Clustering standard errors at industry level

[Anny Hung]

Dear Statalist members,

Currently my data size is about 12,000 observations.
The topic of my thesis is trade credit and peer effect,
I already xtset firm year, and include year and industry fixed effects.
My reference paper used standard error cluster by firm,
but when I cluster by firm then the result is not significant.
So I wonder if I can cluster by industry in this case?
And what kind of situation is suitable for use cluster by firm or industry?


Thanks in advance

[Jeff Wooldridge]

Anny: How many industries do you have? How many firms per industry? How many time periods?

[Anny Hung]

Anny: How many industries do you have? How many firms per industry? How many time periods?

there are 826 firms , 49 industries and 20 years.

[Joro Kolev]

You can at least try and and see what happens when you cluster at the industry level--the computer will not break, and you have enough clusters at industry level.

My view is that you should cluster at the highest level at which your main regressor and your errors are correlated. Professor Wooldridge is not quite agreed with that view though, and indeed this is not a settled matter in the literature.

You can read more on this thread here:
https://www.statalist.org/forums/for...tandard-errors

[Anny Hung]

You can at least try and and see what happens when you cluster at the industry level--the computer will not break, and you have enough clusters at industry level.

My view is that you should cluster at the highest level at which your main regressor and your errors are correlated. Professor Wooldridge is not quite agreed with that view though, and indeed this is not a settled matter in the literature.

You can read more on this thread here:
https://www.statalist.org/forums/for...tandard-errors

I tried cluster at industry level , the p-value is 0.061
"cluster at the highest level" is that means cluster by firm is better because firms observations are more than industries ?

[Joro Kolev]

No, I meant by "cluster at the highest level," cluster at the most aggregate level at which you suspect that there is still correlation between your errors and your major regressor. And you still should have enough clusters.

Also the previous literature is a guidance of what researchers consider reasonable. In your case if everybody is clustering at the firm level, this should be the benchmark.

Also your result--that standard errors are larger at the lower level of clustering (firm level), and smaller at the higher (more aggregate level, industry) of clustering is suspicious. You can try to bootstrap.

I tried cluster at industry level , the p-value is 0.061
"cluster at the highest level" is that means cluster by firm is better because firms observations are more than industries ?

[Anny Hung]

No, I meant by "cluster at the highest level," cluster at the most aggregate level at which you suspect that there is still correlation between your errors and your major regressor. And you still should have enough clusters.

Also the previous literature is a guidance of what researchers consider reasonable. In your case if everybody is clustering at the firm level, this should be the benchmark.

Also your result--that standard errors are larger at the lower level of clustering (firm level), and smaller at the higher (more aggregate level, industry) of clustering is suspicious. You can try to bootstrap.

thanks for your reply,
I understand what you mean.
"that standard errors are larger at the lower level of clustering (firm level), and smaller at the higher (more aggregate level, industry) of clustering is suspicious" is it possible that I used xtivreg2 to solve endogenous problems?

this is my code:
xi:xtivreg2 AR (peer_AR = L.ir_peer) L.firm_shock L.peer_ppe L.peer_cash L.peer_debt L.peer_sale L.peer_ROA L.peer_size L.debt L.ppe L.size L.cash L.ROA L.sale i.ind_code i.year ,fe robust cluster(company_code)