Physician's Prescribing Preference as an Instrumental Variable: Exploring Assumptions Using Survey Data.

BACKGROUND Physician's prescribing preference is increasingly used as an instrumental variable in studies of therapeutic effects. However, differences in prescribing patterns among physicians may reflect differences in preferences or in case-mix. Furthermore, there is debate regarding the possible assumptions for point estimation using physician's preference as an instrument. METHODS A survey was sent to general practitioners (GPs) in The Netherlands, the United Kingdom, New Zealand, Ireland, Switzerland, and Germany, asking whether they would prescribe levothyroxine to eight fictitious patients with subclinical hypothyroidism. We investigated (1) whether variation in physician's preference was observable and to what extent it was explained by characteristics of GPs and their patient populations and (2) whether the data were compatible with deterministic and stochastic monotonicity assumptions. RESULTS Levothyroxine prescriptions varied substantially among the 526 responding GPs. Between-GP variance in levothyroxine prescriptions (logit scale) was 9.9 (95% confidence interval: 8.0, 12) in the initial mixed effects logistic model, 8.3 (6.7, 10) after adding a fixed effect for country and 8.2 (6.6, 10) after adding GP characteristics. The occurring prescription patterns falsified the deterministic monotonicity assumption. All cases in all countries were more likely to receive levothyroxine if a different case of the same GP received levothyroxine, which is compatible with the stochastic monotonicity assumption. The data were incompatible with this assumption for a different definition of the instrument. CONCLUSIONS Our study supports the existence of physician's preference as a determinant in treatment decisions. Deterministic monotonicity will generally not be plausible for physician's preference as an instrument. Depending on the definition of the instrument, stochastic monotonicity may be plausible.