BACKGROUND: Intimate partner violence (IPV) against women is pervasive throughout the world, with profound consequences for women's health. Research to understand the extent, causes and consequences of IPV relies on self-reported data on violence, and yet there is a paucity of research into the consistency with which women report lifetime IPV over time. METHODS: We use data from the control group of the cluster randomised trial and a follow-on longitudinal study in Tanzania to examine discrepancies in women's reported experience of lifetime physical IPV and sexual IPV over three time-points (T0, T29, T53 months). Among those reporting lifetime history of IPV at T0, we calculate the proportion who subsequently report no lifetime history at T29 and/or T53 ('discrepant' reporting). We use logistic regression to explore associations between discrepant reporting and respondent baseline characteristics, the nature of their IPV experiences at baseline, and situational factors at T53. RESULTS: Complete IPV data were available for 301 women. At T0, 154 (51%) women reported lifetime history of physical IPV, of whom 62% gave a discrepant 'never' report in a subsequent round. Among 93 (31%) with lifetime history of sexual IPV at T0, 73% provided a subsequent discrepant report. 73% of women reported lifetime physical IPV, and 55% lifetime sexual IPV in at least one survey round. For both IPV outcomes, women were less likely to provide discrepant reports if they had recent IPV at baseline, poor mental health (T53) and poor communication with partner (T53). For physical IPV only, reduced discrepant reporting was also associated with baseline household-level financial hardship and more severe or extensive experience of IPV. CONCLUSIONS: A large proportion of women provided discrepant reports over the course of the study. Prevalence estimates of lifetime IPV from one-off cross-sectional surveys are likely to be underestimates, biased towards more recent and severe cases. To improve the stability of IPV measures, researchers should explicitly clarify the meaning of reference periods such as 'ever', consider using shorter reference periods (e.g. past-year), and avoid filter questions that use positive reports of lifetime IPV as a gateway to asking about more recent experiences. TRIAL REGISTRATION: Maisha CRT01 registered at ClinicalTrials.gov #NCT02592252, registered retrospectively (13/08/2015).