Year: 2020
Type: letter
Abstract: In the article “Addressing Extreme Propensity Scores via the Overlap Weights” by Li et al. (1), the asymmetric propensity score (PS) trimming of Stürmer et al. (2) was incorrectly implemented as a comparative approach in the simulation study. This incorrect implementation led to large bias for estimating the average treatment effect. In this erratum, we provide corrected simulation results for asymmetric PS trimming. Bias of the Asymmetric Trimming Estimators in the Presence of Increasingly Strong Tails in the Propensity Score Distributions (n = 2,000) Bias of the Asymmetric Trimming Estimators in the Presence of Increasingly Strong Tails in the Propensity Score Distributions (n = 2,000) The correct implementation of asymmetric trimming involves the following steps. First, one excludes patients with PS outside of the common PS range formed by the treated and control patients. Then, one further excludes all patients (treated and untreated) whose PS is below the |$q$| quantile of the treated units, as well as all patients (treated and untreated) whose PS is above the |$1-q$| quantile of the control units. The correct implementation of asymmetric trimming should always ensure a common range of the PS among the treated and untreated subjects and therefore is expected to provide approximately unbiased point estimates in the simulation design with a constant treatment effect. Using the same simulation seed as was used in our original article (1), we summarize the corrected results for asymmetric trimming in our Web Tables 2–4. These 3 tables can be read together with Tables 2–4 in Li et al. (1). Under lack of overlap, we confirm that asymmetric trimming leads to approximately unbiased point estimates with trimming thresholds |$q=0.01$| and |$0.05$|. The results for |$q=0$| are unchanged compared to those in Li et al. (1) but are included in Web Tables 2–4 for completeness. The current results correct the message in Li et al. (1) regarding asymmetric trimming and demonstrate the validity of asymmetric trimming under lack of overlap. On the other hand, the corrected simulation results for asymmetric trimming still confirm that the overlap weighting is the most statistically efficient among its competitors. The authors regret the incorrect implementation of asymmetric trimming in the original simulation study and thank Dr. Til Stürmer and his colleagues for personal communication that lead to this important clarification. Relative Efficiency (Relative to the Crude Estimator) of the Asymmetric Trimming Estimators in the Presence of Increasingly Strong Tails in the Propensity Score Distributions (n = 2,000) Relative Efficiency (Relative to the Crude Estimator) of the Asymmetric Trimming Estimators in the Presence of Increasingly Strong Tails in the Propensity Score Distributions (n = 2,000) The 95% Confidence Interval Coverage of the Asymmetric Trimming Estimators in the Presence of Increasingly Strong Tails in the Propensity Score Distributions (n = 2,000) The 95% Confidence Interval Coverage of the Asymmetric Trimming Estimators in the Presence of Increasingly Strong Tails in the Propensity Score Distributions (n = 2,000) Conflict of interest: none declared.
Source: American Journal of Epidemiology
DOI: 10.1093/aje/kwaa229
Authors: Fan Li, Laine E Thomas, Fan Li
Institutions: Duke University, Yale University
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4244621058