This paper considers the use of instrumental variables to estimate the mean effect of treatment on the treated. It reviews previous work on this topic by Heckman and Robb (1985, 1986) and demonstrates that (a) unless the effect of treatment is the same for everyone (conditional on observables),...