BACKGROUND: In nonrandomized studies examining treatment comparisons, propensity scores (PS) are frequently used to account for measured confounding. As part of this methodology, the PS distribution is typically reviewed, and patients with a PS in the areas of nonoverlap and/or with extreme PS values may be excluded with aim of improving validity; a practice commonly referred to as "trimming." In studies examining safety events, trimming can lead to an incomplete safety profile of the treated patients and reduce the generalizability of results. It is recommended that information on treated patients who were excluded from the analysis (i.e., trimmed) be presented to provide a more complete understanding of the study population (Sturmer et al. 2010).
OBJECTIVES: By performing a targeted literature review, we aim to quantify the PS methods used in nonrandomized safety cohort studies in recent years and determine if summary information is provided on the trimmed patients.
METHODS: Articles published during 2014 and 2015 in six leading pharmacoepidemiology journals were identified in PubMed using the keyword "propensity score." Articles were required to be nonrandomized cohort studies with the primary objective of examining treatment safety. Two independent reviewers examined each article to determine eligibility and classification.
RESULTS: Of the 59 articles identified in the PubMed search, the majority were methods articles, reviews, or editorials and were excluded. The 18 eligible articles used the following PS methods: 9 matching, 2 inverse probability weighting, 5 regression (covariate), and 2 stratification. A total of 12 studies trimmed patients from the overall sample. Researchers commonly reported information on the overall sample and the analysis sample. However, none of the 12 articles presented summary information on the trimmed patients.
CONCLUSIONS: When PS trimming is conducted in safety cohort studies, researchers often report information on the overall and analysis sample without presenting information on the excluded patients, thus missing an important piece of information when evaluating the study's generalizability and a treatment's safety profile.
