BACKGROUND AND PURPOSE In 2010, Greater Manchester and London centralized acute stroke care into ‘hyperacute’ units (Greater Manchester=3, London=8), with additional units providing ongoing specialist stroke care nearer patients’ homes. Greater Manchester patients presenting within four hours of symptom onset were eligible for hyperacute unit admission; all London patients were eligible. Research indicates that, post-centralization, only London’s stroke mortality fell significantly more than elsewhere in England. This paper attempts to explain this difference by analyzing how centralization affects provision of evidence-based clinical interventions. METHODS Controlled before and after analysis was conducted, using national audit data covering Greater Manchester, London, and a non-centralized urban comparator (38,623 adult stroke patients, April 2008-December 2012). Likelihood of receiving all interventions measured reliably in pre- and post-centralization audits (brain scan; stroke unit admission; receiving antiplatelet; physiotherapist, nutrition, and swallow assessments) was calculated, adjusting for age, sex, stroke-type, consciousness, and whether stroke occurred in-hospital. RESULTS Post-centralization, likelihood of receiving interventions increased in all areas. London patients were overall significantly more likely to receive interventions, e.g. brain scan within three hours: Greater Manchester=65.2%[95% Confidence Interval=64.3-66.2]; London=72.1%[71.4-72.8]; comparator=55.5%[54.8-56.3]. Hyperacute units were significantly more likely to provide interventions, but fewer Greater Manchester patients were admitted to these (Greater Manchester=39%; London=93%). Differences resulted from contrasting hyperacute unit referral criteria, and how reliably they were followed. CONCLUSIONS Centralized systems admitting all stroke patients to hyperacute units, as in London, are significantly more likely to provide evidence-based clinical interventions. This may help explain previous research showing better outcomes associated with fully-centralized models