OBJECTIVES: Ambulance clinicians use prealert calls to advise emergency departments (ED) of the arrival of patients requiring immediate review or intervention. Consistency of prealert practice is important in ensuring appropriate ED response to prealert calls. We used routine data to describe prealert practice and explore factors affecting variation in practice. DESIGN AND SETTING: We undertook a retrospective observational study in three UK ambulance services using a linked dataset incorporating 12 months' ambulance patient records, ambulance clinician data and emergency call data. OUTCOME MEASURES: We used least absolute shrinkage and selection operator regression to identify candidate variables for multivariate logistic regression models to predict variation in prealert use, analysing clinician factors (role, experience, qualification, time of prealert during shift), patient factors (National Early Warning Score version 2, clinical working impression, age, sex) and hospital factors (receiving ED, ED handover delay status). RESULTS: From the dataset of 1 363 274 patients conveyed to ED, 142 795 (10.5%) were prealerted, of whom 42 362 (30%) were for conditions with clear prealert pathways (eg, sepsis, stroke, ST-elevation myocardial infarction, major trauma). Prealert rates varied across and within different ambulance services. Casemix (illness acuity score, clinical diagnostic impression) was the strongest predictor of prealert use, but male patient sex, clinician role, receiving hospital and hospital turnaround delay at receiving hospitals were also statistically significant predictors, after adjusting for casemix. There was no evidence that prealert rates are higher during the final hour of shift. CONCLUSIONS: Prealert decisions are influenced by factors other than illness acuity and clinical diagnostic impression alone. Variation in prealert practice suggests that procedures and processes for prealerting may lack clarity and improved prealert protocols may be required. Research is required to understand whether our findings are reproducible elsewhere and why non-clinical factors (eg, patient gender) may influence prealert practice.