Are there ways to estimate the finite sample bias with instrumental variables? I guess this would be conditional on assuming some structure to the problem and also would involve simulation, but, at least in applied econometrics papers, I have never seen people do it.

PS: If possible, the answer could be illustrated with R code.


Using the model with one regressor and many instruments(vector-matrix notation) $$y = x\beta +u$$ $$x = Z\gamma +v$$

Bound, Jaeger & Baker (1995) give the following approximation to the 2SLS finite sample bias:

$$B(\hat \beta_{IV}) = \frac {\operatorname{Cov}(u,v)}{\gamma'Z'Z\gamma}(K-2)$$

where $K$ is the number of instrument (the column dimension of matrix Z). Naturally, one uses the estimated quantities to calculate the above. This as a special case of results in A.L.Nagar (1959).
Other relevant literature is A.Buse (1992) and Staiger and Stock (1997).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.