Controlling for systematic selection in retrospective analyses: an application to fluoxetine and sertraline prescribing in the United Kingdom.

BACKGROUND

Criticism has been made of observational studies in clinical practice because of their failure to control for unobserved factors that correlate with both initial treatment selection and observed outcomes.

METHOD

A two-stage statistical model was applied to data obtained from a large general practitioner medical records database (DIN-LINK) to estimate the effect of initial antidepressant selection on the duration of antidepressant therapy and on the likelihood of being prescribed an average daily dose above the minimum recommended dose. The statistical model controlled for unobserved factors correlated with initial treatment selection and the observed outcomes as well as for observed confounders.

RESULTS

Unobserved factors correlated with treatment selection were not a statistically significant determinant of the number of days of antidepressant therapy. However, unobserved factors correlated with treatment selection were a statistically significant determinant of the likelihood of receiving an average dose during therapy greater than the minimum recommended. After controlling for relevant confounders, those patients who began treatment with sertraline as opposed to fluoxetine had fewer days of antidepressant therapy and were more likely to receive average doses greater than the minimum recommended during therapy.

CONCLUSION

Unobserved factors correlated with treatment selection can impact outcomes in observational studies and should be tested and controlled for whenever possible.