Evaluating the validity of an instrumental variable study of neuroleptics: can between-physician differences in prescribing patterns be used to estimate treatment effects?

BACKGROUND: Postmarketing studies of prescription drugs are challenging because prognostic variables that determine treatment choices are often unmeasured. In this setting, instrumental variable (IV) methods that exploit differences in prescribing patterns between physicians may be used to estimate treatment effects; however, IV methods require strong assumptions to yield consistent estimates. We sought to explore the validity of physician-level IV in a comparative study of short-term mortality risk among elderly users of conventional versus atypical antipsychotic medications (APM).

METHODS: We studied a cohort of patients initiating APMs in Pennsylvania who were eligible for Medicare and a state-funded pharmaceutical benefit plan. The IV was defined as the type of the APM prescription written by each physician before the index prescription. To evaluate whether the IV was related to other therapeutic decisions that could affect mortality, we explored the association between the instrument and 2 types of potentially hazardous coprescriptions: a tricyclic antidepressant (TCA) not recommended for use in the elderly or a long-acting benzodiazepine. To insure that the IV analysis was not biased by case-mix differences between physicians, we examined the associations between the observed patient characteristics and the IV.

RESULTS: The cohort consisted of 15,389 new users of APMs. Our multivariable model indicated that physicians who had most recently prescribed a conventional APM were not significantly more or less likely to coprescribe a potentially hazardous TCA [odds ratio (OR), 0.78; 95% confidence interval (CI), 0.58-1.02] but were less likely to prescribe a long-acting benzodiazepine (OR, 0.57; 95% CI, 0.45-0.72) with their current APM prescription. The association between long-acting benzodiazepine prescribing and APM preference was no longer significant when the analysis was restricted to primary care physicians (OR, 0.84; 95% CI, 0.62-1.15). Multivariable regression indicated that important medical comorbidities (eg, cancer, hypertension, stroke) were unrelated to the IV.

CONCLUSIONS: The previous APM prescription written by the physician was unassociated with major medical comorbidities in the current patient, suggesting that the IV estimates were not biased by case-mix differences between physicians. However, we did find that the IV was associated with the use of long-acting benzodiazepines. This association disappeared when the study was restricted to the patients treated by primary care physicians. Our study illustrates how internal validation approaches may be used to improve the design of quasi-experimental studies.