Levels of clustered standard errors

[Chris McDonald]

Hi All,

I have pooled data with 2500 funds, 2200 firms, and 20 time periods. Each fund could hold shares at many firms and in multiple time periods. Also, each firm can attract many funds in multiple time periods.

I am thinking of using the REGHDFE command to run OLS regressions with multiple fixed effects including fund FEs, firm FEs, and time FEs. In terms of clustering, there are several cluster levels, ranging from the interaction fund_id#firm_id, firm_id, fund_id, 2-way clustering (firm_id fund_id), 2-way clustering (firm_id time), 2-way clustering (fund_id time), and 3-way clustering (firm_id fund_id time).

I am not sure which level of clusters is appropriate after controlling for the above three fixed effects. May I have your views on which cluster levels I should go for in this case?

reghdfe LHS RHS, absorb(fund_id firm_id time) vce(cluster ???)

Many thanks,
Chris

[Fei Wang]

For dyadic panel data, the 3-way clustering would be appropriate. BTW, one-way FE would be insufficient. Need to include fund-firm paired dummies (fund_id#firm_id in -absorb()-).

[Chris McDonald]

Hi @Fei Wang,

Thanks for your suggestions. Do you think the small number of time clusters, relative to the number of funds and firms, could be a problem when doing the 3-way clustering?

Kind regards,
Chris

[Fei Wang]

Yes, Chris, a small number of clusters may result in inaccurate inference. In a classic paper (http://cameron.econ.ucdavis.edu/rese...13sept2006.pdf), Table 1 shows simulation results with different numbers of clusters. 20 clusters seem ok particularly when critical values for testing are adjusted (I think Stata command will do this).

[Chris McDonald]

Hi @Fei Wang,

Thanks for the helpful advice.