A retrospective cohort study of mortality rates and potential predictors of death was conducted in public-sector patients initiating highly active antiretroviral therapy (HAART) between October 2005 and September 2007 in a rural, under-resourced region of South Africa. The aims were to determine the relative contribution of death to cohort exit and the causes and predictors of mortality among HAART initiators. A community outreach programme traced non-attenders. Patients categorised as dying at home underwent a verbal autopsy (by interviewing family members) and case-file review, and those dying in hospital a case-file review, to determine the probable cause of death. At 24 months 1131 (83.6%) patients were retained on treatment in the programme, 124 (9.2%) had died, 63 (4.7%) had transferred out, and 35 (2.6%) were lost to follow-up. The most common causes of death were tuberculosis (44.3%) and diarrhoeal diseases (24.5%). Death was the major reason for cohort exit. As experience is gained with rural HAART programmes mortality rates may decrease. These results draw attention to the need for early HIV diagnosis, increased access to HAART services with earlier treatment initiation, and routine screening and aggressive management of opportunistic infections, particularly tuberculosis.