Propensity score in observational studies and its application for biostatistics

In observational and/or nonrandomized studies, investigators have no control on the randomallocation of subjects to treatment and control groups. Because of this reason, somedifferences may be observed on the covariates and the demographic characteristics in patientswho belong to case- control groups, therefore bias in the estimates might have accurred bythese differences.Propensity score, which can be determined as a balancing score, is the conditional probabilityof the observed covariates of case relative to control groups. Propensity score is usedprimarily to reduce bias in retrospective studies, increase the precision of the estimates as inthe prospective studies and determine the effects of some covariates. Propensity score iscalculated by using logistic regression analysis. Once the propensity score is estimated, it isaimed to reduce sampling bias by means of matching, stratification (quantiles), regressionadjustment or combination of three methods.This study included 478 patients who were operated by a single thoracic surgeon in MarmaraUniversity Hospital between 1996 and 2003. With regard to 10 preoperative risk factors and14 postoperative risk factors based on clinical data, logistic regression was performedregarding 18 patients with postoperative delirium (POD) after thoracic surgery and 460patients without POD. Then propensity score was calculated and patients who had similarpropensity score in case and control groups were matched by using stratification methods,therefore some patients from case and control groups were excluded to form the new sample.After this new sampling, case-control groups had similar characteristics in risk factors andbias in retrospective studies was reduced.As a result, the significance in risk factors, when logistic regression was used for the sampleof case-control groups? patients before and after adjusting the propensity score, was observedto be different.Keywords: Propensity score, observational studies, reduction bias