Abstract Changing and often declining health among elderly individuals makes interpreting the long-standing association between self-reported health (SRH) and mortality potentially problematic. This analysis of the Longitudinal Study of Aging from 1984 through 1986 explores changes over time in the association between a single self-report of health and survival among 4380 noninstitutionalized individuals aged 70 and older. Health was reported as excellent or very good (excellent/very good), good, fair or poor. The association between SRH and survival was modeled controlling for age, race, education, marital status, body mass index, difficulty performing activities of daily living, social contacts, self-reported stroke, heart disease, cancer, diabetes and recent hospitalization. Among women, SRH and survival were associated in a nonproportional model, with relative hazard declining over time. Women with poor vs excellent/very good SRH had adjusted relative hazards at 5, 14, 23 and 32 months of 3.8 [95% confidence interval (CI) 2.0–7.1], 2.7 (95% CI: 1.8–4.1), 2.0 (95% CI: 1.3–3.0), and 1.4 (95% CI: 0.7–2.7). The declining relative hazard was due to an increasing mortality rate over time among women initially reporting excellent/very good health. SRH was associated with survival among men in a proportional model (constant relative hazard over time). Men with poor vs excellent/very good SRH had an adjusted relative hazard of 1.7 (95% CI: 1.1–2.6) over the entire follow-up. The relative hazard of lesser magnitude among men reflects the weaker SRH-survival association, possibly too weak for any interaction with time to be detected. The constant relative hazard is also consistent with a rapid decline in health before death among men. The diminishing SRH-survival association among women is consistent with their longer period of declining health before death.