PURPOSE: Instrumental variables (IV) methodology removes bias due to unobserved confounding by replacing in the analysis the treatment with another variable--the instrument--that is well correlated with the treatment and independent of confounders. Recently, physician drug preference, operationalized as the treatment prescribed to the previous patient of the same physician, was proposed as an instrument in database studies comparing two competing drugs. We assessed, in simulations, how the performance of the IV estimates depends on the strength of this instrument. METHODS: The 'physician preference' instrument correlates well with the treatment only if physician preferences affect the treatment received by a large fraction of patients. Yet, often there is a subgroup of patients whose treatment cannot be affected by physician's preferences. The larger this subgroup is, the weaker the instrument. We investigated the impact of weakening this instrument on the performance of IV estimates, by comparing risk difference estimates from the conventional and IV analyses in the presence of an unobserved confounder, for both continuous and binary outcomes. RESULTS: The IV estimates were uniformly less biased than the conventional estimates, but had higher variance. Accordingly, the bias-variance trade-off favors the IV estimates only when physician' preference is a strong instrument. Still, the coverage rate of the 95%CI for the IV estimates was very close to the nominal 95%, while it was consistently lower for the conventional estimates. CONCLUSION: Researchers should consider using the 'physician preference' instrument when comparing two competing drugs, but should be aware of the underlying assumptions. (c) 2009 John Wiley & Sons, Ltd.
PURPOSE: Instrumental variables (IV) methodology removes bias due to unobserved confounding by replacing in the analysis the treatment with another variable--the instrument--that is well correlated with the treatment and independent of confounders. Recently, physician drug preference, operationalized as the treatment prescribed to the previous patient of the same physician, was proposed as an instrument in database studies comparing two competing drugs. We assessed, in simulations, how the performance of the IV estimates depends on the strength of this instrument. METHODS: The 'physician preference' instrument correlates well with the treatment only if physician preferences affect the treatment received by a large fraction of patients. Yet, often there is a subgroup of patients whose treatment cannot be affected by physician's preferences. The larger this subgroup is, the weaker the instrument. We investigated the impact of weakening this instrument on the performance of IV estimates, by comparing risk difference estimates from the conventional and IV analyses in the presence of an unobserved confounder, for both continuous and binary outcomes. RESULTS: The IV estimates were uniformly less biased than the conventional estimates, but had higher variance. Accordingly, the bias-variance trade-off favors the IV estimates only when physician' preference is a strong instrument. Still, the coverage rate of the 95%CI for the IV estimates was very close to the nominal 95%, while it was consistently lower for the conventional estimates. CONCLUSION: Researchers should consider using the 'physician preference' instrument when comparing two competing drugs, but should be aware of the underlying assumptions. (c) 2009 John Wiley & Sons, Ltd.
Authors: Krista F Huybrechts; Tobias Gerhard; Jessica M Franklin; Raisa Levin; Stephen Crystal; Sebastian Schneeweiss Journal: Pharmacoepidemiol Drug Saf Date: 2014-03-24 Impact factor: 2.890