Depression is a frequent health problem in old age, although the detection of such cases in clinical practice is often difficult. The systematic use of depression rating scales may increase diagnostic rates of depression amongst the elderly. This study aimed to assess the test-retest reliability of short versions of the Geriatric Depression Scale (GDS) with 1, 4, 10, and 15 items. Sixty-four consecutive patients aged 60 or over attending the outpatient clinic for the elderly (UNID) at the Department of Mental Health of Santa Casa of São Paulo were recruited for the study between February and May 1998. All subjects fulfilled criteria for the diagnosis of a depressive disorder (current or in remission) according to ICD-10, and had Mini Mental State scores greater than 10. They were evaluated twice in 48 to 72 hours with the GDS-15. Fifty-one patients completed both assessments. Agreement between scores for individual items was evaluated with Kappa statistic. Kappa coefficients ranged from 0.04 to 0.49, indicating that there was much variation within individual items. Total GDS-15 scores were reasonably stable, as assessed by paired Wilcoxon (z = 1.60, p = 0.109), Spearman correlation coefficient (rho = 0.86, p < 0.001), and weighted Kappa (Kappa = 0.64). The same pattern was also observed for the total scores of the GDS-10 on the paired Wilcoxon (z = 0.85, p = 0.402), Spearman correlation coefficient (rho = 0.81, p < 0.001), and weighted Kappa (Kappa = 0.60). Total score for the GDS-4 showed significant changes from test to retest (z = 3.75, p < 0.001; rho = 0.56, p < 0.001; Kappa = 0.37). These results indicate that the short GDS versions with 1 and 4 items are unreliable for use in clinical practice. In contrast, the GDS with 10 and 15 items produced consistent results in the assessment of elderly patients when total scores were used as clinical guidelines.