Estimation of antipsychotic effects on hospitalization risk in a naturalistic study with selection on unobservables.

Estimates of effects of antipsychotic medication on hospitalization risk based on nonexperimental data may be affected by selection bias from either observable or unobservable factors. This study applies a statistical method, using instrumental variables, that controls for both types of possible selection bias. We use data from a large observational study of people under treatment for schizophrenia to estimate models of drug choice and hospitalization, including atypical (versus typical) medication effects on 12-month hospitalization risk. Results for younger patients (<age 45 years) indicate that unobservable factors bias the atypical effect estimate in a positive direction; correcting for this bias yields a significant negative effect on hospitalization risk. With data for older patients, our instrumental variables performed poorly and provided little information about possible selection bias. Obtaining detailed information on treatment history and other determinants of medication choice in future studies is critical for deriving more accurate estimates of medication effects from nonexperimental data.