Use of observational databases to evaluate the effectiveness of antiretroviral therapy for HIV infection: Comparison of cohort studies with randomized trials

Objectives: It is important to assess the extent of bias when comparing the clinical efficacy of antiretroviral regimens in observational databases because, with the current lack of planned large trials, such analyses may represent the only means of assessing the risk of serious clinical events associated with new regimens. We aimed to compare the results from observational databases with those from randomized trials. Methods: Three treatment comparisons from randomized trials [Delta, AIDS Clinical Trials Group (ACTG) 175, Community Programs for Clinical Research on AIDS (CPCRA) 007 and ACTG 320] were mimicked in cohorts: (i) zidovudine monotherapy versus combination regimens of two nucleoside analogues; (ii) zidovudine combined with either didanosine or zalcitabine; and (iii) a dual combination versus a triple regimen including a protease inhibitor. Data for over 10,000 patients from the French Hospital Database on HIV, the EuroSIDA study and the Swiss HIV cohort study were analysed for each of the comparisons. Progression to AIDS disease or death was analysed in Cox models, adjusting for baseline differences, and results compared, with randomized trials. Results: For comparison (i) the adjusted relative risk estimates from cohorts were between 0.61 and 0.84, favouring combinations over monotherapy, compared with 0.57 to 0.63 for trials. For comparison (ii) relative risk estimates from cohorts ranged from 0.81 to 1.01 compared with 0.77 to 0.92 for trials. For comparison (iii), two of the cohorts showed similar results to the ACTG 320 trial but one indicated a higher risk of progression on triple therapy [relative risk 1.20, 95% confidence interval (CI) 1.01-1.44], in direct contrast to the trial result (relative risk 0.50, 95% CI 0.33-0.76). Conclusion: Serious biases can be present when comparing outcomes from the use of antiretroviral regimens in observational studies. However, such bias is not inevitable and careful interpretation of the results from several observational studies considered together is likely to be informative, guiding the design of new trials.