Aims: To extrapolate, from the proportion of subjects with observable retinopathy at diagnosis of type 2 diabetes mellitus in routine clinical practice, the mean duration of undiagnosed diabetes. Methods: On 1 October 1999, there were 4313 patients with type 2 diabetes in the 41 participating practices in the Tayside region (registered with one of 166 GPs). 501 (12%; 95% CI 11 to 13%) patients were selected using a pseudo-random number allocation algorithm, and practice lists checked for recently deceased, non-residents (45 exclusions). Retinopathy was graded by validated slit lamp biomicroscopy and four-field stereo photography. Date of first diagnosis of diabetes was ascertained from the regional diabetes register created using multiple source data capture. Results: Of living Tayside resident patients, 295 from 456 invited type 2 patients (65%) were examined. 14.68% (95% CI 12.48 to 16.88%) were found to have retinopathy at diagnosis. Assuming a linear model, these data suggest that the onset of detectable retinopathy occurs 5.77 years (95% CI 4.6 to 7 years) before diagnosis. Comparison using the log rank test with survival to onset of sight threatening retinopathy/maculopathy in 291 patients with type 1 diabetes mellitus also examined from the same population cohort showed the 95% CIs of length of preclinical diabetes to be between 3.0 and 9.4 years. Conclusion: There is accumulating evidence to question the assumption of linearity as a model of choice. The authors' understanding of a distinct glycaemic threshold for retinal change is also overly simplistic and consequently the bounds of uncertainty concerning the preclinical duration of disease are considerable.