Matching on the propensity score allows to estimate treatment effects using randomized inference when randomized experimentation is not feasible, e.g., when we have observational data. A fundamental assumption for the obtained estimates to be unbiased is the ignorable treatment assignment assumption. This assumption requires that individuals are selected into treatment or control groups based only on observable covariates. However, this is often not the case, and in real applications we often encounter selection on unobservables that generates hidden bias. The objective of this work is to explore the consequences of having unobserved variables that affect treatment selection. We use simulated and real data to evaluate how different propensity score matching techniques perform when they are used in the estimation of a treatment effect in the presence of unobserved variables.