Overlap weighting
Overlap weighting emphasizes patients who could plausibly receive either treatment, then compares weighted outcomes in that overlap population.
Fit
When to use it, and when not to
Use it when
- You compare two baseline treatments and the overlap population is clinically meaningful.
- Extreme treatment chances make full-population weighting unstable.
Not when
- You need the average effect for everyone eligible. Compare propensity weighting targets before choosing the analysis. Read: Choosing propensity matching or weighting
- Treatment and confounders change together over time. Consider a marginal structural model. Read: Choosing an analysis for time-varying treatment
Before you start
What you need
- A defined overlap populationExplain whose treatment choice the result will inform.
- Baseline treatment chances and observed outcomesUse factors measured before treatment and a shared outcome horizon.
- Balance and weighted sample summariesDescribe which patient profiles receive most weight and how many independent people contribute.
- Measured baseline factors suffice to control confounding.How to check: Use clinical knowledge to identify common causes. Weighting cannot test whether important causes are missing.
- Both treatments occur in the target region.How to check: Inspect treatment support. Smooth weights do not supply outcomes where a treatment is impossible.
- The treatment contrast is well defined and relevant interference is absent.How to check: Align treatment versions, eligibility and time zero.
- The score model supports the intended target and contrast.How to check: An unpenalized logistic fit with an intercept balances included covariate means exactly when a finite solution exists.
- Multiple imputation may be justified under missing at random given recorded predictors.How to check: Refit scores within each imputed dataset and assess departures from the missingness assumption.
Plain language
How it works
A treated person receives one minus their propensity score, the estimated chance of treatment. An untreated person receives their score. Normalize weights within each arm, then subtract weighted outcome means. These bounded weights emphasize the region where treatment choice is least predictable. The target population has density proportional to the original density times the score times one minus the score. It differs from the full eligible population.
Li and colleagues derive the overlap estimator and its uncertainty. Zhou and colleagues examine limited overlap and model misspecification. Good balance is useful but does not establish causal validity. Use an estimator-specific sandwich variance, including score fitting, or a percentile bootstrap that refits the score and weighted contrast. Resample clusters when observations are clustered.
Worked example
A worked example
- Setting
- A four-record pain-score example supplies baseline treatment chances of 0.2 and 0.8 in each treatment group.
- Question
- What is the weighted difference in mean pain for the overlap target?
- Approach
- Treated pain scores are 2 and 8; untreated scores are 4 and 10. Use weights 0.8 and 0.2 for treated records, reversed for untreated records.
- Result
- Weighted mean pain is 3.2 for treatment and 8.8 for control. The treatment-minus-control difference is -5.6 points.
- What it means
- The tiny dataset teaches the weighting arithmetic, not a reliable clinical estimate. No confidence interval is inferred from these means.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Overlap weighting changes the target to patients with support for either treatment. It does not identify the full-population effect or remove unmeasured confounding.
- Calling the result an effect for all patients.Instead: Describe the overlap target and its weighted patient profile.
- Taking exact mean balance as proof of no bias.Instead: Check distributions and omitted confounders; exact balance covers only included model terms.
- Using the usual unweighted standard error.Instead: Use an overlap-weight estimator with score-fitting uncertainty, or refit the whole analysis in a bootstrap and report percentile bounds.
Inside Mallard
How Mallard uses this
Plans do not link to this method yet.
overlap-weights-ato).analysis.overlap_weighting, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Addressing Extreme Propensity Scores via the Overlap Weights (opens in a new tab) · Am J Epidemiol · doi 10.1093/aje/kwy201
- Propensity score weighting under limited overlap and model misspecification (opens in a new tab) · Stat Methods Med Res · doi 10.1177/0962280220940334