Identification and classification of principal features for analyzing unwarranted clinical variation

Abstract

AbstractRationale, Aims, and ObjectiveUnwarranted clinical variation (UCV) is an undesirable aspect of a healthcare system, but analyzing for UCV can be difficult and time‐consuming. No analytic feature guidelines currently exist to aid researchers. We performed a systematic review of UCV literature to identify and classify the features researchers have identified as necessary for the analysis of UCV.MethodsThe literature search followed the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses. We looked for articles with the terms ‘medical practice variation’ and ‘unwarranted clinical variation’ from four databases: Medline, Web of Science, EMBASE and CINAHL. The search was performed on 24 March 2023. The articles selected were original research articles in the English language reporting on UCV analysis in adult populations. Most of the studies were retrospective cohort analyses. We excluded studies reporting geographic variation based on the Atlas of Variation or small‐area analysis methods. We used ASReview Lab software to assist in identifying articles for abstract review. We also conducted subsequent reference searches of the primary articles to retrieve additional articles.ResultsThe search yielded 499 articles, and we reviewed 46. We identified 28 principal analytic features utilized to analyze for unwarranted variation, categorised under patient‐related or local healthcare context factors. Within the patient‐related factors, we identified three subcategories: patient sociodemographics, clinical characteristics, and preferences, and classified 17 features into seven subcategories. In the local context category, 11 features are classified under two subcategories. Examples are provided on the usage of each feature for analysis.ConclusionTwenty‐eight analytic features have been identified, and a categorisation has been established showing the relationships between features. Identifying and classifying features provides guidelines for known confounders during analysis and reduces the steps required when performing UCV analysis; there is no longer a need for a UCV researcher to engage in time‐consuming feature engineering activities.