Using robust standard errors or clustered standard errors in iv estimation (ivreg2)

[Luiz Morga]

Hello guys,

I have the following problem and I would appreciate any help:

So Im analyzing the relationship between the quality of care (dependent variable) and the price of care (treatment / independent variable) using stata. Both variables are on firm level. Moreover, I add regional data on county level such as income, number of physicians etc (300 counties in total) and state fixed effects (10 states in total).
To account for endogeneity, I perform a IV-estimation (using ivreg2) where I instrument the price.
The data is a cross sectional data, so every observation appears once (in total 8000 observations).

My question would be if I do need to cluster standard errors on county level or if robust standard errors are sufficient. Because it severly effects the f-statistic of the excluded instruments if I use clustered standard errors on county level (its then below the magical number of 10).
Since its a cross section and the dependent and treatment variable are on firm level, I would assume using robust standard errors should be the right thing?
I would appreciate any help!

Thanks in advance.

[Joro Kolev]

You do not say anything about your instrumental variable.

I would say that if your instrument is varying at county level only, you need to cluster at county level.

If your instrument is varying at firm level, you do not need to cluster.

[Luiz Morga]

Thanks for the answer, Mr. Kolev!
The instrument variable is at county level.
Unfortunately, when I cluster at the county level, I have an f-statistic less than 10, which indicates weak instruments. It is also difficult to draw alternative instruments from my data set.
Does the estimation suffer a lot when I do not cluster at the county level and instead use robust standard errors?

[Joro Kolev]

In my view the first stage threshold of 10 is a bit of a shenanigan in general, and it is definitely inappropriate when you are clustering.

The appropriate test when you are clustering is Montiel Olea & Pflueger (2013)'s test for weak instruments called an "effective first-stage F-statistic".

Check out for a Stata package that implements it, -weakivtest-, described in this article
https://www.stata-journal.com/articl...article=st0377

Thanks for the answer, Mr. Kolev!
The instrument variable is at county level.
Unfortunately, when I cluster at the county level, I have an f-statistic less than 10, which indicates weak instruments. It is also difficult to draw alternative instruments from my data set.
Does the estimation suffer a lot when I do not cluster at the county level and instead use robust standard errors?