OBJECTIVES: Confounding is a common source of bias in outcome studies involving observational non-randomized data. The propensity scores methodology has been suggested as a good analytical approach for handling this problem without any indication on whether a threshold exits on its predictive ability. We investigate the usefulness of the C-statistic in this regard using simulated data.METHODS: In each simulation, we generated 100 sets of 10,000 patients; each patient being assigned probabilities of being treated and of experiencing the outcome of interest. The process involved two logistic models, one that related treatment to a set of 10 independent covariates and the other relating outcome to treatment and the same 10 covariates, using Bernoulli distributions that assumed an odd ratio (OR) for treatment between 0.14 to 1.00 for each dataset. Propensity scores from each dataset were estimated and propensity scores-matched analysis conducted using conditional logistic regression to estimate the OR and from the 100 sets, we obtained the mean, median and bias in the estimate. Bias was defined as the difference between actual and estimated ORs as a proportion of actual. RESULTS: We found evidence of correlation between the levels of bias in the OR estimates and the C-statistics, with level often exceeding 300% when the C-statistic was less than 80%.   CONCLUSIONS: Where as an elevated value of the C-Statistic may not guarantee effective correction of confounding by the resultant propensity scores derived from a given data, our study indicates that a lower value does indicate a poor capability. We suggest the C-Statistic can be adopted as a simple reporting tool on the propensity scores model in respect of its efficiency.