The Impact of Nurse Turnover on Quality of Care and Mortality in Nursing Homes, Yaa Antwi and John Bowblis, in the American Journal of Health Economics, 2018, right here: https://www.journals.uchicago.edu/doi/10.1162/ajhe_a_00096
This uncomplicated article involves three types of regressions: a basic pooled model, one adding fixed effects, and another adding instruments. The issue involves closing the loop. The direction of possible bias is clear, and addition of fixed effects (which is expected to reduce bias) moves the coefficient estimate accordingly, but the use of IV moves the estimate very far in the other, unexpected direction. This unexpected result is unexplained.
In addition, the issue of experimental content arises, not in the main regression, but in the first-stage regression using the instrument, reinforcing the point made in article #5, that issues with experimental content can take many forms and arise in a variety of circumstances, even those designed to produce "good" estimates that can be given causal interpretations.