Objective: To investigate the association between estimated GFR (eGFR) and all-cause mortality, including the contribution of temporal eGFR changes, in well-characterised community-based patients with type 2 diabetes. Design: Longitudinal observational study. Methods: Participants from the Fremantle Diabetes Study Phase 1 were assessed between 1993 and 1996 and followed until end-December 2012. Cox proportional hazards modelling was used to assess the relationship between baseline eGFR category (Stage 1-5) and all-cause death, and between eGFR trajectories assigned by semiparametric group-based modelling (GBM) and all-cause death in patients with five post-baseline annual eGFR measurements. Results: In the full cohort (1296 patients; mean ± s.d. age 64.1 ± 11.3 years, 48.6% males), 738 (56.9%) died during 12.9 ± 6.1 years of follow-up. There was a U-shaped relationship between all-cause death and eGFR category. With Stage 3 (45-59 mL/min/1.73 m 2 ) as reference, the strongest association was for eGFR ≥90 mL/min/1.73 m 2 (hazard ratio (95% CI) 2.01 (1.52-2.66); P < 0.001). GBM identified four linear trajectories ('low', 'medium', 'high', 'high/declining') in 532 patients with serial eGFR measurements. With medium trajectory as reference, eGFR trajectory displaced baseline eGFR category as an independent predictor of death, with low and high/declining trajectories associated with more than double the risk (2.03 (1.30-3.18) and 2.24 (1.31-3.83) respectively, P ≤ 0.003) and associated median reductions in survival of 6.5 and 8.7 years respectively. Conclusion: There is a nonlinear relationship between eGFR and death in type 2 diabetes, which is at least partially explained by a sub-group of patients with an initially high but then rapidly declining eGFR.