Dear Profs and Colleagues,
I am working on a 2SLS estimation using a Card (2001)-style shift-share instrument. For simplicity, I won’t go into the full model details, as my question is more general and conceptual.
In my setup, a referee raised a concern that the Kleibergen-Paap LM and Wald statistics in my results are unusually large — often exceeding 1000 — and suggested this might indicate a problem with instrument validity.
However, my sample contains over 1.5 million observations, and both KP statistics are highly significant (p < 0.01). My understanding is that these values indicate strong instrument relevance, and that large KP statistics are not necessarily problematic in large samples.
I would appreciate clarification on the following:
- Is it problematic in IV settings to observe very large Kleibergen-Paap statistics (e.g., >1000) in the presence of a large sample?
- Are there any guidelines or references in the literature on interpreting such high values?
- Are there published examples where similarly large KP statistics are reported in large datasets?
Cheers,
Paris
