What the evidence says
Propensity score methods can only compare units that have a realistic chance of receiving either treatment. Without adequate overlap, "causal comparisons cannot be made reliably because certain subgroups are observed under only one treatment condition" (Ahn 2026). Ahn also states that validity and generalizability of propensity-score estimates depend critically on sufficient overlap, and that without it the causal interpretation is unreliable "regardless of the apparent success of the statistical adjustment." The sources give no required amount of overlap.
The usual remedy is to drop units, and dropping units changes the question. In matching, unmatched individuals are excluded. The estimate then applies only to a restricted region of the covariate space, not the whole study population (Ahn 2026). Yang (2017) makes the same point for propensity-score trimming: dropping subjects with scores near zero or one "generally alters the estimand by changing the reference population." Researchers who trim first and then describe the target population from the surviving sample define the effect after the fact, and "using different samples may change the target estimand." Keele (2022), in an abstract-only claim, says steps to achieve overlap can yield very narrowly defined effects with little policy relevance.
Trimming is also a response to a variance problem. Extreme propensity scores produce large weights, which "can result in a large variance and poor finite sample properties" (Yang 2017). Khan (2024) says limited overlap is common in observational data, especially in high dimensions, and leads to extreme scores and high-variance estimates. D'Amour (2017) shows theoretically why high dimensions make this worse. Strict overlap bounds on covariate-mean imbalance grow more restrictive as dimension grows. The average unique discriminating information per covariate converges to zero. In the large-p limit, if the Bayes optimal classifier is consistent, the expected share of the sample that must be discarded to reach any fixed overlap bound approaches one. D'Amour concludes that trimming may need to be used more often when covariate dimension is large.
The evidence on how to trim is narrower. Yang proposes smooth weighting as an approximation to hard trimming. It is asymptotically linear, supports bootstrap confidence intervals, and shows slightly smaller variance than indicator weighting in simulations. Khan finds that trimming on propensity scores alone reduced variance little in an NHANES example, whereas trimming on both propensities and conditional variances reduced variance by 17%. Yang reports that for effects on the treated, the optimal support restriction trims scores close to one. Yang's simulations also show all estimators biased when the propensity model is misspecified.
Evidence is thin on real-world applications. Most variance results come from simulations or a single NHANES example. The sources do not measure how much of a sample is typically lost to trimming in applied studies.