A fundamental identification problem in program evaluation arises when idiosyncratic gains from participation and the treatment decision depend on each other. Imbens and Angrist (1994) were the first to exploit a monotonicity condition in order to identify an average treatment effect parameter using instrumental variables. More recently, Heckman and Vytlacil (1999) suggested estimation of a variety of treatment effect parameters using a local version of their approach. However, identification hinges on the same monotonicity assumption that is fundamentally untestable. We investigate the sensitivity of respective estimates to reasonable departures from monotonicity that are likely to be encountered in practice and relate it to properties of a structural parameter. One of our results is that the bias vanishes under a testable linearity condition. Our findings are illustrated in a Monte Carlo analysis.