The availability of large observational data bases allow empirical scientists to consider estimating treatment effects without conducting costly and/or unethical experiments where the treatment would be randomized. The Neyman-Rubin model (potential outcome framework) and the associated matching estimators have become increasingly popular, because they allow for the non-parametric estimation of average treatment effects. Like parametric models (e.g., ANCOVA), matching estimators control for a set of covariates (pre-treatment characteristics) in order to estimate the effect of a non-randomized treatment. However, unlike regression models, the selection of the covariates to be used with matching estimators has attracted little attention in the literature. This talk discusses why, when using matching estimators, the set of covariates used has to be "minimal". A set of covariates is said minimal if it cannot be reduced without violating the assumptions of the Neyman-Rubin model. Moreover, sufficient conditions are given for the identifiability of a minimal set of covariates. In order to obtain such conditions we use graphical models to impose restrictions on the set of conditional independence statements holding for the random variables involved. Finally, data-driven methods for the selection of the covariates are discussed.