Instrumental variables for multilevel data.

[Nerea Gomez]

Good morning,

I am writing to ask you a question about the use of instrumental variables if my data has a hierarchical structure. Specifically, I am working with educational data in which the selected students have variable individual characteristics, but also common characteristics that have to do with the school and the teachers who teach them, and therefore there is a multilevel structure.

I want to use instrumental variables to see how an independent variable that has to do with the teaching staff affects the academic performance of the students. My question is which command should I apply, if there is any special command for multilevel models or if I should use the general IV making use of OLS.

Thanks in advance.



Translated with www.DeepL.com/Translator (free version)

[Phil Bromiley]

You didn't get a quick answer. You will increase your chances of useful answer by following the FAQ on asking questions – provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

If you go to the Stata window, and click on help, then select PDF documentation, you will see the range of documentation and procedures that are provided with Stata. There is an entire manual multilevel analysis. You need to read it first before proceeding. I

[Jeff Wooldridge]

Nerea: Doing true IV in Stata accounting for a nested structure won't be easy. You'd have to program it, I think.

I have two suggestions.

1. Ignore the heirarchy in estimation and just use usual 2SLS: ivregress 2sls. But then cluster your standard errors at the school level.

2. Use a control function approach where you can then use HLM in the second stage.

The equation is

y = a + b*w + x*c + u

and w is endogenous.

(i) Regress w on x and z and obtain the residuals, vhat.

(ii) Add vhat to the above equation and estimate a standard HLM. w and x appear as usual. A t-test on vhat tests the null that w is exogenous.

(iii) You'd have to bootstrap the two steps at the school level to get a proper standard error because vhat was estimated in a first stage.

JW

(i)

[Juan Felipe Contreras]

Hello,

In the procedure posted by Wooldrige, what would the commands be to boostrap the standard errores at the school level?

Kind regards,

Juan