Because the first COX-2 inhibitors (celecoxib) were released in January 1999, we used descriptive statistics, applying patient sampling weights, to examine trends in the use of COX-2 inhibitors (rofecoxib, celecoxib, and valdecoxib) and NSAIDs (ibuprofen, indomethacin, naproxen, diclofenac, etodolac, fluribiprofen, ketoprofen, ketorolac, meclofenamate, meloxicam, nabumetone, oxaprozin, piroxicam, sulindac, tolmetin) from 1999 through 2002. To examine predictors of COX-2 inhibitor use, we defined our outcome variable as the choice between a COX-2 inhibitor and an NSAID, conditional on having received 1 of these 2 medication classes. We excluded the 0.4% of patients with prescriptions for both an NSAID and a COX-2 inhibitor. We used χ2 analyses to evaluate the bivariate association between hypothesized characteristics of patients (age, sex, race, comorbid conditions, and GI risk score), physicians (specialty and employment status), and visits (source of payment, region of country, new vs established patient, year, type of office, owner of practice, and solo vs group practice) and the outcome of interest. We performed weighted logistic regression for complex survey data to examine the multivariate association between the predictor variables and our outcome variable.27 Initial models included basic demographic characteristics of patient visits, variables significant on bivariate analysis (P<.25), and an interaction term between year and GI risk score to assess whether the association between GI risk and the likelihood of COX-2 inhibitor receipt was independently modified by the year of observation. We also examined several other interaction terms that we hypothesized a priori might be important explanatory variables (eg, race interacted with source of payment, physician specialty interacted with time spent with physician, and physician specialty interacted with patient GI risk), but they did not add considerably to the model’s goodness of fit. Because most comorbid conditions were present in fewer than 5% of all patient visits, we aggregated these comorbid conditions into 2 categories—those that might be an indication for COX-2 inhibitors rather than NSAIDs (GI bleeding, peptic ulcer disease, rheumatoid arthritis, corticosteroid use, coagulation defects, history of heartburn, stomach pain, nausea, or vomiting) and those that might be a relative contraindication for either COX-2 inhibitors or NSAIDs (congestive heart failure, liver dysfunction, or renal dysfunction). We refined our multivariate model using the Hosmer-Lemeshow goodness-of-fit test and the Pregibon Linktest for Nonlinearity. Our final models were robust, maximized goodness of fit, and included basic sociodemographic variables (eg, race) and variables that were statistically significant in the earlier models (P<.05). We examined the impact of omitting alternative analgesics (opioids, opioid analogues, and acetaminophen) from our analyses. There were no statistically significant trends in the use of these medicines from 1999 through 2002, nor was the use of these analgesics associated with GI risk score. Thus, we reasoned these alternatives were not important substitutes for COX-2 inhibitors or NSAIDs in our analyses. We also conducted sensitivity analyses that examined multivariate models limited to the elements of the GI risk score and that dichotomized patients based on the presence of any of the main risk factors for adverse events from NSAIDs (eg, age >65 years). All analyses were conducted using Stata statistical software (Stata Corporation, College Station, Tex).