Abstract
Chenhall and Moers (European Accounting Review, this issue, pp. 173–195) provide an excellent overview of the econometrics of endogeneity. In response to their discussion I argue that researchers should be courageous enough to set aside endogeneity concerns when their research question is important. Theory does not admit a definite answer to the question whether endogeneity is present in a particular model and econometrics has few technical solutions to offer. Since we cannot be sure endogeneity exists, and if we were to be sure of its existence, there is little we can do about it, researchers are well advised to move on to more serious problems.
Acknowledgments
I am grateful for helpful suggestions from Margaret Abernethy, Jan Bouwens, Willem Buijink, Eddy Cardinaels, and especially, Valeri Nikolaev.
Notes
1On the other hand, exploring how ‘clusters of choices’ respond to exogenous variables (e.g. environmental changes) is a line of inquiry that deserves closer attention.
2Clearly, robustness of the results against different operational choices (e.g. when using two different proxies for the same variable in the structural model) is something to aim for. Here I mean to suggest that the nature of sensitivity analysis across structural models is distinctly unlike that of the more common robustness checks, as in this case the aim is to find plausible models that produce significantly different results.
3The literature suggests statistical procedures to take specification uncertainty into account (Bartels, Citation1997). Intuitively, the results from a variety of different model specifications are averaged while discounting the data-mining effect of estimating many different specifications, especially if the results tend to change significantly with the specification.
4If there are no valid instruments available, sensitivity tests can also be helpful. For illustrative purposes, consider the following simple case. Suppose we have two variables, x and y, which are interrelated: y = α x + ϵ and x = β y + ν. In the absence of a valid instrument variable, we cannot estimate α, but if the researcher has a good idea about the range of reasonable values of β (and if cov(ϵ, ν) = 0), it is possible to explore how α behaves at different values within this range. By using ν = x − β y as an instrument for x, α can be identified. If α does not change much or retains its sign at different ‘reasonable’ values of β, then we can be much more comfortable about the results.