Whether we must cluster by country with staggered laws implementation?

Question

When trying to find a way of avoiding using clustering, I saw that Abadie, 2017 have a great paper mentioned when we should cluster, summarized by McKenzie here.

I used the paper of Dasgupta,2019 to link to the summarized work of McKenzie. So, in Dasgupta's paper, he examines the impact of anticollusion laws of firms' asset growth in a standard Difference-in-Differences (DID) setting with multiple groups and periods. In specific, each country will pass the anticollusion laws in different years, and he examine the impact of such law implementation on firms' asset growth.

It seems to me that Dasgupta, 2019 does not need to cluster based on their setting. However, from this comment of @Björn , it seems that at least, Dasgupta need to cluster by countries.

As treatments are applied at the country level, sure clustering at least at the country level is obvious?

Part of the intuition is that you cannot tell apart a general country trend (pretty plausible) from an effect of the treatment. Instead, if within the country the treatment was applied at different times per different firms you could tell it apart (and a general country trend could to some extent be absorbed into the general noise). Or if regulation is applied to different industries at different times within a country, then clustering by sector within country could be an alternative

So, it leads to another story here that maybe good for a new question rather than answering in the comment part: So, we must, at least cluster by country in the case above? I do not fully understand the argument.

Answer 1

1. First of all, you actually do not need to use clustering specifically in that case. The point of clustering is to correct for heteroskedasticity or/and autocorrelation at some level. You have to correct for heteroskedasticity or/and autocorrelation somehow otherwise your test statistics (base on which you calculate p-values) will be wrong. But you do not need to necessarily use clustering for that.

   Clustering is not the only game in the town, you can correct for heteroskedasticity and/or autocorrelation in various ways. For example, in case of autocorrelation if you have long panel you can make your model dynamic and get rid of autocorrelation by specifying appropriate lag structure. Alternatively, one might use various bootstrapping techniques to estimate correct errors. Various techniques exists to dealing with these problems, you can have look at Yaffee (2003) A primer for panel data analysis or Pesaran (2015) Time series and panel data econometrics for an overview of various ways of dealing with this issue.

2. So, we must, at least cluster by country in the case above? I do not fully understand the argument.

   Again you do not have to cluster specifically just adjust for heteroskedasticity and autocorrelation somehow. The comment merely points out that the it would be surprising not find some heteroskedasticity or autocorrelation at country level. For example, if in one country economy is booming then it would not be surprising to find that year after year assets of companies in that country grow in value and so there will be an autocorrelation for firms assets at an country level. Or as that comment pointed out if the laws are not applied in exactly the same way across all industries there might be heteroskedasticity or autocorrelation at an industry level.