Background: Clinical research aiming to test treatments in humans should a priori specify outcomes related to both benefits and harms in protocol and final reports. However, how often assessment of harm is pre-specified in protocols and reported in final publication is not known. We performed this assessment in a cohort of randomized controlled trials (RCTs) conducted by Southwest Oncology Group (SWOG). Methods: We extracted data on elements addressing assessment of harms both from research protocols and publications for all consecutive phase III RCTs for the years 1960 to 2003. Chi-square statistics at p < 0.05 level was employed to assess statistical differences between protocols and publications. Results: We reviewed 117 RCTs of which matching protocols were available for 105 RCTs involving 129 comparisons (n=58,908 patients). Definition of treatment related harms was stated clearly in 80% (103/129) of protocols versus 45% (58/129) of publications only (p < 0.001). Standardized harms assessment instrument was denoted in 80% (103/129) of protocols versus 46% (59/129) of publications (p < 0.001). Seventy-three percent (94/129) of protocols clearly mentioned seriousness grades of harms compared to 10% (13/129) of publications (p=0.04). Severity of harms was clearly mentioned in 78% (100/129) of protocols versus 71% (92/129) of publications (p=0.30). Plan of analysis for harms was outlined in only 12% (15/129) of protocols versus 15% (19/129) of publications (p < 0.001). Statistical significance for harms was discussed in 5% (7/129) of protocols versus 42% (54/129) of publications (p < 0.001). Eight percent (10/129) of protocols reported sample size calculations for harms detection versus < 1% (1/129) of publications (p=0.11). Conclusion: Although the SWOG trials were well planned, the quality of harms reporting needs improvement. This, in turn, may contribute to better decision-making by clinicians and patients about treatment choices.