Among adults eligible for antiretroviral therapy (ART) in Thyolo (rural Malawi) and Kibera (Nairobi, Kenya), this study (a) reports on retention and attrition during the preparation phase and after starting ART and (b) identifies risk factors associated with attrition. 'Retention' implies being alive and on follow-up, whilst 'attrition' implies loss to follow-up, death or stopping treatment (if on ART). There were 11,309 ART-eligible patients from Malawi and 3633 from Kenya, of whom 8421 (74%) and 2792 (77%), respectively, went through the preparation phase and started ART. In Malawi, 2649 patients (23%) were lost to attrition in the preparation phase and 2189 (26%) after starting ART. Similarly, in Kenya 546 patients (15%) were lost to attrition in the ART preparation phase and 647 (23%) while on ART. Overall programme attrition was 43% (4838/11,309) for Malawi and 33% (1193/3633) for Kenya. Restricting cohort evaluation to 'on ART' (as is usually done) underestimates overall programme attrition by 38% in Malawi and 36% in Kenya. Risk factors associated with attrition in the preparation phase included male sex, age <35 years, advanced HIV/AIDS disease and increasing malnutrition. Considerable attrition occurs during the preparation phase of ART, and programme evaluations confined to on-treatment analysis significantly underestimate attrition. This has important operational implications, which are discussed here.