Propensity-score matching is increasingly being used to reduce the impact of treatment-selection bias when estimating causal treatment effects using observational data. Matching on the propensity score creates sets of treated and untreated subjects who have a similar distribution of baseline covariates. Propensity-score matching frequently relies upon calipers, such that matched treated and untreated subjects must have propensity scores that lie within a specified caliper distance of each other. We define the ‘marginally matched’ subject as a subject who would be matched using the specified caliper width, but who would not have been matched had calipers with a narrower width been employed. Using patients hospitalized with an acute myocardial infarction (or heart attack) and with exposure to a statin prescription at discharge, we demonstrate that the inclusion of marginally matched subjects can have both a quantitative and qualitative impact upon the estimated treatment effect. Furthermore, marginally matched treated subjects can differ from marginally matched untreated subjects to a substantially greater degree than the differences between non-marginally matched treated and untreated subjects in the propensity-score matched sample. The concept of the marginally matched subject can be used as a sensitivity analysis to examine the impact of the matching method on the estimates of treatment effectiveness. Copyright © 2009 John Wiley & Sons, Ltd.