Objectives: To investigate the determinants of patient withdrawal from our study, and the effect of these withdrawals on the outcome of treatment with inhaled corticosteroids in patients with COPD. Design: A double-blind, placebo-controlled, randomized trial. Setting: Eighteen outpatient centers in the United Kingdom. Participants: Seven hundred fifty-one patients with stable COPD defined clinically and as baseline postbronchodilator FEV1 > 0.8 L and < 85% predicted, FEV1/FVC ratio < 70%, and FEV1 change after albuterol < 10% of predicted. Intervention: Random assignment of either 500g bid of inhaled fluticasone propionate (FP) using a spacer device or an identical placebo inhaler. Treatment was continued for 3 years or until patients withdrew from follow-up. Measurements and results: Postbronchodilator FEV1 was measured on three occasions before randomization and every 3 months thereafter. Health status was assessed by the disease-specific St. George Respiratory Questionnaire (SGRQ) and the modified short-form 36 questionnaire (SF-36) at baseline and every 6 months. Three hundred thirty-nine patients withdrew, of whom 156 patients received FP. Prescription of frequent courses of oral prednisolone was the most common reason for withdrawing as specified in the protocol (69 patients in the FP group withdrew due to respiratory symptoms, compared with 93 patients in the placebo group). This explained the significantly greater dropout of placebo-treated patients that was most evident when FEV1 was < 50% predicted. Patients withdrawing had a significantly more rapid decline in health status, measured by both the SGRQ and the SF-36 (p < 0.001). Those withdrawing from the placebo group had a more rapid decline in FEV1 and more exacerbations than the FP-treated groups. Baseline FEV1 was lower in dropouts than in patients completing the study receiving placebo, but there was no difference between the respective groups receiving FP. Conclusions: Patients who withdrew from follow-up were those with the most rapidly deteriorating health status and lung function. Losing these patients from the final analysis can reduce the power of a study to achieve its primary end point.