Nisha Gottfredson writes:

I am using the instrumental variables approach to model the effect of teacher compliance to prevention curriculum treatment on student substance use. Since there are 43 teachers, I need to account for nesting. Around pg. 511 in your textbook you discuss using the IV approach with multilevel data; however, you do not mention the problem of biased standard errors in this section as you did on pg. 222 when you discussed IV in the non-multilevel context.

My question (which I hope is simple enough) is this: Are the standard errors still biased in the multilevel context, requiring the adjustment you discussed in the earlier chapter, or are they correct?

My reply: I don’t know; Jennifer wrote that bit! I’ve actually never done an instrumental variables analysis myself (but I did learn a lot about the method while writing the book).

P.S. Jennifer writes:

The reason the standard errors had to be corrected in that original discussion was because the two equations weren’t being modeled simultaneously… we fit one model, took the results (predictions) from it and used those in the next model. The multilevel model version that we describe in the book models the two equations simultaneously thus avoiding this problem — s.e.’s should be correct.