OBJECTIVES: Without resistance tests, deciding which patients with virological failure should switch to second-line antiretroviral therapy (ART) is difficult. The factors influencing this decision are poorly understood. We assess predictors of switching regimens after virological failure. DESIGN: Retrospective cohort study using clinical data from a South African ART program with 6-monthly viral load (VL) monitoring. METHODS: We constructed a dataset of patient visits occurring following first-line virological failure, and used random effects logistic regression (accounting for individual-level and clinic-level clustering) to assess predictors of switching at each visit. RESULTS: One thousand six hundred sixty-eight patients with virological failure (73% male, mean age 41 years, median CD4 184 cells/mm, mean log10 VL 4.3) contributed 1922 person-years of viremia. 12 months after failure, the cumulative incidence of switching regimen, viral resuppression, or death was 16.9%, 13.2%, and 4.6%, respectively. In adjusted analysis, switching was more likely at the third or subsequent visit after failure; in visits occurring in 2008 versus 2003 to 2007; and in patients with ART experience pre-programme, current high VL or low CD4 count. Switching was less likely in patients with no clinic contact for 4 months, or declining VL. Switching rates varied between clinics with clinic-level clustering evident in the final model (P <0.001). CONCLUSIONS: Despite 6-monthly virological monitoring and recommendations to switch after adherence interventions and confirmed viremia, patients experienced delayed switching. Individual-level covariates influenced switching but did not account for variable switching rates between clinics, suggesting differences in guideline implementation. In certain circumstances delays may be warranted; however understanding barriers to guideline implementation will limit unnecessary delays.