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), or (b) treatment effects are variable across persons but the person-specific component of the variability not forecastable by observables does not determine participation in the program, widely-used instrumental variable methods produce inconsistent estimators of the parameter of interest. Neither assumption is very palatable. The first assumes a homogeneity that is implausible. The second assumes either very rich data available to the econometrician or that the persons being studied either do not have better information than the econometrician or that they do not use it. Instrumental variable methods do not provide a general solution to the evaluation problem.
