- Published
- Updated
Applying a propensity score method does not guarantee that every person has a comparable counterpart or sufficient comparison information. Lack of overlap in important characteristics can change which people an estimated effect can meaningfully describe.
Define the question and assumptions
Define the target population and intended effect first. Examine score distributions, covariate balance and weights together. Extreme weights or unmatched people are not merely technical inconveniences: they may indicate regions where the data do not support the comparison. Justify exclusion or trimming rules in relation to the scientific purpose and resulting population.
Hypothetical study example
In a hypothetical dataset, severely affected people almost always receive one intervention. Matching leaves most of them outside the analysis. The team does not generalise the result in the retained sample to the entire original population; it reports the changed target and the profiles that lack adequate representation.
Design review checks
- Display score distributions and nonoverlapping regions separately for both groups.
- Inspect large weights and concentration of influence on the resulting estimate.
- Describe target-population changes after matching, trimming or other selection steps.
Record the decision and its limits
Accepted reporting includes balance and overlap evidence alongside the method name. Balance on measured covariates does not establish balance on unmeasured confounders. Do not select analytical rules after seeing which result they produce. Discuss remaining data support, the intended effect and generalisability together with the numerical uncertainty of the estimate.
Frequently asked questions
- Does a good score model establish a good causal analysis?
- Not alone. Overlap, measurement and confounding assumptions also matter.
- Are unmatched people unimportant?
- No. Their exclusion can change the population to which inference applies.