Explaining variation in antibiotic prescribing for common infections: a three-way variance decomposition using UK primary care data

BackgroundBenchmarking has been used to target clinically unwarranted variation in antimicrobial prescribing in UK primary care. However, variation in antibiotic prescribing between general practices may be partly explained by differences in case-mix. We aimed to quantify how much variation in antibiotic prescribing for common infections was attributable to differences in prescribing between practices after accounting for case mix. MethodsWe used the UK Clinical Practice Research Datalink (CPRD) Aurum database to identify GP consultations for 11 common infections. Three-way variance decomposition quantified the proportions of total variance attributable to patient case-mix, between practice variation, and residual unexplained variance using variables recorded in electronic medical records, across three models (no, minimal (age/sex), and full case mix adjustment). For lower respiratory tract infection (LRTI) and sore throat, external data to impute illness severity were used to estimate the potential effect of unmeasured infection severity. FindingsWe identified 3,820,806 consultations in 2019. There was clear variability in antibiotic prescribing across practices for most conditions. In fully adjusted models, between practice variation explained 5.8-32.6% of total variance, exceeding variation attributed to case-mix in 9 out of 11 infections. Compared to no adjustment, full case-mix adjustment reduced between-practice differences, lowering their contribution to total explained variation by more than 20% in 6 of 11 infections and by 10-20% in 4 others. Minimal age-sex adjustment had little impact, with changes below 5% in 8 of 11 infections. Imputing infection severity in addition to full case-mix adjustment further reduced contribution of between-practice variance to the total variance (by 25.9% for LRTI and 8.5% for sore throat). InterpretationDifferences in practice-level prescribing, beyond patient case-mix, call for targeted interventions and highlight the value of providing feedback at the practice level. Comprehensive case-mix adjustment, including imputed infection severity, improves the assessment of prescribing variation and supports fairer performance comparisons. FundingWellcome Trust; NIHR RESEARCH IN CONTEXTO_ST_ABSEvidence before this studyC_ST_ABSWe searched MEDLINE for articles published between 1 January 2005 and 31 July 2025, using a combination of key terms including "antibiotic" (or "antimicrobial" or "antibacterial"), "prescribing" (or "prescription" or "use" or "utilisation" or "utilization"), "primary health care" (or "primary care" or "general practice" or "general practitioner" or "GP"), and "United Kingdom" (or "UK" or "England"). We focused on studies using patient-level data to compare antibiotic prescribing between general practices (GP practices). Most studies assessed overall prescribing or focused on a small subset of infections. Only a few examined condition-specific measures across a broader range of infections. We found no studies that decomposed variance into that caused by patient case-mix versus practice performance adjusting for case-mix across a wide range of infections. Added value of this studyUsing individual-level data from 3.8 million consultations for eleven common infections in the UK Clinical Practice Research Datalink (CPRD) Aurum, we applied three-way variance decomposition to quantify the proportions of total variance attributable to patient case-mix, between-practice differences after adjusting for case-mix, and residual variation under three adjustment strategies (non, age-sex only, and full case-mix adjustment). There was clear variability in antibiotic prescribing across practices for most conditions. The total variance attributable to between-practice differences exceeded that attributed to case-mix in 9 out of 11 infections according, according to the fully adjusted case-mix models. Fully adjusting for case-mix based on routinely collected data substantially reduced between-practice differences, lowering their contribution to explained variance (the sum of the patient case-mix variance and between-practice variance) by more than 20% in 6 of 11 infections and by 10-20% in 4 others, whereas minimal age-sex adjustment had little impact. Between-practice differences were reduced further incorporating external information to simulate unmeasured infection severity. Implications of all the available evidenceDifferences in practice-level prescribing, beyond patient case-mix, call for targeted interventions and highlight the value of providing feedback at the practice level. Full case-mix adjustment substantially reduces the risk of overstating between-practice differences, performing far better than adjusting for age/sex alone. Condition-specific indicators with sufficient case-mix adjustment may be more effective benchmarks of practice performance than aggregated total antibiotic use levels as general practitioners (GPs) are more likely to respond positively to comparisons they perceive as fair. In particular, acute otitis media and upper respiratory tract infection, conditions with substantial variability in antibiotic prescribing across GP practices and the highest variance attributable to adjusted between-practice differences (12.6% and 10.3%, respectively), are promising candidates for fair prescribing indicators.