Katsiaryna Bykov1,2, Murray A Mittleman2, Robert J Glynn1,3, Sebastian Schneeweiss1,2, Joshua J Gagne1,2. 1. From the Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women's Hospital and Harvard Medical School, Boston, MA. 2. Department of Epidemiology, Harvard T.H. Chan School of Public Health, Boston, MA. 3. Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, MA.
Abstract
BACKGROUND: The case-crossover design may be useful for evaluating the clinical impact of drug-drug interactions in electronic healthcare data; however, experience with the design in this context is limited. METHODS: Using US healthcare claims data (1994-2013), we evaluated two examples of interacting drugs with prior evidence of harm: (1) cytochrome P450 (CYP)3A4-metabolized statins + clarithromycin or erythromycin and rhabdomyolysis; and (2) clopidogrel + fluoxetine or fluvoxamine and ischemic events. We conducted case-crossover analyses with (1) a three-parameter model with a product term and a six-parameter saturated model that distinguished initiation order of the two drugs; and (2) with or without active comparators. RESULTS: In the statin example, the three-parameter model produced estimates consistent with prior evidence with the active comparator (product term odds ratio [OR] = 2.05, 95% confidence interval [CI] = 1.00, 4.23) and without (OR = 1.99, 95% CI = 1.04, 3.81). In the clopidogrel example, this model produced results opposite of expectation (OR = 0.78, 95% = 0.68, 0.89), but closer to what was observed in prior studies when active comparator was used (OR = 1.03, 95% CI = 0.90, 1.19). The saturated model revealed heterogeneity of estimates across strata and considerable confounding; strata with concordant clopidogrel exposure likely produced the least biased estimates. CONCLUSION: The three-parameter model assumes a common drug-drug interaction effect, whereas the saturated model is useful for identifying potential effect heterogeneity or differential confounding across strata. Restriction to certain strata or use of an active comparator may be necessary in the presence of within-person confounding.
BACKGROUND: The case-crossover design may be useful for evaluating the clinical impact of drug-drug interactions in electronic healthcare data; however, experience with the design in this context is limited. METHODS: Using US healthcare claims data (1994-2013), we evaluated two examples of interacting drugs with prior evidence of harm: (1) cytochrome P450 (CYP)3A4-metabolized statins + clarithromycin or erythromycin and rhabdomyolysis; and (2) clopidogrel + fluoxetine or fluvoxamine and ischemic events. We conducted case-crossover analyses with (1) a three-parameter model with a product term and a six-parameter saturated model that distinguished initiation order of the two drugs; and (2) with or without active comparators. RESULTS: In the statin example, the three-parameter model produced estimates consistent with prior evidence with the active comparator (product term odds ratio [OR] = 2.05, 95% confidence interval [CI] = 1.00, 4.23) and without (OR = 1.99, 95% CI = 1.04, 3.81). In the clopidogrel example, this model produced results opposite of expectation (OR = 0.78, 95% = 0.68, 0.89), but closer to what was observed in prior studies when active comparator was used (OR = 1.03, 95% CI = 0.90, 1.19). The saturated model revealed heterogeneity of estimates across strata and considerable confounding; strata with concordant clopidogrel exposure likely produced the least biased estimates. CONCLUSION: The three-parameter model assumes a common drug-drug interaction effect, whereas the saturated model is useful for identifying potential effect heterogeneity or differential confounding across strata. Restriction to certain strata or use of an active comparator may be necessary in the presence of within-person confounding.