OBJECTIVES: Selection bias is common in observational studies.  When treatment selection is non-random , cases and controls frequently show imbalances in patient characteristics.  To reduce such imbalances in a retrospective case-control study, we used propensity score matching.  We present the derivation of propensity scores, selection of controls, and compare patient characteristics before and after matching. METHODS: We utilized the largest hospital service-level database in the U.S. We identified 2,588,722 adult patients undergoing inpatient echocardiography between January 2003 and October 2005 of which 2,900 had diagnoses for critical illness (heart failure, acute myocardial infarction, arrhythmia, respiratory failure, pulmonary embolism, emphysema , and pulmonary hypertension) and who also received a contrast agent (Perflutren Protein-Type A Microspheres Injectable Suspension, USP).  Patients receiving other contrast agents were excluded from the study.  To control for differences between patients receiving contrast echocardiography (cases) to those who received non-contrast echocardiography (controls), we used propensity score matching.  A stepwise logistic regression was used to model treatment choice (contrast vs. non-contrast). Variables used in the construction of the propensity score included comorbidities , demographic factors , hospital-specific factors , level of care, and mechanical ventilation status.  Cases were matched to 4 controls using the nearest neighbor matching algorithm based on differences in propensity scores among cases and controls. RESULTS: The nearest neighbor matching algorithm successfully identified 4 matches for each of the 2900 contrast patients.  Prior to matching, 23 out of the 26 patient characteristics showed statistically significant differences between cases and controls (P<0.01).  These characteristics included mechanical ventilation status, ICU status, and the Deyo-Charlson Comorbidity Score.  After matching, one variable remained statistically significant (higher concomitant medication usage among cases; P=0.006). CONCLUSIONS: The use of propensity score matching can reduce selection bias in a retrospective case-control study and, thus, create well-balanced groups of cases and controls for the analysis.