Adherence to antiretroviral therapy (ART) is critical for successful treatment of HIV, but comparisons across settings are difficult because adherence is measured in different ways. We examined utility of different adherence measures for identification of patients at risk of viral failure (VF). Eight cohorts in the ART Cohort Collaboration contributed data from pharmacy refills or self-report questionnaires collected between 1996-2013 (N=11689). For pharmacy data (N=7156), we examined associations of percentage adherence during 1st year of ART with VF (>500 copies/ml) at 1 year. For self-report data (N=4533), we examined 28-day adherence with VF based on closest viral load measure within 6 months after questionnaire date. Because adherence differed markedly by measurement type we defined different cut-points for pharmacy (lower <45%, medium 45-99%, higher 100%) and self-report (lower ≤95%, medium 96-99%, higher 100%) data. Adjusted odds ratios (OR) for VF in lower and medium compared to higher adherence groups were 23.04 (95% CI:18.44-28.78) and 3.84 (3.36-4.39) for pharmacy data. For self-report data they were 3.19 (2.31-4.40) and 1.08 (0.80-1.46). Both types of measure were strongly associated with VF. Although adherence measurements over longer time-frames are preferable for prediction, they are less useful for intervention.