Estimating systems of equations with different instruments for different equations

For applications where different instruments are required for different equations, this paper offers a relatively primitive set of assumptions under which the nonlinear three-stage least squares estimator is the asymptotically efficient generalized method of moments estimator. These conditions can be verified for a general class of panel data models under reasonable assumptions about the dynamics in the model. On the other hand, the 3SLS estimator is necessarily inefficient for models such as hedonic price systems and for certain simultaneous equations models with measurement error or omitted variables in some equations.