Our starting point for discussing instrumental variables is that we would like to be able to regress an outcome variable Y on explanatory variables X and U , with particular interest in the role of X in the conditional distribution of (Y |X ,U). That is, we would like to infer the X-Y association adjusted for U , and we might refer to U as a confounder, or at least a potential confounder. A serious challenge arises, however, in that while X and Y are observable, U is unobservable. The hope is that a different observable variable Z, having special properties, can save the day.