OBJECTIVES: In the absence of well-controlled clinical studies, medical records provide a potential wealth of information about the value of treatments; however, differences in pretreatment patient or other characteristics may influence treatment assignment and lead to biased estimates of treatment effects. Several strategies are available to reduce treatment selection bias. These include multivariable regression (MR) and propensity score (PS) techniques. Cepeda and colleagues (Am J Epidemiol 2003;158:280-7) demonstrated that PS is less biased than MR when the ratio of number of events to number of confounders (REC) is less than 8 by simulation. Using methods deemed more appropriate than Cepeda, we set out to evaluate conditions upon which their conclusions may be incorrect. METHODS: Monte Carlo simulation was performed in which each subject: 1) had 10 confounders (Zk:k=1,...,10) generated using normal and Bernoulli distributions; 2) was assigned to exposure or non-exposure with probability p determined by confounding variables; and 3) was given a binary response variable with probability g determined by confounder and exposure strength of association. For each simulation, binary logistic regression was used to: 1) generate individual PSs by regressing exposure variable on the confounder variables Z; and 2) estimate PS- and MR-adjusted treatment effects. Process was repeated 1000 times to evaluate bias and power of the statistical test. RESULTS: MR method produces asymptotically unbiased estimate of treatment effect; a result that is only marginally affected by the REC. Even when REC was 4.5, MR produced unbiased estimate of treatment effect with larger sample size. Contrary to the MR method, PS produces estimates that are consistently lower than the true effect regardless of sample size or REC. Power is always lower using the PS method. CONCLUSION: Results suggest PS method provides no statistical advantage over traditional MR; a conclusion that is contrary to Cepeda et al recommendations.