Evaluation of different approaches for confounding in nonrandomised observational data: a case-study of antipsychotics treatment.

Although randomised controlled trials are regarded as the gold standard for treatments efficacy, evidence from observational studies remains relevant. To address the problem of possible confounding in these studies, investigators must employ analysis methods that adjust for confounders and lead to an unbiased estimation of the treatment effect. In this paper, the authors describe two relevant statistical methods. The first method represents the classical approach consisting of a multiple regression model including the effects of treatment and covariates. This approach considers the relation between prognostic factors and the outcome variable as a relevant criterion for adjustment. The second method is based on the propensity score, and focuses on the relation between prognostic factors and treatment assignment. These approaches were applied to a cohort of 183 French schizophrenic patients who were followed for a 2-year period (from 1998 to 2000). The probability of relapse according to antipsychotic treatment exposure was modelled using Cox regression models with the two statistical methods. Goodness-of-fit criteria were used to compare the modelling approaches. This study demonstrates that the propensity score, a predicted probability, has an important balancing property that underscores its value in strengthening the results of nonrandomised observational studies.