample received treatment. A treated observation has estimated treatment probability 0.1. Its stabilized ATE weight is 0.3/0.1 = 3, compared with unstabilized weight 10.

A control observation has estimated treatment probability 0.8, so its probability of the observed control status is 0.2. Its stabilized weight is 0.7/0.2 = 3.5, compared with unstabilized weight 5. Using 0.3 in the control numerator or 0.8 in its denominator would represent a different and incorrect calculation for this stated scheme.

Align the outcome analysis and interpretation

For longitudinal treatment, products of time-specific probabilities and time-varying covariate histories require a more detailed specification. The simple baseline example is not a substitute for that analysis. Explain how censoring weights, if used, combine with treatment weights and which assumptions each component introduces.

Use an uncertainty procedure appropriate to the estimated weights and outcome estimator. Preserve numerator and denominator model versions, diagnostics and any truncation sensitivity analyses. Report the target population and remaining assumptions about exchangeability, consistency and positivity. A less variable weight distribution can improve numerical behavior, but it does not turn an observational comparison into a randomized experiment or prove that the chosen models are correct.

Source verification: 6 November 2026. Examples are hypothetical.