Remission of depressive symptoms is the goal of all antidepressant therapy. Rating scales define remission in clinical trials, but it is unclear how well these definitions predict risk of later relapse. We measured the sensitivity and specificity of a range of Montgomery-Asberg Depression Rating Scale (MADRS) cutoff scores at 3- and 6-months, wherein scores above a given cutoff would predict relapse over an 18-month period. We examined 153 elderly depressed subjects exhibiting a MADRS < or = 15 after 3 or 6 months of antidepressant therapy. Subjects who subsequently exhibited a MADRS > 15 during the 18-month study period were defined as relapsed. Receiver operating characteristic (ROC) curves were developed and area under the curve (AUC) values calculated for the sensitivity and specificity of 3- and 6-month MADRS scores to predict future relapse. The 3-month ROC had an AUC value of 0.63; the 6-month ROC had an AUC value of 0.66. There was no MADRS cutoff found that could predict likelihood of relapse with good sensitivity and specificity. A post hoc analysis where relapse rate was adjusted by controlling for medical comorbidity, disability, and social support showed no change in the ROCs or AUC values. The higher the MADRS score at 3 and 6 months, the greater the likelihood of relapse. With no clean MADRS cutoff score, the goal of antidepressant therapy should be the lowest possible degree of depressive symptomatology to minimize risk of later relapse. Definitions of remission that are better associated with longer-term outcomes are needed.