OBJECTIVES: To critically evaluate published cost-effectiveness (CE) models and identify model elements contributing to the large variability in results. METHODS: A literature search of MEDLINE and EMBASE from 1985-2004 identified eight English-language CE models comparing coxibs to a nonsteroidal anti-inflammatory drug (NSAID)-alone regimen. Two studies were excluded due to unavailable model input data. Model time horizons ranged from six months to lifetime, and primary outcomes ranged from gastrointestinal (GI) events averted to life-years gained and quality-adjusted life-years gained. Common elements across models were minor GI discomfort/dyspepsia, moderate GI events/symptomatic ulcer, and severe GI events. Only two of the analyses included cardiovascular side effects. To compare model inputs we standardized all analyses to a six-month tree structure with the three GI side effects. Study probabilities were converted to six-month rates where necessary and costs were converted to USD using the purchasing power parity index. Cost offsets between coxibs and NSAIDs were calculated by multiplying the probability of the GI event by cost per event. RESULTS: The relative price used for coxibs compared with NSAIDs differed widely across studies (median over six months, $156; range, $14-$387). Differences in total GI event cost offsets were small (median, -$41; range -$53 to -$18). Moderate GI events provided the greatest GI event cost offset in four of the six studies (median, -$21; range -$35 to -$0.50). In one case, the GI event offsets (due to the unusually high cost of treating minor GI) were greater than the additional cost of the coxib. CONCLUSIONS: Variation in drug acquisition cost of NSAIDs relative to coxibs was more important in contributing to the variation in results but the variations in clinical inputs and in costs of GI events were also important. More investigation into the reasons for differences in costs and clinical input is needed.